4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
73 #include <asm/irq_regs.h>
76 * Scheduler clock - returns current time in nanosec units.
77 * This is default implementation.
78 * Architectures and sub-architectures can override this.
80 unsigned long long __attribute__((weak)) sched_clock(void)
82 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
126 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
127 * Since cpu_power is a 'constant', we can use a reciprocal divide.
129 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
131 return reciprocal_divide(load, sg->reciprocal_cpu_power);
135 * Each time a sched group cpu_power is changed,
136 * we must compute its reciprocal value
138 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
140 sg->__cpu_power += val;
141 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
145 static inline int rt_policy(int policy)
147 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
152 static inline int task_has_rt_policy(struct task_struct *p)
154 return rt_policy(p->policy);
158 * This is the priority-queue data structure of the RT scheduling class:
160 struct rt_prio_array {
161 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
162 struct list_head queue[MAX_RT_PRIO];
165 struct rt_bandwidth {
166 /* nests inside the rq lock: */
167 spinlock_t rt_runtime_lock;
170 struct hrtimer rt_period_timer;
173 static struct rt_bandwidth def_rt_bandwidth;
175 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
177 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
179 struct rt_bandwidth *rt_b =
180 container_of(timer, struct rt_bandwidth, rt_period_timer);
186 now = hrtimer_cb_get_time(timer);
187 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
192 idle = do_sched_rt_period_timer(rt_b, overrun);
195 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
199 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
201 rt_b->rt_period = ns_to_ktime(period);
202 rt_b->rt_runtime = runtime;
204 spin_lock_init(&rt_b->rt_runtime_lock);
206 hrtimer_init(&rt_b->rt_period_timer,
207 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
208 rt_b->rt_period_timer.function = sched_rt_period_timer;
209 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
212 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
216 if (rt_b->rt_runtime == RUNTIME_INF)
219 if (hrtimer_active(&rt_b->rt_period_timer))
222 spin_lock(&rt_b->rt_runtime_lock);
224 if (hrtimer_active(&rt_b->rt_period_timer))
227 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
228 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
229 hrtimer_start(&rt_b->rt_period_timer,
230 rt_b->rt_period_timer.expires,
233 spin_unlock(&rt_b->rt_runtime_lock);
236 #ifdef CONFIG_RT_GROUP_SCHED
237 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
239 hrtimer_cancel(&rt_b->rt_period_timer);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity **se;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq **cfs_rq;
262 unsigned long shares;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity **rt_se;
267 struct rt_rq **rt_rq;
269 struct rt_bandwidth rt_bandwidth;
273 struct list_head list;
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* Default task group's sched entity on each cpu */
278 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
279 /* Default task group's cfs_rq on each cpu */
280 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
283 #ifdef CONFIG_RT_GROUP_SCHED
284 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
285 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
288 /* task_group_lock serializes add/remove of task groups and also changes to
289 * a task group's cpu shares.
291 static DEFINE_SPINLOCK(task_group_lock);
293 /* doms_cur_mutex serializes access to doms_cur[] array */
294 static DEFINE_MUTEX(doms_cur_mutex);
296 #ifdef CONFIG_FAIR_GROUP_SCHED
297 #ifdef CONFIG_USER_SCHED
298 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
300 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
303 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
306 /* Default task group.
307 * Every task in system belong to this group at bootup.
309 struct task_group init_task_group;
311 /* return group to which a task belongs */
312 static inline struct task_group *task_group(struct task_struct *p)
314 struct task_group *tg;
316 #ifdef CONFIG_USER_SCHED
318 #elif defined(CONFIG_CGROUP_SCHED)
319 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
320 struct task_group, css);
322 tg = &init_task_group;
327 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
328 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
330 #ifdef CONFIG_FAIR_GROUP_SCHED
331 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
332 p->se.parent = task_group(p)->se[cpu];
335 #ifdef CONFIG_RT_GROUP_SCHED
336 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
337 p->rt.parent = task_group(p)->rt_se[cpu];
341 static inline void lock_doms_cur(void)
343 mutex_lock(&doms_cur_mutex);
346 static inline void unlock_doms_cur(void)
348 mutex_unlock(&doms_cur_mutex);
353 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
354 static inline void lock_doms_cur(void) { }
355 static inline void unlock_doms_cur(void) { }
357 #endif /* CONFIG_GROUP_SCHED */
359 /* CFS-related fields in a runqueue */
361 struct load_weight load;
362 unsigned long nr_running;
367 struct rb_root tasks_timeline;
368 struct rb_node *rb_leftmost;
369 struct rb_node *rb_load_balance_curr;
370 /* 'curr' points to currently running entity on this cfs_rq.
371 * It is set to NULL otherwise (i.e when none are currently running).
373 struct sched_entity *curr, *next;
375 unsigned long nr_spread_over;
377 #ifdef CONFIG_FAIR_GROUP_SCHED
378 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
381 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
382 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
383 * (like users, containers etc.)
385 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
386 * list is used during load balance.
388 struct list_head leaf_cfs_rq_list;
389 struct task_group *tg; /* group that "owns" this runqueue */
393 /* Real-Time classes' related field in a runqueue: */
395 struct rt_prio_array active;
396 unsigned long rt_nr_running;
397 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
398 int highest_prio; /* highest queued rt task prio */
401 unsigned long rt_nr_migratory;
407 /* Nests inside the rq lock: */
408 spinlock_t rt_runtime_lock;
410 #ifdef CONFIG_RT_GROUP_SCHED
411 unsigned long rt_nr_boosted;
414 struct list_head leaf_rt_rq_list;
415 struct task_group *tg;
416 struct sched_rt_entity *rt_se;
423 * We add the notion of a root-domain which will be used to define per-domain
424 * variables. Each exclusive cpuset essentially defines an island domain by
425 * fully partitioning the member cpus from any other cpuset. Whenever a new
426 * exclusive cpuset is created, we also create and attach a new root-domain
436 * The "RT overload" flag: it gets set if a CPU has more than
437 * one runnable RT task.
444 * By default the system creates a single root-domain with all cpus as
445 * members (mimicking the global state we have today).
447 static struct root_domain def_root_domain;
452 * This is the main, per-CPU runqueue data structure.
454 * Locking rule: those places that want to lock multiple runqueues
455 * (such as the load balancing or the thread migration code), lock
456 * acquire operations must be ordered by ascending &runqueue.
463 * nr_running and cpu_load should be in the same cacheline because
464 * remote CPUs use both these fields when doing load calculation.
466 unsigned long nr_running;
467 #define CPU_LOAD_IDX_MAX 5
468 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
469 unsigned char idle_at_tick;
471 unsigned long last_tick_seen;
472 unsigned char in_nohz_recently;
474 /* capture load from *all* tasks on this cpu: */
475 struct load_weight load;
476 unsigned long nr_load_updates;
482 #ifdef CONFIG_FAIR_GROUP_SCHED
483 /* list of leaf cfs_rq on this cpu: */
484 struct list_head leaf_cfs_rq_list;
486 #ifdef CONFIG_RT_GROUP_SCHED
487 struct list_head leaf_rt_rq_list;
491 * This is part of a global counter where only the total sum
492 * over all CPUs matters. A task can increase this counter on
493 * one CPU and if it got migrated afterwards it may decrease
494 * it on another CPU. Always updated under the runqueue lock:
496 unsigned long nr_uninterruptible;
498 struct task_struct *curr, *idle;
499 unsigned long next_balance;
500 struct mm_struct *prev_mm;
502 u64 clock, prev_clock_raw;
505 unsigned int clock_warps, clock_overflows, clock_underflows;
507 unsigned int clock_deep_idle_events;
513 struct root_domain *rd;
514 struct sched_domain *sd;
516 /* For active balancing */
519 /* cpu of this runqueue: */
522 struct task_struct *migration_thread;
523 struct list_head migration_queue;
526 #ifdef CONFIG_SCHED_HRTICK
527 unsigned long hrtick_flags;
528 ktime_t hrtick_expire;
529 struct hrtimer hrtick_timer;
532 #ifdef CONFIG_SCHEDSTATS
534 struct sched_info rq_sched_info;
536 /* sys_sched_yield() stats */
537 unsigned int yld_exp_empty;
538 unsigned int yld_act_empty;
539 unsigned int yld_both_empty;
540 unsigned int yld_count;
542 /* schedule() stats */
543 unsigned int sched_switch;
544 unsigned int sched_count;
545 unsigned int sched_goidle;
547 /* try_to_wake_up() stats */
548 unsigned int ttwu_count;
549 unsigned int ttwu_local;
552 unsigned int bkl_count;
554 struct lock_class_key rq_lock_key;
557 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
559 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
561 rq->curr->sched_class->check_preempt_curr(rq, p);
564 static inline int cpu_of(struct rq *rq)
574 static inline bool nohz_on(int cpu)
576 return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
579 static inline u64 max_skipped_ticks(struct rq *rq)
581 return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
584 static inline void update_last_tick_seen(struct rq *rq)
586 rq->last_tick_seen = jiffies;
589 static inline u64 max_skipped_ticks(struct rq *rq)
594 static inline void update_last_tick_seen(struct rq *rq)
600 * Update the per-runqueue clock, as finegrained as the platform can give
601 * us, but without assuming monotonicity, etc.:
603 static void __update_rq_clock(struct rq *rq)
605 u64 prev_raw = rq->prev_clock_raw;
606 u64 now = sched_clock();
607 s64 delta = now - prev_raw;
608 u64 clock = rq->clock;
610 #ifdef CONFIG_SCHED_DEBUG
611 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
614 * Protect against sched_clock() occasionally going backwards:
616 if (unlikely(delta < 0)) {
621 * Catch too large forward jumps too:
623 u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
624 u64 max_time = rq->tick_timestamp + max_jump;
626 if (unlikely(clock + delta > max_time)) {
627 if (clock < max_time)
631 rq->clock_overflows++;
633 if (unlikely(delta > rq->clock_max_delta))
634 rq->clock_max_delta = delta;
639 rq->prev_clock_raw = now;
643 static void update_rq_clock(struct rq *rq)
645 if (likely(smp_processor_id() == cpu_of(rq)))
646 __update_rq_clock(rq);
650 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
651 * See detach_destroy_domains: synchronize_sched for details.
653 * The domain tree of any CPU may only be accessed from within
654 * preempt-disabled sections.
656 #define for_each_domain(cpu, __sd) \
657 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
659 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
660 #define this_rq() (&__get_cpu_var(runqueues))
661 #define task_rq(p) cpu_rq(task_cpu(p))
662 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
665 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
667 #ifdef CONFIG_SCHED_DEBUG
668 # define const_debug __read_mostly
670 # define const_debug static const
674 * Debugging: various feature bits
677 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
678 SCHED_FEAT_WAKEUP_PREEMPT = 2,
679 SCHED_FEAT_START_DEBIT = 4,
680 SCHED_FEAT_AFFINE_WAKEUPS = 8,
681 SCHED_FEAT_CACHE_HOT_BUDDY = 16,
682 SCHED_FEAT_SYNC_WAKEUPS = 32,
683 SCHED_FEAT_HRTICK = 64,
684 SCHED_FEAT_DOUBLE_TICK = 128,
685 SCHED_FEAT_NORMALIZED_SLEEPER = 256,
688 const_debug unsigned int sysctl_sched_features =
689 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
690 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
691 SCHED_FEAT_START_DEBIT * 1 |
692 SCHED_FEAT_AFFINE_WAKEUPS * 1 |
693 SCHED_FEAT_CACHE_HOT_BUDDY * 1 |
694 SCHED_FEAT_SYNC_WAKEUPS * 1 |
695 SCHED_FEAT_HRTICK * 1 |
696 SCHED_FEAT_DOUBLE_TICK * 0 |
697 SCHED_FEAT_NORMALIZED_SLEEPER * 1;
699 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
702 * Number of tasks to iterate in a single balance run.
703 * Limited because this is done with IRQs disabled.
705 const_debug unsigned int sysctl_sched_nr_migrate = 32;
708 * period over which we measure -rt task cpu usage in us.
711 unsigned int sysctl_sched_rt_period = 1000000;
713 static __read_mostly int scheduler_running;
716 * part of the period that we allow rt tasks to run in us.
719 int sysctl_sched_rt_runtime = 950000;
721 static inline u64 global_rt_period(void)
723 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
726 static inline u64 global_rt_runtime(void)
728 if (sysctl_sched_rt_period < 0)
731 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
734 static const unsigned long long time_sync_thresh = 100000;
736 static DEFINE_PER_CPU(unsigned long long, time_offset);
737 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
740 * Global lock which we take every now and then to synchronize
741 * the CPUs time. This method is not warp-safe, but it's good
742 * enough to synchronize slowly diverging time sources and thus
743 * it's good enough for tracing:
745 static DEFINE_SPINLOCK(time_sync_lock);
746 static unsigned long long prev_global_time;
748 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
752 spin_lock_irqsave(&time_sync_lock, flags);
754 if (time < prev_global_time) {
755 per_cpu(time_offset, cpu) += prev_global_time - time;
756 time = prev_global_time;
758 prev_global_time = time;
761 spin_unlock_irqrestore(&time_sync_lock, flags);
766 static unsigned long long __cpu_clock(int cpu)
768 unsigned long long now;
773 * Only call sched_clock() if the scheduler has already been
774 * initialized (some code might call cpu_clock() very early):
776 if (unlikely(!scheduler_running))
779 local_irq_save(flags);
783 local_irq_restore(flags);
789 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
790 * clock constructed from sched_clock():
792 unsigned long long cpu_clock(int cpu)
794 unsigned long long prev_cpu_time, time, delta_time;
796 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
797 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
798 delta_time = time-prev_cpu_time;
800 if (unlikely(delta_time > time_sync_thresh))
801 time = __sync_cpu_clock(time, cpu);
805 EXPORT_SYMBOL_GPL(cpu_clock);
807 #ifndef prepare_arch_switch
808 # define prepare_arch_switch(next) do { } while (0)
810 #ifndef finish_arch_switch
811 # define finish_arch_switch(prev) do { } while (0)
814 static inline int task_current(struct rq *rq, struct task_struct *p)
816 return rq->curr == p;
819 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
820 static inline int task_running(struct rq *rq, struct task_struct *p)
822 return task_current(rq, p);
825 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
829 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
831 #ifdef CONFIG_DEBUG_SPINLOCK
832 /* this is a valid case when another task releases the spinlock */
833 rq->lock.owner = current;
836 * If we are tracking spinlock dependencies then we have to
837 * fix up the runqueue lock - which gets 'carried over' from
840 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
842 spin_unlock_irq(&rq->lock);
845 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
846 static inline int task_running(struct rq *rq, struct task_struct *p)
851 return task_current(rq, p);
855 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
859 * We can optimise this out completely for !SMP, because the
860 * SMP rebalancing from interrupt is the only thing that cares
865 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
866 spin_unlock_irq(&rq->lock);
868 spin_unlock(&rq->lock);
872 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
876 * After ->oncpu is cleared, the task can be moved to a different CPU.
877 * We must ensure this doesn't happen until the switch is completely
883 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
887 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
890 * __task_rq_lock - lock the runqueue a given task resides on.
891 * Must be called interrupts disabled.
893 static inline struct rq *__task_rq_lock(struct task_struct *p)
897 struct rq *rq = task_rq(p);
898 spin_lock(&rq->lock);
899 if (likely(rq == task_rq(p)))
901 spin_unlock(&rq->lock);
906 * task_rq_lock - lock the runqueue a given task resides on and disable
907 * interrupts. Note the ordering: we can safely lookup the task_rq without
908 * explicitly disabling preemption.
910 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
916 local_irq_save(*flags);
918 spin_lock(&rq->lock);
919 if (likely(rq == task_rq(p)))
921 spin_unlock_irqrestore(&rq->lock, *flags);
925 static void __task_rq_unlock(struct rq *rq)
928 spin_unlock(&rq->lock);
931 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
934 spin_unlock_irqrestore(&rq->lock, *flags);
938 * this_rq_lock - lock this runqueue and disable interrupts.
940 static struct rq *this_rq_lock(void)
947 spin_lock(&rq->lock);
953 * We are going deep-idle (irqs are disabled):
955 void sched_clock_idle_sleep_event(void)
957 struct rq *rq = cpu_rq(smp_processor_id());
959 spin_lock(&rq->lock);
960 __update_rq_clock(rq);
961 spin_unlock(&rq->lock);
962 rq->clock_deep_idle_events++;
964 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
967 * We just idled delta nanoseconds (called with irqs disabled):
969 void sched_clock_idle_wakeup_event(u64 delta_ns)
971 struct rq *rq = cpu_rq(smp_processor_id());
972 u64 now = sched_clock();
974 rq->idle_clock += delta_ns;
976 * Override the previous timestamp and ignore all
977 * sched_clock() deltas that occured while we idled,
978 * and use the PM-provided delta_ns to advance the
981 spin_lock(&rq->lock);
982 rq->prev_clock_raw = now;
983 rq->clock += delta_ns;
984 spin_unlock(&rq->lock);
985 touch_softlockup_watchdog();
987 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
989 static void __resched_task(struct task_struct *p, int tif_bit);
991 static inline void resched_task(struct task_struct *p)
993 __resched_task(p, TIF_NEED_RESCHED);
996 #ifdef CONFIG_SCHED_HRTICK
998 * Use HR-timers to deliver accurate preemption points.
1000 * Its all a bit involved since we cannot program an hrt while holding the
1001 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1004 * When we get rescheduled we reprogram the hrtick_timer outside of the
1007 static inline void resched_hrt(struct task_struct *p)
1009 __resched_task(p, TIF_HRTICK_RESCHED);
1012 static inline void resched_rq(struct rq *rq)
1014 unsigned long flags;
1016 spin_lock_irqsave(&rq->lock, flags);
1017 resched_task(rq->curr);
1018 spin_unlock_irqrestore(&rq->lock, flags);
1022 HRTICK_SET, /* re-programm hrtick_timer */
1023 HRTICK_RESET, /* not a new slice */
1028 * - enabled by features
1029 * - hrtimer is actually high res
1031 static inline int hrtick_enabled(struct rq *rq)
1033 if (!sched_feat(HRTICK))
1035 return hrtimer_is_hres_active(&rq->hrtick_timer);
1039 * Called to set the hrtick timer state.
1041 * called with rq->lock held and irqs disabled
1043 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1045 assert_spin_locked(&rq->lock);
1048 * preempt at: now + delay
1051 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1053 * indicate we need to program the timer
1055 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1057 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1060 * New slices are called from the schedule path and don't need a
1061 * forced reschedule.
1064 resched_hrt(rq->curr);
1067 static void hrtick_clear(struct rq *rq)
1069 if (hrtimer_active(&rq->hrtick_timer))
1070 hrtimer_cancel(&rq->hrtick_timer);
1074 * Update the timer from the possible pending state.
1076 static void hrtick_set(struct rq *rq)
1080 unsigned long flags;
1082 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1084 spin_lock_irqsave(&rq->lock, flags);
1085 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1086 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1087 time = rq->hrtick_expire;
1088 clear_thread_flag(TIF_HRTICK_RESCHED);
1089 spin_unlock_irqrestore(&rq->lock, flags);
1092 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1093 if (reset && !hrtimer_active(&rq->hrtick_timer))
1100 * High-resolution timer tick.
1101 * Runs from hardirq context with interrupts disabled.
1103 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1105 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1107 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1109 spin_lock(&rq->lock);
1110 __update_rq_clock(rq);
1111 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1112 spin_unlock(&rq->lock);
1114 return HRTIMER_NORESTART;
1117 static inline void init_rq_hrtick(struct rq *rq)
1119 rq->hrtick_flags = 0;
1120 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1121 rq->hrtick_timer.function = hrtick;
1122 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1125 void hrtick_resched(void)
1128 unsigned long flags;
1130 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1133 local_irq_save(flags);
1134 rq = cpu_rq(smp_processor_id());
1136 local_irq_restore(flags);
1139 static inline void hrtick_clear(struct rq *rq)
1143 static inline void hrtick_set(struct rq *rq)
1147 static inline void init_rq_hrtick(struct rq *rq)
1151 void hrtick_resched(void)
1157 * resched_task - mark a task 'to be rescheduled now'.
1159 * On UP this means the setting of the need_resched flag, on SMP it
1160 * might also involve a cross-CPU call to trigger the scheduler on
1165 #ifndef tsk_is_polling
1166 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1169 static void __resched_task(struct task_struct *p, int tif_bit)
1173 assert_spin_locked(&task_rq(p)->lock);
1175 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1178 set_tsk_thread_flag(p, tif_bit);
1181 if (cpu == smp_processor_id())
1184 /* NEED_RESCHED must be visible before we test polling */
1186 if (!tsk_is_polling(p))
1187 smp_send_reschedule(cpu);
1190 static void resched_cpu(int cpu)
1192 struct rq *rq = cpu_rq(cpu);
1193 unsigned long flags;
1195 if (!spin_trylock_irqsave(&rq->lock, flags))
1197 resched_task(cpu_curr(cpu));
1198 spin_unlock_irqrestore(&rq->lock, flags);
1203 * When add_timer_on() enqueues a timer into the timer wheel of an
1204 * idle CPU then this timer might expire before the next timer event
1205 * which is scheduled to wake up that CPU. In case of a completely
1206 * idle system the next event might even be infinite time into the
1207 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1208 * leaves the inner idle loop so the newly added timer is taken into
1209 * account when the CPU goes back to idle and evaluates the timer
1210 * wheel for the next timer event.
1212 void wake_up_idle_cpu(int cpu)
1214 struct rq *rq = cpu_rq(cpu);
1216 if (cpu == smp_processor_id())
1220 * This is safe, as this function is called with the timer
1221 * wheel base lock of (cpu) held. When the CPU is on the way
1222 * to idle and has not yet set rq->curr to idle then it will
1223 * be serialized on the timer wheel base lock and take the new
1224 * timer into account automatically.
1226 if (rq->curr != rq->idle)
1230 * We can set TIF_RESCHED on the idle task of the other CPU
1231 * lockless. The worst case is that the other CPU runs the
1232 * idle task through an additional NOOP schedule()
1234 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1236 /* NEED_RESCHED must be visible before we test polling */
1238 if (!tsk_is_polling(rq->idle))
1239 smp_send_reschedule(cpu);
1244 static void __resched_task(struct task_struct *p, int tif_bit)
1246 assert_spin_locked(&task_rq(p)->lock);
1247 set_tsk_thread_flag(p, tif_bit);
1251 #if BITS_PER_LONG == 32
1252 # define WMULT_CONST (~0UL)
1254 # define WMULT_CONST (1UL << 32)
1257 #define WMULT_SHIFT 32
1260 * Shift right and round:
1262 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1264 static unsigned long
1265 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1266 struct load_weight *lw)
1270 if (unlikely(!lw->inv_weight))
1271 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1273 tmp = (u64)delta_exec * weight;
1275 * Check whether we'd overflow the 64-bit multiplication:
1277 if (unlikely(tmp > WMULT_CONST))
1278 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1281 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1283 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1286 static inline unsigned long
1287 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1289 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1292 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1298 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1305 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1306 * of tasks with abnormal "nice" values across CPUs the contribution that
1307 * each task makes to its run queue's load is weighted according to its
1308 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1309 * scaled version of the new time slice allocation that they receive on time
1313 #define WEIGHT_IDLEPRIO 2
1314 #define WMULT_IDLEPRIO (1 << 31)
1317 * Nice levels are multiplicative, with a gentle 10% change for every
1318 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1319 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1320 * that remained on nice 0.
