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) ((unsigned long)(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 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
116 * Since cpu_power is a 'constant', we can use a reciprocal divide.
118 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
120 return reciprocal_divide(load, sg->reciprocal_cpu_power);
124 * Each time a sched group cpu_power is changed,
125 * we must compute its reciprocal value
127 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
129 sg->__cpu_power += val;
130 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
134 static inline int rt_policy(int policy)
136 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
141 static inline int task_has_rt_policy(struct task_struct *p)
143 return rt_policy(p->policy);
147 * This is the priority-queue data structure of the RT scheduling class:
149 struct rt_prio_array {
150 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
151 struct list_head queue[MAX_RT_PRIO];
154 #ifdef CONFIG_FAIR_GROUP_SCHED
158 /* task group related information */
160 /* schedulable entities of this group on each cpu */
161 struct sched_entity **se;
162 /* runqueue "owned" by this group on each cpu */
163 struct cfs_rq **cfs_rq;
164 unsigned long shares;
167 /* Default task group's sched entity on each cpu */
168 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
169 /* Default task group's cfs_rq on each cpu */
170 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
172 static struct sched_entity *init_sched_entity_p[NR_CPUS];
173 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
175 /* Default task group.
176 * Every task in system belong to this group at bootup.
178 struct task_group init_task_group = {
179 .se = init_sched_entity_p,
180 .cfs_rq = init_cfs_rq_p,
183 #ifdef CONFIG_FAIR_USER_SCHED
184 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
186 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
189 static int init_task_group_load = INIT_TASK_GRP_LOAD;
191 /* return group to which a task belongs */
192 static inline struct task_group *task_group(struct task_struct *p)
194 struct task_group *tg;
196 #ifdef CONFIG_FAIR_USER_SCHED
199 tg = &init_task_group;
205 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
206 static inline void set_task_cfs_rq(struct task_struct *p)
208 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
209 p->se.parent = task_group(p)->se[task_cpu(p)];
214 static inline void set_task_cfs_rq(struct task_struct *p) { }
216 #endif /* CONFIG_FAIR_GROUP_SCHED */
218 /* CFS-related fields in a runqueue */
220 struct load_weight load;
221 unsigned long nr_running;
226 struct rb_root tasks_timeline;
227 struct rb_node *rb_leftmost;
228 struct rb_node *rb_load_balance_curr;
229 /* 'curr' points to currently running entity on this cfs_rq.
230 * It is set to NULL otherwise (i.e when none are currently running).
232 struct sched_entity *curr;
234 unsigned long nr_spread_over;
236 #ifdef CONFIG_FAIR_GROUP_SCHED
237 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
239 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
240 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
241 * (like users, containers etc.)
243 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
244 * list is used during load balance.
246 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
247 struct task_group *tg; /* group that "owns" this runqueue */
252 /* Real-Time classes' related field in a runqueue: */
254 struct rt_prio_array active;
255 int rt_load_balance_idx;
256 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
260 * This is the main, per-CPU runqueue data structure.
262 * Locking rule: those places that want to lock multiple runqueues
263 * (such as the load balancing or the thread migration code), lock
264 * acquire operations must be ordered by ascending &runqueue.
267 spinlock_t lock; /* runqueue lock */
270 * nr_running and cpu_load should be in the same cacheline because
271 * remote CPUs use both these fields when doing load calculation.
273 unsigned long nr_running;
274 #define CPU_LOAD_IDX_MAX 5
275 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
276 unsigned char idle_at_tick;
278 unsigned char in_nohz_recently;
280 struct load_weight load; /* capture load from *all* tasks on this cpu */
281 unsigned long nr_load_updates;
285 #ifdef CONFIG_FAIR_GROUP_SCHED
286 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
291 * This is part of a global counter where only the total sum
292 * over all CPUs matters. A task can increase this counter on
293 * one CPU and if it got migrated afterwards it may decrease
294 * it on another CPU. Always updated under the runqueue lock:
296 unsigned long nr_uninterruptible;
298 struct task_struct *curr, *idle;
299 unsigned long next_balance;
300 struct mm_struct *prev_mm;
302 u64 clock, prev_clock_raw;
305 unsigned int clock_warps, clock_overflows;
307 unsigned int clock_deep_idle_events;
313 struct sched_domain *sd;
315 /* For active balancing */
318 int cpu; /* cpu of this runqueue */
320 struct task_struct *migration_thread;
321 struct list_head migration_queue;
324 #ifdef CONFIG_SCHEDSTATS
326 struct sched_info rq_sched_info;
328 /* sys_sched_yield() stats */
329 unsigned long yld_exp_empty;
330 unsigned long yld_act_empty;
331 unsigned long yld_both_empty;
332 unsigned long yld_count;
334 /* schedule() stats */
335 unsigned long sched_switch;
336 unsigned long sched_count;
337 unsigned long sched_goidle;
339 /* try_to_wake_up() stats */
340 unsigned long ttwu_count;
341 unsigned long ttwu_local;
344 unsigned long bkl_count;
346 struct lock_class_key rq_lock_key;
349 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
350 static DEFINE_MUTEX(sched_hotcpu_mutex);
352 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
354 rq->curr->sched_class->check_preempt_curr(rq, p);
357 static inline int cpu_of(struct rq *rq)
367 * Update the per-runqueue clock, as finegrained as the platform can give
368 * us, but without assuming monotonicity, etc.:
370 static void __update_rq_clock(struct rq *rq)
372 u64 prev_raw = rq->prev_clock_raw;
373 u64 now = sched_clock();
374 s64 delta = now - prev_raw;
375 u64 clock = rq->clock;
377 #ifdef CONFIG_SCHED_DEBUG
378 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
381 * Protect against sched_clock() occasionally going backwards:
383 if (unlikely(delta < 0)) {
388 * Catch too large forward jumps too:
390 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
391 if (clock < rq->tick_timestamp + TICK_NSEC)
392 clock = rq->tick_timestamp + TICK_NSEC;
395 rq->clock_overflows++;
397 if (unlikely(delta > rq->clock_max_delta))
398 rq->clock_max_delta = delta;
403 rq->prev_clock_raw = now;
407 static void update_rq_clock(struct rq *rq)
409 if (likely(smp_processor_id() == cpu_of(rq)))
410 __update_rq_clock(rq);
414 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
415 * See detach_destroy_domains: synchronize_sched for details.
417 * The domain tree of any CPU may only be accessed from within
418 * preempt-disabled sections.
420 #define for_each_domain(cpu, __sd) \
421 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
423 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
424 #define this_rq() (&__get_cpu_var(runqueues))
425 #define task_rq(p) cpu_rq(task_cpu(p))
426 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
429 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
431 #ifdef CONFIG_SCHED_DEBUG
432 # define const_debug __read_mostly
434 # define const_debug static const
438 * Debugging: various feature bits
441 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
442 SCHED_FEAT_START_DEBIT = 2,
443 SCHED_FEAT_TREE_AVG = 4,
444 SCHED_FEAT_APPROX_AVG = 8,
447 const_debug unsigned int sysctl_sched_features =
448 SCHED_FEAT_NEW_FAIR_SLEEPERS *1 |
449 SCHED_FEAT_START_DEBIT *1 |
450 SCHED_FEAT_TREE_AVG *0 |
451 SCHED_FEAT_APPROX_AVG *0;
453 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
456 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
457 * clock constructed from sched_clock():
459 unsigned long long cpu_clock(int cpu)
461 unsigned long long now;
465 local_irq_save(flags);
469 local_irq_restore(flags);
473 EXPORT_SYMBOL_GPL(cpu_clock);
475 #ifndef prepare_arch_switch
476 # define prepare_arch_switch(next) do { } while (0)
478 #ifndef finish_arch_switch
479 # define finish_arch_switch(prev) do { } while (0)
482 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
483 static inline int task_running(struct rq *rq, struct task_struct *p)
485 return rq->curr == p;
488 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
492 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
494 #ifdef CONFIG_DEBUG_SPINLOCK
495 /* this is a valid case when another task releases the spinlock */
496 rq->lock.owner = current;
499 * If we are tracking spinlock dependencies then we have to
500 * fix up the runqueue lock - which gets 'carried over' from
503 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
505 spin_unlock_irq(&rq->lock);
508 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
509 static inline int task_running(struct rq *rq, struct task_struct *p)
514 return rq->curr == p;
518 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
522 * We can optimise this out completely for !SMP, because the
523 * SMP rebalancing from interrupt is the only thing that cares
528 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
529 spin_unlock_irq(&rq->lock);
531 spin_unlock(&rq->lock);
535 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
539 * After ->oncpu is cleared, the task can be moved to a different CPU.
