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
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak)) sched_clock(void)
75 return (unsigned long long)jiffies * (1000000000 / HZ);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio)
145 if (static_prio == NICE_TO_PRIO(19))
148 if (static_prio < NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
151 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
154 static inline int rt_policy(int policy)
156 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
161 static inline int task_has_rt_policy(struct task_struct *p)
163 return rt_policy(p->policy);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array {
170 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
171 struct list_head queue[MAX_RT_PRIO];
174 #ifdef CONFIG_FAIR_GROUP_SCHED
178 /* task group related information */
180 /* schedulable entities of this group on each cpu */
181 struct sched_entity **se;
182 /* runqueue "owned" by this group on each cpu */
183 struct cfs_rq **cfs_rq;
184 unsigned long shares;
187 /* Default task group's sched entity on each cpu */
188 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
189 /* Default task group's cfs_rq on each cpu */
190 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
192 static struct sched_entity *init_sched_entity_p[NR_CPUS];
193 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
195 /* Default task group.
196 * Every task in system belong to this group at bootup.
198 struct task_grp init_task_grp = {
199 .se = init_sched_entity_p,
200 .cfs_rq = init_cfs_rq_p,
203 #define INIT_TASK_GRP_LOAD NICE_0_LOAD
204 static int init_task_grp_load = INIT_TASK_GRP_LOAD;
206 /* return group to which a task belongs */
207 static inline struct task_grp *task_grp(struct task_struct *p)
216 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
217 static inline void set_task_cfs_rq(struct task_struct *p)
219 p->se.cfs_rq = task_grp(p)->cfs_rq[task_cpu(p)];
220 p->se.parent = task_grp(p)->se[task_cpu(p)];
225 static inline void set_task_cfs_rq(struct task_struct *p) { }
227 #endif /* CONFIG_FAIR_GROUP_SCHED */
229 /* CFS-related fields in a runqueue */
231 struct load_weight load;
232 unsigned long nr_running;
237 struct rb_root tasks_timeline;
238 struct rb_node *rb_leftmost;
239 struct rb_node *rb_load_balance_curr;
240 /* 'curr' points to currently running entity on this cfs_rq.
241 * It is set to NULL otherwise (i.e when none are currently running).
243 struct sched_entity *curr;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
247 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
248 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
249 * (like users, containers etc.)
251 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
252 * list is used during load balance.
254 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
255 struct task_grp *tg; /* group that "owns" this runqueue */
260 /* Real-Time classes' related field in a runqueue: */
262 struct rt_prio_array active;
263 int rt_load_balance_idx;
264 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
268 * This is the main, per-CPU runqueue data structure.
270 * Locking rule: those places that want to lock multiple runqueues
271 * (such as the load balancing or the thread migration code), lock
272 * acquire operations must be ordered by ascending &runqueue.
275 spinlock_t lock; /* runqueue lock */
278 * nr_running and cpu_load should be in the same cacheline because
279 * remote CPUs use both these fields when doing load calculation.
281 unsigned long nr_running;
282 #define CPU_LOAD_IDX_MAX 5
283 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
284 unsigned char idle_at_tick;
286 unsigned char in_nohz_recently;
288 struct load_weight load; /* capture load from *all* tasks on this cpu */
289 unsigned long nr_load_updates;
293 #ifdef CONFIG_FAIR_GROUP_SCHED
294 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
299 * This is part of a global counter where only the total sum
300 * over all CPUs matters. A task can increase this counter on
301 * one CPU and if it got migrated afterwards it may decrease
302 * it on another CPU. Always updated under the runqueue lock:
304 unsigned long nr_uninterruptible;
306 struct task_struct *curr, *idle;
307 unsigned long next_balance;
308 struct mm_struct *prev_mm;
310 u64 clock, prev_clock_raw;
313 unsigned int clock_warps, clock_overflows;
315 unsigned int clock_deep_idle_events;
321 struct sched_domain *sd;
323 /* For active balancing */
326 int cpu; /* cpu of this runqueue */
328 struct task_struct *migration_thread;
329 struct list_head migration_queue;
332 #ifdef CONFIG_SCHEDSTATS
334 struct sched_info rq_sched_info;
336 /* sys_sched_yield() stats */
337 unsigned long yld_exp_empty;
338 unsigned long yld_act_empty;
339 unsigned long yld_both_empty;
340 unsigned long yld_cnt;
342 /* schedule() stats */
343 unsigned long sched_switch;
344 unsigned long sched_cnt;
345 unsigned long sched_goidle;
347 /* try_to_wake_up() stats */
348 unsigned long ttwu_cnt;
349 unsigned long ttwu_local;
351 struct lock_class_key rq_lock_key;
354 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
355 static DEFINE_MUTEX(sched_hotcpu_mutex);
357 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
359 rq->curr->sched_class->check_preempt_curr(rq, p);
362 static inline int cpu_of(struct rq *rq)
372 * Update the per-runqueue clock, as finegrained as the platform can give
373 * us, but without assuming monotonicity, etc.:
375 static void __update_rq_clock(struct rq *rq)
377 u64 prev_raw = rq->prev_clock_raw;
378 u64 now = sched_clock();
379 s64 delta = now - prev_raw;
380 u64 clock = rq->clock;
382 #ifdef CONFIG_SCHED_DEBUG
383 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
386 * Protect against sched_clock() occasionally going backwards:
388 if (unlikely(delta < 0)) {
393 * Catch too large forward jumps too:
395 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
396 if (clock < rq->tick_timestamp + TICK_NSEC)
397 clock = rq->tick_timestamp + TICK_NSEC;
400 rq->clock_overflows++;
402 if (unlikely(delta > rq->clock_max_delta))
403 rq->clock_max_delta = delta;
408 rq->prev_clock_raw = now;
412 static void update_rq_clock(struct rq *rq)
414 if (likely(smp_processor_id() == cpu_of(rq)))
415 __update_rq_clock(rq);
419 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
420 * See detach_destroy_domains: synchronize_sched for details.
422 * The domain tree of any CPU may only be accessed from within
423 * preempt-disabled sections.
425 #define for_each_domain(cpu, __sd) \
426 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
428 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
429 #define this_rq() (&__get_cpu_var(runqueues))
430 #define task_rq(p) cpu_rq(task_cpu(p))
431 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
434 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
436 #ifdef CONFIG_SCHED_DEBUG
437 # define const_debug __read_mostly
439 # define const_debug static const
443 * Debugging: various feature bits
446 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
447 SCHED_FEAT_START_DEBIT = 2,
448 SCHED_FEAT_USE_TREE_AVG = 4,
449 SCHED_FEAT_APPROX_AVG = 8,
452 const_debug unsigned int sysctl_sched_features =
453 SCHED_FEAT_NEW_FAIR_SLEEPERS *1 |
454 SCHED_FEAT_START_DEBIT *1 |
455 SCHED_FEAT_USE_TREE_AVG *0 |
456 SCHED_FEAT_APPROX_AVG *0;
458 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
461 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
462 * clock constructed from sched_clock():
464 unsigned long long cpu_clock(int cpu)
466 unsigned long long now;
470 local_irq_save(flags);
474 local_irq_restore(flags);
479 #ifndef prepare_arch_switch
480 # define prepare_arch_switch(next) do { } while (0)
482 #ifndef finish_arch_switch
483 # define finish_arch_switch(prev) do { } while (0)
486 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
487 static inline int task_running(struct rq *rq, struct task_struct *p)
489 return rq->curr == p;
492 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
496 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
498 #ifdef CONFIG_DEBUG_SPINLOCK
499 /* this is a valid case when another task releases the spinlock */
500 rq->lock.owner = current;
503 * If we are tracking spinlock dependencies then we have to
504 * fix up the runqueue lock - which gets 'carried over' from
507 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
509 spin_unlock_irq(&rq->lock);
512 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
513 static inline int task_running(struct rq *rq, struct task_struct *p)
518 return rq->curr == p;
522 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
526 * We can optimise this out completely for !SMP, because the
527 * SMP rebalancing from interrupt is the only thing that cares
532 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
533 spin_unlock_irq(&rq->lock);
535 spin_unlock(&rq->lock);
539 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
543 * After ->oncpu is cleared, the task can be moved to a different CPU.
544 * We must ensure this doesn't happen until the switch is completely
550 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
554 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
557 * __task_rq_lock - lock the runqueue a given task resides on.
558 * Must be called interrupts disabled.
560 static inline struct rq *__task_rq_lock(struct task_struct *p)
567 spin_lock(&rq->lock);
568 if (unlikely(rq != task_rq(p))) {
569 spin_unlock(&rq->lock);
570 goto repeat_lock_task;
576 * task_rq_lock - lock the runqueue a given task resides on and disable
577 * interrupts. Note the ordering: we can safely lookup the task_rq without
578 * explicitly disabling preemption.
580 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
586 local_irq_save(*flags);
588 spin_lock(&rq->lock);
589 if (unlikely(rq != task_rq(p))) {
590 spin_unlock_irqrestore(&rq->lock, *flags);
591 goto repeat_lock_task;
596 static inline void __task_rq_unlock(struct rq *rq)
599 spin_unlock(&rq->lock);
602 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
605 spin_unlock_irqrestore(&rq->lock, *flags);
609 * this_rq_lock - lock this runqueue and disable interrupts.
611 static inline struct rq *this_rq_lock(void)
618 spin_lock(&rq->lock);
624 * We are going deep-idle (irqs are disabled):
626 void sched_clock_idle_sleep_event(void)
628 struct rq *rq = cpu_rq(smp_processor_id());
630 spin_lock(&rq->lock);
631 __update_rq_clock(rq);
632 spin_unlock(&rq->lock);
633 rq->clock_deep_idle_events++;
635 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
638 * We just idled delta nanoseconds (called with irqs disabled):
640 void sched_clock_idle_wakeup_event(u64 delta_ns)
642 struct rq *rq = cpu_rq(smp_processor_id());
643 u64 now = sched_clock();
645 rq->idle_clock += delta_ns;
647 * Override the previous timestamp and ignore all
648 * sched_clock() deltas that occured while we idled,
649 * and use the PM-provided delta_ns to advance the
652 spin_lock(&rq->lock);
653 rq->prev_clock_raw = now;
654 rq->clock += delta_ns;
655 spin_unlock(&rq->lock);
657 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
660 * resched_task - mark a task 'to be rescheduled now'.
662 * On UP this means the setting of the need_resched flag, on SMP it
663 * might also involve a cross-CPU call to trigger the scheduler on
668 #ifndef tsk_is_polling
669 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
672 static void resched_task(struct task_struct *p)
676 assert_spin_locked(&task_rq(p)->lock);
678 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
681 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
684 if (cpu == smp_processor_id())
687 /* NEED_RESCHED must be visible before we test polling */
689 if (!tsk_is_polling(p))
690 smp_send_reschedule(cpu);
693 static void resched_cpu(int cpu)
695 struct rq *rq = cpu_rq(cpu);
698 if (!spin_trylock_irqsave(&rq->lock, flags))
700 resched_task(cpu_curr(cpu));
701 spin_unlock_irqrestore(&rq->lock, flags);
704 static inline void resched_task(struct task_struct *p)
706 assert_spin_locked(&task_rq(p)->lock);
707 set_tsk_need_resched(p);
711 #if BITS_PER_LONG == 32
712 # define WMULT_CONST (~0UL)
714 # define WMULT_CONST (1UL << 32)
717 #define WMULT_SHIFT 32
720 * Shift right and round:
722 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
725 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
726 struct load_weight *lw)
730 if (unlikely(!lw->inv_weight))
731 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
733 tmp = (u64)delta_exec * weight;
735 * Check whether we'd overflow the 64-bit multiplication:
737 if (unlikely(tmp > WMULT_CONST))
738 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
741 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
743 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
746 static inline unsigned long
747 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
749 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
752 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
757 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
763 * To aid in avoiding the subversion of "niceness" due to uneven distribution
764 * of tasks with abnormal "nice" values across CPUs the contribution that
765 * each task makes to its run queue's load is weighted according to its
766 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
767 * scaled version of the new time slice allocation that they receive on time
771 #define WEIGHT_IDLEPRIO 2
772 #define WMULT_IDLEPRIO (1 << 31)
775 * Nice levels are multiplicative, with a gentle 10% change for every
776 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
777 * nice 1, it will get ~10% less CPU time than another CPU-bound task
778 * that remained on nice 0.
780 * The "10% effect" is relative and cumulative: from _any_ nice level,
781 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
782 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
783 * If a task goes up by ~10% and another task goes down by ~10% then
784 * the relative distance between them is ~25%.)
786 static const int prio_to_weight[40] = {
787 /* -20 */ 88761, 71755, 56483, 46273, 36291,
788 /* -15 */ 29154, 23254, 18705, 14949, 11916,
789 /* -10 */ 9548, 7620, 6100, 4904, 3906,
790 /* -5 */ 3121, 2501, 1991, 1586, 1277,
791 /* 0 */ 1024, 820, 655, 526, 423,
792 /* 5 */ 335, 272, 215, 172, 137,
793 /* 10 */ 110, 87, 70, 56, 45,
794 /* 15 */ 36, 29, 23, 18, 15,
798 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
800 * In cases where the weight does not change often, we can use the
801 * precalculated inverse to speed up arithmetics by turning divisions
802 * into multiplications:
804 static const u32 prio_to_wmult[40] = {
805 /* -20 */ 48388, 59856, 76040, 92818, 118348,
806 /* -15 */ 147320, 184698, 229616, 287308, 360437,
807 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
808 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
809 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
810 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
811 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
812 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
815 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
818 * runqueue iterator, to support SMP load-balancing between different
819 * scheduling classes, without having to expose their internal data
820 * structures to the load-balancing proper:
824 struct task_struct *(*start)(void *);
825 struct task_struct *(*next)(void *);
828 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
829 unsigned long max_nr_move, unsigned long max_load_move,
830 struct sched_domain *sd, enum cpu_idle_type idle,
831 int *all_pinned, unsigned long *load_moved,
832 int *this_best_prio, struct rq_iterator *iterator);
834 #include "sched_stats.h"
835 #include "sched_rt.c"
836 #include "sched_fair.c"
837 #include "sched_idletask.c"
838 #ifdef CONFIG_SCHED_DEBUG
839 # include "sched_debug.c"
842 #define sched_class_highest (&rt_sched_class)
845 * Update delta_exec, delta_fair fields for rq.
847 * delta_fair clock advances at a rate inversely proportional to
848 * total load (rq->load.weight) on the runqueue, while
849 * delta_exec advances at the same rate as wall-clock (provided
852 * delta_exec / delta_fair is a measure of the (smoothened) load on this
853 * runqueue over any given interval. This (smoothened) load is used
854 * during load balance.
856 * This function is called /before/ updating rq->load
857 * and when switching tasks.
859 static inline void inc_load(struct rq *rq, const struct task_struct *p)
861 update_load_add(&rq->load, p->se.load.weight);
864 static inline void dec_load(struct rq *rq, const struct task_struct *p)
866 update_load_sub(&rq->load, p->se.load.weight);
869 static void inc_nr_running(struct task_struct *p, struct rq *rq)
875 static void dec_nr_running(struct task_struct *p, struct rq *rq)
881 static void set_load_weight(struct task_struct *p)
883 if (task_has_rt_policy(p)) {
884 p->se.load.weight = prio_to_weight[0] * 2;
885 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
890 * SCHED_IDLE tasks get minimal weight:
892 if (p->policy == SCHED_IDLE) {
893 p->se.load.weight = WEIGHT_IDLEPRIO;
894 p->se.load.inv_weight = WMULT_IDLEPRIO;
898 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
899 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
902 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
904 sched_info_queued(p);
905 p->sched_class->enqueue_task(rq, p, wakeup);
909 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
911 p->sched_class->dequeue_task(rq, p, sleep);
916 * __normal_prio - return the priority that is based on the static prio
918 static inline int __normal_prio(struct task_struct *p)
920 return p->static_prio;
924 * Calculate the expected normal priority: i.e. priority
925 * without taking RT-inheritance into account. Might be
926 * boosted by interactivity modifiers. Changes upon fork,
927 * setprio syscalls, and whenever the interactivity
928 * estimator recalculates.
930 static inline int normal_prio(struct task_struct *p)
934 if (task_has_rt_policy(p))
935 prio = MAX_RT_PRIO-1 - p->rt_priority;
937 prio = __normal_prio(p);
942 * Calculate the current priority, i.e. the priority
943 * taken into account by the scheduler. This value might
944 * be boosted by RT tasks, or might be boosted by
945 * interactivity modifiers. Will be RT if the task got
946 * RT-boosted. If not then it returns p->normal_prio.
948 static int effective_prio(struct task_struct *p)
950 p->normal_prio = normal_prio(p);
952 * If we are RT tasks or we were boosted to RT priority,
953 * keep the priority unchanged. Otherwise, update priority
954 * to the normal priority:
956 if (!rt_prio(p->prio))
957 return p->normal_prio;
962 * activate_task - move a task to the runqueue.
964 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
966 if (p->state == TASK_UNINTERRUPTIBLE)
967 rq->nr_uninterruptible--;
969 enqueue_task(rq, p, wakeup);
970 inc_nr_running(p, rq);
974 * activate_idle_task - move idle task to the _front_ of runqueue.
976 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
980 if (p->state == TASK_UNINTERRUPTIBLE)
981 rq->nr_uninterruptible--;
983 enqueue_task(rq, p, 0);
984 inc_nr_running(p, rq);
988 * deactivate_task - remove a task from the runqueue.
990 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
992 if (p->state == TASK_UNINTERRUPTIBLE)
993 rq->nr_uninterruptible++;
995 dequeue_task(rq, p, sleep);
996 dec_nr_running(p, rq);
1000 * task_curr - is this task currently executing on a CPU?
1001 * @p: the task in question.
1003 inline int task_curr(const struct task_struct *p)
1005 return cpu_curr(task_cpu(p)) == p;
1008 /* Used instead of source_load when we know the type == 0 */
1009 unsigned long weighted_cpuload(const int cpu)
1011 return cpu_rq(cpu)->load.weight;
1014 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1017 task_thread_info(p)->cpu = cpu;
1024 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1026 int old_cpu = task_cpu(p);
1027 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1030 clock_offset = old_rq->clock - new_rq->clock;
1032 #ifdef CONFIG_SCHEDSTATS
1033 if (p->se.wait_start)
1034 p->se.wait_start -= clock_offset;
1035 if (p->se.sleep_start)
1036 p->se.sleep_start -= clock_offset;
1037 if (p->se.block_start)
1038 p->se.block_start -= clock_offset;
1040 if (likely(new_rq->cfs.min_vruntime))
1041 p->se.vruntime -= old_rq->cfs.min_vruntime -
1042 new_rq->cfs.min_vruntime;
1044 __set_task_cpu(p, new_cpu);
1047 struct migration_req {
1048 struct list_head list;
1050 struct task_struct *task;
1053 struct completion done;
1057 * The task's runqueue lock must be held.
1058 * Returns true if you have to wait for migration thread.
1061 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1063 struct rq *rq = task_rq(p);
1066 * If the task is not on a runqueue (and not running), then
1067 * it is sufficient to simply update the task's cpu field.