1322 * The "10% effect" is relative and cumulative: from _any_ nice level,
1323 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1324 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1325 * If a task goes up by ~10% and another task goes down by ~10% then
1326 * the relative distance between them is ~25%.)
1328 static const int prio_to_weight[40] = {
1329 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1330 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1331 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1332 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1333 /* 0 */ 1024, 820, 655, 526, 423,
1334 /* 5 */ 335, 272, 215, 172, 137,
1335 /* 10 */ 110, 87, 70, 56, 45,
1336 /* 15 */ 36, 29, 23, 18, 15,
1340 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1342 * In cases where the weight does not change often, we can use the
1343 * precalculated inverse to speed up arithmetics by turning divisions
1344 * into multiplications:
1346 static const u32 prio_to_wmult[40] = {
1347 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1348 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1349 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1350 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1351 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1352 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1353 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1354 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1357 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1360 * runqueue iterator, to support SMP load-balancing between different
1361 * scheduling classes, without having to expose their internal data
1362 * structures to the load-balancing proper:
1364 struct rq_iterator {
1366 struct task_struct *(*start)(void *);
1367 struct task_struct *(*next)(void *);
1371 static unsigned long
1372 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1373 unsigned long max_load_move, struct sched_domain *sd,
1374 enum cpu_idle_type idle, int *all_pinned,
1375 int *this_best_prio, struct rq_iterator *iterator);
1378 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1379 struct sched_domain *sd, enum cpu_idle_type idle,
1380 struct rq_iterator *iterator);
1383 #ifdef CONFIG_CGROUP_CPUACCT
1384 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1386 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1390 static unsigned long source_load(int cpu, int type);
1391 static unsigned long target_load(int cpu, int type);
1392 static unsigned long cpu_avg_load_per_task(int cpu);
1393 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1394 #endif /* CONFIG_SMP */
1396 #include "sched_stats.h"
1397 #include "sched_idletask.c"
1398 #include "sched_fair.c"
1399 #include "sched_rt.c"
1400 #ifdef CONFIG_SCHED_DEBUG
1401 # include "sched_debug.c"
1404 #define sched_class_highest (&rt_sched_class)
1406 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1408 update_load_add(&rq->load, p->se.load.weight);
1411 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1413 update_load_sub(&rq->load, p->se.load.weight);
1416 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1422 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1428 static void set_load_weight(struct task_struct *p)
1430 if (task_has_rt_policy(p)) {
1431 p->se.load.weight = prio_to_weight[0] * 2;
1432 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1437 * SCHED_IDLE tasks get minimal weight:
1439 if (p->policy == SCHED_IDLE) {
1440 p->se.load.weight = WEIGHT_IDLEPRIO;
1441 p->se.load.inv_weight = WMULT_IDLEPRIO;
1445 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1446 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1449 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1451 sched_info_queued(p);
1452 p->sched_class->enqueue_task(rq, p, wakeup);
1456 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1458 p->sched_class->dequeue_task(rq, p, sleep);
1463 * __normal_prio - return the priority that is based on the static prio
1465 static inline int __normal_prio(struct task_struct *p)
1467 return p->static_prio;
1471 * Calculate the expected normal priority: i.e. priority
1472 * without taking RT-inheritance into account. Might be
1473 * boosted by interactivity modifiers. Changes upon fork,
1474 * setprio syscalls, and whenever the interactivity
1475 * estimator recalculates.
1477 static inline int normal_prio(struct task_struct *p)
1481 if (task_has_rt_policy(p))
1482 prio = MAX_RT_PRIO-1 - p->rt_priority;
1484 prio = __normal_prio(p);
1489 * Calculate the current priority, i.e. the priority
1490 * taken into account by the scheduler. This value might
1491 * be boosted by RT tasks, or might be boosted by
1492 * interactivity modifiers. Will be RT if the task got
1493 * RT-boosted. If not then it returns p->normal_prio.
1495 static int effective_prio(struct task_struct *p)
1497 p->normal_prio = normal_prio(p);
1499 * If we are RT tasks or we were boosted to RT priority,
1500 * keep the priority unchanged. Otherwise, update priority
1501 * to the normal priority:
1503 if (!rt_prio(p->prio))
1504 return p->normal_prio;
1509 * activate_task - move a task to the runqueue.
1511 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1513 if (task_contributes_to_load(p))
1514 rq->nr_uninterruptible--;
1516 enqueue_task(rq, p, wakeup);
1517 inc_nr_running(p, rq);
1521 * deactivate_task - remove a task from the runqueue.
1523 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1525 if (task_contributes_to_load(p))
1526 rq->nr_uninterruptible++;
1528 dequeue_task(rq, p, sleep);
1529 dec_nr_running(p, rq);
1533 * task_curr - is this task currently executing on a CPU?
1534 * @p: the task in question.
1536 inline int task_curr(const struct task_struct *p)
1538 return cpu_curr(task_cpu(p)) == p;
1541 /* Used instead of source_load when we know the type == 0 */
1542 unsigned long weighted_cpuload(const int cpu)
1544 return cpu_rq(cpu)->load.weight;
1547 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1549 set_task_rq(p, cpu);
1552 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1553 * successfuly executed on another CPU. We must ensure that updates of
1554 * per-task data have been completed by this moment.
1557 task_thread_info(p)->cpu = cpu;
1561 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1562 const struct sched_class *prev_class,
1563 int oldprio, int running)
1565 if (prev_class != p->sched_class) {
1566 if (prev_class->switched_from)
1567 prev_class->switched_from(rq, p, running);
1568 p->sched_class->switched_to(rq, p, running);
1570 p->sched_class->prio_changed(rq, p, oldprio, running);
1576 * Is this task likely cache-hot:
1579 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1584 * Buddy candidates are cache hot:
1586 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1589 if (p->sched_class != &fair_sched_class)
1592 if (sysctl_sched_migration_cost == -1)
1594 if (sysctl_sched_migration_cost == 0)
1597 delta = now - p->se.exec_start;
1599 return delta < (s64)sysctl_sched_migration_cost;
1603 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1605 int old_cpu = task_cpu(p);
1606 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1607 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1608 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1611 clock_offset = old_rq->clock - new_rq->clock;
1613 #ifdef CONFIG_SCHEDSTATS
1614 if (p->se.wait_start)
1615 p->se.wait_start -= clock_offset;
1616 if (p->se.sleep_start)
1617 p->se.sleep_start -= clock_offset;
1618 if (p->se.block_start)
1619 p->se.block_start -= clock_offset;
1620 if (old_cpu != new_cpu) {
1621 schedstat_inc(p, se.nr_migrations);
1622 if (task_hot(p, old_rq->clock, NULL))
1623 schedstat_inc(p, se.nr_forced2_migrations);
1626 p->se.vruntime -= old_cfsrq->min_vruntime -
1627 new_cfsrq->min_vruntime;
1629 __set_task_cpu(p, new_cpu);
1632 struct migration_req {
1633 struct list_head list;
1635 struct task_struct *task;
1638 struct completion done;
1642 * The task's runqueue lock must be held.
1643 * Returns true if you have to wait for migration thread.
1646 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1648 struct rq *rq = task_rq(p);
1651 * If the task is not on a runqueue (and not running), then
1652 * it is sufficient to simply update the task's cpu field.
1654 if (!p->se.on_rq && !task_running(rq, p)) {
1655 set_task_cpu(p, dest_cpu);
1659 init_completion(&req->done);
1661 req->dest_cpu = dest_cpu;
1662 list_add(&req->list, &rq->migration_queue);
1668 * wait_task_inactive - wait for a thread to unschedule.
1670 * The caller must ensure that the task *will* unschedule sometime soon,
1671 * else this function might spin for a *long* time. This function can't
1672 * be called with interrupts off, or it may introduce deadlock with
1673 * smp_call_function() if an IPI is sent by the same process we are
1674 * waiting to become inactive.
1676 void wait_task_inactive(struct task_struct *p)
1678 unsigned long flags;
1684 * We do the initial early heuristics without holding
1685 * any task-queue locks at all. We'll only try to get
1686 * the runqueue lock when things look like they will
1692 * If the task is actively running on another CPU
1693 * still, just relax and busy-wait without holding
1696 * NOTE! Since we don't hold any locks, it's not
1697 * even sure that "rq" stays as the right runqueue!
1698 * But we don't care, since "task_running()" will
1699 * return false if the runqueue has changed and p
1700 * is actually now running somewhere else!
1702 while (task_running(rq, p))
1706 * Ok, time to look more closely! We need the rq
1707 * lock now, to be *sure*. If we're wrong, we'll
1708 * just go back and repeat.
1710 rq = task_rq_lock(p, &flags);
1711 running = task_running(rq, p);
1712 on_rq = p->se.on_rq;
1713 task_rq_unlock(rq, &flags);
1716 * Was it really running after all now that we
1717 * checked with the proper locks actually held?
1719 * Oops. Go back and try again..
1721 if (unlikely(running)) {
1727 * It's not enough that it's not actively running,
1728 * it must be off the runqueue _entirely_, and not
1731 * So if it wa still runnable (but just not actively
1732 * running right now), it's preempted, and we should
1733 * yield - it could be a while.
1735 if (unlikely(on_rq)) {
1736 schedule_timeout_uninterruptible(1);
1741 * Ahh, all good. It wasn't running, and it wasn't
1742 * runnable, which means that it will never become
1743 * running in the future either. We're all done!
1750 * kick_process - kick a running thread to enter/exit the kernel
1751 * @p: the to-be-kicked thread
1753 * Cause a process which is running on another CPU to enter
1754 * kernel-mode, without any delay. (to get signals handled.)
1756 * NOTE: this function doesnt have to take the runqueue lock,
1757 * because all it wants to ensure is that the remote task enters
1758 * the kernel. If the IPI races and the task has been migrated
1759 * to another CPU then no harm is done and the purpose has been
1762 void kick_process(struct task_struct *p)
1768 if ((cpu != smp_processor_id()) && task_curr(p))
1769 smp_send_reschedule(cpu);
1774 * Return a low guess at the load of a migration-source cpu weighted
1775 * according to the scheduling class and "nice" value.
1777 * We want to under-estimate the load of migration sources, to
1778 * balance conservatively.
1780 static unsigned long source_load(int cpu, int type)
1782 struct rq *rq = cpu_rq(cpu);
1783 unsigned long total = weighted_cpuload(cpu);
1788 return min(rq->cpu_load[type-1], total);
1792 * Return a high guess at the load of a migration-target cpu weighted
1793 * according to the scheduling class and "nice" value.
1795 static unsigned long target_load(int cpu, int type)
1797 struct rq *rq = cpu_rq(cpu);
1798 unsigned long total = weighted_cpuload(cpu);
1803 return max(rq->cpu_load[type-1], total);
1807 * Return the average load per task on the cpu's run queue
1809 static unsigned long cpu_avg_load_per_task(int cpu)
1811 struct rq *rq = cpu_rq(cpu);
1812 unsigned long total = weighted_cpuload(cpu);
1813 unsigned long n = rq->nr_running;
1815 return n ? total / n : SCHED_LOAD_SCALE;
1819 * find_idlest_group finds and returns the least busy CPU group within the
1822 static struct sched_group *
1823 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1825 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1826 unsigned long min_load = ULONG_MAX, this_load = 0;
1827 int load_idx = sd->forkexec_idx;
1828 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1831 unsigned long load, avg_load;
1835 /* Skip over this group if it has no CPUs allowed */
1836 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1839 local_group = cpu_isset(this_cpu, group->cpumask);
1841 /* Tally up the load of all CPUs in the group */
1844 for_each_cpu_mask(i, group->cpumask) {
1845 /* Bias balancing toward cpus of our domain */
1847 load = source_load(i, load_idx);
1849 load = target_load(i, load_idx);
1854 /* Adjust by relative CPU power of the group */
1855 avg_load = sg_div_cpu_power(group,
1856 avg_load * SCHED_LOAD_SCALE);
1859 this_load = avg_load;
1861 } else if (avg_load < min_load) {
1862 min_load = avg_load;
1865 } while (group = group->next, group != sd->groups);
1867 if (!idlest || 100*this_load < imbalance*min_load)
1873 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1876 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1879 unsigned long load, min_load = ULONG_MAX;
1883 /* Traverse only the allowed CPUs */
1884 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1886 for_each_cpu_mask(i, *tmp) {
1887 load = weighted_cpuload(i);
1889 if (load < min_load || (load == min_load && i == this_cpu)) {
1899 * sched_balance_self: balance the current task (running on cpu) in domains
1900 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1903 * Balance, ie. select the least loaded group.
1905 * Returns the target CPU number, or the same CPU if no balancing is needed.
1907 * preempt must be disabled.
1909 static int sched_balance_self(int cpu, int flag)
1911 struct task_struct *t = current;
1912 struct sched_domain *tmp, *sd = NULL;
1914 for_each_domain(cpu, tmp) {
1916 * If power savings logic is enabled for a domain, stop there.
1918 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1920 if (tmp->flags & flag)
1925 cpumask_t span, tmpmask;
1926 struct sched_group *group;
1927 int new_cpu, weight;
1929 if (!(sd->flags & flag)) {
1935 group = find_idlest_group(sd, t, cpu);
1941 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
1942 if (new_cpu == -1 || new_cpu == cpu) {
1943 /* Now try balancing at a lower domain level of cpu */
1948 /* Now try balancing at a lower domain level of new_cpu */
1951 weight = cpus_weight(span);
1952 for_each_domain(cpu, tmp) {
1953 if (weight <= cpus_weight(tmp->span))
1955 if (tmp->flags & flag)
1958 /* while loop will break here if sd == NULL */
1964 #endif /* CONFIG_SMP */
1967 * try_to_wake_up - wake up a thread
1968 * @p: the to-be-woken-up thread
1969 * @state: the mask of task states that can be woken
1970 * @sync: do a synchronous wakeup?
1972 * Put it on the run-queue if it's not already there. The "current"
1973 * thread is always on the run-queue (except when the actual
1974 * re-schedule is in progress), and as such you're allowed to do
1975 * the simpler "current->state = TASK_RUNNING" to mark yourself
1976 * runnable without the overhead of this.
1978 * returns failure only if the task is already active.
1980 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1982 int cpu, orig_cpu, this_cpu, success = 0;
1983 unsigned long flags;
1987 if (!sched_feat(SYNC_WAKEUPS))
1991 rq = task_rq_lock(p, &flags);
1992 old_state = p->state;
1993 if (!(old_state & state))
2001 this_cpu = smp_processor_id();
2004 if (unlikely(task_running(rq, p)))
2007 cpu = p->sched_class->select_task_rq(p, sync);
2008 if (cpu != orig_cpu) {
2009 set_task_cpu(p, cpu);
2010 task_rq_unlock(rq, &flags);
2011 /* might preempt at this point */
2012 rq = task_rq_lock(p, &flags);
2013 old_state = p->state;
2014 if (!(old_state & state))
2019 this_cpu = smp_processor_id();
2023 #ifdef CONFIG_SCHEDSTATS
2024 schedstat_inc(rq, ttwu_count);
2025 if (cpu == this_cpu)
2026 schedstat_inc(rq, ttwu_local);
2028 struct sched_domain *sd;
2029 for_each_domain(this_cpu, sd) {
2030 if (cpu_isset(cpu, sd->span)) {
2031 schedstat_inc(sd, ttwu_wake_remote);
2039 #endif /* CONFIG_SMP */
2040 schedstat_inc(p, se.nr_wakeups);
2042 schedstat_inc(p, se.nr_wakeups_sync);
2043 if (orig_cpu != cpu)
2044 schedstat_inc(p, se.nr_wakeups_migrate);
2045 if (cpu == this_cpu)
2046 schedstat_inc(p, se.nr_wakeups_local);
2048 schedstat_inc(p, se.nr_wakeups_remote);
2049 update_rq_clock(rq);
2050 activate_task(rq, p, 1);
2054 check_preempt_curr(rq, p);
2056 p->state = TASK_RUNNING;
2058 if (p->sched_class->task_wake_up)
2059 p->sched_class->task_wake_up(rq, p);
2062 task_rq_unlock(rq, &flags);
2067 int wake_up_process(struct task_struct *p)
2069 return try_to_wake_up(p, TASK_ALL, 0);
2071 EXPORT_SYMBOL(wake_up_process);
2073 int wake_up_state(struct task_struct *p, unsigned int state)
2075 return try_to_wake_up(p, state, 0);
2079 * Perform scheduler related setup for a newly forked process p.
2080 * p is forked by current.
2082 * __sched_fork() is basic setup used by init_idle() too:
2084 static void __sched_fork(struct task_struct *p)
2086 p->se.exec_start = 0;
2087 p->se.sum_exec_runtime = 0;
2088 p->se.prev_sum_exec_runtime = 0;
2089 p->se.last_wakeup = 0;
2090 p->se.avg_overlap = 0;
2092 #ifdef CONFIG_SCHEDSTATS
2093 p->se.wait_start = 0;
2094 p->se.sum_sleep_runtime = 0;
2095 p->se.sleep_start = 0;
2096 p->se.block_start = 0;
2097 p->se.sleep_max = 0;
2098 p->se.block_max = 0;
2100 p->se.slice_max = 0;
2104 INIT_LIST_HEAD(&p->rt.run_list);
2107 #ifdef CONFIG_PREEMPT_NOTIFIERS
2108 INIT_HLIST_HEAD(&p->preempt_notifiers);
2112 * We mark the process as running here, but have not actually
2113 * inserted it onto the runqueue yet. This guarantees that
2114 * nobody will actually run it, and a signal or other external
2115 * event cannot wake it up and insert it on the runqueue either.
2117 p->state = TASK_RUNNING;
2121 * fork()/clone()-time setup:
2123 void sched_fork(struct task_struct *p, int clone_flags)
2125 int cpu = get_cpu();
2130 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2132 set_task_cpu(p, cpu);
2135 * Make sure we do not leak PI boosting priority to the child:
2137 p->prio = current->normal_prio;
2138 if (!rt_prio(p->prio))
2139 p->sched_class = &fair_sched_class;
2141 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2142 if (likely(sched_info_on()))
2143 memset(&p->sched_info, 0, sizeof(p->sched_info));
2145 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2148 #ifdef CONFIG_PREEMPT
2149 /* Want to start with kernel preemption disabled. */
2150 task_thread_info(p)->preempt_count = 1;
2156 * wake_up_new_task - wake up a newly created task for the first time.
2158 * This function will do some initial scheduler statistics housekeeping
2159 * that must be done for every newly created context, then puts the task
2160 * on the runqueue and wakes it.
2162 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2164 unsigned long flags;
2167 rq = task_rq_lock(p, &flags);
2168 BUG_ON(p->state != TASK_RUNNING);
2169 update_rq_clock(rq);
2171 p->prio = effective_prio(p);
2173 if (!p->sched_class->task_new || !current->se.on_rq) {
2174 activate_task(rq, p, 0);
2177 * Let the scheduling class do new task startup
2178 * management (if any):
2180 p->sched_class->task_new(rq, p);
2181 inc_nr_running(p, rq);
2183 check_preempt_curr(rq, p);
2185 if (p->sched_class->task_wake_up)
2186 p->sched_class->task_wake_up(rq, p);
2188 task_rq_unlock(rq, &flags);
2191 #ifdef CONFIG_PREEMPT_NOTIFIERS
2194 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2195 * @notifier: notifier struct to register
2197 void preempt_notifier_register(struct preempt_notifier *notifier)
2199 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2201 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2204 * preempt_notifier_unregister - no longer interested in preemption notifications
2205 * @notifier: notifier struct to unregister
2207 * This is safe to call from within a preemption notifier.
2209 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2211 hlist_del(¬ifier->link);
2213 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2215 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2217 struct preempt_notifier *notifier;
2218 struct hlist_node *node;
2220 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2221 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2225 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2226 struct task_struct *next)
2228 struct preempt_notifier *notifier;
2229 struct hlist_node *node;
2231 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2232 notifier->ops->sched_out(notifier, next);
2237 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2242 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2243 struct task_struct *next)
2250 * prepare_task_switch - prepare to switch tasks
2251 * @rq: the runqueue preparing to switch
2252 * @prev: the current task that is being switched out
2253 * @next: the task we are going to switch to.
2255 * This is called with the rq lock held and interrupts off. It must
2256 * be paired with a subsequent finish_task_switch after the context
2259 * prepare_task_switch sets up locking and calls architecture specific
2263 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2264 struct task_struct *next)
2266 fire_sched_out_preempt_notifiers(prev, next);
2267 prepare_lock_switch(rq, next);
2268 prepare_arch_switch(next);
2272 * finish_task_switch - clean up after a task-switch
2273 * @rq: runqueue associated with task-switch
2274 * @prev: the thread we just switched away from.
2276 * finish_task_switch must be called after the context switch, paired
2277 * with a prepare_task_switch call before the context switch.
2278 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2279 * and do any other architecture-specific cleanup actions.
2281 * Note that we may have delayed dropping an mm in context_switch(). If
2282 * so, we finish that here outside of the runqueue lock. (Doing it
2283 * with the lock held can cause deadlocks; see schedule() for
2286 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2287 __releases(rq->lock)
2289 struct mm_struct *mm = rq->prev_mm;
2295 * A task struct has one reference for the use as "current".
2296 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2297 * schedule one last time. The schedule call will never return, and
2298 * the scheduled task must drop that reference.
2299 * The test for TASK_DEAD must occur while the runqueue locks are
2300 * still held, otherwise prev could be scheduled on another cpu, die
2301 * there before we look at prev->state, and then the reference would
2303 * Manfred Spraul <manfred@colorfullife.com>
2305 prev_state = prev->state;
2306 finish_arch_switch(prev);
2307 finish_lock_switch(rq, prev);
2309 if (current->sched_class->post_schedule)
2310 current->sched_class->post_schedule(rq);
2313 fire_sched_in_preempt_notifiers(current);
2316 if (unlikely(prev_state == TASK_DEAD)) {
2318 * Remove function-return probe instances associated with this
2319 * task and put them back on the free list.
2321 kprobe_flush_task(prev);
2322 put_task_struct(prev);
2327 * schedule_tail - first thing a freshly forked thread must call.
2328 * @prev: the thread we just switched away from.
2330 asmlinkage void schedule_tail(struct task_struct *prev)
2331 __releases(rq->lock)
2333 struct rq *rq = this_rq();
2335 finish_task_switch(rq, prev);
2336 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2337 /* In this case, finish_task_switch does not reenable preemption */
2340 if (current->set_child_tid)
2341 put_user(task_pid_vnr(current), current->set_child_tid);
2345 * context_switch - switch to the new MM and the new
2346 * thread's register state.