540 * We must ensure this doesn't happen until the switch is completely
546 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
550 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
553 * __task_rq_lock - lock the runqueue a given task resides on.
554 * Must be called interrupts disabled.
556 static inline struct rq *__task_rq_lock(struct task_struct *p)
563 spin_lock(&rq->lock);
564 if (unlikely(rq != task_rq(p))) {
565 spin_unlock(&rq->lock);
566 goto repeat_lock_task;
572 * task_rq_lock - lock the runqueue a given task resides on and disable
573 * interrupts. Note the ordering: we can safely lookup the task_rq without
574 * explicitly disabling preemption.
576 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
582 local_irq_save(*flags);
584 spin_lock(&rq->lock);
585 if (unlikely(rq != task_rq(p))) {
586 spin_unlock_irqrestore(&rq->lock, *flags);
587 goto repeat_lock_task;
592 static void __task_rq_unlock(struct rq *rq)
595 spin_unlock(&rq->lock);
598 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
601 spin_unlock_irqrestore(&rq->lock, *flags);
605 * this_rq_lock - lock this runqueue and disable interrupts.
607 static struct rq *this_rq_lock(void)
614 spin_lock(&rq->lock);
620 * We are going deep-idle (irqs are disabled):
622 void sched_clock_idle_sleep_event(void)
624 struct rq *rq = cpu_rq(smp_processor_id());
626 spin_lock(&rq->lock);
627 __update_rq_clock(rq);
628 spin_unlock(&rq->lock);
629 rq->clock_deep_idle_events++;
631 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
634 * We just idled delta nanoseconds (called with irqs disabled):
636 void sched_clock_idle_wakeup_event(u64 delta_ns)
638 struct rq *rq = cpu_rq(smp_processor_id());
639 u64 now = sched_clock();
641 rq->idle_clock += delta_ns;
643 * Override the previous timestamp and ignore all
644 * sched_clock() deltas that occured while we idled,
645 * and use the PM-provided delta_ns to advance the
648 spin_lock(&rq->lock);
649 rq->prev_clock_raw = now;
650 rq->clock += delta_ns;
651 spin_unlock(&rq->lock);
653 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
656 * resched_task - mark a task 'to be rescheduled now'.
658 * On UP this means the setting of the need_resched flag, on SMP it
659 * might also involve a cross-CPU call to trigger the scheduler on
664 #ifndef tsk_is_polling
665 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
668 static void resched_task(struct task_struct *p)
672 assert_spin_locked(&task_rq(p)->lock);
674 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
677 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
680 if (cpu == smp_processor_id())
683 /* NEED_RESCHED must be visible before we test polling */
685 if (!tsk_is_polling(p))
686 smp_send_reschedule(cpu);
689 static void resched_cpu(int cpu)
691 struct rq *rq = cpu_rq(cpu);
694 if (!spin_trylock_irqsave(&rq->lock, flags))
696 resched_task(cpu_curr(cpu));
697 spin_unlock_irqrestore(&rq->lock, flags);
700 static inline void resched_task(struct task_struct *p)
702 assert_spin_locked(&task_rq(p)->lock);
703 set_tsk_need_resched(p);
707 #if BITS_PER_LONG == 32
708 # define WMULT_CONST (~0UL)
710 # define WMULT_CONST (1UL << 32)
713 #define WMULT_SHIFT 32
716 * Shift right and round:
718 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
721 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
722 struct load_weight *lw)
726 if (unlikely(!lw->inv_weight))
727 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
729 tmp = (u64)delta_exec * weight;
731 * Check whether we'd overflow the 64-bit multiplication:
733 if (unlikely(tmp > WMULT_CONST))
734 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
737 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
739 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
742 static inline unsigned long
743 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
745 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
748 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
753 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
759 * To aid in avoiding the subversion of "niceness" due to uneven distribution
760 * of tasks with abnormal "nice" values across CPUs the contribution that
761 * each task makes to its run queue's load is weighted according to its
762 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
763 * scaled version of the new time slice allocation that they receive on time
767 #define WEIGHT_IDLEPRIO 2
768 #define WMULT_IDLEPRIO (1 << 31)
771 * Nice levels are multiplicative, with a gentle 10% change for every
772 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
773 * nice 1, it will get ~10% less CPU time than another CPU-bound task
774 * that remained on nice 0.
776 * The "10% effect" is relative and cumulative: from _any_ nice level,
777 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
778 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
779 * If a task goes up by ~10% and another task goes down by ~10% then
780 * the relative distance between them is ~25%.)
782 static const int prio_to_weight[40] = {
783 /* -20 */ 88761, 71755, 56483, 46273, 36291,
784 /* -15 */ 29154, 23254, 18705, 14949, 11916,
785 /* -10 */ 9548, 7620, 6100, 4904, 3906,
786 /* -5 */ 3121, 2501, 1991, 1586, 1277,
787 /* 0 */ 1024, 820, 655, 526, 423,
788 /* 5 */ 335, 272, 215, 172, 137,
789 /* 10 */ 110, 87, 70, 56, 45,
790 /* 15 */ 36, 29, 23, 18, 15,
794 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
796 * In cases where the weight does not change often, we can use the
797 * precalculated inverse to speed up arithmetics by turning divisions
798 * into multiplications:
800 static const u32 prio_to_wmult[40] = {
801 /* -20 */ 48388, 59856, 76040, 92818, 118348,
802 /* -15 */ 147320, 184698, 229616, 287308, 360437,
803 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
804 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
805 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
806 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
807 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
808 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
811 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
814 * runqueue iterator, to support SMP load-balancing between different
815 * scheduling classes, without having to expose their internal data
816 * structures to the load-balancing proper:
820 struct task_struct *(*start)(void *);
821 struct task_struct *(*next)(void *);
824 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
825 unsigned long max_nr_move, unsigned long max_load_move,
826 struct sched_domain *sd, enum cpu_idle_type idle,
827 int *all_pinned, unsigned long *load_moved,
828 int *this_best_prio, struct rq_iterator *iterator);
830 #include "sched_stats.h"
831 #include "sched_idletask.c"
832 #include "sched_fair.c"
833 #include "sched_rt.c"
834 #ifdef CONFIG_SCHED_DEBUG
835 # include "sched_debug.c"
838 #define sched_class_highest (&rt_sched_class)
841 * Update delta_exec, delta_fair fields for rq.
843 * delta_fair clock advances at a rate inversely proportional to
844 * total load (rq->load.weight) on the runqueue, while
845 * delta_exec advances at the same rate as wall-clock (provided
848 * delta_exec / delta_fair is a measure of the (smoothened) load on this
849 * runqueue over any given interval. This (smoothened) load is used
850 * during load balance.
