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;
275 struct task_group *parent;
276 struct list_head siblings;
277 struct list_head children;
280 #ifdef CONFIG_USER_SCHED
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
287 struct task_group root_task_group;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* Default task group's sched entity on each cpu */
291 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
292 /* Default task group's cfs_rq on each cpu */
293 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
296 #ifdef CONFIG_RT_GROUP_SCHED
297 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
298 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
301 #define root_task_group init_task_group
304 /* task_group_lock serializes add/remove of task groups and also changes to
305 * a task group's cpu shares.
307 static DEFINE_SPINLOCK(task_group_lock);
309 /* doms_cur_mutex serializes access to doms_cur[] array */
310 static DEFINE_MUTEX(doms_cur_mutex);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
313 #ifdef CONFIG_USER_SCHED
314 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
316 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
321 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
324 /* Default task group.
325 * Every task in system belong to this group at bootup.
327 struct task_group init_task_group;
329 /* return group to which a task belongs */
330 static inline struct task_group *task_group(struct task_struct *p)
332 struct task_group *tg;
334 #ifdef CONFIG_USER_SCHED
336 #elif defined(CONFIG_CGROUP_SCHED)
337 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
338 struct task_group, css);
340 tg = &init_task_group;
345 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
346 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
348 #ifdef CONFIG_FAIR_GROUP_SCHED
349 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
350 p->se.parent = task_group(p)->se[cpu];
353 #ifdef CONFIG_RT_GROUP_SCHED
354 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
355 p->rt.parent = task_group(p)->rt_se[cpu];
359 static inline void lock_doms_cur(void)
361 mutex_lock(&doms_cur_mutex);
364 static inline void unlock_doms_cur(void)
366 mutex_unlock(&doms_cur_mutex);
371 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
372 static inline void lock_doms_cur(void) { }
373 static inline void unlock_doms_cur(void) { }
375 #endif /* CONFIG_GROUP_SCHED */
377 /* CFS-related fields in a runqueue */
379 struct load_weight load;
380 unsigned long nr_running;
385 struct rb_root tasks_timeline;
386 struct rb_node *rb_leftmost;
388 struct list_head tasks;
389 struct list_head *balance_iterator;
392 * 'curr' points to currently running entity on this cfs_rq.
393 * It is set to NULL otherwise (i.e when none are currently running).
395 struct sched_entity *curr, *next;
397 unsigned long nr_spread_over;
399 #ifdef CONFIG_FAIR_GROUP_SCHED
400 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
403 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
404 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
405 * (like users, containers etc.)
407 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
408 * list is used during load balance.
410 struct list_head leaf_cfs_rq_list;
411 struct task_group *tg; /* group that "owns" this runqueue */
414 unsigned long task_weight;
415 unsigned long shares;
417 * We need space to build a sched_domain wide view of the full task
418 * group tree, in order to avoid depending on dynamic memory allocation
419 * during the load balancing we place this in the per cpu task group
420 * hierarchy. This limits the load balancing to one instance per cpu,
421 * but more should not be needed anyway.
423 struct aggregate_struct {
425 * load = weight(cpus) * f(tg)
427 * Where f(tg) is the recursive weight fraction assigned to
433 * part of the group weight distributed to this span.
435 unsigned long shares;
438 * The sum of all runqueue weights within this span.
440 unsigned long rq_weight;
443 * Weight contributed by tasks; this is the part we can
444 * influence by moving tasks around.
446 unsigned long task_weight;
452 /* Real-Time classes' related field in a runqueue: */
454 struct rt_prio_array active;
455 unsigned long rt_nr_running;
456 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int highest_prio; /* highest queued rt task prio */
460 unsigned long rt_nr_migratory;
466 /* Nests inside the rq lock: */
467 spinlock_t rt_runtime_lock;
469 #ifdef CONFIG_RT_GROUP_SCHED
470 unsigned long rt_nr_boosted;
473 struct list_head leaf_rt_rq_list;
474 struct task_group *tg;
475 struct sched_rt_entity *rt_se;
482 * We add the notion of a root-domain which will be used to define per-domain
483 * variables. Each exclusive cpuset essentially defines an island domain by
484 * fully partitioning the member cpus from any other cpuset. Whenever a new
485 * exclusive cpuset is created, we also create and attach a new root-domain
495 * The "RT overload" flag: it gets set if a CPU has more than
496 * one runnable RT task.
503 * By default the system creates a single root-domain with all cpus as
504 * members (mimicking the global state we have today).
506 static struct root_domain def_root_domain;
511 * This is the main, per-CPU runqueue data structure.
513 * Locking rule: those places that want to lock multiple runqueues
514 * (such as the load balancing or the thread migration code), lock
515 * acquire operations must be ordered by ascending &runqueue.
522 * nr_running and cpu_load should be in the same cacheline because
523 * remote CPUs use both these fields when doing load calculation.
525 unsigned long nr_running;
526 #define CPU_LOAD_IDX_MAX 5
527 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
528 unsigned char idle_at_tick;
530 unsigned long last_tick_seen;
531 unsigned char in_nohz_recently;
533 /* capture load from *all* tasks on this cpu: */
534 struct load_weight load;
535 unsigned long nr_load_updates;
541 #ifdef CONFIG_FAIR_GROUP_SCHED
542 /* list of leaf cfs_rq on this cpu: */
543 struct list_head leaf_cfs_rq_list;
545 #ifdef CONFIG_RT_GROUP_SCHED
546 struct list_head leaf_rt_rq_list;
550 * This is part of a global counter where only the total sum
551 * over all CPUs matters. A task can increase this counter on
552 * one CPU and if it got migrated afterwards it may decrease
553 * it on another CPU. Always updated under the runqueue lock:
555 unsigned long nr_uninterruptible;
557 struct task_struct *curr, *idle;
558 unsigned long next_balance;
559 struct mm_struct *prev_mm;
561 u64 clock, prev_clock_raw;
564 unsigned int clock_warps, clock_overflows, clock_underflows;
566 unsigned int clock_deep_idle_events;
572 struct root_domain *rd;
573 struct sched_domain *sd;
575 /* For active balancing */
578 /* cpu of this runqueue: */
581 struct task_struct *migration_thread;
582 struct list_head migration_queue;
585 #ifdef CONFIG_SCHED_HRTICK
586 unsigned long hrtick_flags;
587 ktime_t hrtick_expire;
588 struct hrtimer hrtick_timer;
591 #ifdef CONFIG_SCHEDSTATS
593 struct sched_info rq_sched_info;
595 /* sys_sched_yield() stats */
596 unsigned int yld_exp_empty;
597 unsigned int yld_act_empty;
598 unsigned int yld_both_empty;
599 unsigned int yld_count;
601 /* schedule() stats */
602 unsigned int sched_switch;
603 unsigned int sched_count;
604 unsigned int sched_goidle;
606 /* try_to_wake_up() stats */
607 unsigned int ttwu_count;
608 unsigned int ttwu_local;
611 unsigned int bkl_count;
613 struct lock_class_key rq_lock_key;
616 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
618 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
620 rq->curr->sched_class->check_preempt_curr(rq, p);
623 static inline int cpu_of(struct rq *rq)
633 static inline bool nohz_on(int cpu)
635 return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
638 static inline u64 max_skipped_ticks(struct rq *rq)
640 return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
643 static inline void update_last_tick_seen(struct rq *rq)
645 rq->last_tick_seen = jiffies;
648 static inline u64 max_skipped_ticks(struct rq *rq)
653 static inline void update_last_tick_seen(struct rq *rq)
659 * Update the per-runqueue clock, as finegrained as the platform can give
660 * us, but without assuming monotonicity, etc.:
662 static void __update_rq_clock(struct rq *rq)
664 u64 prev_raw = rq->prev_clock_raw;
665 u64 now = sched_clock();
666 s64 delta = now - prev_raw;
667 u64 clock = rq->clock;
669 #ifdef CONFIG_SCHED_DEBUG
670 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
673 * Protect against sched_clock() occasionally going backwards:
675 if (unlikely(delta < 0)) {
680 * Catch too large forward jumps too:
682 u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
683 u64 max_time = rq->tick_timestamp + max_jump;
685 if (unlikely(clock + delta > max_time)) {
686 if (clock < max_time)
690 rq->clock_overflows++;
692 if (unlikely(delta > rq->clock_max_delta))
693 rq->clock_max_delta = delta;
698 rq->prev_clock_raw = now;
702 static void update_rq_clock(struct rq *rq)
704 if (likely(smp_processor_id() == cpu_of(rq)))
705 __update_rq_clock(rq);
709 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
710 * See detach_destroy_domains: synchronize_sched for details.
712 * The domain tree of any CPU may only be accessed from within
713 * preempt-disabled sections.
715 #define for_each_domain(cpu, __sd) \
716 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
718 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
719 #define this_rq() (&__get_cpu_var(runqueues))
720 #define task_rq(p) cpu_rq(task_cpu(p))
721 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
724 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
726 #ifdef CONFIG_SCHED_DEBUG
727 # define const_debug __read_mostly
729 # define const_debug static const
733 * Debugging: various feature bits
736 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
737 SCHED_FEAT_WAKEUP_PREEMPT = 2,
738 SCHED_FEAT_START_DEBIT = 4,
739 SCHED_FEAT_AFFINE_WAKEUPS = 8,
740 SCHED_FEAT_CACHE_HOT_BUDDY = 16,
741 SCHED_FEAT_SYNC_WAKEUPS = 32,
742 SCHED_FEAT_HRTICK = 64,
743 SCHED_FEAT_DOUBLE_TICK = 128,
744 SCHED_FEAT_NORMALIZED_SLEEPER = 256,
747 const_debug unsigned int sysctl_sched_features =
748 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
749 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
750 SCHED_FEAT_START_DEBIT * 1 |
751 SCHED_FEAT_AFFINE_WAKEUPS * 1 |
752 SCHED_FEAT_CACHE_HOT_BUDDY * 1 |
753 SCHED_FEAT_SYNC_WAKEUPS * 1 |
754 SCHED_FEAT_HRTICK * 1 |
755 SCHED_FEAT_DOUBLE_TICK * 0 |
756 SCHED_FEAT_NORMALIZED_SLEEPER * 1;
758 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
761 * Number of tasks to iterate in a single balance run.
762 * Limited because this is done with IRQs disabled.
764 const_debug unsigned int sysctl_sched_nr_migrate = 32;
767 * period over which we measure -rt task cpu usage in us.
770 unsigned int sysctl_sched_rt_period = 1000000;
772 static __read_mostly int scheduler_running;
775 * part of the period that we allow rt tasks to run in us.
778 int sysctl_sched_rt_runtime = 950000;
780 static inline u64 global_rt_period(void)
782 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
785 static inline u64 global_rt_runtime(void)
787 if (sysctl_sched_rt_period < 0)
790 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
793 static const unsigned long long time_sync_thresh = 100000;
795 static DEFINE_PER_CPU(unsigned long long, time_offset);
796 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
799 * Global lock which we take every now and then to synchronize
800 * the CPUs time. This method is not warp-safe, but it's good
801 * enough to synchronize slowly diverging time sources and thus
802 * it's good enough for tracing:
804 static DEFINE_SPINLOCK(time_sync_lock);
805 static unsigned long long prev_global_time;
807 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
811 spin_lock_irqsave(&time_sync_lock, flags);
813 if (time < prev_global_time) {
814 per_cpu(time_offset, cpu) += prev_global_time - time;
815 time = prev_global_time;
817 prev_global_time = time;
820 spin_unlock_irqrestore(&time_sync_lock, flags);
825 static unsigned long long __cpu_clock(int cpu)
827 unsigned long long now;
832 * Only call sched_clock() if the scheduler has already been
833 * initialized (some code might call cpu_clock() very early):
835 if (unlikely(!scheduler_running))
838 local_irq_save(flags);
842 local_irq_restore(flags);
848 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
849 * clock constructed from sched_clock():
851 unsigned long long cpu_clock(int cpu)
853 unsigned long long prev_cpu_time, time, delta_time;
855 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
856 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
857 delta_time = time-prev_cpu_time;
859 if (unlikely(delta_time > time_sync_thresh))
860 time = __sync_cpu_clock(time, cpu);
864 EXPORT_SYMBOL_GPL(cpu_clock);
866 #ifndef prepare_arch_switch
867 # define prepare_arch_switch(next) do { } while (0)
869 #ifndef finish_arch_switch
870 # define finish_arch_switch(prev) do { } while (0)
873 static inline int task_current(struct rq *rq, struct task_struct *p)
875 return rq->curr == p;
878 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
879 static inline int task_running(struct rq *rq, struct task_struct *p)
881 return task_current(rq, p);
884 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
888 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
890 #ifdef CONFIG_DEBUG_SPINLOCK
891 /* this is a valid case when another task releases the spinlock */
892 rq->lock.owner = current;
895 * If we are tracking spinlock dependencies then we have to
896 * fix up the runqueue lock - which gets 'carried over' from
899 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
901 spin_unlock_irq(&rq->lock);
904 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
905 static inline int task_running(struct rq *rq, struct task_struct *p)
910 return task_current(rq, p);
914 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
918 * We can optimise this out completely for !SMP, because the
919 * SMP rebalancing from interrupt is the only thing that cares
924 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
925 spin_unlock_irq(&rq->lock);
927 spin_unlock(&rq->lock);
931 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
935 * After ->oncpu is cleared, the task can be moved to a different CPU.
936 * We must ensure this doesn't happen until the switch is completely
942 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
946 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
949 * __task_rq_lock - lock the runqueue a given task resides on.
950 * Must be called interrupts disabled.
952 static inline struct rq *__task_rq_lock(struct task_struct *p)
956 struct rq *rq = task_rq(p);
957 spin_lock(&rq->lock);
958 if (likely(rq == task_rq(p)))
960 spin_unlock(&rq->lock);
965 * task_rq_lock - lock the runqueue a given task resides on and disable
966 * interrupts. Note the ordering: we can safely lookup the task_rq without
967 * explicitly disabling preemption.
969 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
975 local_irq_save(*flags);
977 spin_lock(&rq->lock);
978 if (likely(rq == task_rq(p)))
980 spin_unlock_irqrestore(&rq->lock, *flags);
984 static void __task_rq_unlock(struct rq *rq)
987 spin_unlock(&rq->lock);
990 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
993 spin_unlock_irqrestore(&rq->lock, *flags);
997 * this_rq_lock - lock this runqueue and disable interrupts.
999 static struct rq *this_rq_lock(void)
1000 __acquires(rq->lock)
1004 local_irq_disable();
1006 spin_lock(&rq->lock);
1012 * We are going deep-idle (irqs are disabled):
1014 void sched_clock_idle_sleep_event(void)
1016 struct rq *rq = cpu_rq(smp_processor_id());
1018 spin_lock(&rq->lock);
1019 __update_rq_clock(rq);
1020 spin_unlock(&rq->lock);
1021 rq->clock_deep_idle_events++;
1023 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
1026 * We just idled delta nanoseconds (called with irqs disabled):
1028 void sched_clock_idle_wakeup_event(u64 delta_ns)
1030 struct rq *rq = cpu_rq(smp_processor_id());
1031 u64 now = sched_clock();
1033 rq->idle_clock += delta_ns;
1035 * Override the previous timestamp and ignore all
1036 * sched_clock() deltas that occured while we idled,
1037 * and use the PM-provided delta_ns to advance the
1040 spin_lock(&rq->lock);
1041 rq->prev_clock_raw = now;
1042 rq->clock += delta_ns;
1043 spin_unlock(&rq->lock);
1044 touch_softlockup_watchdog();
1046 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
1048 static void __resched_task(struct task_struct *p, int tif_bit);
1050 static inline void resched_task(struct task_struct *p)
1052 __resched_task(p, TIF_NEED_RESCHED);
1055 #ifdef CONFIG_SCHED_HRTICK
1057 * Use HR-timers to deliver accurate preemption points.
1059 * Its all a bit involved since we cannot program an hrt while holding the
1060 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1063 * When we get rescheduled we reprogram the hrtick_timer outside of the
1066 static inline void resched_hrt(struct task_struct *p)
1068 __resched_task(p, TIF_HRTICK_RESCHED);
1071 static inline void resched_rq(struct rq *rq)
1073 unsigned long flags;
1075 spin_lock_irqsave(&rq->lock, flags);
1076 resched_task(rq->curr);
1077 spin_unlock_irqrestore(&rq->lock, flags);
1081 HRTICK_SET, /* re-programm hrtick_timer */
1082 HRTICK_RESET, /* not a new slice */
1087 * - enabled by features
1088 * - hrtimer is actually high res
1090 static inline int hrtick_enabled(struct rq *rq)
1092 if (!sched_feat(HRTICK))
1094 return hrtimer_is_hres_active(&rq->hrtick_timer);
1098 * Called to set the hrtick timer state.
1100 * called with rq->lock held and irqs disabled
1102 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1104 assert_spin_locked(&rq->lock);
1107 * preempt at: now + delay
1110 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1112 * indicate we need to program the timer
1114 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1116 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1119 * New slices are called from the schedule path and don't need a
1120 * forced reschedule.
1123 resched_hrt(rq->curr);
1126 static void hrtick_clear(struct rq *rq)
1128 if (hrtimer_active(&rq->hrtick_timer))
1129 hrtimer_cancel(&rq->hrtick_timer);
1133 * Update the timer from the possible pending state.
1135 static void hrtick_set(struct rq *rq)
1139 unsigned long flags;
1141 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1143 spin_lock_irqsave(&rq->lock, flags);
1144 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1145 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1146 time = rq->hrtick_expire;
1147 clear_thread_flag(TIF_HRTICK_RESCHED);
1148 spin_unlock_irqrestore(&rq->lock, flags);
1151 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1152 if (reset && !hrtimer_active(&rq->hrtick_timer))
1159 * High-resolution timer tick.
1160 * Runs from hardirq context with interrupts disabled.
1162 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1164 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1166 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1168 spin_lock(&rq->lock);
1169 __update_rq_clock(rq);
1170 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1171 spin_unlock(&rq->lock);
1173 return HRTIMER_NORESTART;
1176 static inline void init_rq_hrtick(struct rq *rq)
1178 rq->hrtick_flags = 0;
1179 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1180 rq->hrtick_timer.function = hrtick;
1181 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1184 void hrtick_resched(void)
1187 unsigned long flags;
1189 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1192 local_irq_save(flags);
1193 rq = cpu_rq(smp_processor_id());
1195 local_irq_restore(flags);
1198 static inline void hrtick_clear(struct rq *rq)
1202 static inline void hrtick_set(struct rq *rq)
1206 static inline void init_rq_hrtick(struct rq *rq)
1210 void hrtick_resched(void)
1216 * resched_task - mark a task 'to be rescheduled now'.
1218 * On UP this means the setting of the need_resched flag, on SMP it
1219 * might also involve a cross-CPU call to trigger the scheduler on
1224 #ifndef tsk_is_polling
1225 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1228 static void __resched_task(struct task_struct *p, int tif_bit)
1232 assert_spin_locked(&task_rq(p)->lock);
1234 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1237 set_tsk_thread_flag(p, tif_bit);
1240 if (cpu == smp_processor_id())
1243 /* NEED_RESCHED must be visible before we test polling */
1245 if (!tsk_is_polling(p))
1246 smp_send_reschedule(cpu);
1249 static void resched_cpu(int cpu)
1251 struct rq *rq = cpu_rq(cpu);
1252 unsigned long flags;
1254 if (!spin_trylock_irqsave(&rq->lock, flags))
1256 resched_task(cpu_curr(cpu));
1257 spin_unlock_irqrestore(&rq->lock, flags);
1262 * When add_timer_on() enqueues a timer into the timer wheel of an
1263 * idle CPU then this timer might expire before the next timer event
1264 * which is scheduled to wake up that CPU. In case of a completely
1265 * idle system the next event might even be infinite time into the
1266 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1267 * leaves the inner idle loop so the newly added timer is taken into
1268 * account when the CPU goes back to idle and evaluates the timer
1269 * wheel for the next timer event.
1271 void wake_up_idle_cpu(int cpu)
1273 struct rq *rq = cpu_rq(cpu);
1275 if (cpu == smp_processor_id())
1279 * This is safe, as this function is called with the timer
1280 * wheel base lock of (cpu) held. When the CPU is on the way
1281 * to idle and has not yet set rq->curr to idle then it will
1282 * be serialized on the timer wheel base lock and take the new
1283 * timer into account automatically.
1285 if (rq->curr != rq->idle)
1289 * We can set TIF_RESCHED on the idle task of the other CPU
1290 * lockless. The worst case is that the other CPU runs the
1291 * idle task through an additional NOOP schedule()
1293 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1295 /* NEED_RESCHED must be visible before we test polling */
1297 if (!tsk_is_polling(rq->idle))
1298 smp_send_reschedule(cpu);
1303 static void __resched_task(struct task_struct *p, int tif_bit)
1305 assert_spin_locked(&task_rq(p)->lock);
1306 set_tsk_thread_flag(p, tif_bit);
1310 #if BITS_PER_LONG == 32
1311 # define WMULT_CONST (~0UL)
1313 # define WMULT_CONST (1UL << 32)
1316 #define WMULT_SHIFT 32
1319 * Shift right and round:
1321 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1323 static unsigned long
1324 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1325 struct load_weight *lw)
1329 if (unlikely(!lw->inv_weight))
1330 lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1332 tmp = (u64)delta_exec * weight;
1334 * Check whether we'd overflow the 64-bit multiplication:
1336 if (unlikely(tmp > WMULT_CONST))
1337 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1340 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1342 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1345 static inline unsigned long
1346 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1348 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1351 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1357 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1364 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1365 * of tasks with abnormal "nice" values across CPUs the contribution that
1366 * each task makes to its run queue's load is weighted according to its
1367 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1368 * scaled version of the new time slice allocation that they receive on time
1372 #define WEIGHT_IDLEPRIO 2
1373 #define WMULT_IDLEPRIO (1 << 31)
1376 * Nice levels are multiplicative, with a gentle 10% change for every
1377 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1378 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1379 * that remained on nice 0.
1381 * The "10% effect" is relative and cumulative: from _any_ nice level,
1382 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1383 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1384 * If a task goes up by ~10% and another task goes down by ~10% then
1385 * the relative distance between them is ~25%.)
1387 static const int prio_to_weight[40] = {
1388 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1389 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1390 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1391 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1392 /* 0 */ 1024, 820, 655, 526, 423,
1393 /* 5 */ 335, 272, 215, 172, 137,
1394 /* 10 */ 110, 87, 70, 56, 45,
1395 /* 15 */ 36, 29, 23, 18, 15,
1399 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1401 * In cases where the weight does not change often, we can use the
1402 * precalculated inverse to speed up arithmetics by turning divisions
1403 * into multiplications:
1405 static const u32 prio_to_wmult[40] = {
1406 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1407 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1408 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1409 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1410 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1411 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1412 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1413 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1416 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1419 * runqueue iterator, to support SMP load-balancing between different
1420 * scheduling classes, without having to expose their internal data
1421 * structures to the load-balancing proper:
1423 struct rq_iterator {
1425 struct task_struct *(*start)(void *);
1426 struct task_struct *(*next)(void *);
1430 static unsigned long
1431 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1432 unsigned long max_load_move, struct sched_domain *sd,
1433 enum cpu_idle_type idle, int *all_pinned,
1434 int *this_best_prio, struct rq_iterator *iterator);
1437 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1438 struct sched_domain *sd, enum cpu_idle_type idle,
1439 struct rq_iterator *iterator);
1442 #ifdef CONFIG_CGROUP_CPUACCT
1443 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1445 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1448 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1450 update_load_add(&rq->load, load);
1453 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1455 update_load_sub(&rq->load, load);
1459 static unsigned long source_load(int cpu, int type);
1460 static unsigned long target_load(int cpu, int type);
1461 static unsigned long cpu_avg_load_per_task(int cpu);
1462 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1464 #ifdef CONFIG_FAIR_GROUP_SCHED
1467 * Group load balancing.
1469 * We calculate a few balance domain wide aggregate numbers; load and weight.
1470 * Given the pictures below, and assuming each item has equal weight:
1481 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1482 * which equals 1/9-th of the total load.
1485 * The weight of this group on the selected cpus.
1488 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1492 * Part of the rq_weight contributed by tasks; all groups except B would
1496 static inline struct aggregate_struct *
1497 aggregate(struct task_group *tg, struct sched_domain *sd)
1499 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1502 typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1505 * Iterate the full tree, calling @down when first entering a node and @up when
1506 * leaving it for the final time.
