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,
632 const_debug unsigned int sysctl_sched_features =
633 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
634 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
635 SCHED_FEAT_START_DEBIT * 1 |
636 SCHED_FEAT_HRTICK * 1 |
637 SCHED_FEAT_DOUBLE_TICK * 0;
639 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
642 * Number of tasks to iterate in a single balance run.
643 * Limited because this is done with IRQs disabled.
645 const_debug unsigned int sysctl_sched_nr_migrate = 32;
648 * period over which we measure -rt task cpu usage in us.
651 unsigned int sysctl_sched_rt_period = 1000000;
653 static __read_mostly int scheduler_running;
656 * part of the period that we allow rt tasks to run in us.
659 int sysctl_sched_rt_runtime = 950000;
662 * single value that denotes runtime == period, ie unlimited time.
664 #define RUNTIME_INF ((u64)~0ULL)
666 static const unsigned long long time_sync_thresh = 100000;
668 static DEFINE_PER_CPU(unsigned long long, time_offset);
669 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
672 * Global lock which we take every now and then to synchronize
673 * the CPUs time. This method is not warp-safe, but it's good
674 * enough to synchronize slowly diverging time sources and thus
675 * it's good enough for tracing:
677 static DEFINE_SPINLOCK(time_sync_lock);
678 static unsigned long long prev_global_time;
680 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
684 spin_lock_irqsave(&time_sync_lock, flags);
686 if (time < prev_global_time) {
687 per_cpu(time_offset, cpu) += prev_global_time - time;
688 time = prev_global_time;
690 prev_global_time = time;
693 spin_unlock_irqrestore(&time_sync_lock, flags);
698 static unsigned long long __cpu_clock(int cpu)
700 unsigned long long now;
705 * Only call sched_clock() if the scheduler has already been
706 * initialized (some code might call cpu_clock() very early):
708 if (unlikely(!scheduler_running))
711 local_irq_save(flags);
715 local_irq_restore(flags);
721 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
722 * clock constructed from sched_clock():
724 unsigned long long cpu_clock(int cpu)
726 unsigned long long prev_cpu_time, time, delta_time;
728 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
729 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
730 delta_time = time-prev_cpu_time;
732 if (unlikely(delta_time > time_sync_thresh))
733 time = __sync_cpu_clock(time, cpu);
737 EXPORT_SYMBOL_GPL(cpu_clock);
739 #ifndef prepare_arch_switch
740 # define prepare_arch_switch(next) do { } while (0)
742 #ifndef finish_arch_switch
743 # define finish_arch_switch(prev) do { } while (0)
746 static inline int task_current(struct rq *rq, struct task_struct *p)
748 return rq->curr == p;
751 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
752 static inline int task_running(struct rq *rq, struct task_struct *p)
754 return task_current(rq, p);
757 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
761 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
763 #ifdef CONFIG_DEBUG_SPINLOCK
764 /* this is a valid case when another task releases the spinlock */
765 rq->lock.owner = current;
768 * If we are tracking spinlock dependencies then we have to
769 * fix up the runqueue lock - which gets 'carried over' from
772 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
774 spin_unlock_irq(&rq->lock);
777 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
778 static inline int task_running(struct rq *rq, struct task_struct *p)
783 return task_current(rq, p);
787 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
791 * We can optimise this out completely for !SMP, because the
792 * SMP rebalancing from interrupt is the only thing that cares
797 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
798 spin_unlock_irq(&rq->lock);
800 spin_unlock(&rq->lock);
804 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
808 * After ->oncpu is cleared, the task can be moved to a different CPU.
809 * We must ensure this doesn't happen until the switch is completely
815 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
819 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
822 * __task_rq_lock - lock the runqueue a given task resides on.
823 * Must be called interrupts disabled.
825 static inline struct rq *__task_rq_lock(struct task_struct *p)
829 struct rq *rq = task_rq(p);
830 spin_lock(&rq->lock);
831 if (likely(rq == task_rq(p)))
833 spin_unlock(&rq->lock);
838 * task_rq_lock - lock the runqueue a given task resides on and disable
839 * interrupts. Note the ordering: we can safely lookup the task_rq without
840 * explicitly disabling preemption.
842 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
848 local_irq_save(*flags);
850 spin_lock(&rq->lock);
851 if (likely(rq == task_rq(p)))
853 spin_unlock_irqrestore(&rq->lock, *flags);
857 static void __task_rq_unlock(struct rq *rq)
860 spin_unlock(&rq->lock);
863 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
866 spin_unlock_irqrestore(&rq->lock, *flags);
870 * this_rq_lock - lock this runqueue and disable interrupts.
872 static struct rq *this_rq_lock(void)
879 spin_lock(&rq->lock);
885 * We are going deep-idle (irqs are disabled):
887 void sched_clock_idle_sleep_event(void)
889 struct rq *rq = cpu_rq(smp_processor_id());
891 spin_lock(&rq->lock);
892 __update_rq_clock(rq);
893 spin_unlock(&rq->lock);
894 rq->clock_deep_idle_events++;
896 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
899 * We just idled delta nanoseconds (called with irqs disabled):
901 void sched_clock_idle_wakeup_event(u64 delta_ns)
903 struct rq *rq = cpu_rq(smp_processor_id());
904 u64 now = sched_clock();
906 rq->idle_clock += delta_ns;
908 * Override the previous timestamp and ignore all
909 * sched_clock() deltas that occured while we idled,
910 * and use the PM-provided delta_ns to advance the
913 spin_lock(&rq->lock);
914 rq->prev_clock_raw = now;
915 rq->clock += delta_ns;
916 spin_unlock(&rq->lock);
917 touch_softlockup_watchdog();
919 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
921 static void __resched_task(struct task_struct *p, int tif_bit);
923 static inline void resched_task(struct task_struct *p)
925 __resched_task(p, TIF_NEED_RESCHED);
928 #ifdef CONFIG_SCHED_HRTICK
930 * Use HR-timers to deliver accurate preemption points.
932 * Its all a bit involved since we cannot program an hrt while holding the
933 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
936 * When we get rescheduled we reprogram the hrtick_timer outside of the
939 static inline void resched_hrt(struct task_struct *p)
941 __resched_task(p, TIF_HRTICK_RESCHED);
944 static inline void resched_rq(struct rq *rq)
948 spin_lock_irqsave(&rq->lock, flags);
949 resched_task(rq->curr);
950 spin_unlock_irqrestore(&rq->lock, flags);
954 HRTICK_SET, /* re-programm hrtick_timer */
955 HRTICK_RESET, /* not a new slice */
960 * - enabled by features
961 * - hrtimer is actually high res
963 static inline int hrtick_enabled(struct rq *rq)
965 if (!sched_feat(HRTICK))
967 return hrtimer_is_hres_active(&rq->hrtick_timer);
971 * Called to set the hrtick timer state.
973 * called with rq->lock held and irqs disabled
975 static void hrtick_start(struct rq *rq, u64 delay, int reset)
977 assert_spin_locked(&rq->lock);
980 * preempt at: now + delay
983 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
985 * indicate we need to program the timer
987 __set_bit(HRTICK_SET, &rq->hrtick_flags);
989 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
992 * New slices are called from the schedule path and don't need a
996 resched_hrt(rq->curr);
999 static void hrtick_clear(struct rq *rq)
1001 if (hrtimer_active(&rq->hrtick_timer))
1002 hrtimer_cancel(&rq->hrtick_timer);
1006 * Update the timer from the possible pending state.
1008 static void hrtick_set(struct rq *rq)
1012 unsigned long flags;
1014 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1016 spin_lock_irqsave(&rq->lock, flags);
1017 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1018 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1019 time = rq->hrtick_expire;
1020 clear_thread_flag(TIF_HRTICK_RESCHED);
1021 spin_unlock_irqrestore(&rq->lock, flags);
1024 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1025 if (reset && !hrtimer_active(&rq->hrtick_timer))
1032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1035 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1037 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1039 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1041 spin_lock(&rq->lock);
1042 __update_rq_clock(rq);
1043 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1044 spin_unlock(&rq->lock);
1046 return HRTIMER_NORESTART;
1049 static inline void init_rq_hrtick(struct rq *rq)
1051 rq->hrtick_flags = 0;
1052 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1053 rq->hrtick_timer.function = hrtick;
1054 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1057 void hrtick_resched(void)
1060 unsigned long flags;
1062 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1065 local_irq_save(flags);
1066 rq = cpu_rq(smp_processor_id());
1068 local_irq_restore(flags);
1071 static inline void hrtick_clear(struct rq *rq)
1075 static inline void hrtick_set(struct rq *rq)
1079 static inline void init_rq_hrtick(struct rq *rq)
1083 void hrtick_resched(void)
1089 * resched_task - mark a task 'to be rescheduled now'.
1091 * On UP this means the setting of the need_resched flag, on SMP it
1092 * might also involve a cross-CPU call to trigger the scheduler on
1097 #ifndef tsk_is_polling
1098 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1101 static void __resched_task(struct task_struct *p, int tif_bit)
1105 assert_spin_locked(&task_rq(p)->lock);
1107 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1110 set_tsk_thread_flag(p, tif_bit);
1113 if (cpu == smp_processor_id())
1116 /* NEED_RESCHED must be visible before we test polling */
1118 if (!tsk_is_polling(p))
1119 smp_send_reschedule(cpu);
1122 static void resched_cpu(int cpu)
1124 struct rq *rq = cpu_rq(cpu);
1125 unsigned long flags;
1127 if (!spin_trylock_irqsave(&rq->lock, flags))
1129 resched_task(cpu_curr(cpu));
1130 spin_unlock_irqrestore(&rq->lock, flags);
1135 * When add_timer_on() enqueues a timer into the timer wheel of an
1136 * idle CPU then this timer might expire before the next timer event
1137 * which is scheduled to wake up that CPU. In case of a completely
1138 * idle system the next event might even be infinite time into the
1139 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1140 * leaves the inner idle loop so the newly added timer is taken into
1141 * account when the CPU goes back to idle and evaluates the timer
1142 * wheel for the next timer event.
1144 void wake_up_idle_cpu(int cpu)
1146 struct rq *rq = cpu_rq(cpu);
1148 if (cpu == smp_processor_id())
1152 * This is safe, as this function is called with the timer
1153 * wheel base lock of (cpu) held. When the CPU is on the way
1154 * to idle and has not yet set rq->curr to idle then it will
1155 * be serialized on the timer wheel base lock and take the new
1156 * timer into account automatically.
1158 if (rq->curr != rq->idle)
1162 * We can set TIF_RESCHED on the idle task of the other CPU
1163 * lockless. The worst case is that the other CPU runs the
1164 * idle task through an additional NOOP schedule()
1166 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1168 /* NEED_RESCHED must be visible before we test polling */
1170 if (!tsk_is_polling(rq->idle))
1171 smp_send_reschedule(cpu);
1176 static void __resched_task(struct task_struct *p, int tif_bit)
1178 assert_spin_locked(&task_rq(p)->lock);
1179 set_tsk_thread_flag(p, tif_bit);
1183 #if BITS_PER_LONG == 32
1184 # define WMULT_CONST (~0UL)
1186 # define WMULT_CONST (1UL << 32)
1189 #define WMULT_SHIFT 32
1192 * Shift right and round:
1194 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1196 static unsigned long
1197 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1198 struct load_weight *lw)
1202 if (unlikely(!lw->inv_weight))
1203 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1205 tmp = (u64)delta_exec * weight;
1207 * Check whether we'd overflow the 64-bit multiplication:
1209 if (unlikely(tmp > WMULT_CONST))
1210 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1213 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1215 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1218 static inline unsigned long
1219 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1221 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1224 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1230 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1237 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1238 * of tasks with abnormal "nice" values across CPUs the contribution that
1239 * each task makes to its run queue's load is weighted according to its
1240 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1241 * scaled version of the new time slice allocation that they receive on time
1245 #define WEIGHT_IDLEPRIO 2
1246 #define WMULT_IDLEPRIO (1 << 31)
1249 * Nice levels are multiplicative, with a gentle 10% change for every
1250 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1251 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1252 * that remained on nice 0.
1254 * The "10% effect" is relative and cumulative: from _any_ nice level,
1255 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1256 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1257 * If a task goes up by ~10% and another task goes down by ~10% then
1258 * the relative distance between them is ~25%.)
1260 static const int prio_to_weight[40] = {
1261 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1262 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1263 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1264 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1265 /* 0 */ 1024, 820, 655, 526, 423,
1266 /* 5 */ 335, 272, 215, 172, 137,
1267 /* 10 */ 110, 87, 70, 56, 45,
1268 /* 15 */ 36, 29, 23, 18, 15,
1272 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1274 * In cases where the weight does not change often, we can use the
1275 * precalculated inverse to speed up arithmetics by turning divisions
1276 * into multiplications:
1278 static const u32 prio_to_wmult[40] = {
1279 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1280 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1281 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1282 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1283 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1284 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1285 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1286 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1289 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1292 * runqueue iterator, to support SMP load-balancing between different
1293 * scheduling classes, without having to expose their internal data
1294 * structures to the load-balancing proper:
1296 struct rq_iterator {
1298 struct task_struct *(*start)(void *);
1299 struct task_struct *(*next)(void *);
1303 static unsigned long
1304 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1305 unsigned long max_load_move, struct sched_domain *sd,
1306 enum cpu_idle_type idle, int *all_pinned,
1307 int *this_best_prio, struct rq_iterator *iterator);
1310 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1311 struct sched_domain *sd, enum cpu_idle_type idle,
1312 struct rq_iterator *iterator);
1315 #ifdef CONFIG_CGROUP_CPUACCT
1316 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1318 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1322 static unsigned long source_load(int cpu, int type);
1323 static unsigned long target_load(int cpu, int type);
1324 static unsigned long cpu_avg_load_per_task(int cpu);
1325 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1326 #endif /* CONFIG_SMP */
1328 #include "sched_stats.h"
1329 #include "sched_idletask.c"
1330 #include "sched_fair.c"
1331 #include "sched_rt.c"
1332 #ifdef CONFIG_SCHED_DEBUG
1333 # include "sched_debug.c"
1336 #define sched_class_highest (&rt_sched_class)
1338 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1340 update_load_add(&rq->load, p->se.load.weight);
1343 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1345 update_load_sub(&rq->load, p->se.load.weight);
1348 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1354 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1360 static void set_load_weight(struct task_struct *p)
1362 if (task_has_rt_policy(p)) {
1363 p->se.load.weight = prio_to_weight[0] * 2;
1364 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1369 * SCHED_IDLE tasks get minimal weight:
1371 if (p->policy == SCHED_IDLE) {
1372 p->se.load.weight = WEIGHT_IDLEPRIO;
1373 p->se.load.inv_weight = WMULT_IDLEPRIO;
1377 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1378 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1381 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1383 sched_info_queued(p);
1384 p->sched_class->enqueue_task(rq, p, wakeup);
1388 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1390 p->sched_class->dequeue_task(rq, p, sleep);
1395 * __normal_prio - return the priority that is based on the static prio
1397 static inline int __normal_prio(struct task_struct *p)
1399 return p->static_prio;
1403 * Calculate the expected normal priority: i.e. priority
1404 * without taking RT-inheritance into account. Might be
1405 * boosted by interactivity modifiers. Changes upon fork,
1406 * setprio syscalls, and whenever the interactivity
1407 * estimator recalculates.
1409 static inline int normal_prio(struct task_struct *p)
1413 if (task_has_rt_policy(p))
1414 prio = MAX_RT_PRIO-1 - p->rt_priority;
1416 prio = __normal_prio(p);
1421 * Calculate the current priority, i.e. the priority
1422 * taken into account by the scheduler. This value might
1423 * be boosted by RT tasks, or might be boosted by
1424 * interactivity modifiers. Will be RT if the task got
1425 * RT-boosted. If not then it returns p->normal_prio.
1427 static int effective_prio(struct task_struct *p)
1429 p->normal_prio = normal_prio(p);
1431 * If we are RT tasks or we were boosted to RT priority,
1432 * keep the priority unchanged. Otherwise, update priority
1433 * to the normal priority:
1435 if (!rt_prio(p->prio))
1436 return p->normal_prio;
1441 * activate_task - move a task to the runqueue.
1443 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1445 if (task_contributes_to_load(p))
1446 rq->nr_uninterruptible--;
1448 enqueue_task(rq, p, wakeup);
1449 inc_nr_running(p, rq);
1453 * deactivate_task - remove a task from the runqueue.
1455 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1457 if (task_contributes_to_load(p))
1458 rq->nr_uninterruptible++;
1460 dequeue_task(rq, p, sleep);
1461 dec_nr_running(p, rq);
1465 * task_curr - is this task currently executing on a CPU?
1466 * @p: the task in question.
1468 inline int task_curr(const struct task_struct *p)
1470 return cpu_curr(task_cpu(p)) == p;
1473 /* Used instead of source_load when we know the type == 0 */
1474 unsigned long weighted_cpuload(const int cpu)
1476 return cpu_rq(cpu)->load.weight;
1479 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1481 set_task_rq(p, cpu);
1484 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1485 * successfuly executed on another CPU. We must ensure that updates of
1486 * per-task data have been completed by this moment.
1489 task_thread_info(p)->cpu = cpu;
1493 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1494 const struct sched_class *prev_class,
1495 int oldprio, int running)
1497 if (prev_class != p->sched_class) {
1498 if (prev_class->switched_from)
1499 prev_class->switched_from(rq, p, running);
1500 p->sched_class->switched_to(rq, p, running);
1502 p->sched_class->prio_changed(rq, p, oldprio, running);
1508 * Is this task likely cache-hot:
1511 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1516 * Buddy candidates are cache hot:
1518 if (&p->se == cfs_rq_of(&p->se)->next)
1521 if (p->sched_class != &fair_sched_class)
1524 if (sysctl_sched_migration_cost == -1)
1526 if (sysctl_sched_migration_cost == 0)
1529 delta = now - p->se.exec_start;
1531 return delta < (s64)sysctl_sched_migration_cost;
1535 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1537 int old_cpu = task_cpu(p);
1538 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1539 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1540 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1543 clock_offset = old_rq->clock - new_rq->clock;
1545 #ifdef CONFIG_SCHEDSTATS
1546 if (p->se.wait_start)
1547 p->se.wait_start -= clock_offset;
1548 if (p->se.sleep_start)
1549 p->se.sleep_start -= clock_offset;
1550 if (p->se.block_start)
1551 p->se.block_start -= clock_offset;
1552 if (old_cpu != new_cpu) {
1553 schedstat_inc(p, se.nr_migrations);
1554 if (task_hot(p, old_rq->clock, NULL))
1555 schedstat_inc(p, se.nr_forced2_migrations);
1558 p->se.vruntime -= old_cfsrq->min_vruntime -
1559 new_cfsrq->min_vruntime;
1561 __set_task_cpu(p, new_cpu);
1564 struct migration_req {
1565 struct list_head list;
1567 struct task_struct *task;
1570 struct completion done;
1574 * The task's runqueue lock must be held.
1575 * Returns true if you have to wait for migration thread.
1578 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1580 struct rq *rq = task_rq(p);
1583 * If the task is not on a runqueue (and not running), then
1584 * it is sufficient to simply update the task's cpu field.
1586 if (!p->se.on_rq && !task_running(rq, p)) {
1587 set_task_cpu(p, dest_cpu);
1591 init_completion(&req->done);
1593 req->dest_cpu = dest_cpu;
1594 list_add(&req->list, &rq->migration_queue);
1600 * wait_task_inactive - wait for a thread to unschedule.
1602 * The caller must ensure that the task *will* unschedule sometime soon,
1603 * else this function might spin for a *long* time. This function can't
1604 * be called with interrupts off, or it may introduce deadlock with
1605 * smp_call_function() if an IPI is sent by the same process we are
1606 * waiting to become inactive.
1608 void wait_task_inactive(struct task_struct *p)
1610 unsigned long flags;
1616 * We do the initial early heuristics without holding
1617 * any task-queue locks at all. We'll only try to get
1618 * the runqueue lock when things look like they will
1624 * If the task is actively running on another CPU
1625 * still, just relax and busy-wait without holding
1628 * NOTE! Since we don't hold any locks, it's not
1629 * even sure that "rq" stays as the right runqueue!
1630 * But we don't care, since "task_running()" will
1631 * return false if the runqueue has changed and p
1632 * is actually now running somewhere else!
1634 while (task_running(rq, p))
1638 * Ok, time to look more closely! We need the rq
1639 * lock now, to be *sure*. If we're wrong, we'll
1640 * just go back and repeat.
1642 rq = task_rq_lock(p, &flags);
1643 running = task_running(rq, p);
1644 on_rq = p->se.on_rq;
1645 task_rq_unlock(rq, &flags);
1648 * Was it really running after all now that we
1649 * checked with the proper locks actually held?
1651 * Oops. Go back and try again..