852 * This function is called /before/ updating rq->load
853 * and when switching tasks.
855 static inline void inc_load(struct rq *rq, const struct task_struct *p)
857 update_load_add(&rq->load, p->se.load.weight);
860 static inline void dec_load(struct rq *rq, const struct task_struct *p)
862 update_load_sub(&rq->load, p->se.load.weight);
865 static void inc_nr_running(struct task_struct *p, struct rq *rq)
871 static void dec_nr_running(struct task_struct *p, struct rq *rq)
877 static void set_load_weight(struct task_struct *p)
879 if (task_has_rt_policy(p)) {
880 p->se.load.weight = prio_to_weight[0] * 2;
881 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
886 * SCHED_IDLE tasks get minimal weight:
888 if (p->policy == SCHED_IDLE) {
889 p->se.load.weight = WEIGHT_IDLEPRIO;
890 p->se.load.inv_weight = WMULT_IDLEPRIO;
894 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
895 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
898 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
900 sched_info_queued(p);
901 p->sched_class->enqueue_task(rq, p, wakeup);
905 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
907 p->sched_class->dequeue_task(rq, p, sleep);
912 * __normal_prio - return the priority that is based on the static prio
914 static inline int __normal_prio(struct task_struct *p)
916 return p->static_prio;
920 * Calculate the expected normal priority: i.e. priority
921 * without taking RT-inheritance into account. Might be
922 * boosted by interactivity modifiers. Changes upon fork,
923 * setprio syscalls, and whenever the interactivity
924 * estimator recalculates.
926 static inline int normal_prio(struct task_struct *p)
930 if (task_has_rt_policy(p))
931 prio = MAX_RT_PRIO-1 - p->rt_priority;
933 prio = __normal_prio(p);
938 * Calculate the current priority, i.e. the priority
939 * taken into account by the scheduler. This value might
940 * be boosted by RT tasks, or might be boosted by
941 * interactivity modifiers. Will be RT if the task got
942 * RT-boosted. If not then it returns p->normal_prio.
944 static int effective_prio(struct task_struct *p)
946 p->normal_prio = normal_prio(p);
948 * If we are RT tasks or we were boosted to RT priority,
949 * keep the priority unchanged. Otherwise, update priority
950 * to the normal priority:
952 if (!rt_prio(p->prio))
953 return p->normal_prio;
958 * activate_task - move a task to the runqueue.
960 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
962 if (p->state == TASK_UNINTERRUPTIBLE)
963 rq->nr_uninterruptible--;
965 enqueue_task(rq, p, wakeup);
966 inc_nr_running(p, rq);
970 * deactivate_task - remove a task from the runqueue.
972 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
974 if (p->state == TASK_UNINTERRUPTIBLE)
975 rq->nr_uninterruptible++;
977 dequeue_task(rq, p, sleep);
978 dec_nr_running(p, rq);
982 * task_curr - is this task currently executing on a CPU?
983 * @p: the task in question.
985 inline int task_curr(const struct task_struct *p)
987 return cpu_curr(task_cpu(p)) == p;
990 /* Used instead of source_load when we know the type == 0 */
991 unsigned long weighted_cpuload(const int cpu)
993 return cpu_rq(cpu)->load.weight;
996 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
999 task_thread_info(p)->cpu = cpu;
1006 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1008 int old_cpu = task_cpu(p);
1009 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1010 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1011 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1014 clock_offset = old_rq->clock - new_rq->clock;
1016 #ifdef CONFIG_SCHEDSTATS
1017 if (p->se.wait_start)
1018 p->se.wait_start -= clock_offset;
1019 if (p->se.sleep_start)
1020 p->se.sleep_start -= clock_offset;
1021 if (p->se.block_start)
1022 p->se.block_start -= clock_offset;
1024 p->se.vruntime -= old_cfsrq->min_vruntime -
1025 new_cfsrq->min_vruntime;
1027 __set_task_cpu(p, new_cpu);
1030 struct migration_req {
1031 struct list_head list;
1033 struct task_struct *task;
1036 struct completion done;
1040 * The task's runqueue lock must be held.
1041 * Returns true if you have to wait for migration thread.
1044 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1046 struct rq *rq = task_rq(p);
1049 * If the task is not on a runqueue (and not running), then
1050 * it is sufficient to simply update the task's cpu field.
1052 if (!p->se.on_rq && !task_running(rq, p)) {
1053 set_task_cpu(p, dest_cpu);
1057 init_completion(&req->done);
1059 req->dest_cpu = dest_cpu;
1060 list_add(&req->list, &rq->migration_queue);
1066 * wait_task_inactive - wait for a thread to unschedule.
1068 * The caller must ensure that the task *will* unschedule sometime soon,
1069 * else this function might spin for a *long* time. This function can't
1070 * be called with interrupts off, or it may introduce deadlock with
1071 * smp_call_function() if an IPI is sent by the same process we are
1072 * waiting to become inactive.
1074 void wait_task_inactive(struct task_struct *p)
1076 unsigned long flags;
1082 * We do the initial early heuristics without holding
1083 * any task-queue locks at all. We'll only try to get
1084 * the runqueue lock when things look like they will
1090 * If the task is actively running on another CPU
1091 * still, just relax and busy-wait without holding
1094 * NOTE! Since we don't hold any locks, it's not
1095 * even sure that "rq" stays as the right runqueue!
1096 * But we don't care, since "task_running()" will
1097 * return false if the runqueue has changed and p
1098 * is actually now running somewhere else!
1100 while (task_running(rq, p))
1104 * Ok, time to look more closely! We need the rq
1105 * lock now, to be *sure*. If we're wrong, we'll
1106 * just go back and repeat.
1108 rq = task_rq_lock(p, &flags);
1109 running = task_running(rq, p);
1110 on_rq = p->se.on_rq;
1111 task_rq_unlock(rq, &flags);
1114 * Was it really running after all now that we
1115 * checked with the proper locks actually held?
1117 * Oops. Go back and try again..
1119 if (unlikely(running)) {
1125 * It's not enough that it's not actively running,
1126 * it must be off the runqueue _entirely_, and not
1129 * So if it wa still runnable (but just not actively
1130 * running right now), it's preempted, and we should
1131 * yield - it could be a while.
1133 if (unlikely(on_rq)) {
1139 * Ahh, all good. It wasn't running, and it wasn't
1140 * runnable, which means that it will never become
1141 * running in the future either. We're all done!
1146 * kick_process - kick a running thread to enter/exit the kernel
1147 * @p: the to-be-kicked thread
1149 * Cause a process which is running on another CPU to enter
1150 * kernel-mode, without any delay. (to get signals handled.)
1152 * NOTE: this function doesnt have to take the runqueue lock,
1153 * because all it wants to ensure is that the remote task enters
1154 * the kernel. If the IPI races and the task has been migrated
1155 * to another CPU then no harm is done and the purpose has been
1158 void kick_process(struct task_struct *p)
1164 if ((cpu != smp_processor_id()) && task_curr(p))
1165 smp_send_reschedule(cpu);
1170 * Return a low guess at the load of a migration-source cpu weighted
1171 * according to the scheduling class and "nice" value.
1173 * We want to under-estimate the load of migration sources, to
1174 * balance conservatively.
1176 static unsigned long source_load(int cpu, int type)
1178 struct rq *rq = cpu_rq(cpu);
1179 unsigned long total = weighted_cpuload(cpu);
1184 return min(rq->cpu_load[type-1], total);
1188 * Return a high guess at the load of a migration-target cpu weighted
1189 * according to the scheduling class and "nice" value.