1069 if (!p->se.on_rq && !task_running(rq, p)) {
1070 set_task_cpu(p, dest_cpu);
1074 init_completion(&req->done);
1076 req->dest_cpu = dest_cpu;
1077 list_add(&req->list, &rq->migration_queue);
1083 * wait_task_inactive - wait for a thread to unschedule.
1085 * The caller must ensure that the task *will* unschedule sometime soon,
1086 * else this function might spin for a *long* time. This function can't
1087 * be called with interrupts off, or it may introduce deadlock with
1088 * smp_call_function() if an IPI is sent by the same process we are
1089 * waiting to become inactive.
1091 void wait_task_inactive(struct task_struct *p)
1093 unsigned long flags;
1099 * We do the initial early heuristics without holding
1100 * any task-queue locks at all. We'll only try to get
1101 * the runqueue lock when things look like they will
1107 * If the task is actively running on another CPU
1108 * still, just relax and busy-wait without holding
1111 * NOTE! Since we don't hold any locks, it's not
1112 * even sure that "rq" stays as the right runqueue!
1113 * But we don't care, since "task_running()" will
1114 * return false if the runqueue has changed and p
1115 * is actually now running somewhere else!
1117 while (task_running(rq, p))
1121 * Ok, time to look more closely! We need the rq
1122 * lock now, to be *sure*. If we're wrong, we'll
1123 * just go back and repeat.
1125 rq = task_rq_lock(p, &flags);
1126 running = task_running(rq, p);
1127 on_rq = p->se.on_rq;
1128 task_rq_unlock(rq, &flags);
1131 * Was it really running after all now that we
1132 * checked with the proper locks actually held?
1134 * Oops. Go back and try again..
1136 if (unlikely(running)) {
1142 * It's not enough that it's not actively running,
1143 * it must be off the runqueue _entirely_, and not
1146 * So if it wa still runnable (but just not actively
1147 * running right now), it's preempted, and we should
1148 * yield - it could be a while.
1150 if (unlikely(on_rq)) {
1156 * Ahh, all good. It wasn't running, and it wasn't
1157 * runnable, which means that it will never become
1158 * running in the future either. We're all done!
1163 * kick_process - kick a running thread to enter/exit the kernel
1164 * @p: the to-be-kicked thread
1166 * Cause a process which is running on another CPU to enter
1167 * kernel-mode, without any delay. (to get signals handled.)
1169 * NOTE: this function doesnt have to take the runqueue lock,
1170 * because all it wants to ensure is that the remote task enters
1171 * the kernel. If the IPI races and the task has been migrated
1172 * to another CPU then no harm is done and the purpose has been
1175 void kick_process(struct task_struct *p)
1181 if ((cpu != smp_processor_id()) && task_curr(p))
1182 smp_send_reschedule(cpu);
1187 * Return a low guess at the load of a migration-source cpu weighted
1188 * according to the scheduling class and "nice" value.
1190 * We want to under-estimate the load of migration sources, to
1191 * balance conservatively.
1193 static inline unsigned long source_load(int cpu, int type)
1195 struct rq *rq = cpu_rq(cpu);
1196 unsigned long total = weighted_cpuload(cpu);
1201 return min(rq->cpu_load[type-1], total);
1205 * Return a high guess at the load of a migration-target cpu weighted
1206 * according to the scheduling class and "nice" value.
1208 static inline unsigned long target_load(int cpu, int type)
1210 struct rq *rq = cpu_rq(cpu);
1211 unsigned long total = weighted_cpuload(cpu);
1216 return max(rq->cpu_load[type-1], total);
1220 * Return the average load per task on the cpu's run queue
1222 static inline unsigned long cpu_avg_load_per_task(int cpu)
1224 struct rq *rq = cpu_rq(cpu);
1225 unsigned long total = weighted_cpuload(cpu);
1226 unsigned long n = rq->nr_running;
1228 return n ? total / n : SCHED_LOAD_SCALE;
1232 * find_idlest_group finds and returns the least busy CPU group within the
1235 static struct sched_group *
1236 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1238 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1239 unsigned long min_load = ULONG_MAX, this_load = 0;
1240 int load_idx = sd->forkexec_idx;
1241 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1244 unsigned long load, avg_load;
1248 /* Skip over this group if it has no CPUs allowed */
1249 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1252 local_group = cpu_isset(this_cpu, group->cpumask);
1254 /* Tally up the load of all CPUs in the group */
1257 for_each_cpu_mask(i, group->cpumask) {
1258 /* Bias balancing toward cpus of our domain */
1260 load = source_load(i, load_idx);
1262 load = target_load(i, load_idx);
1267 /* Adjust by relative CPU power of the group */
1268 avg_load = sg_div_cpu_power(group,
1269 avg_load * SCHED_LOAD_SCALE);
1272 this_load = avg_load;
1274 } else if (avg_load < min_load) {
1275 min_load = avg_load;
1279 group = group->next;
1280 } while (group != sd->groups);
1282 if (!idlest || 100*this_load < imbalance*min_load)
1288 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1291 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1294 unsigned long load, min_load = ULONG_MAX;
1298 /* Traverse only the allowed CPUs */
1299 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1301 for_each_cpu_mask(i, tmp) {
1302 load = weighted_cpuload(i);
1304 if (load < min_load || (load == min_load && i == this_cpu)) {
1314 * sched_balance_self: balance the current task (running on cpu) in domains
1315 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1318 * Balance, ie. select the least loaded group.
1320 * Returns the target CPU number, or the same CPU if no balancing is needed.
1322 * preempt must be disabled.
1324 static int sched_balance_self(int cpu, int flag)
1326 struct task_struct *t = current;
1327 struct sched_domain *tmp, *sd = NULL;
1329 for_each_domain(cpu, tmp) {
1331 * If power savings logic is enabled for a domain, stop there.
1333 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1335 if (tmp->flags & flag)
1341 struct sched_group *group;
1342 int new_cpu, weight;
1344 if (!(sd->flags & flag)) {
1350 group = find_idlest_group(sd, t, cpu);
1356 new_cpu = find_idlest_cpu(group, t, cpu);
1357 if (new_cpu == -1 || new_cpu == cpu) {
1358 /* Now try balancing at a lower domain level of cpu */
1363 /* Now try balancing at a lower domain level of new_cpu */
1366 weight = cpus_weight(span);
1367 for_each_domain(cpu, tmp) {
1368 if (weight <= cpus_weight(tmp->span))
1370 if (tmp->flags & flag)
1373 /* while loop will break here if sd == NULL */
1379 #endif /* CONFIG_SMP */
1382 * wake_idle() will wake a task on an idle cpu if task->cpu is
1383 * not idle and an idle cpu is available. The span of cpus to
1384 * search starts with cpus closest then further out as needed,
1385 * so we always favor a closer, idle cpu.
1387 * Returns the CPU we should wake onto.
1389 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1390 static int wake_idle(int cpu, struct task_struct *p)
1393 struct sched_domain *sd;
1397 * If it is idle, then it is the best cpu to run this task.
1399 * This cpu is also the best, if it has more than one task already.
1400 * Siblings must be also busy(in most cases) as they didn't already
1401 * pickup the extra load from this cpu and hence we need not check
1402 * sibling runqueue info. This will avoid the checks and cache miss
1403 * penalities associated with that.
1405 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1408 for_each_domain(cpu, sd) {
1409 if (sd->flags & SD_WAKE_IDLE) {
1410 cpus_and(tmp, sd->span, p->cpus_allowed);
1411 for_each_cpu_mask(i, tmp) {
1422 static inline int wake_idle(int cpu, struct task_struct *p)
1429 * try_to_wake_up - wake up a thread
1430 * @p: the to-be-woken-up thread
1431 * @state: the mask of task states that can be woken
1432 * @sync: do a synchronous wakeup?
1434 * Put it on the run-queue if it's not already there. The "current"
1435 * thread is always on the run-queue (except when the actual
1436 * re-schedule is in progress), and as such you're allowed to do
1437 * the simpler "current->state = TASK_RUNNING" to mark yourself
1438 * runnable without the overhead of this.
1440 * returns failure only if the task is already active.
1442 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1444 int cpu, this_cpu, success = 0;
1445 unsigned long flags;
1449 struct sched_domain *sd, *this_sd = NULL;
1450 unsigned long load, this_load;
1454 rq = task_rq_lock(p, &flags);
1455 old_state = p->state;
1456 if (!(old_state & state))
1463 this_cpu = smp_processor_id();
1466 if (unlikely(task_running(rq, p)))
1471 schedstat_inc(rq, ttwu_cnt);
1472 if (cpu == this_cpu) {
1473 schedstat_inc(rq, ttwu_local);
1477 for_each_domain(this_cpu, sd) {
1478 if (cpu_isset(cpu, sd->span)) {
1479 schedstat_inc(sd, ttwu_wake_remote);
1485 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1489 * Check for affine wakeup and passive balancing possibilities.
1492 int idx = this_sd->wake_idx;
1493 unsigned int imbalance;
1495 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1497 load = source_load(cpu, idx);
1498 this_load = target_load(this_cpu, idx);
1500 new_cpu = this_cpu; /* Wake to this CPU if we can */
1502 if (this_sd->flags & SD_WAKE_AFFINE) {
1503 unsigned long tl = this_load;
1504 unsigned long tl_per_task;
1506 tl_per_task = cpu_avg_load_per_task(this_cpu);
1509 * If sync wakeup then subtract the (maximum possible)
1510 * effect of the currently running task from the load
1511 * of the current CPU:
1514 tl -= current->se.load.weight;
1517 tl + target_load(cpu, idx) <= tl_per_task) ||
1518 100*(tl + p->se.load.weight) <= imbalance*load) {
1520 * This domain has SD_WAKE_AFFINE and
1521 * p is cache cold in this domain, and
1522 * there is no bad imbalance.
1524 schedstat_inc(this_sd, ttwu_move_affine);
1530 * Start passive balancing when half the imbalance_pct
1533 if (this_sd->flags & SD_WAKE_BALANCE) {
1534 if (imbalance*this_load <= 100*load) {
1535 schedstat_inc(this_sd, ttwu_move_balance);
1541 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1543 new_cpu = wake_idle(new_cpu, p);
1544 if (new_cpu != cpu) {
1545 set_task_cpu(p, new_cpu);
1546 task_rq_unlock(rq, &flags);
1547 /* might preempt at this point */
1548 rq = task_rq_lock(p, &flags);
1549 old_state = p->state;
1550 if (!(old_state & state))
1555 this_cpu = smp_processor_id();
1560 #endif /* CONFIG_SMP */
1561 update_rq_clock(rq);
1562 activate_task(rq, p, 1);
1564 * Sync wakeups (i.e. those types of wakeups where the waker
1565 * has indicated that it will leave the CPU in short order)
1566 * don't trigger a preemption, if the woken up task will run on
1567 * this cpu. (in this case the 'I will reschedule' promise of
1568 * the waker guarantees that the freshly woken up task is going
1569 * to be considered on this CPU.)
1571 if (!sync || cpu != this_cpu)
1572 check_preempt_curr(rq, p);
1576 p->state = TASK_RUNNING;
1578 task_rq_unlock(rq, &flags);
1583 int fastcall wake_up_process(struct task_struct *p)
1585 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1586 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1588 EXPORT_SYMBOL(wake_up_process);
1590 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1592 return try_to_wake_up(p, state, 0);
1596 * Perform scheduler related setup for a newly forked process p.
1597 * p is forked by current.
1599 * __sched_fork() is basic setup used by init_idle() too:
1601 static void __sched_fork(struct task_struct *p)
1603 p->se.exec_start = 0;
1604 p->se.sum_exec_runtime = 0;
1605 p->se.prev_sum_exec_runtime = 0;
1607 #ifdef CONFIG_SCHEDSTATS
1608 p->se.wait_start = 0;
1609 p->se.sum_sleep_runtime = 0;
1610 p->se.sleep_start = 0;
1611 p->se.block_start = 0;
1612 p->se.sleep_max = 0;
1613 p->se.block_max = 0;
1615 p->se.slice_max = 0;
1619 INIT_LIST_HEAD(&p->run_list);
1622 #ifdef CONFIG_PREEMPT_NOTIFIERS
1623 INIT_HLIST_HEAD(&p->preempt_notifiers);
1627 * We mark the process as running here, but have not actually
1628 * inserted it onto the runqueue yet. This guarantees that
1629 * nobody will actually run it, and a signal or other external
1630 * event cannot wake it up and insert it on the runqueue either.
1632 p->state = TASK_RUNNING;
1636 * fork()/clone()-time setup:
1638 void sched_fork(struct task_struct *p, int clone_flags)
1640 int cpu = get_cpu();
1645 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1647 __set_task_cpu(p, cpu);
1650 * Make sure we do not leak PI boosting priority to the child:
1652 p->prio = current->normal_prio;
1654 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1655 if (likely(sched_info_on()))
1656 memset(&p->sched_info, 0, sizeof(p->sched_info));
1658 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1661 #ifdef CONFIG_PREEMPT
1662 /* Want to start with kernel preemption disabled. */
1663 task_thread_info(p)->preempt_count = 1;
1669 * wake_up_new_task - wake up a newly created task for the first time.
1671 * This function will do some initial scheduler statistics housekeeping
1672 * that must be done for every newly created context, then puts the task
1673 * on the runqueue and wakes it.
1675 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1677 unsigned long flags;
1681 rq = task_rq_lock(p, &flags);
1682 BUG_ON(p->state != TASK_RUNNING);
1683 this_cpu = smp_processor_id(); /* parent's CPU */
1684 update_rq_clock(rq);
1686 p->prio = effective_prio(p);
1688 if (rt_prio(p->prio))
1689 p->sched_class = &rt_sched_class;
1691 p->sched_class = &fair_sched_class;
1693 if (task_cpu(p) != this_cpu || !p->sched_class->task_new ||
1694 !current->se.on_rq) {
1695 activate_task(rq, p, 0);
1698 * Let the scheduling class do new task startup
1699 * management (if any):
1701 p->sched_class->task_new(rq, p);
1702 inc_nr_running(p, rq);
1704 check_preempt_curr(rq, p);
1705 task_rq_unlock(rq, &flags);
1708 #ifdef CONFIG_PREEMPT_NOTIFIERS
1711 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1712 * @notifier: notifier struct to register
1714 void preempt_notifier_register(struct preempt_notifier *notifier)
1716 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1718 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1721 * preempt_notifier_unregister - no longer interested in preemption notifications
1722 * @notifier: notifier struct to unregister
1724 * This is safe to call from within a preemption notifier.
1726 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1728 hlist_del(¬ifier->link);
1730 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1732 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1734 struct preempt_notifier *notifier;
1735 struct hlist_node *node;
1737 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1738 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1742 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1743 struct task_struct *next)
1745 struct preempt_notifier *notifier;
1746 struct hlist_node *node;
1748 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1749 notifier->ops->sched_out(notifier, next);
1754 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1759 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1760 struct task_struct *next)
1767 * prepare_task_switch - prepare to switch tasks
1768 * @rq: the runqueue preparing to switch
1769 * @prev: the current task that is being switched out
1770 * @next: the task we are going to switch to.
1772 * This is called with the rq lock held and interrupts off. It must
1773 * be paired with a subsequent finish_task_switch after the context
1776 * prepare_task_switch sets up locking and calls architecture specific
1780 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1781 struct task_struct *next)
1783 fire_sched_out_preempt_notifiers(prev, next);
1784 prepare_lock_switch(rq, next);
1785 prepare_arch_switch(next);
1789 * finish_task_switch - clean up after a task-switch
1790 * @rq: runqueue associated with task-switch
1791 * @prev: the thread we just switched away from.
1793 * finish_task_switch must be called after the context switch, paired
1794 * with a prepare_task_switch call before the context switch.
1795 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1796 * and do any other architecture-specific cleanup actions.
1798 * Note that we may have delayed dropping an mm in context_switch(). If
1799 * so, we finish that here outside of the runqueue lock. (Doing it
1800 * with the lock held can cause deadlocks; see schedule() for
1803 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1804 __releases(rq->lock)
1806 struct mm_struct *mm = rq->prev_mm;
1812 * A task struct has one reference for the use as "current".
1813 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1814 * schedule one last time. The schedule call will never return, and
1815 * the scheduled task must drop that reference.
1816 * The test for TASK_DEAD must occur while the runqueue locks are
1817 * still held, otherwise prev could be scheduled on another cpu, die
1818 * there before we look at prev->state, and then the reference would
1820 * Manfred Spraul <manfred@colorfullife.com>
1822 prev_state = prev->state;
1823 finish_arch_switch(prev);
1824 finish_lock_switch(rq, prev);
1825 fire_sched_in_preempt_notifiers(current);
1828 if (unlikely(prev_state == TASK_DEAD)) {
1830 * Remove function-return probe instances associated with this
1831 * task and put them back on the free list.
1833 kprobe_flush_task(prev);
1834 put_task_struct(prev);
1839 * schedule_tail - first thing a freshly forked thread must call.
1840 * @prev: the thread we just switched away from.
1842 asmlinkage void schedule_tail(struct task_struct *prev)
1843 __releases(rq->lock)
1845 struct rq *rq = this_rq();
1847 finish_task_switch(rq, prev);
1848 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1849 /* In this case, finish_task_switch does not reenable preemption */
1852 if (current->set_child_tid)
1853 put_user(current->pid, current->set_child_tid);
1857 * context_switch - switch to the new MM and the new
1858 * thread's register state.
1861 context_switch(struct rq *rq, struct task_struct *prev,
1862 struct task_struct *next)
1864 struct mm_struct *mm, *oldmm;
1866 prepare_task_switch(rq, prev, next);
1868 oldmm = prev->active_mm;
1870 * For paravirt, this is coupled with an exit in switch_to to
1871 * combine the page table reload and the switch backend into
1874 arch_enter_lazy_cpu_mode();
1876 if (unlikely(!mm)) {
1877 next->active_mm = oldmm;
1878 atomic_inc(&oldmm->mm_count);
1879 enter_lazy_tlb(oldmm, next);
1881 switch_mm(oldmm, mm, next);
1883 if (unlikely(!prev->mm)) {
1884 prev->active_mm = NULL;
1885 rq->prev_mm = oldmm;
1888 * Since the runqueue lock will be released by the next
1889 * task (which is an invalid locking op but in the case
1890 * of the scheduler it's an obvious special-case), so we
1891 * do an early lockdep release here:
1893 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1894 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1897 /* Here we just switch the register state and the stack. */
1898 switch_to(prev, next, prev);
1902 * this_rq must be evaluated again because prev may have moved
1903 * CPUs since it called schedule(), thus the 'rq' on its stack
1904 * frame will be invalid.
1906 finish_task_switch(this_rq(), prev);
1910 * nr_running, nr_uninterruptible and nr_context_switches:
1912 * externally visible scheduler statistics: current number of runnable
1913 * threads, current number of uninterruptible-sleeping threads, total
1914 * number of context switches performed since bootup.