1509 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1510 struct sched_domain *sd)
1512 struct task_group *parent, *child;
1515 parent = &root_task_group;
1517 (*down)(parent, sd);
1518 list_for_each_entry_rcu(child, &parent->children, siblings) {
1528 parent = parent->parent;
1535 * Calculate the aggregate runqueue weight.
1538 void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1540 unsigned long rq_weight = 0;
1541 unsigned long task_weight = 0;
1544 for_each_cpu_mask(i, sd->span) {
1545 rq_weight += tg->cfs_rq[i]->load.weight;
1546 task_weight += tg->cfs_rq[i]->task_weight;
1549 aggregate(tg, sd)->rq_weight = rq_weight;
1550 aggregate(tg, sd)->task_weight = task_weight;
1554 * Redistribute tg->shares amongst all tg->cfs_rq[]s.
1556 static void __aggregate_redistribute_shares(struct task_group *tg)
1558 int i, max_cpu = smp_processor_id();
1559 unsigned long rq_weight = 0;
1560 unsigned long shares, max_shares = 0, shares_rem = tg->shares;
1562 for_each_possible_cpu(i)
1563 rq_weight += tg->cfs_rq[i]->load.weight;
1565 for_each_possible_cpu(i) {
1567 * divide shares proportional to the rq_weights.
1569 shares = tg->shares * tg->cfs_rq[i]->load.weight;
1570 shares /= rq_weight + 1;
1572 tg->cfs_rq[i]->shares = shares;
1574 if (shares > max_shares) {
1575 max_shares = shares;
1578 shares_rem -= shares;
1582 * Ensure it all adds up to tg->shares; we can loose a few
1583 * due to rounding down when computing the per-cpu shares.
1586 tg->cfs_rq[max_cpu]->shares += shares_rem;
1590 * Compute the weight of this group on the given cpus.
1593 void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1595 unsigned long shares = 0;
1599 for_each_cpu_mask(i, sd->span)
1600 shares += tg->cfs_rq[i]->shares;
1603 * When the span doesn't have any shares assigned, but does have
1604 * tasks to run do a machine wide rebalance (should be rare).
1606 if (unlikely(!shares && aggregate(tg, sd)->rq_weight)) {
1607 __aggregate_redistribute_shares(tg);
1611 aggregate(tg, sd)->shares = shares;
1615 * Compute the load fraction assigned to this group, relies on the aggregate
1616 * weight and this group's parent's load, i.e. top-down.
1619 void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1627 for_each_cpu_mask(i, sd->span)
1628 load += cpu_rq(i)->load.weight;
1631 load = aggregate(tg->parent, sd)->load;
1634 * shares is our weight in the parent's rq so
1635 * shares/parent->rq_weight gives our fraction of the load
1637 load *= aggregate(tg, sd)->shares;
1638 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1641 aggregate(tg, sd)->load = load;
1644 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1647 * Calculate and set the cpu's group shares.
1650 __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1654 unsigned long shares;
1655 unsigned long rq_weight;
1660 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1663 * If there are currently no tasks on the cpu pretend there is one of
1664 * average load so that when a new task gets to run here it will not
1665 * get delayed by group starvation.
1669 rq_weight = NICE_0_LOAD;
1673 * \Sum shares * rq_weight
1674 * shares = -----------------------
1678 shares = aggregate(tg, sd)->shares * rq_weight;
1679 shares /= aggregate(tg, sd)->rq_weight + 1;
1682 * record the actual number of shares, not the boosted amount.
1684 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1686 if (shares < MIN_SHARES)
1687 shares = MIN_SHARES;
1689 __set_se_shares(tg->se[tcpu], shares);
1693 * Re-adjust the weights on the cpu the task came from and on the cpu the
1697 __move_group_shares(struct task_group *tg, struct sched_domain *sd,
1700 unsigned long shares;
1702 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1704 __update_group_shares_cpu(tg, sd, scpu);
1705 __update_group_shares_cpu(tg, sd, dcpu);
1708 * ensure we never loose shares due to rounding errors in the
1709 * above redistribution.
1711 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1713 tg->cfs_rq[dcpu]->shares += shares;
1717 * Because changing a group's shares changes the weight of the super-group
1718 * we need to walk up the tree and change all shares until we hit the root.
1721 move_group_shares(struct task_group *tg, struct sched_domain *sd,
1725 __move_group_shares(tg, sd, scpu, dcpu);
1731 void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1733 unsigned long shares = aggregate(tg, sd)->shares;
1736 for_each_cpu_mask(i, sd->span) {
1737 struct rq *rq = cpu_rq(i);
1738 unsigned long flags;
1740 spin_lock_irqsave(&rq->lock, flags);
1741 __update_group_shares_cpu(tg, sd, i);
1742 spin_unlock_irqrestore(&rq->lock, flags);
1745 aggregate_group_shares(tg, sd);
1748 * ensure we never loose shares due to rounding errors in the
1749 * above redistribution.
1751 shares -= aggregate(tg, sd)->shares;
1753 tg->cfs_rq[sd->first_cpu]->shares += shares;
1754 aggregate(tg, sd)->shares += shares;
1759 * Calculate the accumulative weight and recursive load of each task group
1760 * while walking down the tree.
1763 void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1765 aggregate_group_weight(tg, sd);
1766 aggregate_group_shares(tg, sd);
1767 aggregate_group_load(tg, sd);
1771 * Rebalance the cpu shares while walking back up the tree.
1774 void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1776 aggregate_group_set_shares(tg, sd);
1779 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1781 static void __init init_aggregate(void)
1785 for_each_possible_cpu(i)
1786 spin_lock_init(&per_cpu(aggregate_lock, i));
1789 static int get_aggregate(struct sched_domain *sd)
1791 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1794 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1798 static void put_aggregate(struct sched_domain *sd)
1800 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1803 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1805 cfs_rq->shares = shares;
1810 static inline void init_aggregate(void)
1814 static inline int get_aggregate(struct sched_domain *sd)
1819 static inline void put_aggregate(struct sched_domain *sd)
1824 #else /* CONFIG_SMP */
1826 #ifdef CONFIG_FAIR_GROUP_SCHED
1827 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1832 #endif /* CONFIG_SMP */
1834 #include "sched_stats.h"
1835 #include "sched_idletask.c"
1836 #include "sched_fair.c"
1837 #include "sched_rt.c"
1838 #ifdef CONFIG_SCHED_DEBUG
1839 # include "sched_debug.c"
1842 #define sched_class_highest (&rt_sched_class)
1844 static void inc_nr_running(struct rq *rq)
1849 static void dec_nr_running(struct rq *rq)
1854 static void set_load_weight(struct task_struct *p)
1856 if (task_has_rt_policy(p)) {
1857 p->se.load.weight = prio_to_weight[0] * 2;
1858 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1863 * SCHED_IDLE tasks get minimal weight:
1865 if (p->policy == SCHED_IDLE) {
1866 p->se.load.weight = WEIGHT_IDLEPRIO;
1867 p->se.load.inv_weight = WMULT_IDLEPRIO;
1871 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1872 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1875 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1877 sched_info_queued(p);
1878 p->sched_class->enqueue_task(rq, p, wakeup);
1882 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1884 p->sched_class->dequeue_task(rq, p, sleep);
1889 * __normal_prio - return the priority that is based on the static prio
1891 static inline int __normal_prio(struct task_struct *p)
1893 return p->static_prio;
1897 * Calculate the expected normal priority: i.e. priority
1898 * without taking RT-inheritance into account. Might be
1899 * boosted by interactivity modifiers. Changes upon fork,
1900 * setprio syscalls, and whenever the interactivity
1901 * estimator recalculates.
1903 static inline int normal_prio(struct task_struct *p)
1907 if (task_has_rt_policy(p))
1908 prio = MAX_RT_PRIO-1 - p->rt_priority;
1910 prio = __normal_prio(p);
1915 * Calculate the current priority, i.e. the priority
1916 * taken into account by the scheduler. This value might
1917 * be boosted by RT tasks, or might be boosted by
1918 * interactivity modifiers. Will be RT if the task got
1919 * RT-boosted. If not then it returns p->normal_prio.
1921 static int effective_prio(struct task_struct *p)
1923 p->normal_prio = normal_prio(p);
1925 * If we are RT tasks or we were boosted to RT priority,
1926 * keep the priority unchanged. Otherwise, update priority
1927 * to the normal priority:
1929 if (!rt_prio(p->prio))
1930 return p->normal_prio;
1935 * activate_task - move a task to the runqueue.
1937 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1939 if (task_contributes_to_load(p))
1940 rq->nr_uninterruptible--;
1942 enqueue_task(rq, p, wakeup);
1947 * deactivate_task - remove a task from the runqueue.
1949 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1951 if (task_contributes_to_load(p))
1952 rq->nr_uninterruptible++;
1954 dequeue_task(rq, p, sleep);
1959 * task_curr - is this task currently executing on a CPU?
1960 * @p: the task in question.
1962 inline int task_curr(const struct task_struct *p)
1964 return cpu_curr(task_cpu(p)) == p;
1967 /* Used instead of source_load when we know the type == 0 */
1968 unsigned long weighted_cpuload(const int cpu)
1970 return cpu_rq(cpu)->load.weight;
1973 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1975 set_task_rq(p, cpu);
1978 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1979 * successfuly executed on another CPU. We must ensure that updates of
1980 * per-task data have been completed by this moment.
1983 task_thread_info(p)->cpu = cpu;
1987 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1988 const struct sched_class *prev_class,
1989 int oldprio, int running)
1991 if (prev_class != p->sched_class) {
1992 if (prev_class->switched_from)
1993 prev_class->switched_from(rq, p, running);
1994 p->sched_class->switched_to(rq, p, running);
1996 p->sched_class->prio_changed(rq, p, oldprio, running);
2002 * Is this task likely cache-hot:
2005 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2010 * Buddy candidates are cache hot:
2012 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
2015 if (p->sched_class != &fair_sched_class)
2018 if (sysctl_sched_migration_cost == -1)
2020 if (sysctl_sched_migration_cost == 0)
2023 delta = now - p->se.exec_start;
2025 return delta < (s64)sysctl_sched_migration_cost;
2029 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2031 int old_cpu = task_cpu(p);
2032 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2033 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2034 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2037 clock_offset = old_rq->clock - new_rq->clock;
2039 #ifdef CONFIG_SCHEDSTATS
2040 if (p->se.wait_start)
2041 p->se.wait_start -= clock_offset;
2042 if (p->se.sleep_start)
2043 p->se.sleep_start -= clock_offset;
2044 if (p->se.block_start)
2045 p->se.block_start -= clock_offset;
2046 if (old_cpu != new_cpu) {
2047 schedstat_inc(p, se.nr_migrations);
2048 if (task_hot(p, old_rq->clock, NULL))
2049 schedstat_inc(p, se.nr_forced2_migrations);
2052 p->se.vruntime -= old_cfsrq->min_vruntime -
2053 new_cfsrq->min_vruntime;
2055 __set_task_cpu(p, new_cpu);
2058 struct migration_req {
2059 struct list_head list;
2061 struct task_struct *task;
2064 struct completion done;
2068 * The task's runqueue lock must be held.
2069 * Returns true if you have to wait for migration thread.
2072 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2074 struct rq *rq = task_rq(p);
2077 * If the task is not on a runqueue (and not running), then
2078 * it is sufficient to simply update the task's cpu field.
2080 if (!p->se.on_rq && !task_running(rq, p)) {
2081 set_task_cpu(p, dest_cpu);
2085 init_completion(&req->done);
2087 req->dest_cpu = dest_cpu;
2088 list_add(&req->list, &rq->migration_queue);
2094 * wait_task_inactive - wait for a thread to unschedule.
2096 * The caller must ensure that the task *will* unschedule sometime soon,
2097 * else this function might spin for a *long* time. This function can't
2098 * be called with interrupts off, or it may introduce deadlock with
2099 * smp_call_function() if an IPI is sent by the same process we are
2100 * waiting to become inactive.
2102 void wait_task_inactive(struct task_struct *p)
2104 unsigned long flags;
2110 * We do the initial early heuristics without holding
2111 * any task-queue locks at all. We'll only try to get
2112 * the runqueue lock when things look like they will
2118 * If the task is actively running on another CPU
2119 * still, just relax and busy-wait without holding
2122 * NOTE! Since we don't hold any locks, it's not
2123 * even sure that "rq" stays as the right runqueue!
2124 * But we don't care, since "task_running()" will
2125 * return false if the runqueue has changed and p
2126 * is actually now running somewhere else!
2128 while (task_running(rq, p))
2132 * Ok, time to look more closely! We need the rq
2133 * lock now, to be *sure*. If we're wrong, we'll
2134 * just go back and repeat.
2136 rq = task_rq_lock(p, &flags);
2137 running = task_running(rq, p);
2138 on_rq = p->se.on_rq;
2139 task_rq_unlock(rq, &flags);
2142 * Was it really running after all now that we
2143 * checked with the proper locks actually held?
2145 * Oops. Go back and try again..
2147 if (unlikely(running)) {
2153 * It's not enough that it's not actively running,
2154 * it must be off the runqueue _entirely_, and not
2157 * So if it wa still runnable (but just not actively
2158 * running right now), it's preempted, and we should
2159 * yield - it could be a while.
2161 if (unlikely(on_rq)) {
2162 schedule_timeout_uninterruptible(1);
2167 * Ahh, all good. It wasn't running, and it wasn't
2168 * runnable, which means that it will never become
2169 * running in the future either. We're all done!
2176 * kick_process - kick a running thread to enter/exit the kernel
2177 * @p: the to-be-kicked thread
2179 * Cause a process which is running on another CPU to enter
2180 * kernel-mode, without any delay. (to get signals handled.)
2182 * NOTE: this function doesnt have to take the runqueue lock,
2183 * because all it wants to ensure is that the remote task enters
2184 * the kernel. If the IPI races and the task has been migrated
2185 * to another CPU then no harm is done and the purpose has been
2188 void kick_process(struct task_struct *p)
2194 if ((cpu != smp_processor_id()) && task_curr(p))
2195 smp_send_reschedule(cpu);
2200 * Return a low guess at the load of a migration-source cpu weighted
2201 * according to the scheduling class and "nice" value.
2203 * We want to under-estimate the load of migration sources, to
2204 * balance conservatively.
2206 static unsigned long source_load(int cpu, int type)
2208 struct rq *rq = cpu_rq(cpu);
2209 unsigned long total = weighted_cpuload(cpu);
2214 return min(rq->cpu_load[type-1], total);
2218 * Return a high guess at the load of a migration-target cpu weighted
2219 * according to the scheduling class and "nice" value.
2221 static unsigned long target_load(int cpu, int type)
2223 struct rq *rq = cpu_rq(cpu);
2224 unsigned long total = weighted_cpuload(cpu);
2229 return max(rq->cpu_load[type-1], total);
2233 * Return the average load per task on the cpu's run queue
2235 static unsigned long cpu_avg_load_per_task(int cpu)
2237 struct rq *rq = cpu_rq(cpu);
2238 unsigned long total = weighted_cpuload(cpu);
2239 unsigned long n = rq->nr_running;
2241 return n ? total / n : SCHED_LOAD_SCALE;
2245 * find_idlest_group finds and returns the least busy CPU group within the
2248 static struct sched_group *
2249 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2251 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2252 unsigned long min_load = ULONG_MAX, this_load = 0;
2253 int load_idx = sd->forkexec_idx;
2254 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2257 unsigned long load, avg_load;
2261 /* Skip over this group if it has no CPUs allowed */
2262 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2265 local_group = cpu_isset(this_cpu, group->cpumask);
2267 /* Tally up the load of all CPUs in the group */
2270 for_each_cpu_mask(i, group->cpumask) {
2271 /* Bias balancing toward cpus of our domain */
2273 load = source_load(i, load_idx);
2275 load = target_load(i, load_idx);
2280 /* Adjust by relative CPU power of the group */
2281 avg_load = sg_div_cpu_power(group,
2282 avg_load * SCHED_LOAD_SCALE);
2285 this_load = avg_load;
2287 } else if (avg_load < min_load) {
2288 min_load = avg_load;
2291 } while (group = group->next, group != sd->groups);
2293 if (!idlest || 100*this_load < imbalance*min_load)
2299 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2302 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2305 unsigned long load, min_load = ULONG_MAX;
2309 /* Traverse only the allowed CPUs */
2310 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2312 for_each_cpu_mask(i, *tmp) {
2313 load = weighted_cpuload(i);
2315 if (load < min_load || (load == min_load && i == this_cpu)) {
2325 * sched_balance_self: balance the current task (running on cpu) in domains
2326 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2329 * Balance, ie. select the least loaded group.
2331 * Returns the target CPU number, or the same CPU if no balancing is needed.
2333 * preempt must be disabled.
2335 static int sched_balance_self(int cpu, int flag)
2337 struct task_struct *t = current;
2338 struct sched_domain *tmp, *sd = NULL;
2340 for_each_domain(cpu, tmp) {
2342 * If power savings logic is enabled for a domain, stop there.
2344 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2346 if (tmp->flags & flag)
2351 cpumask_t span, tmpmask;
2352 struct sched_group *group;
2353 int new_cpu, weight;
2355 if (!(sd->flags & flag)) {
2361 group = find_idlest_group(sd, t, cpu);
2367 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2368 if (new_cpu == -1 || new_cpu == cpu) {
2369 /* Now try balancing at a lower domain level of cpu */
2374 /* Now try balancing at a lower domain level of new_cpu */
2377 weight = cpus_weight(span);
2378 for_each_domain(cpu, tmp) {
2379 if (weight <= cpus_weight(tmp->span))
2381 if (tmp->flags & flag)
2384 /* while loop will break here if sd == NULL */
2390 #endif /* CONFIG_SMP */
2393 * try_to_wake_up - wake up a thread
2394 * @p: the to-be-woken-up thread
2395 * @state: the mask of task states that can be woken
2396 * @sync: do a synchronous wakeup?
2398 * Put it on the run-queue if it's not already there. The "current"
2399 * thread is always on the run-queue (except when the actual
2400 * re-schedule is in progress), and as such you're allowed to do
2401 * the simpler "current->state = TASK_RUNNING" to mark yourself
2402 * runnable without the overhead of this.
2404 * returns failure only if the task is already active.
2406 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2408 int cpu, orig_cpu, this_cpu, success = 0;
2409 unsigned long flags;
2413 if (!sched_feat(SYNC_WAKEUPS))
2417 rq = task_rq_lock(p, &flags);
2418 old_state = p->state;
2419 if (!(old_state & state))
2427 this_cpu = smp_processor_id();
2430 if (unlikely(task_running(rq, p)))
2433 cpu = p->sched_class->select_task_rq(p, sync);
2434 if (cpu != orig_cpu) {
2435 set_task_cpu(p, cpu);
2436 task_rq_unlock(rq, &flags);
2437 /* might preempt at this point */
2438 rq = task_rq_lock(p, &flags);
2439 old_state = p->state;
2440 if (!(old_state & state))
2445 this_cpu = smp_processor_id();
2449 #ifdef CONFIG_SCHEDSTATS
2450 schedstat_inc(rq, ttwu_count);
2451 if (cpu == this_cpu)
2452 schedstat_inc(rq, ttwu_local);
2454 struct sched_domain *sd;
2455 for_each_domain(this_cpu, sd) {
2456 if (cpu_isset(cpu, sd->span)) {
2457 schedstat_inc(sd, ttwu_wake_remote);
2465 #endif /* CONFIG_SMP */
2466 schedstat_inc(p, se.nr_wakeups);
2468 schedstat_inc(p, se.nr_wakeups_sync);
2469 if (orig_cpu != cpu)
2470 schedstat_inc(p, se.nr_wakeups_migrate);
2471 if (cpu == this_cpu)
2472 schedstat_inc(p, se.nr_wakeups_local);
2474 schedstat_inc(p, se.nr_wakeups_remote);
2475 update_rq_clock(rq);
2476 activate_task(rq, p, 1);
2480 check_preempt_curr(rq, p);
2482 p->state = TASK_RUNNING;
2484 if (p->sched_class->task_wake_up)
2485 p->sched_class->task_wake_up(rq, p);
2488 task_rq_unlock(rq, &flags);
2493 int wake_up_process(struct task_struct *p)
2495 return try_to_wake_up(p, TASK_ALL, 0);
2497 EXPORT_SYMBOL(wake_up_process);
2499 int wake_up_state(struct task_struct *p, unsigned int state)
2501 return try_to_wake_up(p, state, 0);
2505 * Perform scheduler related setup for a newly forked process p.
2506 * p is forked by current.
2508 * __sched_fork() is basic setup used by init_idle() too:
2510 static void __sched_fork(struct task_struct *p)
2512 p->se.exec_start = 0;
2513 p->se.sum_exec_runtime = 0;
2514 p->se.prev_sum_exec_runtime = 0;
2515 p->se.last_wakeup = 0;
2516 p->se.avg_overlap = 0;
2518 #ifdef CONFIG_SCHEDSTATS
2519 p->se.wait_start = 0;
2520 p->se.sum_sleep_runtime = 0;
2521 p->se.sleep_start = 0;
2522 p->se.block_start = 0;
2523 p->se.sleep_max = 0;
2524 p->se.block_max = 0;
2526 p->se.slice_max = 0;
2530 INIT_LIST_HEAD(&p->rt.run_list);
2532 INIT_LIST_HEAD(&p->se.group_node);
2534 #ifdef CONFIG_PREEMPT_NOTIFIERS
2535 INIT_HLIST_HEAD(&p->preempt_notifiers);
2539 * We mark the process as running here, but have not actually
2540 * inserted it onto the runqueue yet. This guarantees that
2541 * nobody will actually run it, and a signal or other external
2542 * event cannot wake it up and insert it on the runqueue either.
2544 p->state = TASK_RUNNING;
2548 * fork()/clone()-time setup:
2550 void sched_fork(struct task_struct *p, int clone_flags)
2552 int cpu = get_cpu();
2557 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2559 set_task_cpu(p, cpu);
2562 * Make sure we do not leak PI boosting priority to the child:
2564 p->prio = current->normal_prio;
2565 if (!rt_prio(p->prio))
2566 p->sched_class = &fair_sched_class;
2568 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2569 if (likely(sched_info_on()))
2570 memset(&p->sched_info, 0, sizeof(p->sched_info));
2572 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2575 #ifdef CONFIG_PREEMPT
2576 /* Want to start with kernel preemption disabled. */
2577 task_thread_info(p)->preempt_count = 1;
2583 * wake_up_new_task - wake up a newly created task for the first time.
2585 * This function will do some initial scheduler statistics housekeeping
2586 * that must be done for every newly created context, then puts the task
2587 * on the runqueue and wakes it.
2589 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2591 unsigned long flags;
2594 rq = task_rq_lock(p, &flags);
2595 BUG_ON(p->state != TASK_RUNNING);
2596 update_rq_clock(rq);
2598 p->prio = effective_prio(p);
2600 if (!p->sched_class->task_new || !current->se.on_rq) {
2601 activate_task(rq, p, 0);
2604 * Let the scheduling class do new task startup
2605 * management (if any):
2607 p->sched_class->task_new(rq, p);
2610 check_preempt_curr(rq, p);
2612 if (p->sched_class->task_wake_up)
2613 p->sched_class->task_wake_up(rq, p);
2615 task_rq_unlock(rq, &flags);
2618 #ifdef CONFIG_PREEMPT_NOTIFIERS
2621 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2622 * @notifier: notifier struct to register
2624 void preempt_notifier_register(struct preempt_notifier *notifier)
2626 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2628 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2631 * preempt_notifier_unregister - no longer interested in preemption notifications
2632 * @notifier: notifier struct to unregister
2634 * This is safe to call from within a preemption notifier.
2636 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2638 hlist_del(¬ifier->link);
2640 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2642 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2644 struct preempt_notifier *notifier;
2645 struct hlist_node *node;
2647 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2648 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2652 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2653 struct task_struct *next)
2655 struct preempt_notifier *notifier;
2656 struct hlist_node *node;
2658 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2659 notifier->ops->sched_out(notifier, next);
2664 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2669 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2670 struct task_struct *next)
2677 * prepare_task_switch - prepare to switch tasks
2678 * @rq: the runqueue preparing to switch
2679 * @prev: the current task that is being switched out
2680 * @next: the task we are going to switch to.