1191 static unsigned long target_load(int cpu, int type)
1193 struct rq *rq = cpu_rq(cpu);
1194 unsigned long total = weighted_cpuload(cpu);
1199 return max(rq->cpu_load[type-1], total);
1203 * Return the average load per task on the cpu's run queue
1205 static inline unsigned long cpu_avg_load_per_task(int cpu)
1207 struct rq *rq = cpu_rq(cpu);
1208 unsigned long total = weighted_cpuload(cpu);
1209 unsigned long n = rq->nr_running;
1211 return n ? total / n : SCHED_LOAD_SCALE;
1215 * find_idlest_group finds and returns the least busy CPU group within the
1218 static struct sched_group *
1219 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1221 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1222 unsigned long min_load = ULONG_MAX, this_load = 0;
1223 int load_idx = sd->forkexec_idx;
1224 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1227 unsigned long load, avg_load;
1231 /* Skip over this group if it has no CPUs allowed */
1232 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1235 local_group = cpu_isset(this_cpu, group->cpumask);
1237 /* Tally up the load of all CPUs in the group */
1240 for_each_cpu_mask(i, group->cpumask) {
1241 /* Bias balancing toward cpus of our domain */
1243 load = source_load(i, load_idx);
1245 load = target_load(i, load_idx);
1250 /* Adjust by relative CPU power of the group */
1251 avg_load = sg_div_cpu_power(group,
1252 avg_load * SCHED_LOAD_SCALE);
1255 this_load = avg_load;
1257 } else if (avg_load < min_load) {
1258 min_load = avg_load;
1262 group = group->next;
1263 } while (group != sd->groups);
1265 if (!idlest || 100*this_load < imbalance*min_load)
1271 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1274 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1277 unsigned long load, min_load = ULONG_MAX;
1281 /* Traverse only the allowed CPUs */
1282 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1284 for_each_cpu_mask(i, tmp) {
1285 load = weighted_cpuload(i);
1287 if (load < min_load || (load == min_load && i == this_cpu)) {
1297 * sched_balance_self: balance the current task (running on cpu) in domains
1298 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1301 * Balance, ie. select the least loaded group.
1303 * Returns the target CPU number, or the same CPU if no balancing is needed.
1305 * preempt must be disabled.
1307 static int sched_balance_self(int cpu, int flag)
1309 struct task_struct *t = current;
1310 struct sched_domain *tmp, *sd = NULL;
1312 for_each_domain(cpu, tmp) {
1314 * If power savings logic is enabled for a domain, stop there.
1316 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1318 if (tmp->flags & flag)
1324 struct sched_group *group;
1325 int new_cpu, weight;
1327 if (!(sd->flags & flag)) {
1333 group = find_idlest_group(sd, t, cpu);
1339 new_cpu = find_idlest_cpu(group, t, cpu);
1340 if (new_cpu == -1 || new_cpu == cpu) {
1341 /* Now try balancing at a lower domain level of cpu */
1346 /* Now try balancing at a lower domain level of new_cpu */
1349 weight = cpus_weight(span);
1350 for_each_domain(cpu, tmp) {
1351 if (weight <= cpus_weight(tmp->span))
1353 if (tmp->flags & flag)
1356 /* while loop will break here if sd == NULL */
1362 #endif /* CONFIG_SMP */
1365 * wake_idle() will wake a task on an idle cpu if task->cpu is
1366 * not idle and an idle cpu is available. The span of cpus to
1367 * search starts with cpus closest then further out as needed,
1368 * so we always favor a closer, idle cpu.
1370 * Returns the CPU we should wake onto.
1372 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1373 static int wake_idle(int cpu, struct task_struct *p)
1376 struct sched_domain *sd;
1380 * If it is idle, then it is the best cpu to run this task.
1382 * This cpu is also the best, if it has more than one task already.
1383 * Siblings must be also busy(in most cases) as they didn't already
1384 * pickup the extra load from this cpu and hence we need not check
1385 * sibling runqueue info. This will avoid the checks and cache miss
1386 * penalities associated with that.
1388 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1391 for_each_domain(cpu, sd) {
1392 if (sd->flags & SD_WAKE_IDLE) {
1393 cpus_and(tmp, sd->span, p->cpus_allowed);
1394 for_each_cpu_mask(i, tmp) {
1405 static inline int wake_idle(int cpu, struct task_struct *p)
1412 * try_to_wake_up - wake up a thread
1413 * @p: the to-be-woken-up thread
1414 * @state: the mask of task states that can be woken
1415 * @sync: do a synchronous wakeup?
1417 * Put it on the run-queue if it's not already there. The "current"
1418 * thread is always on the run-queue (except when the actual
1419 * re-schedule is in progress), and as such you're allowed to do
1420 * the simpler "current->state = TASK_RUNNING" to mark yourself
1421 * runnable without the overhead of this.
1423 * returns failure only if the task is already active.
1425 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1427 int cpu, this_cpu, success = 0;
1428 unsigned long flags;
1432 struct sched_domain *sd, *this_sd = NULL;
1433 unsigned long load, this_load;
1437 rq = task_rq_lock(p, &flags);
1438 old_state = p->state;
1439 if (!(old_state & state))
1446 this_cpu = smp_processor_id();
1449 if (unlikely(task_running(rq, p)))
1454 schedstat_inc(rq, ttwu_count);
1455 if (cpu == this_cpu) {
1456 schedstat_inc(rq, ttwu_local);
1460 for_each_domain(this_cpu, sd) {
1461 if (cpu_isset(cpu, sd->span)) {
1462 schedstat_inc(sd, ttwu_wake_remote);
1468 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1472 * Check for affine wakeup and passive balancing possibilities.
1475 int idx = this_sd->wake_idx;
1476 unsigned int imbalance;
1478 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1480 load = source_load(cpu, idx);
1481 this_load = target_load(this_cpu, idx);
1483 new_cpu = this_cpu; /* Wake to this CPU if we can */
1485 if (this_sd->flags & SD_WAKE_AFFINE) {
1486 unsigned long tl = this_load;
1487 unsigned long tl_per_task;
1489 tl_per_task = cpu_avg_load_per_task(this_cpu);
1492 * If sync wakeup then subtract the (maximum possible)
1493 * effect of the currently running task from the load
1494 * of the current CPU:
1497 tl -= current->se.load.weight;
1500 tl + target_load(cpu, idx) <= tl_per_task) ||
1501 100*(tl + p->se.load.weight) <= imbalance*load) {
1503 * This domain has SD_WAKE_AFFINE and
1504 * p is cache cold in this domain, and
1505 * there is no bad imbalance.
1507 schedstat_inc(this_sd, ttwu_move_affine);
1513 * Start passive balancing when half the imbalance_pct
1516 if (this_sd->flags & SD_WAKE_BALANCE) {
1517 if (imbalance*this_load <= 100*load) {
1518 schedstat_inc(this_sd, ttwu_move_balance);
1524 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1526 new_cpu = wake_idle(new_cpu, p);
1527 if (new_cpu != cpu) {
1528 set_task_cpu(p, new_cpu);
1529 task_rq_unlock(rq, &flags);
1530 /* might preempt at this point */
1531 rq = task_rq_lock(p, &flags);
1532 old_state = p->state;
1533 if (!(old_state & state))
1538 this_cpu = smp_processor_id();
1543 #endif /* CONFIG_SMP */
1544 update_rq_clock(rq);
1545 activate_task(rq, p, 1);
1547 * Sync wakeups (i.e. those types of wakeups where the waker
1548 * has indicated that it will leave the CPU in short order)
1549 * don't trigger a preemption, if the woken up task will run on
1550 * this cpu. (in this case the 'I will reschedule' promise of
1551 * the waker guarantees that the freshly woken up task is going
1552 * to be considered on this CPU.)