2349 context_switch(struct rq *rq, struct task_struct *prev,
2350 struct task_struct *next)
2352 struct mm_struct *mm, *oldmm;
2354 prepare_task_switch(rq, prev, next);
2356 oldmm = prev->active_mm;
2358 * For paravirt, this is coupled with an exit in switch_to to
2359 * combine the page table reload and the switch backend into
2362 arch_enter_lazy_cpu_mode();
2364 if (unlikely(!mm)) {
2365 next->active_mm = oldmm;
2366 atomic_inc(&oldmm->mm_count);
2367 enter_lazy_tlb(oldmm, next);
2369 switch_mm(oldmm, mm, next);
2371 if (unlikely(!prev->mm)) {
2372 prev->active_mm = NULL;
2373 rq->prev_mm = oldmm;
2376 * Since the runqueue lock will be released by the next
2377 * task (which is an invalid locking op but in the case
2378 * of the scheduler it's an obvious special-case), so we
2379 * do an early lockdep release here:
2381 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2382 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2385 /* Here we just switch the register state and the stack. */
2386 switch_to(prev, next, prev);
2390 * this_rq must be evaluated again because prev may have moved
2391 * CPUs since it called schedule(), thus the 'rq' on its stack
2392 * frame will be invalid.
2394 finish_task_switch(this_rq(), prev);
2398 * nr_running, nr_uninterruptible and nr_context_switches:
2400 * externally visible scheduler statistics: current number of runnable
2401 * threads, current number of uninterruptible-sleeping threads, total
2402 * number of context switches performed since bootup.
2404 unsigned long nr_running(void)
2406 unsigned long i, sum = 0;
2408 for_each_online_cpu(i)
2409 sum += cpu_rq(i)->nr_running;
2414 unsigned long nr_uninterruptible(void)
2416 unsigned long i, sum = 0;
2418 for_each_possible_cpu(i)
2419 sum += cpu_rq(i)->nr_uninterruptible;
2422 * Since we read the counters lockless, it might be slightly
2423 * inaccurate. Do not allow it to go below zero though:
2425 if (unlikely((long)sum < 0))
2431 unsigned long long nr_context_switches(void)
2434 unsigned long long sum = 0;
2436 for_each_possible_cpu(i)
2437 sum += cpu_rq(i)->nr_switches;
2442 unsigned long nr_iowait(void)
2444 unsigned long i, sum = 0;
2446 for_each_possible_cpu(i)
2447 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2452 unsigned long nr_active(void)
2454 unsigned long i, running = 0, uninterruptible = 0;
2456 for_each_online_cpu(i) {
2457 running += cpu_rq(i)->nr_running;
2458 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2461 if (unlikely((long)uninterruptible < 0))
2462 uninterruptible = 0;
2464 return running + uninterruptible;
2468 * Update rq->cpu_load[] statistics. This function is usually called every
2469 * scheduler tick (TICK_NSEC).
2471 static void update_cpu_load(struct rq *this_rq)
2473 unsigned long this_load = this_rq->load.weight;
2476 this_rq->nr_load_updates++;
2478 /* Update our load: */
2479 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2480 unsigned long old_load, new_load;
2482 /* scale is effectively 1 << i now, and >> i divides by scale */
2484 old_load = this_rq->cpu_load[i];
2485 new_load = this_load;
2487 * Round up the averaging division if load is increasing. This
2488 * prevents us from getting stuck on 9 if the load is 10, for
2491 if (new_load > old_load)
2492 new_load += scale-1;
2493 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2500 * double_rq_lock - safely lock two runqueues
2502 * Note this does not disable interrupts like task_rq_lock,
2503 * you need to do so manually before calling.
2505 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2506 __acquires(rq1->lock)
2507 __acquires(rq2->lock)
2509 BUG_ON(!irqs_disabled());
2511 spin_lock(&rq1->lock);
2512 __acquire(rq2->lock); /* Fake it out ;) */
2515 spin_lock(&rq1->lock);
2516 spin_lock(&rq2->lock);
2518 spin_lock(&rq2->lock);
2519 spin_lock(&rq1->lock);
2522 update_rq_clock(rq1);
2523 update_rq_clock(rq2);
2527 * double_rq_unlock - safely unlock two runqueues
2529 * Note this does not restore interrupts like task_rq_unlock,
2530 * you need to do so manually after calling.
2532 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2533 __releases(rq1->lock)
2534 __releases(rq2->lock)
2536 spin_unlock(&rq1->lock);
2538 spin_unlock(&rq2->lock);
2540 __release(rq2->lock);
2544 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2546 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2547 __releases(this_rq->lock)
2548 __acquires(busiest->lock)
2549 __acquires(this_rq->lock)
2553 if (unlikely(!irqs_disabled())) {
2554 /* printk() doesn't work good under rq->lock */
2555 spin_unlock(&this_rq->lock);
2558 if (unlikely(!spin_trylock(&busiest->lock))) {
2559 if (busiest < this_rq) {
2560 spin_unlock(&this_rq->lock);
2561 spin_lock(&busiest->lock);
2562 spin_lock(&this_rq->lock);
2565 spin_lock(&busiest->lock);
2571 * If dest_cpu is allowed for this process, migrate the task to it.
2572 * This is accomplished by forcing the cpu_allowed mask to only
2573 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2574 * the cpu_allowed mask is restored.
2576 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2578 struct migration_req req;
2579 unsigned long flags;
2582 rq = task_rq_lock(p, &flags);
2583 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2584 || unlikely(cpu_is_offline(dest_cpu)))
2587 /* force the process onto the specified CPU */
2588 if (migrate_task(p, dest_cpu, &req)) {
2589 /* Need to wait for migration thread (might exit: take ref). */
2590 struct task_struct *mt = rq->migration_thread;
2592 get_task_struct(mt);
2593 task_rq_unlock(rq, &flags);
2594 wake_up_process(mt);
2595 put_task_struct(mt);
2596 wait_for_completion(&req.done);
2601 task_rq_unlock(rq, &flags);
2605 * sched_exec - execve() is a valuable balancing opportunity, because at
2606 * this point the task has the smallest effective memory and cache footprint.
2608 void sched_exec(void)
2610 int new_cpu, this_cpu = get_cpu();
2611 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2613 if (new_cpu != this_cpu)
2614 sched_migrate_task(current, new_cpu);
2618 * pull_task - move a task from a remote runqueue to the local runqueue.
2619 * Both runqueues must be locked.
2621 static void pull_task(struct rq *src_rq, struct task_struct *p,
2622 struct rq *this_rq, int this_cpu)
2624 deactivate_task(src_rq, p, 0);
2625 set_task_cpu(p, this_cpu);
2626 activate_task(this_rq, p, 0);
2628 * Note that idle threads have a prio of MAX_PRIO, for this test
2629 * to be always true for them.
2631 check_preempt_curr(this_rq, p);
2635 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2638 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2639 struct sched_domain *sd, enum cpu_idle_type idle,
2643 * We do not migrate tasks that are:
2644 * 1) running (obviously), or
2645 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2646 * 3) are cache-hot on their current CPU.
2648 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2649 schedstat_inc(p, se.nr_failed_migrations_affine);
2654 if (task_running(rq, p)) {
2655 schedstat_inc(p, se.nr_failed_migrations_running);
2660 * Aggressive migration if:
2661 * 1) task is cache cold, or
2662 * 2) too many balance attempts have failed.
2665 if (!task_hot(p, rq->clock, sd) ||
2666 sd->nr_balance_failed > sd->cache_nice_tries) {
2667 #ifdef CONFIG_SCHEDSTATS
2668 if (task_hot(p, rq->clock, sd)) {
2669 schedstat_inc(sd, lb_hot_gained[idle]);
2670 schedstat_inc(p, se.nr_forced_migrations);
2676 if (task_hot(p, rq->clock, sd)) {
2677 schedstat_inc(p, se.nr_failed_migrations_hot);
2683 static unsigned long
2684 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2685 unsigned long max_load_move, struct sched_domain *sd,
2686 enum cpu_idle_type idle, int *all_pinned,
2687 int *this_best_prio, struct rq_iterator *iterator)
2689 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2690 struct task_struct *p;
2691 long rem_load_move = max_load_move;
2693 if (max_load_move == 0)
2699 * Start the load-balancing iterator:
2701 p = iterator->start(iterator->arg);
2703 if (!p || loops++ > sysctl_sched_nr_migrate)
2706 * To help distribute high priority tasks across CPUs we don't
2707 * skip a task if it will be the highest priority task (i.e. smallest
2708 * prio value) on its new queue regardless of its load weight
2710 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2711 SCHED_LOAD_SCALE_FUZZ;
2712 if ((skip_for_load && p->prio >= *this_best_prio) ||
2713 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2714 p = iterator->next(iterator->arg);
2718 pull_task(busiest, p, this_rq, this_cpu);
2720 rem_load_move -= p->se.load.weight;
2723 * We only want to steal up to the prescribed amount of weighted load.
2725 if (rem_load_move > 0) {
2726 if (p->prio < *this_best_prio)
2727 *this_best_prio = p->prio;
2728 p = iterator->next(iterator->arg);
2733 * Right now, this is one of only two places pull_task() is called,
2734 * so we can safely collect pull_task() stats here rather than
2735 * inside pull_task().
2737 schedstat_add(sd, lb_gained[idle], pulled);
2740 *all_pinned = pinned;
2742 return max_load_move - rem_load_move;
2746 * move_tasks tries to move up to max_load_move weighted load from busiest to
2747 * this_rq, as part of a balancing operation within domain "sd".
2748 * Returns 1 if successful and 0 otherwise.
2750 * Called with both runqueues locked.
2752 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2753 unsigned long max_load_move,
2754 struct sched_domain *sd, enum cpu_idle_type idle,
2757 const struct sched_class *class = sched_class_highest;
2758 unsigned long total_load_moved = 0;
2759 int this_best_prio = this_rq->curr->prio;
2763 class->load_balance(this_rq, this_cpu, busiest,
2764 max_load_move - total_load_moved,
2765 sd, idle, all_pinned, &this_best_prio);
2766 class = class->next;
2767 } while (class && max_load_move > total_load_moved);
2769 return total_load_moved > 0;
2773 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2774 struct sched_domain *sd, enum cpu_idle_type idle,
2775 struct rq_iterator *iterator)
2777 struct task_struct *p = iterator->start(iterator->arg);
2781 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2782 pull_task(busiest, p, this_rq, this_cpu);
2784 * Right now, this is only the second place pull_task()
2785 * is called, so we can safely collect pull_task()
2786 * stats here rather than inside pull_task().
2788 schedstat_inc(sd, lb_gained[idle]);
2792 p = iterator->next(iterator->arg);
2799 * move_one_task tries to move exactly one task from busiest to this_rq, as
2800 * part of active balancing operations within "domain".
2801 * Returns 1 if successful and 0 otherwise.
2803 * Called with both runqueues locked.
2805 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2806 struct sched_domain *sd, enum cpu_idle_type idle)
2808 const struct sched_class *class;
2810 for (class = sched_class_highest; class; class = class->next)
2811 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2818 * find_busiest_group finds and returns the busiest CPU group within the
2819 * domain. It calculates and returns the amount of weighted load which
2820 * should be moved to restore balance via the imbalance parameter.
2822 static struct sched_group *
2823 find_busiest_group(struct sched_domain *sd, int this_cpu,
2824 unsigned long *imbalance, enum cpu_idle_type idle,
2825 int *sd_idle, const cpumask_t *cpus, int *balance)
2827 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2828 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2829 unsigned long max_pull;
2830 unsigned long busiest_load_per_task, busiest_nr_running;
2831 unsigned long this_load_per_task, this_nr_running;
2832 int load_idx, group_imb = 0;
2833 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2834 int power_savings_balance = 1;
2835 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2836 unsigned long min_nr_running = ULONG_MAX;
2837 struct sched_group *group_min = NULL, *group_leader = NULL;
2840 max_load = this_load = total_load = total_pwr = 0;
2841 busiest_load_per_task = busiest_nr_running = 0;
2842 this_load_per_task = this_nr_running = 0;
2843 if (idle == CPU_NOT_IDLE)
2844 load_idx = sd->busy_idx;
2845 else if (idle == CPU_NEWLY_IDLE)
2846 load_idx = sd->newidle_idx;
2848 load_idx = sd->idle_idx;
2851 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2854 int __group_imb = 0;
2855 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2856 unsigned long sum_nr_running, sum_weighted_load;
2858 local_group = cpu_isset(this_cpu, group->cpumask);
2861 balance_cpu = first_cpu(group->cpumask);
2863 /* Tally up the load of all CPUs in the group */
2864 sum_weighted_load = sum_nr_running = avg_load = 0;
2866 min_cpu_load = ~0UL;
2868 for_each_cpu_mask(i, group->cpumask) {
2871 if (!cpu_isset(i, *cpus))
2876 if (*sd_idle && rq->nr_running)
2879 /* Bias balancing toward cpus of our domain */
2881 if (idle_cpu(i) && !first_idle_cpu) {
2886 load = target_load(i, load_idx);
2888 load = source_load(i, load_idx);
2889 if (load > max_cpu_load)
2890 max_cpu_load = load;
2891 if (min_cpu_load > load)
2892 min_cpu_load = load;
2896 sum_nr_running += rq->nr_running;
2897 sum_weighted_load += weighted_cpuload(i);
2901 * First idle cpu or the first cpu(busiest) in this sched group
2902 * is eligible for doing load balancing at this and above
2903 * domains. In the newly idle case, we will allow all the cpu's
2904 * to do the newly idle load balance.
2906 if (idle != CPU_NEWLY_IDLE && local_group &&
2907 balance_cpu != this_cpu && balance) {
2912 total_load += avg_load;
2913 total_pwr += group->__cpu_power;
2915 /* Adjust by relative CPU power of the group */
2916 avg_load = sg_div_cpu_power(group,
2917 avg_load * SCHED_LOAD_SCALE);
2919 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2922 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2925 this_load = avg_load;
2927 this_nr_running = sum_nr_running;
2928 this_load_per_task = sum_weighted_load;
2929 } else if (avg_load > max_load &&
2930 (sum_nr_running > group_capacity || __group_imb)) {
2931 max_load = avg_load;
2933 busiest_nr_running = sum_nr_running;
2934 busiest_load_per_task = sum_weighted_load;
2935 group_imb = __group_imb;
2938 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2940 * Busy processors will not participate in power savings
2943 if (idle == CPU_NOT_IDLE ||
2944 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2948 * If the local group is idle or completely loaded
2949 * no need to do power savings balance at this domain
2951 if (local_group && (this_nr_running >= group_capacity ||
2953 power_savings_balance = 0;
2956 * If a group is already running at full capacity or idle,
2957 * don't include that group in power savings calculations
2959 if (!power_savings_balance || sum_nr_running >= group_capacity
2964 * Calculate the group which has the least non-idle load.
2965 * This is the group from where we need to pick up the load
2968 if ((sum_nr_running < min_nr_running) ||
2969 (sum_nr_running == min_nr_running &&
2970 first_cpu(group->cpumask) <
2971 first_cpu(group_min->cpumask))) {
2973 min_nr_running = sum_nr_running;
2974 min_load_per_task = sum_weighted_load /
2979 * Calculate the group which is almost near its
2980 * capacity but still has some space to pick up some load
2981 * from other group and save more power
2983 if (sum_nr_running <= group_capacity - 1) {
2984 if (sum_nr_running > leader_nr_running ||
2985 (sum_nr_running == leader_nr_running &&
2986 first_cpu(group->cpumask) >
2987 first_cpu(group_leader->cpumask))) {
2988 group_leader = group;
2989 leader_nr_running = sum_nr_running;
2994 group = group->next;
2995 } while (group != sd->groups);
2997 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3000 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3002 if (this_load >= avg_load ||
3003 100*max_load <= sd->imbalance_pct*this_load)
3006 busiest_load_per_task /= busiest_nr_running;
3008 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3011 * We're trying to get all the cpus to the average_load, so we don't
3012 * want to push ourselves above the average load, nor do we wish to
3013 * reduce the max loaded cpu below the average load, as either of these
3014 * actions would just result in more rebalancing later, and ping-pong
3015 * tasks around. Thus we look for the minimum possible imbalance.
3016 * Negative imbalances (*we* are more loaded than anyone else) will
3017 * be counted as no imbalance for these purposes -- we can't fix that
3018 * by pulling tasks to us. Be careful of negative numbers as they'll
3019 * appear as very large values with unsigned longs.
3021 if (max_load <= busiest_load_per_task)
3025 * In the presence of smp nice balancing, certain scenarios can have
3026 * max load less than avg load(as we skip the groups at or below
3027 * its cpu_power, while calculating max_load..)
3029 if (max_load < avg_load) {
3031 goto small_imbalance;
3034 /* Don't want to pull so many tasks that a group would go idle */
3035 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3037 /* How much load to actually move to equalise the imbalance */
3038 *imbalance = min(max_pull * busiest->__cpu_power,
3039 (avg_load - this_load) * this->__cpu_power)
3043 * if *imbalance is less than the average load per runnable task
3044 * there is no gaurantee that any tasks will be moved so we'll have
3045 * a think about bumping its value to force at least one task to be
3048 if (*imbalance < busiest_load_per_task) {
3049 unsigned long tmp, pwr_now, pwr_move;
3053 pwr_move = pwr_now = 0;
3055 if (this_nr_running) {
3056 this_load_per_task /= this_nr_running;
3057 if (busiest_load_per_task > this_load_per_task)
3060 this_load_per_task = SCHED_LOAD_SCALE;
3062 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3063 busiest_load_per_task * imbn) {
3064 *imbalance = busiest_load_per_task;
3069 * OK, we don't have enough imbalance to justify moving tasks,
3070 * however we may be able to increase total CPU power used by
3074 pwr_now += busiest->__cpu_power *
3075 min(busiest_load_per_task, max_load);
3076 pwr_now += this->__cpu_power *
3077 min(this_load_per_task, this_load);
3078 pwr_now /= SCHED_LOAD_SCALE;
3080 /* Amount of load we'd subtract */
3081 tmp = sg_div_cpu_power(busiest,
3082 busiest_load_per_task * SCHED_LOAD_SCALE);
3084 pwr_move += busiest->__cpu_power *
3085 min(busiest_load_per_task, max_load - tmp);
3087 /* Amount of load we'd add */
3088 if (max_load * busiest->__cpu_power <
3089 busiest_load_per_task * SCHED_LOAD_SCALE)
3090 tmp = sg_div_cpu_power(this,
3091 max_load * busiest->__cpu_power);
3093 tmp = sg_div_cpu_power(this,
3094 busiest_load_per_task * SCHED_LOAD_SCALE);
3095 pwr_move += this->__cpu_power *
3096 min(this_load_per_task, this_load + tmp);
3097 pwr_move /= SCHED_LOAD_SCALE;
3099 /* Move if we gain throughput */
3100 if (pwr_move > pwr_now)
3101 *imbalance = busiest_load_per_task;
3107 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3108 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3111 if (this == group_leader && group_leader != group_min) {
3112 *imbalance = min_load_per_task;
3122 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3125 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3126 unsigned long imbalance, const cpumask_t *cpus)
3128 struct rq *busiest = NULL, *rq;
3129 unsigned long max_load = 0;
3132 for_each_cpu_mask(i, group->cpumask) {
3135 if (!cpu_isset(i, *cpus))
3139 wl = weighted_cpuload(i);
3141 if (rq->nr_running == 1 && wl > imbalance)
3144 if (wl > max_load) {
3154 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3155 * so long as it is large enough.
3157 #define MAX_PINNED_INTERVAL 512
3160 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3161 * tasks if there is an imbalance.
3163 static int load_balance(int this_cpu, struct rq *this_rq,
3164 struct sched_domain *sd, enum cpu_idle_type idle,
3165 int *balance, cpumask_t *cpus)
3167 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3168 struct sched_group *group;
3169 unsigned long imbalance;
3171 unsigned long flags;
3176 * When power savings policy is enabled for the parent domain, idle
3177 * sibling can pick up load irrespective of busy siblings. In this case,
3178 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3179 * portraying it as CPU_NOT_IDLE.
3181 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3182 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3185 schedstat_inc(sd, lb_count[idle]);
3188 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3195 schedstat_inc(sd, lb_nobusyg[idle]);
3199 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3201 schedstat_inc(sd, lb_nobusyq[idle]);
3205 BUG_ON(busiest == this_rq);
3207 schedstat_add(sd, lb_imbalance[idle], imbalance);
3210 if (busiest->nr_running > 1) {
3212 * Attempt to move tasks. If find_busiest_group has found
3213 * an imbalance but busiest->nr_running <= 1, the group is
3214 * still unbalanced. ld_moved simply stays zero, so it is
3215 * correctly treated as an imbalance.
3217 local_irq_save(flags);
3218 double_rq_lock(this_rq, busiest);
3219 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3220 imbalance, sd, idle, &all_pinned);
3221 double_rq_unlock(this_rq, busiest);
3222 local_irq_restore(flags);
3225 * some other cpu did the load balance for us.
3227 if (ld_moved && this_cpu != smp_processor_id())
3228 resched_cpu(this_cpu);
3230 /* All tasks on this runqueue were pinned by CPU affinity */
3231 if (unlikely(all_pinned)) {
3232 cpu_clear(cpu_of(busiest), *cpus);
3233 if (!cpus_empty(*cpus))
3240 schedstat_inc(sd, lb_failed[idle]);
3241 sd->nr_balance_failed++;
3243 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3245 spin_lock_irqsave(&busiest->lock, flags);
3247 /* don't kick the migration_thread, if the curr
3248 * task on busiest cpu can't be moved to this_cpu
3250 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3251 spin_unlock_irqrestore(&busiest->lock, flags);
3253 goto out_one_pinned;
3256 if (!busiest->active_balance) {
3257 busiest->active_balance = 1;
3258 busiest->push_cpu = this_cpu;
3261 spin_unlock_irqrestore(&busiest->lock, flags);
3263 wake_up_process(busiest->migration_thread);
3266 * We've kicked active balancing, reset the failure
3269 sd->nr_balance_failed = sd->cache_nice_tries+1;
3272 sd->nr_balance_failed = 0;
3274 if (likely(!active_balance)) {
3275 /* We were unbalanced, so reset the balancing interval */
3276 sd->balance_interval = sd->min_interval;
3279 * If we've begun active balancing, start to back off. This
3280 * case may not be covered by the all_pinned logic if there
3281 * is only 1 task on the busy runqueue (because we don't call
3284 if (sd->balance_interval < sd->max_interval)
3285 sd->balance_interval *= 2;
3288 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3289 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3294 schedstat_inc(sd, lb_balanced[idle]);
3296 sd->nr_balance_failed = 0;
3299 /* tune up the balancing interval */
3300 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3301 (sd->balance_interval < sd->max_interval))
3302 sd->balance_interval *= 2;
3304 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3305 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3311 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3312 * tasks if there is an imbalance.
3314 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3315 * this_rq is locked.
3318 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3321 struct sched_group *group;
3322 struct rq *busiest = NULL;
3323 unsigned long imbalance;
3331 * When power savings policy is enabled for the parent domain, idle
3332 * sibling can pick up load irrespective of busy siblings. In this case,
3333 * let the state of idle sibling percolate up as IDLE, instead of
3334 * portraying it as CPU_NOT_IDLE.