2682 * This is called with the rq lock held and interrupts off. It must
2683 * be paired with a subsequent finish_task_switch after the context
2686 * prepare_task_switch sets up locking and calls architecture specific
2690 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2691 struct task_struct *next)
2693 fire_sched_out_preempt_notifiers(prev, next);
2694 prepare_lock_switch(rq, next);
2695 prepare_arch_switch(next);
2699 * finish_task_switch - clean up after a task-switch
2700 * @rq: runqueue associated with task-switch
2701 * @prev: the thread we just switched away from.
2703 * finish_task_switch must be called after the context switch, paired
2704 * with a prepare_task_switch call before the context switch.
2705 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2706 * and do any other architecture-specific cleanup actions.
2708 * Note that we may have delayed dropping an mm in context_switch(). If
2709 * so, we finish that here outside of the runqueue lock. (Doing it
2710 * with the lock held can cause deadlocks; see schedule() for
2713 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2714 __releases(rq->lock)
2716 struct mm_struct *mm = rq->prev_mm;
2722 * A task struct has one reference for the use as "current".
2723 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2724 * schedule one last time. The schedule call will never return, and
2725 * the scheduled task must drop that reference.
2726 * The test for TASK_DEAD must occur while the runqueue locks are
2727 * still held, otherwise prev could be scheduled on another cpu, die
2728 * there before we look at prev->state, and then the reference would
2730 * Manfred Spraul <manfred@colorfullife.com>
2732 prev_state = prev->state;
2733 finish_arch_switch(prev);
2734 finish_lock_switch(rq, prev);
2736 if (current->sched_class->post_schedule)
2737 current->sched_class->post_schedule(rq);
2740 fire_sched_in_preempt_notifiers(current);
2743 if (unlikely(prev_state == TASK_DEAD)) {
2745 * Remove function-return probe instances associated with this
2746 * task and put them back on the free list.
2748 kprobe_flush_task(prev);
2749 put_task_struct(prev);
2754 * schedule_tail - first thing a freshly forked thread must call.
2755 * @prev: the thread we just switched away from.
2757 asmlinkage void schedule_tail(struct task_struct *prev)
2758 __releases(rq->lock)
2760 struct rq *rq = this_rq();
2762 finish_task_switch(rq, prev);
2763 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2764 /* In this case, finish_task_switch does not reenable preemption */
2767 if (current->set_child_tid)
2768 put_user(task_pid_vnr(current), current->set_child_tid);
2772 * context_switch - switch to the new MM and the new
2773 * thread's register state.
2776 context_switch(struct rq *rq, struct task_struct *prev,
2777 struct task_struct *next)
2779 struct mm_struct *mm, *oldmm;
2781 prepare_task_switch(rq, prev, next);
2783 oldmm = prev->active_mm;
2785 * For paravirt, this is coupled with an exit in switch_to to
2786 * combine the page table reload and the switch backend into
2789 arch_enter_lazy_cpu_mode();
2791 if (unlikely(!mm)) {
2792 next->active_mm = oldmm;
2793 atomic_inc(&oldmm->mm_count);
2794 enter_lazy_tlb(oldmm, next);
2796 switch_mm(oldmm, mm, next);
2798 if (unlikely(!prev->mm)) {
2799 prev->active_mm = NULL;
2800 rq->prev_mm = oldmm;
2803 * Since the runqueue lock will be released by the next
2804 * task (which is an invalid locking op but in the case
2805 * of the scheduler it's an obvious special-case), so we
2806 * do an early lockdep release here:
2808 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2809 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2812 /* Here we just switch the register state and the stack. */
2813 switch_to(prev, next, prev);
2817 * this_rq must be evaluated again because prev may have moved
2818 * CPUs since it called schedule(), thus the 'rq' on its stack
2819 * frame will be invalid.
2821 finish_task_switch(this_rq(), prev);
2825 * nr_running, nr_uninterruptible and nr_context_switches:
2827 * externally visible scheduler statistics: current number of runnable
2828 * threads, current number of uninterruptible-sleeping threads, total
2829 * number of context switches performed since bootup.
2831 unsigned long nr_running(void)
2833 unsigned long i, sum = 0;
2835 for_each_online_cpu(i)
2836 sum += cpu_rq(i)->nr_running;
2841 unsigned long nr_uninterruptible(void)
2843 unsigned long i, sum = 0;
2845 for_each_possible_cpu(i)
2846 sum += cpu_rq(i)->nr_uninterruptible;
2849 * Since we read the counters lockless, it might be slightly
2850 * inaccurate. Do not allow it to go below zero though:
2852 if (unlikely((long)sum < 0))
2858 unsigned long long nr_context_switches(void)
2861 unsigned long long sum = 0;
2863 for_each_possible_cpu(i)
2864 sum += cpu_rq(i)->nr_switches;
2869 unsigned long nr_iowait(void)
2871 unsigned long i, sum = 0;
2873 for_each_possible_cpu(i)
2874 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2879 unsigned long nr_active(void)
2881 unsigned long i, running = 0, uninterruptible = 0;
2883 for_each_online_cpu(i) {
2884 running += cpu_rq(i)->nr_running;
2885 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2888 if (unlikely((long)uninterruptible < 0))
2889 uninterruptible = 0;
2891 return running + uninterruptible;
2895 * Update rq->cpu_load[] statistics. This function is usually called every
2896 * scheduler tick (TICK_NSEC).
2898 static void update_cpu_load(struct rq *this_rq)
2900 unsigned long this_load = this_rq->load.weight;
2903 this_rq->nr_load_updates++;
2905 /* Update our load: */
2906 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2907 unsigned long old_load, new_load;
2909 /* scale is effectively 1 << i now, and >> i divides by scale */
2911 old_load = this_rq->cpu_load[i];
2912 new_load = this_load;
2914 * Round up the averaging division if load is increasing. This
2915 * prevents us from getting stuck on 9 if the load is 10, for
2918 if (new_load > old_load)
2919 new_load += scale-1;
2920 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2927 * double_rq_lock - safely lock two runqueues
2929 * Note this does not disable interrupts like task_rq_lock,
2930 * you need to do so manually before calling.
2932 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2933 __acquires(rq1->lock)
2934 __acquires(rq2->lock)
2936 BUG_ON(!irqs_disabled());
2938 spin_lock(&rq1->lock);
2939 __acquire(rq2->lock); /* Fake it out ;) */
2942 spin_lock(&rq1->lock);
2943 spin_lock(&rq2->lock);
2945 spin_lock(&rq2->lock);
2946 spin_lock(&rq1->lock);
2949 update_rq_clock(rq1);
2950 update_rq_clock(rq2);
2954 * double_rq_unlock - safely unlock two runqueues
2956 * Note this does not restore interrupts like task_rq_unlock,
2957 * you need to do so manually after calling.
2959 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2960 __releases(rq1->lock)
2961 __releases(rq2->lock)
2963 spin_unlock(&rq1->lock);
2965 spin_unlock(&rq2->lock);
2967 __release(rq2->lock);
2971 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2973 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2974 __releases(this_rq->lock)
2975 __acquires(busiest->lock)
2976 __acquires(this_rq->lock)
2980 if (unlikely(!irqs_disabled())) {
2981 /* printk() doesn't work good under rq->lock */
2982 spin_unlock(&this_rq->lock);
2985 if (unlikely(!spin_trylock(&busiest->lock))) {
2986 if (busiest < this_rq) {
2987 spin_unlock(&this_rq->lock);
2988 spin_lock(&busiest->lock);
2989 spin_lock(&this_rq->lock);
2992 spin_lock(&busiest->lock);
2998 * If dest_cpu is allowed for this process, migrate the task to it.
2999 * This is accomplished by forcing the cpu_allowed mask to only
3000 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3001 * the cpu_allowed mask is restored.
3003 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3005 struct migration_req req;
3006 unsigned long flags;
3009 rq = task_rq_lock(p, &flags);
3010 if (!cpu_isset(dest_cpu, p->cpus_allowed)
3011 || unlikely(cpu_is_offline(dest_cpu)))
3014 /* force the process onto the specified CPU */
3015 if (migrate_task(p, dest_cpu, &req)) {
3016 /* Need to wait for migration thread (might exit: take ref). */
3017 struct task_struct *mt = rq->migration_thread;
3019 get_task_struct(mt);
3020 task_rq_unlock(rq, &flags);
3021 wake_up_process(mt);
3022 put_task_struct(mt);
3023 wait_for_completion(&req.done);
3028 task_rq_unlock(rq, &flags);
3032 * sched_exec - execve() is a valuable balancing opportunity, because at
3033 * this point the task has the smallest effective memory and cache footprint.
3035 void sched_exec(void)
3037 int new_cpu, this_cpu = get_cpu();
3038 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3040 if (new_cpu != this_cpu)
3041 sched_migrate_task(current, new_cpu);
3045 * pull_task - move a task from a remote runqueue to the local runqueue.
3046 * Both runqueues must be locked.
3048 static void pull_task(struct rq *src_rq, struct task_struct *p,
3049 struct rq *this_rq, int this_cpu)
3051 deactivate_task(src_rq, p, 0);
3052 set_task_cpu(p, this_cpu);
3053 activate_task(this_rq, p, 0);
3055 * Note that idle threads have a prio of MAX_PRIO, for this test
3056 * to be always true for them.
3058 check_preempt_curr(this_rq, p);
3062 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3065 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3066 struct sched_domain *sd, enum cpu_idle_type idle,
3070 * We do not migrate tasks that are:
3071 * 1) running (obviously), or
3072 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3073 * 3) are cache-hot on their current CPU.
3075 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3076 schedstat_inc(p, se.nr_failed_migrations_affine);
3081 if (task_running(rq, p)) {
3082 schedstat_inc(p, se.nr_failed_migrations_running);
3087 * Aggressive migration if:
3088 * 1) task is cache cold, or
3089 * 2) too many balance attempts have failed.
3092 if (!task_hot(p, rq->clock, sd) ||
3093 sd->nr_balance_failed > sd->cache_nice_tries) {
3094 #ifdef CONFIG_SCHEDSTATS
3095 if (task_hot(p, rq->clock, sd)) {
3096 schedstat_inc(sd, lb_hot_gained[idle]);
3097 schedstat_inc(p, se.nr_forced_migrations);
3103 if (task_hot(p, rq->clock, sd)) {
3104 schedstat_inc(p, se.nr_failed_migrations_hot);
3110 static unsigned long
3111 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3112 unsigned long max_load_move, struct sched_domain *sd,
3113 enum cpu_idle_type idle, int *all_pinned,
3114 int *this_best_prio, struct rq_iterator *iterator)
3116 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3117 struct task_struct *p;
3118 long rem_load_move = max_load_move;
3120 if (max_load_move == 0)
3126 * Start the load-balancing iterator:
3128 p = iterator->start(iterator->arg);
3130 if (!p || loops++ > sysctl_sched_nr_migrate)
3133 * To help distribute high priority tasks across CPUs we don't
3134 * skip a task if it will be the highest priority task (i.e. smallest
3135 * prio value) on its new queue regardless of its load weight
3137 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3138 SCHED_LOAD_SCALE_FUZZ;
3139 if ((skip_for_load && p->prio >= *this_best_prio) ||
3140 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3141 p = iterator->next(iterator->arg);
3145 pull_task(busiest, p, this_rq, this_cpu);
3147 rem_load_move -= p->se.load.weight;
3150 * We only want to steal up to the prescribed amount of weighted load.
3152 if (rem_load_move > 0) {
3153 if (p->prio < *this_best_prio)
3154 *this_best_prio = p->prio;
3155 p = iterator->next(iterator->arg);
3160 * Right now, this is one of only two places pull_task() is called,
3161 * so we can safely collect pull_task() stats here rather than
3162 * inside pull_task().
3164 schedstat_add(sd, lb_gained[idle], pulled);
3167 *all_pinned = pinned;
3169 return max_load_move - rem_load_move;
3173 * move_tasks tries to move up to max_load_move weighted load from busiest to
3174 * this_rq, as part of a balancing operation within domain "sd".
3175 * Returns 1 if successful and 0 otherwise.
3177 * Called with both runqueues locked.
3179 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3180 unsigned long max_load_move,
3181 struct sched_domain *sd, enum cpu_idle_type idle,
3184 const struct sched_class *class = sched_class_highest;
3185 unsigned long total_load_moved = 0;
3186 int this_best_prio = this_rq->curr->prio;
3190 class->load_balance(this_rq, this_cpu, busiest,
3191 max_load_move - total_load_moved,
3192 sd, idle, all_pinned, &this_best_prio);
3193 class = class->next;
3194 } while (class && max_load_move > total_load_moved);
3196 return total_load_moved > 0;
3200 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3201 struct sched_domain *sd, enum cpu_idle_type idle,
3202 struct rq_iterator *iterator)
3204 struct task_struct *p = iterator->start(iterator->arg);
3208 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3209 pull_task(busiest, p, this_rq, this_cpu);
3211 * Right now, this is only the second place pull_task()
3212 * is called, so we can safely collect pull_task()
3213 * stats here rather than inside pull_task().
3215 schedstat_inc(sd, lb_gained[idle]);
3219 p = iterator->next(iterator->arg);
3226 * move_one_task tries to move exactly one task from busiest to this_rq, as
3227 * part of active balancing operations within "domain".
3228 * Returns 1 if successful and 0 otherwise.
3230 * Called with both runqueues locked.
3232 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3233 struct sched_domain *sd, enum cpu_idle_type idle)
3235 const struct sched_class *class;
3237 for (class = sched_class_highest; class; class = class->next)
3238 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3245 * find_busiest_group finds and returns the busiest CPU group within the
3246 * domain. It calculates and returns the amount of weighted load which
3247 * should be moved to restore balance via the imbalance parameter.
3249 static struct sched_group *
3250 find_busiest_group(struct sched_domain *sd, int this_cpu,
3251 unsigned long *imbalance, enum cpu_idle_type idle,
3252 int *sd_idle, const cpumask_t *cpus, int *balance)
3254 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3255 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3256 unsigned long max_pull;
3257 unsigned long busiest_load_per_task, busiest_nr_running;
3258 unsigned long this_load_per_task, this_nr_running;
3259 int load_idx, group_imb = 0;
3260 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3261 int power_savings_balance = 1;
3262 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3263 unsigned long min_nr_running = ULONG_MAX;
3264 struct sched_group *group_min = NULL, *group_leader = NULL;
3267 max_load = this_load = total_load = total_pwr = 0;
3268 busiest_load_per_task = busiest_nr_running = 0;
3269 this_load_per_task = this_nr_running = 0;
3270 if (idle == CPU_NOT_IDLE)
3271 load_idx = sd->busy_idx;
3272 else if (idle == CPU_NEWLY_IDLE)
3273 load_idx = sd->newidle_idx;
3275 load_idx = sd->idle_idx;
3278 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3281 int __group_imb = 0;
3282 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3283 unsigned long sum_nr_running, sum_weighted_load;
3285 local_group = cpu_isset(this_cpu, group->cpumask);
3288 balance_cpu = first_cpu(group->cpumask);
3290 /* Tally up the load of all CPUs in the group */
3291 sum_weighted_load = sum_nr_running = avg_load = 0;
3293 min_cpu_load = ~0UL;
3295 for_each_cpu_mask(i, group->cpumask) {
3298 if (!cpu_isset(i, *cpus))
3303 if (*sd_idle && rq->nr_running)
3306 /* Bias balancing toward cpus of our domain */
3308 if (idle_cpu(i) && !first_idle_cpu) {
3313 load = target_load(i, load_idx);
3315 load = source_load(i, load_idx);
3316 if (load > max_cpu_load)
3317 max_cpu_load = load;
3318 if (min_cpu_load > load)
3319 min_cpu_load = load;
3323 sum_nr_running += rq->nr_running;
3324 sum_weighted_load += weighted_cpuload(i);
3328 * First idle cpu or the first cpu(busiest) in this sched group
3329 * is eligible for doing load balancing at this and above
3330 * domains. In the newly idle case, we will allow all the cpu's
3331 * to do the newly idle load balance.
3333 if (idle != CPU_NEWLY_IDLE && local_group &&
3334 balance_cpu != this_cpu && balance) {
3339 total_load += avg_load;
3340 total_pwr += group->__cpu_power;
3342 /* Adjust by relative CPU power of the group */
3343 avg_load = sg_div_cpu_power(group,
3344 avg_load * SCHED_LOAD_SCALE);
3346 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3349 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3352 this_load = avg_load;
3354 this_nr_running = sum_nr_running;
3355 this_load_per_task = sum_weighted_load;
3356 } else if (avg_load > max_load &&
3357 (sum_nr_running > group_capacity || __group_imb)) {
3358 max_load = avg_load;
3360 busiest_nr_running = sum_nr_running;
3361 busiest_load_per_task = sum_weighted_load;
3362 group_imb = __group_imb;
3365 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3367 * Busy processors will not participate in power savings
3370 if (idle == CPU_NOT_IDLE ||
3371 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3375 * If the local group is idle or completely loaded
3376 * no need to do power savings balance at this domain
3378 if (local_group && (this_nr_running >= group_capacity ||
3380 power_savings_balance = 0;
3383 * If a group is already running at full capacity or idle,
3384 * don't include that group in power savings calculations
3386 if (!power_savings_balance || sum_nr_running >= group_capacity
3391 * Calculate the group which has the least non-idle load.
3392 * This is the group from where we need to pick up the load
3395 if ((sum_nr_running < min_nr_running) ||
3396 (sum_nr_running == min_nr_running &&
3397 first_cpu(group->cpumask) <
3398 first_cpu(group_min->cpumask))) {
3400 min_nr_running = sum_nr_running;
3401 min_load_per_task = sum_weighted_load /
3406 * Calculate the group which is almost near its
3407 * capacity but still has some space to pick up some load
3408 * from other group and save more power
3410 if (sum_nr_running <= group_capacity - 1) {
3411 if (sum_nr_running > leader_nr_running ||
3412 (sum_nr_running == leader_nr_running &&
3413 first_cpu(group->cpumask) >
3414 first_cpu(group_leader->cpumask))) {
3415 group_leader = group;
3416 leader_nr_running = sum_nr_running;
3421 group = group->next;
3422 } while (group != sd->groups);
3424 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3427 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3429 if (this_load >= avg_load ||
3430 100*max_load <= sd->imbalance_pct*this_load)
3433 busiest_load_per_task /= busiest_nr_running;
3435 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3438 * We're trying to get all the cpus to the average_load, so we don't
3439 * want to push ourselves above the average load, nor do we wish to
3440 * reduce the max loaded cpu below the average load, as either of these
3441 * actions would just result in more rebalancing later, and ping-pong
3442 * tasks around. Thus we look for the minimum possible imbalance.
3443 * Negative imbalances (*we* are more loaded than anyone else) will
3444 * be counted as no imbalance for these purposes -- we can't fix that
3445 * by pulling tasks to us. Be careful of negative numbers as they'll
3446 * appear as very large values with unsigned longs.
3448 if (max_load <= busiest_load_per_task)
3452 * In the presence of smp nice balancing, certain scenarios can have
3453 * max load less than avg load(as we skip the groups at or below
3454 * its cpu_power, while calculating max_load..)
3456 if (max_load < avg_load) {
3458 goto small_imbalance;
3461 /* Don't want to pull so many tasks that a group would go idle */
3462 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3464 /* How much load to actually move to equalise the imbalance */
3465 *imbalance = min(max_pull * busiest->__cpu_power,
3466 (avg_load - this_load) * this->__cpu_power)
3470 * if *imbalance is less than the average load per runnable task
3471 * there is no gaurantee that any tasks will be moved so we'll have
3472 * a think about bumping its value to force at least one task to be
3475 if (*imbalance < busiest_load_per_task) {
3476 unsigned long tmp, pwr_now, pwr_move;
3480 pwr_move = pwr_now = 0;
3482 if (this_nr_running) {
3483 this_load_per_task /= this_nr_running;
3484 if (busiest_load_per_task > this_load_per_task)
3487 this_load_per_task = SCHED_LOAD_SCALE;
3489 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3490 busiest_load_per_task * imbn) {
3491 *imbalance = busiest_load_per_task;
3496 * OK, we don't have enough imbalance to justify moving tasks,
3497 * however we may be able to increase total CPU power used by
3501 pwr_now += busiest->__cpu_power *
3502 min(busiest_load_per_task, max_load);
3503 pwr_now += this->__cpu_power *
3504 min(this_load_per_task, this_load);
3505 pwr_now /= SCHED_LOAD_SCALE;
3507 /* Amount of load we'd subtract */
3508 tmp = sg_div_cpu_power(busiest,
3509 busiest_load_per_task * SCHED_LOAD_SCALE);
3511 pwr_move += busiest->__cpu_power *
3512 min(busiest_load_per_task, max_load - tmp);
3514 /* Amount of load we'd add */
3515 if (max_load * busiest->__cpu_power <
3516 busiest_load_per_task * SCHED_LOAD_SCALE)
3517 tmp = sg_div_cpu_power(this,
3518 max_load * busiest->__cpu_power);
3520 tmp = sg_div_cpu_power(this,
3521 busiest_load_per_task * SCHED_LOAD_SCALE);
3522 pwr_move += this->__cpu_power *
3523 min(this_load_per_task, this_load + tmp);
3524 pwr_move /= SCHED_LOAD_SCALE;
3526 /* Move if we gain throughput */
3527 if (pwr_move > pwr_now)
3528 *imbalance = busiest_load_per_task;
3534 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3535 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3538 if (this == group_leader && group_leader != group_min) {
3539 *imbalance = min_load_per_task;
3549 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3552 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3553 unsigned long imbalance, const cpumask_t *cpus)
3555 struct rq *busiest = NULL, *rq;
3556 unsigned long max_load = 0;
3559 for_each_cpu_mask(i, group->cpumask) {
3562 if (!cpu_isset(i, *cpus))
3566 wl = weighted_cpuload(i);
3568 if (rq->nr_running == 1 && wl > imbalance)
3571 if (wl > max_load) {
3581 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3582 * so long as it is large enough.
3584 #define MAX_PINNED_INTERVAL 512
3587 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3588 * tasks if there is an imbalance.
3590 static int load_balance(int this_cpu, struct rq *this_rq,
3591 struct sched_domain *sd, enum cpu_idle_type idle,
3592 int *balance, cpumask_t *cpus)
3594 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3595 struct sched_group *group;
3596 unsigned long imbalance;
3598 unsigned long flags;
3599 int unlock_aggregate;
3603 unlock_aggregate = get_aggregate(sd);
3606 * When power savings policy is enabled for the parent domain, idle
3607 * sibling can pick up load irrespective of busy siblings. In this case,
3608 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3609 * portraying it as CPU_NOT_IDLE.
3611 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3612 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3615 schedstat_inc(sd, lb_count[idle]);
3618 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3625 schedstat_inc(sd, lb_nobusyg[idle]);
3629 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3631 schedstat_inc(sd, lb_nobusyq[idle]);
3635 BUG_ON(busiest == this_rq);
3637 schedstat_add(sd, lb_imbalance[idle], imbalance);
3640 if (busiest->nr_running > 1) {
3642 * Attempt to move tasks. If find_busiest_group has found
3643 * an imbalance but busiest->nr_running <= 1, the group is
3644 * still unbalanced. ld_moved simply stays zero, so it is
3645 * correctly treated as an imbalance.
3647 local_irq_save(flags);
3648 double_rq_lock(this_rq, busiest);
3649 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3650 imbalance, sd, idle, &all_pinned);
3651 double_rq_unlock(this_rq, busiest);
3652 local_irq_restore(flags);
3655 * some other cpu did the load balance for us.
3657 if (ld_moved && this_cpu != smp_processor_id())
3658 resched_cpu(this_cpu);
3660 /* All tasks on this runqueue were pinned by CPU affinity */
3661 if (unlikely(all_pinned)) {
3662 cpu_clear(cpu_of(busiest), *cpus);
3663 if (!cpus_empty(*cpus))
3670 schedstat_inc(sd, lb_failed[idle]);
3671 sd->nr_balance_failed++;
3673 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3675 spin_lock_irqsave(&busiest->lock, flags);
3677 /* don't kick the migration_thread, if the curr
3678 * task on busiest cpu can't be moved to this_cpu
3680 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3681 spin_unlock_irqrestore(&busiest->lock, flags);
3683 goto out_one_pinned;
3686 if (!busiest->active_balance) {
3687 busiest->active_balance = 1;
3688 busiest->push_cpu = this_cpu;
3691 spin_unlock_irqrestore(&busiest->lock, flags);
3693 wake_up_process(busiest->migration_thread);
3696 * We've kicked active balancing, reset the failure
3699 sd->nr_balance_failed = sd->cache_nice_tries+1;
3702 sd->nr_balance_failed = 0;
3704 if (likely(!active_balance)) {
3705 /* We were unbalanced, so reset the balancing interval */
3706 sd->balance_interval = sd->min_interval;
3709 * If we've begun active balancing, start to back off. This
3710 * case may not be covered by the all_pinned logic if there
3711 * is only 1 task on the busy runqueue (because we don't call
3714 if (sd->balance_interval < sd->max_interval)
3715 sd->balance_interval *= 2;
3718 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3719 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3725 schedstat_inc(sd, lb_balanced[idle]);
3727 sd->nr_balance_failed = 0;
3730 /* tune up the balancing interval */
3731 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3732 (sd->balance_interval < sd->max_interval))
3733 sd->balance_interval *= 2;
3735 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3736 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3741 if (unlock_aggregate)
3747 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3748 * tasks if there is an imbalance.