1653 if (unlikely(running)) {
1659 * It's not enough that it's not actively running,
1660 * it must be off the runqueue _entirely_, and not
1663 * So if it wa still runnable (but just not actively
1664 * running right now), it's preempted, and we should
1665 * yield - it could be a while.
1667 if (unlikely(on_rq)) {
1668 schedule_timeout_uninterruptible(1);
1673 * Ahh, all good. It wasn't running, and it wasn't
1674 * runnable, which means that it will never become
1675 * running in the future either. We're all done!
1682 * kick_process - kick a running thread to enter/exit the kernel
1683 * @p: the to-be-kicked thread
1685 * Cause a process which is running on another CPU to enter
1686 * kernel-mode, without any delay. (to get signals handled.)
1688 * NOTE: this function doesnt have to take the runqueue lock,
1689 * because all it wants to ensure is that the remote task enters
1690 * the kernel. If the IPI races and the task has been migrated
1691 * to another CPU then no harm is done and the purpose has been
1694 void kick_process(struct task_struct *p)
1700 if ((cpu != smp_processor_id()) && task_curr(p))
1701 smp_send_reschedule(cpu);
1706 * Return a low guess at the load of a migration-source cpu weighted
1707 * according to the scheduling class and "nice" value.
1709 * We want to under-estimate the load of migration sources, to
1710 * balance conservatively.
1712 static unsigned long source_load(int cpu, int type)
1714 struct rq *rq = cpu_rq(cpu);
1715 unsigned long total = weighted_cpuload(cpu);
1720 return min(rq->cpu_load[type-1], total);
1724 * Return a high guess at the load of a migration-target cpu weighted
1725 * according to the scheduling class and "nice" value.
1727 static unsigned long target_load(int cpu, int type)
1729 struct rq *rq = cpu_rq(cpu);
1730 unsigned long total = weighted_cpuload(cpu);
1735 return max(rq->cpu_load[type-1], total);
1739 * Return the average load per task on the cpu's run queue
1741 static unsigned long cpu_avg_load_per_task(int cpu)
1743 struct rq *rq = cpu_rq(cpu);
1744 unsigned long total = weighted_cpuload(cpu);
1745 unsigned long n = rq->nr_running;
1747 return n ? total / n : SCHED_LOAD_SCALE;
1751 * find_idlest_group finds and returns the least busy CPU group within the
1754 static struct sched_group *
1755 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1757 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1758 unsigned long min_load = ULONG_MAX, this_load = 0;
1759 int load_idx = sd->forkexec_idx;
1760 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1763 unsigned long load, avg_load;
1767 /* Skip over this group if it has no CPUs allowed */
1768 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1771 local_group = cpu_isset(this_cpu, group->cpumask);
1773 /* Tally up the load of all CPUs in the group */
1776 for_each_cpu_mask(i, group->cpumask) {
1777 /* Bias balancing toward cpus of our domain */
1779 load = source_load(i, load_idx);
1781 load = target_load(i, load_idx);
1786 /* Adjust by relative CPU power of the group */
1787 avg_load = sg_div_cpu_power(group,
1788 avg_load * SCHED_LOAD_SCALE);
1791 this_load = avg_load;
1793 } else if (avg_load < min_load) {
1794 min_load = avg_load;
1797 } while (group = group->next, group != sd->groups);
1799 if (!idlest || 100*this_load < imbalance*min_load)
1805 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1808 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1811 unsigned long load, min_load = ULONG_MAX;
1815 /* Traverse only the allowed CPUs */
1816 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1818 for_each_cpu_mask(i, tmp) {
1819 load = weighted_cpuload(i);
1821 if (load < min_load || (load == min_load && i == this_cpu)) {
1831 * sched_balance_self: balance the current task (running on cpu) in domains
1832 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1835 * Balance, ie. select the least loaded group.
1837 * Returns the target CPU number, or the same CPU if no balancing is needed.
1839 * preempt must be disabled.
1841 static int sched_balance_self(int cpu, int flag)
1843 struct task_struct *t = current;
1844 struct sched_domain *tmp, *sd = NULL;
1846 for_each_domain(cpu, tmp) {
1848 * If power savings logic is enabled for a domain, stop there.
1850 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1852 if (tmp->flags & flag)
1858 struct sched_group *group;
1859 int new_cpu, weight;
1861 if (!(sd->flags & flag)) {
1867 group = find_idlest_group(sd, t, cpu);
1873 new_cpu = find_idlest_cpu(group, t, cpu);
1874 if (new_cpu == -1 || new_cpu == cpu) {
1875 /* Now try balancing at a lower domain level of cpu */
1880 /* Now try balancing at a lower domain level of new_cpu */
1883 weight = cpus_weight(span);
1884 for_each_domain(cpu, tmp) {
1885 if (weight <= cpus_weight(tmp->span))
1887 if (tmp->flags & flag)
1890 /* while loop will break here if sd == NULL */
1896 #endif /* CONFIG_SMP */
1899 * try_to_wake_up - wake up a thread
1900 * @p: the to-be-woken-up thread
1901 * @state: the mask of task states that can be woken
1902 * @sync: do a synchronous wakeup?
1904 * Put it on the run-queue if it's not already there. The "current"
1905 * thread is always on the run-queue (except when the actual
1906 * re-schedule is in progress), and as such you're allowed to do
1907 * the simpler "current->state = TASK_RUNNING" to mark yourself
1908 * runnable without the overhead of this.
1910 * returns failure only if the task is already active.
1912 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1914 int cpu, orig_cpu, this_cpu, success = 0;
1915 unsigned long flags;
1920 rq = task_rq_lock(p, &flags);
1921 old_state = p->state;
1922 if (!(old_state & state))
1930 this_cpu = smp_processor_id();
1933 if (unlikely(task_running(rq, p)))
1936 cpu = p->sched_class->select_task_rq(p, sync);
1937 if (cpu != orig_cpu) {
1938 set_task_cpu(p, cpu);
1939 task_rq_unlock(rq, &flags);
1940 /* might preempt at this point */
1941 rq = task_rq_lock(p, &flags);
1942 old_state = p->state;
1943 if (!(old_state & state))
1948 this_cpu = smp_processor_id();
1952 #ifdef CONFIG_SCHEDSTATS
1953 schedstat_inc(rq, ttwu_count);
1954 if (cpu == this_cpu)
1955 schedstat_inc(rq, ttwu_local);
1957 struct sched_domain *sd;
1958 for_each_domain(this_cpu, sd) {
1959 if (cpu_isset(cpu, sd->span)) {
1960 schedstat_inc(sd, ttwu_wake_remote);
1968 #endif /* CONFIG_SMP */
1969 schedstat_inc(p, se.nr_wakeups);
1971 schedstat_inc(p, se.nr_wakeups_sync);
1972 if (orig_cpu != cpu)
1973 schedstat_inc(p, se.nr_wakeups_migrate);
1974 if (cpu == this_cpu)
1975 schedstat_inc(p, se.nr_wakeups_local);
1977 schedstat_inc(p, se.nr_wakeups_remote);
1978 update_rq_clock(rq);
1979 activate_task(rq, p, 1);
1983 check_preempt_curr(rq, p);
1985 p->state = TASK_RUNNING;
1987 if (p->sched_class->task_wake_up)
1988 p->sched_class->task_wake_up(rq, p);
1991 task_rq_unlock(rq, &flags);
1996 int wake_up_process(struct task_struct *p)
1998 return try_to_wake_up(p, TASK_ALL, 0);
2000 EXPORT_SYMBOL(wake_up_process);
2002 int wake_up_state(struct task_struct *p, unsigned int state)
2004 return try_to_wake_up(p, state, 0);
2008 * Perform scheduler related setup for a newly forked process p.
2009 * p is forked by current.
2011 * __sched_fork() is basic setup used by init_idle() too:
2013 static void __sched_fork(struct task_struct *p)
2015 p->se.exec_start = 0;
2016 p->se.sum_exec_runtime = 0;
2017 p->se.prev_sum_exec_runtime = 0;
2018 p->se.last_wakeup = 0;
2019 p->se.avg_overlap = 0;
2021 #ifdef CONFIG_SCHEDSTATS
2022 p->se.wait_start = 0;
2023 p->se.sum_sleep_runtime = 0;
2024 p->se.sleep_start = 0;
2025 p->se.block_start = 0;
2026 p->se.sleep_max = 0;
2027 p->se.block_max = 0;
2029 p->se.slice_max = 0;
2033 INIT_LIST_HEAD(&p->rt.run_list);
2036 #ifdef CONFIG_PREEMPT_NOTIFIERS
2037 INIT_HLIST_HEAD(&p->preempt_notifiers);
2041 * We mark the process as running here, but have not actually
2042 * inserted it onto the runqueue yet. This guarantees that
2043 * nobody will actually run it, and a signal or other external
2044 * event cannot wake it up and insert it on the runqueue either.
2046 p->state = TASK_RUNNING;
2050 * fork()/clone()-time setup:
2052 void sched_fork(struct task_struct *p, int clone_flags)
2054 int cpu = get_cpu();
2059 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2061 set_task_cpu(p, cpu);
2064 * Make sure we do not leak PI boosting priority to the child:
2066 p->prio = current->normal_prio;
2067 if (!rt_prio(p->prio))
2068 p->sched_class = &fair_sched_class;
2070 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2071 if (likely(sched_info_on()))
2072 memset(&p->sched_info, 0, sizeof(p->sched_info));
2074 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2077 #ifdef CONFIG_PREEMPT
2078 /* Want to start with kernel preemption disabled. */
2079 task_thread_info(p)->preempt_count = 1;
2085 * wake_up_new_task - wake up a newly created task for the first time.
2087 * This function will do some initial scheduler statistics housekeeping
2088 * that must be done for every newly created context, then puts the task
2089 * on the runqueue and wakes it.
2091 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2093 unsigned long flags;
2096 rq = task_rq_lock(p, &flags);
2097 BUG_ON(p->state != TASK_RUNNING);
2098 update_rq_clock(rq);
2100 p->prio = effective_prio(p);
2102 if (!p->sched_class->task_new || !current->se.on_rq) {
2103 activate_task(rq, p, 0);
2106 * Let the scheduling class do new task startup
2107 * management (if any):
2109 p->sched_class->task_new(rq, p);
2110 inc_nr_running(p, rq);
2112 check_preempt_curr(rq, p);
2114 if (p->sched_class->task_wake_up)
2115 p->sched_class->task_wake_up(rq, p);
2117 task_rq_unlock(rq, &flags);
2120 #ifdef CONFIG_PREEMPT_NOTIFIERS
2123 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2124 * @notifier: notifier struct to register
2126 void preempt_notifier_register(struct preempt_notifier *notifier)
2128 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2130 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2133 * preempt_notifier_unregister - no longer interested in preemption notifications
2134 * @notifier: notifier struct to unregister
2136 * This is safe to call from within a preemption notifier.
2138 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2140 hlist_del(¬ifier->link);
2142 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2144 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2146 struct preempt_notifier *notifier;
2147 struct hlist_node *node;
2149 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2150 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2154 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2155 struct task_struct *next)
2157 struct preempt_notifier *notifier;
2158 struct hlist_node *node;
2160 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2161 notifier->ops->sched_out(notifier, next);
2166 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2171 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2172 struct task_struct *next)
2179 * prepare_task_switch - prepare to switch tasks
2180 * @rq: the runqueue preparing to switch
2181 * @prev: the current task that is being switched out
2182 * @next: the task we are going to switch to.
2184 * This is called with the rq lock held and interrupts off. It must
2185 * be paired with a subsequent finish_task_switch after the context
2188 * prepare_task_switch sets up locking and calls architecture specific
2192 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2193 struct task_struct *next)
2195 fire_sched_out_preempt_notifiers(prev, next);
2196 prepare_lock_switch(rq, next);
2197 prepare_arch_switch(next);
2201 * finish_task_switch - clean up after a task-switch
2202 * @rq: runqueue associated with task-switch
2203 * @prev: the thread we just switched away from.
2205 * finish_task_switch must be called after the context switch, paired
2206 * with a prepare_task_switch call before the context switch.
2207 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2208 * and do any other architecture-specific cleanup actions.
2210 * Note that we may have delayed dropping an mm in context_switch(). If
2211 * so, we finish that here outside of the runqueue lock. (Doing it
2212 * with the lock held can cause deadlocks; see schedule() for
2215 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2216 __releases(rq->lock)
2218 struct mm_struct *mm = rq->prev_mm;
2224 * A task struct has one reference for the use as "current".
2225 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2226 * schedule one last time. The schedule call will never return, and
2227 * the scheduled task must drop that reference.
2228 * The test for TASK_DEAD must occur while the runqueue locks are
2229 * still held, otherwise prev could be scheduled on another cpu, die
2230 * there before we look at prev->state, and then the reference would
2232 * Manfred Spraul <manfred@colorfullife.com>
2234 prev_state = prev->state;
2235 finish_arch_switch(prev);
2236 finish_lock_switch(rq, prev);
2238 if (current->sched_class->post_schedule)
2239 current->sched_class->post_schedule(rq);
2242 fire_sched_in_preempt_notifiers(current);
2245 if (unlikely(prev_state == TASK_DEAD)) {
2247 * Remove function-return probe instances associated with this
2248 * task and put them back on the free list.
2250 kprobe_flush_task(prev);
2251 put_task_struct(prev);
2256 * schedule_tail - first thing a freshly forked thread must call.
2257 * @prev: the thread we just switched away from.
2259 asmlinkage void schedule_tail(struct task_struct *prev)
2260 __releases(rq->lock)
2262 struct rq *rq = this_rq();
2264 finish_task_switch(rq, prev);
2265 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2266 /* In this case, finish_task_switch does not reenable preemption */
2269 if (current->set_child_tid)
2270 put_user(task_pid_vnr(current), current->set_child_tid);
2274 * context_switch - switch to the new MM and the new
2275 * thread's register state.
2278 context_switch(struct rq *rq, struct task_struct *prev,
2279 struct task_struct *next)
2281 struct mm_struct *mm, *oldmm;
2283 prepare_task_switch(rq, prev, next);
2285 oldmm = prev->active_mm;
2287 * For paravirt, this is coupled with an exit in switch_to to
2288 * combine the page table reload and the switch backend into
2291 arch_enter_lazy_cpu_mode();
2293 if (unlikely(!mm)) {
2294 next->active_mm = oldmm;
2295 atomic_inc(&oldmm->mm_count);
2296 enter_lazy_tlb(oldmm, next);
2298 switch_mm(oldmm, mm, next);
2300 if (unlikely(!prev->mm)) {
2301 prev->active_mm = NULL;
2302 rq->prev_mm = oldmm;
2305 * Since the runqueue lock will be released by the next
2306 * task (which is an invalid locking op but in the case
2307 * of the scheduler it's an obvious special-case), so we
2308 * do an early lockdep release here:
2310 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2311 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2314 /* Here we just switch the register state and the stack. */
2315 switch_to(prev, next, prev);
2319 * this_rq must be evaluated again because prev may have moved
2320 * CPUs since it called schedule(), thus the 'rq' on its stack
2321 * frame will be invalid.
2323 finish_task_switch(this_rq(), prev);
2327 * nr_running, nr_uninterruptible and nr_context_switches:
2329 * externally visible scheduler statistics: current number of runnable
2330 * threads, current number of uninterruptible-sleeping threads, total
2331 * number of context switches performed since bootup.
2333 unsigned long nr_running(void)
2335 unsigned long i, sum = 0;
2337 for_each_online_cpu(i)
2338 sum += cpu_rq(i)->nr_running;
2343 unsigned long nr_uninterruptible(void)
2345 unsigned long i, sum = 0;
2347 for_each_possible_cpu(i)
2348 sum += cpu_rq(i)->nr_uninterruptible;
2351 * Since we read the counters lockless, it might be slightly
2352 * inaccurate. Do not allow it to go below zero though:
2354 if (unlikely((long)sum < 0))
2360 unsigned long long nr_context_switches(void)
2363 unsigned long long sum = 0;
2365 for_each_possible_cpu(i)
2366 sum += cpu_rq(i)->nr_switches;
2371 unsigned long nr_iowait(void)
2373 unsigned long i, sum = 0;
2375 for_each_possible_cpu(i)
2376 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2381 unsigned long nr_active(void)
2383 unsigned long i, running = 0, uninterruptible = 0;
2385 for_each_online_cpu(i) {
2386 running += cpu_rq(i)->nr_running;
2387 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2390 if (unlikely((long)uninterruptible < 0))
2391 uninterruptible = 0;
2393 return running + uninterruptible;
2397 * Update rq->cpu_load[] statistics. This function is usually called every
2398 * scheduler tick (TICK_NSEC).
2400 static void update_cpu_load(struct rq *this_rq)
2402 unsigned long this_load = this_rq->load.weight;
2405 this_rq->nr_load_updates++;
2407 /* Update our load: */
2408 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2409 unsigned long old_load, new_load;
2411 /* scale is effectively 1 << i now, and >> i divides by scale */
2413 old_load = this_rq->cpu_load[i];
2414 new_load = this_load;
2416 * Round up the averaging division if load is increasing. This
2417 * prevents us from getting stuck on 9 if the load is 10, for
2420 if (new_load > old_load)
2421 new_load += scale-1;
2422 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2429 * double_rq_lock - safely lock two runqueues
2431 * Note this does not disable interrupts like task_rq_lock,
2432 * you need to do so manually before calling.
2434 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2435 __acquires(rq1->lock)
2436 __acquires(rq2->lock)
2438 BUG_ON(!irqs_disabled());
2440 spin_lock(&rq1->lock);
2441 __acquire(rq2->lock); /* Fake it out ;) */
2444 spin_lock(&rq1->lock);
2445 spin_lock(&rq2->lock);
2447 spin_lock(&rq2->lock);
2448 spin_lock(&rq1->lock);
2451 update_rq_clock(rq1);
2452 update_rq_clock(rq2);
2456 * double_rq_unlock - safely unlock two runqueues
2458 * Note this does not restore interrupts like task_rq_unlock,
2459 * you need to do so manually after calling.
2461 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2462 __releases(rq1->lock)
2463 __releases(rq2->lock)
2465 spin_unlock(&rq1->lock);
2467 spin_unlock(&rq2->lock);
2469 __release(rq2->lock);
2473 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2475 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2476 __releases(this_rq->lock)
2477 __acquires(busiest->lock)
2478 __acquires(this_rq->lock)
2482 if (unlikely(!irqs_disabled())) {
2483 /* printk() doesn't work good under rq->lock */
2484 spin_unlock(&this_rq->lock);
2487 if (unlikely(!spin_trylock(&busiest->lock))) {
2488 if (busiest < this_rq) {
2489 spin_unlock(&this_rq->lock);
2490 spin_lock(&busiest->lock);
2491 spin_lock(&this_rq->lock);
2494 spin_lock(&busiest->lock);
2500 * If dest_cpu is allowed for this process, migrate the task to it.
2501 * This is accomplished by forcing the cpu_allowed mask to only
2502 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2503 * the cpu_allowed mask is restored.
2505 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2507 struct migration_req req;
2508 unsigned long flags;
2511 rq = task_rq_lock(p, &flags);
2512 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2513 || unlikely(cpu_is_offline(dest_cpu)))
2516 /* force the process onto the specified CPU */
2517 if (migrate_task(p, dest_cpu, &req)) {
2518 /* Need to wait for migration thread (might exit: take ref). */
2519 struct task_struct *mt = rq->migration_thread;
2521 get_task_struct(mt);
2522 task_rq_unlock(rq, &flags);
2523 wake_up_process(mt);
2524 put_task_struct(mt);
2525 wait_for_completion(&req.done);
2530 task_rq_unlock(rq, &flags);
2534 * sched_exec - execve() is a valuable balancing opportunity, because at
2535 * this point the task has the smallest effective memory and cache footprint.
2537 void sched_exec(void)
2539 int new_cpu, this_cpu = get_cpu();
2540 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2542 if (new_cpu != this_cpu)
2543 sched_migrate_task(current, new_cpu);
2547 * pull_task - move a task from a remote runqueue to the local runqueue.
2548 * Both runqueues must be locked.
2550 static void pull_task(struct rq *src_rq, struct task_struct *p,
2551 struct rq *this_rq, int this_cpu)
2553 deactivate_task(src_rq, p, 0);
2554 set_task_cpu(p, this_cpu);
2555 activate_task(this_rq, p, 0);
2557 * Note that idle threads have a prio of MAX_PRIO, for this test
2558 * to be always true for them.
2560 check_preempt_curr(this_rq, p);
2564 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2567 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2568 struct sched_domain *sd, enum cpu_idle_type idle,
2572 * We do not migrate tasks that are:
2573 * 1) running (obviously), or
2574 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2575 * 3) are cache-hot on their current CPU.
2577 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2578 schedstat_inc(p, se.nr_failed_migrations_affine);
2583 if (task_running(rq, p)) {
2584 schedstat_inc(p, se.nr_failed_migrations_running);
2589 * Aggressive migration if:
2590 * 1) task is cache cold, or
2591 * 2) too many balance attempts have failed.