1554 if (!sync || cpu != this_cpu)
1555 check_preempt_curr(rq, p);
1559 p->state = TASK_RUNNING;
1561 task_rq_unlock(rq, &flags);
1566 int fastcall wake_up_process(struct task_struct *p)
1568 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1569 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1571 EXPORT_SYMBOL(wake_up_process);
1573 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1575 return try_to_wake_up(p, state, 0);
1579 * Perform scheduler related setup for a newly forked process p.
1580 * p is forked by current.
1582 * __sched_fork() is basic setup used by init_idle() too:
1584 static void __sched_fork(struct task_struct *p)
1586 p->se.exec_start = 0;
1587 p->se.sum_exec_runtime = 0;
1588 p->se.prev_sum_exec_runtime = 0;
1590 #ifdef CONFIG_SCHEDSTATS
1591 p->se.wait_start = 0;
1592 p->se.sum_sleep_runtime = 0;
1593 p->se.sleep_start = 0;
1594 p->se.block_start = 0;
1595 p->se.sleep_max = 0;
1596 p->se.block_max = 0;
1598 p->se.slice_max = 0;
1602 INIT_LIST_HEAD(&p->run_list);
1605 #ifdef CONFIG_PREEMPT_NOTIFIERS
1606 INIT_HLIST_HEAD(&p->preempt_notifiers);
1610 * We mark the process as running here, but have not actually
1611 * inserted it onto the runqueue yet. This guarantees that
1612 * nobody will actually run it, and a signal or other external
1613 * event cannot wake it up and insert it on the runqueue either.
1615 p->state = TASK_RUNNING;
1619 * fork()/clone()-time setup:
1621 void sched_fork(struct task_struct *p, int clone_flags)
1623 int cpu = get_cpu();
1628 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1630 set_task_cpu(p, cpu);
1633 * Make sure we do not leak PI boosting priority to the child:
1635 p->prio = current->normal_prio;
1636 if (!rt_prio(p->prio))
1637 p->sched_class = &fair_sched_class;
1639 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1640 if (likely(sched_info_on()))
1641 memset(&p->sched_info, 0, sizeof(p->sched_info));
1643 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1646 #ifdef CONFIG_PREEMPT
1647 /* Want to start with kernel preemption disabled. */
1648 task_thread_info(p)->preempt_count = 1;
1654 * wake_up_new_task - wake up a newly created task for the first time.
1656 * This function will do some initial scheduler statistics housekeeping
1657 * that must be done for every newly created context, then puts the task
1658 * on the runqueue and wakes it.
1660 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1662 unsigned long flags;
1665 rq = task_rq_lock(p, &flags);
1666 BUG_ON(p->state != TASK_RUNNING);
1667 update_rq_clock(rq);
1669 p->prio = effective_prio(p);
1671 if (!p->sched_class->task_new || !current->se.on_rq || !rq->cfs.curr) {
1672 activate_task(rq, p, 0);
1675 * Let the scheduling class do new task startup
1676 * management (if any):
1678 p->sched_class->task_new(rq, p);
1679 inc_nr_running(p, rq);
1681 check_preempt_curr(rq, p);
1682 task_rq_unlock(rq, &flags);
1685 #ifdef CONFIG_PREEMPT_NOTIFIERS
1688 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1689 * @notifier: notifier struct to register
1691 void preempt_notifier_register(struct preempt_notifier *notifier)
1693 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1695 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1698 * preempt_notifier_unregister - no longer interested in preemption notifications
1699 * @notifier: notifier struct to unregister
1701 * This is safe to call from within a preemption notifier.
1703 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1705 hlist_del(¬ifier->link);
1707 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1709 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1711 struct preempt_notifier *notifier;
1712 struct hlist_node *node;
1714 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1715 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1719 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1720 struct task_struct *next)
1722 struct preempt_notifier *notifier;
1723 struct hlist_node *node;
1725 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1726 notifier->ops->sched_out(notifier, next);
1731 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1736 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1737 struct task_struct *next)
1744 * prepare_task_switch - prepare to switch tasks
1745 * @rq: the runqueue preparing to switch
1746 * @prev: the current task that is being switched out
1747 * @next: the task we are going to switch to.
1749 * This is called with the rq lock held and interrupts off. It must
1750 * be paired with a subsequent finish_task_switch after the context
1753 * prepare_task_switch sets up locking and calls architecture specific
1757 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1758 struct task_struct *next)
1760 fire_sched_out_preempt_notifiers(prev, next);
1761 prepare_lock_switch(rq, next);
1762 prepare_arch_switch(next);
1766 * finish_task_switch - clean up after a task-switch
1767 * @rq: runqueue associated with task-switch
1768 * @prev: the thread we just switched away from.
1770 * finish_task_switch must be called after the context switch, paired
1771 * with a prepare_task_switch call before the context switch.
1772 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1773 * and do any other architecture-specific cleanup actions.
1775 * Note that we may have delayed dropping an mm in context_switch(). If
1776 * so, we finish that here outside of the runqueue lock. (Doing it
1777 * with the lock held can cause deadlocks; see schedule() for
1780 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1781 __releases(rq->lock)
1783 struct mm_struct *mm = rq->prev_mm;
1789 * A task struct has one reference for the use as "current".
1790 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1791 * schedule one last time. The schedule call will never return, and
1792 * the scheduled task must drop that reference.
1793 * The test for TASK_DEAD must occur while the runqueue locks are
1794 * still held, otherwise prev could be scheduled on another cpu, die
1795 * there before we look at prev->state, and then the reference would
1797 * Manfred Spraul <manfred@colorfullife.com>
1799 prev_state = prev->state;
1800 finish_arch_switch(prev);
1801 finish_lock_switch(rq, prev);
1802 fire_sched_in_preempt_notifiers(current);
1805 if (unlikely(prev_state == TASK_DEAD)) {
1807 * Remove function-return probe instances associated with this
1808 * task and put them back on the free list.
1810 kprobe_flush_task(prev);
1811 put_task_struct(prev);
1816 * schedule_tail - first thing a freshly forked thread must call.
1817 * @prev: the thread we just switched away from.
1819 asmlinkage void schedule_tail(struct task_struct *prev)
1820 __releases(rq->lock)
1822 struct rq *rq = this_rq();
1824 finish_task_switch(rq, prev);
1825 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1826 /* In this case, finish_task_switch does not reenable preemption */
1829 if (current->set_child_tid)
1830 put_user(current->pid, current->set_child_tid);
1834 * context_switch - switch to the new MM and the new
1835 * thread's register state.
1838 context_switch(struct rq *rq, struct task_struct *prev,
1839 struct task_struct *next)
1841 struct mm_struct *mm, *oldmm;
1843 prepare_task_switch(rq, prev, next);
1845 oldmm = prev->active_mm;
1847 * For paravirt, this is coupled with an exit in switch_to to
1848 * combine the page table reload and the switch backend into
1851 arch_enter_lazy_cpu_mode();
1853 if (unlikely(!mm)) {
1854 next->active_mm = oldmm;
1855 atomic_inc(&oldmm->mm_count);
1856 enter_lazy_tlb(oldmm, next);
1858 switch_mm(oldmm, mm, next);
1860 if (unlikely(!prev->mm)) {
1861 prev->active_mm = NULL;
1862 rq->prev_mm = oldmm;
1865 * Since the runqueue lock will be released by the next
1866 * task (which is an invalid locking op but in the case
1867 * of the scheduler it's an obvious special-case), so we
1868 * do an early lockdep release here:
1870 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1871 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1874 /* Here we just switch the register state and the stack. */
1875 switch_to(prev, next, prev);
1879 * this_rq must be evaluated again because prev may have moved
1880 * CPUs since it called schedule(), thus the 'rq' on its stack
1881 * frame will be invalid.