3750 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3751 * this_rq is locked.
3754 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3757 struct sched_group *group;
3758 struct rq *busiest = NULL;
3759 unsigned long imbalance;
3767 * When power savings policy is enabled for the parent domain, idle
3768 * sibling can pick up load irrespective of busy siblings. In this case,
3769 * let the state of idle sibling percolate up as IDLE, instead of
3770 * portraying it as CPU_NOT_IDLE.
3772 if (sd->flags & SD_SHARE_CPUPOWER &&
3773 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3776 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3778 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3779 &sd_idle, cpus, NULL);
3781 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3785 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3787 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3791 BUG_ON(busiest == this_rq);
3793 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3796 if (busiest->nr_running > 1) {
3797 /* Attempt to move tasks */
3798 double_lock_balance(this_rq, busiest);
3799 /* this_rq->clock is already updated */
3800 update_rq_clock(busiest);
3801 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3802 imbalance, sd, CPU_NEWLY_IDLE,
3804 spin_unlock(&busiest->lock);
3806 if (unlikely(all_pinned)) {
3807 cpu_clear(cpu_of(busiest), *cpus);
3808 if (!cpus_empty(*cpus))
3814 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3815 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3816 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3819 sd->nr_balance_failed = 0;
3824 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3825 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3826 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3828 sd->nr_balance_failed = 0;
3834 * idle_balance is called by schedule() if this_cpu is about to become
3835 * idle. Attempts to pull tasks from other CPUs.
3837 static void idle_balance(int this_cpu, struct rq *this_rq)
3839 struct sched_domain *sd;
3840 int pulled_task = -1;
3841 unsigned long next_balance = jiffies + HZ;
3844 for_each_domain(this_cpu, sd) {
3845 unsigned long interval;
3847 if (!(sd->flags & SD_LOAD_BALANCE))
3850 if (sd->flags & SD_BALANCE_NEWIDLE)
3851 /* If we've pulled tasks over stop searching: */
3852 pulled_task = load_balance_newidle(this_cpu, this_rq,
3855 interval = msecs_to_jiffies(sd->balance_interval);
3856 if (time_after(next_balance, sd->last_balance + interval))
3857 next_balance = sd->last_balance + interval;
3861 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3863 * We are going idle. next_balance may be set based on
3864 * a busy processor. So reset next_balance.
3866 this_rq->next_balance = next_balance;
3871 * active_load_balance is run by migration threads. It pushes running tasks
3872 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3873 * running on each physical CPU where possible, and avoids physical /
3874 * logical imbalances.
3876 * Called with busiest_rq locked.
3878 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3880 int target_cpu = busiest_rq->push_cpu;
3881 struct sched_domain *sd;
3882 struct rq *target_rq;
3884 /* Is there any task to move? */
3885 if (busiest_rq->nr_running <= 1)
3888 target_rq = cpu_rq(target_cpu);
3891 * This condition is "impossible", if it occurs
3892 * we need to fix it. Originally reported by
3893 * Bjorn Helgaas on a 128-cpu setup.
3895 BUG_ON(busiest_rq == target_rq);
3897 /* move a task from busiest_rq to target_rq */
3898 double_lock_balance(busiest_rq, target_rq);
3899 update_rq_clock(busiest_rq);
3900 update_rq_clock(target_rq);
3902 /* Search for an sd spanning us and the target CPU. */
3903 for_each_domain(target_cpu, sd) {
3904 if ((sd->flags & SD_LOAD_BALANCE) &&
3905 cpu_isset(busiest_cpu, sd->span))
3910 schedstat_inc(sd, alb_count);
3912 if (move_one_task(target_rq, target_cpu, busiest_rq,
3914 schedstat_inc(sd, alb_pushed);
3916 schedstat_inc(sd, alb_failed);
3918 spin_unlock(&target_rq->lock);
3923 atomic_t load_balancer;
3925 } nohz ____cacheline_aligned = {
3926 .load_balancer = ATOMIC_INIT(-1),
3927 .cpu_mask = CPU_MASK_NONE,
3931 * This routine will try to nominate the ilb (idle load balancing)
3932 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3933 * load balancing on behalf of all those cpus. If all the cpus in the system
3934 * go into this tickless mode, then there will be no ilb owner (as there is
3935 * no need for one) and all the cpus will sleep till the next wakeup event
3938 * For the ilb owner, tick is not stopped. And this tick will be used
3939 * for idle load balancing. ilb owner will still be part of
3942 * While stopping the tick, this cpu will become the ilb owner if there
3943 * is no other owner. And will be the owner till that cpu becomes busy
3944 * or if all cpus in the system stop their ticks at which point
3945 * there is no need for ilb owner.
3947 * When the ilb owner becomes busy, it nominates another owner, during the
3948 * next busy scheduler_tick()
3950 int select_nohz_load_balancer(int stop_tick)
3952 int cpu = smp_processor_id();
3955 cpu_set(cpu, nohz.cpu_mask);
3956 cpu_rq(cpu)->in_nohz_recently = 1;
3959 * If we are going offline and still the leader, give up!
3961 if (cpu_is_offline(cpu) &&
3962 atomic_read(&nohz.load_balancer) == cpu) {
3963 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3968 /* time for ilb owner also to sleep */
3969 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3970 if (atomic_read(&nohz.load_balancer) == cpu)
3971 atomic_set(&nohz.load_balancer, -1);
3975 if (atomic_read(&nohz.load_balancer) == -1) {
3976 /* make me the ilb owner */
3977 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3979 } else if (atomic_read(&nohz.load_balancer) == cpu)
3982 if (!cpu_isset(cpu, nohz.cpu_mask))
3985 cpu_clear(cpu, nohz.cpu_mask);
3987 if (atomic_read(&nohz.load_balancer) == cpu)
3988 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3995 static DEFINE_SPINLOCK(balancing);
3998 * It checks each scheduling domain to see if it is due to be balanced,
3999 * and initiates a balancing operation if so.
4001 * Balancing parameters are set up in arch_init_sched_domains.
4003 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4006 struct rq *rq = cpu_rq(cpu);
4007 unsigned long interval;
4008 struct sched_domain *sd;
4009 /* Earliest time when we have to do rebalance again */
4010 unsigned long next_balance = jiffies + 60*HZ;
4011 int update_next_balance = 0;
4014 for_each_domain(cpu, sd) {
4015 if (!(sd->flags & SD_LOAD_BALANCE))
4018 interval = sd->balance_interval;
4019 if (idle != CPU_IDLE)
4020 interval *= sd->busy_factor;
4022 /* scale ms to jiffies */
4023 interval = msecs_to_jiffies(interval);
4024 if (unlikely(!interval))
4026 if (interval > HZ*NR_CPUS/10)
4027 interval = HZ*NR_CPUS/10;
4030 if (sd->flags & SD_SERIALIZE) {
4031 if (!spin_trylock(&balancing))
4035 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4036 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
4038 * We've pulled tasks over so either we're no
4039 * longer idle, or one of our SMT siblings is
4042 idle = CPU_NOT_IDLE;
4044 sd->last_balance = jiffies;
4046 if (sd->flags & SD_SERIALIZE)
4047 spin_unlock(&balancing);
4049 if (time_after(next_balance, sd->last_balance + interval)) {
4050 next_balance = sd->last_balance + interval;
4051 update_next_balance = 1;
4055 * Stop the load balance at this level. There is another
4056 * CPU in our sched group which is doing load balancing more
4064 * next_balance will be updated only when there is a need.
4065 * When the cpu is attached to null domain for ex, it will not be
4068 if (likely(update_next_balance))
4069 rq->next_balance = next_balance;
4073 * run_rebalance_domains is triggered when needed from the scheduler tick.
4074 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4075 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4077 static void run_rebalance_domains(struct softirq_action *h)
4079 int this_cpu = smp_processor_id();
4080 struct rq *this_rq = cpu_rq(this_cpu);
4081 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4082 CPU_IDLE : CPU_NOT_IDLE;
4084 rebalance_domains(this_cpu, idle);
4088 * If this cpu is the owner for idle load balancing, then do the
4089 * balancing on behalf of the other idle cpus whose ticks are
4092 if (this_rq->idle_at_tick &&
4093 atomic_read(&nohz.load_balancer) == this_cpu) {
4094 cpumask_t cpus = nohz.cpu_mask;
4098 cpu_clear(this_cpu, cpus);
4099 for_each_cpu_mask(balance_cpu, cpus) {
4101 * If this cpu gets work to do, stop the load balancing
4102 * work being done for other cpus. Next load
4103 * balancing owner will pick it up.
4108 rebalance_domains(balance_cpu, CPU_IDLE);
4110 rq = cpu_rq(balance_cpu);
4111 if (time_after(this_rq->next_balance, rq->next_balance))
4112 this_rq->next_balance = rq->next_balance;
4119 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4121 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4122 * idle load balancing owner or decide to stop the periodic load balancing,
4123 * if the whole system is idle.
4125 static inline void trigger_load_balance(struct rq *rq, int cpu)
4129 * If we were in the nohz mode recently and busy at the current
4130 * scheduler tick, then check if we need to nominate new idle
4133 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4134 rq->in_nohz_recently = 0;
4136 if (atomic_read(&nohz.load_balancer) == cpu) {
4137 cpu_clear(cpu, nohz.cpu_mask);
4138 atomic_set(&nohz.load_balancer, -1);
4141 if (atomic_read(&nohz.load_balancer) == -1) {
4143 * simple selection for now: Nominate the
4144 * first cpu in the nohz list to be the next
4147 * TBD: Traverse the sched domains and nominate
4148 * the nearest cpu in the nohz.cpu_mask.
4150 int ilb = first_cpu(nohz.cpu_mask);
4152 if (ilb < nr_cpu_ids)
4158 * If this cpu is idle and doing idle load balancing for all the
4159 * cpus with ticks stopped, is it time for that to stop?
4161 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4162 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4168 * If this cpu is idle and the idle load balancing is done by
4169 * someone else, then no need raise the SCHED_SOFTIRQ
4171 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4172 cpu_isset(cpu, nohz.cpu_mask))
4175 if (time_after_eq(jiffies, rq->next_balance))
4176 raise_softirq(SCHED_SOFTIRQ);
4179 #else /* CONFIG_SMP */
4182 * on UP we do not need to balance between CPUs:
4184 static inline void idle_balance(int cpu, struct rq *rq)
4190 DEFINE_PER_CPU(struct kernel_stat, kstat);
4192 EXPORT_PER_CPU_SYMBOL(kstat);
4195 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4196 * that have not yet been banked in case the task is currently running.
4198 unsigned long long task_sched_runtime(struct task_struct *p)
4200 unsigned long flags;
4204 rq = task_rq_lock(p, &flags);
4205 ns = p->se.sum_exec_runtime;
4206 if (task_current(rq, p)) {
4207 update_rq_clock(rq);
4208 delta_exec = rq->clock - p->se.exec_start;
4209 if ((s64)delta_exec > 0)
4212 task_rq_unlock(rq, &flags);
4218 * Account user cpu time to a process.
4219 * @p: the process that the cpu time gets accounted to
4220 * @cputime: the cpu time spent in user space since the last update
4222 void account_user_time(struct task_struct *p, cputime_t cputime)
4224 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4227 p->utime = cputime_add(p->utime, cputime);
4229 /* Add user time to cpustat. */
4230 tmp = cputime_to_cputime64(cputime);
4231 if (TASK_NICE(p) > 0)
4232 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4234 cpustat->user = cputime64_add(cpustat->user, tmp);
4238 * Account guest cpu time to a process.
4239 * @p: the process that the cpu time gets accounted to
4240 * @cputime: the cpu time spent in virtual machine since the last update
4242 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4245 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4247 tmp = cputime_to_cputime64(cputime);
4249 p->utime = cputime_add(p->utime, cputime);
4250 p->gtime = cputime_add(p->gtime, cputime);
4252 cpustat->user = cputime64_add(cpustat->user, tmp);
4253 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4257 * Account scaled user cpu time to a process.
4258 * @p: the process that the cpu time gets accounted to
4259 * @cputime: the cpu time spent in user space since the last update
4261 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4263 p->utimescaled = cputime_add(p->utimescaled, cputime);
4267 * Account system cpu time to a process.
4268 * @p: the process that the cpu time gets accounted to
4269 * @hardirq_offset: the offset to subtract from hardirq_count()
4270 * @cputime: the cpu time spent in kernel space since the last update
4272 void account_system_time(struct task_struct *p, int hardirq_offset,
4275 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4276 struct rq *rq = this_rq();
4279 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
4280 return account_guest_time(p, cputime);
4282 p->stime = cputime_add(p->stime, cputime);
4284 /* Add system time to cpustat. */
4285 tmp = cputime_to_cputime64(cputime);
4286 if (hardirq_count() - hardirq_offset)
4287 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4288 else if (softirq_count())
4289 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4290 else if (p != rq->idle)
4291 cpustat->system = cputime64_add(cpustat->system, tmp);
4292 else if (atomic_read(&rq->nr_iowait) > 0)
4293 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4295 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4296 /* Account for system time used */
4297 acct_update_integrals(p);
4301 * Account scaled system cpu time to a process.
4302 * @p: the process that the cpu time gets accounted to
4303 * @hardirq_offset: the offset to subtract from hardirq_count()
4304 * @cputime: the cpu time spent in kernel space since the last update
4306 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4308 p->stimescaled = cputime_add(p->stimescaled, cputime);
4312 * Account for involuntary wait time.
4313 * @p: the process from which the cpu time has been stolen
4314 * @steal: the cpu time spent in involuntary wait
4316 void account_steal_time(struct task_struct *p, cputime_t steal)
4318 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4319 cputime64_t tmp = cputime_to_cputime64(steal);
4320 struct rq *rq = this_rq();
4322 if (p == rq->idle) {
4323 p->stime = cputime_add(p->stime, steal);
4324 if (atomic_read(&rq->nr_iowait) > 0)
4325 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4327 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4329 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4333 * This function gets called by the timer code, with HZ frequency.
4334 * We call it with interrupts disabled.
4336 * It also gets called by the fork code, when changing the parent's
4339 void scheduler_tick(void)
4341 int cpu = smp_processor_id();
4342 struct rq *rq = cpu_rq(cpu);
4343 struct task_struct *curr = rq->curr;
4344 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
4346 spin_lock(&rq->lock);
4347 __update_rq_clock(rq);
4349 * Let rq->clock advance by at least TICK_NSEC:
4351 if (unlikely(rq->clock < next_tick)) {
4352 rq->clock = next_tick;
4353 rq->clock_underflows++;
4355 rq->tick_timestamp = rq->clock;
4356 update_last_tick_seen(rq);
4357 update_cpu_load(rq);
4358 curr->sched_class->task_tick(rq, curr, 0);
4359 spin_unlock(&rq->lock);
4362 rq->idle_at_tick = idle_cpu(cpu);
4363 trigger_load_balance(rq, cpu);
4367 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4369 void __kprobes add_preempt_count(int val)
4374 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4376 preempt_count() += val;
4378 * Spinlock count overflowing soon?
4380 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4383 EXPORT_SYMBOL(add_preempt_count);
4385 void __kprobes sub_preempt_count(int val)
4390 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4393 * Is the spinlock portion underflowing?
4395 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4396 !(preempt_count() & PREEMPT_MASK)))
4399 preempt_count() -= val;
4401 EXPORT_SYMBOL(sub_preempt_count);
4406 * Print scheduling while atomic bug:
4408 static noinline void __schedule_bug(struct task_struct *prev)
4410 struct pt_regs *regs = get_irq_regs();
4412 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4413 prev->comm, prev->pid, preempt_count());
4415 debug_show_held_locks(prev);
4416 if (irqs_disabled())
4417 print_irqtrace_events(prev);
4426 * Various schedule()-time debugging checks and statistics:
4428 static inline void schedule_debug(struct task_struct *prev)
4431 * Test if we are atomic. Since do_exit() needs to call into
4432 * schedule() atomically, we ignore that path for now.
4433 * Otherwise, whine if we are scheduling when we should not be.
4435 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4436 __schedule_bug(prev);
4438 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4440 schedstat_inc(this_rq(), sched_count);
4441 #ifdef CONFIG_SCHEDSTATS
4442 if (unlikely(prev->lock_depth >= 0)) {
4443 schedstat_inc(this_rq(), bkl_count);
4444 schedstat_inc(prev, sched_info.bkl_count);
4450 * Pick up the highest-prio task:
4452 static inline struct task_struct *
4453 pick_next_task(struct rq *rq, struct task_struct *prev)
4455 const struct sched_class *class;
4456 struct task_struct *p;
4459 * Optimization: we know that if all tasks are in
4460 * the fair class we can call that function directly:
4462 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4463 p = fair_sched_class.pick_next_task(rq);
4468 class = sched_class_highest;
4470 p = class->pick_next_task(rq);
4474 * Will never be NULL as the idle class always
4475 * returns a non-NULL p:
4477 class = class->next;
4482 * schedule() is the main scheduler function.
4484 asmlinkage void __sched schedule(void)
4486 struct task_struct *prev, *next;
4487 unsigned long *switch_count;
4493 cpu = smp_processor_id();
4497 switch_count = &prev->nivcsw;
4499 release_kernel_lock(prev);
4500 need_resched_nonpreemptible:
4502 schedule_debug(prev);
4507 * Do the rq-clock update outside the rq lock:
4509 local_irq_disable();
4510 __update_rq_clock(rq);
4511 spin_lock(&rq->lock);
4512 clear_tsk_need_resched(prev);
4514 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4515 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4516 signal_pending(prev))) {
4517 prev->state = TASK_RUNNING;
4519 deactivate_task(rq, prev, 1);
4521 switch_count = &prev->nvcsw;
4525 if (prev->sched_class->pre_schedule)
4526 prev->sched_class->pre_schedule(rq, prev);
4529 if (unlikely(!rq->nr_running))
4530 idle_balance(cpu, rq);
4532 prev->sched_class->put_prev_task(rq, prev);
4533 next = pick_next_task(rq, prev);
4535 sched_info_switch(prev, next);
4537 if (likely(prev != next)) {
4542 context_switch(rq, prev, next); /* unlocks the rq */
4544 * the context switch might have flipped the stack from under
4545 * us, hence refresh the local variables.
4547 cpu = smp_processor_id();
4550 spin_unlock_irq(&rq->lock);
4554 if (unlikely(reacquire_kernel_lock(current) < 0))
4555 goto need_resched_nonpreemptible;
4557 preempt_enable_no_resched();
4558 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4561 EXPORT_SYMBOL(schedule);
4563 #ifdef CONFIG_PREEMPT
4565 * this is the entry point to schedule() from in-kernel preemption
4566 * off of preempt_enable. Kernel preemptions off return from interrupt
4567 * occur there and call schedule directly.
4569 asmlinkage void __sched preempt_schedule(void)
4571 struct thread_info *ti = current_thread_info();
4572 struct task_struct *task = current;
4573 int saved_lock_depth;
4576 * If there is a non-zero preempt_count or interrupts are disabled,
4577 * we do not want to preempt the current task. Just return..
4579 if (likely(ti->preempt_count || irqs_disabled()))
4583 add_preempt_count(PREEMPT_ACTIVE);
4586 * We keep the big kernel semaphore locked, but we
4587 * clear ->lock_depth so that schedule() doesnt
4588 * auto-release the semaphore:
4590 saved_lock_depth = task->lock_depth;
4591 task->lock_depth = -1;
4593 task->lock_depth = saved_lock_depth;
4594 sub_preempt_count(PREEMPT_ACTIVE);
4597 * Check again in case we missed a preemption opportunity
4598 * between schedule and now.
4601 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4603 EXPORT_SYMBOL(preempt_schedule);
4606 * this is the entry point to schedule() from kernel preemption
4607 * off of irq context.
4608 * Note, that this is called and return with irqs disabled. This will
4609 * protect us against recursive calling from irq.
4611 asmlinkage void __sched preempt_schedule_irq(void)
4613 struct thread_info *ti = current_thread_info();
4614 struct task_struct *task = current;
4615 int saved_lock_depth;
4617 /* Catch callers which need to be fixed */
4618 BUG_ON(ti->preempt_count || !irqs_disabled());
4621 add_preempt_count(PREEMPT_ACTIVE);
4624 * We keep the big kernel semaphore locked, but we
4625 * clear ->lock_depth so that schedule() doesnt
4626 * auto-release the semaphore:
4628 saved_lock_depth = task->lock_depth;
4629 task->lock_depth = -1;
4632 local_irq_disable();
4633 task->lock_depth = saved_lock_depth;
4634 sub_preempt_count(PREEMPT_ACTIVE);
4637 * Check again in case we missed a preemption opportunity
4638 * between schedule and now.
4641 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4644 #endif /* CONFIG_PREEMPT */
4646 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4649 return try_to_wake_up(curr->private, mode, sync);
4651 EXPORT_SYMBOL(default_wake_function);
4654 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4655 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4656 * number) then we wake all the non-exclusive tasks and one exclusive task.
4658 * There are circumstances in which we can try to wake a task which has already
4659 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4660 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4662 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4663 int nr_exclusive, int sync, void *key)
4665 wait_queue_t *curr, *next;
4667 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4668 unsigned flags = curr->flags;
4670 if (curr->func(curr, mode, sync, key) &&
4671 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4677 * __wake_up - wake up threads blocked on a waitqueue.
4679 * @mode: which threads
4680 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4681 * @key: is directly passed to the wakeup function
4683 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4684 int nr_exclusive, void *key)
4686 unsigned long flags;
4688 spin_lock_irqsave(&q->lock, flags);
4689 __wake_up_common(q, mode, nr_exclusive, 0, key);
4690 spin_unlock_irqrestore(&q->lock, flags);
4692 EXPORT_SYMBOL(__wake_up);
4695 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4697 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4699 __wake_up_common(q, mode, 1, 0, NULL);
4703 * __wake_up_sync - wake up threads blocked on a waitqueue.
4705 * @mode: which threads
4706 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4708 * The sync wakeup differs that the waker knows that it will schedule
4709 * away soon, so while the target thread will be woken up, it will not
4710 * be migrated to another CPU - ie. the two threads are 'synchronized'
4711 * with each other. This can prevent needless bouncing between CPUs.
4713 * On UP it can prevent extra preemption.