1916 unsigned long nr_running(void)
1918 unsigned long i, sum = 0;
1920 for_each_online_cpu(i)
1921 sum += cpu_rq(i)->nr_running;
1926 unsigned long nr_uninterruptible(void)
1928 unsigned long i, sum = 0;
1930 for_each_possible_cpu(i)
1931 sum += cpu_rq(i)->nr_uninterruptible;
1934 * Since we read the counters lockless, it might be slightly
1935 * inaccurate. Do not allow it to go below zero though:
1937 if (unlikely((long)sum < 0))
1943 unsigned long long nr_context_switches(void)
1946 unsigned long long sum = 0;
1948 for_each_possible_cpu(i)
1949 sum += cpu_rq(i)->nr_switches;
1954 unsigned long nr_iowait(void)
1956 unsigned long i, sum = 0;
1958 for_each_possible_cpu(i)
1959 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1964 unsigned long nr_active(void)
1966 unsigned long i, running = 0, uninterruptible = 0;
1968 for_each_online_cpu(i) {
1969 running += cpu_rq(i)->nr_running;
1970 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1973 if (unlikely((long)uninterruptible < 0))
1974 uninterruptible = 0;
1976 return running + uninterruptible;
1980 * Update rq->cpu_load[] statistics. This function is usually called every
1981 * scheduler tick (TICK_NSEC).
1983 static void update_cpu_load(struct rq *this_rq)
1985 unsigned long this_load = this_rq->load.weight;
1988 this_rq->nr_load_updates++;
1990 /* Update our load: */
1991 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1992 unsigned long old_load, new_load;
1994 /* scale is effectively 1 << i now, and >> i divides by scale */
1996 old_load = this_rq->cpu_load[i];
1997 new_load = this_load;
1999 * Round up the averaging division if load is increasing. This
2000 * prevents us from getting stuck on 9 if the load is 10, for
2003 if (new_load > old_load)
2004 new_load += scale-1;
2005 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2012 * double_rq_lock - safely lock two runqueues
2014 * Note this does not disable interrupts like task_rq_lock,
2015 * you need to do so manually before calling.
2017 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2018 __acquires(rq1->lock)
2019 __acquires(rq2->lock)
2021 BUG_ON(!irqs_disabled());
2023 spin_lock(&rq1->lock);
2024 __acquire(rq2->lock); /* Fake it out ;) */
2027 spin_lock(&rq1->lock);
2028 spin_lock(&rq2->lock);
2030 spin_lock(&rq2->lock);
2031 spin_lock(&rq1->lock);
2034 update_rq_clock(rq1);
2035 update_rq_clock(rq2);
2039 * double_rq_unlock - safely unlock two runqueues
2041 * Note this does not restore interrupts like task_rq_unlock,
2042 * you need to do so manually after calling.
2044 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2045 __releases(rq1->lock)
2046 __releases(rq2->lock)
2048 spin_unlock(&rq1->lock);
2050 spin_unlock(&rq2->lock);
2052 __release(rq2->lock);
2056 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2058 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2059 __releases(this_rq->lock)
2060 __acquires(busiest->lock)
2061 __acquires(this_rq->lock)
2063 if (unlikely(!irqs_disabled())) {
2064 /* printk() doesn't work good under rq->lock */
2065 spin_unlock(&this_rq->lock);
2068 if (unlikely(!spin_trylock(&busiest->lock))) {
2069 if (busiest < this_rq) {
2070 spin_unlock(&this_rq->lock);
2071 spin_lock(&busiest->lock);
2072 spin_lock(&this_rq->lock);
2074 spin_lock(&busiest->lock);
2079 * If dest_cpu is allowed for this process, migrate the task to it.
2080 * This is accomplished by forcing the cpu_allowed mask to only
2081 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2082 * the cpu_allowed mask is restored.
2084 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2086 struct migration_req req;
2087 unsigned long flags;
2090 rq = task_rq_lock(p, &flags);
2091 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2092 || unlikely(cpu_is_offline(dest_cpu)))
2095 /* force the process onto the specified CPU */
2096 if (migrate_task(p, dest_cpu, &req)) {
2097 /* Need to wait for migration thread (might exit: take ref). */
2098 struct task_struct *mt = rq->migration_thread;
2100 get_task_struct(mt);
2101 task_rq_unlock(rq, &flags);
2102 wake_up_process(mt);
2103 put_task_struct(mt);
2104 wait_for_completion(&req.done);
2109 task_rq_unlock(rq, &flags);
2113 * sched_exec - execve() is a valuable balancing opportunity, because at
2114 * this point the task has the smallest effective memory and cache footprint.
2116 void sched_exec(void)
2118 int new_cpu, this_cpu = get_cpu();
2119 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2121 if (new_cpu != this_cpu)
2122 sched_migrate_task(current, new_cpu);
2126 * pull_task - move a task from a remote runqueue to the local runqueue.
2127 * Both runqueues must be locked.
2129 static void pull_task(struct rq *src_rq, struct task_struct *p,
2130 struct rq *this_rq, int this_cpu)
2132 deactivate_task(src_rq, p, 0);
2133 set_task_cpu(p, this_cpu);
2134 activate_task(this_rq, p, 0);
2136 * Note that idle threads have a prio of MAX_PRIO, for this test
2137 * to be always true for them.
2139 check_preempt_curr(this_rq, p);
2143 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2146 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2147 struct sched_domain *sd, enum cpu_idle_type idle,
2151 * We do not migrate tasks that are:
2152 * 1) running (obviously), or
2153 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2154 * 3) are cache-hot on their current CPU.
2156 if (!cpu_isset(this_cpu, p->cpus_allowed))
2160 if (task_running(rq, p))
2166 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2167 unsigned long max_nr_move, unsigned long max_load_move,
2168 struct sched_domain *sd, enum cpu_idle_type idle,
2169 int *all_pinned, unsigned long *load_moved,
2170 int *this_best_prio, struct rq_iterator *iterator)
2172 int pulled = 0, pinned = 0, skip_for_load;
2173 struct task_struct *p;
2174 long rem_load_move = max_load_move;
2176 if (max_nr_move == 0 || max_load_move == 0)
2182 * Start the load-balancing iterator:
2184 p = iterator->start(iterator->arg);
2189 * To help distribute high priority tasks accross CPUs we don't
2190 * skip a task if it will be the highest priority task (i.e. smallest
2191 * prio value) on its new queue regardless of its load weight
2193 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2194 SCHED_LOAD_SCALE_FUZZ;
2195 if ((skip_for_load && p->prio >= *this_best_prio) ||
2196 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2197 p = iterator->next(iterator->arg);
2201 pull_task(busiest, p, this_rq, this_cpu);
2203 rem_load_move -= p->se.load.weight;
2206 * We only want to steal up to the prescribed number of tasks
2207 * and the prescribed amount of weighted load.
2209 if (pulled < max_nr_move && rem_load_move > 0) {
2210 if (p->prio < *this_best_prio)
2211 *this_best_prio = p->prio;
2212 p = iterator->next(iterator->arg);
2217 * Right now, this is the only place pull_task() is called,
2218 * so we can safely collect pull_task() stats here rather than
2219 * inside pull_task().
2221 schedstat_add(sd, lb_gained[idle], pulled);
2224 *all_pinned = pinned;
2225 *load_moved = max_load_move - rem_load_move;
2230 * move_tasks tries to move up to max_load_move weighted load from busiest to
2231 * this_rq, as part of a balancing operation within domain "sd".
2232 * Returns 1 if successful and 0 otherwise.
2234 * Called with both runqueues locked.
2236 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2237 unsigned long max_load_move,
2238 struct sched_domain *sd, enum cpu_idle_type idle,
2241 struct sched_class *class = sched_class_highest;
2242 unsigned long total_load_moved = 0;
2243 int this_best_prio = this_rq->curr->prio;
2247 class->load_balance(this_rq, this_cpu, busiest,
2248 ULONG_MAX, max_load_move - total_load_moved,
2249 sd, idle, all_pinned, &this_best_prio);
2250 class = class->next;
2251 } while (class && max_load_move > total_load_moved);
2253 return total_load_moved > 0;
2257 * move_one_task tries to move exactly one task from busiest to this_rq, as
2258 * part of active balancing operations within "domain".
2259 * Returns 1 if successful and 0 otherwise.
2261 * Called with both runqueues locked.
2263 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2264 struct sched_domain *sd, enum cpu_idle_type idle)
2266 struct sched_class *class;
2267 int this_best_prio = MAX_PRIO;
2269 for (class = sched_class_highest; class; class = class->next)
2270 if (class->load_balance(this_rq, this_cpu, busiest,
2271 1, ULONG_MAX, sd, idle, NULL,
2279 * find_busiest_group finds and returns the busiest CPU group within the
2280 * domain. It calculates and returns the amount of weighted load which
2281 * should be moved to restore balance via the imbalance parameter.
2283 static struct sched_group *
2284 find_busiest_group(struct sched_domain *sd, int this_cpu,
2285 unsigned long *imbalance, enum cpu_idle_type idle,
2286 int *sd_idle, cpumask_t *cpus, int *balance)
2288 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2289 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2290 unsigned long max_pull;
2291 unsigned long busiest_load_per_task, busiest_nr_running;
2292 unsigned long this_load_per_task, this_nr_running;
2294 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2295 int power_savings_balance = 1;
2296 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2297 unsigned long min_nr_running = ULONG_MAX;
2298 struct sched_group *group_min = NULL, *group_leader = NULL;
2301 max_load = this_load = total_load = total_pwr = 0;
2302 busiest_load_per_task = busiest_nr_running = 0;
2303 this_load_per_task = this_nr_running = 0;
2304 if (idle == CPU_NOT_IDLE)
2305 load_idx = sd->busy_idx;
2306 else if (idle == CPU_NEWLY_IDLE)
2307 load_idx = sd->newidle_idx;
2309 load_idx = sd->idle_idx;
2312 unsigned long load, group_capacity;
2315 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2316 unsigned long sum_nr_running, sum_weighted_load;
2318 local_group = cpu_isset(this_cpu, group->cpumask);
2321 balance_cpu = first_cpu(group->cpumask);
2323 /* Tally up the load of all CPUs in the group */
2324 sum_weighted_load = sum_nr_running = avg_load = 0;
2326 for_each_cpu_mask(i, group->cpumask) {
2329 if (!cpu_isset(i, *cpus))
2334 if (*sd_idle && rq->nr_running)
2337 /* Bias balancing toward cpus of our domain */
2339 if (idle_cpu(i) && !first_idle_cpu) {
2344 load = target_load(i, load_idx);
2346 load = source_load(i, load_idx);
2349 sum_nr_running += rq->nr_running;
2350 sum_weighted_load += weighted_cpuload(i);
2354 * First idle cpu or the first cpu(busiest) in this sched group
2355 * is eligible for doing load balancing at this and above
2356 * domains. In the newly idle case, we will allow all the cpu's
2357 * to do the newly idle load balance.
2359 if (idle != CPU_NEWLY_IDLE && local_group &&
2360 balance_cpu != this_cpu && balance) {
2365 total_load += avg_load;
2366 total_pwr += group->__cpu_power;
2368 /* Adjust by relative CPU power of the group */
2369 avg_load = sg_div_cpu_power(group,
2370 avg_load * SCHED_LOAD_SCALE);
2372 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2375 this_load = avg_load;
2377 this_nr_running = sum_nr_running;
2378 this_load_per_task = sum_weighted_load;
2379 } else if (avg_load > max_load &&
2380 sum_nr_running > group_capacity) {
2381 max_load = avg_load;
2383 busiest_nr_running = sum_nr_running;
2384 busiest_load_per_task = sum_weighted_load;
2387 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2389 * Busy processors will not participate in power savings
2392 if (idle == CPU_NOT_IDLE ||
2393 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2397 * If the local group is idle or completely loaded
2398 * no need to do power savings balance at this domain
2400 if (local_group && (this_nr_running >= group_capacity ||
2402 power_savings_balance = 0;
2405 * If a group is already running at full capacity or idle,
2406 * don't include that group in power savings calculations
2408 if (!power_savings_balance || sum_nr_running >= group_capacity
2413 * Calculate the group which has the least non-idle load.
2414 * This is the group from where we need to pick up the load
2417 if ((sum_nr_running < min_nr_running) ||
2418 (sum_nr_running == min_nr_running &&
2419 first_cpu(group->cpumask) <
2420 first_cpu(group_min->cpumask))) {
2422 min_nr_running = sum_nr_running;
2423 min_load_per_task = sum_weighted_load /
2428 * Calculate the group which is almost near its
2429 * capacity but still has some space to pick up some load
2430 * from other group and save more power
2432 if (sum_nr_running <= group_capacity - 1) {
2433 if (sum_nr_running > leader_nr_running ||
2434 (sum_nr_running == leader_nr_running &&
2435 first_cpu(group->cpumask) >
2436 first_cpu(group_leader->cpumask))) {
2437 group_leader = group;
2438 leader_nr_running = sum_nr_running;
2443 group = group->next;
2444 } while (group != sd->groups);
2446 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2449 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2451 if (this_load >= avg_load ||
2452 100*max_load <= sd->imbalance_pct*this_load)
2455 busiest_load_per_task /= busiest_nr_running;
2457 * We're trying to get all the cpus to the average_load, so we don't
2458 * want to push ourselves above the average load, nor do we wish to
2459 * reduce the max loaded cpu below the average load, as either of these
2460 * actions would just result in more rebalancing later, and ping-pong
2461 * tasks around. Thus we look for the minimum possible imbalance.
2462 * Negative imbalances (*we* are more loaded than anyone else) will
2463 * be counted as no imbalance for these purposes -- we can't fix that
2464 * by pulling tasks to us. Be careful of negative numbers as they'll
2465 * appear as very large values with unsigned longs.
2467 if (max_load <= busiest_load_per_task)
2471 * In the presence of smp nice balancing, certain scenarios can have
2472 * max load less than avg load(as we skip the groups at or below
2473 * its cpu_power, while calculating max_load..)
2475 if (max_load < avg_load) {
2477 goto small_imbalance;
2480 /* Don't want to pull so many tasks that a group would go idle */
2481 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2483 /* How much load to actually move to equalise the imbalance */
2484 *imbalance = min(max_pull * busiest->__cpu_power,
2485 (avg_load - this_load) * this->__cpu_power)
2489 * if *imbalance is less than the average load per runnable task
2490 * there is no gaurantee that any tasks will be moved so we'll have
2491 * a think about bumping its value to force at least one task to be
2494 if (*imbalance < busiest_load_per_task) {
2495 unsigned long tmp, pwr_now, pwr_move;
2499 pwr_move = pwr_now = 0;
2501 if (this_nr_running) {
2502 this_load_per_task /= this_nr_running;
2503 if (busiest_load_per_task > this_load_per_task)
2506 this_load_per_task = SCHED_LOAD_SCALE;
2508 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2509 busiest_load_per_task * imbn) {
2510 *imbalance = busiest_load_per_task;
2515 * OK, we don't have enough imbalance to justify moving tasks,
2516 * however we may be able to increase total CPU power used by
2520 pwr_now += busiest->__cpu_power *
2521 min(busiest_load_per_task, max_load);
2522 pwr_now += this->__cpu_power *
2523 min(this_load_per_task, this_load);
2524 pwr_now /= SCHED_LOAD_SCALE;
2526 /* Amount of load we'd subtract */
2527 tmp = sg_div_cpu_power(busiest,
2528 busiest_load_per_task * SCHED_LOAD_SCALE);
2530 pwr_move += busiest->__cpu_power *
2531 min(busiest_load_per_task, max_load - tmp);
2533 /* Amount of load we'd add */
2534 if (max_load * busiest->__cpu_power <
2535 busiest_load_per_task * SCHED_LOAD_SCALE)
2536 tmp = sg_div_cpu_power(this,
2537 max_load * busiest->__cpu_power);
2539 tmp = sg_div_cpu_power(this,
2540 busiest_load_per_task * SCHED_LOAD_SCALE);
2541 pwr_move += this->__cpu_power *
2542 min(this_load_per_task, this_load + tmp);
2543 pwr_move /= SCHED_LOAD_SCALE;
2545 /* Move if we gain throughput */
2546 if (pwr_move > pwr_now)
2547 *imbalance = busiest_load_per_task;
2553 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2554 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2557 if (this == group_leader && group_leader != group_min) {
2558 *imbalance = min_load_per_task;
2568 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2571 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2572 unsigned long imbalance, cpumask_t *cpus)
2574 struct rq *busiest = NULL, *rq;
2575 unsigned long max_load = 0;
2578 for_each_cpu_mask(i, group->cpumask) {
2581 if (!cpu_isset(i, *cpus))
2585 wl = weighted_cpuload(i);
2587 if (rq->nr_running == 1 && wl > imbalance)
2590 if (wl > max_load) {
2600 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2601 * so long as it is large enough.
2603 #define MAX_PINNED_INTERVAL 512
2606 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2607 * tasks if there is an imbalance.
2609 static int load_balance(int this_cpu, struct rq *this_rq,
2610 struct sched_domain *sd, enum cpu_idle_type idle,
2613 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2614 struct sched_group *group;
2615 unsigned long imbalance;
2617 cpumask_t cpus = CPU_MASK_ALL;
2618 unsigned long flags;
2621 * When power savings policy is enabled for the parent domain, idle
2622 * sibling can pick up load irrespective of busy siblings. In this case,
2623 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2624 * portraying it as CPU_NOT_IDLE.
2626 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2627 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2630 schedstat_inc(sd, lb_cnt[idle]);
2633 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2640 schedstat_inc(sd, lb_nobusyg[idle]);
2644 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2646 schedstat_inc(sd, lb_nobusyq[idle]);
2650 BUG_ON(busiest == this_rq);
2652 schedstat_add(sd, lb_imbalance[idle], imbalance);
2655 if (busiest->nr_running > 1) {
2657 * Attempt to move tasks. If find_busiest_group has found
2658 * an imbalance but busiest->nr_running <= 1, the group is
2659 * still unbalanced. ld_moved simply stays zero, so it is
2660 * correctly treated as an imbalance.
2662 local_irq_save(flags);
2663 double_rq_lock(this_rq, busiest);
2664 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2665 imbalance, sd, idle, &all_pinned);
2666 double_rq_unlock(this_rq, busiest);
2667 local_irq_restore(flags);
2670 * some other cpu did the load balance for us.