2594 if (!task_hot(p, rq->clock, sd) ||
2595 sd->nr_balance_failed > sd->cache_nice_tries) {
2596 #ifdef CONFIG_SCHEDSTATS
2597 if (task_hot(p, rq->clock, sd)) {
2598 schedstat_inc(sd, lb_hot_gained[idle]);
2599 schedstat_inc(p, se.nr_forced_migrations);
2605 if (task_hot(p, rq->clock, sd)) {
2606 schedstat_inc(p, se.nr_failed_migrations_hot);
2612 static unsigned long
2613 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2614 unsigned long max_load_move, struct sched_domain *sd,
2615 enum cpu_idle_type idle, int *all_pinned,
2616 int *this_best_prio, struct rq_iterator *iterator)
2618 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2619 struct task_struct *p;
2620 long rem_load_move = max_load_move;
2622 if (max_load_move == 0)
2628 * Start the load-balancing iterator:
2630 p = iterator->start(iterator->arg);
2632 if (!p || loops++ > sysctl_sched_nr_migrate)
2635 * To help distribute high priority tasks across CPUs we don't
2636 * skip a task if it will be the highest priority task (i.e. smallest
2637 * prio value) on its new queue regardless of its load weight
2639 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2640 SCHED_LOAD_SCALE_FUZZ;
2641 if ((skip_for_load && p->prio >= *this_best_prio) ||
2642 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2643 p = iterator->next(iterator->arg);
2647 pull_task(busiest, p, this_rq, this_cpu);
2649 rem_load_move -= p->se.load.weight;
2652 * We only want to steal up to the prescribed amount of weighted load.
2654 if (rem_load_move > 0) {
2655 if (p->prio < *this_best_prio)
2656 *this_best_prio = p->prio;
2657 p = iterator->next(iterator->arg);
2662 * Right now, this is one of only two places pull_task() is called,
2663 * so we can safely collect pull_task() stats here rather than
2664 * inside pull_task().
2666 schedstat_add(sd, lb_gained[idle], pulled);
2669 *all_pinned = pinned;
2671 return max_load_move - rem_load_move;
2675 * move_tasks tries to move up to max_load_move weighted load from busiest to
2676 * this_rq, as part of a balancing operation within domain "sd".
2677 * Returns 1 if successful and 0 otherwise.
2679 * Called with both runqueues locked.
2681 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2682 unsigned long max_load_move,
2683 struct sched_domain *sd, enum cpu_idle_type idle,
2686 const struct sched_class *class = sched_class_highest;
2687 unsigned long total_load_moved = 0;
2688 int this_best_prio = this_rq->curr->prio;
2692 class->load_balance(this_rq, this_cpu, busiest,
2693 max_load_move - total_load_moved,
2694 sd, idle, all_pinned, &this_best_prio);
2695 class = class->next;
2696 } while (class && max_load_move > total_load_moved);
2698 return total_load_moved > 0;
2702 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2703 struct sched_domain *sd, enum cpu_idle_type idle,
2704 struct rq_iterator *iterator)
2706 struct task_struct *p = iterator->start(iterator->arg);
2710 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2711 pull_task(busiest, p, this_rq, this_cpu);
2713 * Right now, this is only the second place pull_task()
2714 * is called, so we can safely collect pull_task()
2715 * stats here rather than inside pull_task().
2717 schedstat_inc(sd, lb_gained[idle]);
2721 p = iterator->next(iterator->arg);
2728 * move_one_task tries to move exactly one task from busiest to this_rq, as
2729 * part of active balancing operations within "domain".
2730 * Returns 1 if successful and 0 otherwise.
2732 * Called with both runqueues locked.
2734 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2735 struct sched_domain *sd, enum cpu_idle_type idle)
2737 const struct sched_class *class;
2739 for (class = sched_class_highest; class; class = class->next)
2740 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2747 * find_busiest_group finds and returns the busiest CPU group within the
2748 * domain. It calculates and returns the amount of weighted load which
2749 * should be moved to restore balance via the imbalance parameter.
2751 static struct sched_group *
2752 find_busiest_group(struct sched_domain *sd, int this_cpu,
2753 unsigned long *imbalance, enum cpu_idle_type idle,
2754 int *sd_idle, cpumask_t *cpus, int *balance)
2756 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2757 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2758 unsigned long max_pull;
2759 unsigned long busiest_load_per_task, busiest_nr_running;
2760 unsigned long this_load_per_task, this_nr_running;
2761 int load_idx, group_imb = 0;
2762 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2763 int power_savings_balance = 1;
2764 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2765 unsigned long min_nr_running = ULONG_MAX;
2766 struct sched_group *group_min = NULL, *group_leader = NULL;
2769 max_load = this_load = total_load = total_pwr = 0;
2770 busiest_load_per_task = busiest_nr_running = 0;
2771 this_load_per_task = this_nr_running = 0;
2772 if (idle == CPU_NOT_IDLE)
2773 load_idx = sd->busy_idx;
2774 else if (idle == CPU_NEWLY_IDLE)
2775 load_idx = sd->newidle_idx;
2777 load_idx = sd->idle_idx;
2780 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2783 int __group_imb = 0;
2784 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2785 unsigned long sum_nr_running, sum_weighted_load;
2787 local_group = cpu_isset(this_cpu, group->cpumask);
2790 balance_cpu = first_cpu(group->cpumask);
2792 /* Tally up the load of all CPUs in the group */
2793 sum_weighted_load = sum_nr_running = avg_load = 0;
2795 min_cpu_load = ~0UL;
2797 for_each_cpu_mask(i, group->cpumask) {
2800 if (!cpu_isset(i, *cpus))
2805 if (*sd_idle && rq->nr_running)
2808 /* Bias balancing toward cpus of our domain */
2810 if (idle_cpu(i) && !first_idle_cpu) {
2815 load = target_load(i, load_idx);
2817 load = source_load(i, load_idx);
2818 if (load > max_cpu_load)
2819 max_cpu_load = load;
2820 if (min_cpu_load > load)
2821 min_cpu_load = load;
2825 sum_nr_running += rq->nr_running;
2826 sum_weighted_load += weighted_cpuload(i);
2830 * First idle cpu or the first cpu(busiest) in this sched group
2831 * is eligible for doing load balancing at this and above
2832 * domains. In the newly idle case, we will allow all the cpu's
2833 * to do the newly idle load balance.
2835 if (idle != CPU_NEWLY_IDLE && local_group &&
2836 balance_cpu != this_cpu && balance) {
2841 total_load += avg_load;
2842 total_pwr += group->__cpu_power;
2844 /* Adjust by relative CPU power of the group */
2845 avg_load = sg_div_cpu_power(group,
2846 avg_load * SCHED_LOAD_SCALE);
2848 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2851 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2854 this_load = avg_load;
2856 this_nr_running = sum_nr_running;
2857 this_load_per_task = sum_weighted_load;
2858 } else if (avg_load > max_load &&
2859 (sum_nr_running > group_capacity || __group_imb)) {
2860 max_load = avg_load;
2862 busiest_nr_running = sum_nr_running;
2863 busiest_load_per_task = sum_weighted_load;
2864 group_imb = __group_imb;
2867 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2869 * Busy processors will not participate in power savings
2872 if (idle == CPU_NOT_IDLE ||
2873 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2877 * If the local group is idle or completely loaded
2878 * no need to do power savings balance at this domain
2880 if (local_group && (this_nr_running >= group_capacity ||
2882 power_savings_balance = 0;
2885 * If a group is already running at full capacity or idle,
2886 * don't include that group in power savings calculations
2888 if (!power_savings_balance || sum_nr_running >= group_capacity
2893 * Calculate the group which has the least non-idle load.
2894 * This is the group from where we need to pick up the load
2897 if ((sum_nr_running < min_nr_running) ||
2898 (sum_nr_running == min_nr_running &&
2899 first_cpu(group->cpumask) <
2900 first_cpu(group_min->cpumask))) {
2902 min_nr_running = sum_nr_running;
2903 min_load_per_task = sum_weighted_load /
2908 * Calculate the group which is almost near its
2909 * capacity but still has some space to pick up some load
2910 * from other group and save more power
2912 if (sum_nr_running <= group_capacity - 1) {
2913 if (sum_nr_running > leader_nr_running ||
2914 (sum_nr_running == leader_nr_running &&
2915 first_cpu(group->cpumask) >
2916 first_cpu(group_leader->cpumask))) {
2917 group_leader = group;
2918 leader_nr_running = sum_nr_running;
2923 group = group->next;
2924 } while (group != sd->groups);
2926 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2929 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2931 if (this_load >= avg_load ||
2932 100*max_load <= sd->imbalance_pct*this_load)
2935 busiest_load_per_task /= busiest_nr_running;
2937 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2940 * We're trying to get all the cpus to the average_load, so we don't
2941 * want to push ourselves above the average load, nor do we wish to
2942 * reduce the max loaded cpu below the average load, as either of these
2943 * actions would just result in more rebalancing later, and ping-pong
2944 * tasks around. Thus we look for the minimum possible imbalance.
2945 * Negative imbalances (*we* are more loaded than anyone else) will
2946 * be counted as no imbalance for these purposes -- we can't fix that
2947 * by pulling tasks to us. Be careful of negative numbers as they'll
2948 * appear as very large values with unsigned longs.
2950 if (max_load <= busiest_load_per_task)
2954 * In the presence of smp nice balancing, certain scenarios can have
2955 * max load less than avg load(as we skip the groups at or below
2956 * its cpu_power, while calculating max_load..)
2958 if (max_load < avg_load) {
2960 goto small_imbalance;
2963 /* Don't want to pull so many tasks that a group would go idle */
2964 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2966 /* How much load to actually move to equalise the imbalance */
2967 *imbalance = min(max_pull * busiest->__cpu_power,
2968 (avg_load - this_load) * this->__cpu_power)
2972 * if *imbalance is less than the average load per runnable task
2973 * there is no gaurantee that any tasks will be moved so we'll have
2974 * a think about bumping its value to force at least one task to be
2977 if (*imbalance < busiest_load_per_task) {
2978 unsigned long tmp, pwr_now, pwr_move;
2982 pwr_move = pwr_now = 0;
2984 if (this_nr_running) {
2985 this_load_per_task /= this_nr_running;
2986 if (busiest_load_per_task > this_load_per_task)
2989 this_load_per_task = SCHED_LOAD_SCALE;
2991 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2992 busiest_load_per_task * imbn) {
2993 *imbalance = busiest_load_per_task;
2998 * OK, we don't have enough imbalance to justify moving tasks,
2999 * however we may be able to increase total CPU power used by
3003 pwr_now += busiest->__cpu_power *
3004 min(busiest_load_per_task, max_load);
3005 pwr_now += this->__cpu_power *
3006 min(this_load_per_task, this_load);
3007 pwr_now /= SCHED_LOAD_SCALE;
3009 /* Amount of load we'd subtract */
3010 tmp = sg_div_cpu_power(busiest,
3011 busiest_load_per_task * SCHED_LOAD_SCALE);
3013 pwr_move += busiest->__cpu_power *
3014 min(busiest_load_per_task, max_load - tmp);
3016 /* Amount of load we'd add */
3017 if (max_load * busiest->__cpu_power <
3018 busiest_load_per_task * SCHED_LOAD_SCALE)
3019 tmp = sg_div_cpu_power(this,
3020 max_load * busiest->__cpu_power);
3022 tmp = sg_div_cpu_power(this,
3023 busiest_load_per_task * SCHED_LOAD_SCALE);
3024 pwr_move += this->__cpu_power *
3025 min(this_load_per_task, this_load + tmp);
3026 pwr_move /= SCHED_LOAD_SCALE;
3028 /* Move if we gain throughput */
3029 if (pwr_move > pwr_now)
3030 *imbalance = busiest_load_per_task;
3036 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3037 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3040 if (this == group_leader && group_leader != group_min) {
3041 *imbalance = min_load_per_task;
3051 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3054 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3055 unsigned long imbalance, cpumask_t *cpus)
3057 struct rq *busiest = NULL, *rq;
3058 unsigned long max_load = 0;
3061 for_each_cpu_mask(i, group->cpumask) {
3064 if (!cpu_isset(i, *cpus))
3068 wl = weighted_cpuload(i);
3070 if (rq->nr_running == 1 && wl > imbalance)
3073 if (wl > max_load) {
3083 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3084 * so long as it is large enough.
3086 #define MAX_PINNED_INTERVAL 512
3089 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3090 * tasks if there is an imbalance.
3092 static int load_balance(int this_cpu, struct rq *this_rq,
3093 struct sched_domain *sd, enum cpu_idle_type idle,
3096 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3097 struct sched_group *group;
3098 unsigned long imbalance;
3100 cpumask_t cpus = CPU_MASK_ALL;
3101 unsigned long flags;
3104 * When power savings policy is enabled for the parent domain, idle
3105 * sibling can pick up load irrespective of busy siblings. In this case,
3106 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3107 * portraying it as CPU_NOT_IDLE.
3109 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3110 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3113 schedstat_inc(sd, lb_count[idle]);
3116 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3123 schedstat_inc(sd, lb_nobusyg[idle]);
3127 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
3129 schedstat_inc(sd, lb_nobusyq[idle]);
3133 BUG_ON(busiest == this_rq);
3135 schedstat_add(sd, lb_imbalance[idle], imbalance);
3138 if (busiest->nr_running > 1) {
3140 * Attempt to move tasks. If find_busiest_group has found
3141 * an imbalance but busiest->nr_running <= 1, the group is
3142 * still unbalanced. ld_moved simply stays zero, so it is
3143 * correctly treated as an imbalance.
3145 local_irq_save(flags);
3146 double_rq_lock(this_rq, busiest);
3147 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3148 imbalance, sd, idle, &all_pinned);
3149 double_rq_unlock(this_rq, busiest);
3150 local_irq_restore(flags);
3153 * some other cpu did the load balance for us.
3155 if (ld_moved && this_cpu != smp_processor_id())
3156 resched_cpu(this_cpu);
3158 /* All tasks on this runqueue were pinned by CPU affinity */
3159 if (unlikely(all_pinned)) {
3160 cpu_clear(cpu_of(busiest), cpus);
3161 if (!cpus_empty(cpus))
3168 schedstat_inc(sd, lb_failed[idle]);
3169 sd->nr_balance_failed++;
3171 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3173 spin_lock_irqsave(&busiest->lock, flags);
3175 /* don't kick the migration_thread, if the curr
3176 * task on busiest cpu can't be moved to this_cpu
3178 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3179 spin_unlock_irqrestore(&busiest->lock, flags);
3181 goto out_one_pinned;
3184 if (!busiest->active_balance) {
3185 busiest->active_balance = 1;
3186 busiest->push_cpu = this_cpu;
3189 spin_unlock_irqrestore(&busiest->lock, flags);
3191 wake_up_process(busiest->migration_thread);
3194 * We've kicked active balancing, reset the failure
3197 sd->nr_balance_failed = sd->cache_nice_tries+1;
3200 sd->nr_balance_failed = 0;
3202 if (likely(!active_balance)) {
3203 /* We were unbalanced, so reset the balancing interval */
3204 sd->balance_interval = sd->min_interval;
3207 * If we've begun active balancing, start to back off. This
3208 * case may not be covered by the all_pinned logic if there
3209 * is only 1 task on the busy runqueue (because we don't call
3212 if (sd->balance_interval < sd->max_interval)
3213 sd->balance_interval *= 2;
3216 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3217 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3222 schedstat_inc(sd, lb_balanced[idle]);
3224 sd->nr_balance_failed = 0;
3227 /* tune up the balancing interval */
3228 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3229 (sd->balance_interval < sd->max_interval))
3230 sd->balance_interval *= 2;
3232 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3233 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3239 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3240 * tasks if there is an imbalance.
3242 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3243 * this_rq is locked.
3246 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3248 struct sched_group *group;
3249 struct rq *busiest = NULL;
3250 unsigned long imbalance;
3254 cpumask_t cpus = CPU_MASK_ALL;
3257 * When power savings policy is enabled for the parent domain, idle
3258 * sibling can pick up load irrespective of busy siblings. In this case,
3259 * let the state of idle sibling percolate up as IDLE, instead of
3260 * portraying it as CPU_NOT_IDLE.
3262 if (sd->flags & SD_SHARE_CPUPOWER &&
3263 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3266 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3268 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3269 &sd_idle, &cpus, NULL);
3271 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3275 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
3278 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3282 BUG_ON(busiest == this_rq);
3284 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3287 if (busiest->nr_running > 1) {
3288 /* Attempt to move tasks */
3289 double_lock_balance(this_rq, busiest);
3290 /* this_rq->clock is already updated */
3291 update_rq_clock(busiest);
3292 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3293 imbalance, sd, CPU_NEWLY_IDLE,
3295 spin_unlock(&busiest->lock);
3297 if (unlikely(all_pinned)) {
3298 cpu_clear(cpu_of(busiest), cpus);
3299 if (!cpus_empty(cpus))
3305 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3306 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3307 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3310 sd->nr_balance_failed = 0;
3315 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3316 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3317 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3319 sd->nr_balance_failed = 0;
3325 * idle_balance is called by schedule() if this_cpu is about to become
3326 * idle. Attempts to pull tasks from other CPUs.
3328 static void idle_balance(int this_cpu, struct rq *this_rq)
3330 struct sched_domain *sd;
3331 int pulled_task = -1;
3332 unsigned long next_balance = jiffies + HZ;
3334 for_each_domain(this_cpu, sd) {
3335 unsigned long interval;
3337 if (!(sd->flags & SD_LOAD_BALANCE))
3340 if (sd->flags & SD_BALANCE_NEWIDLE)
3341 /* If we've pulled tasks over stop searching: */
3342 pulled_task = load_balance_newidle(this_cpu,
3345 interval = msecs_to_jiffies(sd->balance_interval);
3346 if (time_after(next_balance, sd->last_balance + interval))
3347 next_balance = sd->last_balance + interval;
3351 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3353 * We are going idle. next_balance may be set based on
3354 * a busy processor. So reset next_balance.
3356 this_rq->next_balance = next_balance;
3361 * active_load_balance is run by migration threads. It pushes running tasks
3362 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3363 * running on each physical CPU where possible, and avoids physical /
3364 * logical imbalances.
3366 * Called with busiest_rq locked.
3368 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3370 int target_cpu = busiest_rq->push_cpu;
3371 struct sched_domain *sd;
3372 struct rq *target_rq;
3374 /* Is there any task to move? */
3375 if (busiest_rq->nr_running <= 1)
3378 target_rq = cpu_rq(target_cpu);
3381 * This condition is "impossible", if it occurs
3382 * we need to fix it. Originally reported by
3383 * Bjorn Helgaas on a 128-cpu setup.
3385 BUG_ON(busiest_rq == target_rq);
3387 /* move a task from busiest_rq to target_rq */
3388 double_lock_balance(busiest_rq, target_rq);
3389 update_rq_clock(busiest_rq);
3390 update_rq_clock(target_rq);
3392 /* Search for an sd spanning us and the target CPU. */
3393 for_each_domain(target_cpu, sd) {
3394 if ((sd->flags & SD_LOAD_BALANCE) &&
3395 cpu_isset(busiest_cpu, sd->span))
3400 schedstat_inc(sd, alb_count);
3402 if (move_one_task(target_rq, target_cpu, busiest_rq,
3404 schedstat_inc(sd, alb_pushed);
3406 schedstat_inc(sd, alb_failed);
3408 spin_unlock(&target_rq->lock);
3413 atomic_t load_balancer;
3415 } nohz ____cacheline_aligned = {
3416 .load_balancer = ATOMIC_INIT(-1),
3417 .cpu_mask = CPU_MASK_NONE,
3421 * This routine will try to nominate the ilb (idle load balancing)
3422 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3423 * load balancing on behalf of all those cpus. If all the cpus in the system
3424 * go into this tickless mode, then there will be no ilb owner (as there is
3425 * no need for one) and all the cpus will sleep till the next wakeup event
3428 * For the ilb owner, tick is not stopped. And this tick will be used
3429 * for idle load balancing. ilb owner will still be part of
3432 * While stopping the tick, this cpu will become the ilb owner if there
3433 * is no other owner. And will be the owner till that cpu becomes busy
3434 * or if all cpus in the system stop their ticks at which point
3435 * there is no need for ilb owner.
3437 * When the ilb owner becomes busy, it nominates another owner, during the
3438 * next busy scheduler_tick()
3440 int select_nohz_load_balancer(int stop_tick)
3442 int cpu = smp_processor_id();
3445 cpu_set(cpu, nohz.cpu_mask);
3446 cpu_rq(cpu)->in_nohz_recently = 1;
3449 * If we are going offline and still the leader, give up!