4716 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4718 unsigned long flags;
4724 if (unlikely(!nr_exclusive))
4727 spin_lock_irqsave(&q->lock, flags);
4728 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4729 spin_unlock_irqrestore(&q->lock, flags);
4731 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4733 void complete(struct completion *x)
4735 unsigned long flags;
4737 spin_lock_irqsave(&x->wait.lock, flags);
4739 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4740 spin_unlock_irqrestore(&x->wait.lock, flags);
4742 EXPORT_SYMBOL(complete);
4744 void complete_all(struct completion *x)
4746 unsigned long flags;
4748 spin_lock_irqsave(&x->wait.lock, flags);
4749 x->done += UINT_MAX/2;
4750 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4751 spin_unlock_irqrestore(&x->wait.lock, flags);
4753 EXPORT_SYMBOL(complete_all);
4755 static inline long __sched
4756 do_wait_for_common(struct completion *x, long timeout, int state)
4759 DECLARE_WAITQUEUE(wait, current);
4761 wait.flags |= WQ_FLAG_EXCLUSIVE;
4762 __add_wait_queue_tail(&x->wait, &wait);
4764 if ((state == TASK_INTERRUPTIBLE &&
4765 signal_pending(current)) ||
4766 (state == TASK_KILLABLE &&
4767 fatal_signal_pending(current))) {
4768 __remove_wait_queue(&x->wait, &wait);
4769 return -ERESTARTSYS;
4771 __set_current_state(state);
4772 spin_unlock_irq(&x->wait.lock);
4773 timeout = schedule_timeout(timeout);
4774 spin_lock_irq(&x->wait.lock);
4776 __remove_wait_queue(&x->wait, &wait);
4780 __remove_wait_queue(&x->wait, &wait);
4787 wait_for_common(struct completion *x, long timeout, int state)
4791 spin_lock_irq(&x->wait.lock);
4792 timeout = do_wait_for_common(x, timeout, state);
4793 spin_unlock_irq(&x->wait.lock);
4797 void __sched wait_for_completion(struct completion *x)
4799 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4801 EXPORT_SYMBOL(wait_for_completion);
4803 unsigned long __sched
4804 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4806 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4808 EXPORT_SYMBOL(wait_for_completion_timeout);
4810 int __sched wait_for_completion_interruptible(struct completion *x)
4812 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4813 if (t == -ERESTARTSYS)
4817 EXPORT_SYMBOL(wait_for_completion_interruptible);
4819 unsigned long __sched
4820 wait_for_completion_interruptible_timeout(struct completion *x,
4821 unsigned long timeout)
4823 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4825 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4827 int __sched wait_for_completion_killable(struct completion *x)
4829 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4830 if (t == -ERESTARTSYS)
4834 EXPORT_SYMBOL(wait_for_completion_killable);
4837 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4839 unsigned long flags;
4842 init_waitqueue_entry(&wait, current);
4844 __set_current_state(state);
4846 spin_lock_irqsave(&q->lock, flags);
4847 __add_wait_queue(q, &wait);
4848 spin_unlock(&q->lock);
4849 timeout = schedule_timeout(timeout);
4850 spin_lock_irq(&q->lock);
4851 __remove_wait_queue(q, &wait);
4852 spin_unlock_irqrestore(&q->lock, flags);
4857 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4859 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4861 EXPORT_SYMBOL(interruptible_sleep_on);
4864 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4866 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4868 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4870 void __sched sleep_on(wait_queue_head_t *q)
4872 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4874 EXPORT_SYMBOL(sleep_on);
4876 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4878 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4880 EXPORT_SYMBOL(sleep_on_timeout);
4882 #ifdef CONFIG_RT_MUTEXES
4885 * rt_mutex_setprio - set the current priority of a task
4887 * @prio: prio value (kernel-internal form)
4889 * This function changes the 'effective' priority of a task. It does
4890 * not touch ->normal_prio like __setscheduler().
4892 * Used by the rt_mutex code to implement priority inheritance logic.
4894 void rt_mutex_setprio(struct task_struct *p, int prio)
4896 unsigned long flags;
4897 int oldprio, on_rq, running;
4899 const struct sched_class *prev_class = p->sched_class;
4901 BUG_ON(prio < 0 || prio > MAX_PRIO);
4903 rq = task_rq_lock(p, &flags);
4904 update_rq_clock(rq);
4907 on_rq = p->se.on_rq;
4908 running = task_current(rq, p);
4910 dequeue_task(rq, p, 0);
4912 p->sched_class->put_prev_task(rq, p);
4915 p->sched_class = &rt_sched_class;
4917 p->sched_class = &fair_sched_class;
4922 p->sched_class->set_curr_task(rq);
4924 enqueue_task(rq, p, 0);
4926 check_class_changed(rq, p, prev_class, oldprio, running);
4928 task_rq_unlock(rq, &flags);
4933 void set_user_nice(struct task_struct *p, long nice)
4935 int old_prio, delta, on_rq;
4936 unsigned long flags;
4939 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4942 * We have to be careful, if called from sys_setpriority(),
4943 * the task might be in the middle of scheduling on another CPU.
4945 rq = task_rq_lock(p, &flags);
4946 update_rq_clock(rq);
4948 * The RT priorities are set via sched_setscheduler(), but we still
4949 * allow the 'normal' nice value to be set - but as expected
4950 * it wont have any effect on scheduling until the task is
4951 * SCHED_FIFO/SCHED_RR:
4953 if (task_has_rt_policy(p)) {
4954 p->static_prio = NICE_TO_PRIO(nice);
4957 on_rq = p->se.on_rq;
4959 dequeue_task(rq, p, 0);
4961 p->static_prio = NICE_TO_PRIO(nice);
4964 p->prio = effective_prio(p);
4965 delta = p->prio - old_prio;
4968 enqueue_task(rq, p, 0);
4970 * If the task increased its priority or is running and
4971 * lowered its priority, then reschedule its CPU:
4973 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4974 resched_task(rq->curr);
4977 task_rq_unlock(rq, &flags);
4979 EXPORT_SYMBOL(set_user_nice);
4982 * can_nice - check if a task can reduce its nice value
4986 int can_nice(const struct task_struct *p, const int nice)
4988 /* convert nice value [19,-20] to rlimit style value [1,40] */
4989 int nice_rlim = 20 - nice;
4991 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4992 capable(CAP_SYS_NICE));
4995 #ifdef __ARCH_WANT_SYS_NICE
4998 * sys_nice - change the priority of the current process.
4999 * @increment: priority increment
5001 * sys_setpriority is a more generic, but much slower function that
5002 * does similar things.
5004 asmlinkage long sys_nice(int increment)
5009 * Setpriority might change our priority at the same moment.
5010 * We don't have to worry. Conceptually one call occurs first
5011 * and we have a single winner.
5013 if (increment < -40)
5018 nice = PRIO_TO_NICE(current->static_prio) + increment;
5024 if (increment < 0 && !can_nice(current, nice))
5027 retval = security_task_setnice(current, nice);
5031 set_user_nice(current, nice);
5038 * task_prio - return the priority value of a given task.
5039 * @p: the task in question.
5041 * This is the priority value as seen by users in /proc.
5042 * RT tasks are offset by -200. Normal tasks are centered
5043 * around 0, value goes from -16 to +15.
5045 int task_prio(const struct task_struct *p)
5047 return p->prio - MAX_RT_PRIO;
5051 * task_nice - return the nice value of a given task.
5052 * @p: the task in question.
5054 int task_nice(const struct task_struct *p)
5056 return TASK_NICE(p);
5058 EXPORT_SYMBOL(task_nice);
5061 * idle_cpu - is a given cpu idle currently?
5062 * @cpu: the processor in question.
5064 int idle_cpu(int cpu)
5066 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5070 * idle_task - return the idle task for a given cpu.
5071 * @cpu: the processor in question.
5073 struct task_struct *idle_task(int cpu)
5075 return cpu_rq(cpu)->idle;
5079 * find_process_by_pid - find a process with a matching PID value.
5080 * @pid: the pid in question.
5082 static struct task_struct *find_process_by_pid(pid_t pid)
5084 return pid ? find_task_by_vpid(pid) : current;
5087 /* Actually do priority change: must hold rq lock. */
5089 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5091 BUG_ON(p->se.on_rq);
5094 switch (p->policy) {
5098 p->sched_class = &fair_sched_class;
5102 p->sched_class = &rt_sched_class;
5106 p->rt_priority = prio;
5107 p->normal_prio = normal_prio(p);
5108 /* we are holding p->pi_lock already */
5109 p->prio = rt_mutex_getprio(p);
5114 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5115 * @p: the task in question.
5116 * @policy: new policy.
5117 * @param: structure containing the new RT priority.
5119 * NOTE that the task may be already dead.
5121 int sched_setscheduler(struct task_struct *p, int policy,
5122 struct sched_param *param)
5124 int retval, oldprio, oldpolicy = -1, on_rq, running;
5125 unsigned long flags;
5126 const struct sched_class *prev_class = p->sched_class;
5129 /* may grab non-irq protected spin_locks */
5130 BUG_ON(in_interrupt());
5132 /* double check policy once rq lock held */
5134 policy = oldpolicy = p->policy;
5135 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5136 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5137 policy != SCHED_IDLE)
5140 * Valid priorities for SCHED_FIFO and SCHED_RR are
5141 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5142 * SCHED_BATCH and SCHED_IDLE is 0.
5144 if (param->sched_priority < 0 ||
5145 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5146 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5148 if (rt_policy(policy) != (param->sched_priority != 0))
5152 * Allow unprivileged RT tasks to decrease priority:
5154 if (!capable(CAP_SYS_NICE)) {
5155 if (rt_policy(policy)) {
5156 unsigned long rlim_rtprio;
5158 if (!lock_task_sighand(p, &flags))
5160 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5161 unlock_task_sighand(p, &flags);
5163 /* can't set/change the rt policy */
5164 if (policy != p->policy && !rlim_rtprio)
5167 /* can't increase priority */
5168 if (param->sched_priority > p->rt_priority &&
5169 param->sched_priority > rlim_rtprio)
5173 * Like positive nice levels, dont allow tasks to
5174 * move out of SCHED_IDLE either:
5176 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5179 /* can't change other user's priorities */
5180 if ((current->euid != p->euid) &&
5181 (current->euid != p->uid))
5185 #ifdef CONFIG_RT_GROUP_SCHED
5187 * Do not allow realtime tasks into groups that have no runtime
5190 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5194 retval = security_task_setscheduler(p, policy, param);
5198 * make sure no PI-waiters arrive (or leave) while we are
5199 * changing the priority of the task:
5201 spin_lock_irqsave(&p->pi_lock, flags);
5203 * To be able to change p->policy safely, the apropriate
5204 * runqueue lock must be held.
5206 rq = __task_rq_lock(p);
5207 /* recheck policy now with rq lock held */
5208 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5209 policy = oldpolicy = -1;
5210 __task_rq_unlock(rq);
5211 spin_unlock_irqrestore(&p->pi_lock, flags);
5214 update_rq_clock(rq);
5215 on_rq = p->se.on_rq;
5216 running = task_current(rq, p);
5218 deactivate_task(rq, p, 0);
5220 p->sched_class->put_prev_task(rq, p);
5223 __setscheduler(rq, p, policy, param->sched_priority);
5226 p->sched_class->set_curr_task(rq);
5228 activate_task(rq, p, 0);
5230 check_class_changed(rq, p, prev_class, oldprio, running);
5232 __task_rq_unlock(rq);
5233 spin_unlock_irqrestore(&p->pi_lock, flags);
5235 rt_mutex_adjust_pi(p);
5239 EXPORT_SYMBOL_GPL(sched_setscheduler);
5242 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5244 struct sched_param lparam;
5245 struct task_struct *p;
5248 if (!param || pid < 0)
5250 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5255 p = find_process_by_pid(pid);
5257 retval = sched_setscheduler(p, policy, &lparam);
5264 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5265 * @pid: the pid in question.
5266 * @policy: new policy.
5267 * @param: structure containing the new RT priority.
5270 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5272 /* negative values for policy are not valid */
5276 return do_sched_setscheduler(pid, policy, param);
5280 * sys_sched_setparam - set/change the RT priority of a thread
5281 * @pid: the pid in question.
5282 * @param: structure containing the new RT priority.
5284 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5286 return do_sched_setscheduler(pid, -1, param);
5290 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5291 * @pid: the pid in question.
5293 asmlinkage long sys_sched_getscheduler(pid_t pid)
5295 struct task_struct *p;
5302 read_lock(&tasklist_lock);
5303 p = find_process_by_pid(pid);
5305 retval = security_task_getscheduler(p);
5309 read_unlock(&tasklist_lock);
5314 * sys_sched_getscheduler - get the RT priority of a thread
5315 * @pid: the pid in question.
5316 * @param: structure containing the RT priority.
5318 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5320 struct sched_param lp;
5321 struct task_struct *p;
5324 if (!param || pid < 0)
5327 read_lock(&tasklist_lock);
5328 p = find_process_by_pid(pid);
5333 retval = security_task_getscheduler(p);
5337 lp.sched_priority = p->rt_priority;
5338 read_unlock(&tasklist_lock);
5341 * This one might sleep, we cannot do it with a spinlock held ...
5343 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5348 read_unlock(&tasklist_lock);
5352 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5354 cpumask_t cpus_allowed;
5355 cpumask_t new_mask = *in_mask;
5356 struct task_struct *p;
5360 read_lock(&tasklist_lock);
5362 p = find_process_by_pid(pid);
5364 read_unlock(&tasklist_lock);
5370 * It is not safe to call set_cpus_allowed with the
5371 * tasklist_lock held. We will bump the task_struct's
5372 * usage count and then drop tasklist_lock.
5375 read_unlock(&tasklist_lock);
5378 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5379 !capable(CAP_SYS_NICE))
5382 retval = security_task_setscheduler(p, 0, NULL);
5386 cpuset_cpus_allowed(p, &cpus_allowed);
5387 cpus_and(new_mask, new_mask, cpus_allowed);
5389 retval = set_cpus_allowed_ptr(p, &new_mask);
5392 cpuset_cpus_allowed(p, &cpus_allowed);
5393 if (!cpus_subset(new_mask, cpus_allowed)) {
5395 * We must have raced with a concurrent cpuset
5396 * update. Just reset the cpus_allowed to the
5397 * cpuset's cpus_allowed
5399 new_mask = cpus_allowed;
5409 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5410 cpumask_t *new_mask)
5412 if (len < sizeof(cpumask_t)) {
5413 memset(new_mask, 0, sizeof(cpumask_t));
5414 } else if (len > sizeof(cpumask_t)) {
5415 len = sizeof(cpumask_t);
5417 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5421 * sys_sched_setaffinity - set the cpu affinity of a process
5422 * @pid: pid of the process
5423 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5424 * @user_mask_ptr: user-space pointer to the new cpu mask
5426 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5427 unsigned long __user *user_mask_ptr)
5432 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5436 return sched_setaffinity(pid, &new_mask);
5440 * Represents all cpu's present in the system
5441 * In systems capable of hotplug, this map could dynamically grow
5442 * as new cpu's are detected in the system via any platform specific
5443 * method, such as ACPI for e.g.
5446 cpumask_t cpu_present_map __read_mostly;
5447 EXPORT_SYMBOL(cpu_present_map);
5450 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5451 EXPORT_SYMBOL(cpu_online_map);
5453 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5454 EXPORT_SYMBOL(cpu_possible_map);
5457 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5459 struct task_struct *p;
5463 read_lock(&tasklist_lock);
5466 p = find_process_by_pid(pid);
5470 retval = security_task_getscheduler(p);
5474 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5477 read_unlock(&tasklist_lock);
5484 * sys_sched_getaffinity - get the cpu affinity of a process
5485 * @pid: pid of the process
5486 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5487 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5489 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5490 unsigned long __user *user_mask_ptr)
5495 if (len < sizeof(cpumask_t))
5498 ret = sched_getaffinity(pid, &mask);
5502 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5505 return sizeof(cpumask_t);
5509 * sys_sched_yield - yield the current processor to other threads.
5511 * This function yields the current CPU to other tasks. If there are no
5512 * other threads running on this CPU then this function will return.
5514 asmlinkage long sys_sched_yield(void)
5516 struct rq *rq = this_rq_lock();
5518 schedstat_inc(rq, yld_count);
5519 current->sched_class->yield_task(rq);
5522 * Since we are going to call schedule() anyway, there's
5523 * no need to preempt or enable interrupts:
5525 __release(rq->lock);
5526 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5527 _raw_spin_unlock(&rq->lock);
5528 preempt_enable_no_resched();
5535 static void __cond_resched(void)
5537 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5538 __might_sleep(__FILE__, __LINE__);
5541 * The BKS might be reacquired before we have dropped
5542 * PREEMPT_ACTIVE, which could trigger a second
5543 * cond_resched() call.
5546 add_preempt_count(PREEMPT_ACTIVE);
5548 sub_preempt_count(PREEMPT_ACTIVE);
5549 } while (need_resched());
5552 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5553 int __sched _cond_resched(void)
5555 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5556 system_state == SYSTEM_RUNNING) {
5562 EXPORT_SYMBOL(_cond_resched);
5566 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5567 * call schedule, and on return reacquire the lock.
5569 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5570 * operations here to prevent schedule() from being called twice (once via
5571 * spin_unlock(), once by hand).
5573 int cond_resched_lock(spinlock_t *lock)
5575 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5578 if (spin_needbreak(lock) || resched) {
5580 if (resched && need_resched())
5589 EXPORT_SYMBOL(cond_resched_lock);
5591 int __sched cond_resched_softirq(void)
5593 BUG_ON(!in_softirq());
5595 if (need_resched() && system_state == SYSTEM_RUNNING) {
5603 EXPORT_SYMBOL(cond_resched_softirq);
5606 * yield - yield the current processor to other threads.
5608 * This is a shortcut for kernel-space yielding - it marks the
5609 * thread runnable and calls sys_sched_yield().
5611 void __sched yield(void)
5613 set_current_state(TASK_RUNNING);
5616 EXPORT_SYMBOL(yield);
5619 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5620 * that process accounting knows that this is a task in IO wait state.
5622 * But don't do that if it is a deliberate, throttling IO wait (this task
5623 * has set its backing_dev_info: the queue against which it should throttle)
5625 void __sched io_schedule(void)
5627 struct rq *rq = &__raw_get_cpu_var(runqueues);
5629 delayacct_blkio_start();
5630 atomic_inc(&rq->nr_iowait);
5632 atomic_dec(&rq->nr_iowait);
5633 delayacct_blkio_end();
5635 EXPORT_SYMBOL(io_schedule);
5637 long __sched io_schedule_timeout(long timeout)
5639 struct rq *rq = &__raw_get_cpu_var(runqueues);
5642 delayacct_blkio_start();
5643 atomic_inc(&rq->nr_iowait);
5644 ret = schedule_timeout(timeout);
5645 atomic_dec(&rq->nr_iowait);
5646 delayacct_blkio_end();
5651 * sys_sched_get_priority_max - return maximum RT priority.
5652 * @policy: scheduling class.
5654 * this syscall returns the maximum rt_priority that can be used
5655 * by a given scheduling class.
5657 asmlinkage long sys_sched_get_priority_max(int policy)
5664 ret = MAX_USER_RT_PRIO-1;
5676 * sys_sched_get_priority_min - return minimum RT priority.
5677 * @policy: scheduling class.
5679 * this syscall returns the minimum rt_priority that can be used
5680 * by a given scheduling class.
5682 asmlinkage long sys_sched_get_priority_min(int policy)
5700 * sys_sched_rr_get_interval - return the default timeslice of a process.
5701 * @pid: pid of the process.
5702 * @interval: userspace pointer to the timeslice value.
5704 * this syscall writes the default timeslice value of a given process
5705 * into the user-space timespec buffer. A value of '0' means infinity.
5708 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5710 struct task_struct *p;
5711 unsigned int time_slice;
5719 read_lock(&tasklist_lock);
5720 p = find_process_by_pid(pid);
5724 retval = security_task_getscheduler(p);
5729 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5730 * tasks that are on an otherwise idle runqueue:
5733 if (p->policy == SCHED_RR) {
5734 time_slice = DEF_TIMESLICE;
5735 } else if (p->policy != SCHED_FIFO) {
5736 struct sched_entity *se = &p->se;
5737 unsigned long flags;
5740 rq = task_rq_lock(p, &flags);
5741 if (rq->cfs.load.weight)
5742 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5743 task_rq_unlock(rq, &flags);
5745 read_unlock(&tasklist_lock);
5746 jiffies_to_timespec(time_slice, &t);
5747 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5751 read_unlock(&tasklist_lock);
5755 static const char stat_nam[] = "RSDTtZX";
5757 void sched_show_task(struct task_struct *p)
5759 unsigned long free = 0;
5762 state = p->state ? __ffs(p->state) + 1 : 0;
5763 printk(KERN_INFO "%-13.13s %c", p->comm,
5764 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5765 #if BITS_PER_LONG == 32
5766 if (state == TASK_RUNNING)
5767 printk(KERN_CONT " running ");
5769 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5771 if (state == TASK_RUNNING)
5772 printk(KERN_CONT " running task ");
5774 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5776 #ifdef CONFIG_DEBUG_STACK_USAGE
5778 unsigned long *n = end_of_stack(p);
5781 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5784 printk(KERN_CONT "%5lu %5d %6d\n", free,
5785 task_pid_nr(p), task_pid_nr(p->real_parent));
5787 show_stack(p, NULL);
5790 void show_state_filter(unsigned long state_filter)
5792 struct task_struct *g, *p;
5794 #if BITS_PER_LONG == 32
5796 " task PC stack pid father\n");
5799 " task PC stack pid father\n");
5801 read_lock(&tasklist_lock);
5802 do_each_thread(g, p) {
5804 * reset the NMI-timeout, listing all files on a slow
5805 * console might take alot of time:
5807 touch_nmi_watchdog();
5808 if (!state_filter || (p->state & state_filter))
5810 } while_each_thread(g, p);
5812 touch_all_softlockup_watchdogs();
5814 #ifdef CONFIG_SCHED_DEBUG
5815 sysrq_sched_debug_show();
5817 read_unlock(&tasklist_lock);
5819 * Only show locks if all tasks are dumped:
5821 if (state_filter == -1)
5822 debug_show_all_locks();
5825 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5827 idle->sched_class = &idle_sched_class;
5831 * init_idle - set up an idle thread for a given CPU
5832 * @idle: task in question
5833 * @cpu: cpu the idle task belongs to
5835 * NOTE: this function does not set the idle thread's NEED_RESCHED
5836 * flag, to make booting more robust.
5838 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5840 struct rq *rq = cpu_rq(cpu);
5841 unsigned long flags;
5844 idle->se.exec_start = sched_clock();
5846 idle->prio = idle->normal_prio = MAX_PRIO;
5847 idle->cpus_allowed = cpumask_of_cpu(cpu);
5848 __set_task_cpu(idle, cpu);
5850 spin_lock_irqsave(&rq->lock, flags);
5851 rq->curr = rq->idle = idle;
5852 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5855 spin_unlock_irqrestore(&rq->lock, flags);
5857 /* Set the preempt count _outside_ the spinlocks! */
5858 task_thread_info(idle)->preempt_count = 0;
5861 * The idle tasks have their own, simple scheduling class:
5863 idle->sched_class = &idle_sched_class;
5867 * In a system that switches off the HZ timer nohz_cpu_mask
5868 * indicates which cpus entered this state. This is used
5869 * in the rcu update to wait only for active cpus. For system
5870 * which do not switch off the HZ timer nohz_cpu_mask should
5871 * always be CPU_MASK_NONE.
5873 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5876 * Increase the granularity value when there are more CPUs,
5877 * because with more CPUs the 'effective latency' as visible
5878 * to users decreases. But the relationship is not linear,
5879 * so pick a second-best guess by going with the log2 of the
5882 * This idea comes from the SD scheduler of Con Kolivas:
5884 static inline void sched_init_granularity(void)
5886 unsigned int factor = 1 + ilog2(num_online_cpus());
5887 const unsigned long limit = 200000000;
5889 sysctl_sched_min_granularity *= factor;
5890 if (sysctl_sched_min_granularity > limit)
5891 sysctl_sched_min_granularity = limit;
5893 sysctl_sched_latency *= factor;
5894 if (sysctl_sched_latency > limit)
5895 sysctl_sched_latency = limit;
5897 sysctl_sched_wakeup_granularity *= factor;
5902 * This is how migration works:
5904 * 1) we queue a struct migration_req structure in the source CPU's
5905 * runqueue and wake up that CPU's migration thread.
5906 * 2) we down() the locked semaphore => thread blocks.
5907 * 3) migration thread wakes up (implicitly it forces the migrated
5908 * thread off the CPU)
5909 * 4) it gets the migration request and checks whether the migrated
5910 * task is still in the wrong runqueue.
5911 * 5) if it's in the wrong runqueue then the migration thread removes
5912 * it and puts it into the right queue.
5913 * 6) migration thread up()s the semaphore.
5914 * 7) we wake up and the migration is done.
5918 * Change a given task's CPU affinity. Migrate the thread to a
5919 * proper CPU and schedule it away if the CPU it's executing on
5920 * is removed from the allowed bitmask.
5922 * NOTE: the caller must have a valid reference to the task, the
5923 * task must not exit() & deallocate itself prematurely. The
5924 * call is not atomic; no spinlocks may be held.
5926 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5928 struct migration_req req;
5929 unsigned long flags;
5933 rq = task_rq_lock(p, &flags);
5934 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5939 if (p->sched_class->set_cpus_allowed)
5940 p->sched_class->set_cpus_allowed(p, new_mask);
5942 p->cpus_allowed = *new_mask;
5943 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5946 /* Can the task run on the task's current CPU? If so, we're done */
5947 if (cpu_isset(task_cpu(p), *new_mask))
5950 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5951 /* Need help from migration thread: drop lock and wait. */
5952 task_rq_unlock(rq, &flags);
5953 wake_up_process(rq->migration_thread);
5954 wait_for_completion(&req.done);
5955 tlb_migrate_finish(p->mm);
5959 task_rq_unlock(rq, &flags);
5963 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5966 * Move (not current) task off this cpu, onto dest cpu. We're doing
5967 * this because either it can't run here any more (set_cpus_allowed()
5968 * away from this CPU, or CPU going down), or because we're
5969 * attempting to rebalance this task on exec (sched_exec).