2672 if (ld_moved && this_cpu != smp_processor_id())
2673 resched_cpu(this_cpu);
2675 /* All tasks on this runqueue were pinned by CPU affinity */
2676 if (unlikely(all_pinned)) {
2677 cpu_clear(cpu_of(busiest), cpus);
2678 if (!cpus_empty(cpus))
2685 schedstat_inc(sd, lb_failed[idle]);
2686 sd->nr_balance_failed++;
2688 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2690 spin_lock_irqsave(&busiest->lock, flags);
2692 /* don't kick the migration_thread, if the curr
2693 * task on busiest cpu can't be moved to this_cpu
2695 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2696 spin_unlock_irqrestore(&busiest->lock, flags);
2698 goto out_one_pinned;
2701 if (!busiest->active_balance) {
2702 busiest->active_balance = 1;
2703 busiest->push_cpu = this_cpu;
2706 spin_unlock_irqrestore(&busiest->lock, flags);
2708 wake_up_process(busiest->migration_thread);
2711 * We've kicked active balancing, reset the failure
2714 sd->nr_balance_failed = sd->cache_nice_tries+1;
2717 sd->nr_balance_failed = 0;
2719 if (likely(!active_balance)) {
2720 /* We were unbalanced, so reset the balancing interval */
2721 sd->balance_interval = sd->min_interval;
2724 * If we've begun active balancing, start to back off. This
2725 * case may not be covered by the all_pinned logic if there
2726 * is only 1 task on the busy runqueue (because we don't call
2729 if (sd->balance_interval < sd->max_interval)
2730 sd->balance_interval *= 2;
2733 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2734 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2739 schedstat_inc(sd, lb_balanced[idle]);
2741 sd->nr_balance_failed = 0;
2744 /* tune up the balancing interval */
2745 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2746 (sd->balance_interval < sd->max_interval))
2747 sd->balance_interval *= 2;
2749 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2750 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2756 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2757 * tasks if there is an imbalance.
2759 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2760 * this_rq is locked.
2763 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2765 struct sched_group *group;
2766 struct rq *busiest = NULL;
2767 unsigned long imbalance;
2771 cpumask_t cpus = CPU_MASK_ALL;
2774 * When power savings policy is enabled for the parent domain, idle
2775 * sibling can pick up load irrespective of busy siblings. In this case,
2776 * let the state of idle sibling percolate up as IDLE, instead of
2777 * portraying it as CPU_NOT_IDLE.
2779 if (sd->flags & SD_SHARE_CPUPOWER &&
2780 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2783 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2785 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2786 &sd_idle, &cpus, NULL);
2788 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2792 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2795 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2799 BUG_ON(busiest == this_rq);
2801 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2804 if (busiest->nr_running > 1) {
2805 /* Attempt to move tasks */
2806 double_lock_balance(this_rq, busiest);
2807 /* this_rq->clock is already updated */
2808 update_rq_clock(busiest);
2809 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2810 imbalance, sd, CPU_NEWLY_IDLE,
2812 spin_unlock(&busiest->lock);
2814 if (unlikely(all_pinned)) {
2815 cpu_clear(cpu_of(busiest), cpus);
2816 if (!cpus_empty(cpus))
2822 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2823 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2824 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2827 sd->nr_balance_failed = 0;
2832 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2833 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2834 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2836 sd->nr_balance_failed = 0;
2842 * idle_balance is called by schedule() if this_cpu is about to become
2843 * idle. Attempts to pull tasks from other CPUs.
2845 static void idle_balance(int this_cpu, struct rq *this_rq)
2847 struct sched_domain *sd;
2848 int pulled_task = -1;
2849 unsigned long next_balance = jiffies + HZ;
2851 for_each_domain(this_cpu, sd) {
2852 unsigned long interval;
2854 if (!(sd->flags & SD_LOAD_BALANCE))
2857 if (sd->flags & SD_BALANCE_NEWIDLE)
2858 /* If we've pulled tasks over stop searching: */
2859 pulled_task = load_balance_newidle(this_cpu,
2862 interval = msecs_to_jiffies(sd->balance_interval);
2863 if (time_after(next_balance, sd->last_balance + interval))
2864 next_balance = sd->last_balance + interval;
2868 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2870 * We are going idle. next_balance may be set based on
2871 * a busy processor. So reset next_balance.
2873 this_rq->next_balance = next_balance;
2878 * active_load_balance is run by migration threads. It pushes running tasks
2879 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2880 * running on each physical CPU where possible, and avoids physical /
2881 * logical imbalances.
2883 * Called with busiest_rq locked.
2885 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2887 int target_cpu = busiest_rq->push_cpu;
2888 struct sched_domain *sd;
2889 struct rq *target_rq;
2891 /* Is there any task to move? */
2892 if (busiest_rq->nr_running <= 1)
2895 target_rq = cpu_rq(target_cpu);
2898 * This condition is "impossible", if it occurs
2899 * we need to fix it. Originally reported by
2900 * Bjorn Helgaas on a 128-cpu setup.
2902 BUG_ON(busiest_rq == target_rq);
2904 /* move a task from busiest_rq to target_rq */
2905 double_lock_balance(busiest_rq, target_rq);
2906 update_rq_clock(busiest_rq);
2907 update_rq_clock(target_rq);
2909 /* Search for an sd spanning us and the target CPU. */
2910 for_each_domain(target_cpu, sd) {
2911 if ((sd->flags & SD_LOAD_BALANCE) &&
2912 cpu_isset(busiest_cpu, sd->span))
2917 schedstat_inc(sd, alb_cnt);
2919 if (move_one_task(target_rq, target_cpu, busiest_rq,
2921 schedstat_inc(sd, alb_pushed);
2923 schedstat_inc(sd, alb_failed);
2925 spin_unlock(&target_rq->lock);
2930 atomic_t load_balancer;
2932 } nohz ____cacheline_aligned = {
2933 .load_balancer = ATOMIC_INIT(-1),
2934 .cpu_mask = CPU_MASK_NONE,
2938 * This routine will try to nominate the ilb (idle load balancing)
2939 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2940 * load balancing on behalf of all those cpus. If all the cpus in the system
2941 * go into this tickless mode, then there will be no ilb owner (as there is
2942 * no need for one) and all the cpus will sleep till the next wakeup event
2945 * For the ilb owner, tick is not stopped. And this tick will be used
2946 * for idle load balancing. ilb owner will still be part of
2949 * While stopping the tick, this cpu will become the ilb owner if there
2950 * is no other owner. And will be the owner till that cpu becomes busy
2951 * or if all cpus in the system stop their ticks at which point
2952 * there is no need for ilb owner.
2954 * When the ilb owner becomes busy, it nominates another owner, during the
2955 * next busy scheduler_tick()
2957 int select_nohz_load_balancer(int stop_tick)
2959 int cpu = smp_processor_id();
2962 cpu_set(cpu, nohz.cpu_mask);
2963 cpu_rq(cpu)->in_nohz_recently = 1;
2966 * If we are going offline and still the leader, give up!
2968 if (cpu_is_offline(cpu) &&
2969 atomic_read(&nohz.load_balancer) == cpu) {
2970 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2975 /* time for ilb owner also to sleep */
2976 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2977 if (atomic_read(&nohz.load_balancer) == cpu)
2978 atomic_set(&nohz.load_balancer, -1);
2982 if (atomic_read(&nohz.load_balancer) == -1) {
2983 /* make me the ilb owner */
2984 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2986 } else if (atomic_read(&nohz.load_balancer) == cpu)
2989 if (!cpu_isset(cpu, nohz.cpu_mask))
2992 cpu_clear(cpu, nohz.cpu_mask);
2994 if (atomic_read(&nohz.load_balancer) == cpu)
2995 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3002 static DEFINE_SPINLOCK(balancing);
3005 * It checks each scheduling domain to see if it is due to be balanced,
3006 * and initiates a balancing operation if so.
3008 * Balancing parameters are set up in arch_init_sched_domains.
3010 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3013 struct rq *rq = cpu_rq(cpu);
3014 unsigned long interval;
3015 struct sched_domain *sd;
3016 /* Earliest time when we have to do rebalance again */
3017 unsigned long next_balance = jiffies + 60*HZ;
3018 int update_next_balance = 0;
3020 for_each_domain(cpu, sd) {
3021 if (!(sd->flags & SD_LOAD_BALANCE))
3024 interval = sd->balance_interval;
3025 if (idle != CPU_IDLE)
3026 interval *= sd->busy_factor;
3028 /* scale ms to jiffies */
3029 interval = msecs_to_jiffies(interval);
3030 if (unlikely(!interval))
3032 if (interval > HZ*NR_CPUS/10)
3033 interval = HZ*NR_CPUS/10;
3036 if (sd->flags & SD_SERIALIZE) {
3037 if (!spin_trylock(&balancing))
3041 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3042 if (load_balance(cpu, rq, sd, idle, &balance)) {
3044 * We've pulled tasks over so either we're no
3045 * longer idle, or one of our SMT siblings is
3048 idle = CPU_NOT_IDLE;
3050 sd->last_balance = jiffies;
3052 if (sd->flags & SD_SERIALIZE)
3053 spin_unlock(&balancing);
3055 if (time_after(next_balance, sd->last_balance + interval)) {
3056 next_balance = sd->last_balance + interval;
3057 update_next_balance = 1;
3061 * Stop the load balance at this level. There is another
3062 * CPU in our sched group which is doing load balancing more
3070 * next_balance will be updated only when there is a need.
3071 * When the cpu is attached to null domain for ex, it will not be
3074 if (likely(update_next_balance))
3075 rq->next_balance = next_balance;
3079 * run_rebalance_domains is triggered when needed from the scheduler tick.
3080 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3081 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3083 static void run_rebalance_domains(struct softirq_action *h)
3085 int this_cpu = smp_processor_id();
3086 struct rq *this_rq = cpu_rq(this_cpu);
3087 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3088 CPU_IDLE : CPU_NOT_IDLE;
3090 rebalance_domains(this_cpu, idle);
3094 * If this cpu is the owner for idle load balancing, then do the
3095 * balancing on behalf of the other idle cpus whose ticks are
3098 if (this_rq->idle_at_tick &&
3099 atomic_read(&nohz.load_balancer) == this_cpu) {
3100 cpumask_t cpus = nohz.cpu_mask;
3104 cpu_clear(this_cpu, cpus);
3105 for_each_cpu_mask(balance_cpu, cpus) {
3107 * If this cpu gets work to do, stop the load balancing
3108 * work being done for other cpus. Next load
3109 * balancing owner will pick it up.
3114 rebalance_domains(balance_cpu, CPU_IDLE);
3116 rq = cpu_rq(balance_cpu);
3117 if (time_after(this_rq->next_balance, rq->next_balance))
3118 this_rq->next_balance = rq->next_balance;
3125 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3127 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3128 * idle load balancing owner or decide to stop the periodic load balancing,
3129 * if the whole system is idle.
3131 static inline void trigger_load_balance(struct rq *rq, int cpu)
3135 * If we were in the nohz mode recently and busy at the current
3136 * scheduler tick, then check if we need to nominate new idle
3139 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3140 rq->in_nohz_recently = 0;
3142 if (atomic_read(&nohz.load_balancer) == cpu) {
3143 cpu_clear(cpu, nohz.cpu_mask);
3144 atomic_set(&nohz.load_balancer, -1);
3147 if (atomic_read(&nohz.load_balancer) == -1) {
3149 * simple selection for now: Nominate the
3150 * first cpu in the nohz list to be the next
3153 * TBD: Traverse the sched domains and nominate
3154 * the nearest cpu in the nohz.cpu_mask.
3156 int ilb = first_cpu(nohz.cpu_mask);
3164 * If this cpu is idle and doing idle load balancing for all the
3165 * cpus with ticks stopped, is it time for that to stop?
3167 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3168 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3174 * If this cpu is idle and the idle load balancing is done by
3175 * someone else, then no need raise the SCHED_SOFTIRQ
3177 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3178 cpu_isset(cpu, nohz.cpu_mask))
3181 if (time_after_eq(jiffies, rq->next_balance))
3182 raise_softirq(SCHED_SOFTIRQ);
3185 #else /* CONFIG_SMP */
3188 * on UP we do not need to balance between CPUs:
3190 static inline void idle_balance(int cpu, struct rq *rq)
3194 /* Avoid "used but not defined" warning on UP */
3195 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3196 unsigned long max_nr_move, unsigned long max_load_move,
3197 struct sched_domain *sd, enum cpu_idle_type idle,
3198 int *all_pinned, unsigned long *load_moved,
3199 int *this_best_prio, struct rq_iterator *iterator)
3208 DEFINE_PER_CPU(struct kernel_stat, kstat);
3210 EXPORT_PER_CPU_SYMBOL(kstat);
3213 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3214 * that have not yet been banked in case the task is currently running.
3216 unsigned long long task_sched_runtime(struct task_struct *p)
3218 unsigned long flags;
3222 rq = task_rq_lock(p, &flags);
3223 ns = p->se.sum_exec_runtime;
3224 if (rq->curr == p) {
3225 update_rq_clock(rq);
3226 delta_exec = rq->clock - p->se.exec_start;
3227 if ((s64)delta_exec > 0)
3230 task_rq_unlock(rq, &flags);
3236 * Account user cpu time to a process.
3237 * @p: the process that the cpu time gets accounted to
3238 * @hardirq_offset: the offset to subtract from hardirq_count()
3239 * @cputime: the cpu time spent in user space since the last update
3241 void account_user_time(struct task_struct *p, cputime_t cputime)
3243 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3246 p->utime = cputime_add(p->utime, cputime);
3248 /* Add user time to cpustat. */
3249 tmp = cputime_to_cputime64(cputime);
3250 if (TASK_NICE(p) > 0)
3251 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3253 cpustat->user = cputime64_add(cpustat->user, tmp);
3257 * Account system cpu time to a process.
3258 * @p: the process that the cpu time gets accounted to
3259 * @hardirq_offset: the offset to subtract from hardirq_count()
3260 * @cputime: the cpu time spent in kernel space since the last update
3262 void account_system_time(struct task_struct *p, int hardirq_offset,
3265 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3266 struct rq *rq = this_rq();
3269 p->stime = cputime_add(p->stime, cputime);
3271 /* Add system time to cpustat. */
3272 tmp = cputime_to_cputime64(cputime);
3273 if (hardirq_count() - hardirq_offset)
3274 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3275 else if (softirq_count())
3276 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3277 else if (p != rq->idle)
3278 cpustat->system = cputime64_add(cpustat->system, tmp);
3279 else if (atomic_read(&rq->nr_iowait) > 0)
3280 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3282 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3283 /* Account for system time used */
3284 acct_update_integrals(p);
3288 * Account for involuntary wait time.
3289 * @p: the process from which the cpu time has been stolen
3290 * @steal: the cpu time spent in involuntary wait
3292 void account_steal_time(struct task_struct *p, cputime_t steal)
3294 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3295 cputime64_t tmp = cputime_to_cputime64(steal);
3296 struct rq *rq = this_rq();
3298 if (p == rq->idle) {
3299 p->stime = cputime_add(p->stime, steal);
3300 if (atomic_read(&rq->nr_iowait) > 0)
3301 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3303 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3305 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3309 * This function gets called by the timer code, with HZ frequency.
3310 * We call it with interrupts disabled.
3312 * It also gets called by the fork code, when changing the parent's
3315 void scheduler_tick(void)
3317 int cpu = smp_processor_id();
3318 struct rq *rq = cpu_rq(cpu);
3319 struct task_struct *curr = rq->curr;
3320 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3322 spin_lock(&rq->lock);
3323 __update_rq_clock(rq);
3325 * Let rq->clock advance by at least TICK_NSEC:
3327 if (unlikely(rq->clock < next_tick))
3328 rq->clock = next_tick;
3329 rq->tick_timestamp = rq->clock;
3330 update_cpu_load(rq);
3331 if (curr != rq->idle) /* FIXME: needed? */
3332 curr->sched_class->task_tick(rq, curr);
3333 spin_unlock(&rq->lock);
3336 rq->idle_at_tick = idle_cpu(cpu);
3337 trigger_load_balance(rq, cpu);
3341 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3343 void fastcall add_preempt_count(int val)
3348 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3350 preempt_count() += val;
3352 * Spinlock count overflowing soon?
3354 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3357 EXPORT_SYMBOL(add_preempt_count);
3359 void fastcall sub_preempt_count(int val)
3364 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3367 * Is the spinlock portion underflowing?
3369 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3370 !(preempt_count() & PREEMPT_MASK)))
3373 preempt_count() -= val;
3375 EXPORT_SYMBOL(sub_preempt_count);
3380 * Print scheduling while atomic bug:
3382 static noinline void __schedule_bug(struct task_struct *prev)
3384 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3385 prev->comm, preempt_count(), prev->pid);
3386 debug_show_held_locks(prev);
3387 if (irqs_disabled())
3388 print_irqtrace_events(prev);
3393 * Various schedule()-time debugging checks and statistics:
3395 static inline void schedule_debug(struct task_struct *prev)
3398 * Test if we are atomic. Since do_exit() needs to call into
3399 * schedule() atomically, we ignore that path for now.
3400 * Otherwise, whine if we are scheduling when we should not be.
3402 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3403 __schedule_bug(prev);
3405 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3407 schedstat_inc(this_rq(), sched_cnt);
3411 * Pick up the highest-prio task:
3413 static inline struct task_struct *
3414 pick_next_task(struct rq *rq, struct task_struct *prev)
3416 struct sched_class *class;
3417 struct task_struct *p;
3420 * Optimization: we know that if all tasks are in
3421 * the fair class we can call that function directly:
3423 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3424 p = fair_sched_class.pick_next_task(rq);
3429 class = sched_class_highest;
3431 p = class->pick_next_task(rq);
3435 * Will never be NULL as the idle class always
3436 * returns a non-NULL p:
3438 class = class->next;
3443 * schedule() is the main scheduler function.
3445 asmlinkage void __sched schedule(void)
3447 struct task_struct *prev, *next;
3454 cpu = smp_processor_id();
3458 switch_count = &prev->nivcsw;
3460 release_kernel_lock(prev);
3461 need_resched_nonpreemptible:
3463 schedule_debug(prev);
3465 spin_lock_irq(&rq->lock);
3466 clear_tsk_need_resched(prev);
3467 __update_rq_clock(rq);
3469 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3470 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3471 unlikely(signal_pending(prev)))) {
3472 prev->state = TASK_RUNNING;
3474 deactivate_task(rq, prev, 1);
3476 switch_count = &prev->nvcsw;
3479 if (unlikely(!rq->nr_running))
3480 idle_balance(cpu, rq);
3482 prev->sched_class->put_prev_task(rq, prev);
3483 next = pick_next_task(rq, prev);
3485 sched_info_switch(prev, next);
3487 if (likely(prev != next)) {
3492 context_switch(rq, prev, next); /* unlocks the rq */
3494 spin_unlock_irq(&rq->lock);
3496 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3497 cpu = smp_processor_id();
3499 goto need_resched_nonpreemptible;
3501 preempt_enable_no_resched();
3502 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3505 EXPORT_SYMBOL(schedule);
3507 #ifdef CONFIG_PREEMPT
3509 * this is the entry point to schedule() from in-kernel preemption
3510 * off of preempt_enable. Kernel preemptions off return from interrupt
3511 * occur there and call schedule directly.
3513 asmlinkage void __sched preempt_schedule(void)
3515 struct thread_info *ti = current_thread_info();
3516 #ifdef CONFIG_PREEMPT_BKL
3517 struct task_struct *task = current;
3518 int saved_lock_depth;
3521 * If there is a non-zero preempt_count or interrupts are disabled,
3522 * we do not want to preempt the current task. Just return..