3451 if (cpu_is_offline(cpu) &&
3452 atomic_read(&nohz.load_balancer) == cpu) {
3453 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3458 /* time for ilb owner also to sleep */
3459 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3460 if (atomic_read(&nohz.load_balancer) == cpu)
3461 atomic_set(&nohz.load_balancer, -1);
3465 if (atomic_read(&nohz.load_balancer) == -1) {
3466 /* make me the ilb owner */
3467 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3469 } else if (atomic_read(&nohz.load_balancer) == cpu)
3472 if (!cpu_isset(cpu, nohz.cpu_mask))
3475 cpu_clear(cpu, nohz.cpu_mask);
3477 if (atomic_read(&nohz.load_balancer) == cpu)
3478 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3485 static DEFINE_SPINLOCK(balancing);
3488 * It checks each scheduling domain to see if it is due to be balanced,
3489 * and initiates a balancing operation if so.
3491 * Balancing parameters are set up in arch_init_sched_domains.
3493 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3496 struct rq *rq = cpu_rq(cpu);
3497 unsigned long interval;
3498 struct sched_domain *sd;
3499 /* Earliest time when we have to do rebalance again */
3500 unsigned long next_balance = jiffies + 60*HZ;
3501 int update_next_balance = 0;
3503 for_each_domain(cpu, sd) {
3504 if (!(sd->flags & SD_LOAD_BALANCE))
3507 interval = sd->balance_interval;
3508 if (idle != CPU_IDLE)
3509 interval *= sd->busy_factor;
3511 /* scale ms to jiffies */
3512 interval = msecs_to_jiffies(interval);
3513 if (unlikely(!interval))
3515 if (interval > HZ*NR_CPUS/10)
3516 interval = HZ*NR_CPUS/10;
3519 if (sd->flags & SD_SERIALIZE) {
3520 if (!spin_trylock(&balancing))
3524 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3525 if (load_balance(cpu, rq, sd, idle, &balance)) {
3527 * We've pulled tasks over so either we're no
3528 * longer idle, or one of our SMT siblings is
3531 idle = CPU_NOT_IDLE;
3533 sd->last_balance = jiffies;
3535 if (sd->flags & SD_SERIALIZE)
3536 spin_unlock(&balancing);
3538 if (time_after(next_balance, sd->last_balance + interval)) {
3539 next_balance = sd->last_balance + interval;
3540 update_next_balance = 1;
3544 * Stop the load balance at this level. There is another
3545 * CPU in our sched group which is doing load balancing more
3553 * next_balance will be updated only when there is a need.
3554 * When the cpu is attached to null domain for ex, it will not be
3557 if (likely(update_next_balance))
3558 rq->next_balance = next_balance;
3562 * run_rebalance_domains is triggered when needed from the scheduler tick.
3563 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3564 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3566 static void run_rebalance_domains(struct softirq_action *h)
3568 int this_cpu = smp_processor_id();
3569 struct rq *this_rq = cpu_rq(this_cpu);
3570 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3571 CPU_IDLE : CPU_NOT_IDLE;
3573 rebalance_domains(this_cpu, idle);
3577 * If this cpu is the owner for idle load balancing, then do the
3578 * balancing on behalf of the other idle cpus whose ticks are
3581 if (this_rq->idle_at_tick &&
3582 atomic_read(&nohz.load_balancer) == this_cpu) {
3583 cpumask_t cpus = nohz.cpu_mask;
3587 cpu_clear(this_cpu, cpus);
3588 for_each_cpu_mask(balance_cpu, cpus) {
3590 * If this cpu gets work to do, stop the load balancing
3591 * work being done for other cpus. Next load
3592 * balancing owner will pick it up.
3597 rebalance_domains(balance_cpu, CPU_IDLE);
3599 rq = cpu_rq(balance_cpu);
3600 if (time_after(this_rq->next_balance, rq->next_balance))
3601 this_rq->next_balance = rq->next_balance;
3608 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3610 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3611 * idle load balancing owner or decide to stop the periodic load balancing,
3612 * if the whole system is idle.
3614 static inline void trigger_load_balance(struct rq *rq, int cpu)
3618 * If we were in the nohz mode recently and busy at the current
3619 * scheduler tick, then check if we need to nominate new idle
3622 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3623 rq->in_nohz_recently = 0;
3625 if (atomic_read(&nohz.load_balancer) == cpu) {
3626 cpu_clear(cpu, nohz.cpu_mask);
3627 atomic_set(&nohz.load_balancer, -1);
3630 if (atomic_read(&nohz.load_balancer) == -1) {
3632 * simple selection for now: Nominate the
3633 * first cpu in the nohz list to be the next
3636 * TBD: Traverse the sched domains and nominate
3637 * the nearest cpu in the nohz.cpu_mask.
3639 int ilb = first_cpu(nohz.cpu_mask);
3647 * If this cpu is idle and doing idle load balancing for all the
3648 * cpus with ticks stopped, is it time for that to stop?
3650 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3651 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3657 * If this cpu is idle and the idle load balancing is done by
3658 * someone else, then no need raise the SCHED_SOFTIRQ
3660 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3661 cpu_isset(cpu, nohz.cpu_mask))
3664 if (time_after_eq(jiffies, rq->next_balance))
3665 raise_softirq(SCHED_SOFTIRQ);
3668 #else /* CONFIG_SMP */
3671 * on UP we do not need to balance between CPUs:
3673 static inline void idle_balance(int cpu, struct rq *rq)
3679 DEFINE_PER_CPU(struct kernel_stat, kstat);
3681 EXPORT_PER_CPU_SYMBOL(kstat);
3684 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3685 * that have not yet been banked in case the task is currently running.
3687 unsigned long long task_sched_runtime(struct task_struct *p)
3689 unsigned long flags;
3693 rq = task_rq_lock(p, &flags);
3694 ns = p->se.sum_exec_runtime;
3695 if (task_current(rq, p)) {
3696 update_rq_clock(rq);
3697 delta_exec = rq->clock - p->se.exec_start;
3698 if ((s64)delta_exec > 0)
3701 task_rq_unlock(rq, &flags);
3707 * Account user cpu time to a process.
3708 * @p: the process that the cpu time gets accounted to
3709 * @cputime: the cpu time spent in user space since the last update
3711 void account_user_time(struct task_struct *p, cputime_t cputime)
3713 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3716 p->utime = cputime_add(p->utime, cputime);
3718 /* Add user time to cpustat. */
3719 tmp = cputime_to_cputime64(cputime);
3720 if (TASK_NICE(p) > 0)
3721 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3723 cpustat->user = cputime64_add(cpustat->user, tmp);
3727 * Account guest cpu time to a process.
3728 * @p: the process that the cpu time gets accounted to
3729 * @cputime: the cpu time spent in virtual machine since the last update
3731 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3734 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3736 tmp = cputime_to_cputime64(cputime);
3738 p->utime = cputime_add(p->utime, cputime);
3739 p->gtime = cputime_add(p->gtime, cputime);
3741 cpustat->user = cputime64_add(cpustat->user, tmp);
3742 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3746 * Account scaled user cpu time to a process.
3747 * @p: the process that the cpu time gets accounted to
3748 * @cputime: the cpu time spent in user space since the last update
3750 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3752 p->utimescaled = cputime_add(p->utimescaled, cputime);
3756 * Account system cpu time to a process.
3757 * @p: the process that the cpu time gets accounted to
3758 * @hardirq_offset: the offset to subtract from hardirq_count()
3759 * @cputime: the cpu time spent in kernel space since the last update
3761 void account_system_time(struct task_struct *p, int hardirq_offset,
3764 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3765 struct rq *rq = this_rq();
3768 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3769 return account_guest_time(p, cputime);
3771 p->stime = cputime_add(p->stime, cputime);
3773 /* Add system time to cpustat. */
3774 tmp = cputime_to_cputime64(cputime);
3775 if (hardirq_count() - hardirq_offset)
3776 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3777 else if (softirq_count())
3778 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3779 else if (p != rq->idle)
3780 cpustat->system = cputime64_add(cpustat->system, tmp);
3781 else if (atomic_read(&rq->nr_iowait) > 0)
3782 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3784 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3785 /* Account for system time used */
3786 acct_update_integrals(p);
3790 * Account scaled system cpu time to a process.
3791 * @p: the process that the cpu time gets accounted to
3792 * @hardirq_offset: the offset to subtract from hardirq_count()
3793 * @cputime: the cpu time spent in kernel space since the last update
3795 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3797 p->stimescaled = cputime_add(p->stimescaled, cputime);
3801 * Account for involuntary wait time.
3802 * @p: the process from which the cpu time has been stolen
3803 * @steal: the cpu time spent in involuntary wait
3805 void account_steal_time(struct task_struct *p, cputime_t steal)
3807 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3808 cputime64_t tmp = cputime_to_cputime64(steal);
3809 struct rq *rq = this_rq();
3811 if (p == rq->idle) {
3812 p->stime = cputime_add(p->stime, steal);
3813 if (atomic_read(&rq->nr_iowait) > 0)
3814 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3816 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3818 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3822 * This function gets called by the timer code, with HZ frequency.
3823 * We call it with interrupts disabled.
3825 * It also gets called by the fork code, when changing the parent's
3828 void scheduler_tick(void)
3830 int cpu = smp_processor_id();
3831 struct rq *rq = cpu_rq(cpu);
3832 struct task_struct *curr = rq->curr;
3833 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3835 spin_lock(&rq->lock);
3836 __update_rq_clock(rq);
3838 * Let rq->clock advance by at least TICK_NSEC:
3840 if (unlikely(rq->clock < next_tick)) {
3841 rq->clock = next_tick;
3842 rq->clock_underflows++;
3844 rq->tick_timestamp = rq->clock;
3845 update_last_tick_seen(rq);
3846 update_cpu_load(rq);
3847 curr->sched_class->task_tick(rq, curr, 0);
3848 update_sched_rt_period(rq);
3849 spin_unlock(&rq->lock);
3852 rq->idle_at_tick = idle_cpu(cpu);
3853 trigger_load_balance(rq, cpu);
3857 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3859 void __kprobes add_preempt_count(int val)
3864 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3866 preempt_count() += val;
3868 * Spinlock count overflowing soon?
3870 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3873 EXPORT_SYMBOL(add_preempt_count);
3875 void __kprobes sub_preempt_count(int val)
3880 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3883 * Is the spinlock portion underflowing?
3885 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3886 !(preempt_count() & PREEMPT_MASK)))
3889 preempt_count() -= val;
3891 EXPORT_SYMBOL(sub_preempt_count);
3896 * Print scheduling while atomic bug:
3898 static noinline void __schedule_bug(struct task_struct *prev)
3900 struct pt_regs *regs = get_irq_regs();
3902 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3903 prev->comm, prev->pid, preempt_count());
3905 debug_show_held_locks(prev);
3906 if (irqs_disabled())
3907 print_irqtrace_events(prev);
3916 * Various schedule()-time debugging checks and statistics:
3918 static inline void schedule_debug(struct task_struct *prev)
3921 * Test if we are atomic. Since do_exit() needs to call into
3922 * schedule() atomically, we ignore that path for now.
3923 * Otherwise, whine if we are scheduling when we should not be.
3925 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3926 __schedule_bug(prev);
3928 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3930 schedstat_inc(this_rq(), sched_count);
3931 #ifdef CONFIG_SCHEDSTATS
3932 if (unlikely(prev->lock_depth >= 0)) {
3933 schedstat_inc(this_rq(), bkl_count);
3934 schedstat_inc(prev, sched_info.bkl_count);
3940 * Pick up the highest-prio task:
3942 static inline struct task_struct *
3943 pick_next_task(struct rq *rq, struct task_struct *prev)
3945 const struct sched_class *class;
3946 struct task_struct *p;
3949 * Optimization: we know that if all tasks are in
3950 * the fair class we can call that function directly:
3952 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3953 p = fair_sched_class.pick_next_task(rq);
3958 class = sched_class_highest;
3960 p = class->pick_next_task(rq);
3964 * Will never be NULL as the idle class always
3965 * returns a non-NULL p:
3967 class = class->next;
3972 * schedule() is the main scheduler function.
3974 asmlinkage void __sched schedule(void)
3976 struct task_struct *prev, *next;
3977 unsigned long *switch_count;
3983 cpu = smp_processor_id();
3987 switch_count = &prev->nivcsw;
3989 release_kernel_lock(prev);
3990 need_resched_nonpreemptible:
3992 schedule_debug(prev);
3997 * Do the rq-clock update outside the rq lock:
3999 local_irq_disable();
4000 __update_rq_clock(rq);
4001 spin_lock(&rq->lock);
4002 clear_tsk_need_resched(prev);
4004 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4005 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4006 signal_pending(prev))) {
4007 prev->state = TASK_RUNNING;
4009 deactivate_task(rq, prev, 1);
4011 switch_count = &prev->nvcsw;
4015 if (prev->sched_class->pre_schedule)
4016 prev->sched_class->pre_schedule(rq, prev);
4019 if (unlikely(!rq->nr_running))
4020 idle_balance(cpu, rq);
4022 prev->sched_class->put_prev_task(rq, prev);
4023 next = pick_next_task(rq, prev);
4025 sched_info_switch(prev, next);
4027 if (likely(prev != next)) {
4032 context_switch(rq, prev, next); /* unlocks the rq */
4034 * the context switch might have flipped the stack from under
4035 * us, hence refresh the local variables.
4037 cpu = smp_processor_id();
4040 spin_unlock_irq(&rq->lock);
4044 if (unlikely(reacquire_kernel_lock(current) < 0))
4045 goto need_resched_nonpreemptible;
4047 preempt_enable_no_resched();
4048 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4051 EXPORT_SYMBOL(schedule);
4053 #ifdef CONFIG_PREEMPT
4055 * this is the entry point to schedule() from in-kernel preemption
4056 * off of preempt_enable. Kernel preemptions off return from interrupt
4057 * occur there and call schedule directly.
4059 asmlinkage void __sched preempt_schedule(void)
4061 struct thread_info *ti = current_thread_info();
4062 struct task_struct *task = current;
4063 int saved_lock_depth;
4066 * If there is a non-zero preempt_count or interrupts are disabled,
4067 * we do not want to preempt the current task. Just return..
4069 if (likely(ti->preempt_count || irqs_disabled()))
4073 add_preempt_count(PREEMPT_ACTIVE);
4076 * We keep the big kernel semaphore locked, but we
4077 * clear ->lock_depth so that schedule() doesnt
4078 * auto-release the semaphore:
4080 saved_lock_depth = task->lock_depth;
4081 task->lock_depth = -1;
4083 task->lock_depth = saved_lock_depth;
4084 sub_preempt_count(PREEMPT_ACTIVE);
4087 * Check again in case we missed a preemption opportunity
4088 * between schedule and now.
4091 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4093 EXPORT_SYMBOL(preempt_schedule);
4096 * this is the entry point to schedule() from kernel preemption
4097 * off of irq context.
4098 * Note, that this is called and return with irqs disabled. This will
4099 * protect us against recursive calling from irq.
4101 asmlinkage void __sched preempt_schedule_irq(void)
4103 struct thread_info *ti = current_thread_info();
4104 struct task_struct *task = current;
4105 int saved_lock_depth;
4107 /* Catch callers which need to be fixed */
4108 BUG_ON(ti->preempt_count || !irqs_disabled());
4111 add_preempt_count(PREEMPT_ACTIVE);
4114 * We keep the big kernel semaphore locked, but we
4115 * clear ->lock_depth so that schedule() doesnt
4116 * auto-release the semaphore:
4118 saved_lock_depth = task->lock_depth;
4119 task->lock_depth = -1;
4122 local_irq_disable();
4123 task->lock_depth = saved_lock_depth;
4124 sub_preempt_count(PREEMPT_ACTIVE);
4127 * Check again in case we missed a preemption opportunity
4128 * between schedule and now.
4131 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4134 #endif /* CONFIG_PREEMPT */
4136 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4139 return try_to_wake_up(curr->private, mode, sync);
4141 EXPORT_SYMBOL(default_wake_function);
4144 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4145 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4146 * number) then we wake all the non-exclusive tasks and one exclusive task.
4148 * There are circumstances in which we can try to wake a task which has already
4149 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4150 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4152 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4153 int nr_exclusive, int sync, void *key)
4155 wait_queue_t *curr, *next;
4157 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4158 unsigned flags = curr->flags;
4160 if (curr->func(curr, mode, sync, key) &&
4161 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4167 * __wake_up - wake up threads blocked on a waitqueue.
4169 * @mode: which threads
4170 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4171 * @key: is directly passed to the wakeup function
4173 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4174 int nr_exclusive, void *key)
4176 unsigned long flags;
4178 spin_lock_irqsave(&q->lock, flags);
4179 __wake_up_common(q, mode, nr_exclusive, 0, key);
4180 spin_unlock_irqrestore(&q->lock, flags);
4182 EXPORT_SYMBOL(__wake_up);
4185 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4187 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4189 __wake_up_common(q, mode, 1, 0, NULL);
4193 * __wake_up_sync - wake up threads blocked on a waitqueue.
4195 * @mode: which threads
4196 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4198 * The sync wakeup differs that the waker knows that it will schedule
4199 * away soon, so while the target thread will be woken up, it will not
4200 * be migrated to another CPU - ie. the two threads are 'synchronized'
4201 * with each other. This can prevent needless bouncing between CPUs.
4203 * On UP it can prevent extra preemption.
4206 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4208 unsigned long flags;
4214 if (unlikely(!nr_exclusive))
4217 spin_lock_irqsave(&q->lock, flags);
4218 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4219 spin_unlock_irqrestore(&q->lock, flags);
4221 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4223 void complete(struct completion *x)
4225 unsigned long flags;
4227 spin_lock_irqsave(&x->wait.lock, flags);
4229 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4230 spin_unlock_irqrestore(&x->wait.lock, flags);
4232 EXPORT_SYMBOL(complete);
4234 void complete_all(struct completion *x)
4236 unsigned long flags;
4238 spin_lock_irqsave(&x->wait.lock, flags);
4239 x->done += UINT_MAX/2;
4240 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4241 spin_unlock_irqrestore(&x->wait.lock, flags);
4243 EXPORT_SYMBOL(complete_all);
4245 static inline long __sched
4246 do_wait_for_common(struct completion *x, long timeout, int state)
4249 DECLARE_WAITQUEUE(wait, current);
4251 wait.flags |= WQ_FLAG_EXCLUSIVE;
4252 __add_wait_queue_tail(&x->wait, &wait);
4254 if ((state == TASK_INTERRUPTIBLE &&
4255 signal_pending(current)) ||
4256 (state == TASK_KILLABLE &&
4257 fatal_signal_pending(current))) {
4258 __remove_wait_queue(&x->wait, &wait);
4259 return -ERESTARTSYS;
4261 __set_current_state(state);
4262 spin_unlock_irq(&x->wait.lock);
4263 timeout = schedule_timeout(timeout);
4264 spin_lock_irq(&x->wait.lock);
4266 __remove_wait_queue(&x->wait, &wait);
4270 __remove_wait_queue(&x->wait, &wait);
4277 wait_for_common(struct completion *x, long timeout, int state)
4281 spin_lock_irq(&x->wait.lock);
4282 timeout = do_wait_for_common(x, timeout, state);
4283 spin_unlock_irq(&x->wait.lock);
4287 void __sched wait_for_completion(struct completion *x)
4289 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4291 EXPORT_SYMBOL(wait_for_completion);
4293 unsigned long __sched
4294 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4296 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4298 EXPORT_SYMBOL(wait_for_completion_timeout);
4300 int __sched wait_for_completion_interruptible(struct completion *x)
4302 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4303 if (t == -ERESTARTSYS)
4307 EXPORT_SYMBOL(wait_for_completion_interruptible);
4309 unsigned long __sched
4310 wait_for_completion_interruptible_timeout(struct completion *x,
4311 unsigned long timeout)
4313 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4315 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4317 int __sched wait_for_completion_killable(struct completion *x)
4319 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4320 if (t == -ERESTARTSYS)
4324 EXPORT_SYMBOL(wait_for_completion_killable);
4327 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4329 unsigned long flags;
4332 init_waitqueue_entry(&wait, current);
4334 __set_current_state(state);
4336 spin_lock_irqsave(&q->lock, flags);
4337 __add_wait_queue(q, &wait);
4338 spin_unlock(&q->lock);
4339 timeout = schedule_timeout(timeout);
4340 spin_lock_irq(&q->lock);
4341 __remove_wait_queue(q, &wait);
4342 spin_unlock_irqrestore(&q->lock, flags);
4347 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4349 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4351 EXPORT_SYMBOL(interruptible_sleep_on);
4354 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4356 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4358 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4360 void __sched sleep_on(wait_queue_head_t *q)
4362 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4364 EXPORT_SYMBOL(sleep_on);
4366 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4368 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4370 EXPORT_SYMBOL(sleep_on_timeout);
4372 #ifdef CONFIG_RT_MUTEXES
4375 * rt_mutex_setprio - set the current priority of a task
4377 * @prio: prio value (kernel-internal form)
4379 * This function changes the 'effective' priority of a task. It does
4380 * not touch ->normal_prio like __setscheduler().