3336 if (sd->flags & SD_SHARE_CPUPOWER &&
3337 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3340 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3342 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3343 &sd_idle, cpus, NULL);
3345 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3349 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3351 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3355 BUG_ON(busiest == this_rq);
3357 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3360 if (busiest->nr_running > 1) {
3361 /* Attempt to move tasks */
3362 double_lock_balance(this_rq, busiest);
3363 /* this_rq->clock is already updated */
3364 update_rq_clock(busiest);
3365 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3366 imbalance, sd, CPU_NEWLY_IDLE,
3368 spin_unlock(&busiest->lock);
3370 if (unlikely(all_pinned)) {
3371 cpu_clear(cpu_of(busiest), *cpus);
3372 if (!cpus_empty(*cpus))
3378 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3379 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3380 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3383 sd->nr_balance_failed = 0;
3388 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3389 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3390 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3392 sd->nr_balance_failed = 0;
3398 * idle_balance is called by schedule() if this_cpu is about to become
3399 * idle. Attempts to pull tasks from other CPUs.
3401 static void idle_balance(int this_cpu, struct rq *this_rq)
3403 struct sched_domain *sd;
3404 int pulled_task = -1;
3405 unsigned long next_balance = jiffies + HZ;
3408 for_each_domain(this_cpu, sd) {
3409 unsigned long interval;
3411 if (!(sd->flags & SD_LOAD_BALANCE))
3414 if (sd->flags & SD_BALANCE_NEWIDLE)
3415 /* If we've pulled tasks over stop searching: */
3416 pulled_task = load_balance_newidle(this_cpu, this_rq,
3419 interval = msecs_to_jiffies(sd->balance_interval);
3420 if (time_after(next_balance, sd->last_balance + interval))
3421 next_balance = sd->last_balance + interval;
3425 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3427 * We are going idle. next_balance may be set based on
3428 * a busy processor. So reset next_balance.
3430 this_rq->next_balance = next_balance;
3435 * active_load_balance is run by migration threads. It pushes running tasks
3436 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3437 * running on each physical CPU where possible, and avoids physical /
3438 * logical imbalances.
3440 * Called with busiest_rq locked.
3442 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3444 int target_cpu = busiest_rq->push_cpu;
3445 struct sched_domain *sd;
3446 struct rq *target_rq;
3448 /* Is there any task to move? */
3449 if (busiest_rq->nr_running <= 1)
3452 target_rq = cpu_rq(target_cpu);
3455 * This condition is "impossible", if it occurs
3456 * we need to fix it. Originally reported by
3457 * Bjorn Helgaas on a 128-cpu setup.
3459 BUG_ON(busiest_rq == target_rq);
3461 /* move a task from busiest_rq to target_rq */
3462 double_lock_balance(busiest_rq, target_rq);
3463 update_rq_clock(busiest_rq);
3464 update_rq_clock(target_rq);
3466 /* Search for an sd spanning us and the target CPU. */
3467 for_each_domain(target_cpu, sd) {
3468 if ((sd->flags & SD_LOAD_BALANCE) &&
3469 cpu_isset(busiest_cpu, sd->span))
3474 schedstat_inc(sd, alb_count);
3476 if (move_one_task(target_rq, target_cpu, busiest_rq,
3478 schedstat_inc(sd, alb_pushed);
3480 schedstat_inc(sd, alb_failed);
3482 spin_unlock(&target_rq->lock);
3487 atomic_t load_balancer;
3489 } nohz ____cacheline_aligned = {
3490 .load_balancer = ATOMIC_INIT(-1),
3491 .cpu_mask = CPU_MASK_NONE,
3495 * This routine will try to nominate the ilb (idle load balancing)
3496 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3497 * load balancing on behalf of all those cpus. If all the cpus in the system
3498 * go into this tickless mode, then there will be no ilb owner (as there is
3499 * no need for one) and all the cpus will sleep till the next wakeup event
3502 * For the ilb owner, tick is not stopped. And this tick will be used
3503 * for idle load balancing. ilb owner will still be part of
3506 * While stopping the tick, this cpu will become the ilb owner if there
3507 * is no other owner. And will be the owner till that cpu becomes busy
3508 * or if all cpus in the system stop their ticks at which point
3509 * there is no need for ilb owner.
3511 * When the ilb owner becomes busy, it nominates another owner, during the
3512 * next busy scheduler_tick()
3514 int select_nohz_load_balancer(int stop_tick)
3516 int cpu = smp_processor_id();
3519 cpu_set(cpu, nohz.cpu_mask);
3520 cpu_rq(cpu)->in_nohz_recently = 1;
3523 * If we are going offline and still the leader, give up!
3525 if (cpu_is_offline(cpu) &&
3526 atomic_read(&nohz.load_balancer) == cpu) {
3527 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3532 /* time for ilb owner also to sleep */
3533 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3534 if (atomic_read(&nohz.load_balancer) == cpu)
3535 atomic_set(&nohz.load_balancer, -1);
3539 if (atomic_read(&nohz.load_balancer) == -1) {
3540 /* make me the ilb owner */
3541 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3543 } else if (atomic_read(&nohz.load_balancer) == cpu)
3546 if (!cpu_isset(cpu, nohz.cpu_mask))
3549 cpu_clear(cpu, nohz.cpu_mask);
3551 if (atomic_read(&nohz.load_balancer) == cpu)
3552 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3559 static DEFINE_SPINLOCK(balancing);
3562 * It checks each scheduling domain to see if it is due to be balanced,
3563 * and initiates a balancing operation if so.
3565 * Balancing parameters are set up in arch_init_sched_domains.
3567 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3570 struct rq *rq = cpu_rq(cpu);
3571 unsigned long interval;
3572 struct sched_domain *sd;
3573 /* Earliest time when we have to do rebalance again */
3574 unsigned long next_balance = jiffies + 60*HZ;
3575 int update_next_balance = 0;
3578 for_each_domain(cpu, sd) {
3579 if (!(sd->flags & SD_LOAD_BALANCE))
3582 interval = sd->balance_interval;
3583 if (idle != CPU_IDLE)
3584 interval *= sd->busy_factor;
3586 /* scale ms to jiffies */
3587 interval = msecs_to_jiffies(interval);
3588 if (unlikely(!interval))
3590 if (interval > HZ*NR_CPUS/10)
3591 interval = HZ*NR_CPUS/10;
3594 if (sd->flags & SD_SERIALIZE) {
3595 if (!spin_trylock(&balancing))
3599 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3600 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3602 * We've pulled tasks over so either we're no
3603 * longer idle, or one of our SMT siblings is
3606 idle = CPU_NOT_IDLE;
3608 sd->last_balance = jiffies;
3610 if (sd->flags & SD_SERIALIZE)
3611 spin_unlock(&balancing);
3613 if (time_after(next_balance, sd->last_balance + interval)) {
3614 next_balance = sd->last_balance + interval;
3615 update_next_balance = 1;
3619 * Stop the load balance at this level. There is another
3620 * CPU in our sched group which is doing load balancing more
3628 * next_balance will be updated only when there is a need.
3629 * When the cpu is attached to null domain for ex, it will not be
3632 if (likely(update_next_balance))
3633 rq->next_balance = next_balance;
3637 * run_rebalance_domains is triggered when needed from the scheduler tick.
3638 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3639 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3641 static void run_rebalance_domains(struct softirq_action *h)
3643 int this_cpu = smp_processor_id();
3644 struct rq *this_rq = cpu_rq(this_cpu);
3645 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3646 CPU_IDLE : CPU_NOT_IDLE;
3648 rebalance_domains(this_cpu, idle);
3652 * If this cpu is the owner for idle load balancing, then do the
3653 * balancing on behalf of the other idle cpus whose ticks are
3656 if (this_rq->idle_at_tick &&
3657 atomic_read(&nohz.load_balancer) == this_cpu) {
3658 cpumask_t cpus = nohz.cpu_mask;
3662 cpu_clear(this_cpu, cpus);
3663 for_each_cpu_mask(balance_cpu, cpus) {
3665 * If this cpu gets work to do, stop the load balancing
3666 * work being done for other cpus. Next load
3667 * balancing owner will pick it up.
3672 rebalance_domains(balance_cpu, CPU_IDLE);
3674 rq = cpu_rq(balance_cpu);
3675 if (time_after(this_rq->next_balance, rq->next_balance))
3676 this_rq->next_balance = rq->next_balance;
3683 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3685 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3686 * idle load balancing owner or decide to stop the periodic load balancing,
3687 * if the whole system is idle.
3689 static inline void trigger_load_balance(struct rq *rq, int cpu)
3693 * If we were in the nohz mode recently and busy at the current
3694 * scheduler tick, then check if we need to nominate new idle
3697 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3698 rq->in_nohz_recently = 0;
3700 if (atomic_read(&nohz.load_balancer) == cpu) {
3701 cpu_clear(cpu, nohz.cpu_mask);
3702 atomic_set(&nohz.load_balancer, -1);
3705 if (atomic_read(&nohz.load_balancer) == -1) {
3707 * simple selection for now: Nominate the
3708 * first cpu in the nohz list to be the next
3711 * TBD: Traverse the sched domains and nominate
3712 * the nearest cpu in the nohz.cpu_mask.
3714 int ilb = first_cpu(nohz.cpu_mask);
3716 if (ilb < nr_cpu_ids)
3722 * If this cpu is idle and doing idle load balancing for all the
3723 * cpus with ticks stopped, is it time for that to stop?
3725 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3726 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3732 * If this cpu is idle and the idle load balancing is done by
3733 * someone else, then no need raise the SCHED_SOFTIRQ
3735 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3736 cpu_isset(cpu, nohz.cpu_mask))
3739 if (time_after_eq(jiffies, rq->next_balance))
3740 raise_softirq(SCHED_SOFTIRQ);
3743 #else /* CONFIG_SMP */
3746 * on UP we do not need to balance between CPUs:
3748 static inline void idle_balance(int cpu, struct rq *rq)
3754 DEFINE_PER_CPU(struct kernel_stat, kstat);
3756 EXPORT_PER_CPU_SYMBOL(kstat);
3759 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3760 * that have not yet been banked in case the task is currently running.
3762 unsigned long long task_sched_runtime(struct task_struct *p)
3764 unsigned long flags;
3768 rq = task_rq_lock(p, &flags);
3769 ns = p->se.sum_exec_runtime;
3770 if (task_current(rq, p)) {
3771 update_rq_clock(rq);
3772 delta_exec = rq->clock - p->se.exec_start;
3773 if ((s64)delta_exec > 0)
3776 task_rq_unlock(rq, &flags);
3782 * Account user cpu time to a process.
3783 * @p: the process that the cpu time gets accounted to
3784 * @cputime: the cpu time spent in user space since the last update
3786 void account_user_time(struct task_struct *p, cputime_t cputime)
3788 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3791 p->utime = cputime_add(p->utime, cputime);
3793 /* Add user time to cpustat. */
3794 tmp = cputime_to_cputime64(cputime);
3795 if (TASK_NICE(p) > 0)
3796 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3798 cpustat->user = cputime64_add(cpustat->user, tmp);
3802 * Account guest cpu time to a process.
3803 * @p: the process that the cpu time gets accounted to
3804 * @cputime: the cpu time spent in virtual machine since the last update
3806 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3809 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3811 tmp = cputime_to_cputime64(cputime);
3813 p->utime = cputime_add(p->utime, cputime);
3814 p->gtime = cputime_add(p->gtime, cputime);
3816 cpustat->user = cputime64_add(cpustat->user, tmp);
3817 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3821 * Account scaled user cpu time to a process.
3822 * @p: the process that the cpu time gets accounted to
3823 * @cputime: the cpu time spent in user space since the last update
3825 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3827 p->utimescaled = cputime_add(p->utimescaled, cputime);
3831 * Account system cpu time to a process.
3832 * @p: the process that the cpu time gets accounted to
3833 * @hardirq_offset: the offset to subtract from hardirq_count()
3834 * @cputime: the cpu time spent in kernel space since the last update
3836 void account_system_time(struct task_struct *p, int hardirq_offset,
3839 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3840 struct rq *rq = this_rq();
3843 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3844 return account_guest_time(p, cputime);
3846 p->stime = cputime_add(p->stime, cputime);
3848 /* Add system time to cpustat. */
3849 tmp = cputime_to_cputime64(cputime);
3850 if (hardirq_count() - hardirq_offset)
3851 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3852 else if (softirq_count())
3853 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3854 else if (p != rq->idle)
3855 cpustat->system = cputime64_add(cpustat->system, tmp);
3856 else if (atomic_read(&rq->nr_iowait) > 0)
3857 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3859 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3860 /* Account for system time used */
3861 acct_update_integrals(p);
3865 * Account scaled system cpu time to a process.
3866 * @p: the process that the cpu time gets accounted to
3867 * @hardirq_offset: the offset to subtract from hardirq_count()
3868 * @cputime: the cpu time spent in kernel space since the last update
3870 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3872 p->stimescaled = cputime_add(p->stimescaled, cputime);
3876 * Account for involuntary wait time.
3877 * @p: the process from which the cpu time has been stolen
3878 * @steal: the cpu time spent in involuntary wait
3880 void account_steal_time(struct task_struct *p, cputime_t steal)
3882 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3883 cputime64_t tmp = cputime_to_cputime64(steal);
3884 struct rq *rq = this_rq();
3886 if (p == rq->idle) {
3887 p->stime = cputime_add(p->stime, steal);
3888 if (atomic_read(&rq->nr_iowait) > 0)
3889 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3891 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3893 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3897 * This function gets called by the timer code, with HZ frequency.
3898 * We call it with interrupts disabled.
3900 * It also gets called by the fork code, when changing the parent's
3903 void scheduler_tick(void)
3905 int cpu = smp_processor_id();
3906 struct rq *rq = cpu_rq(cpu);
3907 struct task_struct *curr = rq->curr;
3908 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3910 spin_lock(&rq->lock);
3911 __update_rq_clock(rq);
3913 * Let rq->clock advance by at least TICK_NSEC:
3915 if (unlikely(rq->clock < next_tick)) {
3916 rq->clock = next_tick;
3917 rq->clock_underflows++;
3919 rq->tick_timestamp = rq->clock;
3920 update_last_tick_seen(rq);
3921 update_cpu_load(rq);
3922 curr->sched_class->task_tick(rq, curr, 0);
3923 spin_unlock(&rq->lock);
3926 rq->idle_at_tick = idle_cpu(cpu);
3927 trigger_load_balance(rq, cpu);
3931 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3933 void __kprobes add_preempt_count(int val)
3938 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3940 preempt_count() += val;
3942 * Spinlock count overflowing soon?
3944 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3947 EXPORT_SYMBOL(add_preempt_count);
3949 void __kprobes sub_preempt_count(int val)
3954 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3957 * Is the spinlock portion underflowing?
3959 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3960 !(preempt_count() & PREEMPT_MASK)))
3963 preempt_count() -= val;
3965 EXPORT_SYMBOL(sub_preempt_count);
3970 * Print scheduling while atomic bug:
3972 static noinline void __schedule_bug(struct task_struct *prev)
3974 struct pt_regs *regs = get_irq_regs();
3976 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3977 prev->comm, prev->pid, preempt_count());
3979 debug_show_held_locks(prev);
3980 if (irqs_disabled())
3981 print_irqtrace_events(prev);
3990 * Various schedule()-time debugging checks and statistics:
3992 static inline void schedule_debug(struct task_struct *prev)
3995 * Test if we are atomic. Since do_exit() needs to call into
3996 * schedule() atomically, we ignore that path for now.
3997 * Otherwise, whine if we are scheduling when we should not be.
3999 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4000 __schedule_bug(prev);
4002 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4004 schedstat_inc(this_rq(), sched_count);
4005 #ifdef CONFIG_SCHEDSTATS
4006 if (unlikely(prev->lock_depth >= 0)) {
4007 schedstat_inc(this_rq(), bkl_count);
4008 schedstat_inc(prev, sched_info.bkl_count);
4014 * Pick up the highest-prio task:
4016 static inline struct task_struct *
4017 pick_next_task(struct rq *rq, struct task_struct *prev)
4019 const struct sched_class *class;
4020 struct task_struct *p;
4023 * Optimization: we know that if all tasks are in
4024 * the fair class we can call that function directly:
4026 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4027 p = fair_sched_class.pick_next_task(rq);
4032 class = sched_class_highest;
4034 p = class->pick_next_task(rq);
4038 * Will never be NULL as the idle class always
4039 * returns a non-NULL p:
4041 class = class->next;
4046 * schedule() is the main scheduler function.
4048 asmlinkage void __sched schedule(void)
4050 struct task_struct *prev, *next;
4051 unsigned long *switch_count;
4057 cpu = smp_processor_id();
4061 switch_count = &prev->nivcsw;
4063 release_kernel_lock(prev);
4064 need_resched_nonpreemptible:
4066 schedule_debug(prev);
4071 * Do the rq-clock update outside the rq lock:
4073 local_irq_disable();
4074 __update_rq_clock(rq);
4075 spin_lock(&rq->lock);
4076 clear_tsk_need_resched(prev);
4078 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4079 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4080 signal_pending(prev))) {
4081 prev->state = TASK_RUNNING;
4083 deactivate_task(rq, prev, 1);
4085 switch_count = &prev->nvcsw;
4089 if (prev->sched_class->pre_schedule)
4090 prev->sched_class->pre_schedule(rq, prev);
4093 if (unlikely(!rq->nr_running))
4094 idle_balance(cpu, rq);
4096 prev->sched_class->put_prev_task(rq, prev);
4097 next = pick_next_task(rq, prev);
4099 sched_info_switch(prev, next);
4101 if (likely(prev != next)) {
4106 context_switch(rq, prev, next); /* unlocks the rq */
4108 * the context switch might have flipped the stack from under
4109 * us, hence refresh the local variables.
4111 cpu = smp_processor_id();
4114 spin_unlock_irq(&rq->lock);
4118 if (unlikely(reacquire_kernel_lock(current) < 0))
4119 goto need_resched_nonpreemptible;
4121 preempt_enable_no_resched();
4122 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4125 EXPORT_SYMBOL(schedule);
4127 #ifdef CONFIG_PREEMPT
4129 * this is the entry point to schedule() from in-kernel preemption
4130 * off of preempt_enable. Kernel preemptions off return from interrupt
4131 * occur there and call schedule directly.
4133 asmlinkage void __sched preempt_schedule(void)
4135 struct thread_info *ti = current_thread_info();
4136 struct task_struct *task = current;
4137 int saved_lock_depth;
4140 * If there is a non-zero preempt_count or interrupts are disabled,
4141 * we do not want to preempt the current task. Just return..
4143 if (likely(ti->preempt_count || irqs_disabled()))
4147 add_preempt_count(PREEMPT_ACTIVE);
4150 * We keep the big kernel semaphore locked, but we
4151 * clear ->lock_depth so that schedule() doesnt
4152 * auto-release the semaphore:
4154 saved_lock_depth = task->lock_depth;
4155 task->lock_depth = -1;
4157 task->lock_depth = saved_lock_depth;
4158 sub_preempt_count(PREEMPT_ACTIVE);
4161 * Check again in case we missed a preemption opportunity
4162 * between schedule and now.
4165 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4167 EXPORT_SYMBOL(preempt_schedule);
4170 * this is the entry point to schedule() from kernel preemption
4171 * off of irq context.
4172 * Note, that this is called and return with irqs disabled. This will
4173 * protect us against recursive calling from irq.
4175 asmlinkage void __sched preempt_schedule_irq(void)
4177 struct thread_info *ti = current_thread_info();
4178 struct task_struct *task = current;
4179 int saved_lock_depth;
4181 /* Catch callers which need to be fixed */
4182 BUG_ON(ti->preempt_count || !irqs_disabled());
4185 add_preempt_count(PREEMPT_ACTIVE);
4188 * We keep the big kernel semaphore locked, but we
4189 * clear ->lock_depth so that schedule() doesnt
4190 * auto-release the semaphore:
4192 saved_lock_depth = task->lock_depth;
4193 task->lock_depth = -1;
4196 local_irq_disable();
4197 task->lock_depth = saved_lock_depth;
4198 sub_preempt_count(PREEMPT_ACTIVE);
4201 * Check again in case we missed a preemption opportunity
4202 * between schedule and now.
4205 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4208 #endif /* CONFIG_PREEMPT */
4210 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4213 return try_to_wake_up(curr->private, mode, sync);
4215 EXPORT_SYMBOL(default_wake_function);
4218 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4219 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4220 * number) then we wake all the non-exclusive tasks and one exclusive task.
4222 * There are circumstances in which we can try to wake a task which has already
4223 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4224 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4226 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4227 int nr_exclusive, int sync, void *key)
4229 wait_queue_t *curr, *next;
4231 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4232 unsigned flags = curr->flags;
4234 if (curr->func(curr, mode, sync, key) &&
4235 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4241 * __wake_up - wake up threads blocked on a waitqueue.
4243 * @mode: which threads
4244 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4245 * @key: is directly passed to the wakeup function
4247 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4248 int nr_exclusive, void *key)
4250 unsigned long flags;
4252 spin_lock_irqsave(&q->lock, flags);
4253 __wake_up_common(q, mode, nr_exclusive, 0, key);
4254 spin_unlock_irqrestore(&q->lock, flags);
4256 EXPORT_SYMBOL(__wake_up);
4259 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4261 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4263 __wake_up_common(q, mode, 1, 0, NULL);
4267 * __wake_up_sync - wake up threads blocked on a waitqueue.
4269 * @mode: which threads
4270 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4272 * The sync wakeup differs that the waker knows that it will schedule
4273 * away soon, so while the target thread will be woken up, it will not
4274 * be migrated to another CPU - ie. the two threads are 'synchronized'
4275 * with each other. This can prevent needless bouncing between CPUs.
4277 * On UP it can prevent extra preemption.