1883 finish_task_switch(this_rq(), prev);
1887 * nr_running, nr_uninterruptible and nr_context_switches:
1889 * externally visible scheduler statistics: current number of runnable
1890 * threads, current number of uninterruptible-sleeping threads, total
1891 * number of context switches performed since bootup.
1893 unsigned long nr_running(void)
1895 unsigned long i, sum = 0;
1897 for_each_online_cpu(i)
1898 sum += cpu_rq(i)->nr_running;
1903 unsigned long nr_uninterruptible(void)
1905 unsigned long i, sum = 0;
1907 for_each_possible_cpu(i)
1908 sum += cpu_rq(i)->nr_uninterruptible;
1911 * Since we read the counters lockless, it might be slightly
1912 * inaccurate. Do not allow it to go below zero though:
1914 if (unlikely((long)sum < 0))
1920 unsigned long long nr_context_switches(void)
1923 unsigned long long sum = 0;
1925 for_each_possible_cpu(i)
1926 sum += cpu_rq(i)->nr_switches;
1931 unsigned long nr_iowait(void)
1933 unsigned long i, sum = 0;
1935 for_each_possible_cpu(i)
1936 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1941 unsigned long nr_active(void)
1943 unsigned long i, running = 0, uninterruptible = 0;
1945 for_each_online_cpu(i) {
1946 running += cpu_rq(i)->nr_running;
1947 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1950 if (unlikely((long)uninterruptible < 0))
1951 uninterruptible = 0;
1953 return running + uninterruptible;
1957 * Update rq->cpu_load[] statistics. This function is usually called every
1958 * scheduler tick (TICK_NSEC).
1960 static void update_cpu_load(struct rq *this_rq)
1962 unsigned long this_load = this_rq->load.weight;
1965 this_rq->nr_load_updates++;
1967 /* Update our load: */
1968 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1969 unsigned long old_load, new_load;
1971 /* scale is effectively 1 << i now, and >> i divides by scale */
1973 old_load = this_rq->cpu_load[i];
1974 new_load = this_load;
1976 * Round up the averaging division if load is increasing. This
1977 * prevents us from getting stuck on 9 if the load is 10, for
1980 if (new_load > old_load)
1981 new_load += scale-1;
1982 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
1989 * double_rq_lock - safely lock two runqueues
1991 * Note this does not disable interrupts like task_rq_lock,
1992 * you need to do so manually before calling.
1994 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1995 __acquires(rq1->lock)
1996 __acquires(rq2->lock)
1998 BUG_ON(!irqs_disabled());
2000 spin_lock(&rq1->lock);
2001 __acquire(rq2->lock); /* Fake it out ;) */
2004 spin_lock(&rq1->lock);
2005 spin_lock(&rq2->lock);
2007 spin_lock(&rq2->lock);
2008 spin_lock(&rq1->lock);
2011 update_rq_clock(rq1);
2012 update_rq_clock(rq2);
2016 * double_rq_unlock - safely unlock two runqueues
2018 * Note this does not restore interrupts like task_rq_unlock,
2019 * you need to do so manually after calling.
2021 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2022 __releases(rq1->lock)
2023 __releases(rq2->lock)
2025 spin_unlock(&rq1->lock);
2027 spin_unlock(&rq2->lock);
2029 __release(rq2->lock);
2033 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2035 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2036 __releases(this_rq->lock)
2037 __acquires(busiest->lock)
2038 __acquires(this_rq->lock)
2040 if (unlikely(!irqs_disabled())) {
2041 /* printk() doesn't work good under rq->lock */
2042 spin_unlock(&this_rq->lock);
2045 if (unlikely(!spin_trylock(&busiest->lock))) {
2046 if (busiest < this_rq) {
2047 spin_unlock(&this_rq->lock);
2048 spin_lock(&busiest->lock);
2049 spin_lock(&this_rq->lock);
2051 spin_lock(&busiest->lock);
2056 * If dest_cpu is allowed for this process, migrate the task to it.
2057 * This is accomplished by forcing the cpu_allowed mask to only
2058 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2059 * the cpu_allowed mask is restored.
2061 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2063 struct migration_req req;
2064 unsigned long flags;
2067 rq = task_rq_lock(p, &flags);
2068 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2069 || unlikely(cpu_is_offline(dest_cpu)))
2072 /* force the process onto the specified CPU */
2073 if (migrate_task(p, dest_cpu, &req)) {
2074 /* Need to wait for migration thread (might exit: take ref). */
2075 struct task_struct *mt = rq->migration_thread;
2077 get_task_struct(mt);
2078 task_rq_unlock(rq, &flags);
2079 wake_up_process(mt);
2080 put_task_struct(mt);
2081 wait_for_completion(&req.done);
2086 task_rq_unlock(rq, &flags);
2090 * sched_exec - execve() is a valuable balancing opportunity, because at
2091 * this point the task has the smallest effective memory and cache footprint.
2093 void sched_exec(void)
2095 int new_cpu, this_cpu = get_cpu();
2096 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2098 if (new_cpu != this_cpu)
2099 sched_migrate_task(current, new_cpu);
2103 * pull_task - move a task from a remote runqueue to the local runqueue.
2104 * Both runqueues must be locked.
2106 static void pull_task(struct rq *src_rq, struct task_struct *p,
2107 struct rq *this_rq, int this_cpu)
2109 deactivate_task(src_rq, p, 0);
2110 set_task_cpu(p, this_cpu);
2111 activate_task(this_rq, p, 0);
2113 * Note that idle threads have a prio of MAX_PRIO, for this test
2114 * to be always true for them.
2116 check_preempt_curr(this_rq, p);
2120 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2123 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2124 struct sched_domain *sd, enum cpu_idle_type idle,
2128 * We do not migrate tasks that are:
2129 * 1) running (obviously), or
2130 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2131 * 3) are cache-hot on their current CPU.
2133 if (!cpu_isset(this_cpu, p->cpus_allowed))
2137 if (task_running(rq, p))
2143 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2144 unsigned long max_nr_move, unsigned long max_load_move,
2145 struct sched_domain *sd, enum cpu_idle_type idle,
2146 int *all_pinned, unsigned long *load_moved,
2147 int *this_best_prio, struct rq_iterator *iterator)
2149 int pulled = 0, pinned = 0, skip_for_load;
2150 struct task_struct *p;
2151 long rem_load_move = max_load_move;
2153 if (max_nr_move == 0 || max_load_move == 0)
2159 * Start the load-balancing iterator:
2161 p = iterator->start(iterator->arg);
2166 * To help distribute high priority tasks accross CPUs we don't
2167 * skip a task if it will be the highest priority task (i.e. smallest
2168 * prio value) on its new queue regardless of its load weight
2170 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2171 SCHED_LOAD_SCALE_FUZZ;
2172 if ((skip_for_load && p->prio >= *this_best_prio) ||
2173 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2174 p = iterator->next(iterator->arg);
2178 pull_task(busiest, p, this_rq, this_cpu);
2180 rem_load_move -= p->se.load.weight;
2183 * We only want to steal up to the prescribed number of tasks
2184 * and the prescribed amount of weighted load.
2186 if (pulled < max_nr_move && rem_load_move > 0) {
2187 if (p->prio < *this_best_prio)
2188 *this_best_prio = p->prio;
2189 p = iterator->next(iterator->arg);
2194 * Right now, this is the only place pull_task() is called,
2195 * so we can safely collect pull_task() stats here rather than
2196 * inside pull_task().