4382 * Used by the rt_mutex code to implement priority inheritance logic.
4384 void rt_mutex_setprio(struct task_struct *p, int prio)
4386 unsigned long flags;
4387 int oldprio, on_rq, running;
4389 const struct sched_class *prev_class = p->sched_class;
4391 BUG_ON(prio < 0 || prio > MAX_PRIO);
4393 rq = task_rq_lock(p, &flags);
4394 update_rq_clock(rq);
4397 on_rq = p->se.on_rq;
4398 running = task_current(rq, p);
4400 dequeue_task(rq, p, 0);
4402 p->sched_class->put_prev_task(rq, p);
4405 p->sched_class = &rt_sched_class;
4407 p->sched_class = &fair_sched_class;
4412 p->sched_class->set_curr_task(rq);
4414 enqueue_task(rq, p, 0);
4416 check_class_changed(rq, p, prev_class, oldprio, running);
4418 task_rq_unlock(rq, &flags);
4423 void set_user_nice(struct task_struct *p, long nice)
4425 int old_prio, delta, on_rq;
4426 unsigned long flags;
4429 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4432 * We have to be careful, if called from sys_setpriority(),
4433 * the task might be in the middle of scheduling on another CPU.
4435 rq = task_rq_lock(p, &flags);
4436 update_rq_clock(rq);
4438 * The RT priorities are set via sched_setscheduler(), but we still
4439 * allow the 'normal' nice value to be set - but as expected
4440 * it wont have any effect on scheduling until the task is
4441 * SCHED_FIFO/SCHED_RR:
4443 if (task_has_rt_policy(p)) {
4444 p->static_prio = NICE_TO_PRIO(nice);
4447 on_rq = p->se.on_rq;
4449 dequeue_task(rq, p, 0);
4453 p->static_prio = NICE_TO_PRIO(nice);
4456 p->prio = effective_prio(p);
4457 delta = p->prio - old_prio;
4460 enqueue_task(rq, p, 0);
4463 * If the task increased its priority or is running and
4464 * lowered its priority, then reschedule its CPU:
4466 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4467 resched_task(rq->curr);
4470 task_rq_unlock(rq, &flags);
4472 EXPORT_SYMBOL(set_user_nice);
4475 * can_nice - check if a task can reduce its nice value
4479 int can_nice(const struct task_struct *p, const int nice)
4481 /* convert nice value [19,-20] to rlimit style value [1,40] */
4482 int nice_rlim = 20 - nice;
4484 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4485 capable(CAP_SYS_NICE));
4488 #ifdef __ARCH_WANT_SYS_NICE
4491 * sys_nice - change the priority of the current process.
4492 * @increment: priority increment
4494 * sys_setpriority is a more generic, but much slower function that
4495 * does similar things.
4497 asmlinkage long sys_nice(int increment)
4502 * Setpriority might change our priority at the same moment.
4503 * We don't have to worry. Conceptually one call occurs first
4504 * and we have a single winner.
4506 if (increment < -40)
4511 nice = PRIO_TO_NICE(current->static_prio) + increment;
4517 if (increment < 0 && !can_nice(current, nice))
4520 retval = security_task_setnice(current, nice);
4524 set_user_nice(current, nice);
4531 * task_prio - return the priority value of a given task.
4532 * @p: the task in question.
4534 * This is the priority value as seen by users in /proc.
4535 * RT tasks are offset by -200. Normal tasks are centered
4536 * around 0, value goes from -16 to +15.
4538 int task_prio(const struct task_struct *p)
4540 return p->prio - MAX_RT_PRIO;
4544 * task_nice - return the nice value of a given task.
4545 * @p: the task in question.
4547 int task_nice(const struct task_struct *p)
4549 return TASK_NICE(p);
4551 EXPORT_SYMBOL(task_nice);
4554 * idle_cpu - is a given cpu idle currently?
4555 * @cpu: the processor in question.
4557 int idle_cpu(int cpu)
4559 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4563 * idle_task - return the idle task for a given cpu.
4564 * @cpu: the processor in question.
4566 struct task_struct *idle_task(int cpu)
4568 return cpu_rq(cpu)->idle;
4572 * find_process_by_pid - find a process with a matching PID value.
4573 * @pid: the pid in question.
4575 static struct task_struct *find_process_by_pid(pid_t pid)
4577 return pid ? find_task_by_vpid(pid) : current;
4580 /* Actually do priority change: must hold rq lock. */
4582 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4584 BUG_ON(p->se.on_rq);
4587 switch (p->policy) {
4591 p->sched_class = &fair_sched_class;
4595 p->sched_class = &rt_sched_class;
4599 p->rt_priority = prio;
4600 p->normal_prio = normal_prio(p);
4601 /* we are holding p->pi_lock already */
4602 p->prio = rt_mutex_getprio(p);
4607 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4608 * @p: the task in question.
4609 * @policy: new policy.
4610 * @param: structure containing the new RT priority.
4612 * NOTE that the task may be already dead.
4614 int sched_setscheduler(struct task_struct *p, int policy,
4615 struct sched_param *param)
4617 int retval, oldprio, oldpolicy = -1, on_rq, running;
4618 unsigned long flags;
4619 const struct sched_class *prev_class = p->sched_class;
4622 /* may grab non-irq protected spin_locks */
4623 BUG_ON(in_interrupt());
4625 /* double check policy once rq lock held */
4627 policy = oldpolicy = p->policy;
4628 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4629 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4630 policy != SCHED_IDLE)
4633 * Valid priorities for SCHED_FIFO and SCHED_RR are
4634 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4635 * SCHED_BATCH and SCHED_IDLE is 0.
4637 if (param->sched_priority < 0 ||
4638 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4639 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4641 if (rt_policy(policy) != (param->sched_priority != 0))
4645 * Allow unprivileged RT tasks to decrease priority:
4647 if (!capable(CAP_SYS_NICE)) {
4648 if (rt_policy(policy)) {
4649 unsigned long rlim_rtprio;
4651 if (!lock_task_sighand(p, &flags))
4653 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4654 unlock_task_sighand(p, &flags);
4656 /* can't set/change the rt policy */
4657 if (policy != p->policy && !rlim_rtprio)
4660 /* can't increase priority */
4661 if (param->sched_priority > p->rt_priority &&
4662 param->sched_priority > rlim_rtprio)
4666 * Like positive nice levels, dont allow tasks to
4667 * move out of SCHED_IDLE either:
4669 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4672 /* can't change other user's priorities */
4673 if ((current->euid != p->euid) &&
4674 (current->euid != p->uid))
4678 #ifdef CONFIG_RT_GROUP_SCHED
4680 * Do not allow realtime tasks into groups that have no runtime
4683 if (rt_policy(policy) && task_group(p)->rt_runtime == 0)
4687 retval = security_task_setscheduler(p, policy, param);
4691 * make sure no PI-waiters arrive (or leave) while we are
4692 * changing the priority of the task:
4694 spin_lock_irqsave(&p->pi_lock, flags);
4696 * To be able to change p->policy safely, the apropriate
4697 * runqueue lock must be held.
4699 rq = __task_rq_lock(p);
4700 /* recheck policy now with rq lock held */
4701 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4702 policy = oldpolicy = -1;
4703 __task_rq_unlock(rq);
4704 spin_unlock_irqrestore(&p->pi_lock, flags);
4707 update_rq_clock(rq);
4708 on_rq = p->se.on_rq;
4709 running = task_current(rq, p);
4711 deactivate_task(rq, p, 0);
4713 p->sched_class->put_prev_task(rq, p);
4716 __setscheduler(rq, p, policy, param->sched_priority);
4719 p->sched_class->set_curr_task(rq);
4721 activate_task(rq, p, 0);
4723 check_class_changed(rq, p, prev_class, oldprio, running);
4725 __task_rq_unlock(rq);
4726 spin_unlock_irqrestore(&p->pi_lock, flags);
4728 rt_mutex_adjust_pi(p);
4732 EXPORT_SYMBOL_GPL(sched_setscheduler);
4735 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4737 struct sched_param lparam;
4738 struct task_struct *p;
4741 if (!param || pid < 0)
4743 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4748 p = find_process_by_pid(pid);
4750 retval = sched_setscheduler(p, policy, &lparam);
4757 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4758 * @pid: the pid in question.
4759 * @policy: new policy.
4760 * @param: structure containing the new RT priority.
4763 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4765 /* negative values for policy are not valid */
4769 return do_sched_setscheduler(pid, policy, param);
4773 * sys_sched_setparam - set/change the RT priority of a thread
4774 * @pid: the pid in question.
4775 * @param: structure containing the new RT priority.
4777 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4779 return do_sched_setscheduler(pid, -1, param);
4783 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4784 * @pid: the pid in question.
4786 asmlinkage long sys_sched_getscheduler(pid_t pid)
4788 struct task_struct *p;
4795 read_lock(&tasklist_lock);
4796 p = find_process_by_pid(pid);
4798 retval = security_task_getscheduler(p);
4802 read_unlock(&tasklist_lock);
4807 * sys_sched_getscheduler - get the RT priority of a thread
4808 * @pid: the pid in question.
4809 * @param: structure containing the RT priority.
4811 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4813 struct sched_param lp;
4814 struct task_struct *p;
4817 if (!param || pid < 0)
4820 read_lock(&tasklist_lock);
4821 p = find_process_by_pid(pid);
4826 retval = security_task_getscheduler(p);
4830 lp.sched_priority = p->rt_priority;
4831 read_unlock(&tasklist_lock);
4834 * This one might sleep, we cannot do it with a spinlock held ...
4836 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4841 read_unlock(&tasklist_lock);
4845 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4847 cpumask_t cpus_allowed;
4848 struct task_struct *p;
4852 read_lock(&tasklist_lock);
4854 p = find_process_by_pid(pid);
4856 read_unlock(&tasklist_lock);
4862 * It is not safe to call set_cpus_allowed with the
4863 * tasklist_lock held. We will bump the task_struct's
4864 * usage count and then drop tasklist_lock.
4867 read_unlock(&tasklist_lock);
4870 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4871 !capable(CAP_SYS_NICE))
4874 retval = security_task_setscheduler(p, 0, NULL);
4878 cpus_allowed = cpuset_cpus_allowed(p);
4879 cpus_and(new_mask, new_mask, cpus_allowed);
4881 retval = set_cpus_allowed(p, new_mask);
4884 cpus_allowed = cpuset_cpus_allowed(p);
4885 if (!cpus_subset(new_mask, cpus_allowed)) {
4887 * We must have raced with a concurrent cpuset
4888 * update. Just reset the cpus_allowed to the
4889 * cpuset's cpus_allowed
4891 new_mask = cpus_allowed;
4901 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4902 cpumask_t *new_mask)
4904 if (len < sizeof(cpumask_t)) {
4905 memset(new_mask, 0, sizeof(cpumask_t));
4906 } else if (len > sizeof(cpumask_t)) {
4907 len = sizeof(cpumask_t);
4909 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4913 * sys_sched_setaffinity - set the cpu affinity of a process
4914 * @pid: pid of the process
4915 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4916 * @user_mask_ptr: user-space pointer to the new cpu mask
4918 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4919 unsigned long __user *user_mask_ptr)
4924 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4928 return sched_setaffinity(pid, new_mask);
4932 * Represents all cpu's present in the system
4933 * In systems capable of hotplug, this map could dynamically grow
4934 * as new cpu's are detected in the system via any platform specific
4935 * method, such as ACPI for e.g.
4938 cpumask_t cpu_present_map __read_mostly;
4939 EXPORT_SYMBOL(cpu_present_map);
4942 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4943 EXPORT_SYMBOL(cpu_online_map);
4945 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4946 EXPORT_SYMBOL(cpu_possible_map);
4949 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4951 struct task_struct *p;
4955 read_lock(&tasklist_lock);
4958 p = find_process_by_pid(pid);
4962 retval = security_task_getscheduler(p);
4966 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4969 read_unlock(&tasklist_lock);
4976 * sys_sched_getaffinity - get the cpu affinity of a process
4977 * @pid: pid of the process
4978 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4979 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4981 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4982 unsigned long __user *user_mask_ptr)
4987 if (len < sizeof(cpumask_t))
4990 ret = sched_getaffinity(pid, &mask);
4994 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4997 return sizeof(cpumask_t);
5001 * sys_sched_yield - yield the current processor to other threads.
5003 * This function yields the current CPU to other tasks. If there are no
5004 * other threads running on this CPU then this function will return.
5006 asmlinkage long sys_sched_yield(void)
5008 struct rq *rq = this_rq_lock();
5010 schedstat_inc(rq, yld_count);
5011 current->sched_class->yield_task(rq);
5014 * Since we are going to call schedule() anyway, there's
5015 * no need to preempt or enable interrupts:
5017 __release(rq->lock);
5018 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5019 _raw_spin_unlock(&rq->lock);
5020 preempt_enable_no_resched();
5027 static void __cond_resched(void)
5029 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5030 __might_sleep(__FILE__, __LINE__);
5033 * The BKS might be reacquired before we have dropped
5034 * PREEMPT_ACTIVE, which could trigger a second
5035 * cond_resched() call.
5038 add_preempt_count(PREEMPT_ACTIVE);
5040 sub_preempt_count(PREEMPT_ACTIVE);
5041 } while (need_resched());
5044 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5045 int __sched _cond_resched(void)
5047 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5048 system_state == SYSTEM_RUNNING) {
5054 EXPORT_SYMBOL(_cond_resched);
5058 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5059 * call schedule, and on return reacquire the lock.
5061 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5062 * operations here to prevent schedule() from being called twice (once via
5063 * spin_unlock(), once by hand).
5065 int cond_resched_lock(spinlock_t *lock)
5067 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5070 if (spin_needbreak(lock) || resched) {
5072 if (resched && need_resched())
5081 EXPORT_SYMBOL(cond_resched_lock);
5083 int __sched cond_resched_softirq(void)
5085 BUG_ON(!in_softirq());
5087 if (need_resched() && system_state == SYSTEM_RUNNING) {
5095 EXPORT_SYMBOL(cond_resched_softirq);
5098 * yield - yield the current processor to other threads.
5100 * This is a shortcut for kernel-space yielding - it marks the
5101 * thread runnable and calls sys_sched_yield().
5103 void __sched yield(void)
5105 set_current_state(TASK_RUNNING);
5108 EXPORT_SYMBOL(yield);
5111 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5112 * that process accounting knows that this is a task in IO wait state.
5114 * But don't do that if it is a deliberate, throttling IO wait (this task
5115 * has set its backing_dev_info: the queue against which it should throttle)
5117 void __sched io_schedule(void)
5119 struct rq *rq = &__raw_get_cpu_var(runqueues);
5121 delayacct_blkio_start();
5122 atomic_inc(&rq->nr_iowait);
5124 atomic_dec(&rq->nr_iowait);
5125 delayacct_blkio_end();
5127 EXPORT_SYMBOL(io_schedule);
5129 long __sched io_schedule_timeout(long timeout)
5131 struct rq *rq = &__raw_get_cpu_var(runqueues);
5134 delayacct_blkio_start();
5135 atomic_inc(&rq->nr_iowait);
5136 ret = schedule_timeout(timeout);
5137 atomic_dec(&rq->nr_iowait);
5138 delayacct_blkio_end();
5143 * sys_sched_get_priority_max - return maximum RT priority.
5144 * @policy: scheduling class.
5146 * this syscall returns the maximum rt_priority that can be used
5147 * by a given scheduling class.
5149 asmlinkage long sys_sched_get_priority_max(int policy)
5156 ret = MAX_USER_RT_PRIO-1;
5168 * sys_sched_get_priority_min - return minimum RT priority.
5169 * @policy: scheduling class.
5171 * this syscall returns the minimum rt_priority that can be used
5172 * by a given scheduling class.
5174 asmlinkage long sys_sched_get_priority_min(int policy)
5192 * sys_sched_rr_get_interval - return the default timeslice of a process.
5193 * @pid: pid of the process.
5194 * @interval: userspace pointer to the timeslice value.
5196 * this syscall writes the default timeslice value of a given process
5197 * into the user-space timespec buffer. A value of '0' means infinity.
5200 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5202 struct task_struct *p;
5203 unsigned int time_slice;
5211 read_lock(&tasklist_lock);
5212 p = find_process_by_pid(pid);
5216 retval = security_task_getscheduler(p);
5221 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5222 * tasks that are on an otherwise idle runqueue:
5225 if (p->policy == SCHED_RR) {
5226 time_slice = DEF_TIMESLICE;
5227 } else if (p->policy != SCHED_FIFO) {
5228 struct sched_entity *se = &p->se;
5229 unsigned long flags;
5232 rq = task_rq_lock(p, &flags);
5233 if (rq->cfs.load.weight)
5234 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5235 task_rq_unlock(rq, &flags);
5237 read_unlock(&tasklist_lock);
5238 jiffies_to_timespec(time_slice, &t);
5239 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5243 read_unlock(&tasklist_lock);
5247 static const char stat_nam[] = "RSDTtZX";
5249 void sched_show_task(struct task_struct *p)
5251 unsigned long free = 0;
5254 state = p->state ? __ffs(p->state) + 1 : 0;
5255 printk(KERN_INFO "%-13.13s %c", p->comm,
5256 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5257 #if BITS_PER_LONG == 32
5258 if (state == TASK_RUNNING)
5259 printk(KERN_CONT " running ");
5261 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5263 if (state == TASK_RUNNING)
5264 printk(KERN_CONT " running task ");
5266 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5268 #ifdef CONFIG_DEBUG_STACK_USAGE
5270 unsigned long *n = end_of_stack(p);
5273 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5276 printk(KERN_CONT "%5lu %5d %6d\n", free,
5277 task_pid_nr(p), task_pid_nr(p->real_parent));
5279 show_stack(p, NULL);
5282 void show_state_filter(unsigned long state_filter)
5284 struct task_struct *g, *p;
5286 #if BITS_PER_LONG == 32
5288 " task PC stack pid father\n");
5291 " task PC stack pid father\n");
5293 read_lock(&tasklist_lock);
5294 do_each_thread(g, p) {
5296 * reset the NMI-timeout, listing all files on a slow
5297 * console might take alot of time:
5299 touch_nmi_watchdog();
5300 if (!state_filter || (p->state & state_filter))
5302 } while_each_thread(g, p);
5304 touch_all_softlockup_watchdogs();
5306 #ifdef CONFIG_SCHED_DEBUG
5307 sysrq_sched_debug_show();
5309 read_unlock(&tasklist_lock);
5311 * Only show locks if all tasks are dumped:
5313 if (state_filter == -1)
5314 debug_show_all_locks();
5317 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5319 idle->sched_class = &idle_sched_class;
5323 * init_idle - set up an idle thread for a given CPU
5324 * @idle: task in question
5325 * @cpu: cpu the idle task belongs to
5327 * NOTE: this function does not set the idle thread's NEED_RESCHED
5328 * flag, to make booting more robust.
5330 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5332 struct rq *rq = cpu_rq(cpu);
5333 unsigned long flags;
5336 idle->se.exec_start = sched_clock();
5338 idle->prio = idle->normal_prio = MAX_PRIO;
5339 idle->cpus_allowed = cpumask_of_cpu(cpu);
5340 __set_task_cpu(idle, cpu);
5342 spin_lock_irqsave(&rq->lock, flags);
5343 rq->curr = rq->idle = idle;
5344 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5347 spin_unlock_irqrestore(&rq->lock, flags);
5349 /* Set the preempt count _outside_ the spinlocks! */
5350 task_thread_info(idle)->preempt_count = 0;
5353 * The idle tasks have their own, simple scheduling class:
5355 idle->sched_class = &idle_sched_class;
5359 * In a system that switches off the HZ timer nohz_cpu_mask
5360 * indicates which cpus entered this state. This is used
5361 * in the rcu update to wait only for active cpus. For system
5362 * which do not switch off the HZ timer nohz_cpu_mask should
5363 * always be CPU_MASK_NONE.
5365 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5368 * Increase the granularity value when there are more CPUs,
5369 * because with more CPUs the 'effective latency' as visible
5370 * to users decreases. But the relationship is not linear,
5371 * so pick a second-best guess by going with the log2 of the
5374 * This idea comes from the SD scheduler of Con Kolivas:
5376 static inline void sched_init_granularity(void)
5378 unsigned int factor = 1 + ilog2(num_online_cpus());
5379 const unsigned long limit = 200000000;
5381 sysctl_sched_min_granularity *= factor;
5382 if (sysctl_sched_min_granularity > limit)
5383 sysctl_sched_min_granularity = limit;
5385 sysctl_sched_latency *= factor;
5386 if (sysctl_sched_latency > limit)
5387 sysctl_sched_latency = limit;
5389 sysctl_sched_wakeup_granularity *= factor;
5390 sysctl_sched_batch_wakeup_granularity *= factor;
5395 * This is how migration works:
5397 * 1) we queue a struct migration_req structure in the source CPU's
5398 * runqueue and wake up that CPU's migration thread.
5399 * 2) we down() the locked semaphore => thread blocks.