5971 * So we race with normal scheduler movements, but that's OK, as long
5972 * as the task is no longer on this CPU.
5974 * Returns non-zero if task was successfully migrated.
5976 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5978 struct rq *rq_dest, *rq_src;
5981 if (unlikely(cpu_is_offline(dest_cpu)))
5984 rq_src = cpu_rq(src_cpu);
5985 rq_dest = cpu_rq(dest_cpu);
5987 double_rq_lock(rq_src, rq_dest);
5988 /* Already moved. */
5989 if (task_cpu(p) != src_cpu)
5991 /* Affinity changed (again). */
5992 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5995 on_rq = p->se.on_rq;
5997 deactivate_task(rq_src, p, 0);
5999 set_task_cpu(p, dest_cpu);
6001 activate_task(rq_dest, p, 0);
6002 check_preempt_curr(rq_dest, p);
6006 double_rq_unlock(rq_src, rq_dest);
6011 * migration_thread - this is a highprio system thread that performs
6012 * thread migration by bumping thread off CPU then 'pushing' onto
6015 static int migration_thread(void *data)
6017 int cpu = (long)data;
6021 BUG_ON(rq->migration_thread != current);
6023 set_current_state(TASK_INTERRUPTIBLE);
6024 while (!kthread_should_stop()) {
6025 struct migration_req *req;
6026 struct list_head *head;
6028 spin_lock_irq(&rq->lock);
6030 if (cpu_is_offline(cpu)) {
6031 spin_unlock_irq(&rq->lock);
6035 if (rq->active_balance) {
6036 active_load_balance(rq, cpu);
6037 rq->active_balance = 0;
6040 head = &rq->migration_queue;
6042 if (list_empty(head)) {
6043 spin_unlock_irq(&rq->lock);
6045 set_current_state(TASK_INTERRUPTIBLE);
6048 req = list_entry(head->next, struct migration_req, list);
6049 list_del_init(head->next);
6051 spin_unlock(&rq->lock);
6052 __migrate_task(req->task, cpu, req->dest_cpu);
6055 complete(&req->done);
6057 __set_current_state(TASK_RUNNING);
6061 /* Wait for kthread_stop */
6062 set_current_state(TASK_INTERRUPTIBLE);
6063 while (!kthread_should_stop()) {
6065 set_current_state(TASK_INTERRUPTIBLE);
6067 __set_current_state(TASK_RUNNING);
6071 #ifdef CONFIG_HOTPLUG_CPU
6073 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6077 local_irq_disable();
6078 ret = __migrate_task(p, src_cpu, dest_cpu);
6084 * Figure out where task on dead CPU should go, use force if necessary.
6085 * NOTE: interrupts should be disabled by the caller
6087 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6089 unsigned long flags;
6096 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6097 cpus_and(mask, mask, p->cpus_allowed);
6098 dest_cpu = any_online_cpu(mask);
6100 /* On any allowed CPU? */
6101 if (dest_cpu >= nr_cpu_ids)
6102 dest_cpu = any_online_cpu(p->cpus_allowed);
6104 /* No more Mr. Nice Guy. */
6105 if (dest_cpu >= nr_cpu_ids) {
6106 cpumask_t cpus_allowed;
6108 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6110 * Try to stay on the same cpuset, where the
6111 * current cpuset may be a subset of all cpus.
6112 * The cpuset_cpus_allowed_locked() variant of
6113 * cpuset_cpus_allowed() will not block. It must be
6114 * called within calls to cpuset_lock/cpuset_unlock.
6116 rq = task_rq_lock(p, &flags);
6117 p->cpus_allowed = cpus_allowed;
6118 dest_cpu = any_online_cpu(p->cpus_allowed);
6119 task_rq_unlock(rq, &flags);
6122 * Don't tell them about moving exiting tasks or
6123 * kernel threads (both mm NULL), since they never
6126 if (p->mm && printk_ratelimit()) {
6127 printk(KERN_INFO "process %d (%s) no "
6128 "longer affine to cpu%d\n",
6129 task_pid_nr(p), p->comm, dead_cpu);
6132 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6136 * While a dead CPU has no uninterruptible tasks queued at this point,
6137 * it might still have a nonzero ->nr_uninterruptible counter, because
6138 * for performance reasons the counter is not stricly tracking tasks to
6139 * their home CPUs. So we just add the counter to another CPU's counter,
6140 * to keep the global sum constant after CPU-down:
6142 static void migrate_nr_uninterruptible(struct rq *rq_src)
6144 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6145 unsigned long flags;
6147 local_irq_save(flags);
6148 double_rq_lock(rq_src, rq_dest);
6149 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6150 rq_src->nr_uninterruptible = 0;
6151 double_rq_unlock(rq_src, rq_dest);
6152 local_irq_restore(flags);
6155 /* Run through task list and migrate tasks from the dead cpu. */
6156 static void migrate_live_tasks(int src_cpu)
6158 struct task_struct *p, *t;
6160 read_lock(&tasklist_lock);
6162 do_each_thread(t, p) {
6166 if (task_cpu(p) == src_cpu)
6167 move_task_off_dead_cpu(src_cpu, p);
6168 } while_each_thread(t, p);
6170 read_unlock(&tasklist_lock);
6174 * Schedules idle task to be the next runnable task on current CPU.
6175 * It does so by boosting its priority to highest possible.
6176 * Used by CPU offline code.
6178 void sched_idle_next(void)
6180 int this_cpu = smp_processor_id();
6181 struct rq *rq = cpu_rq(this_cpu);
6182 struct task_struct *p = rq->idle;
6183 unsigned long flags;
6185 /* cpu has to be offline */
6186 BUG_ON(cpu_online(this_cpu));
6189 * Strictly not necessary since rest of the CPUs are stopped by now
6190 * and interrupts disabled on the current cpu.
6192 spin_lock_irqsave(&rq->lock, flags);
6194 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6196 update_rq_clock(rq);
6197 activate_task(rq, p, 0);
6199 spin_unlock_irqrestore(&rq->lock, flags);
6203 * Ensures that the idle task is using init_mm right before its cpu goes
6206 void idle_task_exit(void)
6208 struct mm_struct *mm = current->active_mm;
6210 BUG_ON(cpu_online(smp_processor_id()));
6213 switch_mm(mm, &init_mm, current);
6217 /* called under rq->lock with disabled interrupts */
6218 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6220 struct rq *rq = cpu_rq(dead_cpu);
6222 /* Must be exiting, otherwise would be on tasklist. */
6223 BUG_ON(!p->exit_state);
6225 /* Cannot have done final schedule yet: would have vanished. */
6226 BUG_ON(p->state == TASK_DEAD);
6231 * Drop lock around migration; if someone else moves it,
6232 * that's OK. No task can be added to this CPU, so iteration is
6235 spin_unlock_irq(&rq->lock);
6236 move_task_off_dead_cpu(dead_cpu, p);
6237 spin_lock_irq(&rq->lock);
6242 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6243 static void migrate_dead_tasks(unsigned int dead_cpu)
6245 struct rq *rq = cpu_rq(dead_cpu);
6246 struct task_struct *next;
6249 if (!rq->nr_running)
6251 update_rq_clock(rq);
6252 next = pick_next_task(rq, rq->curr);
6255 migrate_dead(dead_cpu, next);
6259 #endif /* CONFIG_HOTPLUG_CPU */
6261 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6263 static struct ctl_table sd_ctl_dir[] = {
6265 .procname = "sched_domain",
6271 static struct ctl_table sd_ctl_root[] = {
6273 .ctl_name = CTL_KERN,
6274 .procname = "kernel",
6276 .child = sd_ctl_dir,
6281 static struct ctl_table *sd_alloc_ctl_entry(int n)
6283 struct ctl_table *entry =
6284 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6289 static void sd_free_ctl_entry(struct ctl_table **tablep)
6291 struct ctl_table *entry;
6294 * In the intermediate directories, both the child directory and
6295 * procname are dynamically allocated and could fail but the mode
6296 * will always be set. In the lowest directory the names are
6297 * static strings and all have proc handlers.
6299 for (entry = *tablep; entry->mode; entry++) {
6301 sd_free_ctl_entry(&entry->child);
6302 if (entry->proc_handler == NULL)
6303 kfree(entry->procname);
6311 set_table_entry(struct ctl_table *entry,
6312 const char *procname, void *data, int maxlen,
6313 mode_t mode, proc_handler *proc_handler)
6315 entry->procname = procname;
6317 entry->maxlen = maxlen;
6319 entry->proc_handler = proc_handler;
6322 static struct ctl_table *
6323 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6325 struct ctl_table *table = sd_alloc_ctl_entry(12);
6330 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6331 sizeof(long), 0644, proc_doulongvec_minmax);
6332 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6333 sizeof(long), 0644, proc_doulongvec_minmax);
6334 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6335 sizeof(int), 0644, proc_dointvec_minmax);
6336 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6337 sizeof(int), 0644, proc_dointvec_minmax);
6338 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6339 sizeof(int), 0644, proc_dointvec_minmax);
6340 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6341 sizeof(int), 0644, proc_dointvec_minmax);
6342 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6343 sizeof(int), 0644, proc_dointvec_minmax);
6344 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6345 sizeof(int), 0644, proc_dointvec_minmax);
6346 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6347 sizeof(int), 0644, proc_dointvec_minmax);
6348 set_table_entry(&table[9], "cache_nice_tries",
6349 &sd->cache_nice_tries,
6350 sizeof(int), 0644, proc_dointvec_minmax);
6351 set_table_entry(&table[10], "flags", &sd->flags,
6352 sizeof(int), 0644, proc_dointvec_minmax);
6353 /* &table[11] is terminator */
6358 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6360 struct ctl_table *entry, *table;
6361 struct sched_domain *sd;
6362 int domain_num = 0, i;
6365 for_each_domain(cpu, sd)
6367 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6372 for_each_domain(cpu, sd) {
6373 snprintf(buf, 32, "domain%d", i);
6374 entry->procname = kstrdup(buf, GFP_KERNEL);
6376 entry->child = sd_alloc_ctl_domain_table(sd);
6383 static struct ctl_table_header *sd_sysctl_header;
6384 static void register_sched_domain_sysctl(void)
6386 int i, cpu_num = num_online_cpus();
6387 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6390 WARN_ON(sd_ctl_dir[0].child);
6391 sd_ctl_dir[0].child = entry;
6396 for_each_online_cpu(i) {
6397 snprintf(buf, 32, "cpu%d", i);
6398 entry->procname = kstrdup(buf, GFP_KERNEL);
6400 entry->child = sd_alloc_ctl_cpu_table(i);
6404 WARN_ON(sd_sysctl_header);
6405 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6408 /* may be called multiple times per register */
6409 static void unregister_sched_domain_sysctl(void)
6411 if (sd_sysctl_header)
6412 unregister_sysctl_table(sd_sysctl_header);
6413 sd_sysctl_header = NULL;
6414 if (sd_ctl_dir[0].child)
6415 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6418 static void register_sched_domain_sysctl(void)
6421 static void unregister_sched_domain_sysctl(void)
6427 * migration_call - callback that gets triggered when a CPU is added.
6428 * Here we can start up the necessary migration thread for the new CPU.
6430 static int __cpuinit
6431 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6433 struct task_struct *p;
6434 int cpu = (long)hcpu;
6435 unsigned long flags;
6440 case CPU_UP_PREPARE:
6441 case CPU_UP_PREPARE_FROZEN:
6442 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6445 kthread_bind(p, cpu);
6446 /* Must be high prio: stop_machine expects to yield to it. */
6447 rq = task_rq_lock(p, &flags);
6448 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6449 task_rq_unlock(rq, &flags);
6450 cpu_rq(cpu)->migration_thread = p;
6454 case CPU_ONLINE_FROZEN:
6455 /* Strictly unnecessary, as first user will wake it. */
6456 wake_up_process(cpu_rq(cpu)->migration_thread);
6458 /* Update our root-domain */
6460 spin_lock_irqsave(&rq->lock, flags);
6462 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6463 cpu_set(cpu, rq->rd->online);
6465 spin_unlock_irqrestore(&rq->lock, flags);
6468 #ifdef CONFIG_HOTPLUG_CPU
6469 case CPU_UP_CANCELED:
6470 case CPU_UP_CANCELED_FROZEN:
6471 if (!cpu_rq(cpu)->migration_thread)
6473 /* Unbind it from offline cpu so it can run. Fall thru. */
6474 kthread_bind(cpu_rq(cpu)->migration_thread,
6475 any_online_cpu(cpu_online_map));
6476 kthread_stop(cpu_rq(cpu)->migration_thread);
6477 cpu_rq(cpu)->migration_thread = NULL;
6481 case CPU_DEAD_FROZEN:
6482 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6483 migrate_live_tasks(cpu);
6485 kthread_stop(rq->migration_thread);
6486 rq->migration_thread = NULL;
6487 /* Idle task back to normal (off runqueue, low prio) */
6488 spin_lock_irq(&rq->lock);
6489 update_rq_clock(rq);
6490 deactivate_task(rq, rq->idle, 0);
6491 rq->idle->static_prio = MAX_PRIO;
6492 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6493 rq->idle->sched_class = &idle_sched_class;
6494 migrate_dead_tasks(cpu);
6495 spin_unlock_irq(&rq->lock);
6497 migrate_nr_uninterruptible(rq);
6498 BUG_ON(rq->nr_running != 0);
6501 * No need to migrate the tasks: it was best-effort if
6502 * they didn't take sched_hotcpu_mutex. Just wake up
6505 spin_lock_irq(&rq->lock);
6506 while (!list_empty(&rq->migration_queue)) {
6507 struct migration_req *req;
6509 req = list_entry(rq->migration_queue.next,
6510 struct migration_req, list);
6511 list_del_init(&req->list);
6512 complete(&req->done);
6514 spin_unlock_irq(&rq->lock);
6518 case CPU_DYING_FROZEN:
6519 /* Update our root-domain */
6521 spin_lock_irqsave(&rq->lock, flags);
6523 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6524 cpu_clear(cpu, rq->rd->online);
6526 spin_unlock_irqrestore(&rq->lock, flags);
6533 /* Register at highest priority so that task migration (migrate_all_tasks)
6534 * happens before everything else.
6536 static struct notifier_block __cpuinitdata migration_notifier = {
6537 .notifier_call = migration_call,
6541 void __init migration_init(void)
6543 void *cpu = (void *)(long)smp_processor_id();
6546 /* Start one for the boot CPU: */
6547 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6548 BUG_ON(err == NOTIFY_BAD);
6549 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6550 register_cpu_notifier(&migration_notifier);
6556 #ifdef CONFIG_SCHED_DEBUG
6558 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6559 cpumask_t *groupmask)
6561 struct sched_group *group = sd->groups;
6564 cpulist_scnprintf(str, sizeof(str), sd->span);
6565 cpus_clear(*groupmask);
6567 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6569 if (!(sd->flags & SD_LOAD_BALANCE)) {
6570 printk("does not load-balance\n");
6572 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6577 printk(KERN_CONT "span %s\n", str);
6579 if (!cpu_isset(cpu, sd->span)) {
6580 printk(KERN_ERR "ERROR: domain->span does not contain "
6583 if (!cpu_isset(cpu, group->cpumask)) {
6584 printk(KERN_ERR "ERROR: domain->groups does not contain"
6588 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6592 printk(KERN_ERR "ERROR: group is NULL\n");
6596 if (!group->__cpu_power) {
6597 printk(KERN_CONT "\n");
6598 printk(KERN_ERR "ERROR: domain->cpu_power not "
6603 if (!cpus_weight(group->cpumask)) {
6604 printk(KERN_CONT "\n");
6605 printk(KERN_ERR "ERROR: empty group\n");
6609 if (cpus_intersects(*groupmask, group->cpumask)) {
6610 printk(KERN_CONT "\n");
6611 printk(KERN_ERR "ERROR: repeated CPUs\n");
6615 cpus_or(*groupmask, *groupmask, group->cpumask);
6617 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6618 printk(KERN_CONT " %s", str);
6620 group = group->next;
6621 } while (group != sd->groups);
6622 printk(KERN_CONT "\n");
6624 if (!cpus_equal(sd->span, *groupmask))
6625 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6627 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6628 printk(KERN_ERR "ERROR: parent span is not a superset "
6629 "of domain->span\n");
6633 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6635 cpumask_t *groupmask;
6639 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6643 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6645 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6647 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6652 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6662 # define sched_domain_debug(sd, cpu) do { } while (0)
6665 static int sd_degenerate(struct sched_domain *sd)
6667 if (cpus_weight(sd->span) == 1)
6670 /* Following flags need at least 2 groups */
6671 if (sd->flags & (SD_LOAD_BALANCE |
6672 SD_BALANCE_NEWIDLE |
6676 SD_SHARE_PKG_RESOURCES)) {
6677 if (sd->groups != sd->groups->next)
6681 /* Following flags don't use groups */
6682 if (sd->flags & (SD_WAKE_IDLE |
6691 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6693 unsigned long cflags = sd->flags, pflags = parent->flags;
6695 if (sd_degenerate(parent))
6698 if (!cpus_equal(sd->span, parent->span))
6701 /* Does parent contain flags not in child? */
6702 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6703 if (cflags & SD_WAKE_AFFINE)
6704 pflags &= ~SD_WAKE_BALANCE;
6705 /* Flags needing groups don't count if only 1 group in parent */
6706 if (parent->groups == parent->groups->next) {
6707 pflags &= ~(SD_LOAD_BALANCE |
6708 SD_BALANCE_NEWIDLE |
6712 SD_SHARE_PKG_RESOURCES);
6714 if (~cflags & pflags)
6720 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6722 unsigned long flags;
6723 const struct sched_class *class;
6725 spin_lock_irqsave(&rq->lock, flags);
6728 struct root_domain *old_rd = rq->rd;
6730 for (class = sched_class_highest; class; class = class->next) {
6731 if (class->leave_domain)
6732 class->leave_domain(rq);
6735 cpu_clear(rq->cpu, old_rd->span);
6736 cpu_clear(rq->cpu, old_rd->online);
6738 if (atomic_dec_and_test(&old_rd->refcount))
6742 atomic_inc(&rd->refcount);
6745 cpu_set(rq->cpu, rd->span);
6746 if (cpu_isset(rq->cpu, cpu_online_map))
6747 cpu_set(rq->cpu, rd->online);
6749 for (class = sched_class_highest; class; class = class->next) {
6750 if (class->join_domain)
6751 class->join_domain(rq);
6754 spin_unlock_irqrestore(&rq->lock, flags);
6757 static void init_rootdomain(struct root_domain *rd)
6759 memset(rd, 0, sizeof(*rd));
6761 cpus_clear(rd->span);
6762 cpus_clear(rd->online);
6765 static void init_defrootdomain(void)
6767 init_rootdomain(&def_root_domain);
6768 atomic_set(&def_root_domain.refcount, 1);
6771 static struct root_domain *alloc_rootdomain(void)
6773 struct root_domain *rd;
6775 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6779 init_rootdomain(rd);
6785 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6786 * hold the hotplug lock.
6789 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6791 struct rq *rq = cpu_rq(cpu);
6792 struct sched_domain *tmp;
6794 /* Remove the sched domains which do not contribute to scheduling. */
6795 for (tmp = sd; tmp; tmp = tmp->parent) {
6796 struct sched_domain *parent = tmp->parent;
6799 if (sd_parent_degenerate(tmp, parent)) {
6800 tmp->parent = parent->parent;
6802 parent->parent->child = tmp;
6806 if (sd && sd_degenerate(sd)) {
6812 sched_domain_debug(sd, cpu);
6814 rq_attach_root(rq, rd);
6815 rcu_assign_pointer(rq->sd, sd);
6818 /* cpus with isolated domains */
6819 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6821 /* Setup the mask of cpus configured for isolated domains */
6822 static int __init isolated_cpu_setup(char *str)
6824 int ints[NR_CPUS], i;
6826 str = get_options(str, ARRAY_SIZE(ints), ints);
6827 cpus_clear(cpu_isolated_map);
6828 for (i = 1; i <= ints[0]; i++)
6829 if (ints[i] < NR_CPUS)
6830 cpu_set(ints[i], cpu_isolated_map);
6834 __setup("isolcpus=", isolated_cpu_setup);
6837 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6838 * to a function which identifies what group(along with sched group) a CPU
6839 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6840 * (due to the fact that we keep track of groups covered with a cpumask_t).
6842 * init_sched_build_groups will build a circular linked list of the groups
6843 * covered by the given span, and will set each group's ->cpumask correctly,
6844 * and ->cpu_power to 0.
6847 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6848 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6849 struct sched_group **sg,
6850 cpumask_t *tmpmask),
6851 cpumask_t *covered, cpumask_t *tmpmask)
6853 struct sched_group *first = NULL, *last = NULL;
6856 cpus_clear(*covered);
6858 for_each_cpu_mask(i, *span) {
6859 struct sched_group *sg;
6860 int group = group_fn(i, cpu_map, &sg, tmpmask);
6863 if (cpu_isset(i, *covered))
6866 cpus_clear(sg->cpumask);
6867 sg->__cpu_power = 0;
6869 for_each_cpu_mask(j, *span) {
6870 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6873 cpu_set(j, *covered);
6874 cpu_set(j, sg->cpumask);
6885 #define SD_NODES_PER_DOMAIN 16
6890 * find_next_best_node - find the next node to include in a sched_domain
6891 * @node: node whose sched_domain we're building
6892 * @used_nodes: nodes already in the sched_domain
6894 * Find the next node to include in a given scheduling domain. Simply
6895 * finds the closest node not already in the @used_nodes map.
6897 * Should use nodemask_t.
6899 static int find_next_best_node(int node, nodemask_t *used_nodes)
6901 int i, n, val, min_val, best_node = 0;
6905 for (i = 0; i < MAX_NUMNODES; i++) {
6906 /* Start at @node */
6907 n = (node + i) % MAX_NUMNODES;
6909 if (!nr_cpus_node(n))
6912 /* Skip already used nodes */
6913 if (node_isset(n, *used_nodes))
6916 /* Simple min distance search */
6917 val = node_distance(node, n);
6919 if (val < min_val) {
6925 node_set(best_node, *used_nodes);
6930 * sched_domain_node_span - get a cpumask for a node's sched_domain
6931 * @node: node whose cpumask we're constructing
6933 * Given a node, construct a good cpumask for its sched_domain to span. It
6934 * should be one that prevents unnecessary balancing, but also spreads tasks
6937 static void sched_domain_node_span(int node, cpumask_t *span)
6939 nodemask_t used_nodes;
6940 node_to_cpumask_ptr(nodemask, node);
6944 nodes_clear(used_nodes);
6946 cpus_or(*span, *span, *nodemask);
6947 node_set(node, used_nodes);
6949 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6950 int next_node = find_next_best_node(node, &used_nodes);
6952 node_to_cpumask_ptr_next(nodemask, next_node);
6953 cpus_or(*span, *span, *nodemask);
6958 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6961 * SMT sched-domains:
6963 #ifdef CONFIG_SCHED_SMT
6964 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6965 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6968 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6972 *sg = &per_cpu(sched_group_cpus, cpu);
6978 * multi-core sched-domains:
6980 #ifdef CONFIG_SCHED_MC
6981 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6982 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6985 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6987 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6992 *mask = per_cpu(cpu_sibling_map, cpu);
6993 cpus_and(*mask, *mask, *cpu_map);
6994 group = first_cpu(*mask);
6996 *sg = &per_cpu(sched_group_core, group);
6999 #elif defined(CONFIG_SCHED_MC)
7001 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7005 *sg = &per_cpu(sched_group_core, cpu);
7010 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7011 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7014 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7018 #ifdef CONFIG_SCHED_MC
7019 *mask = cpu_coregroup_map(cpu);
7020 cpus_and(*mask, *mask, *cpu_map);
7021 group = first_cpu(*mask);
7022 #elif defined(CONFIG_SCHED_SMT)
7023 *mask = per_cpu(cpu_sibling_map, cpu);
7024 cpus_and(*mask, *mask, *cpu_map);
7025 group = first_cpu(*mask);
7030 *sg = &per_cpu(sched_group_phys, group);
7036 * The init_sched_build_groups can't handle what we want to do with node
7037 * groups, so roll our own. Now each node has its own list of groups which
7038 * gets dynamically allocated.