4280 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4282 unsigned long flags;
4288 if (unlikely(!nr_exclusive))
4291 spin_lock_irqsave(&q->lock, flags);
4292 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4293 spin_unlock_irqrestore(&q->lock, flags);
4295 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4297 void complete(struct completion *x)
4299 unsigned long flags;
4301 spin_lock_irqsave(&x->wait.lock, flags);
4303 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4304 spin_unlock_irqrestore(&x->wait.lock, flags);
4306 EXPORT_SYMBOL(complete);
4308 void complete_all(struct completion *x)
4310 unsigned long flags;
4312 spin_lock_irqsave(&x->wait.lock, flags);
4313 x->done += UINT_MAX/2;
4314 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4315 spin_unlock_irqrestore(&x->wait.lock, flags);
4317 EXPORT_SYMBOL(complete_all);
4319 static inline long __sched
4320 do_wait_for_common(struct completion *x, long timeout, int state)
4323 DECLARE_WAITQUEUE(wait, current);
4325 wait.flags |= WQ_FLAG_EXCLUSIVE;
4326 __add_wait_queue_tail(&x->wait, &wait);
4328 if ((state == TASK_INTERRUPTIBLE &&
4329 signal_pending(current)) ||
4330 (state == TASK_KILLABLE &&
4331 fatal_signal_pending(current))) {
4332 __remove_wait_queue(&x->wait, &wait);
4333 return -ERESTARTSYS;
4335 __set_current_state(state);
4336 spin_unlock_irq(&x->wait.lock);
4337 timeout = schedule_timeout(timeout);
4338 spin_lock_irq(&x->wait.lock);
4340 __remove_wait_queue(&x->wait, &wait);
4344 __remove_wait_queue(&x->wait, &wait);
4351 wait_for_common(struct completion *x, long timeout, int state)
4355 spin_lock_irq(&x->wait.lock);
4356 timeout = do_wait_for_common(x, timeout, state);
4357 spin_unlock_irq(&x->wait.lock);
4361 void __sched wait_for_completion(struct completion *x)
4363 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4365 EXPORT_SYMBOL(wait_for_completion);
4367 unsigned long __sched
4368 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4370 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4372 EXPORT_SYMBOL(wait_for_completion_timeout);
4374 int __sched wait_for_completion_interruptible(struct completion *x)
4376 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4377 if (t == -ERESTARTSYS)
4381 EXPORT_SYMBOL(wait_for_completion_interruptible);
4383 unsigned long __sched
4384 wait_for_completion_interruptible_timeout(struct completion *x,
4385 unsigned long timeout)
4387 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4389 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4391 int __sched wait_for_completion_killable(struct completion *x)
4393 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4394 if (t == -ERESTARTSYS)
4398 EXPORT_SYMBOL(wait_for_completion_killable);
4401 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4403 unsigned long flags;
4406 init_waitqueue_entry(&wait, current);
4408 __set_current_state(state);
4410 spin_lock_irqsave(&q->lock, flags);
4411 __add_wait_queue(q, &wait);
4412 spin_unlock(&q->lock);
4413 timeout = schedule_timeout(timeout);
4414 spin_lock_irq(&q->lock);
4415 __remove_wait_queue(q, &wait);
4416 spin_unlock_irqrestore(&q->lock, flags);
4421 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4423 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4425 EXPORT_SYMBOL(interruptible_sleep_on);
4428 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4430 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4432 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4434 void __sched sleep_on(wait_queue_head_t *q)
4436 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4438 EXPORT_SYMBOL(sleep_on);
4440 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4442 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4444 EXPORT_SYMBOL(sleep_on_timeout);
4446 #ifdef CONFIG_RT_MUTEXES
4449 * rt_mutex_setprio - set the current priority of a task
4451 * @prio: prio value (kernel-internal form)
4453 * This function changes the 'effective' priority of a task. It does
4454 * not touch ->normal_prio like __setscheduler().
4456 * Used by the rt_mutex code to implement priority inheritance logic.
4458 void rt_mutex_setprio(struct task_struct *p, int prio)
4460 unsigned long flags;
4461 int oldprio, on_rq, running;
4463 const struct sched_class *prev_class = p->sched_class;
4465 BUG_ON(prio < 0 || prio > MAX_PRIO);
4467 rq = task_rq_lock(p, &flags);
4468 update_rq_clock(rq);
4471 on_rq = p->se.on_rq;
4472 running = task_current(rq, p);
4474 dequeue_task(rq, p, 0);
4476 p->sched_class->put_prev_task(rq, p);
4479 p->sched_class = &rt_sched_class;
4481 p->sched_class = &fair_sched_class;
4486 p->sched_class->set_curr_task(rq);
4488 enqueue_task(rq, p, 0);
4490 check_class_changed(rq, p, prev_class, oldprio, running);
4492 task_rq_unlock(rq, &flags);
4497 void set_user_nice(struct task_struct *p, long nice)
4499 int old_prio, delta, on_rq;
4500 unsigned long flags;
4503 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4506 * We have to be careful, if called from sys_setpriority(),
4507 * the task might be in the middle of scheduling on another CPU.
4509 rq = task_rq_lock(p, &flags);
4510 update_rq_clock(rq);
4512 * The RT priorities are set via sched_setscheduler(), but we still
4513 * allow the 'normal' nice value to be set - but as expected
4514 * it wont have any effect on scheduling until the task is
4515 * SCHED_FIFO/SCHED_RR:
4517 if (task_has_rt_policy(p)) {
4518 p->static_prio = NICE_TO_PRIO(nice);
4521 on_rq = p->se.on_rq;
4523 dequeue_task(rq, p, 0);
4527 p->static_prio = NICE_TO_PRIO(nice);
4530 p->prio = effective_prio(p);
4531 delta = p->prio - old_prio;
4534 enqueue_task(rq, p, 0);
4537 * If the task increased its priority or is running and
4538 * lowered its priority, then reschedule its CPU:
4540 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4541 resched_task(rq->curr);
4544 task_rq_unlock(rq, &flags);
4546 EXPORT_SYMBOL(set_user_nice);
4549 * can_nice - check if a task can reduce its nice value
4553 int can_nice(const struct task_struct *p, const int nice)
4555 /* convert nice value [19,-20] to rlimit style value [1,40] */
4556 int nice_rlim = 20 - nice;
4558 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4559 capable(CAP_SYS_NICE));
4562 #ifdef __ARCH_WANT_SYS_NICE
4565 * sys_nice - change the priority of the current process.
4566 * @increment: priority increment
4568 * sys_setpriority is a more generic, but much slower function that
4569 * does similar things.
4571 asmlinkage long sys_nice(int increment)
4576 * Setpriority might change our priority at the same moment.
4577 * We don't have to worry. Conceptually one call occurs first
4578 * and we have a single winner.
4580 if (increment < -40)
4585 nice = PRIO_TO_NICE(current->static_prio) + increment;
4591 if (increment < 0 && !can_nice(current, nice))
4594 retval = security_task_setnice(current, nice);
4598 set_user_nice(current, nice);
4605 * task_prio - return the priority value of a given task.
4606 * @p: the task in question.
4608 * This is the priority value as seen by users in /proc.
4609 * RT tasks are offset by -200. Normal tasks are centered
4610 * around 0, value goes from -16 to +15.
4612 int task_prio(const struct task_struct *p)
4614 return p->prio - MAX_RT_PRIO;
4618 * task_nice - return the nice value of a given task.
4619 * @p: the task in question.
4621 int task_nice(const struct task_struct *p)
4623 return TASK_NICE(p);
4625 EXPORT_SYMBOL(task_nice);
4628 * idle_cpu - is a given cpu idle currently?
4629 * @cpu: the processor in question.
4631 int idle_cpu(int cpu)
4633 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4637 * idle_task - return the idle task for a given cpu.
4638 * @cpu: the processor in question.
4640 struct task_struct *idle_task(int cpu)
4642 return cpu_rq(cpu)->idle;
4646 * find_process_by_pid - find a process with a matching PID value.
4647 * @pid: the pid in question.
4649 static struct task_struct *find_process_by_pid(pid_t pid)
4651 return pid ? find_task_by_vpid(pid) : current;
4654 /* Actually do priority change: must hold rq lock. */
4656 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4658 BUG_ON(p->se.on_rq);
4661 switch (p->policy) {
4665 p->sched_class = &fair_sched_class;
4669 p->sched_class = &rt_sched_class;
4673 p->rt_priority = prio;
4674 p->normal_prio = normal_prio(p);
4675 /* we are holding p->pi_lock already */
4676 p->prio = rt_mutex_getprio(p);
4681 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4682 * @p: the task in question.
4683 * @policy: new policy.
4684 * @param: structure containing the new RT priority.
4686 * NOTE that the task may be already dead.
4688 int sched_setscheduler(struct task_struct *p, int policy,
4689 struct sched_param *param)
4691 int retval, oldprio, oldpolicy = -1, on_rq, running;
4692 unsigned long flags;
4693 const struct sched_class *prev_class = p->sched_class;
4696 /* may grab non-irq protected spin_locks */
4697 BUG_ON(in_interrupt());
4699 /* double check policy once rq lock held */
4701 policy = oldpolicy = p->policy;
4702 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4703 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4704 policy != SCHED_IDLE)
4707 * Valid priorities for SCHED_FIFO and SCHED_RR are
4708 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4709 * SCHED_BATCH and SCHED_IDLE is 0.
4711 if (param->sched_priority < 0 ||
4712 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4713 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4715 if (rt_policy(policy) != (param->sched_priority != 0))
4719 * Allow unprivileged RT tasks to decrease priority:
4721 if (!capable(CAP_SYS_NICE)) {
4722 if (rt_policy(policy)) {
4723 unsigned long rlim_rtprio;
4725 if (!lock_task_sighand(p, &flags))
4727 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4728 unlock_task_sighand(p, &flags);
4730 /* can't set/change the rt policy */
4731 if (policy != p->policy && !rlim_rtprio)
4734 /* can't increase priority */
4735 if (param->sched_priority > p->rt_priority &&
4736 param->sched_priority > rlim_rtprio)
4740 * Like positive nice levels, dont allow tasks to
4741 * move out of SCHED_IDLE either:
4743 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4746 /* can't change other user's priorities */
4747 if ((current->euid != p->euid) &&
4748 (current->euid != p->uid))
4752 #ifdef CONFIG_RT_GROUP_SCHED
4754 * Do not allow realtime tasks into groups that have no runtime
4757 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4761 retval = security_task_setscheduler(p, policy, param);
4765 * make sure no PI-waiters arrive (or leave) while we are
4766 * changing the priority of the task:
4768 spin_lock_irqsave(&p->pi_lock, flags);
4770 * To be able to change p->policy safely, the apropriate
4771 * runqueue lock must be held.
4773 rq = __task_rq_lock(p);
4774 /* recheck policy now with rq lock held */
4775 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4776 policy = oldpolicy = -1;
4777 __task_rq_unlock(rq);
4778 spin_unlock_irqrestore(&p->pi_lock, flags);
4781 update_rq_clock(rq);
4782 on_rq = p->se.on_rq;
4783 running = task_current(rq, p);
4785 deactivate_task(rq, p, 0);
4787 p->sched_class->put_prev_task(rq, p);
4790 __setscheduler(rq, p, policy, param->sched_priority);
4793 p->sched_class->set_curr_task(rq);
4795 activate_task(rq, p, 0);
4797 check_class_changed(rq, p, prev_class, oldprio, running);
4799 __task_rq_unlock(rq);
4800 spin_unlock_irqrestore(&p->pi_lock, flags);
4802 rt_mutex_adjust_pi(p);
4806 EXPORT_SYMBOL_GPL(sched_setscheduler);
4809 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4811 struct sched_param lparam;
4812 struct task_struct *p;
4815 if (!param || pid < 0)
4817 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4822 p = find_process_by_pid(pid);
4824 retval = sched_setscheduler(p, policy, &lparam);
4831 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4832 * @pid: the pid in question.
4833 * @policy: new policy.
4834 * @param: structure containing the new RT priority.
4837 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4839 /* negative values for policy are not valid */
4843 return do_sched_setscheduler(pid, policy, param);
4847 * sys_sched_setparam - set/change the RT priority of a thread
4848 * @pid: the pid in question.
4849 * @param: structure containing the new RT priority.
4851 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4853 return do_sched_setscheduler(pid, -1, param);
4857 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4858 * @pid: the pid in question.
4860 asmlinkage long sys_sched_getscheduler(pid_t pid)
4862 struct task_struct *p;
4869 read_lock(&tasklist_lock);
4870 p = find_process_by_pid(pid);
4872 retval = security_task_getscheduler(p);
4876 read_unlock(&tasklist_lock);
4881 * sys_sched_getscheduler - get the RT priority of a thread
4882 * @pid: the pid in question.
4883 * @param: structure containing the RT priority.
4885 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4887 struct sched_param lp;
4888 struct task_struct *p;
4891 if (!param || pid < 0)
4894 read_lock(&tasklist_lock);
4895 p = find_process_by_pid(pid);
4900 retval = security_task_getscheduler(p);
4904 lp.sched_priority = p->rt_priority;
4905 read_unlock(&tasklist_lock);
4908 * This one might sleep, we cannot do it with a spinlock held ...
4910 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4915 read_unlock(&tasklist_lock);
4919 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4921 cpumask_t cpus_allowed;
4922 cpumask_t new_mask = *in_mask;
4923 struct task_struct *p;
4927 read_lock(&tasklist_lock);
4929 p = find_process_by_pid(pid);
4931 read_unlock(&tasklist_lock);
4937 * It is not safe to call set_cpus_allowed with the
4938 * tasklist_lock held. We will bump the task_struct's
4939 * usage count and then drop tasklist_lock.
4942 read_unlock(&tasklist_lock);
4945 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4946 !capable(CAP_SYS_NICE))
4949 retval = security_task_setscheduler(p, 0, NULL);
4953 cpuset_cpus_allowed(p, &cpus_allowed);
4954 cpus_and(new_mask, new_mask, cpus_allowed);
4956 retval = set_cpus_allowed_ptr(p, &new_mask);
4959 cpuset_cpus_allowed(p, &cpus_allowed);
4960 if (!cpus_subset(new_mask, cpus_allowed)) {
4962 * We must have raced with a concurrent cpuset
4963 * update. Just reset the cpus_allowed to the
4964 * cpuset's cpus_allowed
4966 new_mask = cpus_allowed;
4976 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4977 cpumask_t *new_mask)
4979 if (len < sizeof(cpumask_t)) {
4980 memset(new_mask, 0, sizeof(cpumask_t));
4981 } else if (len > sizeof(cpumask_t)) {
4982 len = sizeof(cpumask_t);
4984 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4988 * sys_sched_setaffinity - set the cpu affinity of a process
4989 * @pid: pid of the process
4990 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4991 * @user_mask_ptr: user-space pointer to the new cpu mask
4993 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4994 unsigned long __user *user_mask_ptr)
4999 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5003 return sched_setaffinity(pid, &new_mask);
5007 * Represents all cpu's present in the system
5008 * In systems capable of hotplug, this map could dynamically grow
5009 * as new cpu's are detected in the system via any platform specific
5010 * method, such as ACPI for e.g.
5013 cpumask_t cpu_present_map __read_mostly;
5014 EXPORT_SYMBOL(cpu_present_map);
5017 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5018 EXPORT_SYMBOL(cpu_online_map);
5020 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5021 EXPORT_SYMBOL(cpu_possible_map);
5024 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5026 struct task_struct *p;
5030 read_lock(&tasklist_lock);
5033 p = find_process_by_pid(pid);
5037 retval = security_task_getscheduler(p);
5041 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5044 read_unlock(&tasklist_lock);
5051 * sys_sched_getaffinity - get the cpu affinity of a process
5052 * @pid: pid of the process
5053 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5054 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5056 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5057 unsigned long __user *user_mask_ptr)
5062 if (len < sizeof(cpumask_t))
5065 ret = sched_getaffinity(pid, &mask);
5069 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5072 return sizeof(cpumask_t);
5076 * sys_sched_yield - yield the current processor to other threads.
5078 * This function yields the current CPU to other tasks. If there are no
5079 * other threads running on this CPU then this function will return.
5081 asmlinkage long sys_sched_yield(void)
5083 struct rq *rq = this_rq_lock();
5085 schedstat_inc(rq, yld_count);
5086 current->sched_class->yield_task(rq);
5089 * Since we are going to call schedule() anyway, there's
5090 * no need to preempt or enable interrupts:
5092 __release(rq->lock);
5093 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5094 _raw_spin_unlock(&rq->lock);
5095 preempt_enable_no_resched();
5102 static void __cond_resched(void)
5104 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5105 __might_sleep(__FILE__, __LINE__);
5108 * The BKS might be reacquired before we have dropped
5109 * PREEMPT_ACTIVE, which could trigger a second
5110 * cond_resched() call.
5113 add_preempt_count(PREEMPT_ACTIVE);
5115 sub_preempt_count(PREEMPT_ACTIVE);
5116 } while (need_resched());
5119 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5120 int __sched _cond_resched(void)
5122 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5123 system_state == SYSTEM_RUNNING) {
5129 EXPORT_SYMBOL(_cond_resched);
5133 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5134 * call schedule, and on return reacquire the lock.
5136 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5137 * operations here to prevent schedule() from being called twice (once via
5138 * spin_unlock(), once by hand).
5140 int cond_resched_lock(spinlock_t *lock)
5142 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5145 if (spin_needbreak(lock) || resched) {
5147 if (resched && need_resched())
5156 EXPORT_SYMBOL(cond_resched_lock);
5158 int __sched cond_resched_softirq(void)
5160 BUG_ON(!in_softirq());
5162 if (need_resched() && system_state == SYSTEM_RUNNING) {
5170 EXPORT_SYMBOL(cond_resched_softirq);
5173 * yield - yield the current processor to other threads.
5175 * This is a shortcut for kernel-space yielding - it marks the
5176 * thread runnable and calls sys_sched_yield().
5178 void __sched yield(void)
5180 set_current_state(TASK_RUNNING);
5183 EXPORT_SYMBOL(yield);
5186 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5187 * that process accounting knows that this is a task in IO wait state.
5189 * But don't do that if it is a deliberate, throttling IO wait (this task
5190 * has set its backing_dev_info: the queue against which it should throttle)
5192 void __sched io_schedule(void)
5194 struct rq *rq = &__raw_get_cpu_var(runqueues);
5196 delayacct_blkio_start();
5197 atomic_inc(&rq->nr_iowait);
5199 atomic_dec(&rq->nr_iowait);
5200 delayacct_blkio_end();
5202 EXPORT_SYMBOL(io_schedule);
5204 long __sched io_schedule_timeout(long timeout)
5206 struct rq *rq = &__raw_get_cpu_var(runqueues);
5209 delayacct_blkio_start();
5210 atomic_inc(&rq->nr_iowait);
5211 ret = schedule_timeout(timeout);
5212 atomic_dec(&rq->nr_iowait);
5213 delayacct_blkio_end();
5218 * sys_sched_get_priority_max - return maximum RT priority.
5219 * @policy: scheduling class.
5221 * this syscall returns the maximum rt_priority that can be used
5222 * by a given scheduling class.
5224 asmlinkage long sys_sched_get_priority_max(int policy)
5231 ret = MAX_USER_RT_PRIO-1;
5243 * sys_sched_get_priority_min - return minimum RT priority.
5244 * @policy: scheduling class.
5246 * this syscall returns the minimum rt_priority that can be used
5247 * by a given scheduling class.
5249 asmlinkage long sys_sched_get_priority_min(int policy)
5267 * sys_sched_rr_get_interval - return the default timeslice of a process.
5268 * @pid: pid of the process.
5269 * @interval: userspace pointer to the timeslice value.
5271 * this syscall writes the default timeslice value of a given process
5272 * into the user-space timespec buffer. A value of '0' means infinity.
5275 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5277 struct task_struct *p;
5278 unsigned int time_slice;
5286 read_lock(&tasklist_lock);
5287 p = find_process_by_pid(pid);
5291 retval = security_task_getscheduler(p);
5296 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5297 * tasks that are on an otherwise idle runqueue:
5300 if (p->policy == SCHED_RR) {
5301 time_slice = DEF_TIMESLICE;
5302 } else if (p->policy != SCHED_FIFO) {
5303 struct sched_entity *se = &p->se;
5304 unsigned long flags;
5307 rq = task_rq_lock(p, &flags);
5308 if (rq->cfs.load.weight)
5309 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5310 task_rq_unlock(rq, &flags);
5312 read_unlock(&tasklist_lock);
5313 jiffies_to_timespec(time_slice, &t);
5314 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5318 read_unlock(&tasklist_lock);
5322 static const char stat_nam[] = "RSDTtZX";
5324 void sched_show_task(struct task_struct *p)
5326 unsigned long free = 0;
5329 state = p->state ? __ffs(p->state) + 1 : 0;
5330 printk(KERN_INFO "%-13.13s %c", p->comm,
5331 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5332 #if BITS_PER_LONG == 32
5333 if (state == TASK_RUNNING)
5334 printk(KERN_CONT " running ");
5336 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5338 if (state == TASK_RUNNING)
5339 printk(KERN_CONT " running task ");
5341 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5343 #ifdef CONFIG_DEBUG_STACK_USAGE
5345 unsigned long *n = end_of_stack(p);
5348 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5351 printk(KERN_CONT "%5lu %5d %6d\n", free,
5352 task_pid_nr(p), task_pid_nr(p->real_parent));
5354 show_stack(p, NULL);
5357 void show_state_filter(unsigned long state_filter)
5359 struct task_struct *g, *p;
5361 #if BITS_PER_LONG == 32
5363 " task PC stack pid father\n");
5366 " task PC stack pid father\n");
5368 read_lock(&tasklist_lock);
5369 do_each_thread(g, p) {
5371 * reset the NMI-timeout, listing all files on a slow
5372 * console might take alot of time:
5374 touch_nmi_watchdog();
5375 if (!state_filter || (p->state & state_filter))
5377 } while_each_thread(g, p);
5379 touch_all_softlockup_watchdogs();
5381 #ifdef CONFIG_SCHED_DEBUG
5382 sysrq_sched_debug_show();
5384 read_unlock(&tasklist_lock);
5386 * Only show locks if all tasks are dumped:
5388 if (state_filter == -1)
5389 debug_show_all_locks();
5392 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5394 idle->sched_class = &idle_sched_class;
5398 * init_idle - set up an idle thread for a given CPU
5399 * @idle: task in question
5400 * @cpu: cpu the idle task belongs to
5402 * NOTE: this function does not set the idle thread's NEED_RESCHED
5403 * flag, to make booting more robust.
5405 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5407 struct rq *rq = cpu_rq(cpu);
5408 unsigned long flags;
5411 idle->se.exec_start = sched_clock();
5413 idle->prio = idle->normal_prio = MAX_PRIO;
5414 idle->cpus_allowed = cpumask_of_cpu(cpu);
5415 __set_task_cpu(idle, cpu);
5417 spin_lock_irqsave(&rq->lock, flags);
5418 rq->curr = rq->idle = idle;
5419 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5422 spin_unlock_irqrestore(&rq->lock, flags);
5424 /* Set the preempt count _outside_ the spinlocks! */
5425 task_thread_info(idle)->preempt_count = 0;
5428 * The idle tasks have their own, simple scheduling class:
5430 idle->sched_class = &idle_sched_class;
5434 * In a system that switches off the HZ timer nohz_cpu_mask
5435 * indicates which cpus entered this state. This is used
5436 * in the rcu update to wait only for active cpus. For system
5437 * which do not switch off the HZ timer nohz_cpu_mask should
5438 * always be CPU_MASK_NONE.
5440 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5443 * Increase the granularity value when there are more CPUs,
5444 * because with more CPUs the 'effective latency' as visible
5445 * to users decreases. But the relationship is not linear,
5446 * so pick a second-best guess by going with the log2 of the
5449 * This idea comes from the SD scheduler of Con Kolivas:
5451 static inline void sched_init_granularity(void)
5453 unsigned int factor = 1 + ilog2(num_online_cpus());
5454 const unsigned long limit = 200000000;
5456 sysctl_sched_min_granularity *= factor;
5457 if (sysctl_sched_min_granularity > limit)
5458 sysctl_sched_min_granularity = limit;
5460 sysctl_sched_latency *= factor;
5461 if (sysctl_sched_latency > limit)
5462 sysctl_sched_latency = limit;
5464 sysctl_sched_wakeup_granularity *= factor;
5469 * This is how migration works:
5471 * 1) we queue a struct migration_req structure in the source CPU's
5472 * runqueue and wake up that CPU's migration thread.