5400 * 3) migration thread wakes up (implicitly it forces the migrated
5401 * thread off the CPU)
5402 * 4) it gets the migration request and checks whether the migrated
5403 * task is still in the wrong runqueue.
5404 * 5) if it's in the wrong runqueue then the migration thread removes
5405 * it and puts it into the right queue.
5406 * 6) migration thread up()s the semaphore.
5407 * 7) we wake up and the migration is done.
5411 * Change a given task's CPU affinity. Migrate the thread to a
5412 * proper CPU and schedule it away if the CPU it's executing on
5413 * is removed from the allowed bitmask.
5415 * NOTE: the caller must have a valid reference to the task, the
5416 * task must not exit() & deallocate itself prematurely. The
5417 * call is not atomic; no spinlocks may be held.
5419 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5421 struct migration_req req;
5422 unsigned long flags;
5426 rq = task_rq_lock(p, &flags);
5427 if (!cpus_intersects(new_mask, cpu_online_map)) {
5432 if (p->sched_class->set_cpus_allowed)
5433 p->sched_class->set_cpus_allowed(p, &new_mask);
5435 p->cpus_allowed = new_mask;
5436 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5439 /* Can the task run on the task's current CPU? If so, we're done */
5440 if (cpu_isset(task_cpu(p), new_mask))
5443 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5444 /* Need help from migration thread: drop lock and wait. */
5445 task_rq_unlock(rq, &flags);
5446 wake_up_process(rq->migration_thread);
5447 wait_for_completion(&req.done);
5448 tlb_migrate_finish(p->mm);
5452 task_rq_unlock(rq, &flags);
5456 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5459 * Move (not current) task off this cpu, onto dest cpu. We're doing
5460 * this because either it can't run here any more (set_cpus_allowed()
5461 * away from this CPU, or CPU going down), or because we're
5462 * attempting to rebalance this task on exec (sched_exec).
5464 * So we race with normal scheduler movements, but that's OK, as long
5465 * as the task is no longer on this CPU.
5467 * Returns non-zero if task was successfully migrated.
5469 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5471 struct rq *rq_dest, *rq_src;
5474 if (unlikely(cpu_is_offline(dest_cpu)))
5477 rq_src = cpu_rq(src_cpu);
5478 rq_dest = cpu_rq(dest_cpu);
5480 double_rq_lock(rq_src, rq_dest);
5481 /* Already moved. */
5482 if (task_cpu(p) != src_cpu)
5484 /* Affinity changed (again). */
5485 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5488 on_rq = p->se.on_rq;
5490 deactivate_task(rq_src, p, 0);
5492 set_task_cpu(p, dest_cpu);
5494 activate_task(rq_dest, p, 0);
5495 check_preempt_curr(rq_dest, p);
5499 double_rq_unlock(rq_src, rq_dest);
5504 * migration_thread - this is a highprio system thread that performs
5505 * thread migration by bumping thread off CPU then 'pushing' onto
5508 static int migration_thread(void *data)
5510 int cpu = (long)data;
5514 BUG_ON(rq->migration_thread != current);
5516 set_current_state(TASK_INTERRUPTIBLE);
5517 while (!kthread_should_stop()) {
5518 struct migration_req *req;
5519 struct list_head *head;
5521 spin_lock_irq(&rq->lock);
5523 if (cpu_is_offline(cpu)) {
5524 spin_unlock_irq(&rq->lock);
5528 if (rq->active_balance) {
5529 active_load_balance(rq, cpu);
5530 rq->active_balance = 0;
5533 head = &rq->migration_queue;
5535 if (list_empty(head)) {
5536 spin_unlock_irq(&rq->lock);
5538 set_current_state(TASK_INTERRUPTIBLE);
5541 req = list_entry(head->next, struct migration_req, list);
5542 list_del_init(head->next);
5544 spin_unlock(&rq->lock);
5545 __migrate_task(req->task, cpu, req->dest_cpu);
5548 complete(&req->done);
5550 __set_current_state(TASK_RUNNING);
5554 /* Wait for kthread_stop */
5555 set_current_state(TASK_INTERRUPTIBLE);
5556 while (!kthread_should_stop()) {
5558 set_current_state(TASK_INTERRUPTIBLE);
5560 __set_current_state(TASK_RUNNING);
5564 #ifdef CONFIG_HOTPLUG_CPU
5566 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5570 local_irq_disable();
5571 ret = __migrate_task(p, src_cpu, dest_cpu);
5577 * Figure out where task on dead CPU should go, use force if necessary.
5578 * NOTE: interrupts should be disabled by the caller
5580 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5582 unsigned long flags;
5589 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5590 cpus_and(mask, mask, p->cpus_allowed);
5591 dest_cpu = any_online_cpu(mask);
5593 /* On any allowed CPU? */
5594 if (dest_cpu == NR_CPUS)
5595 dest_cpu = any_online_cpu(p->cpus_allowed);
5597 /* No more Mr. Nice Guy. */
5598 if (dest_cpu == NR_CPUS) {
5599 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5601 * Try to stay on the same cpuset, where the
5602 * current cpuset may be a subset of all cpus.
5603 * The cpuset_cpus_allowed_locked() variant of
5604 * cpuset_cpus_allowed() will not block. It must be
5605 * called within calls to cpuset_lock/cpuset_unlock.
5607 rq = task_rq_lock(p, &flags);
5608 p->cpus_allowed = cpus_allowed;
5609 dest_cpu = any_online_cpu(p->cpus_allowed);
5610 task_rq_unlock(rq, &flags);
5613 * Don't tell them about moving exiting tasks or
5614 * kernel threads (both mm NULL), since they never
5617 if (p->mm && printk_ratelimit()) {
5618 printk(KERN_INFO "process %d (%s) no "
5619 "longer affine to cpu%d\n",
5620 task_pid_nr(p), p->comm, dead_cpu);
5623 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5627 * While a dead CPU has no uninterruptible tasks queued at this point,
5628 * it might still have a nonzero ->nr_uninterruptible counter, because
5629 * for performance reasons the counter is not stricly tracking tasks to
5630 * their home CPUs. So we just add the counter to another CPU's counter,
5631 * to keep the global sum constant after CPU-down:
5633 static void migrate_nr_uninterruptible(struct rq *rq_src)
5635 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5636 unsigned long flags;
5638 local_irq_save(flags);
5639 double_rq_lock(rq_src, rq_dest);
5640 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5641 rq_src->nr_uninterruptible = 0;
5642 double_rq_unlock(rq_src, rq_dest);
5643 local_irq_restore(flags);
5646 /* Run through task list and migrate tasks from the dead cpu. */
5647 static void migrate_live_tasks(int src_cpu)
5649 struct task_struct *p, *t;
5651 read_lock(&tasklist_lock);
5653 do_each_thread(t, p) {
5657 if (task_cpu(p) == src_cpu)
5658 move_task_off_dead_cpu(src_cpu, p);
5659 } while_each_thread(t, p);
5661 read_unlock(&tasklist_lock);
5665 * Schedules idle task to be the next runnable task on current CPU.
5666 * It does so by boosting its priority to highest possible.
5667 * Used by CPU offline code.
5669 void sched_idle_next(void)
5671 int this_cpu = smp_processor_id();
5672 struct rq *rq = cpu_rq(this_cpu);
5673 struct task_struct *p = rq->idle;
5674 unsigned long flags;
5676 /* cpu has to be offline */
5677 BUG_ON(cpu_online(this_cpu));
5680 * Strictly not necessary since rest of the CPUs are stopped by now
5681 * and interrupts disabled on the current cpu.
5683 spin_lock_irqsave(&rq->lock, flags);
5685 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5687 update_rq_clock(rq);
5688 activate_task(rq, p, 0);
5690 spin_unlock_irqrestore(&rq->lock, flags);
5694 * Ensures that the idle task is using init_mm right before its cpu goes
5697 void idle_task_exit(void)
5699 struct mm_struct *mm = current->active_mm;
5701 BUG_ON(cpu_online(smp_processor_id()));
5704 switch_mm(mm, &init_mm, current);
5708 /* called under rq->lock with disabled interrupts */
5709 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5711 struct rq *rq = cpu_rq(dead_cpu);
5713 /* Must be exiting, otherwise would be on tasklist. */
5714 BUG_ON(!p->exit_state);
5716 /* Cannot have done final schedule yet: would have vanished. */
5717 BUG_ON(p->state == TASK_DEAD);
5722 * Drop lock around migration; if someone else moves it,
5723 * that's OK. No task can be added to this CPU, so iteration is
5726 spin_unlock_irq(&rq->lock);
5727 move_task_off_dead_cpu(dead_cpu, p);
5728 spin_lock_irq(&rq->lock);
5733 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5734 static void migrate_dead_tasks(unsigned int dead_cpu)
5736 struct rq *rq = cpu_rq(dead_cpu);
5737 struct task_struct *next;
5740 if (!rq->nr_running)
5742 update_rq_clock(rq);
5743 next = pick_next_task(rq, rq->curr);
5746 migrate_dead(dead_cpu, next);
5750 #endif /* CONFIG_HOTPLUG_CPU */
5752 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5754 static struct ctl_table sd_ctl_dir[] = {
5756 .procname = "sched_domain",
5762 static struct ctl_table sd_ctl_root[] = {
5764 .ctl_name = CTL_KERN,
5765 .procname = "kernel",
5767 .child = sd_ctl_dir,
5772 static struct ctl_table *sd_alloc_ctl_entry(int n)
5774 struct ctl_table *entry =
5775 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5780 static void sd_free_ctl_entry(struct ctl_table **tablep)
5782 struct ctl_table *entry;
5785 * In the intermediate directories, both the child directory and
5786 * procname are dynamically allocated and could fail but the mode
5787 * will always be set. In the lowest directory the names are
5788 * static strings and all have proc handlers.
5790 for (entry = *tablep; entry->mode; entry++) {
5792 sd_free_ctl_entry(&entry->child);
5793 if (entry->proc_handler == NULL)
5794 kfree(entry->procname);
5802 set_table_entry(struct ctl_table *entry,
5803 const char *procname, void *data, int maxlen,
5804 mode_t mode, proc_handler *proc_handler)
5806 entry->procname = procname;
5808 entry->maxlen = maxlen;
5810 entry->proc_handler = proc_handler;
5813 static struct ctl_table *
5814 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5816 struct ctl_table *table = sd_alloc_ctl_entry(12);
5821 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5822 sizeof(long), 0644, proc_doulongvec_minmax);
5823 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5824 sizeof(long), 0644, proc_doulongvec_minmax);
5825 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5826 sizeof(int), 0644, proc_dointvec_minmax);
5827 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5828 sizeof(int), 0644, proc_dointvec_minmax);
5829 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5830 sizeof(int), 0644, proc_dointvec_minmax);
5831 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5832 sizeof(int), 0644, proc_dointvec_minmax);
5833 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5834 sizeof(int), 0644, proc_dointvec_minmax);
5835 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5836 sizeof(int), 0644, proc_dointvec_minmax);
5837 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5838 sizeof(int), 0644, proc_dointvec_minmax);
5839 set_table_entry(&table[9], "cache_nice_tries",
5840 &sd->cache_nice_tries,
5841 sizeof(int), 0644, proc_dointvec_minmax);
5842 set_table_entry(&table[10], "flags", &sd->flags,
5843 sizeof(int), 0644, proc_dointvec_minmax);
5844 /* &table[11] is terminator */
5849 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5851 struct ctl_table *entry, *table;
5852 struct sched_domain *sd;
5853 int domain_num = 0, i;
5856 for_each_domain(cpu, sd)
5858 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5863 for_each_domain(cpu, sd) {
5864 snprintf(buf, 32, "domain%d", i);
5865 entry->procname = kstrdup(buf, GFP_KERNEL);
5867 entry->child = sd_alloc_ctl_domain_table(sd);
5874 static struct ctl_table_header *sd_sysctl_header;
5875 static void register_sched_domain_sysctl(void)
5877 int i, cpu_num = num_online_cpus();
5878 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5881 WARN_ON(sd_ctl_dir[0].child);
5882 sd_ctl_dir[0].child = entry;
5887 for_each_online_cpu(i) {
5888 snprintf(buf, 32, "cpu%d", i);
5889 entry->procname = kstrdup(buf, GFP_KERNEL);
5891 entry->child = sd_alloc_ctl_cpu_table(i);
5895 WARN_ON(sd_sysctl_header);
5896 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5899 /* may be called multiple times per register */
5900 static void unregister_sched_domain_sysctl(void)
5902 if (sd_sysctl_header)
5903 unregister_sysctl_table(sd_sysctl_header);
5904 sd_sysctl_header = NULL;
5905 if (sd_ctl_dir[0].child)
5906 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5909 static void register_sched_domain_sysctl(void)
5912 static void unregister_sched_domain_sysctl(void)
5918 * migration_call - callback that gets triggered when a CPU is added.
5919 * Here we can start up the necessary migration thread for the new CPU.
5921 static int __cpuinit
5922 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5924 struct task_struct *p;
5925 int cpu = (long)hcpu;
5926 unsigned long flags;
5931 case CPU_UP_PREPARE:
5932 case CPU_UP_PREPARE_FROZEN:
5933 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5936 kthread_bind(p, cpu);
5937 /* Must be high prio: stop_machine expects to yield to it. */
5938 rq = task_rq_lock(p, &flags);
5939 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5940 task_rq_unlock(rq, &flags);
5941 cpu_rq(cpu)->migration_thread = p;
5945 case CPU_ONLINE_FROZEN:
5946 /* Strictly unnecessary, as first user will wake it. */
5947 wake_up_process(cpu_rq(cpu)->migration_thread);
5949 /* Update our root-domain */
5951 spin_lock_irqsave(&rq->lock, flags);
5953 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5954 cpu_set(cpu, rq->rd->online);
5956 spin_unlock_irqrestore(&rq->lock, flags);
5959 #ifdef CONFIG_HOTPLUG_CPU
5960 case CPU_UP_CANCELED:
5961 case CPU_UP_CANCELED_FROZEN:
5962 if (!cpu_rq(cpu)->migration_thread)
5964 /* Unbind it from offline cpu so it can run. Fall thru. */
5965 kthread_bind(cpu_rq(cpu)->migration_thread,
5966 any_online_cpu(cpu_online_map));
5967 kthread_stop(cpu_rq(cpu)->migration_thread);
5968 cpu_rq(cpu)->migration_thread = NULL;
5972 case CPU_DEAD_FROZEN:
5973 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5974 migrate_live_tasks(cpu);
5976 kthread_stop(rq->migration_thread);
5977 rq->migration_thread = NULL;
5978 /* Idle task back to normal (off runqueue, low prio) */
5979 spin_lock_irq(&rq->lock);
5980 update_rq_clock(rq);
5981 deactivate_task(rq, rq->idle, 0);
5982 rq->idle->static_prio = MAX_PRIO;
5983 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5984 rq->idle->sched_class = &idle_sched_class;
5985 migrate_dead_tasks(cpu);
5986 spin_unlock_irq(&rq->lock);
5988 migrate_nr_uninterruptible(rq);
5989 BUG_ON(rq->nr_running != 0);
5992 * No need to migrate the tasks: it was best-effort if
5993 * they didn't take sched_hotcpu_mutex. Just wake up
5996 spin_lock_irq(&rq->lock);
5997 while (!list_empty(&rq->migration_queue)) {
5998 struct migration_req *req;
6000 req = list_entry(rq->migration_queue.next,
6001 struct migration_req, list);
6002 list_del_init(&req->list);
6003 complete(&req->done);
6005 spin_unlock_irq(&rq->lock);
6009 case CPU_DYING_FROZEN:
6010 /* Update our root-domain */
6012 spin_lock_irqsave(&rq->lock, flags);
6014 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6015 cpu_clear(cpu, rq->rd->online);
6017 spin_unlock_irqrestore(&rq->lock, flags);
6024 /* Register at highest priority so that task migration (migrate_all_tasks)
6025 * happens before everything else.
6027 static struct notifier_block __cpuinitdata migration_notifier = {
6028 .notifier_call = migration_call,
6032 void __init migration_init(void)
6034 void *cpu = (void *)(long)smp_processor_id();
6037 /* Start one for the boot CPU: */
6038 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6039 BUG_ON(err == NOTIFY_BAD);
6040 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6041 register_cpu_notifier(&migration_notifier);
6047 /* Number of possible processor ids */
6048 int nr_cpu_ids __read_mostly = NR_CPUS;
6049 EXPORT_SYMBOL(nr_cpu_ids);
6051 #ifdef CONFIG_SCHED_DEBUG
6053 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
6055 struct sched_group *group = sd->groups;
6056 cpumask_t groupmask;
6059 cpumask_scnprintf(str, NR_CPUS, sd->span);
6060 cpus_clear(groupmask);
6062 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6064 if (!(sd->flags & SD_LOAD_BALANCE)) {
6065 printk("does not load-balance\n");
6067 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6072 printk(KERN_CONT "span %s\n", str);
6074 if (!cpu_isset(cpu, sd->span)) {
6075 printk(KERN_ERR "ERROR: domain->span does not contain "
6078 if (!cpu_isset(cpu, group->cpumask)) {
6079 printk(KERN_ERR "ERROR: domain->groups does not contain"
6083 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6087 printk(KERN_ERR "ERROR: group is NULL\n");
6091 if (!group->__cpu_power) {
6092 printk(KERN_CONT "\n");
6093 printk(KERN_ERR "ERROR: domain->cpu_power not "
6098 if (!cpus_weight(group->cpumask)) {
6099 printk(KERN_CONT "\n");
6100 printk(KERN_ERR "ERROR: empty group\n");
6104 if (cpus_intersects(groupmask, group->cpumask)) {
6105 printk(KERN_CONT "\n");
6106 printk(KERN_ERR "ERROR: repeated CPUs\n");
6110 cpus_or(groupmask, groupmask, group->cpumask);
6112 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
6113 printk(KERN_CONT " %s", str);
6115 group = group->next;
6116 } while (group != sd->groups);
6117 printk(KERN_CONT "\n");
6119 if (!cpus_equal(sd->span, groupmask))
6120 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6122 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
6123 printk(KERN_ERR "ERROR: parent span is not a superset "
6124 "of domain->span\n");
6128 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6133 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6137 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6140 if (sched_domain_debug_one(sd, cpu, level))
6149 # define sched_domain_debug(sd, cpu) do { } while (0)
6152 static int sd_degenerate(struct sched_domain *sd)
6154 if (cpus_weight(sd->span) == 1)
6157 /* Following flags need at least 2 groups */
6158 if (sd->flags & (SD_LOAD_BALANCE |
6159 SD_BALANCE_NEWIDLE |
6163 SD_SHARE_PKG_RESOURCES)) {
6164 if (sd->groups != sd->groups->next)
6168 /* Following flags don't use groups */
6169 if (sd->flags & (SD_WAKE_IDLE |
6178 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6180 unsigned long cflags = sd->flags, pflags = parent->flags;
6182 if (sd_degenerate(parent))
6185 if (!cpus_equal(sd->span, parent->span))
6188 /* Does parent contain flags not in child? */
6189 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6190 if (cflags & SD_WAKE_AFFINE)
6191 pflags &= ~SD_WAKE_BALANCE;
6192 /* Flags needing groups don't count if only 1 group in parent */
6193 if (parent->groups == parent->groups->next) {
6194 pflags &= ~(SD_LOAD_BALANCE |
6195 SD_BALANCE_NEWIDLE |
6199 SD_SHARE_PKG_RESOURCES);
6201 if (~cflags & pflags)
6207 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6209 unsigned long flags;
6210 const struct sched_class *class;
6212 spin_lock_irqsave(&rq->lock, flags);
6215 struct root_domain *old_rd = rq->rd;
6217 for (class = sched_class_highest; class; class = class->next) {
6218 if (class->leave_domain)
6219 class->leave_domain(rq);
6222 cpu_clear(rq->cpu, old_rd->span);
6223 cpu_clear(rq->cpu, old_rd->online);
6225 if (atomic_dec_and_test(&old_rd->refcount))
6229 atomic_inc(&rd->refcount);
6232 cpu_set(rq->cpu, rd->span);
6233 if (cpu_isset(rq->cpu, cpu_online_map))
6234 cpu_set(rq->cpu, rd->online);
6236 for (class = sched_class_highest; class; class = class->next) {
6237 if (class->join_domain)
6238 class->join_domain(rq);
6241 spin_unlock_irqrestore(&rq->lock, flags);
6244 static void init_rootdomain(struct root_domain *rd)
6246 memset(rd, 0, sizeof(*rd));
6248 cpus_clear(rd->span);
6249 cpus_clear(rd->online);
6252 static void init_defrootdomain(void)
6254 init_rootdomain(&def_root_domain);
6255 atomic_set(&def_root_domain.refcount, 1);
6258 static struct root_domain *alloc_rootdomain(void)
6260 struct root_domain *rd;
6262 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6266 init_rootdomain(rd);
6272 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6273 * hold the hotplug lock.