2198 schedstat_add(sd, lb_gained[idle], pulled);
2201 *all_pinned = pinned;
2202 *load_moved = max_load_move - rem_load_move;
2207 * move_tasks tries to move up to max_load_move weighted load from busiest to
2208 * this_rq, as part of a balancing operation within domain "sd".
2209 * Returns 1 if successful and 0 otherwise.
2211 * Called with both runqueues locked.
2213 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2214 unsigned long max_load_move,
2215 struct sched_domain *sd, enum cpu_idle_type idle,
2218 const struct sched_class *class = sched_class_highest;
2219 unsigned long total_load_moved = 0;
2220 int this_best_prio = this_rq->curr->prio;
2224 class->load_balance(this_rq, this_cpu, busiest,
2225 ULONG_MAX, max_load_move - total_load_moved,
2226 sd, idle, all_pinned, &this_best_prio);
2227 class = class->next;
2228 } while (class && max_load_move > total_load_moved);
2230 return total_load_moved > 0;
2234 * move_one_task tries to move exactly one task from busiest to this_rq, as
2235 * part of active balancing operations within "domain".
2236 * Returns 1 if successful and 0 otherwise.
2238 * Called with both runqueues locked.
2240 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2241 struct sched_domain *sd, enum cpu_idle_type idle)
2243 const struct sched_class *class;
2244 int this_best_prio = MAX_PRIO;
2246 for (class = sched_class_highest; class; class = class->next)
2247 if (class->load_balance(this_rq, this_cpu, busiest,
2248 1, ULONG_MAX, sd, idle, NULL,
2256 * find_busiest_group finds and returns the busiest CPU group within the
2257 * domain. It calculates and returns the amount of weighted load which
2258 * should be moved to restore balance via the imbalance parameter.
2260 static struct sched_group *
2261 find_busiest_group(struct sched_domain *sd, int this_cpu,
2262 unsigned long *imbalance, enum cpu_idle_type idle,
2263 int *sd_idle, cpumask_t *cpus, int *balance)
2265 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2266 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2267 unsigned long max_pull;
2268 unsigned long busiest_load_per_task, busiest_nr_running;
2269 unsigned long this_load_per_task, this_nr_running;
2271 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2272 int power_savings_balance = 1;
2273 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2274 unsigned long min_nr_running = ULONG_MAX;
2275 struct sched_group *group_min = NULL, *group_leader = NULL;
2278 max_load = this_load = total_load = total_pwr = 0;
2279 busiest_load_per_task = busiest_nr_running = 0;
2280 this_load_per_task = this_nr_running = 0;
2281 if (idle == CPU_NOT_IDLE)
2282 load_idx = sd->busy_idx;
2283 else if (idle == CPU_NEWLY_IDLE)
2284 load_idx = sd->newidle_idx;
2286 load_idx = sd->idle_idx;
2289 unsigned long load, group_capacity;
2292 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2293 unsigned long sum_nr_running, sum_weighted_load;
2295 local_group = cpu_isset(this_cpu, group->cpumask);
2298 balance_cpu = first_cpu(group->cpumask);
2300 /* Tally up the load of all CPUs in the group */
2301 sum_weighted_load = sum_nr_running = avg_load = 0;
2303 for_each_cpu_mask(i, group->cpumask) {
2306 if (!cpu_isset(i, *cpus))
2311 if (*sd_idle && rq->nr_running)
2314 /* Bias balancing toward cpus of our domain */
2316 if (idle_cpu(i) && !first_idle_cpu) {
2321 load = target_load(i, load_idx);
2323 load = source_load(i, load_idx);
2326 sum_nr_running += rq->nr_running;
2327 sum_weighted_load += weighted_cpuload(i);
2331 * First idle cpu or the first cpu(busiest) in this sched group
2332 * is eligible for doing load balancing at this and above
2333 * domains. In the newly idle case, we will allow all the cpu's
2334 * to do the newly idle load balance.
2336 if (idle != CPU_NEWLY_IDLE && local_group &&
2337 balance_cpu != this_cpu && balance) {
2342 total_load += avg_load;
2343 total_pwr += group->__cpu_power;
2345 /* Adjust by relative CPU power of the group */
2346 avg_load = sg_div_cpu_power(group,
2347 avg_load * SCHED_LOAD_SCALE);
2349 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2352 this_load = avg_load;
2354 this_nr_running = sum_nr_running;
2355 this_load_per_task = sum_weighted_load;
2356 } else if (avg_load > max_load &&
2357 sum_nr_running > group_capacity) {
2358 max_load = avg_load;
2360 busiest_nr_running = sum_nr_running;
2361 busiest_load_per_task = sum_weighted_load;
2364 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2366 * Busy processors will not participate in power savings
2369 if (idle == CPU_NOT_IDLE ||
2370 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2374 * If the local group is idle or completely loaded
2375 * no need to do power savings balance at this domain
2377 if (local_group && (this_nr_running >= group_capacity ||
2379 power_savings_balance = 0;
2382 * If a group is already running at full capacity or idle,
2383 * don't include that group in power savings calculations
2385 if (!power_savings_balance || sum_nr_running >= group_capacity
2390 * Calculate the group which has the least non-idle load.
2391 * This is the group from where we need to pick up the load
2394 if ((sum_nr_running < min_nr_running) ||
2395 (sum_nr_running == min_nr_running &&
2396 first_cpu(group->cpumask) <
2397 first_cpu(group_min->cpumask))) {
2399 min_nr_running = sum_nr_running;
2400 min_load_per_task = sum_weighted_load /
2405 * Calculate the group which is almost near its
2406 * capacity but still has some space to pick up some load
2407 * from other group and save more power
2409 if (sum_nr_running <= group_capacity - 1) {
2410 if (sum_nr_running > leader_nr_running ||
2411 (sum_nr_running == leader_nr_running &&
2412 first_cpu(group->cpumask) >
2413 first_cpu(group_leader->cpumask))) {
2414 group_leader = group;
2415 leader_nr_running = sum_nr_running;
2420 group = group->next;
2421 } while (group != sd->groups);
2423 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2426 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2428 if (this_load >= avg_load ||
2429 100*max_load <= sd->imbalance_pct*this_load)
2432 busiest_load_per_task /= busiest_nr_running;
2434 * We're trying to get all the cpus to the average_load, so we don't
2435 * want to push ourselves above the average load, nor do we wish to
2436 * reduce the max loaded cpu below the average load, as either of these
2437 * actions would just result in more rebalancing later, and ping-pong
2438 * tasks around. Thus we look for the minimum possible imbalance.
2439 * Negative imbalances (*we* are more loaded than anyone else) will
2440 * be counted as no imbalance for these purposes -- we can't fix that
2441 * by pulling tasks to us. Be careful of negative numbers as they'll
2442 * appear as very large values with unsigned longs.
2444 if (max_load <= busiest_load_per_task)
2448 * In the presence of smp nice balancing, certain scenarios can have
2449 * max load less than avg load(as we skip the groups at or below
2450 * its cpu_power, while calculating max_load..)