7040 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7041 static struct sched_group ***sched_group_nodes_bycpu;
7043 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7044 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7046 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7047 struct sched_group **sg, cpumask_t *nodemask)
7051 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7052 cpus_and(*nodemask, *nodemask, *cpu_map);
7053 group = first_cpu(*nodemask);
7056 *sg = &per_cpu(sched_group_allnodes, group);
7060 static void init_numa_sched_groups_power(struct sched_group *group_head)
7062 struct sched_group *sg = group_head;
7068 for_each_cpu_mask(j, sg->cpumask) {
7069 struct sched_domain *sd;
7071 sd = &per_cpu(phys_domains, j);
7072 if (j != first_cpu(sd->groups->cpumask)) {
7074 * Only add "power" once for each
7080 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7083 } while (sg != group_head);
7088 /* Free memory allocated for various sched_group structures */
7089 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7093 for_each_cpu_mask(cpu, *cpu_map) {
7094 struct sched_group **sched_group_nodes
7095 = sched_group_nodes_bycpu[cpu];
7097 if (!sched_group_nodes)
7100 for (i = 0; i < MAX_NUMNODES; i++) {
7101 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7103 *nodemask = node_to_cpumask(i);
7104 cpus_and(*nodemask, *nodemask, *cpu_map);
7105 if (cpus_empty(*nodemask))
7115 if (oldsg != sched_group_nodes[i])
7118 kfree(sched_group_nodes);
7119 sched_group_nodes_bycpu[cpu] = NULL;
7123 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7129 * Initialize sched groups cpu_power.
7131 * cpu_power indicates the capacity of sched group, which is used while
7132 * distributing the load between different sched groups in a sched domain.
7133 * Typically cpu_power for all the groups in a sched domain will be same unless
7134 * there are asymmetries in the topology. If there are asymmetries, group
7135 * having more cpu_power will pickup more load compared to the group having
7138 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7139 * the maximum number of tasks a group can handle in the presence of other idle
7140 * or lightly loaded groups in the same sched domain.
7142 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7144 struct sched_domain *child;
7145 struct sched_group *group;
7147 WARN_ON(!sd || !sd->groups);
7149 if (cpu != first_cpu(sd->groups->cpumask))
7154 sd->groups->__cpu_power = 0;
7157 * For perf policy, if the groups in child domain share resources
7158 * (for example cores sharing some portions of the cache hierarchy
7159 * or SMT), then set this domain groups cpu_power such that each group
7160 * can handle only one task, when there are other idle groups in the
7161 * same sched domain.
7163 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7165 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7166 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7171 * add cpu_power of each child group to this groups cpu_power
7173 group = child->groups;
7175 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7176 group = group->next;
7177 } while (group != child->groups);
7181 * Initializers for schedule domains
7182 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7185 #define SD_INIT(sd, type) sd_init_##type(sd)
7186 #define SD_INIT_FUNC(type) \
7187 static noinline void sd_init_##type(struct sched_domain *sd) \
7189 memset(sd, 0, sizeof(*sd)); \
7190 *sd = SD_##type##_INIT; \
7191 sd->level = SD_LV_##type; \
7196 SD_INIT_FUNC(ALLNODES)
7199 #ifdef CONFIG_SCHED_SMT
7200 SD_INIT_FUNC(SIBLING)
7202 #ifdef CONFIG_SCHED_MC
7207 * To minimize stack usage kmalloc room for cpumasks and share the
7208 * space as the usage in build_sched_domains() dictates. Used only
7209 * if the amount of space is significant.
7212 cpumask_t tmpmask; /* make this one first */
7215 cpumask_t this_sibling_map;
7216 cpumask_t this_core_map;
7218 cpumask_t send_covered;
7221 cpumask_t domainspan;
7223 cpumask_t notcovered;
7228 #define SCHED_CPUMASK_ALLOC 1
7229 #define SCHED_CPUMASK_FREE(v) kfree(v)
7230 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7232 #define SCHED_CPUMASK_ALLOC 0
7233 #define SCHED_CPUMASK_FREE(v)
7234 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7237 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7238 ((unsigned long)(a) + offsetof(struct allmasks, v))
7240 static int default_relax_domain_level = -1;
7242 static int __init setup_relax_domain_level(char *str)
7244 default_relax_domain_level = simple_strtoul(str, NULL, 0);
7247 __setup("relax_domain_level=", setup_relax_domain_level);
7249 static void set_domain_attribute(struct sched_domain *sd,
7250 struct sched_domain_attr *attr)
7254 if (!attr || attr->relax_domain_level < 0) {
7255 if (default_relax_domain_level < 0)
7258 request = default_relax_domain_level;
7260 request = attr->relax_domain_level;
7261 if (request < sd->level) {
7262 /* turn off idle balance on this domain */
7263 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7265 /* turn on idle balance on this domain */
7266 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7271 * Build sched domains for a given set of cpus and attach the sched domains
7272 * to the individual cpus
7274 static int __build_sched_domains(const cpumask_t *cpu_map,
7275 struct sched_domain_attr *attr)
7278 struct root_domain *rd;
7279 SCHED_CPUMASK_DECLARE(allmasks);
7282 struct sched_group **sched_group_nodes = NULL;
7283 int sd_allnodes = 0;
7286 * Allocate the per-node list of sched groups
7288 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7290 if (!sched_group_nodes) {
7291 printk(KERN_WARNING "Can not alloc sched group node list\n");
7296 rd = alloc_rootdomain();
7298 printk(KERN_WARNING "Cannot alloc root domain\n");
7300 kfree(sched_group_nodes);
7305 #if SCHED_CPUMASK_ALLOC
7306 /* get space for all scratch cpumask variables */
7307 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7309 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7312 kfree(sched_group_nodes);
7317 tmpmask = (cpumask_t *)allmasks;
7321 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7325 * Set up domains for cpus specified by the cpu_map.
7327 for_each_cpu_mask(i, *cpu_map) {
7328 struct sched_domain *sd = NULL, *p;
7329 SCHED_CPUMASK_VAR(nodemask, allmasks);
7331 *nodemask = node_to_cpumask(cpu_to_node(i));
7332 cpus_and(*nodemask, *nodemask, *cpu_map);
7335 if (cpus_weight(*cpu_map) >
7336 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7337 sd = &per_cpu(allnodes_domains, i);
7338 SD_INIT(sd, ALLNODES);
7339 set_domain_attribute(sd, attr);
7340 sd->span = *cpu_map;
7341 sd->first_cpu = first_cpu(sd->span);
7342 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7348 sd = &per_cpu(node_domains, i);
7350 set_domain_attribute(sd, attr);
7351 sched_domain_node_span(cpu_to_node(i), &sd->span);
7352 sd->first_cpu = first_cpu(sd->span);
7356 cpus_and(sd->span, sd->span, *cpu_map);
7360 sd = &per_cpu(phys_domains, i);
7362 set_domain_attribute(sd, attr);
7363 sd->span = *nodemask;
7364 sd->first_cpu = first_cpu(sd->span);
7368 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7370 #ifdef CONFIG_SCHED_MC
7372 sd = &per_cpu(core_domains, i);
7374 set_domain_attribute(sd, attr);
7375 sd->span = cpu_coregroup_map(i);
7376 sd->first_cpu = first_cpu(sd->span);
7377 cpus_and(sd->span, sd->span, *cpu_map);
7380 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7383 #ifdef CONFIG_SCHED_SMT
7385 sd = &per_cpu(cpu_domains, i);
7386 SD_INIT(sd, SIBLING);
7387 set_domain_attribute(sd, attr);
7388 sd->span = per_cpu(cpu_sibling_map, i);
7389 sd->first_cpu = first_cpu(sd->span);
7390 cpus_and(sd->span, sd->span, *cpu_map);
7393 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7397 #ifdef CONFIG_SCHED_SMT
7398 /* Set up CPU (sibling) groups */
7399 for_each_cpu_mask(i, *cpu_map) {
7400 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7401 SCHED_CPUMASK_VAR(send_covered, allmasks);
7403 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7404 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7405 if (i != first_cpu(*this_sibling_map))
7408 init_sched_build_groups(this_sibling_map, cpu_map,
7410 send_covered, tmpmask);
7414 #ifdef CONFIG_SCHED_MC
7415 /* Set up multi-core groups */
7416 for_each_cpu_mask(i, *cpu_map) {
7417 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7418 SCHED_CPUMASK_VAR(send_covered, allmasks);
7420 *this_core_map = cpu_coregroup_map(i);
7421 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7422 if (i != first_cpu(*this_core_map))
7425 init_sched_build_groups(this_core_map, cpu_map,
7427 send_covered, tmpmask);
7431 /* Set up physical groups */
7432 for (i = 0; i < MAX_NUMNODES; i++) {
7433 SCHED_CPUMASK_VAR(nodemask, allmasks);
7434 SCHED_CPUMASK_VAR(send_covered, allmasks);
7436 *nodemask = node_to_cpumask(i);
7437 cpus_and(*nodemask, *nodemask, *cpu_map);
7438 if (cpus_empty(*nodemask))
7441 init_sched_build_groups(nodemask, cpu_map,
7443 send_covered, tmpmask);
7447 /* Set up node groups */
7449 SCHED_CPUMASK_VAR(send_covered, allmasks);
7451 init_sched_build_groups(cpu_map, cpu_map,
7452 &cpu_to_allnodes_group,
7453 send_covered, tmpmask);
7456 for (i = 0; i < MAX_NUMNODES; i++) {
7457 /* Set up node groups */
7458 struct sched_group *sg, *prev;
7459 SCHED_CPUMASK_VAR(nodemask, allmasks);
7460 SCHED_CPUMASK_VAR(domainspan, allmasks);
7461 SCHED_CPUMASK_VAR(covered, allmasks);
7464 *nodemask = node_to_cpumask(i);
7465 cpus_clear(*covered);
7467 cpus_and(*nodemask, *nodemask, *cpu_map);
7468 if (cpus_empty(*nodemask)) {
7469 sched_group_nodes[i] = NULL;
7473 sched_domain_node_span(i, domainspan);
7474 cpus_and(*domainspan, *domainspan, *cpu_map);
7476 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7478 printk(KERN_WARNING "Can not alloc domain group for "
7482 sched_group_nodes[i] = sg;
7483 for_each_cpu_mask(j, *nodemask) {
7484 struct sched_domain *sd;
7486 sd = &per_cpu(node_domains, j);
7489 sg->__cpu_power = 0;
7490 sg->cpumask = *nodemask;
7492 cpus_or(*covered, *covered, *nodemask);
7495 for (j = 0; j < MAX_NUMNODES; j++) {
7496 SCHED_CPUMASK_VAR(notcovered, allmasks);
7497 int n = (i + j) % MAX_NUMNODES;
7498 node_to_cpumask_ptr(pnodemask, n);
7500 cpus_complement(*notcovered, *covered);
7501 cpus_and(*tmpmask, *notcovered, *cpu_map);
7502 cpus_and(*tmpmask, *tmpmask, *domainspan);
7503 if (cpus_empty(*tmpmask))
7506 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7507 if (cpus_empty(*tmpmask))
7510 sg = kmalloc_node(sizeof(struct sched_group),
7514 "Can not alloc domain group for node %d\n", j);
7517 sg->__cpu_power = 0;
7518 sg->cpumask = *tmpmask;
7519 sg->next = prev->next;
7520 cpus_or(*covered, *covered, *tmpmask);
7527 /* Calculate CPU power for physical packages and nodes */
7528 #ifdef CONFIG_SCHED_SMT
7529 for_each_cpu_mask(i, *cpu_map) {
7530 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7532 init_sched_groups_power(i, sd);
7535 #ifdef CONFIG_SCHED_MC
7536 for_each_cpu_mask(i, *cpu_map) {
7537 struct sched_domain *sd = &per_cpu(core_domains, i);
7539 init_sched_groups_power(i, sd);
7543 for_each_cpu_mask(i, *cpu_map) {
7544 struct sched_domain *sd = &per_cpu(phys_domains, i);
7546 init_sched_groups_power(i, sd);
7550 for (i = 0; i < MAX_NUMNODES; i++)
7551 init_numa_sched_groups_power(sched_group_nodes[i]);
7554 struct sched_group *sg;
7556 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7558 init_numa_sched_groups_power(sg);
7562 /* Attach the domains */
7563 for_each_cpu_mask(i, *cpu_map) {
7564 struct sched_domain *sd;
7565 #ifdef CONFIG_SCHED_SMT
7566 sd = &per_cpu(cpu_domains, i);
7567 #elif defined(CONFIG_SCHED_MC)
7568 sd = &per_cpu(core_domains, i);
7570 sd = &per_cpu(phys_domains, i);
7572 cpu_attach_domain(sd, rd, i);
7575 SCHED_CPUMASK_FREE((void *)allmasks);
7580 free_sched_groups(cpu_map, tmpmask);
7581 SCHED_CPUMASK_FREE((void *)allmasks);
7586 static int build_sched_domains(const cpumask_t *cpu_map)
7588 return __build_sched_domains(cpu_map, NULL);
7591 static cpumask_t *doms_cur; /* current sched domains */
7592 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7593 static struct sched_domain_attr *dattr_cur; /* attribues of custom domains
7597 * Special case: If a kmalloc of a doms_cur partition (array of
7598 * cpumask_t) fails, then fallback to a single sched domain,
7599 * as determined by the single cpumask_t fallback_doms.
7601 static cpumask_t fallback_doms;
7603 void __attribute__((weak)) arch_update_cpu_topology(void)
7608 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7609 * For now this just excludes isolated cpus, but could be used to
7610 * exclude other special cases in the future.
7612 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7616 arch_update_cpu_topology();
7618 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7620 doms_cur = &fallback_doms;
7621 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7623 err = build_sched_domains(doms_cur);
7624 register_sched_domain_sysctl();
7629 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7632 free_sched_groups(cpu_map, tmpmask);
7636 * Detach sched domains from a group of cpus specified in cpu_map
7637 * These cpus will now be attached to the NULL domain
7639 static void detach_destroy_domains(const cpumask_t *cpu_map)
7644 unregister_sched_domain_sysctl();
7646 for_each_cpu_mask(i, *cpu_map)
7647 cpu_attach_domain(NULL, &def_root_domain, i);
7648 synchronize_sched();
7649 arch_destroy_sched_domains(cpu_map, &tmpmask);
7652 /* handle null as "default" */
7653 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7654 struct sched_domain_attr *new, int idx_new)
7656 struct sched_domain_attr tmp;
7663 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7664 new ? (new + idx_new) : &tmp,
7665 sizeof(struct sched_domain_attr));
7669 * Partition sched domains as specified by the 'ndoms_new'
7670 * cpumasks in the array doms_new[] of cpumasks. This compares
7671 * doms_new[] to the current sched domain partitioning, doms_cur[].
7672 * It destroys each deleted domain and builds each new domain.
7674 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7675 * The masks don't intersect (don't overlap.) We should setup one
7676 * sched domain for each mask. CPUs not in any of the cpumasks will
7677 * not be load balanced. If the same cpumask appears both in the
7678 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7681 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7682 * ownership of it and will kfree it when done with it. If the caller
7683 * failed the kmalloc call, then it can pass in doms_new == NULL,
7684 * and partition_sched_domains() will fallback to the single partition
7687 * Call with hotplug lock held
7689 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7690 struct sched_domain_attr *dattr_new)
7696 /* always unregister in case we don't destroy any domains */
7697 unregister_sched_domain_sysctl();
7699 if (doms_new == NULL) {
7701 doms_new = &fallback_doms;
7702 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7706 /* Destroy deleted domains */
7707 for (i = 0; i < ndoms_cur; i++) {
7708 for (j = 0; j < ndoms_new; j++) {
7709 if (cpus_equal(doms_cur[i], doms_new[j])
7710 && dattrs_equal(dattr_cur, i, dattr_new, j))
7713 /* no match - a current sched domain not in new doms_new[] */
7714 detach_destroy_domains(doms_cur + i);
7719 /* Build new domains */
7720 for (i = 0; i < ndoms_new; i++) {
7721 for (j = 0; j < ndoms_cur; j++) {
7722 if (cpus_equal(doms_new[i], doms_cur[j])
7723 && dattrs_equal(dattr_new, i, dattr_cur, j))
7726 /* no match - add a new doms_new */
7727 __build_sched_domains(doms_new + i,
7728 dattr_new ? dattr_new + i : NULL);
7733 /* Remember the new sched domains */
7734 if (doms_cur != &fallback_doms)
7736 kfree(dattr_cur); /* kfree(NULL) is safe */
7737 doms_cur = doms_new;
7738 dattr_cur = dattr_new;
7739 ndoms_cur = ndoms_new;
7741 register_sched_domain_sysctl();
7746 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7747 int arch_reinit_sched_domains(void)
7752 detach_destroy_domains(&cpu_online_map);
7753 err = arch_init_sched_domains(&cpu_online_map);
7759 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7763 if (buf[0] != '0' && buf[0] != '1')
7767 sched_smt_power_savings = (buf[0] == '1');
7769 sched_mc_power_savings = (buf[0] == '1');
7771 ret = arch_reinit_sched_domains();
7773 return ret ? ret : count;
7776 #ifdef CONFIG_SCHED_MC
7777 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7779 return sprintf(page, "%u\n", sched_mc_power_savings);
7781 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7782 const char *buf, size_t count)
7784 return sched_power_savings_store(buf, count, 0);
7786 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7787 sched_mc_power_savings_store);
7790 #ifdef CONFIG_SCHED_SMT
7791 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7793 return sprintf(page, "%u\n", sched_smt_power_savings);
7795 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7796 const char *buf, size_t count)
7798 return sched_power_savings_store(buf, count, 1);
7800 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7801 sched_smt_power_savings_store);
7804 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7808 #ifdef CONFIG_SCHED_SMT
7810 err = sysfs_create_file(&cls->kset.kobj,
7811 &attr_sched_smt_power_savings.attr);
7813 #ifdef CONFIG_SCHED_MC
7814 if (!err && mc_capable())
7815 err = sysfs_create_file(&cls->kset.kobj,
7816 &attr_sched_mc_power_savings.attr);
7823 * Force a reinitialization of the sched domains hierarchy. The domains
7824 * and groups cannot be updated in place without racing with the balancing
7825 * code, so we temporarily attach all running cpus to the NULL domain
7826 * which will prevent rebalancing while the sched domains are recalculated.
7828 static int update_sched_domains(struct notifier_block *nfb,
7829 unsigned long action, void *hcpu)
7832 case CPU_UP_PREPARE:
7833 case CPU_UP_PREPARE_FROZEN:
7834 case CPU_DOWN_PREPARE:
7835 case CPU_DOWN_PREPARE_FROZEN:
7836 detach_destroy_domains(&cpu_online_map);
7839 case CPU_UP_CANCELED:
7840 case CPU_UP_CANCELED_FROZEN:
7841 case CPU_DOWN_FAILED:
7842 case CPU_DOWN_FAILED_FROZEN:
7844 case CPU_ONLINE_FROZEN:
7846 case CPU_DEAD_FROZEN:
7848 * Fall through and re-initialise the domains.
7855 /* The hotplug lock is already held by cpu_up/cpu_down */
7856 arch_init_sched_domains(&cpu_online_map);
7861 void __init sched_init_smp(void)
7863 cpumask_t non_isolated_cpus;
7865 #if defined(CONFIG_NUMA)
7866 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7868 BUG_ON(sched_group_nodes_bycpu == NULL);
7871 arch_init_sched_domains(&cpu_online_map);
7872 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7873 if (cpus_empty(non_isolated_cpus))
7874 cpu_set(smp_processor_id(), non_isolated_cpus);
7876 /* XXX: Theoretical race here - CPU may be hotplugged now */
7877 hotcpu_notifier(update_sched_domains, 0);
7879 /* Move init over to a non-isolated CPU */
7880 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7882 sched_init_granularity();
7885 void __init sched_init_smp(void)
7887 #if defined(CONFIG_NUMA)
7888 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7890 BUG_ON(sched_group_nodes_bycpu == NULL);
7892 sched_init_granularity();
7894 #endif /* CONFIG_SMP */
7896 int in_sched_functions(unsigned long addr)
7898 return in_lock_functions(addr) ||
7899 (addr >= (unsigned long)__sched_text_start
7900 && addr < (unsigned long)__sched_text_end);
7903 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7905 cfs_rq->tasks_timeline = RB_ROOT;
7906 INIT_LIST_HEAD(&cfs_rq->tasks);
7907 #ifdef CONFIG_FAIR_GROUP_SCHED
7910 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7913 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7915 struct rt_prio_array *array;
7918 array = &rt_rq->active;
7919 for (i = 0; i < MAX_RT_PRIO; i++) {
7920 INIT_LIST_HEAD(array->queue + i);
7921 __clear_bit(i, array->bitmap);
7923 /* delimiter for bitsearch: */
7924 __set_bit(MAX_RT_PRIO, array->bitmap);
7926 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7927 rt_rq->highest_prio = MAX_RT_PRIO;
7930 rt_rq->rt_nr_migratory = 0;
7931 rt_rq->overloaded = 0;
7935 rt_rq->rt_throttled = 0;
7936 rt_rq->rt_runtime = 0;
7937 spin_lock_init(&rt_rq->rt_runtime_lock);
7939 #ifdef CONFIG_RT_GROUP_SCHED
7940 rt_rq->rt_nr_boosted = 0;
7945 #ifdef CONFIG_FAIR_GROUP_SCHED
7946 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7947 struct sched_entity *se, int cpu, int add,
7948 struct sched_entity *parent)
7950 struct rq *rq = cpu_rq(cpu);
7951 tg->cfs_rq[cpu] = cfs_rq;
7952 init_cfs_rq(cfs_rq, rq);
7955 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7958 /* se could be NULL for init_task_group */
7963 se->cfs_rq = &rq->cfs;
7965 se->cfs_rq = parent->my_q;
7968 se->load.weight = tg->shares;
7969 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7970 se->parent = parent;
7974 #ifdef CONFIG_RT_GROUP_SCHED
7975 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7976 struct sched_rt_entity *rt_se, int cpu, int add,
7977 struct sched_rt_entity *parent)
7979 struct rq *rq = cpu_rq(cpu);
7981 tg->rt_rq[cpu] = rt_rq;
7982 init_rt_rq(rt_rq, rq);
7984 rt_rq->rt_se = rt_se;
7985 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7987 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7989 tg->rt_se[cpu] = rt_se;
7994 rt_se->rt_rq = &rq->rt;
7996 rt_se->rt_rq = parent->my_q;
7998 rt_se->rt_rq = &rq->rt;
7999 rt_se->my_q = rt_rq;
8000 rt_se->parent = parent;
8001 INIT_LIST_HEAD(&rt_se->run_list);
8005 void __init sched_init(void)
8008 unsigned long alloc_size = 0, ptr;
8010 #ifdef CONFIG_FAIR_GROUP_SCHED
8011 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8013 #ifdef CONFIG_RT_GROUP_SCHED
8014 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8016 #ifdef CONFIG_USER_SCHED
8020 * As sched_init() is called before page_alloc is setup,
8021 * we use alloc_bootmem().
8024 ptr = (unsigned long)alloc_bootmem_low(alloc_size);
8026 #ifdef CONFIG_FAIR_GROUP_SCHED
8027 init_task_group.se = (struct sched_entity **)ptr;
8028 ptr += nr_cpu_ids * sizeof(void **);
8030 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8031 ptr += nr_cpu_ids * sizeof(void **);
8033 #ifdef CONFIG_USER_SCHED
8034 root_task_group.se = (struct sched_entity **)ptr;
8035 ptr += nr_cpu_ids * sizeof(void **);
8037 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8038 ptr += nr_cpu_ids * sizeof(void **);
8041 #ifdef CONFIG_RT_GROUP_SCHED
8042 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8043 ptr += nr_cpu_ids * sizeof(void **);
8045 init_task_group.rt_rq = (struct rt_rq **)ptr;
8046 ptr += nr_cpu_ids * sizeof(void **);
8048 #ifdef CONFIG_USER_SCHED
8049 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8050 ptr += nr_cpu_ids * sizeof(void **);
8052 root_task_group.rt_rq = (struct rt_rq **)ptr;
8053 ptr += nr_cpu_ids * sizeof(void **);
8060 init_defrootdomain();
8063 init_rt_bandwidth(&def_rt_bandwidth,
8064 global_rt_period(), global_rt_runtime());
8066 #ifdef CONFIG_RT_GROUP_SCHED
8067 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8068 global_rt_period(), global_rt_runtime());
8069 #ifdef CONFIG_USER_SCHED
8070 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8071 global_rt_period(), RUNTIME_INF);
8075 #ifdef CONFIG_GROUP_SCHED
8076 list_add(&init_task_group.list, &task_groups);
8077 INIT_LIST_HEAD(&init_task_group.children);
8079 #ifdef CONFIG_USER_SCHED
8080 INIT_LIST_HEAD(&root_task_group.children);
8081 init_task_group.parent = &root_task_group;
8082 list_add(&init_task_group.siblings, &root_task_group.children);
8086 for_each_possible_cpu(i) {
8090 spin_lock_init(&rq->lock);
8091 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8094 update_last_tick_seen(rq);
8095 init_cfs_rq(&rq->cfs, rq);
8096 init_rt_rq(&rq->rt, rq);
8097 #ifdef CONFIG_FAIR_GROUP_SCHED
8098 init_task_group.shares = init_task_group_load;
8099 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8100 #ifdef CONFIG_CGROUP_SCHED
8102 * How much cpu bandwidth does init_task_group get?