6276 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6278 struct rq *rq = cpu_rq(cpu);
6279 struct sched_domain *tmp;
6281 /* Remove the sched domains which do not contribute to scheduling. */
6282 for (tmp = sd; tmp; tmp = tmp->parent) {
6283 struct sched_domain *parent = tmp->parent;
6286 if (sd_parent_degenerate(tmp, parent)) {
6287 tmp->parent = parent->parent;
6289 parent->parent->child = tmp;
6293 if (sd && sd_degenerate(sd)) {
6299 sched_domain_debug(sd, cpu);
6301 rq_attach_root(rq, rd);
6302 rcu_assign_pointer(rq->sd, sd);
6305 /* cpus with isolated domains */
6306 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6308 /* Setup the mask of cpus configured for isolated domains */
6309 static int __init isolated_cpu_setup(char *str)
6311 int ints[NR_CPUS], i;
6313 str = get_options(str, ARRAY_SIZE(ints), ints);
6314 cpus_clear(cpu_isolated_map);
6315 for (i = 1; i <= ints[0]; i++)
6316 if (ints[i] < NR_CPUS)
6317 cpu_set(ints[i], cpu_isolated_map);
6321 __setup("isolcpus=", isolated_cpu_setup);
6324 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6325 * to a function which identifies what group(along with sched group) a CPU
6326 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6327 * (due to the fact that we keep track of groups covered with a cpumask_t).
6329 * init_sched_build_groups will build a circular linked list of the groups
6330 * covered by the given span, and will set each group's ->cpumask correctly,
6331 * and ->cpu_power to 0.
6334 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6335 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6336 struct sched_group **sg))
6338 struct sched_group *first = NULL, *last = NULL;
6339 cpumask_t covered = CPU_MASK_NONE;
6342 for_each_cpu_mask(i, span) {
6343 struct sched_group *sg;
6344 int group = group_fn(i, cpu_map, &sg);
6347 if (cpu_isset(i, covered))
6350 sg->cpumask = CPU_MASK_NONE;
6351 sg->__cpu_power = 0;
6353 for_each_cpu_mask(j, span) {
6354 if (group_fn(j, cpu_map, NULL) != group)
6357 cpu_set(j, covered);
6358 cpu_set(j, sg->cpumask);
6369 #define SD_NODES_PER_DOMAIN 16
6374 * find_next_best_node - find the next node to include in a sched_domain
6375 * @node: node whose sched_domain we're building
6376 * @used_nodes: nodes already in the sched_domain
6378 * Find the next node to include in a given scheduling domain. Simply
6379 * finds the closest node not already in the @used_nodes map.
6381 * Should use nodemask_t.
6383 static int find_next_best_node(int node, unsigned long *used_nodes)
6385 int i, n, val, min_val, best_node = 0;
6389 for (i = 0; i < MAX_NUMNODES; i++) {
6390 /* Start at @node */
6391 n = (node + i) % MAX_NUMNODES;
6393 if (!nr_cpus_node(n))
6396 /* Skip already used nodes */
6397 if (test_bit(n, used_nodes))
6400 /* Simple min distance search */
6401 val = node_distance(node, n);
6403 if (val < min_val) {
6409 set_bit(best_node, used_nodes);
6414 * sched_domain_node_span - get a cpumask for a node's sched_domain
6415 * @node: node whose cpumask we're constructing
6416 * @size: number of nodes to include in this span
6418 * Given a node, construct a good cpumask for its sched_domain to span. It
6419 * should be one that prevents unnecessary balancing, but also spreads tasks
6422 static cpumask_t sched_domain_node_span(int node)
6424 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6425 cpumask_t span, nodemask;
6429 bitmap_zero(used_nodes, MAX_NUMNODES);
6431 nodemask = node_to_cpumask(node);
6432 cpus_or(span, span, nodemask);
6433 set_bit(node, used_nodes);
6435 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6436 int next_node = find_next_best_node(node, used_nodes);
6438 nodemask = node_to_cpumask(next_node);
6439 cpus_or(span, span, nodemask);
6446 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6449 * SMT sched-domains:
6451 #ifdef CONFIG_SCHED_SMT
6452 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6453 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6456 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6459 *sg = &per_cpu(sched_group_cpus, cpu);
6465 * multi-core sched-domains:
6467 #ifdef CONFIG_SCHED_MC
6468 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6469 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6472 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6474 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6477 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6478 cpus_and(mask, mask, *cpu_map);
6479 group = first_cpu(mask);
6481 *sg = &per_cpu(sched_group_core, group);
6484 #elif defined(CONFIG_SCHED_MC)
6486 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6489 *sg = &per_cpu(sched_group_core, cpu);
6494 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6495 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6498 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6501 #ifdef CONFIG_SCHED_MC
6502 cpumask_t mask = cpu_coregroup_map(cpu);
6503 cpus_and(mask, mask, *cpu_map);
6504 group = first_cpu(mask);
6505 #elif defined(CONFIG_SCHED_SMT)
6506 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6507 cpus_and(mask, mask, *cpu_map);
6508 group = first_cpu(mask);
6513 *sg = &per_cpu(sched_group_phys, group);
6519 * The init_sched_build_groups can't handle what we want to do with node
6520 * groups, so roll our own. Now each node has its own list of groups which
6521 * gets dynamically allocated.
6523 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6524 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6526 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6527 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6529 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6530 struct sched_group **sg)
6532 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6535 cpus_and(nodemask, nodemask, *cpu_map);
6536 group = first_cpu(nodemask);
6539 *sg = &per_cpu(sched_group_allnodes, group);
6543 static void init_numa_sched_groups_power(struct sched_group *group_head)
6545 struct sched_group *sg = group_head;
6551 for_each_cpu_mask(j, sg->cpumask) {
6552 struct sched_domain *sd;
6554 sd = &per_cpu(phys_domains, j);
6555 if (j != first_cpu(sd->groups->cpumask)) {
6557 * Only add "power" once for each
6563 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6566 } while (sg != group_head);
6571 /* Free memory allocated for various sched_group structures */
6572 static void free_sched_groups(const cpumask_t *cpu_map)
6576 for_each_cpu_mask(cpu, *cpu_map) {
6577 struct sched_group **sched_group_nodes
6578 = sched_group_nodes_bycpu[cpu];
6580 if (!sched_group_nodes)
6583 for (i = 0; i < MAX_NUMNODES; i++) {
6584 cpumask_t nodemask = node_to_cpumask(i);
6585 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6587 cpus_and(nodemask, nodemask, *cpu_map);
6588 if (cpus_empty(nodemask))
6598 if (oldsg != sched_group_nodes[i])
6601 kfree(sched_group_nodes);
6602 sched_group_nodes_bycpu[cpu] = NULL;
6606 static void free_sched_groups(const cpumask_t *cpu_map)
6612 * Initialize sched groups cpu_power.
6614 * cpu_power indicates the capacity of sched group, which is used while
6615 * distributing the load between different sched groups in a sched domain.
6616 * Typically cpu_power for all the groups in a sched domain will be same unless
6617 * there are asymmetries in the topology. If there are asymmetries, group
6618 * having more cpu_power will pickup more load compared to the group having
6621 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6622 * the maximum number of tasks a group can handle in the presence of other idle
6623 * or lightly loaded groups in the same sched domain.
6625 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6627 struct sched_domain *child;
6628 struct sched_group *group;
6630 WARN_ON(!sd || !sd->groups);
6632 if (cpu != first_cpu(sd->groups->cpumask))
6637 sd->groups->__cpu_power = 0;
6640 * For perf policy, if the groups in child domain share resources
6641 * (for example cores sharing some portions of the cache hierarchy
6642 * or SMT), then set this domain groups cpu_power such that each group
6643 * can handle only one task, when there are other idle groups in the
6644 * same sched domain.
6646 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6648 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6649 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6654 * add cpu_power of each child group to this groups cpu_power
6656 group = child->groups;
6658 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6659 group = group->next;
6660 } while (group != child->groups);
6664 * Build sched domains for a given set of cpus and attach the sched domains
6665 * to the individual cpus
6667 static int build_sched_domains(const cpumask_t *cpu_map)
6670 struct root_domain *rd;
6672 struct sched_group **sched_group_nodes = NULL;
6673 int sd_allnodes = 0;
6676 * Allocate the per-node list of sched groups
6678 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6680 if (!sched_group_nodes) {
6681 printk(KERN_WARNING "Can not alloc sched group node list\n");
6684 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6687 rd = alloc_rootdomain();
6689 printk(KERN_WARNING "Cannot alloc root domain\n");
6694 * Set up domains for cpus specified by the cpu_map.
6696 for_each_cpu_mask(i, *cpu_map) {
6697 struct sched_domain *sd = NULL, *p;
6698 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6700 cpus_and(nodemask, nodemask, *cpu_map);
6703 if (cpus_weight(*cpu_map) >
6704 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6705 sd = &per_cpu(allnodes_domains, i);
6706 *sd = SD_ALLNODES_INIT;
6707 sd->span = *cpu_map;
6708 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6714 sd = &per_cpu(node_domains, i);
6716 sd->span = sched_domain_node_span(cpu_to_node(i));
6720 cpus_and(sd->span, sd->span, *cpu_map);
6724 sd = &per_cpu(phys_domains, i);
6726 sd->span = nodemask;
6730 cpu_to_phys_group(i, cpu_map, &sd->groups);
6732 #ifdef CONFIG_SCHED_MC
6734 sd = &per_cpu(core_domains, i);
6736 sd->span = cpu_coregroup_map(i);
6737 cpus_and(sd->span, sd->span, *cpu_map);
6740 cpu_to_core_group(i, cpu_map, &sd->groups);
6743 #ifdef CONFIG_SCHED_SMT
6745 sd = &per_cpu(cpu_domains, i);
6746 *sd = SD_SIBLING_INIT;
6747 sd->span = per_cpu(cpu_sibling_map, i);
6748 cpus_and(sd->span, sd->span, *cpu_map);
6751 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6755 #ifdef CONFIG_SCHED_SMT
6756 /* Set up CPU (sibling) groups */
6757 for_each_cpu_mask(i, *cpu_map) {
6758 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6759 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6760 if (i != first_cpu(this_sibling_map))
6763 init_sched_build_groups(this_sibling_map, cpu_map,
6768 #ifdef CONFIG_SCHED_MC
6769 /* Set up multi-core groups */
6770 for_each_cpu_mask(i, *cpu_map) {
6771 cpumask_t this_core_map = cpu_coregroup_map(i);
6772 cpus_and(this_core_map, this_core_map, *cpu_map);
6773 if (i != first_cpu(this_core_map))
6775 init_sched_build_groups(this_core_map, cpu_map,
6776 &cpu_to_core_group);
6780 /* Set up physical groups */
6781 for (i = 0; i < MAX_NUMNODES; i++) {
6782 cpumask_t nodemask = node_to_cpumask(i);
6784 cpus_and(nodemask, nodemask, *cpu_map);
6785 if (cpus_empty(nodemask))
6788 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6792 /* Set up node groups */
6794 init_sched_build_groups(*cpu_map, cpu_map,
6795 &cpu_to_allnodes_group);
6797 for (i = 0; i < MAX_NUMNODES; i++) {
6798 /* Set up node groups */
6799 struct sched_group *sg, *prev;
6800 cpumask_t nodemask = node_to_cpumask(i);
6801 cpumask_t domainspan;
6802 cpumask_t covered = CPU_MASK_NONE;
6805 cpus_and(nodemask, nodemask, *cpu_map);
6806 if (cpus_empty(nodemask)) {
6807 sched_group_nodes[i] = NULL;
6811 domainspan = sched_domain_node_span(i);
6812 cpus_and(domainspan, domainspan, *cpu_map);
6814 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6816 printk(KERN_WARNING "Can not alloc domain group for "
6820 sched_group_nodes[i] = sg;
6821 for_each_cpu_mask(j, nodemask) {
6822 struct sched_domain *sd;
6824 sd = &per_cpu(node_domains, j);
6827 sg->__cpu_power = 0;
6828 sg->cpumask = nodemask;
6830 cpus_or(covered, covered, nodemask);
6833 for (j = 0; j < MAX_NUMNODES; j++) {
6834 cpumask_t tmp, notcovered;
6835 int n = (i + j) % MAX_NUMNODES;
6837 cpus_complement(notcovered, covered);
6838 cpus_and(tmp, notcovered, *cpu_map);
6839 cpus_and(tmp, tmp, domainspan);
6840 if (cpus_empty(tmp))
6843 nodemask = node_to_cpumask(n);
6844 cpus_and(tmp, tmp, nodemask);
6845 if (cpus_empty(tmp))
6848 sg = kmalloc_node(sizeof(struct sched_group),
6852 "Can not alloc domain group for node %d\n", j);
6855 sg->__cpu_power = 0;
6857 sg->next = prev->next;
6858 cpus_or(covered, covered, tmp);
6865 /* Calculate CPU power for physical packages and nodes */
6866 #ifdef CONFIG_SCHED_SMT
6867 for_each_cpu_mask(i, *cpu_map) {
6868 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6870 init_sched_groups_power(i, sd);
6873 #ifdef CONFIG_SCHED_MC
6874 for_each_cpu_mask(i, *cpu_map) {
6875 struct sched_domain *sd = &per_cpu(core_domains, i);
6877 init_sched_groups_power(i, sd);
6881 for_each_cpu_mask(i, *cpu_map) {
6882 struct sched_domain *sd = &per_cpu(phys_domains, i);
6884 init_sched_groups_power(i, sd);
6888 for (i = 0; i < MAX_NUMNODES; i++)
6889 init_numa_sched_groups_power(sched_group_nodes[i]);
6892 struct sched_group *sg;
6894 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6895 init_numa_sched_groups_power(sg);
6899 /* Attach the domains */
6900 for_each_cpu_mask(i, *cpu_map) {
6901 struct sched_domain *sd;
6902 #ifdef CONFIG_SCHED_SMT
6903 sd = &per_cpu(cpu_domains, i);
6904 #elif defined(CONFIG_SCHED_MC)
6905 sd = &per_cpu(core_domains, i);
6907 sd = &per_cpu(phys_domains, i);
6909 cpu_attach_domain(sd, rd, i);
6916 free_sched_groups(cpu_map);
6921 static cpumask_t *doms_cur; /* current sched domains */
6922 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6925 * Special case: If a kmalloc of a doms_cur partition (array of
6926 * cpumask_t) fails, then fallback to a single sched domain,
6927 * as determined by the single cpumask_t fallback_doms.
6929 static cpumask_t fallback_doms;
6931 void __attribute__((weak)) arch_update_cpu_topology(void)
6936 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6937 * For now this just excludes isolated cpus, but could be used to
6938 * exclude other special cases in the future.
6940 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6944 arch_update_cpu_topology();
6946 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6948 doms_cur = &fallback_doms;
6949 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6950 err = build_sched_domains(doms_cur);
6951 register_sched_domain_sysctl();
6956 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6958 free_sched_groups(cpu_map);
6962 * Detach sched domains from a group of cpus specified in cpu_map
6963 * These cpus will now be attached to the NULL domain
6965 static void detach_destroy_domains(const cpumask_t *cpu_map)
6969 unregister_sched_domain_sysctl();
6971 for_each_cpu_mask(i, *cpu_map)
6972 cpu_attach_domain(NULL, &def_root_domain, i);
6973 synchronize_sched();
6974 arch_destroy_sched_domains(cpu_map);
6978 * Partition sched domains as specified by the 'ndoms_new'
6979 * cpumasks in the array doms_new[] of cpumasks. This compares
6980 * doms_new[] to the current sched domain partitioning, doms_cur[].
6981 * It destroys each deleted domain and builds each new domain.
6983 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6984 * The masks don't intersect (don't overlap.) We should setup one
6985 * sched domain for each mask. CPUs not in any of the cpumasks will
6986 * not be load balanced. If the same cpumask appears both in the
6987 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6990 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6991 * ownership of it and will kfree it when done with it. If the caller
6992 * failed the kmalloc call, then it can pass in doms_new == NULL,
6993 * and partition_sched_domains() will fallback to the single partition
6996 * Call with hotplug lock held
6998 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7004 /* always unregister in case we don't destroy any domains */
7005 unregister_sched_domain_sysctl();
7007 if (doms_new == NULL) {
7009 doms_new = &fallback_doms;
7010 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7013 /* Destroy deleted domains */
7014 for (i = 0; i < ndoms_cur; i++) {
7015 for (j = 0; j < ndoms_new; j++) {
7016 if (cpus_equal(doms_cur[i], doms_new[j]))
7019 /* no match - a current sched domain not in new doms_new[] */
7020 detach_destroy_domains(doms_cur + i);
7025 /* Build new domains */
7026 for (i = 0; i < ndoms_new; i++) {
7027 for (j = 0; j < ndoms_cur; j++) {
7028 if (cpus_equal(doms_new[i], doms_cur[j]))
7031 /* no match - add a new doms_new */
7032 build_sched_domains(doms_new + i);
7037 /* Remember the new sched domains */
7038 if (doms_cur != &fallback_doms)
7040 doms_cur = doms_new;
7041 ndoms_cur = ndoms_new;
7043 register_sched_domain_sysctl();
7048 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7049 int arch_reinit_sched_domains(void)
7054 detach_destroy_domains(&cpu_online_map);
7055 err = arch_init_sched_domains(&cpu_online_map);
7061 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7065 if (buf[0] != '0' && buf[0] != '1')
7069 sched_smt_power_savings = (buf[0] == '1');
7071 sched_mc_power_savings = (buf[0] == '1');
7073 ret = arch_reinit_sched_domains();
7075 return ret ? ret : count;
7078 #ifdef CONFIG_SCHED_MC
7079 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7081 return sprintf(page, "%u\n", sched_mc_power_savings);
7083 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7084 const char *buf, size_t count)
7086 return sched_power_savings_store(buf, count, 0);
7088 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7089 sched_mc_power_savings_store);
7092 #ifdef CONFIG_SCHED_SMT
7093 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7095 return sprintf(page, "%u\n", sched_smt_power_savings);
7097 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7098 const char *buf, size_t count)
7100 return sched_power_savings_store(buf, count, 1);
7102 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7103 sched_smt_power_savings_store);
7106 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7110 #ifdef CONFIG_SCHED_SMT
7112 err = sysfs_create_file(&cls->kset.kobj,
7113 &attr_sched_smt_power_savings.attr);
7115 #ifdef CONFIG_SCHED_MC
7116 if (!err && mc_capable())
7117 err = sysfs_create_file(&cls->kset.kobj,
7118 &attr_sched_mc_power_savings.attr);
7125 * Force a reinitialization of the sched domains hierarchy. The domains
7126 * and groups cannot be updated in place without racing with the balancing
7127 * code, so we temporarily attach all running cpus to the NULL domain
7128 * which will prevent rebalancing while the sched domains are recalculated.
7130 static int update_sched_domains(struct notifier_block *nfb,
7131 unsigned long action, void *hcpu)
7134 case CPU_UP_PREPARE:
7135 case CPU_UP_PREPARE_FROZEN:
7136 case CPU_DOWN_PREPARE:
7137 case CPU_DOWN_PREPARE_FROZEN:
7138 detach_destroy_domains(&cpu_online_map);
7141 case CPU_UP_CANCELED:
7142 case CPU_UP_CANCELED_FROZEN:
7143 case CPU_DOWN_FAILED:
7144 case CPU_DOWN_FAILED_FROZEN:
7146 case CPU_ONLINE_FROZEN:
7148 case CPU_DEAD_FROZEN:
7150 * Fall through and re-initialise the domains.