5473 * 2) we down() the locked semaphore => thread blocks.
5474 * 3) migration thread wakes up (implicitly it forces the migrated
5475 * thread off the CPU)
5476 * 4) it gets the migration request and checks whether the migrated
5477 * task is still in the wrong runqueue.
5478 * 5) if it's in the wrong runqueue then the migration thread removes
5479 * it and puts it into the right queue.
5480 * 6) migration thread up()s the semaphore.
5481 * 7) we wake up and the migration is done.
5485 * Change a given task's CPU affinity. Migrate the thread to a
5486 * proper CPU and schedule it away if the CPU it's executing on
5487 * is removed from the allowed bitmask.
5489 * NOTE: the caller must have a valid reference to the task, the
5490 * task must not exit() & deallocate itself prematurely. The
5491 * call is not atomic; no spinlocks may be held.
5493 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5495 struct migration_req req;
5496 unsigned long flags;
5500 rq = task_rq_lock(p, &flags);
5501 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5506 if (p->sched_class->set_cpus_allowed)
5507 p->sched_class->set_cpus_allowed(p, new_mask);
5509 p->cpus_allowed = *new_mask;
5510 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5513 /* Can the task run on the task's current CPU? If so, we're done */
5514 if (cpu_isset(task_cpu(p), *new_mask))
5517 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5518 /* Need help from migration thread: drop lock and wait. */
5519 task_rq_unlock(rq, &flags);
5520 wake_up_process(rq->migration_thread);
5521 wait_for_completion(&req.done);
5522 tlb_migrate_finish(p->mm);
5526 task_rq_unlock(rq, &flags);
5530 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5533 * Move (not current) task off this cpu, onto dest cpu. We're doing
5534 * this because either it can't run here any more (set_cpus_allowed()
5535 * away from this CPU, or CPU going down), or because we're
5536 * attempting to rebalance this task on exec (sched_exec).
5538 * So we race with normal scheduler movements, but that's OK, as long
5539 * as the task is no longer on this CPU.
5541 * Returns non-zero if task was successfully migrated.
5543 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5545 struct rq *rq_dest, *rq_src;
5548 if (unlikely(cpu_is_offline(dest_cpu)))
5551 rq_src = cpu_rq(src_cpu);
5552 rq_dest = cpu_rq(dest_cpu);
5554 double_rq_lock(rq_src, rq_dest);
5555 /* Already moved. */
5556 if (task_cpu(p) != src_cpu)
5558 /* Affinity changed (again). */
5559 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5562 on_rq = p->se.on_rq;
5564 deactivate_task(rq_src, p, 0);
5566 set_task_cpu(p, dest_cpu);
5568 activate_task(rq_dest, p, 0);
5569 check_preempt_curr(rq_dest, p);
5573 double_rq_unlock(rq_src, rq_dest);
5578 * migration_thread - this is a highprio system thread that performs
5579 * thread migration by bumping thread off CPU then 'pushing' onto
5582 static int migration_thread(void *data)
5584 int cpu = (long)data;
5588 BUG_ON(rq->migration_thread != current);
5590 set_current_state(TASK_INTERRUPTIBLE);
5591 while (!kthread_should_stop()) {
5592 struct migration_req *req;
5593 struct list_head *head;
5595 spin_lock_irq(&rq->lock);
5597 if (cpu_is_offline(cpu)) {
5598 spin_unlock_irq(&rq->lock);
5602 if (rq->active_balance) {
5603 active_load_balance(rq, cpu);
5604 rq->active_balance = 0;
5607 head = &rq->migration_queue;
5609 if (list_empty(head)) {
5610 spin_unlock_irq(&rq->lock);
5612 set_current_state(TASK_INTERRUPTIBLE);
5615 req = list_entry(head->next, struct migration_req, list);
5616 list_del_init(head->next);
5618 spin_unlock(&rq->lock);
5619 __migrate_task(req->task, cpu, req->dest_cpu);
5622 complete(&req->done);
5624 __set_current_state(TASK_RUNNING);
5628 /* Wait for kthread_stop */
5629 set_current_state(TASK_INTERRUPTIBLE);
5630 while (!kthread_should_stop()) {
5632 set_current_state(TASK_INTERRUPTIBLE);
5634 __set_current_state(TASK_RUNNING);
5638 #ifdef CONFIG_HOTPLUG_CPU
5640 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5644 local_irq_disable();
5645 ret = __migrate_task(p, src_cpu, dest_cpu);
5651 * Figure out where task on dead CPU should go, use force if necessary.
5652 * NOTE: interrupts should be disabled by the caller
5654 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5656 unsigned long flags;
5663 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5664 cpus_and(mask, mask, p->cpus_allowed);
5665 dest_cpu = any_online_cpu(mask);
5667 /* On any allowed CPU? */
5668 if (dest_cpu >= nr_cpu_ids)
5669 dest_cpu = any_online_cpu(p->cpus_allowed);
5671 /* No more Mr. Nice Guy. */
5672 if (dest_cpu >= nr_cpu_ids) {
5673 cpumask_t cpus_allowed;
5675 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5677 * Try to stay on the same cpuset, where the
5678 * current cpuset may be a subset of all cpus.
5679 * The cpuset_cpus_allowed_locked() variant of
5680 * cpuset_cpus_allowed() will not block. It must be
5681 * called within calls to cpuset_lock/cpuset_unlock.
5683 rq = task_rq_lock(p, &flags);
5684 p->cpus_allowed = cpus_allowed;
5685 dest_cpu = any_online_cpu(p->cpus_allowed);
5686 task_rq_unlock(rq, &flags);
5689 * Don't tell them about moving exiting tasks or
5690 * kernel threads (both mm NULL), since they never
5693 if (p->mm && printk_ratelimit()) {
5694 printk(KERN_INFO "process %d (%s) no "
5695 "longer affine to cpu%d\n",
5696 task_pid_nr(p), p->comm, dead_cpu);
5699 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5703 * While a dead CPU has no uninterruptible tasks queued at this point,
5704 * it might still have a nonzero ->nr_uninterruptible counter, because
5705 * for performance reasons the counter is not stricly tracking tasks to
5706 * their home CPUs. So we just add the counter to another CPU's counter,
5707 * to keep the global sum constant after CPU-down:
5709 static void migrate_nr_uninterruptible(struct rq *rq_src)
5711 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5712 unsigned long flags;
5714 local_irq_save(flags);
5715 double_rq_lock(rq_src, rq_dest);
5716 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5717 rq_src->nr_uninterruptible = 0;
5718 double_rq_unlock(rq_src, rq_dest);
5719 local_irq_restore(flags);
5722 /* Run through task list and migrate tasks from the dead cpu. */
5723 static void migrate_live_tasks(int src_cpu)
5725 struct task_struct *p, *t;
5727 read_lock(&tasklist_lock);
5729 do_each_thread(t, p) {
5733 if (task_cpu(p) == src_cpu)
5734 move_task_off_dead_cpu(src_cpu, p);
5735 } while_each_thread(t, p);
5737 read_unlock(&tasklist_lock);
5741 * Schedules idle task to be the next runnable task on current CPU.
5742 * It does so by boosting its priority to highest possible.
5743 * Used by CPU offline code.
5745 void sched_idle_next(void)
5747 int this_cpu = smp_processor_id();
5748 struct rq *rq = cpu_rq(this_cpu);
5749 struct task_struct *p = rq->idle;
5750 unsigned long flags;
5752 /* cpu has to be offline */
5753 BUG_ON(cpu_online(this_cpu));
5756 * Strictly not necessary since rest of the CPUs are stopped by now
5757 * and interrupts disabled on the current cpu.
5759 spin_lock_irqsave(&rq->lock, flags);
5761 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5763 update_rq_clock(rq);
5764 activate_task(rq, p, 0);
5766 spin_unlock_irqrestore(&rq->lock, flags);
5770 * Ensures that the idle task is using init_mm right before its cpu goes
5773 void idle_task_exit(void)
5775 struct mm_struct *mm = current->active_mm;
5777 BUG_ON(cpu_online(smp_processor_id()));
5780 switch_mm(mm, &init_mm, current);
5784 /* called under rq->lock with disabled interrupts */
5785 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5787 struct rq *rq = cpu_rq(dead_cpu);
5789 /* Must be exiting, otherwise would be on tasklist. */
5790 BUG_ON(!p->exit_state);
5792 /* Cannot have done final schedule yet: would have vanished. */
5793 BUG_ON(p->state == TASK_DEAD);
5798 * Drop lock around migration; if someone else moves it,
5799 * that's OK. No task can be added to this CPU, so iteration is
5802 spin_unlock_irq(&rq->lock);
5803 move_task_off_dead_cpu(dead_cpu, p);
5804 spin_lock_irq(&rq->lock);
5809 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5810 static void migrate_dead_tasks(unsigned int dead_cpu)
5812 struct rq *rq = cpu_rq(dead_cpu);
5813 struct task_struct *next;
5816 if (!rq->nr_running)
5818 update_rq_clock(rq);
5819 next = pick_next_task(rq, rq->curr);
5822 migrate_dead(dead_cpu, next);
5826 #endif /* CONFIG_HOTPLUG_CPU */
5828 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5830 static struct ctl_table sd_ctl_dir[] = {
5832 .procname = "sched_domain",
5838 static struct ctl_table sd_ctl_root[] = {
5840 .ctl_name = CTL_KERN,
5841 .procname = "kernel",
5843 .child = sd_ctl_dir,
5848 static struct ctl_table *sd_alloc_ctl_entry(int n)
5850 struct ctl_table *entry =
5851 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5856 static void sd_free_ctl_entry(struct ctl_table **tablep)
5858 struct ctl_table *entry;
5861 * In the intermediate directories, both the child directory and
5862 * procname are dynamically allocated and could fail but the mode
5863 * will always be set. In the lowest directory the names are
5864 * static strings and all have proc handlers.
5866 for (entry = *tablep; entry->mode; entry++) {
5868 sd_free_ctl_entry(&entry->child);
5869 if (entry->proc_handler == NULL)
5870 kfree(entry->procname);
5878 set_table_entry(struct ctl_table *entry,
5879 const char *procname, void *data, int maxlen,
5880 mode_t mode, proc_handler *proc_handler)
5882 entry->procname = procname;
5884 entry->maxlen = maxlen;
5886 entry->proc_handler = proc_handler;
5889 static struct ctl_table *
5890 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5892 struct ctl_table *table = sd_alloc_ctl_entry(12);
5897 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5898 sizeof(long), 0644, proc_doulongvec_minmax);
5899 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5900 sizeof(long), 0644, proc_doulongvec_minmax);
5901 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5902 sizeof(int), 0644, proc_dointvec_minmax);
5903 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5904 sizeof(int), 0644, proc_dointvec_minmax);
5905 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5906 sizeof(int), 0644, proc_dointvec_minmax);
5907 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5908 sizeof(int), 0644, proc_dointvec_minmax);
5909 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5910 sizeof(int), 0644, proc_dointvec_minmax);
5911 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5912 sizeof(int), 0644, proc_dointvec_minmax);
5913 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5914 sizeof(int), 0644, proc_dointvec_minmax);
5915 set_table_entry(&table[9], "cache_nice_tries",
5916 &sd->cache_nice_tries,
5917 sizeof(int), 0644, proc_dointvec_minmax);
5918 set_table_entry(&table[10], "flags", &sd->flags,
5919 sizeof(int), 0644, proc_dointvec_minmax);
5920 /* &table[11] is terminator */
5925 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5927 struct ctl_table *entry, *table;
5928 struct sched_domain *sd;
5929 int domain_num = 0, i;
5932 for_each_domain(cpu, sd)
5934 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5939 for_each_domain(cpu, sd) {
5940 snprintf(buf, 32, "domain%d", i);
5941 entry->procname = kstrdup(buf, GFP_KERNEL);
5943 entry->child = sd_alloc_ctl_domain_table(sd);
5950 static struct ctl_table_header *sd_sysctl_header;
5951 static void register_sched_domain_sysctl(void)
5953 int i, cpu_num = num_online_cpus();
5954 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5957 WARN_ON(sd_ctl_dir[0].child);
5958 sd_ctl_dir[0].child = entry;
5963 for_each_online_cpu(i) {
5964 snprintf(buf, 32, "cpu%d", i);
5965 entry->procname = kstrdup(buf, GFP_KERNEL);
5967 entry->child = sd_alloc_ctl_cpu_table(i);
5971 WARN_ON(sd_sysctl_header);
5972 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5975 /* may be called multiple times per register */
5976 static void unregister_sched_domain_sysctl(void)
5978 if (sd_sysctl_header)
5979 unregister_sysctl_table(sd_sysctl_header);
5980 sd_sysctl_header = NULL;
5981 if (sd_ctl_dir[0].child)
5982 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5985 static void register_sched_domain_sysctl(void)
5988 static void unregister_sched_domain_sysctl(void)
5994 * migration_call - callback that gets triggered when a CPU is added.
5995 * Here we can start up the necessary migration thread for the new CPU.
5997 static int __cpuinit
5998 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6000 struct task_struct *p;
6001 int cpu = (long)hcpu;
6002 unsigned long flags;
6007 case CPU_UP_PREPARE:
6008 case CPU_UP_PREPARE_FROZEN:
6009 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6012 kthread_bind(p, cpu);
6013 /* Must be high prio: stop_machine expects to yield to it. */
6014 rq = task_rq_lock(p, &flags);
6015 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6016 task_rq_unlock(rq, &flags);
6017 cpu_rq(cpu)->migration_thread = p;
6021 case CPU_ONLINE_FROZEN:
6022 /* Strictly unnecessary, as first user will wake it. */
6023 wake_up_process(cpu_rq(cpu)->migration_thread);
6025 /* Update our root-domain */
6027 spin_lock_irqsave(&rq->lock, flags);
6029 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6030 cpu_set(cpu, rq->rd->online);
6032 spin_unlock_irqrestore(&rq->lock, flags);
6035 #ifdef CONFIG_HOTPLUG_CPU
6036 case CPU_UP_CANCELED:
6037 case CPU_UP_CANCELED_FROZEN:
6038 if (!cpu_rq(cpu)->migration_thread)
6040 /* Unbind it from offline cpu so it can run. Fall thru. */
6041 kthread_bind(cpu_rq(cpu)->migration_thread,
6042 any_online_cpu(cpu_online_map));
6043 kthread_stop(cpu_rq(cpu)->migration_thread);
6044 cpu_rq(cpu)->migration_thread = NULL;
6048 case CPU_DEAD_FROZEN:
6049 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6050 migrate_live_tasks(cpu);
6052 kthread_stop(rq->migration_thread);
6053 rq->migration_thread = NULL;
6054 /* Idle task back to normal (off runqueue, low prio) */
6055 spin_lock_irq(&rq->lock);
6056 update_rq_clock(rq);
6057 deactivate_task(rq, rq->idle, 0);
6058 rq->idle->static_prio = MAX_PRIO;
6059 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6060 rq->idle->sched_class = &idle_sched_class;
6061 migrate_dead_tasks(cpu);
6062 spin_unlock_irq(&rq->lock);
6064 migrate_nr_uninterruptible(rq);
6065 BUG_ON(rq->nr_running != 0);
6068 * No need to migrate the tasks: it was best-effort if
6069 * they didn't take sched_hotcpu_mutex. Just wake up
6072 spin_lock_irq(&rq->lock);
6073 while (!list_empty(&rq->migration_queue)) {
6074 struct migration_req *req;
6076 req = list_entry(rq->migration_queue.next,
6077 struct migration_req, list);
6078 list_del_init(&req->list);
6079 complete(&req->done);
6081 spin_unlock_irq(&rq->lock);
6085 case CPU_DYING_FROZEN:
6086 /* Update our root-domain */
6088 spin_lock_irqsave(&rq->lock, flags);
6090 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6091 cpu_clear(cpu, rq->rd->online);
6093 spin_unlock_irqrestore(&rq->lock, flags);
6100 /* Register at highest priority so that task migration (migrate_all_tasks)
6101 * happens before everything else.
6103 static struct notifier_block __cpuinitdata migration_notifier = {
6104 .notifier_call = migration_call,
6108 void __init migration_init(void)
6110 void *cpu = (void *)(long)smp_processor_id();
6113 /* Start one for the boot CPU: */
6114 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6115 BUG_ON(err == NOTIFY_BAD);
6116 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6117 register_cpu_notifier(&migration_notifier);
6123 #ifdef CONFIG_SCHED_DEBUG
6125 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6126 cpumask_t *groupmask)
6128 struct sched_group *group = sd->groups;
6131 cpulist_scnprintf(str, sizeof(str), sd->span);
6132 cpus_clear(*groupmask);
6134 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6136 if (!(sd->flags & SD_LOAD_BALANCE)) {
6137 printk("does not load-balance\n");
6139 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6144 printk(KERN_CONT "span %s\n", str);
6146 if (!cpu_isset(cpu, sd->span)) {
6147 printk(KERN_ERR "ERROR: domain->span does not contain "
6150 if (!cpu_isset(cpu, group->cpumask)) {
6151 printk(KERN_ERR "ERROR: domain->groups does not contain"
6155 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6159 printk(KERN_ERR "ERROR: group is NULL\n");
6163 if (!group->__cpu_power) {
6164 printk(KERN_CONT "\n");
6165 printk(KERN_ERR "ERROR: domain->cpu_power not "
6170 if (!cpus_weight(group->cpumask)) {
6171 printk(KERN_CONT "\n");
6172 printk(KERN_ERR "ERROR: empty group\n");
6176 if (cpus_intersects(*groupmask, group->cpumask)) {
6177 printk(KERN_CONT "\n");
6178 printk(KERN_ERR "ERROR: repeated CPUs\n");
6182 cpus_or(*groupmask, *groupmask, group->cpumask);
6184 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6185 printk(KERN_CONT " %s", str);
6187 group = group->next;
6188 } while (group != sd->groups);
6189 printk(KERN_CONT "\n");
6191 if (!cpus_equal(sd->span, *groupmask))
6192 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6194 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6195 printk(KERN_ERR "ERROR: parent span is not a superset "
6196 "of domain->span\n");
6200 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6202 cpumask_t *groupmask;
6206 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6210 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6212 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6214 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6219 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6229 # define sched_domain_debug(sd, cpu) do { } while (0)
6232 static int sd_degenerate(struct sched_domain *sd)
6234 if (cpus_weight(sd->span) == 1)
6237 /* Following flags need at least 2 groups */
6238 if (sd->flags & (SD_LOAD_BALANCE |
6239 SD_BALANCE_NEWIDLE |
6243 SD_SHARE_PKG_RESOURCES)) {
6244 if (sd->groups != sd->groups->next)
6248 /* Following flags don't use groups */
6249 if (sd->flags & (SD_WAKE_IDLE |
6258 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6260 unsigned long cflags = sd->flags, pflags = parent->flags;
6262 if (sd_degenerate(parent))
6265 if (!cpus_equal(sd->span, parent->span))
6268 /* Does parent contain flags not in child? */
6269 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6270 if (cflags & SD_WAKE_AFFINE)
6271 pflags &= ~SD_WAKE_BALANCE;
6272 /* Flags needing groups don't count if only 1 group in parent */
6273 if (parent->groups == parent->groups->next) {
6274 pflags &= ~(SD_LOAD_BALANCE |
6275 SD_BALANCE_NEWIDLE |
6279 SD_SHARE_PKG_RESOURCES);
6281 if (~cflags & pflags)
6287 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6289 unsigned long flags;
6290 const struct sched_class *class;
6292 spin_lock_irqsave(&rq->lock, flags);
6295 struct root_domain *old_rd = rq->rd;
6297 for (class = sched_class_highest; class; class = class->next) {
6298 if (class->leave_domain)
6299 class->leave_domain(rq);
6302 cpu_clear(rq->cpu, old_rd->span);
6303 cpu_clear(rq->cpu, old_rd->online);
6305 if (atomic_dec_and_test(&old_rd->refcount))
6309 atomic_inc(&rd->refcount);
6312 cpu_set(rq->cpu, rd->span);
6313 if (cpu_isset(rq->cpu, cpu_online_map))
6314 cpu_set(rq->cpu, rd->online);
6316 for (class = sched_class_highest; class; class = class->next) {
6317 if (class->join_domain)
6318 class->join_domain(rq);
6321 spin_unlock_irqrestore(&rq->lock, flags);
6324 static void init_rootdomain(struct root_domain *rd)
6326 memset(rd, 0, sizeof(*rd));
6328 cpus_clear(rd->span);
6329 cpus_clear(rd->online);
6332 static void init_defrootdomain(void)
6334 init_rootdomain(&def_root_domain);
6335 atomic_set(&def_root_domain.refcount, 1);
6338 static struct root_domain *alloc_rootdomain(void)
6340 struct root_domain *rd;
6342 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6346 init_rootdomain(rd);
6352 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6353 * hold the hotplug lock.
6356 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6358 struct rq *rq = cpu_rq(cpu);
6359 struct sched_domain *tmp;
6361 /* Remove the sched domains which do not contribute to scheduling. */
6362 for (tmp = sd; tmp; tmp = tmp->parent) {
6363 struct sched_domain *parent = tmp->parent;
6366 if (sd_parent_degenerate(tmp, parent)) {
6367 tmp->parent = parent->parent;
6369 parent->parent->child = tmp;
6373 if (sd && sd_degenerate(sd)) {
6379 sched_domain_debug(sd, cpu);
6381 rq_attach_root(rq, rd);
6382 rcu_assign_pointer(rq->sd, sd);
6385 /* cpus with isolated domains */
6386 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6388 /* Setup the mask of cpus configured for isolated domains */
6389 static int __init isolated_cpu_setup(char *str)
6391 int ints[NR_CPUS], i;
6393 str = get_options(str, ARRAY_SIZE(ints), ints);
6394 cpus_clear(cpu_isolated_map);
6395 for (i = 1; i <= ints[0]; i++)
6396 if (ints[i] < NR_CPUS)
6397 cpu_set(ints[i], cpu_isolated_map);
6401 __setup("isolcpus=", isolated_cpu_setup);
6404 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6405 * to a function which identifies what group(along with sched group) a CPU
6406 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6407 * (due to the fact that we keep track of groups covered with a cpumask_t).
6409 * init_sched_build_groups will build a circular linked list of the groups
6410 * covered by the given span, and will set each group's ->cpumask correctly,
6411 * and ->cpu_power to 0.