8104 * In case of task-groups formed thr' the cgroup filesystem, it
8105 * gets 100% of the cpu resources in the system. This overall
8106 * system cpu resource is divided among the tasks of
8107 * init_task_group and its child task-groups in a fair manner,
8108 * based on each entity's (task or task-group's) weight
8109 * (se->load.weight).
8111 * In other words, if init_task_group has 10 tasks of weight
8112 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8113 * then A0's share of the cpu resource is:
8115 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8117 * We achieve this by letting init_task_group's tasks sit
8118 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8120 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8121 #elif defined CONFIG_USER_SCHED
8122 root_task_group.shares = NICE_0_LOAD;
8123 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8125 * In case of task-groups formed thr' the user id of tasks,
8126 * init_task_group represents tasks belonging to root user.
8127 * Hence it forms a sibling of all subsequent groups formed.
8128 * In this case, init_task_group gets only a fraction of overall
8129 * system cpu resource, based on the weight assigned to root
8130 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8131 * by letting tasks of init_task_group sit in a separate cfs_rq
8132 * (init_cfs_rq) and having one entity represent this group of
8133 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8135 init_tg_cfs_entry(&init_task_group,
8136 &per_cpu(init_cfs_rq, i),
8137 &per_cpu(init_sched_entity, i), i, 1,
8138 root_task_group.se[i]);
8141 #endif /* CONFIG_FAIR_GROUP_SCHED */
8143 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8144 #ifdef CONFIG_RT_GROUP_SCHED
8145 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8146 #ifdef CONFIG_CGROUP_SCHED
8147 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8148 #elif defined CONFIG_USER_SCHED
8149 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8150 init_tg_rt_entry(&init_task_group,
8151 &per_cpu(init_rt_rq, i),
8152 &per_cpu(init_sched_rt_entity, i), i, 1,
8153 root_task_group.rt_se[i]);
8157 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8158 rq->cpu_load[j] = 0;
8162 rq->active_balance = 0;
8163 rq->next_balance = jiffies;
8166 rq->migration_thread = NULL;
8167 INIT_LIST_HEAD(&rq->migration_queue);
8168 rq_attach_root(rq, &def_root_domain);
8171 atomic_set(&rq->nr_iowait, 0);
8174 set_load_weight(&init_task);
8176 #ifdef CONFIG_PREEMPT_NOTIFIERS
8177 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8181 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8184 #ifdef CONFIG_RT_MUTEXES
8185 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8189 * The boot idle thread does lazy MMU switching as well:
8191 atomic_inc(&init_mm.mm_count);
8192 enter_lazy_tlb(&init_mm, current);
8195 * Make us the idle thread. Technically, schedule() should not be
8196 * called from this thread, however somewhere below it might be,
8197 * but because we are the idle thread, we just pick up running again
8198 * when this runqueue becomes "idle".
8200 init_idle(current, smp_processor_id());
8202 * During early bootup we pretend to be a normal task:
8204 current->sched_class = &fair_sched_class;
8206 scheduler_running = 1;
8209 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8210 void __might_sleep(char *file, int line)
8213 static unsigned long prev_jiffy; /* ratelimiting */
8215 if ((in_atomic() || irqs_disabled()) &&
8216 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8217 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8219 prev_jiffy = jiffies;
8220 printk(KERN_ERR "BUG: sleeping function called from invalid"
8221 " context at %s:%d\n", file, line);
8222 printk("in_atomic():%d, irqs_disabled():%d\n",
8223 in_atomic(), irqs_disabled());
8224 debug_show_held_locks(current);
8225 if (irqs_disabled())
8226 print_irqtrace_events(current);
8231 EXPORT_SYMBOL(__might_sleep);
8234 #ifdef CONFIG_MAGIC_SYSRQ
8235 static void normalize_task(struct rq *rq, struct task_struct *p)
8238 update_rq_clock(rq);
8239 on_rq = p->se.on_rq;
8241 deactivate_task(rq, p, 0);
8242 __setscheduler(rq, p, SCHED_NORMAL, 0);
8244 activate_task(rq, p, 0);
8245 resched_task(rq->curr);
8249 void normalize_rt_tasks(void)
8251 struct task_struct *g, *p;
8252 unsigned long flags;
8255 read_lock_irqsave(&tasklist_lock, flags);
8256 do_each_thread(g, p) {
8258 * Only normalize user tasks:
8263 p->se.exec_start = 0;
8264 #ifdef CONFIG_SCHEDSTATS
8265 p->se.wait_start = 0;
8266 p->se.sleep_start = 0;
8267 p->se.block_start = 0;
8269 task_rq(p)->clock = 0;
8273 * Renice negative nice level userspace
8276 if (TASK_NICE(p) < 0 && p->mm)
8277 set_user_nice(p, 0);
8281 spin_lock(&p->pi_lock);
8282 rq = __task_rq_lock(p);
8284 normalize_task(rq, p);
8286 __task_rq_unlock(rq);
8287 spin_unlock(&p->pi_lock);
8288 } while_each_thread(g, p);
8290 read_unlock_irqrestore(&tasklist_lock, flags);
8293 #endif /* CONFIG_MAGIC_SYSRQ */
8297 * These functions are only useful for the IA64 MCA handling.
8299 * They can only be called when the whole system has been
8300 * stopped - every CPU needs to be quiescent, and no scheduling
8301 * activity can take place. Using them for anything else would
8302 * be a serious bug, and as a result, they aren't even visible
8303 * under any other configuration.
8307 * curr_task - return the current task for a given cpu.
8308 * @cpu: the processor in question.
8310 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8312 struct task_struct *curr_task(int cpu)
8314 return cpu_curr(cpu);
8318 * set_curr_task - set the current task for a given cpu.
8319 * @cpu: the processor in question.
8320 * @p: the task pointer to set.
8322 * Description: This function must only be used when non-maskable interrupts
8323 * are serviced on a separate stack. It allows the architecture to switch the
8324 * notion of the current task on a cpu in a non-blocking manner. This function
8325 * must be called with all CPU's synchronized, and interrupts disabled, the
8326 * and caller must save the original value of the current task (see
8327 * curr_task() above) and restore that value before reenabling interrupts and
8328 * re-starting the system.
8330 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8332 void set_curr_task(int cpu, struct task_struct *p)
8339 #ifdef CONFIG_FAIR_GROUP_SCHED
8340 static void free_fair_sched_group(struct task_group *tg)
8344 for_each_possible_cpu(i) {
8346 kfree(tg->cfs_rq[i]);
8356 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8358 struct cfs_rq *cfs_rq;
8359 struct sched_entity *se, *parent_se;
8363 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8366 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8370 tg->shares = NICE_0_LOAD;
8372 for_each_possible_cpu(i) {
8375 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8376 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8380 se = kmalloc_node(sizeof(struct sched_entity),
8381 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8385 parent_se = parent ? parent->se[i] : NULL;
8386 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8395 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8397 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8398 &cpu_rq(cpu)->leaf_cfs_rq_list);
8401 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8403 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8406 static inline void free_fair_sched_group(struct task_group *tg)
8411 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8416 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8420 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8425 #ifdef CONFIG_RT_GROUP_SCHED
8426 static void free_rt_sched_group(struct task_group *tg)
8430 destroy_rt_bandwidth(&tg->rt_bandwidth);
8432 for_each_possible_cpu(i) {
8434 kfree(tg->rt_rq[i]);
8436 kfree(tg->rt_se[i]);
8444 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8446 struct rt_rq *rt_rq;
8447 struct sched_rt_entity *rt_se, *parent_se;
8451 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8454 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8458 init_rt_bandwidth(&tg->rt_bandwidth,
8459 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8461 for_each_possible_cpu(i) {
8464 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8465 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8469 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8470 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8474 parent_se = parent ? parent->rt_se[i] : NULL;
8475 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8484 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8486 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8487 &cpu_rq(cpu)->leaf_rt_rq_list);
8490 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8492 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8495 static inline void free_rt_sched_group(struct task_group *tg)
8500 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8505 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8509 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8514 #ifdef CONFIG_GROUP_SCHED
8515 static void free_sched_group(struct task_group *tg)
8517 free_fair_sched_group(tg);
8518 free_rt_sched_group(tg);
8522 /* allocate runqueue etc for a new task group */
8523 struct task_group *sched_create_group(struct task_group *parent)
8525 struct task_group *tg;
8526 unsigned long flags;
8529 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8531 return ERR_PTR(-ENOMEM);
8533 if (!alloc_fair_sched_group(tg, parent))
8536 if (!alloc_rt_sched_group(tg, parent))
8539 spin_lock_irqsave(&task_group_lock, flags);
8540 for_each_possible_cpu(i) {
8541 register_fair_sched_group(tg, i);
8542 register_rt_sched_group(tg, i);
8544 list_add_rcu(&tg->list, &task_groups);
8546 WARN_ON(!parent); /* root should already exist */
8548 tg->parent = parent;
8549 list_add_rcu(&tg->siblings, &parent->children);
8550 INIT_LIST_HEAD(&tg->children);
8551 spin_unlock_irqrestore(&task_group_lock, flags);
8556 free_sched_group(tg);
8557 return ERR_PTR(-ENOMEM);
8560 /* rcu callback to free various structures associated with a task group */
8561 static void free_sched_group_rcu(struct rcu_head *rhp)
8563 /* now it should be safe to free those cfs_rqs */
8564 free_sched_group(container_of(rhp, struct task_group, rcu));
8567 /* Destroy runqueue etc associated with a task group */
8568 void sched_destroy_group(struct task_group *tg)
8570 unsigned long flags;
8573 spin_lock_irqsave(&task_group_lock, flags);
8574 for_each_possible_cpu(i) {
8575 unregister_fair_sched_group(tg, i);
8576 unregister_rt_sched_group(tg, i);
8578 list_del_rcu(&tg->list);
8579 list_del_rcu(&tg->siblings);
8580 spin_unlock_irqrestore(&task_group_lock, flags);
8582 /* wait for possible concurrent references to cfs_rqs complete */
8583 call_rcu(&tg->rcu, free_sched_group_rcu);
8586 /* change task's runqueue when it moves between groups.
8587 * The caller of this function should have put the task in its new group
8588 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8589 * reflect its new group.
8591 void sched_move_task(struct task_struct *tsk)
8594 unsigned long flags;
8597 rq = task_rq_lock(tsk, &flags);
8599 update_rq_clock(rq);
8601 running = task_current(rq, tsk);
8602 on_rq = tsk->se.on_rq;
8605 dequeue_task(rq, tsk, 0);
8606 if (unlikely(running))
8607 tsk->sched_class->put_prev_task(rq, tsk);
8609 set_task_rq(tsk, task_cpu(tsk));
8611 #ifdef CONFIG_FAIR_GROUP_SCHED
8612 if (tsk->sched_class->moved_group)
8613 tsk->sched_class->moved_group(tsk);
8616 if (unlikely(running))
8617 tsk->sched_class->set_curr_task(rq);
8619 enqueue_task(rq, tsk, 0);
8621 task_rq_unlock(rq, &flags);
8625 #ifdef CONFIG_FAIR_GROUP_SCHED
8626 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8628 struct cfs_rq *cfs_rq = se->cfs_rq;
8633 dequeue_entity(cfs_rq, se, 0);
8635 se->load.weight = shares;
8636 se->load.inv_weight = div64_64((1ULL<<32), shares);
8639 enqueue_entity(cfs_rq, se, 0);
8642 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8644 struct cfs_rq *cfs_rq = se->cfs_rq;
8645 struct rq *rq = cfs_rq->rq;
8646 unsigned long flags;
8648 spin_lock_irqsave(&rq->lock, flags);
8649 __set_se_shares(se, shares);
8650 spin_unlock_irqrestore(&rq->lock, flags);
8653 static DEFINE_MUTEX(shares_mutex);
8655 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8658 unsigned long flags;
8661 * We can't change the weight of the root cgroup.
8667 * A weight of 0 or 1 can cause arithmetics problems.
8668 * (The default weight is 1024 - so there's no practical
8669 * limitation from this.)
8671 if (shares < MIN_SHARES)
8672 shares = MIN_SHARES;
8674 mutex_lock(&shares_mutex);
8675 if (tg->shares == shares)
8678 spin_lock_irqsave(&task_group_lock, flags);
8679 for_each_possible_cpu(i)
8680 unregister_fair_sched_group(tg, i);
8681 list_del_rcu(&tg->siblings);
8682 spin_unlock_irqrestore(&task_group_lock, flags);
8684 /* wait for any ongoing reference to this group to finish */
8685 synchronize_sched();
8688 * Now we are free to modify the group's share on each cpu
8689 * w/o tripping rebalance_share or load_balance_fair.
8691 tg->shares = shares;
8692 for_each_possible_cpu(i) {
8696 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8697 set_se_shares(tg->se[i], shares/nr_cpu_ids);
8701 * Enable load balance activity on this group, by inserting it back on
8702 * each cpu's rq->leaf_cfs_rq_list.
8704 spin_lock_irqsave(&task_group_lock, flags);
8705 for_each_possible_cpu(i)
8706 register_fair_sched_group(tg, i);
8707 list_add_rcu(&tg->siblings, &tg->parent->children);
8708 spin_unlock_irqrestore(&task_group_lock, flags);
8710 mutex_unlock(&shares_mutex);
8714 unsigned long sched_group_shares(struct task_group *tg)
8720 #ifdef CONFIG_RT_GROUP_SCHED
8722 * Ensure that the real time constraints are schedulable.
8724 static DEFINE_MUTEX(rt_constraints_mutex);
8726 static unsigned long to_ratio(u64 period, u64 runtime)
8728 if (runtime == RUNTIME_INF)
8731 return div64_64(runtime << 16, period);
8734 #ifdef CONFIG_CGROUP_SCHED
8735 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8737 struct task_group *tgi, *parent = tg->parent;
8738 unsigned long total = 0;
8741 if (global_rt_period() < period)
8744 return to_ratio(period, runtime) <
8745 to_ratio(global_rt_period(), global_rt_runtime());
8748 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8752 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8756 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8757 tgi->rt_bandwidth.rt_runtime);
8761 return total + to_ratio(period, runtime) <
8762 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8763 parent->rt_bandwidth.rt_runtime);
8765 #elif defined CONFIG_USER_SCHED
8766 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8768 struct task_group *tgi;
8769 unsigned long total = 0;
8770 unsigned long global_ratio =
8771 to_ratio(global_rt_period(), global_rt_runtime());
8774 list_for_each_entry_rcu(tgi, &task_groups, list) {
8778 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8779 tgi->rt_bandwidth.rt_runtime);
8783 return total + to_ratio(period, runtime) < global_ratio;
8787 /* Must be called with tasklist_lock held */
8788 static inline int tg_has_rt_tasks(struct task_group *tg)
8790 struct task_struct *g, *p;
8791 do_each_thread(g, p) {
8792 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8794 } while_each_thread(g, p);
8798 static int tg_set_bandwidth(struct task_group *tg,
8799 u64 rt_period, u64 rt_runtime)
8803 mutex_lock(&rt_constraints_mutex);
8804 read_lock(&tasklist_lock);
8805 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8809 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8814 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8815 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8816 tg->rt_bandwidth.rt_runtime = rt_runtime;
8818 for_each_possible_cpu(i) {
8819 struct rt_rq *rt_rq = tg->rt_rq[i];
8821 spin_lock(&rt_rq->rt_runtime_lock);
8822 rt_rq->rt_runtime = rt_runtime;
8823 spin_unlock(&rt_rq->rt_runtime_lock);
8825 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8827 read_unlock(&tasklist_lock);
8828 mutex_unlock(&rt_constraints_mutex);
8833 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8835 u64 rt_runtime, rt_period;
8837 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8838 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8839 if (rt_runtime_us < 0)
8840 rt_runtime = RUNTIME_INF;
8842 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8845 long sched_group_rt_runtime(struct task_group *tg)
8849 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8852 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8853 do_div(rt_runtime_us, NSEC_PER_USEC);
8854 return rt_runtime_us;
8857 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8859 u64 rt_runtime, rt_period;
8861 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8862 rt_runtime = tg->rt_bandwidth.rt_runtime;
8864 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8867 long sched_group_rt_period(struct task_group *tg)
8871 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8872 do_div(rt_period_us, NSEC_PER_USEC);
8873 return rt_period_us;
8876 static int sched_rt_global_constraints(void)
8880 mutex_lock(&rt_constraints_mutex);
8881 if (!__rt_schedulable(NULL, 1, 0))
8883 mutex_unlock(&rt_constraints_mutex);
8888 static int sched_rt_global_constraints(void)
8890 unsigned long flags;
8893 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8894 for_each_possible_cpu(i) {
8895 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8897 spin_lock(&rt_rq->rt_runtime_lock);
8898 rt_rq->rt_runtime = global_rt_runtime();
8899 spin_unlock(&rt_rq->rt_runtime_lock);
8901 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8907 int sched_rt_handler(struct ctl_table *table, int write,
8908 struct file *filp, void __user *buffer, size_t *lenp,
8912 int old_period, old_runtime;
8913 static DEFINE_MUTEX(mutex);
8916 old_period = sysctl_sched_rt_period;
8917 old_runtime = sysctl_sched_rt_runtime;
8919 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8921 if (!ret && write) {
8922 ret = sched_rt_global_constraints();
8924 sysctl_sched_rt_period = old_period;
8925 sysctl_sched_rt_runtime = old_runtime;
8927 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8928 def_rt_bandwidth.rt_period =
8929 ns_to_ktime(global_rt_period());
8932 mutex_unlock(&mutex);
8937 #ifdef CONFIG_CGROUP_SCHED
8939 /* return corresponding task_group object of a cgroup */
8940 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8942 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8943 struct task_group, css);
8946 static struct cgroup_subsys_state *
8947 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8949 struct task_group *tg, *parent;
8951 if (!cgrp->parent) {
8952 /* This is early initialization for the top cgroup */
8953 init_task_group.css.cgroup = cgrp;
8954 return &init_task_group.css;
8957 parent = cgroup_tg(cgrp->parent);
8958 tg = sched_create_group(parent);
8960 return ERR_PTR(-ENOMEM);
8962 /* Bind the cgroup to task_group object we just created */
8963 tg->css.cgroup = cgrp;
8969 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8971 struct task_group *tg = cgroup_tg(cgrp);
8973 sched_destroy_group(tg);
8977 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8978 struct task_struct *tsk)
8980 #ifdef CONFIG_RT_GROUP_SCHED
8981 /* Don't accept realtime tasks when there is no way for them to run */
8982 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8985 /* We don't support RT-tasks being in separate groups */
8986 if (tsk->sched_class != &fair_sched_class)
8994 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8995 struct cgroup *old_cont, struct task_struct *tsk)
8997 sched_move_task(tsk);
9000 #ifdef CONFIG_FAIR_GROUP_SCHED
9001 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9004 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9007 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
9009 struct task_group *tg = cgroup_tg(cgrp);
9011 return (u64) tg->shares;
9015 #ifdef CONFIG_RT_GROUP_SCHED
9016 static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9018 const char __user *userbuf,
9019 size_t nbytes, loff_t *unused_ppos)
9028 if (nbytes >= sizeof(buffer))
9030 if (copy_from_user(buffer, userbuf, nbytes))
9033 buffer[nbytes] = 0; /* nul-terminate */
9035 /* strip newline if necessary */
9036 if (nbytes && (buffer[nbytes-1] == '\n'))
9037 buffer[nbytes-1] = 0;
9038 val = simple_strtoll(buffer, &end, 0);
9042 /* Pass to subsystem */
9043 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9049 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
9051 char __user *buf, size_t nbytes,
9055 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
9056 int len = sprintf(tmp, "%ld\n", val);
9058 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
9061 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9064 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9067 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9069 return sched_group_rt_period(cgroup_tg(cgrp));
9073 static struct cftype cpu_files[] = {
9074 #ifdef CONFIG_FAIR_GROUP_SCHED
9077 .read_uint = cpu_shares_read_uint,
9078 .write_uint = cpu_shares_write_uint,
9081 #ifdef CONFIG_RT_GROUP_SCHED
9083 .name = "rt_runtime_us",
9084 .read = cpu_rt_runtime_read,
9085 .write = cpu_rt_runtime_write,
9088 .name = "rt_period_us",
9089 .read_uint = cpu_rt_period_read_uint,
9090 .write_uint = cpu_rt_period_write_uint,
9095 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9097 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9100 struct cgroup_subsys cpu_cgroup_subsys = {
9102 .create = cpu_cgroup_create,
9103 .destroy = cpu_cgroup_destroy,
9104 .can_attach = cpu_cgroup_can_attach,
9105 .attach = cpu_cgroup_attach,
9106 .populate = cpu_cgroup_populate,
9107 .subsys_id = cpu_cgroup_subsys_id,
9111 #endif /* CONFIG_CGROUP_SCHED */
9113 #ifdef CONFIG_CGROUP_CPUACCT
9116 * CPU accounting code for task groups.
9118 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9119 * (balbir@in.ibm.com).
9122 /* track cpu usage of a group of tasks */
9124 struct cgroup_subsys_state css;
9125 /* cpuusage holds pointer to a u64-type object on every cpu */
9129 struct cgroup_subsys cpuacct_subsys;
9131 /* return cpu accounting group corresponding to this container */
9132 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9134 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9135 struct cpuacct, css);
9138 /* return cpu accounting group to which this task belongs */
9139 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9141 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9142 struct cpuacct, css);
9145 /* create a new cpu accounting group */
9146 static struct cgroup_subsys_state *cpuacct_create(
9147 struct cgroup_subsys *ss, struct cgroup *cgrp)
9149 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9152 return ERR_PTR(-ENOMEM);
9154 ca->cpuusage = alloc_percpu(u64);
9155 if (!ca->cpuusage) {
9157 return ERR_PTR(-ENOMEM);
9163 /* destroy an existing cpu accounting group */
9165 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9167 struct cpuacct *ca = cgroup_ca(cgrp);
9169 free_percpu(ca->cpuusage);
9173 /* return total cpu usage (in nanoseconds) of a group */
9174 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9176 struct cpuacct *ca = cgroup_ca(cgrp);
9177 u64 totalcpuusage = 0;
9180 for_each_possible_cpu(i) {
9181 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9184 * Take rq->lock to make 64-bit addition safe on 32-bit
9187 spin_lock_irq(&cpu_rq(i)->lock);
9188 totalcpuusage += *cpuusage;
9189 spin_unlock_irq(&cpu_rq(i)->lock);
9192 return totalcpuusage;
9195 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9198 struct cpuacct *ca = cgroup_ca(cgrp);
9207 for_each_possible_cpu(i) {
9208 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9210 spin_lock_irq(&cpu_rq(i)->lock);
9212 spin_unlock_irq(&cpu_rq(i)->lock);
9218 static struct cftype files[] = {
9221 .read_uint = cpuusage_read,
9222 .write_uint = cpuusage_write,
9226 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9228 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9232 * charge this task's execution time to its accounting group.
9234 * called with rq->lock held.
9236 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9240 if (!cpuacct_subsys.active)
9245 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9247 *cpuusage += cputime;
9251 struct cgroup_subsys cpuacct_subsys = {
9253 .create = cpuacct_create,
9254 .destroy = cpuacct_destroy,
9255 .populate = cpuacct_populate,
9256 .subsys_id = cpuacct_subsys_id,
9258 #endif /* CONFIG_CGROUP_CPUACCT */