3524 if (likely(ti->preempt_count || irqs_disabled()))
3528 add_preempt_count(PREEMPT_ACTIVE);
3530 * We keep the big kernel semaphore locked, but we
3531 * clear ->lock_depth so that schedule() doesnt
3532 * auto-release the semaphore:
3534 #ifdef CONFIG_PREEMPT_BKL
3535 saved_lock_depth = task->lock_depth;
3536 task->lock_depth = -1;
3539 #ifdef CONFIG_PREEMPT_BKL
3540 task->lock_depth = saved_lock_depth;
3542 sub_preempt_count(PREEMPT_ACTIVE);
3544 /* we could miss a preemption opportunity between schedule and now */
3546 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3549 EXPORT_SYMBOL(preempt_schedule);
3552 * this is the entry point to schedule() from kernel preemption
3553 * off of irq context.
3554 * Note, that this is called and return with irqs disabled. This will
3555 * protect us against recursive calling from irq.
3557 asmlinkage void __sched preempt_schedule_irq(void)
3559 struct thread_info *ti = current_thread_info();
3560 #ifdef CONFIG_PREEMPT_BKL
3561 struct task_struct *task = current;
3562 int saved_lock_depth;
3564 /* Catch callers which need to be fixed */
3565 BUG_ON(ti->preempt_count || !irqs_disabled());
3568 add_preempt_count(PREEMPT_ACTIVE);
3570 * We keep the big kernel semaphore locked, but we
3571 * clear ->lock_depth so that schedule() doesnt
3572 * auto-release the semaphore:
3574 #ifdef CONFIG_PREEMPT_BKL
3575 saved_lock_depth = task->lock_depth;
3576 task->lock_depth = -1;
3580 local_irq_disable();
3581 #ifdef CONFIG_PREEMPT_BKL
3582 task->lock_depth = saved_lock_depth;
3584 sub_preempt_count(PREEMPT_ACTIVE);
3586 /* we could miss a preemption opportunity between schedule and now */
3588 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3592 #endif /* CONFIG_PREEMPT */
3594 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3597 return try_to_wake_up(curr->private, mode, sync);
3599 EXPORT_SYMBOL(default_wake_function);
3602 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3603 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3604 * number) then we wake all the non-exclusive tasks and one exclusive task.
3606 * There are circumstances in which we can try to wake a task which has already
3607 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3608 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3610 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3611 int nr_exclusive, int sync, void *key)
3613 wait_queue_t *curr, *next;
3615 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3616 unsigned flags = curr->flags;
3618 if (curr->func(curr, mode, sync, key) &&
3619 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3625 * __wake_up - wake up threads blocked on a waitqueue.
3627 * @mode: which threads
3628 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3629 * @key: is directly passed to the wakeup function
3631 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3632 int nr_exclusive, void *key)
3634 unsigned long flags;
3636 spin_lock_irqsave(&q->lock, flags);
3637 __wake_up_common(q, mode, nr_exclusive, 0, key);
3638 spin_unlock_irqrestore(&q->lock, flags);
3640 EXPORT_SYMBOL(__wake_up);
3643 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3645 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3647 __wake_up_common(q, mode, 1, 0, NULL);
3651 * __wake_up_sync - wake up threads blocked on a waitqueue.
3653 * @mode: which threads
3654 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3656 * The sync wakeup differs that the waker knows that it will schedule
3657 * away soon, so while the target thread will be woken up, it will not
3658 * be migrated to another CPU - ie. the two threads are 'synchronized'
3659 * with each other. This can prevent needless bouncing between CPUs.
3661 * On UP it can prevent extra preemption.
3664 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3666 unsigned long flags;
3672 if (unlikely(!nr_exclusive))
3675 spin_lock_irqsave(&q->lock, flags);
3676 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3677 spin_unlock_irqrestore(&q->lock, flags);
3679 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3681 void fastcall complete(struct completion *x)
3683 unsigned long flags;
3685 spin_lock_irqsave(&x->wait.lock, flags);
3687 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3689 spin_unlock_irqrestore(&x->wait.lock, flags);
3691 EXPORT_SYMBOL(complete);
3693 void fastcall complete_all(struct completion *x)
3695 unsigned long flags;
3697 spin_lock_irqsave(&x->wait.lock, flags);
3698 x->done += UINT_MAX/2;
3699 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3701 spin_unlock_irqrestore(&x->wait.lock, flags);
3703 EXPORT_SYMBOL(complete_all);
3705 void fastcall __sched wait_for_completion(struct completion *x)
3709 spin_lock_irq(&x->wait.lock);
3711 DECLARE_WAITQUEUE(wait, current);
3713 wait.flags |= WQ_FLAG_EXCLUSIVE;
3714 __add_wait_queue_tail(&x->wait, &wait);
3716 __set_current_state(TASK_UNINTERRUPTIBLE);
3717 spin_unlock_irq(&x->wait.lock);
3719 spin_lock_irq(&x->wait.lock);
3721 __remove_wait_queue(&x->wait, &wait);
3724 spin_unlock_irq(&x->wait.lock);
3726 EXPORT_SYMBOL(wait_for_completion);
3728 unsigned long fastcall __sched
3729 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3733 spin_lock_irq(&x->wait.lock);
3735 DECLARE_WAITQUEUE(wait, current);
3737 wait.flags |= WQ_FLAG_EXCLUSIVE;
3738 __add_wait_queue_tail(&x->wait, &wait);
3740 __set_current_state(TASK_UNINTERRUPTIBLE);
3741 spin_unlock_irq(&x->wait.lock);
3742 timeout = schedule_timeout(timeout);
3743 spin_lock_irq(&x->wait.lock);
3745 __remove_wait_queue(&x->wait, &wait);
3749 __remove_wait_queue(&x->wait, &wait);
3753 spin_unlock_irq(&x->wait.lock);
3756 EXPORT_SYMBOL(wait_for_completion_timeout);
3758 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3764 spin_lock_irq(&x->wait.lock);
3766 DECLARE_WAITQUEUE(wait, current);
3768 wait.flags |= WQ_FLAG_EXCLUSIVE;
3769 __add_wait_queue_tail(&x->wait, &wait);
3771 if (signal_pending(current)) {
3773 __remove_wait_queue(&x->wait, &wait);
3776 __set_current_state(TASK_INTERRUPTIBLE);
3777 spin_unlock_irq(&x->wait.lock);
3779 spin_lock_irq(&x->wait.lock);
3781 __remove_wait_queue(&x->wait, &wait);
3785 spin_unlock_irq(&x->wait.lock);
3789 EXPORT_SYMBOL(wait_for_completion_interruptible);
3791 unsigned long fastcall __sched
3792 wait_for_completion_interruptible_timeout(struct completion *x,
3793 unsigned long timeout)
3797 spin_lock_irq(&x->wait.lock);
3799 DECLARE_WAITQUEUE(wait, current);
3801 wait.flags |= WQ_FLAG_EXCLUSIVE;
3802 __add_wait_queue_tail(&x->wait, &wait);
3804 if (signal_pending(current)) {
3805 timeout = -ERESTARTSYS;
3806 __remove_wait_queue(&x->wait, &wait);
3809 __set_current_state(TASK_INTERRUPTIBLE);
3810 spin_unlock_irq(&x->wait.lock);
3811 timeout = schedule_timeout(timeout);
3812 spin_lock_irq(&x->wait.lock);
3814 __remove_wait_queue(&x->wait, &wait);
3818 __remove_wait_queue(&x->wait, &wait);
3822 spin_unlock_irq(&x->wait.lock);
3825 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3828 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3830 spin_lock_irqsave(&q->lock, *flags);
3831 __add_wait_queue(q, wait);
3832 spin_unlock(&q->lock);
3836 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3838 spin_lock_irq(&q->lock);
3839 __remove_wait_queue(q, wait);
3840 spin_unlock_irqrestore(&q->lock, *flags);
3843 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3845 unsigned long flags;
3848 init_waitqueue_entry(&wait, current);
3850 current->state = TASK_INTERRUPTIBLE;
3852 sleep_on_head(q, &wait, &flags);
3854 sleep_on_tail(q, &wait, &flags);
3856 EXPORT_SYMBOL(interruptible_sleep_on);
3859 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3861 unsigned long flags;
3864 init_waitqueue_entry(&wait, current);
3866 current->state = TASK_INTERRUPTIBLE;
3868 sleep_on_head(q, &wait, &flags);
3869 timeout = schedule_timeout(timeout);
3870 sleep_on_tail(q, &wait, &flags);
3874 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3876 void __sched sleep_on(wait_queue_head_t *q)
3878 unsigned long flags;
3881 init_waitqueue_entry(&wait, current);
3883 current->state = TASK_UNINTERRUPTIBLE;
3885 sleep_on_head(q, &wait, &flags);
3887 sleep_on_tail(q, &wait, &flags);
3889 EXPORT_SYMBOL(sleep_on);
3891 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3893 unsigned long flags;
3896 init_waitqueue_entry(&wait, current);
3898 current->state = TASK_UNINTERRUPTIBLE;
3900 sleep_on_head(q, &wait, &flags);
3901 timeout = schedule_timeout(timeout);
3902 sleep_on_tail(q, &wait, &flags);
3906 EXPORT_SYMBOL(sleep_on_timeout);
3908 #ifdef CONFIG_RT_MUTEXES
3911 * rt_mutex_setprio - set the current priority of a task
3913 * @prio: prio value (kernel-internal form)
3915 * This function changes the 'effective' priority of a task. It does
3916 * not touch ->normal_prio like __setscheduler().
3918 * Used by the rt_mutex code to implement priority inheritance logic.
3920 void rt_mutex_setprio(struct task_struct *p, int prio)
3922 unsigned long flags;
3923 int oldprio, on_rq, running;
3926 BUG_ON(prio < 0 || prio > MAX_PRIO);
3928 rq = task_rq_lock(p, &flags);
3929 update_rq_clock(rq);
3932 on_rq = p->se.on_rq;
3933 running = task_running(rq, p);
3935 dequeue_task(rq, p, 0);
3937 p->sched_class->put_prev_task(rq, p);
3941 p->sched_class = &rt_sched_class;
3943 p->sched_class = &fair_sched_class;
3949 p->sched_class->set_curr_task(rq);
3950 enqueue_task(rq, p, 0);
3952 * Reschedule if we are currently running on this runqueue and
3953 * our priority decreased, or if we are not currently running on
3954 * this runqueue and our priority is higher than the current's
3957 if (p->prio > oldprio)
3958 resched_task(rq->curr);
3960 check_preempt_curr(rq, p);
3963 task_rq_unlock(rq, &flags);
3968 void set_user_nice(struct task_struct *p, long nice)
3970 int old_prio, delta, on_rq;
3971 unsigned long flags;
3974 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3977 * We have to be careful, if called from sys_setpriority(),
3978 * the task might be in the middle of scheduling on another CPU.
3980 rq = task_rq_lock(p, &flags);
3981 update_rq_clock(rq);
3983 * The RT priorities are set via sched_setscheduler(), but we still
3984 * allow the 'normal' nice value to be set - but as expected
3985 * it wont have any effect on scheduling until the task is
3986 * SCHED_FIFO/SCHED_RR:
3988 if (task_has_rt_policy(p)) {
3989 p->static_prio = NICE_TO_PRIO(nice);
3992 on_rq = p->se.on_rq;
3994 dequeue_task(rq, p, 0);
3998 p->static_prio = NICE_TO_PRIO(nice);
4001 p->prio = effective_prio(p);
4002 delta = p->prio - old_prio;
4005 enqueue_task(rq, p, 0);
4008 * If the task increased its priority or is running and
4009 * lowered its priority, then reschedule its CPU:
4011 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4012 resched_task(rq->curr);
4015 task_rq_unlock(rq, &flags);
4017 EXPORT_SYMBOL(set_user_nice);
4020 * can_nice - check if a task can reduce its nice value
4024 int can_nice(const struct task_struct *p, const int nice)
4026 /* convert nice value [19,-20] to rlimit style value [1,40] */
4027 int nice_rlim = 20 - nice;
4029 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4030 capable(CAP_SYS_NICE));
4033 #ifdef __ARCH_WANT_SYS_NICE
4036 * sys_nice - change the priority of the current process.
4037 * @increment: priority increment
4039 * sys_setpriority is a more generic, but much slower function that
4040 * does similar things.
4042 asmlinkage long sys_nice(int increment)
4047 * Setpriority might change our priority at the same moment.
4048 * We don't have to worry. Conceptually one call occurs first
4049 * and we have a single winner.
4051 if (increment < -40)
4056 nice = PRIO_TO_NICE(current->static_prio) + increment;
4062 if (increment < 0 && !can_nice(current, nice))
4065 retval = security_task_setnice(current, nice);
4069 set_user_nice(current, nice);
4076 * task_prio - return the priority value of a given task.
4077 * @p: the task in question.
4079 * This is the priority value as seen by users in /proc.
4080 * RT tasks are offset by -200. Normal tasks are centered
4081 * around 0, value goes from -16 to +15.
4083 int task_prio(const struct task_struct *p)
4085 return p->prio - MAX_RT_PRIO;
4089 * task_nice - return the nice value of a given task.
4090 * @p: the task in question.
4092 int task_nice(const struct task_struct *p)
4094 return TASK_NICE(p);
4096 EXPORT_SYMBOL_GPL(task_nice);
4099 * idle_cpu - is a given cpu idle currently?
4100 * @cpu: the processor in question.
4102 int idle_cpu(int cpu)
4104 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4108 * idle_task - return the idle task for a given cpu.
4109 * @cpu: the processor in question.
4111 struct task_struct *idle_task(int cpu)
4113 return cpu_rq(cpu)->idle;
4117 * find_process_by_pid - find a process with a matching PID value.
4118 * @pid: the pid in question.
4120 static inline struct task_struct *find_process_by_pid(pid_t pid)
4122 return pid ? find_task_by_pid(pid) : current;
4125 /* Actually do priority change: must hold rq lock. */
4127 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4129 BUG_ON(p->se.on_rq);
4132 switch (p->policy) {
4136 p->sched_class = &fair_sched_class;
4140 p->sched_class = &rt_sched_class;
4144 p->rt_priority = prio;
4145 p->normal_prio = normal_prio(p);
4146 /* we are holding p->pi_lock already */
4147 p->prio = rt_mutex_getprio(p);
4152 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4153 * @p: the task in question.
4154 * @policy: new policy.
4155 * @param: structure containing the new RT priority.
4157 * NOTE that the task may be already dead.
4159 int sched_setscheduler(struct task_struct *p, int policy,
4160 struct sched_param *param)
4162 int retval, oldprio, oldpolicy = -1, on_rq, running;
4163 unsigned long flags;
4166 /* may grab non-irq protected spin_locks */
4167 BUG_ON(in_interrupt());
4169 /* double check policy once rq lock held */
4171 policy = oldpolicy = p->policy;
4172 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4173 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4174 policy != SCHED_IDLE)
4177 * Valid priorities for SCHED_FIFO and SCHED_RR are
4178 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4179 * SCHED_BATCH and SCHED_IDLE is 0.
4181 if (param->sched_priority < 0 ||
4182 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4183 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4185 if (rt_policy(policy) != (param->sched_priority != 0))
4189 * Allow unprivileged RT tasks to decrease priority:
4191 if (!capable(CAP_SYS_NICE)) {
4192 if (rt_policy(policy)) {
4193 unsigned long rlim_rtprio;
4195 if (!lock_task_sighand(p, &flags))
4197 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4198 unlock_task_sighand(p, &flags);
4200 /* can't set/change the rt policy */
4201 if (policy != p->policy && !rlim_rtprio)
4204 /* can't increase priority */
4205 if (param->sched_priority > p->rt_priority &&
4206 param->sched_priority > rlim_rtprio)
4210 * Like positive nice levels, dont allow tasks to
4211 * move out of SCHED_IDLE either:
4213 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4216 /* can't change other user's priorities */
4217 if ((current->euid != p->euid) &&
4218 (current->euid != p->uid))
4222 retval = security_task_setscheduler(p, policy, param);
4226 * make sure no PI-waiters arrive (or leave) while we are
4227 * changing the priority of the task:
4229 spin_lock_irqsave(&p->pi_lock, flags);
4231 * To be able to change p->policy safely, the apropriate
4232 * runqueue lock must be held.
4234 rq = __task_rq_lock(p);
4235 /* recheck policy now with rq lock held */
4236 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4237 policy = oldpolicy = -1;
4238 __task_rq_unlock(rq);
4239 spin_unlock_irqrestore(&p->pi_lock, flags);
4242 update_rq_clock(rq);
4243 on_rq = p->se.on_rq;
4244 running = task_running(rq, p);
4246 deactivate_task(rq, p, 0);
4248 p->sched_class->put_prev_task(rq, p);
4252 __setscheduler(rq, p, policy, param->sched_priority);
4256 p->sched_class->set_curr_task(rq);
4257 activate_task(rq, p, 0);
4259 * Reschedule if we are currently running on this runqueue and
4260 * our priority decreased, or if we are not currently running on
4261 * this runqueue and our priority is higher than the current's
4264 if (p->prio > oldprio)
4265 resched_task(rq->curr);
4267 check_preempt_curr(rq, p);
4270 __task_rq_unlock(rq);
4271 spin_unlock_irqrestore(&p->pi_lock, flags);
4273 rt_mutex_adjust_pi(p);
4277 EXPORT_SYMBOL_GPL(sched_setscheduler);
4280 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4282 struct sched_param lparam;
4283 struct task_struct *p;
4286 if (!param || pid < 0)
4288 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4293 p = find_process_by_pid(pid);
4295 retval = sched_setscheduler(p, policy, &lparam);
4302 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4303 * @pid: the pid in question.
4304 * @policy: new policy.
4305 * @param: structure containing the new RT priority.
4307 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4308 struct sched_param __user *param)
4310 /* negative values for policy are not valid */
4314 return do_sched_setscheduler(pid, policy, param);
4318 * sys_sched_setparam - set/change the RT priority of a thread
4319 * @pid: the pid in question.
4320 * @param: structure containing the new RT priority.
4322 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4324 return do_sched_setscheduler(pid, -1, param);
4328 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4329 * @pid: the pid in question.
4331 asmlinkage long sys_sched_getscheduler(pid_t pid)
4333 struct task_struct *p;
4334 int retval = -EINVAL;
4340 read_lock(&tasklist_lock);
4341 p = find_process_by_pid(pid);
4343 retval = security_task_getscheduler(p);
4347 read_unlock(&tasklist_lock);
4354 * sys_sched_getscheduler - get the RT priority of a thread
4355 * @pid: the pid in question.
4356 * @param: structure containing the RT priority.
4358 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4360 struct sched_param lp;
4361 struct task_struct *p;
4362 int retval = -EINVAL;
4364 if (!param || pid < 0)
4367 read_lock(&tasklist_lock);
4368 p = find_process_by_pid(pid);
4373 retval = security_task_getscheduler(p);
4377 lp.sched_priority = p->rt_priority;
4378 read_unlock(&tasklist_lock);
4381 * This one might sleep, we cannot do it with a spinlock held ...
4383 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4389 read_unlock(&tasklist_lock);
4393 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4395 cpumask_t cpus_allowed;
4396 struct task_struct *p;
4399 mutex_lock(&sched_hotcpu_mutex);
4400 read_lock(&tasklist_lock);
4402 p = find_process_by_pid(pid);
4404 read_unlock(&tasklist_lock);
4405 mutex_unlock(&sched_hotcpu_mutex);
4410 * It is not safe to call set_cpus_allowed with the
4411 * tasklist_lock held. We will bump the task_struct's
4412 * usage count and then drop tasklist_lock.