6414 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6415 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6416 struct sched_group **sg,
6417 cpumask_t *tmpmask),
6418 cpumask_t *covered, cpumask_t *tmpmask)
6420 struct sched_group *first = NULL, *last = NULL;
6423 cpus_clear(*covered);
6425 for_each_cpu_mask(i, *span) {
6426 struct sched_group *sg;
6427 int group = group_fn(i, cpu_map, &sg, tmpmask);
6430 if (cpu_isset(i, *covered))
6433 cpus_clear(sg->cpumask);
6434 sg->__cpu_power = 0;
6436 for_each_cpu_mask(j, *span) {
6437 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6440 cpu_set(j, *covered);
6441 cpu_set(j, sg->cpumask);
6452 #define SD_NODES_PER_DOMAIN 16
6457 * find_next_best_node - find the next node to include in a sched_domain
6458 * @node: node whose sched_domain we're building
6459 * @used_nodes: nodes already in the sched_domain
6461 * Find the next node to include in a given scheduling domain. Simply
6462 * finds the closest node not already in the @used_nodes map.
6464 * Should use nodemask_t.
6466 static int find_next_best_node(int node, nodemask_t *used_nodes)
6468 int i, n, val, min_val, best_node = 0;
6472 for (i = 0; i < MAX_NUMNODES; i++) {
6473 /* Start at @node */
6474 n = (node + i) % MAX_NUMNODES;
6476 if (!nr_cpus_node(n))
6479 /* Skip already used nodes */
6480 if (node_isset(n, *used_nodes))
6483 /* Simple min distance search */
6484 val = node_distance(node, n);
6486 if (val < min_val) {
6492 node_set(best_node, *used_nodes);
6497 * sched_domain_node_span - get a cpumask for a node's sched_domain
6498 * @node: node whose cpumask we're constructing
6500 * Given a node, construct a good cpumask for its sched_domain to span. It
6501 * should be one that prevents unnecessary balancing, but also spreads tasks
6504 static void sched_domain_node_span(int node, cpumask_t *span)
6506 nodemask_t used_nodes;
6507 node_to_cpumask_ptr(nodemask, node);
6511 nodes_clear(used_nodes);
6513 cpus_or(*span, *span, *nodemask);
6514 node_set(node, used_nodes);
6516 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6517 int next_node = find_next_best_node(node, &used_nodes);
6519 node_to_cpumask_ptr_next(nodemask, next_node);
6520 cpus_or(*span, *span, *nodemask);
6525 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6528 * SMT sched-domains:
6530 #ifdef CONFIG_SCHED_SMT
6531 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6532 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6535 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6539 *sg = &per_cpu(sched_group_cpus, cpu);
6545 * multi-core sched-domains:
6547 #ifdef CONFIG_SCHED_MC
6548 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6549 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6552 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6554 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6559 *mask = per_cpu(cpu_sibling_map, cpu);
6560 cpus_and(*mask, *mask, *cpu_map);
6561 group = first_cpu(*mask);
6563 *sg = &per_cpu(sched_group_core, group);
6566 #elif defined(CONFIG_SCHED_MC)
6568 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6572 *sg = &per_cpu(sched_group_core, cpu);
6577 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6578 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6581 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6585 #ifdef CONFIG_SCHED_MC
6586 *mask = cpu_coregroup_map(cpu);
6587 cpus_and(*mask, *mask, *cpu_map);
6588 group = first_cpu(*mask);
6589 #elif defined(CONFIG_SCHED_SMT)
6590 *mask = per_cpu(cpu_sibling_map, cpu);
6591 cpus_and(*mask, *mask, *cpu_map);
6592 group = first_cpu(*mask);
6597 *sg = &per_cpu(sched_group_phys, group);
6603 * The init_sched_build_groups can't handle what we want to do with node
6604 * groups, so roll our own. Now each node has its own list of groups which
6605 * gets dynamically allocated.
6607 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6608 static struct sched_group ***sched_group_nodes_bycpu;
6610 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6611 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6613 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6614 struct sched_group **sg, cpumask_t *nodemask)
6618 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6619 cpus_and(*nodemask, *nodemask, *cpu_map);
6620 group = first_cpu(*nodemask);
6623 *sg = &per_cpu(sched_group_allnodes, group);
6627 static void init_numa_sched_groups_power(struct sched_group *group_head)
6629 struct sched_group *sg = group_head;
6635 for_each_cpu_mask(j, sg->cpumask) {
6636 struct sched_domain *sd;
6638 sd = &per_cpu(phys_domains, j);
6639 if (j != first_cpu(sd->groups->cpumask)) {
6641 * Only add "power" once for each
6647 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6650 } while (sg != group_head);
6655 /* Free memory allocated for various sched_group structures */
6656 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6660 for_each_cpu_mask(cpu, *cpu_map) {
6661 struct sched_group **sched_group_nodes
6662 = sched_group_nodes_bycpu[cpu];
6664 if (!sched_group_nodes)
6667 for (i = 0; i < MAX_NUMNODES; i++) {
6668 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6670 *nodemask = node_to_cpumask(i);
6671 cpus_and(*nodemask, *nodemask, *cpu_map);
6672 if (cpus_empty(*nodemask))
6682 if (oldsg != sched_group_nodes[i])
6685 kfree(sched_group_nodes);
6686 sched_group_nodes_bycpu[cpu] = NULL;
6690 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6696 * Initialize sched groups cpu_power.
6698 * cpu_power indicates the capacity of sched group, which is used while
6699 * distributing the load between different sched groups in a sched domain.
6700 * Typically cpu_power for all the groups in a sched domain will be same unless
6701 * there are asymmetries in the topology. If there are asymmetries, group
6702 * having more cpu_power will pickup more load compared to the group having
6705 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6706 * the maximum number of tasks a group can handle in the presence of other idle
6707 * or lightly loaded groups in the same sched domain.
6709 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6711 struct sched_domain *child;
6712 struct sched_group *group;
6714 WARN_ON(!sd || !sd->groups);
6716 if (cpu != first_cpu(sd->groups->cpumask))
6721 sd->groups->__cpu_power = 0;
6724 * For perf policy, if the groups in child domain share resources
6725 * (for example cores sharing some portions of the cache hierarchy
6726 * or SMT), then set this domain groups cpu_power such that each group
6727 * can handle only one task, when there are other idle groups in the
6728 * same sched domain.
6730 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6732 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6733 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6738 * add cpu_power of each child group to this groups cpu_power
6740 group = child->groups;
6742 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6743 group = group->next;
6744 } while (group != child->groups);
6748 * Initializers for schedule domains
6749 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6752 #define SD_INIT(sd, type) sd_init_##type(sd)
6753 #define SD_INIT_FUNC(type) \
6754 static noinline void sd_init_##type(struct sched_domain *sd) \
6756 memset(sd, 0, sizeof(*sd)); \
6757 *sd = SD_##type##_INIT; \
6762 SD_INIT_FUNC(ALLNODES)
6765 #ifdef CONFIG_SCHED_SMT
6766 SD_INIT_FUNC(SIBLING)
6768 #ifdef CONFIG_SCHED_MC
6773 * To minimize stack usage kmalloc room for cpumasks and share the
6774 * space as the usage in build_sched_domains() dictates. Used only
6775 * if the amount of space is significant.
6778 cpumask_t tmpmask; /* make this one first */
6781 cpumask_t this_sibling_map;
6782 cpumask_t this_core_map;
6784 cpumask_t send_covered;
6787 cpumask_t domainspan;
6789 cpumask_t notcovered;
6794 #define SCHED_CPUMASK_ALLOC 1
6795 #define SCHED_CPUMASK_FREE(v) kfree(v)
6796 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6798 #define SCHED_CPUMASK_ALLOC 0
6799 #define SCHED_CPUMASK_FREE(v)
6800 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6803 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6804 ((unsigned long)(a) + offsetof(struct allmasks, v))
6807 * Build sched domains for a given set of cpus and attach the sched domains
6808 * to the individual cpus
6810 static int build_sched_domains(const cpumask_t *cpu_map)
6813 struct root_domain *rd;
6814 SCHED_CPUMASK_DECLARE(allmasks);
6817 struct sched_group **sched_group_nodes = NULL;
6818 int sd_allnodes = 0;
6821 * Allocate the per-node list of sched groups
6823 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6825 if (!sched_group_nodes) {
6826 printk(KERN_WARNING "Can not alloc sched group node list\n");
6831 rd = alloc_rootdomain();
6833 printk(KERN_WARNING "Cannot alloc root domain\n");
6835 kfree(sched_group_nodes);
6840 #if SCHED_CPUMASK_ALLOC
6841 /* get space for all scratch cpumask variables */
6842 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6844 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6847 kfree(sched_group_nodes);
6852 tmpmask = (cpumask_t *)allmasks;
6856 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6860 * Set up domains for cpus specified by the cpu_map.
6862 for_each_cpu_mask(i, *cpu_map) {
6863 struct sched_domain *sd = NULL, *p;
6864 SCHED_CPUMASK_VAR(nodemask, allmasks);
6866 *nodemask = node_to_cpumask(cpu_to_node(i));
6867 cpus_and(*nodemask, *nodemask, *cpu_map);
6870 if (cpus_weight(*cpu_map) >
6871 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
6872 sd = &per_cpu(allnodes_domains, i);
6873 SD_INIT(sd, ALLNODES);
6874 sd->span = *cpu_map;
6875 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
6881 sd = &per_cpu(node_domains, i);
6883 sched_domain_node_span(cpu_to_node(i), &sd->span);
6887 cpus_and(sd->span, sd->span, *cpu_map);
6891 sd = &per_cpu(phys_domains, i);
6893 sd->span = *nodemask;
6897 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
6899 #ifdef CONFIG_SCHED_MC
6901 sd = &per_cpu(core_domains, i);
6903 sd->span = cpu_coregroup_map(i);
6904 cpus_and(sd->span, sd->span, *cpu_map);
6907 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
6910 #ifdef CONFIG_SCHED_SMT
6912 sd = &per_cpu(cpu_domains, i);
6913 SD_INIT(sd, SIBLING);
6914 sd->span = per_cpu(cpu_sibling_map, i);
6915 cpus_and(sd->span, sd->span, *cpu_map);
6918 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
6922 #ifdef CONFIG_SCHED_SMT
6923 /* Set up CPU (sibling) groups */
6924 for_each_cpu_mask(i, *cpu_map) {
6925 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
6926 SCHED_CPUMASK_VAR(send_covered, allmasks);
6928 *this_sibling_map = per_cpu(cpu_sibling_map, i);
6929 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
6930 if (i != first_cpu(*this_sibling_map))
6933 init_sched_build_groups(this_sibling_map, cpu_map,
6935 send_covered, tmpmask);
6939 #ifdef CONFIG_SCHED_MC
6940 /* Set up multi-core groups */
6941 for_each_cpu_mask(i, *cpu_map) {
6942 SCHED_CPUMASK_VAR(this_core_map, allmasks);
6943 SCHED_CPUMASK_VAR(send_covered, allmasks);
6945 *this_core_map = cpu_coregroup_map(i);
6946 cpus_and(*this_core_map, *this_core_map, *cpu_map);
6947 if (i != first_cpu(*this_core_map))
6950 init_sched_build_groups(this_core_map, cpu_map,
6952 send_covered, tmpmask);
6956 /* Set up physical groups */
6957 for (i = 0; i < MAX_NUMNODES; i++) {
6958 SCHED_CPUMASK_VAR(nodemask, allmasks);
6959 SCHED_CPUMASK_VAR(send_covered, allmasks);
6961 *nodemask = node_to_cpumask(i);
6962 cpus_and(*nodemask, *nodemask, *cpu_map);
6963 if (cpus_empty(*nodemask))
6966 init_sched_build_groups(nodemask, cpu_map,
6968 send_covered, tmpmask);
6972 /* Set up node groups */
6974 SCHED_CPUMASK_VAR(send_covered, allmasks);
6976 init_sched_build_groups(cpu_map, cpu_map,
6977 &cpu_to_allnodes_group,
6978 send_covered, tmpmask);
6981 for (i = 0; i < MAX_NUMNODES; i++) {
6982 /* Set up node groups */
6983 struct sched_group *sg, *prev;
6984 SCHED_CPUMASK_VAR(nodemask, allmasks);
6985 SCHED_CPUMASK_VAR(domainspan, allmasks);
6986 SCHED_CPUMASK_VAR(covered, allmasks);
6989 *nodemask = node_to_cpumask(i);
6990 cpus_clear(*covered);
6992 cpus_and(*nodemask, *nodemask, *cpu_map);
6993 if (cpus_empty(*nodemask)) {
6994 sched_group_nodes[i] = NULL;
6998 sched_domain_node_span(i, domainspan);
6999 cpus_and(*domainspan, *domainspan, *cpu_map);
7001 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7003 printk(KERN_WARNING "Can not alloc domain group for "
7007 sched_group_nodes[i] = sg;
7008 for_each_cpu_mask(j, *nodemask) {
7009 struct sched_domain *sd;
7011 sd = &per_cpu(node_domains, j);
7014 sg->__cpu_power = 0;
7015 sg->cpumask = *nodemask;
7017 cpus_or(*covered, *covered, *nodemask);
7020 for (j = 0; j < MAX_NUMNODES; j++) {
7021 SCHED_CPUMASK_VAR(notcovered, allmasks);
7022 int n = (i + j) % MAX_NUMNODES;
7023 node_to_cpumask_ptr(pnodemask, n);
7025 cpus_complement(*notcovered, *covered);
7026 cpus_and(*tmpmask, *notcovered, *cpu_map);
7027 cpus_and(*tmpmask, *tmpmask, *domainspan);
7028 if (cpus_empty(*tmpmask))
7031 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7032 if (cpus_empty(*tmpmask))
7035 sg = kmalloc_node(sizeof(struct sched_group),
7039 "Can not alloc domain group for node %d\n", j);
7042 sg->__cpu_power = 0;
7043 sg->cpumask = *tmpmask;
7044 sg->next = prev->next;
7045 cpus_or(*covered, *covered, *tmpmask);
7052 /* Calculate CPU power for physical packages and nodes */
7053 #ifdef CONFIG_SCHED_SMT
7054 for_each_cpu_mask(i, *cpu_map) {
7055 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7057 init_sched_groups_power(i, sd);
7060 #ifdef CONFIG_SCHED_MC
7061 for_each_cpu_mask(i, *cpu_map) {
7062 struct sched_domain *sd = &per_cpu(core_domains, i);
7064 init_sched_groups_power(i, sd);
7068 for_each_cpu_mask(i, *cpu_map) {
7069 struct sched_domain *sd = &per_cpu(phys_domains, i);
7071 init_sched_groups_power(i, sd);
7075 for (i = 0; i < MAX_NUMNODES; i++)
7076 init_numa_sched_groups_power(sched_group_nodes[i]);
7079 struct sched_group *sg;
7081 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7083 init_numa_sched_groups_power(sg);
7087 /* Attach the domains */
7088 for_each_cpu_mask(i, *cpu_map) {
7089 struct sched_domain *sd;
7090 #ifdef CONFIG_SCHED_SMT
7091 sd = &per_cpu(cpu_domains, i);
7092 #elif defined(CONFIG_SCHED_MC)
7093 sd = &per_cpu(core_domains, i);
7095 sd = &per_cpu(phys_domains, i);
7097 cpu_attach_domain(sd, rd, i);
7100 SCHED_CPUMASK_FREE((void *)allmasks);
7105 free_sched_groups(cpu_map, tmpmask);
7106 SCHED_CPUMASK_FREE((void *)allmasks);
7111 static cpumask_t *doms_cur; /* current sched domains */
7112 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7115 * Special case: If a kmalloc of a doms_cur partition (array of
7116 * cpumask_t) fails, then fallback to a single sched domain,
7117 * as determined by the single cpumask_t fallback_doms.
7119 static cpumask_t fallback_doms;
7121 void __attribute__((weak)) arch_update_cpu_topology(void)
7126 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7127 * For now this just excludes isolated cpus, but could be used to
7128 * exclude other special cases in the future.
7130 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7134 arch_update_cpu_topology();
7136 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7138 doms_cur = &fallback_doms;
7139 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7140 err = build_sched_domains(doms_cur);
7141 register_sched_domain_sysctl();
7146 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7149 free_sched_groups(cpu_map, tmpmask);
7153 * Detach sched domains from a group of cpus specified in cpu_map
7154 * These cpus will now be attached to the NULL domain
7156 static void detach_destroy_domains(const cpumask_t *cpu_map)
7161 unregister_sched_domain_sysctl();
7163 for_each_cpu_mask(i, *cpu_map)
7164 cpu_attach_domain(NULL, &def_root_domain, i);
7165 synchronize_sched();
7166 arch_destroy_sched_domains(cpu_map, &tmpmask);
7170 * Partition sched domains as specified by the 'ndoms_new'
7171 * cpumasks in the array doms_new[] of cpumasks. This compares
7172 * doms_new[] to the current sched domain partitioning, doms_cur[].
7173 * It destroys each deleted domain and builds each new domain.
7175 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7176 * The masks don't intersect (don't overlap.) We should setup one
7177 * sched domain for each mask. CPUs not in any of the cpumasks will
7178 * not be load balanced. If the same cpumask appears both in the
7179 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7182 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7183 * ownership of it and will kfree it when done with it. If the caller
7184 * failed the kmalloc call, then it can pass in doms_new == NULL,
7185 * and partition_sched_domains() will fallback to the single partition
7188 * Call with hotplug lock held
7190 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
7196 /* always unregister in case we don't destroy any domains */
7197 unregister_sched_domain_sysctl();
7199 if (doms_new == NULL) {
7201 doms_new = &fallback_doms;
7202 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7205 /* Destroy deleted domains */
7206 for (i = 0; i < ndoms_cur; i++) {
7207 for (j = 0; j < ndoms_new; j++) {
7208 if (cpus_equal(doms_cur[i], doms_new[j]))
7211 /* no match - a current sched domain not in new doms_new[] */
7212 detach_destroy_domains(doms_cur + i);
7217 /* Build new domains */
7218 for (i = 0; i < ndoms_new; i++) {
7219 for (j = 0; j < ndoms_cur; j++) {
7220 if (cpus_equal(doms_new[i], doms_cur[j]))
7223 /* no match - add a new doms_new */
7224 build_sched_domains(doms_new + i);
7229 /* Remember the new sched domains */
7230 if (doms_cur != &fallback_doms)
7232 doms_cur = doms_new;
7233 ndoms_cur = ndoms_new;
7235 register_sched_domain_sysctl();
7240 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7241 int arch_reinit_sched_domains(void)
7246 detach_destroy_domains(&cpu_online_map);
7247 err = arch_init_sched_domains(&cpu_online_map);
7253 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7257 if (buf[0] != '0' && buf[0] != '1')
7261 sched_smt_power_savings = (buf[0] == '1');
7263 sched_mc_power_savings = (buf[0] == '1');
7265 ret = arch_reinit_sched_domains();
7267 return ret ? ret : count;
7270 #ifdef CONFIG_SCHED_MC
7271 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7273 return sprintf(page, "%u\n", sched_mc_power_savings);
7275 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7276 const char *buf, size_t count)
7278 return sched_power_savings_store(buf, count, 0);
7280 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7281 sched_mc_power_savings_store);
7284 #ifdef CONFIG_SCHED_SMT
7285 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7287 return sprintf(page, "%u\n", sched_smt_power_savings);
7289 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7290 const char *buf, size_t count)
7292 return sched_power_savings_store(buf, count, 1);
7294 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7295 sched_smt_power_savings_store);
7298 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7302 #ifdef CONFIG_SCHED_SMT
7304 err = sysfs_create_file(&cls->kset.kobj,
7305 &attr_sched_smt_power_savings.attr);
7307 #ifdef CONFIG_SCHED_MC
7308 if (!err && mc_capable())
7309 err = sysfs_create_file(&cls->kset.kobj,
7310 &attr_sched_mc_power_savings.attr);
7317 * Force a reinitialization of the sched domains hierarchy. The domains
7318 * and groups cannot be updated in place without racing with the balancing
7319 * code, so we temporarily attach all running cpus to the NULL domain
7320 * which will prevent rebalancing while the sched domains are recalculated.
7322 static int update_sched_domains(struct notifier_block *nfb,
7323 unsigned long action, void *hcpu)
7326 case CPU_UP_PREPARE:
7327 case CPU_UP_PREPARE_FROZEN:
7328 case CPU_DOWN_PREPARE:
7329 case CPU_DOWN_PREPARE_FROZEN:
7330 detach_destroy_domains(&cpu_online_map);
7333 case CPU_UP_CANCELED:
7334 case CPU_UP_CANCELED_FROZEN:
7335 case CPU_DOWN_FAILED:
7336 case CPU_DOWN_FAILED_FROZEN:
7338 case CPU_ONLINE_FROZEN:
7340 case CPU_DEAD_FROZEN:
7342 * Fall through and re-initialise the domains.
7349 /* The hotplug lock is already held by cpu_up/cpu_down */
7350 arch_init_sched_domains(&cpu_online_map);
7355 void __init sched_init_smp(void)
7357 cpumask_t non_isolated_cpus;
7359 #if defined(CONFIG_NUMA)
7360 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7362 BUG_ON(sched_group_nodes_bycpu == NULL);
7365 arch_init_sched_domains(&cpu_online_map);
7366 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7367 if (cpus_empty(non_isolated_cpus))
7368 cpu_set(smp_processor_id(), non_isolated_cpus);
7370 /* XXX: Theoretical race here - CPU may be hotplugged now */
7371 hotcpu_notifier(update_sched_domains, 0);
7373 /* Move init over to a non-isolated CPU */
7374 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7376 sched_init_granularity();
7379 void __init sched_init_smp(void)
7381 #if defined(CONFIG_NUMA)
7382 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7384 BUG_ON(sched_group_nodes_bycpu == NULL);
7386 sched_init_granularity();
7388 #endif /* CONFIG_SMP */
7390 int in_sched_functions(unsigned long addr)
7392 return in_lock_functions(addr) ||
7393 (addr >= (unsigned long)__sched_text_start
7394 && addr < (unsigned long)__sched_text_end);
7397 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7399 cfs_rq->tasks_timeline = RB_ROOT;
7400 #ifdef CONFIG_FAIR_GROUP_SCHED
7403 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7406 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7408 struct rt_prio_array *array;
7411 array = &rt_rq->active;
7412 for (i = 0; i < MAX_RT_PRIO; i++) {
7413 INIT_LIST_HEAD(array->queue + i);
7414 __clear_bit(i, array->bitmap);
7416 /* delimiter for bitsearch: */
7417 __set_bit(MAX_RT_PRIO, array->bitmap);
7419 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7420 rt_rq->highest_prio = MAX_RT_PRIO;
7423 rt_rq->rt_nr_migratory = 0;
7424 rt_rq->overloaded = 0;
7428 rt_rq->rt_throttled = 0;
7429 rt_rq->rt_runtime = 0;
7430 spin_lock_init(&rt_rq->rt_runtime_lock);
7432 #ifdef CONFIG_RT_GROUP_SCHED
7433 rt_rq->rt_nr_boosted = 0;
7438 #ifdef CONFIG_FAIR_GROUP_SCHED
7439 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7440 struct cfs_rq *cfs_rq, struct sched_entity *se,
7443 tg->cfs_rq[cpu] = cfs_rq;
7444 init_cfs_rq(cfs_rq, rq);
7447 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7450 se->cfs_rq = &rq->cfs;
7452 se->load.weight = tg->shares;
7453 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7458 #ifdef CONFIG_RT_GROUP_SCHED
7459 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7460 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7463 tg->rt_rq[cpu] = rt_rq;
7464 init_rt_rq(rt_rq, rq);
7466 rt_rq->rt_se = rt_se;
7467 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7469 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7471 tg->rt_se[cpu] = rt_se;
7472 rt_se->rt_rq = &rq->rt;
7473 rt_se->my_q = rt_rq;
7474 rt_se->parent = NULL;
7475 INIT_LIST_HEAD(&rt_se->run_list);
7479 void __init sched_init(void)
7482 unsigned long alloc_size = 0, ptr;
7484 #ifdef CONFIG_FAIR_GROUP_SCHED
7485 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7487 #ifdef CONFIG_RT_GROUP_SCHED
7488 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7491 * As sched_init() is called before page_alloc is setup,
7492 * we use alloc_bootmem().