7157 /* The hotplug lock is already held by cpu_up/cpu_down */
7158 arch_init_sched_domains(&cpu_online_map);
7163 void __init sched_init_smp(void)
7165 cpumask_t non_isolated_cpus;
7168 arch_init_sched_domains(&cpu_online_map);
7169 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7170 if (cpus_empty(non_isolated_cpus))
7171 cpu_set(smp_processor_id(), non_isolated_cpus);
7173 /* XXX: Theoretical race here - CPU may be hotplugged now */
7174 hotcpu_notifier(update_sched_domains, 0);
7176 /* Move init over to a non-isolated CPU */
7177 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7179 sched_init_granularity();
7182 void __init sched_init_smp(void)
7184 sched_init_granularity();
7186 #endif /* CONFIG_SMP */
7188 int in_sched_functions(unsigned long addr)
7190 return in_lock_functions(addr) ||
7191 (addr >= (unsigned long)__sched_text_start
7192 && addr < (unsigned long)__sched_text_end);
7195 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7197 cfs_rq->tasks_timeline = RB_ROOT;
7198 #ifdef CONFIG_FAIR_GROUP_SCHED
7201 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7204 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7206 struct rt_prio_array *array;
7209 array = &rt_rq->active;
7210 for (i = 0; i < MAX_RT_PRIO; i++) {
7211 INIT_LIST_HEAD(array->queue + i);
7212 __clear_bit(i, array->bitmap);
7214 /* delimiter for bitsearch: */
7215 __set_bit(MAX_RT_PRIO, array->bitmap);
7217 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7218 rt_rq->highest_prio = MAX_RT_PRIO;
7221 rt_rq->rt_nr_migratory = 0;
7222 rt_rq->overloaded = 0;
7226 rt_rq->rt_throttled = 0;
7228 #ifdef CONFIG_RT_GROUP_SCHED
7229 rt_rq->rt_nr_boosted = 0;
7234 #ifdef CONFIG_FAIR_GROUP_SCHED
7235 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7236 struct cfs_rq *cfs_rq, struct sched_entity *se,
7239 tg->cfs_rq[cpu] = cfs_rq;
7240 init_cfs_rq(cfs_rq, rq);
7243 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7246 se->cfs_rq = &rq->cfs;
7248 se->load.weight = tg->shares;
7249 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7254 #ifdef CONFIG_RT_GROUP_SCHED
7255 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7256 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7259 tg->rt_rq[cpu] = rt_rq;
7260 init_rt_rq(rt_rq, rq);
7262 rt_rq->rt_se = rt_se;
7264 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7266 tg->rt_se[cpu] = rt_se;
7267 rt_se->rt_rq = &rq->rt;
7268 rt_se->my_q = rt_rq;
7269 rt_se->parent = NULL;
7270 INIT_LIST_HEAD(&rt_se->run_list);
7274 void __init sched_init(void)
7276 int highest_cpu = 0;
7280 init_defrootdomain();
7283 #ifdef CONFIG_GROUP_SCHED
7284 list_add(&init_task_group.list, &task_groups);
7287 for_each_possible_cpu(i) {
7291 spin_lock_init(&rq->lock);
7292 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7295 update_last_tick_seen(rq);
7296 init_cfs_rq(&rq->cfs, rq);
7297 init_rt_rq(&rq->rt, rq);
7298 #ifdef CONFIG_FAIR_GROUP_SCHED
7299 init_task_group.shares = init_task_group_load;
7300 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7301 init_tg_cfs_entry(rq, &init_task_group,
7302 &per_cpu(init_cfs_rq, i),
7303 &per_cpu(init_sched_entity, i), i, 1);
7306 #ifdef CONFIG_RT_GROUP_SCHED
7307 init_task_group.rt_runtime =
7308 sysctl_sched_rt_runtime * NSEC_PER_USEC;
7309 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7310 init_tg_rt_entry(rq, &init_task_group,
7311 &per_cpu(init_rt_rq, i),
7312 &per_cpu(init_sched_rt_entity, i), i, 1);
7314 rq->rt_period_expire = 0;
7315 rq->rt_throttled = 0;
7317 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7318 rq->cpu_load[j] = 0;
7322 rq->active_balance = 0;
7323 rq->next_balance = jiffies;
7326 rq->migration_thread = NULL;
7327 INIT_LIST_HEAD(&rq->migration_queue);
7328 rq_attach_root(rq, &def_root_domain);
7331 atomic_set(&rq->nr_iowait, 0);
7335 set_load_weight(&init_task);
7337 #ifdef CONFIG_PREEMPT_NOTIFIERS
7338 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7342 nr_cpu_ids = highest_cpu + 1;
7343 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7346 #ifdef CONFIG_RT_MUTEXES
7347 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7351 * The boot idle thread does lazy MMU switching as well:
7353 atomic_inc(&init_mm.mm_count);
7354 enter_lazy_tlb(&init_mm, current);
7357 * Make us the idle thread. Technically, schedule() should not be
7358 * called from this thread, however somewhere below it might be,
7359 * but because we are the idle thread, we just pick up running again
7360 * when this runqueue becomes "idle".
7362 init_idle(current, smp_processor_id());
7364 * During early bootup we pretend to be a normal task:
7366 current->sched_class = &fair_sched_class;
7368 scheduler_running = 1;
7371 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7372 void __might_sleep(char *file, int line)
7375 static unsigned long prev_jiffy; /* ratelimiting */
7377 if ((in_atomic() || irqs_disabled()) &&
7378 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7379 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7381 prev_jiffy = jiffies;
7382 printk(KERN_ERR "BUG: sleeping function called from invalid"
7383 " context at %s:%d\n", file, line);
7384 printk("in_atomic():%d, irqs_disabled():%d\n",
7385 in_atomic(), irqs_disabled());
7386 debug_show_held_locks(current);
7387 if (irqs_disabled())
7388 print_irqtrace_events(current);
7393 EXPORT_SYMBOL(__might_sleep);
7396 #ifdef CONFIG_MAGIC_SYSRQ
7397 static void normalize_task(struct rq *rq, struct task_struct *p)
7400 update_rq_clock(rq);
7401 on_rq = p->se.on_rq;
7403 deactivate_task(rq, p, 0);
7404 __setscheduler(rq, p, SCHED_NORMAL, 0);
7406 activate_task(rq, p, 0);
7407 resched_task(rq->curr);
7411 void normalize_rt_tasks(void)
7413 struct task_struct *g, *p;
7414 unsigned long flags;
7417 read_lock_irqsave(&tasklist_lock, flags);
7418 do_each_thread(g, p) {
7420 * Only normalize user tasks:
7425 p->se.exec_start = 0;
7426 #ifdef CONFIG_SCHEDSTATS
7427 p->se.wait_start = 0;
7428 p->se.sleep_start = 0;
7429 p->se.block_start = 0;
7431 task_rq(p)->clock = 0;
7435 * Renice negative nice level userspace
7438 if (TASK_NICE(p) < 0 && p->mm)
7439 set_user_nice(p, 0);
7443 spin_lock(&p->pi_lock);
7444 rq = __task_rq_lock(p);
7446 normalize_task(rq, p);
7448 __task_rq_unlock(rq);
7449 spin_unlock(&p->pi_lock);
7450 } while_each_thread(g, p);
7452 read_unlock_irqrestore(&tasklist_lock, flags);
7455 #endif /* CONFIG_MAGIC_SYSRQ */
7459 * These functions are only useful for the IA64 MCA handling.
7461 * They can only be called when the whole system has been
7462 * stopped - every CPU needs to be quiescent, and no scheduling
7463 * activity can take place. Using them for anything else would
7464 * be a serious bug, and as a result, they aren't even visible
7465 * under any other configuration.
7469 * curr_task - return the current task for a given cpu.
7470 * @cpu: the processor in question.
7472 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7474 struct task_struct *curr_task(int cpu)
7476 return cpu_curr(cpu);
7480 * set_curr_task - set the current task for a given cpu.
7481 * @cpu: the processor in question.
7482 * @p: the task pointer to set.
7484 * Description: This function must only be used when non-maskable interrupts
7485 * are serviced on a separate stack. It allows the architecture to switch the
7486 * notion of the current task on a cpu in a non-blocking manner. This function
7487 * must be called with all CPU's synchronized, and interrupts disabled, the
7488 * and caller must save the original value of the current task (see
7489 * curr_task() above) and restore that value before reenabling interrupts and
7490 * re-starting the system.
7492 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7494 void set_curr_task(int cpu, struct task_struct *p)
7501 #ifdef CONFIG_GROUP_SCHED
7503 #ifdef CONFIG_FAIR_GROUP_SCHED
7504 static void free_fair_sched_group(struct task_group *tg)
7508 for_each_possible_cpu(i) {
7510 kfree(tg->cfs_rq[i]);
7519 static int alloc_fair_sched_group(struct task_group *tg)
7521 struct cfs_rq *cfs_rq;
7522 struct sched_entity *se;
7526 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7529 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7533 tg->shares = NICE_0_LOAD;
7535 for_each_possible_cpu(i) {
7538 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7539 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7543 se = kmalloc_node(sizeof(struct sched_entity),
7544 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7548 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7557 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7559 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7560 &cpu_rq(cpu)->leaf_cfs_rq_list);
7563 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7565 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7568 static inline void free_fair_sched_group(struct task_group *tg)
7572 static inline int alloc_fair_sched_group(struct task_group *tg)
7577 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7581 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7586 #ifdef CONFIG_RT_GROUP_SCHED
7587 static void free_rt_sched_group(struct task_group *tg)
7591 for_each_possible_cpu(i) {
7593 kfree(tg->rt_rq[i]);
7595 kfree(tg->rt_se[i]);
7602 static int alloc_rt_sched_group(struct task_group *tg)
7604 struct rt_rq *rt_rq;
7605 struct sched_rt_entity *rt_se;
7609 tg->rt_rq = kzalloc(sizeof(rt_rq) * NR_CPUS, GFP_KERNEL);
7612 tg->rt_se = kzalloc(sizeof(rt_se) * NR_CPUS, GFP_KERNEL);
7618 for_each_possible_cpu(i) {
7621 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7622 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7626 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7627 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7631 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7640 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7642 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7643 &cpu_rq(cpu)->leaf_rt_rq_list);
7646 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7648 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7651 static inline void free_rt_sched_group(struct task_group *tg)
7655 static inline int alloc_rt_sched_group(struct task_group *tg)
7660 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7664 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7669 static void free_sched_group(struct task_group *tg)
7671 free_fair_sched_group(tg);
7672 free_rt_sched_group(tg);
7676 /* allocate runqueue etc for a new task group */
7677 struct task_group *sched_create_group(void)
7679 struct task_group *tg;
7680 unsigned long flags;
7683 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7685 return ERR_PTR(-ENOMEM);
7687 if (!alloc_fair_sched_group(tg))
7690 if (!alloc_rt_sched_group(tg))
7693 spin_lock_irqsave(&task_group_lock, flags);
7694 for_each_possible_cpu(i) {
7695 register_fair_sched_group(tg, i);
7696 register_rt_sched_group(tg, i);
7698 list_add_rcu(&tg->list, &task_groups);
7699 spin_unlock_irqrestore(&task_group_lock, flags);
7704 free_sched_group(tg);
7705 return ERR_PTR(-ENOMEM);
7708 /* rcu callback to free various structures associated with a task group */
7709 static void free_sched_group_rcu(struct rcu_head *rhp)
7711 /* now it should be safe to free those cfs_rqs */
7712 free_sched_group(container_of(rhp, struct task_group, rcu));
7715 /* Destroy runqueue etc associated with a task group */
7716 void sched_destroy_group(struct task_group *tg)
7718 unsigned long flags;
7721 spin_lock_irqsave(&task_group_lock, flags);
7722 for_each_possible_cpu(i) {
7723 unregister_fair_sched_group(tg, i);
7724 unregister_rt_sched_group(tg, i);
7726 list_del_rcu(&tg->list);
7727 spin_unlock_irqrestore(&task_group_lock, flags);
7729 /* wait for possible concurrent references to cfs_rqs complete */
7730 call_rcu(&tg->rcu, free_sched_group_rcu);
7733 /* change task's runqueue when it moves between groups.
7734 * The caller of this function should have put the task in its new group
7735 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7736 * reflect its new group.
7738 void sched_move_task(struct task_struct *tsk)
7741 unsigned long flags;
7744 rq = task_rq_lock(tsk, &flags);
7746 update_rq_clock(rq);
7748 running = task_current(rq, tsk);
7749 on_rq = tsk->se.on_rq;
7752 dequeue_task(rq, tsk, 0);
7753 if (unlikely(running))
7754 tsk->sched_class->put_prev_task(rq, tsk);
7756 set_task_rq(tsk, task_cpu(tsk));
7758 #ifdef CONFIG_FAIR_GROUP_SCHED
7759 if (tsk->sched_class->moved_group)
7760 tsk->sched_class->moved_group(tsk);
7763 if (unlikely(running))
7764 tsk->sched_class->set_curr_task(rq);
7766 enqueue_task(rq, tsk, 0);
7768 task_rq_unlock(rq, &flags);
7771 #ifdef CONFIG_FAIR_GROUP_SCHED
7772 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7774 struct cfs_rq *cfs_rq = se->cfs_rq;
7775 struct rq *rq = cfs_rq->rq;
7778 spin_lock_irq(&rq->lock);
7782 dequeue_entity(cfs_rq, se, 0);
7784 se->load.weight = shares;
7785 se->load.inv_weight = div64_64((1ULL<<32), shares);
7788 enqueue_entity(cfs_rq, se, 0);
7790 spin_unlock_irq(&rq->lock);
7793 static DEFINE_MUTEX(shares_mutex);
7795 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7798 unsigned long flags;
7801 * A weight of 0 or 1 can cause arithmetics problems.
7802 * (The default weight is 1024 - so there's no practical
7803 * limitation from this.)
7808 mutex_lock(&shares_mutex);
7809 if (tg->shares == shares)
7812 spin_lock_irqsave(&task_group_lock, flags);
7813 for_each_possible_cpu(i)
7814 unregister_fair_sched_group(tg, i);
7815 spin_unlock_irqrestore(&task_group_lock, flags);
7817 /* wait for any ongoing reference to this group to finish */
7818 synchronize_sched();
7821 * Now we are free to modify the group's share on each cpu
7822 * w/o tripping rebalance_share or load_balance_fair.
7824 tg->shares = shares;
7825 for_each_possible_cpu(i)
7826 set_se_shares(tg->se[i], shares);
7829 * Enable load balance activity on this group, by inserting it back on
7830 * each cpu's rq->leaf_cfs_rq_list.
7832 spin_lock_irqsave(&task_group_lock, flags);
7833 for_each_possible_cpu(i)
7834 register_fair_sched_group(tg, i);
7835 spin_unlock_irqrestore(&task_group_lock, flags);
7837 mutex_unlock(&shares_mutex);
7841 unsigned long sched_group_shares(struct task_group *tg)
7847 #ifdef CONFIG_RT_GROUP_SCHED
7849 * Ensure that the real time constraints are schedulable.
7851 static DEFINE_MUTEX(rt_constraints_mutex);
7853 static unsigned long to_ratio(u64 period, u64 runtime)
7855 if (runtime == RUNTIME_INF)
7858 return div64_64(runtime << 16, period);
7861 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7863 struct task_group *tgi;
7864 unsigned long total = 0;
7865 unsigned long global_ratio =
7866 to_ratio(sysctl_sched_rt_period,
7867 sysctl_sched_rt_runtime < 0 ?
7868 RUNTIME_INF : sysctl_sched_rt_runtime);
7871 list_for_each_entry_rcu(tgi, &task_groups, list) {
7875 total += to_ratio(period, tgi->rt_runtime);
7879 return total + to_ratio(period, runtime) < global_ratio;
7882 /* Must be called with tasklist_lock held */
7883 static inline int tg_has_rt_tasks(struct task_group *tg)
7885 struct task_struct *g, *p;
7886 do_each_thread(g, p) {
7887 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
7889 } while_each_thread(g, p);
7893 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7895 u64 rt_runtime, rt_period;
7898 rt_period = (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
7899 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7900 if (rt_runtime_us == -1)
7901 rt_runtime = RUNTIME_INF;
7903 mutex_lock(&rt_constraints_mutex);
7904 read_lock(&tasklist_lock);
7905 if (rt_runtime_us == 0 && tg_has_rt_tasks(tg)) {
7909 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
7913 tg->rt_runtime = rt_runtime;
7915 read_unlock(&tasklist_lock);
7916 mutex_unlock(&rt_constraints_mutex);
7921 long sched_group_rt_runtime(struct task_group *tg)
7925 if (tg->rt_runtime == RUNTIME_INF)
7928 rt_runtime_us = tg->rt_runtime;
7929 do_div(rt_runtime_us, NSEC_PER_USEC);
7930 return rt_runtime_us;
7933 #endif /* CONFIG_GROUP_SCHED */
7935 #ifdef CONFIG_CGROUP_SCHED
7937 /* return corresponding task_group object of a cgroup */
7938 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7940 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7941 struct task_group, css);
7944 static struct cgroup_subsys_state *
7945 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7947 struct task_group *tg;
7949 if (!cgrp->parent) {
7950 /* This is early initialization for the top cgroup */
7951 init_task_group.css.cgroup = cgrp;
7952 return &init_task_group.css;
7955 /* we support only 1-level deep hierarchical scheduler atm */
7956 if (cgrp->parent->parent)
7957 return ERR_PTR(-EINVAL);
7959 tg = sched_create_group();
7961 return ERR_PTR(-ENOMEM);
7963 /* Bind the cgroup to task_group object we just created */
7964 tg->css.cgroup = cgrp;
7970 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7972 struct task_group *tg = cgroup_tg(cgrp);
7974 sched_destroy_group(tg);
7978 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7979 struct task_struct *tsk)
7981 #ifdef CONFIG_RT_GROUP_SCHED
7982 /* Don't accept realtime tasks when there is no way for them to run */
7983 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_runtime == 0)
7986 /* We don't support RT-tasks being in separate groups */
7987 if (tsk->sched_class != &fair_sched_class)
7995 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7996 struct cgroup *old_cont, struct task_struct *tsk)
7998 sched_move_task(tsk);
8001 #ifdef CONFIG_FAIR_GROUP_SCHED
8002 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8005 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8008 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8010 struct task_group *tg = cgroup_tg(cgrp);
8012 return (u64) tg->shares;
8016 #ifdef CONFIG_RT_GROUP_SCHED
8017 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8019 const char __user *userbuf,
8020 size_t nbytes, loff_t *unused_ppos)
8029 if (nbytes >= sizeof(buffer))
8031 if (copy_from_user(buffer, userbuf, nbytes))
8034 buffer[nbytes] = 0; /* nul-terminate */
8036 /* strip newline if necessary */
8037 if (nbytes && (buffer[nbytes-1] == '\n'))
8038 buffer[nbytes-1] = 0;
8039 val = simple_strtoll(buffer, &end, 0);
8043 /* Pass to subsystem */
8044 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8050 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8052 char __user *buf, size_t nbytes,
8056 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8057 int len = sprintf(tmp, "%ld\n", val);
8059 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8063 static struct cftype cpu_files[] = {
8064 #ifdef CONFIG_FAIR_GROUP_SCHED
8067 .read_uint = cpu_shares_read_uint,
8068 .write_uint = cpu_shares_write_uint,
8071 #ifdef CONFIG_RT_GROUP_SCHED
8073 .name = "rt_runtime_us",
8074 .read = cpu_rt_runtime_read,
8075 .write = cpu_rt_runtime_write,
8080 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8082 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8085 struct cgroup_subsys cpu_cgroup_subsys = {
8087 .create = cpu_cgroup_create,
8088 .destroy = cpu_cgroup_destroy,
8089 .can_attach = cpu_cgroup_can_attach,
8090 .attach = cpu_cgroup_attach,
8091 .populate = cpu_cgroup_populate,
8092 .subsys_id = cpu_cgroup_subsys_id,
8096 #endif /* CONFIG_CGROUP_SCHED */
8098 #ifdef CONFIG_CGROUP_CPUACCT
8101 * CPU accounting code for task groups.
8103 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8104 * (balbir@in.ibm.com).
8107 /* track cpu usage of a group of tasks */
8109 struct cgroup_subsys_state css;
8110 /* cpuusage holds pointer to a u64-type object on every cpu */
8114 struct cgroup_subsys cpuacct_subsys;
8116 /* return cpu accounting group corresponding to this container */
8117 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
8119 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
8120 struct cpuacct, css);
8123 /* return cpu accounting group to which this task belongs */
8124 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8126 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8127 struct cpuacct, css);
8130 /* create a new cpu accounting group */
8131 static struct cgroup_subsys_state *cpuacct_create(
8132 struct cgroup_subsys *ss, struct cgroup *cont)
8134 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8137 return ERR_PTR(-ENOMEM);
8139 ca->cpuusage = alloc_percpu(u64);
8140 if (!ca->cpuusage) {
8142 return ERR_PTR(-ENOMEM);
8148 /* destroy an existing cpu accounting group */
8150 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
8152 struct cpuacct *ca = cgroup_ca(cont);
8154 free_percpu(ca->cpuusage);
8158 /* return total cpu usage (in nanoseconds) of a group */
8159 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
8161 struct cpuacct *ca = cgroup_ca(cont);
8162 u64 totalcpuusage = 0;
8165 for_each_possible_cpu(i) {
8166 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8169 * Take rq->lock to make 64-bit addition safe on 32-bit
8172 spin_lock_irq(&cpu_rq(i)->lock);
8173 totalcpuusage += *cpuusage;
8174 spin_unlock_irq(&cpu_rq(i)->lock);
8177 return totalcpuusage;
8180 static struct cftype files[] = {
8183 .read_uint = cpuusage_read,
8187 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8189 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
8193 * charge this task's execution time to its accounting group.
8195 * called with rq->lock held.
8197 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8201 if (!cpuacct_subsys.active)
8206 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8208 *cpuusage += cputime;
8212 struct cgroup_subsys cpuacct_subsys = {
8214 .create = cpuacct_create,
8215 .destroy = cpuacct_destroy,
8216 .populate = cpuacct_populate,
8217 .subsys_id = cpuacct_subsys_id,
8219 #endif /* CONFIG_CGROUP_CPUACCT */