4415 read_unlock(&tasklist_lock);
4418 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4419 !capable(CAP_SYS_NICE))
4422 retval = security_task_setscheduler(p, 0, NULL);
4426 cpus_allowed = cpuset_cpus_allowed(p);
4427 cpus_and(new_mask, new_mask, cpus_allowed);
4428 retval = set_cpus_allowed(p, new_mask);
4432 mutex_unlock(&sched_hotcpu_mutex);
4436 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4437 cpumask_t *new_mask)
4439 if (len < sizeof(cpumask_t)) {
4440 memset(new_mask, 0, sizeof(cpumask_t));
4441 } else if (len > sizeof(cpumask_t)) {
4442 len = sizeof(cpumask_t);
4444 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4448 * sys_sched_setaffinity - set the cpu affinity of a process
4449 * @pid: pid of the process
4450 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4451 * @user_mask_ptr: user-space pointer to the new cpu mask
4453 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4454 unsigned long __user *user_mask_ptr)
4459 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4463 return sched_setaffinity(pid, new_mask);
4467 * Represents all cpu's present in the system
4468 * In systems capable of hotplug, this map could dynamically grow
4469 * as new cpu's are detected in the system via any platform specific
4470 * method, such as ACPI for e.g.
4473 cpumask_t cpu_present_map __read_mostly;
4474 EXPORT_SYMBOL(cpu_present_map);
4477 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4478 EXPORT_SYMBOL(cpu_online_map);
4480 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4481 EXPORT_SYMBOL(cpu_possible_map);
4484 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4486 struct task_struct *p;
4489 mutex_lock(&sched_hotcpu_mutex);
4490 read_lock(&tasklist_lock);
4493 p = find_process_by_pid(pid);
4497 retval = security_task_getscheduler(p);
4501 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4504 read_unlock(&tasklist_lock);
4505 mutex_unlock(&sched_hotcpu_mutex);
4511 * sys_sched_getaffinity - get the cpu affinity of a process
4512 * @pid: pid of the process
4513 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4514 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4516 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4517 unsigned long __user *user_mask_ptr)
4522 if (len < sizeof(cpumask_t))
4525 ret = sched_getaffinity(pid, &mask);
4529 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4532 return sizeof(cpumask_t);
4536 * sys_sched_yield - yield the current processor to other threads.
4538 * This function yields the current CPU to other tasks. If there are no
4539 * other threads running on this CPU then this function will return.
4541 asmlinkage long sys_sched_yield(void)
4543 struct rq *rq = this_rq_lock();
4545 schedstat_inc(rq, yld_cnt);
4546 current->sched_class->yield_task(rq);
4549 * Since we are going to call schedule() anyway, there's
4550 * no need to preempt or enable interrupts:
4552 __release(rq->lock);
4553 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4554 _raw_spin_unlock(&rq->lock);
4555 preempt_enable_no_resched();
4562 static void __cond_resched(void)
4564 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4565 __might_sleep(__FILE__, __LINE__);
4568 * The BKS might be reacquired before we have dropped
4569 * PREEMPT_ACTIVE, which could trigger a second
4570 * cond_resched() call.
4573 add_preempt_count(PREEMPT_ACTIVE);
4575 sub_preempt_count(PREEMPT_ACTIVE);
4576 } while (need_resched());
4579 int __sched cond_resched(void)
4581 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4582 system_state == SYSTEM_RUNNING) {
4588 EXPORT_SYMBOL(cond_resched);
4591 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4592 * call schedule, and on return reacquire the lock.
4594 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4595 * operations here to prevent schedule() from being called twice (once via
4596 * spin_unlock(), once by hand).
4598 int cond_resched_lock(spinlock_t *lock)
4602 if (need_lockbreak(lock)) {
4608 if (need_resched() && system_state == SYSTEM_RUNNING) {
4609 spin_release(&lock->dep_map, 1, _THIS_IP_);
4610 _raw_spin_unlock(lock);
4611 preempt_enable_no_resched();
4618 EXPORT_SYMBOL(cond_resched_lock);
4620 int __sched cond_resched_softirq(void)
4622 BUG_ON(!in_softirq());
4624 if (need_resched() && system_state == SYSTEM_RUNNING) {
4632 EXPORT_SYMBOL(cond_resched_softirq);
4635 * yield - yield the current processor to other threads.
4637 * This is a shortcut for kernel-space yielding - it marks the
4638 * thread runnable and calls sys_sched_yield().
4640 void __sched yield(void)
4642 set_current_state(TASK_RUNNING);
4645 EXPORT_SYMBOL(yield);
4648 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4649 * that process accounting knows that this is a task in IO wait state.
4651 * But don't do that if it is a deliberate, throttling IO wait (this task
4652 * has set its backing_dev_info: the queue against which it should throttle)
4654 void __sched io_schedule(void)
4656 struct rq *rq = &__raw_get_cpu_var(runqueues);
4658 delayacct_blkio_start();
4659 atomic_inc(&rq->nr_iowait);
4661 atomic_dec(&rq->nr_iowait);
4662 delayacct_blkio_end();
4664 EXPORT_SYMBOL(io_schedule);
4666 long __sched io_schedule_timeout(long timeout)
4668 struct rq *rq = &__raw_get_cpu_var(runqueues);
4671 delayacct_blkio_start();
4672 atomic_inc(&rq->nr_iowait);
4673 ret = schedule_timeout(timeout);
4674 atomic_dec(&rq->nr_iowait);
4675 delayacct_blkio_end();
4680 * sys_sched_get_priority_max - return maximum RT priority.
4681 * @policy: scheduling class.
4683 * this syscall returns the maximum rt_priority that can be used
4684 * by a given scheduling class.
4686 asmlinkage long sys_sched_get_priority_max(int policy)
4693 ret = MAX_USER_RT_PRIO-1;
4705 * sys_sched_get_priority_min - return minimum RT priority.
4706 * @policy: scheduling class.
4708 * this syscall returns the minimum rt_priority that can be used
4709 * by a given scheduling class.
4711 asmlinkage long sys_sched_get_priority_min(int policy)
4729 * sys_sched_rr_get_interval - return the default timeslice of a process.
4730 * @pid: pid of the process.
4731 * @interval: userspace pointer to the timeslice value.
4733 * this syscall writes the default timeslice value of a given process
4734 * into the user-space timespec buffer. A value of '0' means infinity.
4737 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4739 struct task_struct *p;
4740 int retval = -EINVAL;
4747 read_lock(&tasklist_lock);
4748 p = find_process_by_pid(pid);
4752 retval = security_task_getscheduler(p);
4756 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4757 0 : static_prio_timeslice(p->static_prio), &t);
4758 read_unlock(&tasklist_lock);
4759 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4763 read_unlock(&tasklist_lock);
4767 static const char stat_nam[] = "RSDTtZX";
4769 static void show_task(struct task_struct *p)
4771 unsigned long free = 0;
4774 state = p->state ? __ffs(p->state) + 1 : 0;
4775 printk("%-13.13s %c", p->comm,
4776 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4777 #if BITS_PER_LONG == 32
4778 if (state == TASK_RUNNING)
4779 printk(" running ");
4781 printk(" %08lx ", thread_saved_pc(p));
4783 if (state == TASK_RUNNING)
4784 printk(" running task ");
4786 printk(" %016lx ", thread_saved_pc(p));
4788 #ifdef CONFIG_DEBUG_STACK_USAGE
4790 unsigned long *n = end_of_stack(p);
4793 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4796 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4798 if (state != TASK_RUNNING)
4799 show_stack(p, NULL);
4802 void show_state_filter(unsigned long state_filter)
4804 struct task_struct *g, *p;
4806 #if BITS_PER_LONG == 32
4808 " task PC stack pid father\n");
4811 " task PC stack pid father\n");
4813 read_lock(&tasklist_lock);
4814 do_each_thread(g, p) {
4816 * reset the NMI-timeout, listing all files on a slow
4817 * console might take alot of time:
4819 touch_nmi_watchdog();
4820 if (!state_filter || (p->state & state_filter))
4822 } while_each_thread(g, p);
4824 touch_all_softlockup_watchdogs();
4826 #ifdef CONFIG_SCHED_DEBUG
4827 sysrq_sched_debug_show();
4829 read_unlock(&tasklist_lock);
4831 * Only show locks if all tasks are dumped:
4833 if (state_filter == -1)
4834 debug_show_all_locks();
4837 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4839 idle->sched_class = &idle_sched_class;
4843 * init_idle - set up an idle thread for a given CPU
4844 * @idle: task in question
4845 * @cpu: cpu the idle task belongs to
4847 * NOTE: this function does not set the idle thread's NEED_RESCHED
4848 * flag, to make booting more robust.
4850 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4852 struct rq *rq = cpu_rq(cpu);
4853 unsigned long flags;
4856 idle->se.exec_start = sched_clock();
4858 idle->prio = idle->normal_prio = MAX_PRIO;
4859 idle->cpus_allowed = cpumask_of_cpu(cpu);
4860 __set_task_cpu(idle, cpu);
4862 spin_lock_irqsave(&rq->lock, flags);
4863 rq->curr = rq->idle = idle;
4864 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4867 spin_unlock_irqrestore(&rq->lock, flags);
4869 /* Set the preempt count _outside_ the spinlocks! */
4870 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4871 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4873 task_thread_info(idle)->preempt_count = 0;
4876 * The idle tasks have their own, simple scheduling class:
4878 idle->sched_class = &idle_sched_class;
4882 * In a system that switches off the HZ timer nohz_cpu_mask
4883 * indicates which cpus entered this state. This is used
4884 * in the rcu update to wait only for active cpus. For system
4885 * which do not switch off the HZ timer nohz_cpu_mask should
4886 * always be CPU_MASK_NONE.
4888 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4892 * This is how migration works:
4894 * 1) we queue a struct migration_req structure in the source CPU's
4895 * runqueue and wake up that CPU's migration thread.
4896 * 2) we down() the locked semaphore => thread blocks.
4897 * 3) migration thread wakes up (implicitly it forces the migrated
4898 * thread off the CPU)
4899 * 4) it gets the migration request and checks whether the migrated
4900 * task is still in the wrong runqueue.
4901 * 5) if it's in the wrong runqueue then the migration thread removes
4902 * it and puts it into the right queue.
4903 * 6) migration thread up()s the semaphore.
4904 * 7) we wake up and the migration is done.
4908 * Change a given task's CPU affinity. Migrate the thread to a
4909 * proper CPU and schedule it away if the CPU it's executing on
4910 * is removed from the allowed bitmask.
4912 * NOTE: the caller must have a valid reference to the task, the
4913 * task must not exit() & deallocate itself prematurely. The
4914 * call is not atomic; no spinlocks may be held.
4916 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4918 struct migration_req req;
4919 unsigned long flags;
4923 rq = task_rq_lock(p, &flags);
4924 if (!cpus_intersects(new_mask, cpu_online_map)) {
4929 p->cpus_allowed = new_mask;
4930 /* Can the task run on the task's current CPU? If so, we're done */
4931 if (cpu_isset(task_cpu(p), new_mask))
4934 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4935 /* Need help from migration thread: drop lock and wait. */
4936 task_rq_unlock(rq, &flags);
4937 wake_up_process(rq->migration_thread);
4938 wait_for_completion(&req.done);
4939 tlb_migrate_finish(p->mm);
4943 task_rq_unlock(rq, &flags);
4947 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4950 * Move (not current) task off this cpu, onto dest cpu. We're doing
4951 * this because either it can't run here any more (set_cpus_allowed()
4952 * away from this CPU, or CPU going down), or because we're
4953 * attempting to rebalance this task on exec (sched_exec).
4955 * So we race with normal scheduler movements, but that's OK, as long
4956 * as the task is no longer on this CPU.
4958 * Returns non-zero if task was successfully migrated.
4960 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4962 struct rq *rq_dest, *rq_src;
4965 if (unlikely(cpu_is_offline(dest_cpu)))
4968 rq_src = cpu_rq(src_cpu);
4969 rq_dest = cpu_rq(dest_cpu);
4971 double_rq_lock(rq_src, rq_dest);
4972 /* Already moved. */
4973 if (task_cpu(p) != src_cpu)
4975 /* Affinity changed (again). */
4976 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4979 on_rq = p->se.on_rq;
4981 deactivate_task(rq_src, p, 0);
4983 set_task_cpu(p, dest_cpu);
4985 activate_task(rq_dest, p, 0);
4986 check_preempt_curr(rq_dest, p);
4990 double_rq_unlock(rq_src, rq_dest);
4995 * migration_thread - this is a highprio system thread that performs
4996 * thread migration by bumping thread off CPU then 'pushing' onto
4999 static int migration_thread(void *data)
5001 int cpu = (long)data;
5005 BUG_ON(rq->migration_thread != current);
5007 set_current_state(TASK_INTERRUPTIBLE);
5008 while (!kthread_should_stop()) {
5009 struct migration_req *req;
5010 struct list_head *head;
5012 spin_lock_irq(&rq->lock);
5014 if (cpu_is_offline(cpu)) {
5015 spin_unlock_irq(&rq->lock);
5019 if (rq->active_balance) {
5020 active_load_balance(rq, cpu);
5021 rq->active_balance = 0;
5024 head = &rq->migration_queue;
5026 if (list_empty(head)) {
5027 spin_unlock_irq(&rq->lock);
5029 set_current_state(TASK_INTERRUPTIBLE);
5032 req = list_entry(head->next, struct migration_req, list);
5033 list_del_init(head->next);
5035 spin_unlock(&rq->lock);
5036 __migrate_task(req->task, cpu, req->dest_cpu);
5039 complete(&req->done);
5041 __set_current_state(TASK_RUNNING);
5045 /* Wait for kthread_stop */
5046 set_current_state(TASK_INTERRUPTIBLE);
5047 while (!kthread_should_stop()) {
5049 set_current_state(TASK_INTERRUPTIBLE);
5051 __set_current_state(TASK_RUNNING);
5055 #ifdef CONFIG_HOTPLUG_CPU
5057 * Figure out where task on dead CPU should go, use force if neccessary.
5058 * NOTE: interrupts should be disabled by the caller
5060 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5062 unsigned long flags;
5069 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5070 cpus_and(mask, mask, p->cpus_allowed);
5071 dest_cpu = any_online_cpu(mask);
5073 /* On any allowed CPU? */
5074 if (dest_cpu == NR_CPUS)
5075 dest_cpu = any_online_cpu(p->cpus_allowed);
5077 /* No more Mr. Nice Guy. */
5078 if (dest_cpu == NR_CPUS) {
5079 rq = task_rq_lock(p, &flags);
5080 cpus_setall(p->cpus_allowed);
5081 dest_cpu = any_online_cpu(p->cpus_allowed);
5082 task_rq_unlock(rq, &flags);
5085 * Don't tell them about moving exiting tasks or
5086 * kernel threads (both mm NULL), since they never
5089 if (p->mm && printk_ratelimit())
5090 printk(KERN_INFO "process %d (%s) no "
5091 "longer affine to cpu%d\n",
5092 p->pid, p->comm, dead_cpu);
5094 if (!__migrate_task(p, dead_cpu, dest_cpu))
5099 * While a dead CPU has no uninterruptible tasks queued at this point,
5100 * it might still have a nonzero ->nr_uninterruptible counter, because
5101 * for performance reasons the counter is not stricly tracking tasks to
5102 * their home CPUs. So we just add the counter to another CPU's counter,
5103 * to keep the global sum constant after CPU-down:
5105 static void migrate_nr_uninterruptible(struct rq *rq_src)
5107 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5108 unsigned long flags;
5110 local_irq_save(flags);
5111 double_rq_lock(rq_src, rq_dest);
5112 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5113 rq_src->nr_uninterruptible = 0;
5114 double_rq_unlock(rq_src, rq_dest);
5115 local_irq_restore(flags);
5118 /* Run through task list and migrate tasks from the dead cpu. */
5119 static void migrate_live_tasks(int src_cpu)
5121 struct task_struct *p, *t;
5123 write_lock_irq(&tasklist_lock);
5125 do_each_thread(t, p) {
5129 if (task_cpu(p) == src_cpu)
5130 move_task_off_dead_cpu(src_cpu, p);
5131 } while_each_thread(t, p);
5133 write_unlock_irq(&tasklist_lock);
5137 * Schedules idle task to be the next runnable task on current CPU.
5138 * It does so by boosting its priority to highest possible and adding it to
5139 * the _front_ of the runqueue. Used by CPU offline code.
5141 void sched_idle_next(void)
5143 int this_cpu = smp_processor_id();
5144 struct rq *rq = cpu_rq(this_cpu);
5145 struct task_struct *p = rq->idle;
5146 unsigned long flags;
5148 /* cpu has to be offline */
5149 BUG_ON(cpu_online(this_cpu));
5152 * Strictly not necessary since rest of the CPUs are stopped by now
5153 * and interrupts disabled on the current cpu.