7495 ptr = (unsigned long)alloc_bootmem_low(alloc_size);
7497 #ifdef CONFIG_FAIR_GROUP_SCHED
7498 init_task_group.se = (struct sched_entity **)ptr;
7499 ptr += nr_cpu_ids * sizeof(void **);
7501 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7502 ptr += nr_cpu_ids * sizeof(void **);
7504 #ifdef CONFIG_RT_GROUP_SCHED
7505 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7506 ptr += nr_cpu_ids * sizeof(void **);
7508 init_task_group.rt_rq = (struct rt_rq **)ptr;
7513 init_defrootdomain();
7516 init_rt_bandwidth(&def_rt_bandwidth,
7517 global_rt_period(), global_rt_runtime());
7519 #ifdef CONFIG_RT_GROUP_SCHED
7520 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7521 global_rt_period(), global_rt_runtime());
7524 #ifdef CONFIG_GROUP_SCHED
7525 list_add(&init_task_group.list, &task_groups);
7528 for_each_possible_cpu(i) {
7532 spin_lock_init(&rq->lock);
7533 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7536 update_last_tick_seen(rq);
7537 init_cfs_rq(&rq->cfs, rq);
7538 init_rt_rq(&rq->rt, rq);
7539 #ifdef CONFIG_FAIR_GROUP_SCHED
7540 init_task_group.shares = init_task_group_load;
7541 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7542 init_tg_cfs_entry(rq, &init_task_group,
7543 &per_cpu(init_cfs_rq, i),
7544 &per_cpu(init_sched_entity, i), i, 1);
7547 #ifdef CONFIG_RT_GROUP_SCHED
7548 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7549 init_tg_rt_entry(rq, &init_task_group,
7550 &per_cpu(init_rt_rq, i),
7551 &per_cpu(init_sched_rt_entity, i), i, 1);
7553 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7556 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7557 rq->cpu_load[j] = 0;
7561 rq->active_balance = 0;
7562 rq->next_balance = jiffies;
7565 rq->migration_thread = NULL;
7566 INIT_LIST_HEAD(&rq->migration_queue);
7567 rq_attach_root(rq, &def_root_domain);
7570 atomic_set(&rq->nr_iowait, 0);
7573 set_load_weight(&init_task);
7575 #ifdef CONFIG_PREEMPT_NOTIFIERS
7576 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7580 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7583 #ifdef CONFIG_RT_MUTEXES
7584 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7588 * The boot idle thread does lazy MMU switching as well:
7590 atomic_inc(&init_mm.mm_count);
7591 enter_lazy_tlb(&init_mm, current);
7594 * Make us the idle thread. Technically, schedule() should not be
7595 * called from this thread, however somewhere below it might be,
7596 * but because we are the idle thread, we just pick up running again
7597 * when this runqueue becomes "idle".
7599 init_idle(current, smp_processor_id());
7601 * During early bootup we pretend to be a normal task:
7603 current->sched_class = &fair_sched_class;
7605 scheduler_running = 1;
7608 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7609 void __might_sleep(char *file, int line)
7612 static unsigned long prev_jiffy; /* ratelimiting */
7614 if ((in_atomic() || irqs_disabled()) &&
7615 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7616 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7618 prev_jiffy = jiffies;
7619 printk(KERN_ERR "BUG: sleeping function called from invalid"
7620 " context at %s:%d\n", file, line);
7621 printk("in_atomic():%d, irqs_disabled():%d\n",
7622 in_atomic(), irqs_disabled());
7623 debug_show_held_locks(current);
7624 if (irqs_disabled())
7625 print_irqtrace_events(current);
7630 EXPORT_SYMBOL(__might_sleep);
7633 #ifdef CONFIG_MAGIC_SYSRQ
7634 static void normalize_task(struct rq *rq, struct task_struct *p)
7637 update_rq_clock(rq);
7638 on_rq = p->se.on_rq;
7640 deactivate_task(rq, p, 0);
7641 __setscheduler(rq, p, SCHED_NORMAL, 0);
7643 activate_task(rq, p, 0);
7644 resched_task(rq->curr);
7648 void normalize_rt_tasks(void)
7650 struct task_struct *g, *p;
7651 unsigned long flags;
7654 read_lock_irqsave(&tasklist_lock, flags);
7655 do_each_thread(g, p) {
7657 * Only normalize user tasks:
7662 p->se.exec_start = 0;
7663 #ifdef CONFIG_SCHEDSTATS
7664 p->se.wait_start = 0;
7665 p->se.sleep_start = 0;
7666 p->se.block_start = 0;
7668 task_rq(p)->clock = 0;
7672 * Renice negative nice level userspace
7675 if (TASK_NICE(p) < 0 && p->mm)
7676 set_user_nice(p, 0);
7680 spin_lock(&p->pi_lock);
7681 rq = __task_rq_lock(p);
7683 normalize_task(rq, p);
7685 __task_rq_unlock(rq);
7686 spin_unlock(&p->pi_lock);
7687 } while_each_thread(g, p);
7689 read_unlock_irqrestore(&tasklist_lock, flags);
7692 #endif /* CONFIG_MAGIC_SYSRQ */
7696 * These functions are only useful for the IA64 MCA handling.
7698 * They can only be called when the whole system has been
7699 * stopped - every CPU needs to be quiescent, and no scheduling
7700 * activity can take place. Using them for anything else would
7701 * be a serious bug, and as a result, they aren't even visible
7702 * under any other configuration.
7706 * curr_task - return the current task for a given cpu.
7707 * @cpu: the processor in question.
7709 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7711 struct task_struct *curr_task(int cpu)
7713 return cpu_curr(cpu);
7717 * set_curr_task - set the current task for a given cpu.
7718 * @cpu: the processor in question.
7719 * @p: the task pointer to set.
7721 * Description: This function must only be used when non-maskable interrupts
7722 * are serviced on a separate stack. It allows the architecture to switch the
7723 * notion of the current task on a cpu in a non-blocking manner. This function
7724 * must be called with all CPU's synchronized, and interrupts disabled, the
7725 * and caller must save the original value of the current task (see
7726 * curr_task() above) and restore that value before reenabling interrupts and
7727 * re-starting the system.
7729 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7731 void set_curr_task(int cpu, struct task_struct *p)
7738 #ifdef CONFIG_FAIR_GROUP_SCHED
7739 static void free_fair_sched_group(struct task_group *tg)
7743 for_each_possible_cpu(i) {
7745 kfree(tg->cfs_rq[i]);
7754 static int alloc_fair_sched_group(struct task_group *tg)
7756 struct cfs_rq *cfs_rq;
7757 struct sched_entity *se;
7761 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7764 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7768 tg->shares = NICE_0_LOAD;
7770 for_each_possible_cpu(i) {
7773 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7774 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7778 se = kmalloc_node(sizeof(struct sched_entity),
7779 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7783 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7792 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7794 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7795 &cpu_rq(cpu)->leaf_cfs_rq_list);
7798 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7800 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7803 static inline void free_fair_sched_group(struct task_group *tg)
7807 static inline int alloc_fair_sched_group(struct task_group *tg)
7812 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7816 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7821 #ifdef CONFIG_RT_GROUP_SCHED
7822 static void free_rt_sched_group(struct task_group *tg)
7826 destroy_rt_bandwidth(&tg->rt_bandwidth);
7828 for_each_possible_cpu(i) {
7830 kfree(tg->rt_rq[i]);
7832 kfree(tg->rt_se[i]);
7839 static int alloc_rt_sched_group(struct task_group *tg)
7841 struct rt_rq *rt_rq;
7842 struct sched_rt_entity *rt_se;
7846 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7849 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7853 init_rt_bandwidth(&tg->rt_bandwidth,
7854 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7856 for_each_possible_cpu(i) {
7859 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7860 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7864 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7865 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7869 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7878 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7880 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7881 &cpu_rq(cpu)->leaf_rt_rq_list);
7884 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7886 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7889 static inline void free_rt_sched_group(struct task_group *tg)
7893 static inline int alloc_rt_sched_group(struct task_group *tg)
7898 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7902 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7907 #ifdef CONFIG_GROUP_SCHED
7908 static void free_sched_group(struct task_group *tg)
7910 free_fair_sched_group(tg);
7911 free_rt_sched_group(tg);
7915 /* allocate runqueue etc for a new task group */
7916 struct task_group *sched_create_group(void)
7918 struct task_group *tg;
7919 unsigned long flags;
7922 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7924 return ERR_PTR(-ENOMEM);
7926 if (!alloc_fair_sched_group(tg))
7929 if (!alloc_rt_sched_group(tg))
7932 spin_lock_irqsave(&task_group_lock, flags);
7933 for_each_possible_cpu(i) {
7934 register_fair_sched_group(tg, i);
7935 register_rt_sched_group(tg, i);
7937 list_add_rcu(&tg->list, &task_groups);
7938 spin_unlock_irqrestore(&task_group_lock, flags);
7943 free_sched_group(tg);
7944 return ERR_PTR(-ENOMEM);
7947 /* rcu callback to free various structures associated with a task group */
7948 static void free_sched_group_rcu(struct rcu_head *rhp)
7950 /* now it should be safe to free those cfs_rqs */
7951 free_sched_group(container_of(rhp, struct task_group, rcu));
7954 /* Destroy runqueue etc associated with a task group */
7955 void sched_destroy_group(struct task_group *tg)
7957 unsigned long flags;
7960 spin_lock_irqsave(&task_group_lock, flags);
7961 for_each_possible_cpu(i) {
7962 unregister_fair_sched_group(tg, i);
7963 unregister_rt_sched_group(tg, i);
7965 list_del_rcu(&tg->list);
7966 spin_unlock_irqrestore(&task_group_lock, flags);
7968 /* wait for possible concurrent references to cfs_rqs complete */
7969 call_rcu(&tg->rcu, free_sched_group_rcu);
7972 /* change task's runqueue when it moves between groups.
7973 * The caller of this function should have put the task in its new group
7974 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7975 * reflect its new group.
7977 void sched_move_task(struct task_struct *tsk)
7980 unsigned long flags;
7983 rq = task_rq_lock(tsk, &flags);
7985 update_rq_clock(rq);
7987 running = task_current(rq, tsk);
7988 on_rq = tsk->se.on_rq;
7991 dequeue_task(rq, tsk, 0);
7992 if (unlikely(running))
7993 tsk->sched_class->put_prev_task(rq, tsk);
7995 set_task_rq(tsk, task_cpu(tsk));
7997 #ifdef CONFIG_FAIR_GROUP_SCHED
7998 if (tsk->sched_class->moved_group)
7999 tsk->sched_class->moved_group(tsk);
8002 if (unlikely(running))
8003 tsk->sched_class->set_curr_task(rq);
8005 enqueue_task(rq, tsk, 0);
8007 task_rq_unlock(rq, &flags);
8011 #ifdef CONFIG_FAIR_GROUP_SCHED
8012 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8014 struct cfs_rq *cfs_rq = se->cfs_rq;
8015 struct rq *rq = cfs_rq->rq;
8018 spin_lock_irq(&rq->lock);
8022 dequeue_entity(cfs_rq, se, 0);
8024 se->load.weight = shares;
8025 se->load.inv_weight = div64_64((1ULL<<32), shares);
8028 enqueue_entity(cfs_rq, se, 0);
8030 spin_unlock_irq(&rq->lock);
8033 static DEFINE_MUTEX(shares_mutex);
8035 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8038 unsigned long flags;
8041 * A weight of 0 or 1 can cause arithmetics problems.
8042 * (The default weight is 1024 - so there's no practical
8043 * limitation from this.)
8048 mutex_lock(&shares_mutex);
8049 if (tg->shares == shares)
8052 spin_lock_irqsave(&task_group_lock, flags);
8053 for_each_possible_cpu(i)
8054 unregister_fair_sched_group(tg, i);
8055 spin_unlock_irqrestore(&task_group_lock, flags);
8057 /* wait for any ongoing reference to this group to finish */
8058 synchronize_sched();
8061 * Now we are free to modify the group's share on each cpu
8062 * w/o tripping rebalance_share or load_balance_fair.
8064 tg->shares = shares;
8065 for_each_possible_cpu(i)
8066 set_se_shares(tg->se[i], shares);
8069 * Enable load balance activity on this group, by inserting it back on
8070 * each cpu's rq->leaf_cfs_rq_list.
8072 spin_lock_irqsave(&task_group_lock, flags);
8073 for_each_possible_cpu(i)
8074 register_fair_sched_group(tg, i);
8075 spin_unlock_irqrestore(&task_group_lock, flags);
8077 mutex_unlock(&shares_mutex);
8081 unsigned long sched_group_shares(struct task_group *tg)
8087 #ifdef CONFIG_RT_GROUP_SCHED
8089 * Ensure that the real time constraints are schedulable.
8091 static DEFINE_MUTEX(rt_constraints_mutex);
8093 static unsigned long to_ratio(u64 period, u64 runtime)
8095 if (runtime == RUNTIME_INF)
8098 return div64_64(runtime << 16, period);
8101 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8103 struct task_group *tgi;
8104 unsigned long total = 0;
8105 unsigned long global_ratio =
8106 to_ratio(global_rt_period(), global_rt_runtime());
8109 list_for_each_entry_rcu(tgi, &task_groups, list) {
8113 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8114 tgi->rt_bandwidth.rt_runtime);
8118 return total + to_ratio(period, runtime) < global_ratio;
8121 /* Must be called with tasklist_lock held */
8122 static inline int tg_has_rt_tasks(struct task_group *tg)
8124 struct task_struct *g, *p;
8125 do_each_thread(g, p) {
8126 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8128 } while_each_thread(g, p);
8132 static int tg_set_bandwidth(struct task_group *tg,
8133 u64 rt_period, u64 rt_runtime)
8137 mutex_lock(&rt_constraints_mutex);
8138 read_lock(&tasklist_lock);
8139 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8143 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8148 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8149 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8150 tg->rt_bandwidth.rt_runtime = rt_runtime;
8152 for_each_possible_cpu(i) {
8153 struct rt_rq *rt_rq = tg->rt_rq[i];
8155 spin_lock(&rt_rq->rt_runtime_lock);
8156 rt_rq->rt_runtime = rt_runtime;
8157 spin_unlock(&rt_rq->rt_runtime_lock);
8159 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8161 read_unlock(&tasklist_lock);
8162 mutex_unlock(&rt_constraints_mutex);
8167 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8169 u64 rt_runtime, rt_period;
8171 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8172 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8173 if (rt_runtime_us < 0)
8174 rt_runtime = RUNTIME_INF;
8176 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8179 long sched_group_rt_runtime(struct task_group *tg)
8183 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8186 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8187 do_div(rt_runtime_us, NSEC_PER_USEC);
8188 return rt_runtime_us;
8191 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8193 u64 rt_runtime, rt_period;
8195 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8196 rt_runtime = tg->rt_bandwidth.rt_runtime;
8198 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8201 long sched_group_rt_period(struct task_group *tg)
8205 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8206 do_div(rt_period_us, NSEC_PER_USEC);
8207 return rt_period_us;
8210 static int sched_rt_global_constraints(void)
8214 mutex_lock(&rt_constraints_mutex);
8215 if (!__rt_schedulable(NULL, 1, 0))
8217 mutex_unlock(&rt_constraints_mutex);
8222 static int sched_rt_global_constraints(void)
8224 unsigned long flags;
8227 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8228 for_each_possible_cpu(i) {
8229 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8231 spin_lock(&rt_rq->rt_runtime_lock);
8232 rt_rq->rt_runtime = global_rt_runtime();
8233 spin_unlock(&rt_rq->rt_runtime_lock);
8235 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8241 int sched_rt_handler(struct ctl_table *table, int write,
8242 struct file *filp, void __user *buffer, size_t *lenp,
8246 int old_period, old_runtime;
8247 static DEFINE_MUTEX(mutex);
8250 old_period = sysctl_sched_rt_period;
8251 old_runtime = sysctl_sched_rt_runtime;
8253 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8255 if (!ret && write) {
8256 ret = sched_rt_global_constraints();
8258 sysctl_sched_rt_period = old_period;
8259 sysctl_sched_rt_runtime = old_runtime;
8261 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8262 def_rt_bandwidth.rt_period =
8263 ns_to_ktime(global_rt_period());
8266 mutex_unlock(&mutex);
8271 #ifdef CONFIG_CGROUP_SCHED
8273 /* return corresponding task_group object of a cgroup */
8274 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8276 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8277 struct task_group, css);
8280 static struct cgroup_subsys_state *
8281 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8283 struct task_group *tg;
8285 if (!cgrp->parent) {
8286 /* This is early initialization for the top cgroup */
8287 init_task_group.css.cgroup = cgrp;
8288 return &init_task_group.css;
8291 /* we support only 1-level deep hierarchical scheduler atm */
8292 if (cgrp->parent->parent)
8293 return ERR_PTR(-EINVAL);
8295 tg = sched_create_group();
8297 return ERR_PTR(-ENOMEM);
8299 /* Bind the cgroup to task_group object we just created */
8300 tg->css.cgroup = cgrp;
8306 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8308 struct task_group *tg = cgroup_tg(cgrp);
8310 sched_destroy_group(tg);
8314 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8315 struct task_struct *tsk)
8317 #ifdef CONFIG_RT_GROUP_SCHED
8318 /* Don't accept realtime tasks when there is no way for them to run */
8319 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8322 /* We don't support RT-tasks being in separate groups */
8323 if (tsk->sched_class != &fair_sched_class)
8331 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8332 struct cgroup *old_cont, struct task_struct *tsk)
8334 sched_move_task(tsk);
8337 #ifdef CONFIG_FAIR_GROUP_SCHED
8338 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8341 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8344 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
8346 struct task_group *tg = cgroup_tg(cgrp);
8348 return (u64) tg->shares;
8352 #ifdef CONFIG_RT_GROUP_SCHED
8353 static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8355 const char __user *userbuf,
8356 size_t nbytes, loff_t *unused_ppos)
8365 if (nbytes >= sizeof(buffer))
8367 if (copy_from_user(buffer, userbuf, nbytes))
8370 buffer[nbytes] = 0; /* nul-terminate */
8372 /* strip newline if necessary */
8373 if (nbytes && (buffer[nbytes-1] == '\n'))
8374 buffer[nbytes-1] = 0;
8375 val = simple_strtoll(buffer, &end, 0);
8379 /* Pass to subsystem */
8380 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8386 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8388 char __user *buf, size_t nbytes,
8392 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8393 int len = sprintf(tmp, "%ld\n", val);
8395 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8398 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8401 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8404 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8406 return sched_group_rt_period(cgroup_tg(cgrp));
8410 static struct cftype cpu_files[] = {
8411 #ifdef CONFIG_FAIR_GROUP_SCHED
8414 .read_uint = cpu_shares_read_uint,
8415 .write_uint = cpu_shares_write_uint,
8418 #ifdef CONFIG_RT_GROUP_SCHED
8420 .name = "rt_runtime_us",
8421 .read = cpu_rt_runtime_read,
8422 .write = cpu_rt_runtime_write,
8425 .name = "rt_period_us",
8426 .read_uint = cpu_rt_period_read_uint,
8427 .write_uint = cpu_rt_period_write_uint,
8432 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8434 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8437 struct cgroup_subsys cpu_cgroup_subsys = {
8439 .create = cpu_cgroup_create,
8440 .destroy = cpu_cgroup_destroy,
8441 .can_attach = cpu_cgroup_can_attach,
8442 .attach = cpu_cgroup_attach,
8443 .populate = cpu_cgroup_populate,
8444 .subsys_id = cpu_cgroup_subsys_id,
8448 #endif /* CONFIG_CGROUP_SCHED */
8450 #ifdef CONFIG_CGROUP_CPUACCT
8453 * CPU accounting code for task groups.
8455 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8456 * (balbir@in.ibm.com).
8459 /* track cpu usage of a group of tasks */
8461 struct cgroup_subsys_state css;
8462 /* cpuusage holds pointer to a u64-type object on every cpu */
8466 struct cgroup_subsys cpuacct_subsys;
8468 /* return cpu accounting group corresponding to this container */
8469 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8471 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8472 struct cpuacct, css);
8475 /* return cpu accounting group to which this task belongs */
8476 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8478 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8479 struct cpuacct, css);
8482 /* create a new cpu accounting group */
8483 static struct cgroup_subsys_state *cpuacct_create(
8484 struct cgroup_subsys *ss, struct cgroup *cgrp)
8486 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8489 return ERR_PTR(-ENOMEM);
8491 ca->cpuusage = alloc_percpu(u64);
8492 if (!ca->cpuusage) {
8494 return ERR_PTR(-ENOMEM);
8500 /* destroy an existing cpu accounting group */
8502 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8504 struct cpuacct *ca = cgroup_ca(cgrp);
8506 free_percpu(ca->cpuusage);
8510 /* return total cpu usage (in nanoseconds) of a group */
8511 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8513 struct cpuacct *ca = cgroup_ca(cgrp);
8514 u64 totalcpuusage = 0;
8517 for_each_possible_cpu(i) {
8518 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8521 * Take rq->lock to make 64-bit addition safe on 32-bit
8524 spin_lock_irq(&cpu_rq(i)->lock);
8525 totalcpuusage += *cpuusage;
8526 spin_unlock_irq(&cpu_rq(i)->lock);
8529 return totalcpuusage;
8532 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8535 struct cpuacct *ca = cgroup_ca(cgrp);
8544 for_each_possible_cpu(i) {
8545 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8547 spin_lock_irq(&cpu_rq(i)->lock);
8549 spin_unlock_irq(&cpu_rq(i)->lock);
8555 static struct cftype files[] = {
8558 .read_uint = cpuusage_read,
8559 .write_uint = cpuusage_write,
8563 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8565 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8569 * charge this task's execution time to its accounting group.
8571 * called with rq->lock held.
8573 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8577 if (!cpuacct_subsys.active)
8582 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8584 *cpuusage += cputime;
8588 struct cgroup_subsys cpuacct_subsys = {
8590 .create = cpuacct_create,
8591 .destroy = cpuacct_destroy,
8592 .populate = cpuacct_populate,
8593 .subsys_id = cpuacct_subsys_id,
8595 #endif /* CONFIG_CGROUP_CPUACCT */