5155 spin_lock_irqsave(&rq->lock, flags);
5157 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5159 /* Add idle task to the _front_ of its priority queue: */
5160 activate_idle_task(p, rq);
5162 spin_unlock_irqrestore(&rq->lock, flags);
5166 * Ensures that the idle task is using init_mm right before its cpu goes
5169 void idle_task_exit(void)
5171 struct mm_struct *mm = current->active_mm;
5173 BUG_ON(cpu_online(smp_processor_id()));
5176 switch_mm(mm, &init_mm, current);
5180 /* called under rq->lock with disabled interrupts */
5181 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5183 struct rq *rq = cpu_rq(dead_cpu);
5185 /* Must be exiting, otherwise would be on tasklist. */
5186 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5188 /* Cannot have done final schedule yet: would have vanished. */
5189 BUG_ON(p->state == TASK_DEAD);
5194 * Drop lock around migration; if someone else moves it,
5195 * that's OK. No task can be added to this CPU, so iteration is
5197 * NOTE: interrupts should be left disabled --dev@
5199 spin_unlock(&rq->lock);
5200 move_task_off_dead_cpu(dead_cpu, p);
5201 spin_lock(&rq->lock);
5206 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5207 static void migrate_dead_tasks(unsigned int dead_cpu)
5209 struct rq *rq = cpu_rq(dead_cpu);
5210 struct task_struct *next;
5213 if (!rq->nr_running)
5215 update_rq_clock(rq);
5216 next = pick_next_task(rq, rq->curr);
5219 migrate_dead(dead_cpu, next);
5223 #endif /* CONFIG_HOTPLUG_CPU */
5225 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5227 static struct ctl_table sd_ctl_dir[] = {
5229 .procname = "sched_domain",
5235 static struct ctl_table sd_ctl_root[] = {
5237 .ctl_name = CTL_KERN,
5238 .procname = "kernel",
5240 .child = sd_ctl_dir,
5245 static struct ctl_table *sd_alloc_ctl_entry(int n)
5247 struct ctl_table *entry =
5248 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5251 memset(entry, 0, n * sizeof(struct ctl_table));
5257 set_table_entry(struct ctl_table *entry,
5258 const char *procname, void *data, int maxlen,
5259 mode_t mode, proc_handler *proc_handler)
5261 entry->procname = procname;
5263 entry->maxlen = maxlen;
5265 entry->proc_handler = proc_handler;
5268 static struct ctl_table *
5269 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5271 struct ctl_table *table = sd_alloc_ctl_entry(14);
5273 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5274 sizeof(long), 0644, proc_doulongvec_minmax);
5275 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5276 sizeof(long), 0644, proc_doulongvec_minmax);
5277 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5278 sizeof(int), 0644, proc_dointvec_minmax);
5279 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5280 sizeof(int), 0644, proc_dointvec_minmax);
5281 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5282 sizeof(int), 0644, proc_dointvec_minmax);
5283 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5284 sizeof(int), 0644, proc_dointvec_minmax);
5285 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5286 sizeof(int), 0644, proc_dointvec_minmax);
5287 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5288 sizeof(int), 0644, proc_dointvec_minmax);
5289 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5290 sizeof(int), 0644, proc_dointvec_minmax);
5291 set_table_entry(&table[10], "cache_nice_tries",
5292 &sd->cache_nice_tries,
5293 sizeof(int), 0644, proc_dointvec_minmax);
5294 set_table_entry(&table[12], "flags", &sd->flags,
5295 sizeof(int), 0644, proc_dointvec_minmax);
5300 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5302 struct ctl_table *entry, *table;
5303 struct sched_domain *sd;
5304 int domain_num = 0, i;
5307 for_each_domain(cpu, sd)
5309 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5312 for_each_domain(cpu, sd) {
5313 snprintf(buf, 32, "domain%d", i);
5314 entry->procname = kstrdup(buf, GFP_KERNEL);
5316 entry->child = sd_alloc_ctl_domain_table(sd);
5323 static struct ctl_table_header *sd_sysctl_header;
5324 static void init_sched_domain_sysctl(void)
5326 int i, cpu_num = num_online_cpus();
5327 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5330 sd_ctl_dir[0].child = entry;
5332 for (i = 0; i < cpu_num; i++, entry++) {
5333 snprintf(buf, 32, "cpu%d", i);
5334 entry->procname = kstrdup(buf, GFP_KERNEL);
5336 entry->child = sd_alloc_ctl_cpu_table(i);
5338 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5341 static void init_sched_domain_sysctl(void)
5347 * migration_call - callback that gets triggered when a CPU is added.
5348 * Here we can start up the necessary migration thread for the new CPU.
5350 static int __cpuinit
5351 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5353 struct task_struct *p;
5354 int cpu = (long)hcpu;
5355 unsigned long flags;
5359 case CPU_LOCK_ACQUIRE:
5360 mutex_lock(&sched_hotcpu_mutex);
5363 case CPU_UP_PREPARE:
5364 case CPU_UP_PREPARE_FROZEN:
5365 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5368 kthread_bind(p, cpu);
5369 /* Must be high prio: stop_machine expects to yield to it. */
5370 rq = task_rq_lock(p, &flags);
5371 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5372 task_rq_unlock(rq, &flags);
5373 cpu_rq(cpu)->migration_thread = p;
5377 case CPU_ONLINE_FROZEN:
5378 /* Strictly unneccessary, as first user will wake it. */
5379 wake_up_process(cpu_rq(cpu)->migration_thread);
5382 #ifdef CONFIG_HOTPLUG_CPU
5383 case CPU_UP_CANCELED:
5384 case CPU_UP_CANCELED_FROZEN:
5385 if (!cpu_rq(cpu)->migration_thread)
5387 /* Unbind it from offline cpu so it can run. Fall thru. */
5388 kthread_bind(cpu_rq(cpu)->migration_thread,
5389 any_online_cpu(cpu_online_map));
5390 kthread_stop(cpu_rq(cpu)->migration_thread);
5391 cpu_rq(cpu)->migration_thread = NULL;
5395 case CPU_DEAD_FROZEN:
5396 migrate_live_tasks(cpu);
5398 kthread_stop(rq->migration_thread);
5399 rq->migration_thread = NULL;
5400 /* Idle task back to normal (off runqueue, low prio) */
5401 rq = task_rq_lock(rq->idle, &flags);
5402 update_rq_clock(rq);
5403 deactivate_task(rq, rq->idle, 0);
5404 rq->idle->static_prio = MAX_PRIO;
5405 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5406 rq->idle->sched_class = &idle_sched_class;
5407 migrate_dead_tasks(cpu);
5408 task_rq_unlock(rq, &flags);
5409 migrate_nr_uninterruptible(rq);
5410 BUG_ON(rq->nr_running != 0);
5412 /* No need to migrate the tasks: it was best-effort if
5413 * they didn't take sched_hotcpu_mutex. Just wake up
5414 * the requestors. */
5415 spin_lock_irq(&rq->lock);
5416 while (!list_empty(&rq->migration_queue)) {
5417 struct migration_req *req;
5419 req = list_entry(rq->migration_queue.next,
5420 struct migration_req, list);
5421 list_del_init(&req->list);
5422 complete(&req->done);
5424 spin_unlock_irq(&rq->lock);
5427 case CPU_LOCK_RELEASE:
5428 mutex_unlock(&sched_hotcpu_mutex);
5434 /* Register at highest priority so that task migration (migrate_all_tasks)
5435 * happens before everything else.
5437 static struct notifier_block __cpuinitdata migration_notifier = {
5438 .notifier_call = migration_call,
5442 int __init migration_init(void)
5444 void *cpu = (void *)(long)smp_processor_id();
5447 /* Start one for the boot CPU: */
5448 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5449 BUG_ON(err == NOTIFY_BAD);
5450 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5451 register_cpu_notifier(&migration_notifier);
5459 /* Number of possible processor ids */
5460 int nr_cpu_ids __read_mostly = NR_CPUS;
5461 EXPORT_SYMBOL(nr_cpu_ids);
5463 #undef SCHED_DOMAIN_DEBUG
5464 #ifdef SCHED_DOMAIN_DEBUG
5465 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5470 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5474 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5479 struct sched_group *group = sd->groups;
5480 cpumask_t groupmask;
5482 cpumask_scnprintf(str, NR_CPUS, sd->span);
5483 cpus_clear(groupmask);
5486 for (i = 0; i < level + 1; i++)
5488 printk("domain %d: ", level);
5490 if (!(sd->flags & SD_LOAD_BALANCE)) {
5491 printk("does not load-balance\n");
5493 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5498 printk("span %s\n", str);
5500 if (!cpu_isset(cpu, sd->span))
5501 printk(KERN_ERR "ERROR: domain->span does not contain "
5503 if (!cpu_isset(cpu, group->cpumask))
5504 printk(KERN_ERR "ERROR: domain->groups does not contain"
5508 for (i = 0; i < level + 2; i++)
5514 printk(KERN_ERR "ERROR: group is NULL\n");
5518 if (!group->__cpu_power) {
5520 printk(KERN_ERR "ERROR: domain->cpu_power not "
5524 if (!cpus_weight(group->cpumask)) {
5526 printk(KERN_ERR "ERROR: empty group\n");
5529 if (cpus_intersects(groupmask, group->cpumask)) {
5531 printk(KERN_ERR "ERROR: repeated CPUs\n");
5534 cpus_or(groupmask, groupmask, group->cpumask);
5536 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5539 group = group->next;
5540 } while (group != sd->groups);
5543 if (!cpus_equal(sd->span, groupmask))
5544 printk(KERN_ERR "ERROR: groups don't span "
5552 if (!cpus_subset(groupmask, sd->span))
5553 printk(KERN_ERR "ERROR: parent span is not a superset "
5554 "of domain->span\n");
5559 # define sched_domain_debug(sd, cpu) do { } while (0)
5562 static int sd_degenerate(struct sched_domain *sd)
5564 if (cpus_weight(sd->span) == 1)
5567 /* Following flags need at least 2 groups */
5568 if (sd->flags & (SD_LOAD_BALANCE |
5569 SD_BALANCE_NEWIDLE |
5573 SD_SHARE_PKG_RESOURCES)) {
5574 if (sd->groups != sd->groups->next)
5578 /* Following flags don't use groups */
5579 if (sd->flags & (SD_WAKE_IDLE |
5588 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5590 unsigned long cflags = sd->flags, pflags = parent->flags;
5592 if (sd_degenerate(parent))
5595 if (!cpus_equal(sd->span, parent->span))
5598 /* Does parent contain flags not in child? */
5599 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5600 if (cflags & SD_WAKE_AFFINE)
5601 pflags &= ~SD_WAKE_BALANCE;
5602 /* Flags needing groups don't count if only 1 group in parent */
5603 if (parent->groups == parent->groups->next) {
5604 pflags &= ~(SD_LOAD_BALANCE |
5605 SD_BALANCE_NEWIDLE |
5609 SD_SHARE_PKG_RESOURCES);
5611 if (~cflags & pflags)
5618 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5619 * hold the hotplug lock.
5621 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5623 struct rq *rq = cpu_rq(cpu);
5624 struct sched_domain *tmp;
5626 /* Remove the sched domains which do not contribute to scheduling. */
5627 for (tmp = sd; tmp; tmp = tmp->parent) {
5628 struct sched_domain *parent = tmp->parent;
5631 if (sd_parent_degenerate(tmp, parent)) {
5632 tmp->parent = parent->parent;
5634 parent->parent->child = tmp;
5638 if (sd && sd_degenerate(sd)) {
5644 sched_domain_debug(sd, cpu);
5646 rcu_assign_pointer(rq->sd, sd);
5649 /* cpus with isolated domains */
5650 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5652 /* Setup the mask of cpus configured for isolated domains */
5653 static int __init isolated_cpu_setup(char *str)
5655 int ints[NR_CPUS], i;
5657 str = get_options(str, ARRAY_SIZE(ints), ints);
5658 cpus_clear(cpu_isolated_map);
5659 for (i = 1; i <= ints[0]; i++)
5660 if (ints[i] < NR_CPUS)
5661 cpu_set(ints[i], cpu_isolated_map);
5665 __setup ("isolcpus=", isolated_cpu_setup);
5668 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5669 * to a function which identifies what group(along with sched group) a CPU
5670 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5671 * (due to the fact that we keep track of groups covered with a cpumask_t).
5673 * init_sched_build_groups will build a circular linked list of the groups
5674 * covered by the given span, and will set each group's ->cpumask correctly,
5675 * and ->cpu_power to 0.
5678 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5679 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5680 struct sched_group **sg))
5682 struct sched_group *first = NULL, *last = NULL;
5683 cpumask_t covered = CPU_MASK_NONE;
5686 for_each_cpu_mask(i, span) {
5687 struct sched_group *sg;
5688 int group = group_fn(i, cpu_map, &sg);
5691 if (cpu_isset(i, covered))
5694 sg->cpumask = CPU_MASK_NONE;
5695 sg->__cpu_power = 0;
5697 for_each_cpu_mask(j, span) {
5698 if (group_fn(j, cpu_map, NULL) != group)
5701 cpu_set(j, covered);
5702 cpu_set(j, sg->cpumask);
5713 #define SD_NODES_PER_DOMAIN 16
5718 * find_next_best_node - find the next node to include in a sched_domain
5719 * @node: node whose sched_domain we're building
5720 * @used_nodes: nodes already in the sched_domain
5722 * Find the next node to include in a given scheduling domain. Simply
5723 * finds the closest node not already in the @used_nodes map.
5725 * Should use nodemask_t.
5727 static int find_next_best_node(int node, unsigned long *used_nodes)
5729 int i, n, val, min_val, best_node = 0;
5733 for (i = 0; i < MAX_NUMNODES; i++) {
5734 /* Start at @node */
5735 n = (node + i) % MAX_NUMNODES;
5737 if (!nr_cpus_node(n))
5740 /* Skip already used nodes */
5741 if (test_bit(n, used_nodes))
5744 /* Simple min distance search */
5745 val = node_distance(node, n);
5747 if (val < min_val) {
5753 set_bit(best_node, used_nodes);
5758 * sched_domain_node_span - get a cpumask for a node's sched_domain
5759 * @node: node whose cpumask we're constructing
5760 * @size: number of nodes to include in this span
5762 * Given a node, construct a good cpumask for its sched_domain to span. It
5763 * should be one that prevents unnecessary balancing, but also spreads tasks
5766 static cpumask_t sched_domain_node_span(int node)
5768 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5769 cpumask_t span, nodemask;
5773 bitmap_zero(used_nodes, MAX_NUMNODES);
5775 nodemask = node_to_cpumask(node);
5776 cpus_or(span, span, nodemask);
5777 set_bit(node, used_nodes);
5779 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5780 int next_node = find_next_best_node(node, used_nodes);
5782 nodemask = node_to_cpumask(next_node);
5783 cpus_or(span, span, nodemask);
5790 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5793 * SMT sched-domains:
5795 #ifdef CONFIG_SCHED_SMT
5796 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5797 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5799 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5800 struct sched_group **sg)
5803 *sg = &per_cpu(sched_group_cpus, cpu);
5809 * multi-core sched-domains:
5811 #ifdef CONFIG_SCHED_MC
5812 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5813 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5816 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5817 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5818 struct sched_group **sg)
5821 cpumask_t mask = cpu_sibling_map[cpu];
5822 cpus_and(mask, mask, *cpu_map);
5823 group = first_cpu(mask);
5825 *sg = &per_cpu(sched_group_core, group);
5828 #elif defined(CONFIG_SCHED_MC)
5829 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5830 struct sched_group **sg)
5833 *sg = &per_cpu(sched_group_core, cpu);
5838 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5839 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5841 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5842 struct sched_group **sg)
5845 #ifdef CONFIG_SCHED_MC
5846 cpumask_t mask = cpu_coregroup_map(cpu);
5847 cpus_and(mask, mask, *cpu_map);
5848 group = first_cpu(mask);
5849 #elif defined(CONFIG_SCHED_SMT)
5850 cpumask_t mask = cpu_sibling_map[cpu];
5851 cpus_and(mask, mask, *cpu_map);
5852 group = first_cpu(mask);
5857 *sg = &per_cpu(sched_group_phys, group);
5863 * The init_sched_build_groups can't handle what we want to do with node
5864 * groups, so roll our own. Now each node has its own list of groups which
5865 * gets dynamically allocated.
5867 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5868 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5870 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5871 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5873 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5874 struct sched_group **sg)
5876 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5879 cpus_and(nodemask, nodemask, *cpu_map);
5880 group = first_cpu(nodemask);
5883 *sg = &per_cpu(sched_group_allnodes, group);
5887 static void init_numa_sched_groups_power(struct sched_group *group_head)
5889 struct sched_group *sg = group_head;
5895 for_each_cpu_mask(j, sg->cpumask) {
5896 struct sched_domain *sd;
5898 sd = &per_cpu(phys_domains, j);
5899 if (j != first_cpu(sd->groups->cpumask)) {
5901 * Only add "power" once for each
5907 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5910 if (sg != group_head)
5916 /* Free memory allocated for various sched_group structures */
5917 static void free_sched_groups(const cpumask_t *cpu_map)
5921 for_each_cpu_mask(cpu, *cpu_map) {
5922 struct sched_group **sched_group_nodes
5923 = sched_group_nodes_bycpu[cpu];
5925 if (!sched_group_nodes)
5928 for (i = 0; i < MAX_NUMNODES; i++) {
5929 cpumask_t nodemask = node_to_cpumask(i);
5930 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5932 cpus_and(nodemask, nodemask, *cpu_map);
5933 if (cpus_empty(nodemask))
5943 if (oldsg != sched_group_nodes[i])
5946 kfree(sched_group_nodes);
5947 sched_group_nodes_bycpu[cpu] = NULL;
5951 static void free_sched_groups(const cpumask_t *cpu_map)
5957 * Initialize sched groups cpu_power.
5959 * cpu_power indicates the capacity of sched group, which is used while
5960 * distributing the load between different sched groups in a sched domain.
5961 * Typically cpu_power for all the groups in a sched domain will be same unless
5962 * there are asymmetries in the topology. If there are asymmetries, group
5963 * having more cpu_power will pickup more load compared to the group having
5966 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5967 * the maximum number of tasks a group can handle in the presence of other idle
5968 * or lightly loaded groups in the same sched domain.
5970 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5972 struct sched_domain *child;
5973 struct sched_group *group;
5975 WARN_ON(!sd || !sd->groups);
5977 if (cpu != first_cpu(sd->groups->cpumask))
5982 sd->groups->__cpu_power = 0;
5985 * For perf policy, if the groups in child domain share resources
5986 * (for example cores sharing some portions of the cache hierarchy
5987 * or SMT), then set this domain groups cpu_power such that each group
5988 * can handle only one task, when there are other idle groups in the
5989 * same sched domain.
5991 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5993 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5994 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5999 * add cpu_power of each child group to this groups cpu_power
6001 group = child->groups;
6003 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6004 group = group->next;
6005 } while (group != child->groups);
6009 * Build sched domains for a given set of cpus and attach the sched domains
6010 * to the individual cpus
6012 static int build_sched_domains(const cpumask_t *cpu_map)
6016 struct sched_group **sched_group_nodes = NULL;
6017 int sd_allnodes = 0;
6020 * Allocate the per-node list of sched groups
6022 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
6024 if (!sched_group_nodes) {
6025 printk(KERN_WARNING "Can not alloc sched group node list\n");
6028 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6032 * Set up domains for cpus specified by the cpu_map.
6034 for_each_cpu_mask(i, *cpu_map) {
6035 struct sched_domain *sd = NULL, *p;
6036 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6038 cpus_and(nodemask, nodemask, *cpu_map);
6041 if (cpus_weight(*cpu_map) >
6042 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6043 sd = &per_cpu(allnodes_domains, i);
6044 *sd = SD_ALLNODES_INIT;
6045 sd->span = *cpu_map;
6046 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6052 sd = &per_cpu(node_domains, i);
6054 sd->span = sched_domain_node_span(cpu_to_node(i));
6058 cpus_and(sd->span, sd->span, *cpu_map);
6062 sd = &per_cpu(phys_domains, i);
6064 sd->span = nodemask;
6068 cpu_to_phys_group(i, cpu_map, &sd->groups);
6070 #ifdef CONFIG_SCHED_MC
6072 sd = &per_cpu(core_domains, i);
6074 sd->span = cpu_coregroup_map(i);
6075 cpus_and(sd->span, sd->span, *cpu_map);
6078 cpu_to_core_group(i, cpu_map, &sd->groups);
6081 #ifdef CONFIG_SCHED_SMT
6083 sd = &per_cpu(cpu_domains, i);
6084 *sd = SD_SIBLING_INIT;
6085 sd->span = cpu_sibling_map[i];
6086 cpus_and(sd->span, sd->span, *cpu_map);
6089 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6093 #ifdef CONFIG_SCHED_SMT
6094 /* Set up CPU (sibling) groups */
6095 for_each_cpu_mask(i, *cpu_map) {
6096 cpumask_t this_sibling_map = cpu_sibling_map[i];
6097 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6098 if (i != first_cpu(this_sibling_map))
6101 init_sched_build_groups(this_sibling_map, cpu_map,
6106 #ifdef CONFIG_SCHED_MC
6107 /* Set up multi-core groups */
6108 for_each_cpu_mask(i, *cpu_map) {
6109 cpumask_t this_core_map = cpu_coregroup_map(i);
6110 cpus_and(this_core_map, this_core_map, *cpu_map);
6111 if (i != first_cpu(this_core_map))
6113 init_sched_build_groups(this_core_map, cpu_map,
6114 &cpu_to_core_group);
6118 /* Set up physical groups */
6119 for (i = 0; i < MAX_NUMNODES; i++) {
6120 cpumask_t nodemask = node_to_cpumask(i);
6122 cpus_and(nodemask, nodemask, *cpu_map);
6123 if (cpus_empty(nodemask))
6126 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6130 /* Set up node groups */
6132 init_sched_build_groups(*cpu_map, cpu_map,
6133 &cpu_to_allnodes_group);
6135 for (i = 0; i < MAX_NUMNODES; i++) {
6136 /* Set up node groups */
6137 struct sched_group *sg, *prev;
6138 cpumask_t nodemask = node_to_cpumask(i);
6139 cpumask_t domainspan;
6140 cpumask_t covered = CPU_MASK_NONE;
6143 cpus_and(nodemask, nodemask, *cpu_map);
6144 if (cpus_empty(nodemask)) {
6145 sched_group_nodes[i] = NULL;
6149 domainspan = sched_domain_node_span(i);
6150 cpus_and(domainspan, domainspan, *cpu_map);
6152 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6154 printk(KERN_WARNING "Can not alloc domain group for "
6158 sched_group_nodes[i] = sg;
6159 for_each_cpu_mask(j, nodemask) {
6160 struct sched_domain *sd;
6162 sd = &per_cpu(node_domains, j);
6165 sg->__cpu_power = 0;
6166 sg->cpumask = nodemask;
6168 cpus_or(covered, covered, nodemask);
6171 for (j = 0; j < MAX_NUMNODES; j++) {
6172 cpumask_t tmp, notcovered;
6173 int n = (i + j) % MAX_NUMNODES;
6175 cpus_complement(notcovered, covered);
6176 cpus_and(tmp, notcovered, *cpu_map);
6177 cpus_and(tmp, tmp, domainspan);
6178 if (cpus_empty(tmp))
6181 nodemask = node_to_cpumask(n);
6182 cpus_and(tmp, tmp, nodemask);
6183 if (cpus_empty(tmp))
6186 sg = kmalloc_node(sizeof(struct sched_group),
6190 "Can not alloc domain group for node %d\n", j);
6193 sg->__cpu_power = 0;
6195 sg->next = prev->next;
6196 cpus_or(covered, covered, tmp);
6203 /* Calculate CPU power for physical packages and nodes */
6204 #ifdef CONFIG_SCHED_SMT
6205 for_each_cpu_mask(i, *cpu_map) {
6206 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6208 init_sched_groups_power(i, sd);
6211 #ifdef CONFIG_SCHED_MC
6212 for_each_cpu_mask(i, *cpu_map) {
6213 struct sched_domain *sd = &per_cpu(core_domains, i);
6215 init_sched_groups_power(i, sd);
6219 for_each_cpu_mask(i, *cpu_map) {
6220 struct sched_domain *sd = &per_cpu(phys_domains, i);
6222 init_sched_groups_power(i, sd);
6226 for (i = 0; i < MAX_NUMNODES; i++)
6227 init_numa_sched_groups_power(sched_group_nodes[i]);
6230 struct sched_group *sg;
6232 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6233 init_numa_sched_groups_power(sg);
6237 /* Attach the domains */
6238 for_each_cpu_mask(i, *cpu_map) {
6239 struct sched_domain *sd;
6240 #ifdef CONFIG_SCHED_SMT
6241 sd = &per_cpu(cpu_domains, i);
6242 #elif defined(CONFIG_SCHED_MC)
6243 sd = &per_cpu(core_domains, i);
6245 sd = &per_cpu(phys_domains, i);
6247 cpu_attach_domain(sd, i);
6254 free_sched_groups(cpu_map);
6259 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6261 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6263 cpumask_t cpu_default_map;
6267 * Setup mask for cpus without special case scheduling requirements.
6268 * For now this just excludes isolated cpus, but could be used to
6269 * exclude other special cases in the future.
6271 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6273 err = build_sched_domains(&cpu_default_map);
6278 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6280 free_sched_groups(cpu_map);
6284 * Detach sched domains from a group of cpus specified in cpu_map
6285 * These cpus will now be attached to the NULL domain
6287 static void detach_destroy_domains(const cpumask_t *cpu_map)
6291 for_each_cpu_mask(i, *cpu_map)
6292 cpu_attach_domain(NULL, i);
6293 synchronize_sched();
6294 arch_destroy_sched_domains(cpu_map);
6298 * Partition sched domains as specified by the cpumasks below.
6299 * This attaches all cpus from the cpumasks to the NULL domain,
6300 * waits for a RCU quiescent period, recalculates sched
6301 * domain information and then attaches them back to the
6302 * correct sched domains
6303 * Call with hotplug lock held
6305 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6307 cpumask_t change_map;
6310 cpus_and(*partition1, *partition1, cpu_online_map);
6311 cpus_and(*partition2, *partition2, cpu_online_map);
6312 cpus_or(change_map, *partition1, *partition2);
6314 /* Detach sched domains from all of the affected cpus */
6315 detach_destroy_domains(&change_map);
6316 if (!cpus_empty(*partition1))
6317 err = build_sched_domains(partition1);
6318 if (!err && !cpus_empty(*partition2))
6319 err = build_sched_domains(partition2);
6324 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6325 static int arch_reinit_sched_domains(void)
6329 mutex_lock(&sched_hotcpu_mutex);
6330 detach_destroy_domains(&cpu_online_map);
6331 err = arch_init_sched_domains(&cpu_online_map);
6332 mutex_unlock(&sched_hotcpu_mutex);
6337 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6341 if (buf[0] != '0' && buf[0] != '1')
6345 sched_smt_power_savings = (buf[0] == '1');
6347 sched_mc_power_savings = (buf[0] == '1');
6349 ret = arch_reinit_sched_domains();
6351 return ret ? ret : count;
6354 #ifdef CONFIG_SCHED_MC
6355 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6357 return sprintf(page, "%u\n", sched_mc_power_savings);
6359 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6360 const char *buf, size_t count)
6362 return sched_power_savings_store(buf, count, 0);
6364 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6365 sched_mc_power_savings_store);
6368 #ifdef CONFIG_SCHED_SMT
6369 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6371 return sprintf(page, "%u\n", sched_smt_power_savings);
6373 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6374 const char *buf, size_t count)
6376 return sched_power_savings_store(buf, count, 1);
6378 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6379 sched_smt_power_savings_store);
6382 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6386 #ifdef CONFIG_SCHED_SMT
6388 err = sysfs_create_file(&cls->kset.kobj,
6389 &attr_sched_smt_power_savings.attr);
6391 #ifdef CONFIG_SCHED_MC
6392 if (!err && mc_capable())
6393 err = sysfs_create_file(&cls->kset.kobj,
6394 &attr_sched_mc_power_savings.attr);
6401 * Force a reinitialization of the sched domains hierarchy. The domains
6402 * and groups cannot be updated in place without racing with the balancing
6403 * code, so we temporarily attach all running cpus to the NULL domain
6404 * which will prevent rebalancing while the sched domains are recalculated.
6406 static int update_sched_domains(struct notifier_block *nfb,
6407 unsigned long action, void *hcpu)
6410 case CPU_UP_PREPARE:
6411 case CPU_UP_PREPARE_FROZEN:
6412 case CPU_DOWN_PREPARE:
6413 case CPU_DOWN_PREPARE_FROZEN:
6414 detach_destroy_domains(&cpu_online_map);
6417 case CPU_UP_CANCELED:
6418 case CPU_UP_CANCELED_FROZEN:
6419 case CPU_DOWN_FAILED:
6420 case CPU_DOWN_FAILED_FROZEN:
6422 case CPU_ONLINE_FROZEN:
6424 case CPU_DEAD_FROZEN:
6426 * Fall through and re-initialise the domains.
6433 /* The hotplug lock is already held by cpu_up/cpu_down */
6434 arch_init_sched_domains(&cpu_online_map);
6439 void __init sched_init_smp(void)
6441 cpumask_t non_isolated_cpus;
6443 mutex_lock(&sched_hotcpu_mutex);
6444 arch_init_sched_domains(&cpu_online_map);
6445 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6446 if (cpus_empty(non_isolated_cpus))
6447 cpu_set(smp_processor_id(), non_isolated_cpus);
6448 mutex_unlock(&sched_hotcpu_mutex);
6449 /* XXX: Theoretical race here - CPU may be hotplugged now */
6450 hotcpu_notifier(update_sched_domains, 0);
6452 init_sched_domain_sysctl();
6454 /* Move init over to a non-isolated CPU */
6455 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6459 void __init sched_init_smp(void)
6462 #endif /* CONFIG_SMP */
6464 int in_sched_functions(unsigned long addr)
6466 /* Linker adds these: start and end of __sched functions */
6467 extern char __sched_text_start[], __sched_text_end[];
6469 return in_lock_functions(addr) ||
6470 (addr >= (unsigned long)__sched_text_start
6471 && addr < (unsigned long)__sched_text_end);
6474 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6476 cfs_rq->tasks_timeline = RB_ROOT;
6477 #ifdef CONFIG_FAIR_GROUP_SCHED
6482 void __init sched_init(void)
6484 int highest_cpu = 0;
6488 * Link up the scheduling class hierarchy:
6490 rt_sched_class.next = &fair_sched_class;
6491 fair_sched_class.next = &idle_sched_class;
6492 idle_sched_class.next = NULL;
6494 for_each_possible_cpu(i) {
6495 struct rt_prio_array *array;
6499 spin_lock_init(&rq->lock);
6500 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6503 init_cfs_rq(&rq->cfs, rq);
6504 #ifdef CONFIG_FAIR_GROUP_SCHED
6505 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6507 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6508 struct sched_entity *se =
6509 &per_cpu(init_sched_entity, i);
6511 init_cfs_rq_p[i] = cfs_rq;
6512 init_cfs_rq(cfs_rq, rq);
6513 cfs_rq->tg = &init_task_grp;
6514 list_add(&cfs_rq->leaf_cfs_rq_list,
6515 &rq->leaf_cfs_rq_list);
6517 init_sched_entity_p[i] = se;
6518 se->cfs_rq = &rq->cfs;
6520 se->load.weight = init_task_grp_load;
6521 se->load.inv_weight =
6522 div64_64(1ULL<<32, init_task_grp_load);
6525 init_task_grp.shares = init_task_grp_load;
6528 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6529 rq->cpu_load[j] = 0;
6532 rq->active_balance = 0;
6533 rq->next_balance = jiffies;
6536 rq->migration_thread = NULL;
6537 INIT_LIST_HEAD(&rq->migration_queue);
6539 atomic_set(&rq->nr_iowait, 0);
6541 array = &rq->rt.active;
6542 for (j = 0; j < MAX_RT_PRIO; j++) {
6543 INIT_LIST_HEAD(array->queue + j);
6544 __clear_bit(j, array->bitmap);
6547 /* delimiter for bitsearch: */
6548 __set_bit(MAX_RT_PRIO, array->bitmap);
6551 set_load_weight(&init_task);
6553 #ifdef CONFIG_PREEMPT_NOTIFIERS
6554 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6558 nr_cpu_ids = highest_cpu + 1;
6559 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6562 #ifdef CONFIG_RT_MUTEXES
6563 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6567 * The boot idle thread does lazy MMU switching as well:
6569 atomic_inc(&init_mm.mm_count);
6570 enter_lazy_tlb(&init_mm, current);
6573 * Make us the idle thread. Technically, schedule() should not be
6574 * called from this thread, however somewhere below it might be,
6575 * but because we are the idle thread, we just pick up running again
6576 * when this runqueue becomes "idle".
6578 init_idle(current, smp_processor_id());
6580 * During early bootup we pretend to be a normal task:
6582 current->sched_class = &fair_sched_class;
6585 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6586 void __might_sleep(char *file, int line)
6589 static unsigned long prev_jiffy; /* ratelimiting */
6591 if ((in_atomic() || irqs_disabled()) &&
6592 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6593 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6595 prev_jiffy = jiffies;
6596 printk(KERN_ERR "BUG: sleeping function called from invalid"
6597 " context at %s:%d\n", file, line);
6598 printk("in_atomic():%d, irqs_disabled():%d\n",
6599 in_atomic(), irqs_disabled());
6600 debug_show_held_locks(current);
6601 if (irqs_disabled())
6602 print_irqtrace_events(current);
6607 EXPORT_SYMBOL(__might_sleep);
6610 #ifdef CONFIG_MAGIC_SYSRQ
6611 void normalize_rt_tasks(void)
6613 struct task_struct *g, *p;
6614 unsigned long flags;
6618 read_lock_irq(&tasklist_lock);
6619 do_each_thread(g, p) {
6620 p->se.exec_start = 0;
6621 #ifdef CONFIG_SCHEDSTATS
6622 p->se.wait_start = 0;
6623 p->se.sleep_start = 0;
6624 p->se.block_start = 0;
6626 task_rq(p)->clock = 0;
6630 * Renice negative nice level userspace
6633 if (TASK_NICE(p) < 0 && p->mm)
6634 set_user_nice(p, 0);
6638 spin_lock_irqsave(&p->pi_lock, flags);
6639 rq = __task_rq_lock(p);
6642 * Do not touch the migration thread:
6644 if (p == rq->migration_thread)
6648 update_rq_clock(rq);
6649 on_rq = p->se.on_rq;
6651 deactivate_task(rq, p, 0);
6652 __setscheduler(rq, p, SCHED_NORMAL, 0);
6654 activate_task(rq, p, 0);
6655 resched_task(rq->curr);
6660 __task_rq_unlock(rq);
6661 spin_unlock_irqrestore(&p->pi_lock, flags);
6662 } while_each_thread(g, p);
6664 read_unlock_irq(&tasklist_lock);
6667 #endif /* CONFIG_MAGIC_SYSRQ */
6671 * These functions are only useful for the IA64 MCA handling.
6673 * They can only be called when the whole system has been
6674 * stopped - every CPU needs to be quiescent, and no scheduling
6675 * activity can take place. Using them for anything else would
6676 * be a serious bug, and as a result, they aren't even visible
6677 * under any other configuration.
6681 * curr_task - return the current task for a given cpu.
6682 * @cpu: the processor in question.
6684 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6686 struct task_struct *curr_task(int cpu)
6688 return cpu_curr(cpu);
6692 * set_curr_task - set the current task for a given cpu.
6693 * @cpu: the processor in question.
6694 * @p: the task pointer to set.
6696 * Description: This function must only be used when non-maskable interrupts
6697 * are serviced on a separate stack. It allows the architecture to switch the
6698 * notion of the current task on a cpu in a non-blocking manner. This function
6699 * must be called with all CPU's synchronized, and interrupts disabled, the
6700 * and caller must save the original value of the current task (see
6701 * curr_task() above) and restore that value before reenabling interrupts and
6702 * re-starting the system.
6704 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6706 void set_curr_task(int cpu, struct task_struct *p)
6713 #ifdef CONFIG_FAIR_GROUP_SCHED
6715 /* allocate runqueue etc for a new task group */
6716 struct task_grp *sched_create_group(void)
6718 struct task_grp *tg;
6719 struct cfs_rq *cfs_rq;
6720 struct sched_entity *se;
6724 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6726 return ERR_PTR(-ENOMEM);
6728 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6731 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6735 for_each_possible_cpu(i) {
6738 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6743 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6748 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6749 memset(se, 0, sizeof(struct sched_entity));
6751 tg->cfs_rq[i] = cfs_rq;
6752 init_cfs_rq(cfs_rq, rq);
6756 se->cfs_rq = &rq->cfs;
6758 se->load.weight = NICE_0_LOAD;
6759 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
6763 for_each_possible_cpu(i) {
6765 cfs_rq = tg->cfs_rq[i];
6766 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6769 tg->shares = NICE_0_LOAD;
6774 for_each_possible_cpu(i) {
6775 if (tg->cfs_rq && tg->cfs_rq[i])
6776 kfree(tg->cfs_rq[i]);
6777 if (tg->se && tg->se[i])
6787 return ERR_PTR(-ENOMEM);
6790 /* rcu callback to free various structures associated with a task group */
6791 static void free_sched_group(struct rcu_head *rhp)
6793 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
6794 struct task_grp *tg = cfs_rq->tg;
6795 struct sched_entity *se;
6798 /* now it should be safe to free those cfs_rqs */
6799 for_each_possible_cpu(i) {
6800 cfs_rq = tg->cfs_rq[i];
6812 /* Destroy runqueue etc associated with a task group */
6813 void sched_destroy_group(struct task_grp *tg)
6815 struct cfs_rq *cfs_rq;
6818 for_each_possible_cpu(i) {
6819 cfs_rq = tg->cfs_rq[i];
6820 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
6823 cfs_rq = tg->cfs_rq[0];
6825 /* wait for possible concurrent references to cfs_rqs complete */
6826 call_rcu(&cfs_rq->rcu, free_sched_group);
6829 /* change task's runqueue when it moves between groups.
6830 * The caller of this function should have put the task in its new group
6831 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6832 * reflect its new group.
6834 void sched_move_task(struct task_struct *tsk)
6837 unsigned long flags;
6840 rq = task_rq_lock(tsk, &flags);
6842 if (tsk->sched_class != &fair_sched_class)
6845 update_rq_clock(rq);
6847 running = task_running(rq, tsk);
6848 on_rq = tsk->se.on_rq;
6851 dequeue_task(rq, tsk, 0);
6852 if (unlikely(running))
6853 tsk->sched_class->put_prev_task(rq, tsk);
6856 set_task_cfs_rq(tsk);
6859 if (unlikely(running))
6860 tsk->sched_class->set_curr_task(rq);
6861 enqueue_task(rq, tsk, 0);
6865 task_rq_unlock(rq, &flags);
6868 static void set_se_shares(struct sched_entity *se, unsigned long shares)
6870 struct cfs_rq *cfs_rq = se->cfs_rq;
6871 struct rq *rq = cfs_rq->rq;
6874 spin_lock_irq(&rq->lock);
6878 dequeue_entity(cfs_rq, se, 0);
6880 se->load.weight = shares;
6881 se->load.inv_weight = div64_64((1ULL<<32), shares);
6884 enqueue_entity(cfs_rq, se, 0);
6886 spin_unlock_irq(&rq->lock);
6889 int sched_group_set_shares(struct task_grp *tg, unsigned long shares)
6893 if (tg->shares == shares)
6896 /* return -EINVAL if the new value is not sane */
6898 tg->shares = shares;
6899 for_each_possible_cpu(i)
6900 set_se_shares(tg->se[i], shares);
6905 #endif /* CONFIG_FAIR_GROUP_SCHED */