2452 if (max_load < avg_load) {
2454 goto small_imbalance;
2457 /* Don't want to pull so many tasks that a group would go idle */
2458 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2460 /* How much load to actually move to equalise the imbalance */
2461 *imbalance = min(max_pull * busiest->__cpu_power,
2462 (avg_load - this_load) * this->__cpu_power)
2466 * if *imbalance is less than the average load per runnable task
2467 * there is no gaurantee that any tasks will be moved so we'll have
2468 * a think about bumping its value to force at least one task to be
2471 if (*imbalance < busiest_load_per_task) {
2472 unsigned long tmp, pwr_now, pwr_move;
2476 pwr_move = pwr_now = 0;
2478 if (this_nr_running) {
2479 this_load_per_task /= this_nr_running;
2480 if (busiest_load_per_task > this_load_per_task)
2483 this_load_per_task = SCHED_LOAD_SCALE;
2485 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2486 busiest_load_per_task * imbn) {
2487 *imbalance = busiest_load_per_task;
2492 * OK, we don't have enough imbalance to justify moving tasks,
2493 * however we may be able to increase total CPU power used by
2497 pwr_now += busiest->__cpu_power *
2498 min(busiest_load_per_task, max_load);
2499 pwr_now += this->__cpu_power *
2500 min(this_load_per_task, this_load);
2501 pwr_now /= SCHED_LOAD_SCALE;
2503 /* Amount of load we'd subtract */
2504 tmp = sg_div_cpu_power(busiest,
2505 busiest_load_per_task * SCHED_LOAD_SCALE);
2507 pwr_move += busiest->__cpu_power *
2508 min(busiest_load_per_task, max_load - tmp);
2510 /* Amount of load we'd add */
2511 if (max_load * busiest->__cpu_power <
2512 busiest_load_per_task * SCHED_LOAD_SCALE)
2513 tmp = sg_div_cpu_power(this,
2514 max_load * busiest->__cpu_power);
2516 tmp = sg_div_cpu_power(this,
2517 busiest_load_per_task * SCHED_LOAD_SCALE);
2518 pwr_move += this->__cpu_power *
2519 min(this_load_per_task, this_load + tmp);
2520 pwr_move /= SCHED_LOAD_SCALE;
2522 /* Move if we gain throughput */
2523 if (pwr_move > pwr_now)
2524 *imbalance = busiest_load_per_task;
2530 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2531 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2534 if (this == group_leader && group_leader != group_min) {
2535 *imbalance = min_load_per_task;
2545 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2548 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2549 unsigned long imbalance, cpumask_t *cpus)
2551 struct rq *busiest = NULL, *rq;
2552 unsigned long max_load = 0;
2555 for_each_cpu_mask(i, group->cpumask) {
2558 if (!cpu_isset(i, *cpus))
2562 wl = weighted_cpuload(i);
2564 if (rq->nr_running == 1 && wl > imbalance)
2567 if (wl > max_load) {
2577 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2578 * so long as it is large enough.
2580 #define MAX_PINNED_INTERVAL 512
2583 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2584 * tasks if there is an imbalance.
2586 static int load_balance(int this_cpu, struct rq *this_rq,
2587 struct sched_domain *sd, enum cpu_idle_type idle,
2590 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2591 struct sched_group *group;
2592 unsigned long imbalance;
2594 cpumask_t cpus = CPU_MASK_ALL;
2595 unsigned long flags;
2598 * When power savings policy is enabled for the parent domain, idle
2599 * sibling can pick up load irrespective of busy siblings. In this case,
2600 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2601 * portraying it as CPU_NOT_IDLE.
2603 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2604 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2607 schedstat_inc(sd, lb_count[idle]);
2610 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2617 schedstat_inc(sd, lb_nobusyg[idle]);
2621 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2623 schedstat_inc(sd, lb_nobusyq[idle]);
2627 BUG_ON(busiest == this_rq);
2629 schedstat_add(sd, lb_imbalance[idle], imbalance);
2632 if (busiest->nr_running > 1) {
2634 * Attempt to move tasks. If find_busiest_group has found
2635 * an imbalance but busiest->nr_running <= 1, the group is
2636 * still unbalanced. ld_moved simply stays zero, so it is
2637 * correctly treated as an imbalance.
2639 local_irq_save(flags);
2640 double_rq_lock(this_rq, busiest);
2641 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2642 imbalance, sd, idle, &all_pinned);
2643 double_rq_unlock(this_rq, busiest);
2644 local_irq_restore(flags);
2647 * some other cpu did the load balance for us.
2649 if (ld_moved && this_cpu != smp_processor_id())
2650 resched_cpu(this_cpu);
2652 /* All tasks on this runqueue were pinned by CPU affinity */
2653 if (unlikely(all_pinned)) {
2654 cpu_clear(cpu_of(busiest), cpus);
2655 if (!cpus_empty(cpus))
2662 schedstat_inc(sd, lb_failed[idle]);
2663 sd->nr_balance_failed++;
2665 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2667 spin_lock_irqsave(&busiest->lock, flags);
2669 /* don't kick the migration_thread, if the curr
2670 * task on busiest cpu can't be moved to this_cpu
2672 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2673 spin_unlock_irqrestore(&busiest->lock, flags);
2675 goto out_one_pinned;
2678 if (!busiest->active_balance) {
2679 busiest->active_balance = 1;
2680 busiest->push_cpu = this_cpu;
2683 spin_unlock_irqrestore(&busiest->lock, flags);
2685 wake_up_process(busiest->migration_thread);
2688 * We've kicked active balancing, reset the failure
2691 sd->nr_balance_failed = sd->cache_nice_tries+1;
2694 sd->nr_balance_failed = 0;
2696 if (likely(!active_balance)) {
2697 /* We were unbalanced, so reset the balancing interval */
2698 sd->balance_interval = sd->min_interval;
2701 * If we've begun active balancing, start to back off. This
2702 * case may not be covered by the all_pinned logic if there
2703 * is only 1 task on the busy runqueue (because we don't call
2706 if (sd->balance_interval < sd->max_interval)
2707 sd->balance_interval *= 2;
2710 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2711 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2716 schedstat_inc(sd, lb_balanced[idle]);
2718 sd->nr_balance_failed = 0;
2721 /* tune up the balancing interval */
2722 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2723 (sd->balance_interval < sd->max_interval))
2724 sd->balance_interval *= 2;
2726 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2727 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2733 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2734 * tasks if there is an imbalance.
2736 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2737 * this_rq is locked.
2740 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2742 struct sched_group *group;
2743 struct rq *busiest = NULL;
2744 unsigned long imbalance;
2748 cpumask_t cpus = CPU_MASK_ALL;
2751 * When power savings policy is enabled for the parent domain, idle
2752 * sibling can pick up load irrespective of busy siblings. In this case,
2753 * let the state of idle sibling percolate up as IDLE, instead of
2754 * portraying it as CPU_NOT_IDLE.
2756 if (sd->flags & SD_SHARE_CPUPOWER &&
2757 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2760 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2762 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2763 &sd_idle, &cpus, NULL);
2765 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2769 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2772 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2776 BUG_ON(busiest == this_rq);
2778 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2781 if (busiest->nr_running > 1) {
2782 /* Attempt to move tasks */
2783 double_lock_balance(this_rq, busiest);
2784 /* this_rq->clock is already updated */
2785 update_rq_clock(busiest);
2786 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2787 imbalance, sd, CPU_NEWLY_IDLE,
2789 spin_unlock(&busiest->lock);
2791 if (unlikely(all_pinned)) {
2792 cpu_clear(cpu_of(busiest), cpus);
2793 if (!cpus_empty(cpus))
2799 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2800 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2801 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2804 sd->nr_balance_failed = 0;
2809 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2810 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2811 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2813 sd->nr_balance_failed = 0;
2819 * idle_balance is called by schedule() if this_cpu is about to become
2820 * idle. Attempts to pull tasks from other CPUs.
2822 static void idle_balance(int this_cpu, struct rq *this_rq)
2824 struct sched_domain *sd;
2825 int pulled_task = -1;
2826 unsigned long next_balance = jiffies + HZ;
2828 for_each_domain(this_cpu, sd) {
2829 unsigned long interval;
2831 if (!(sd->flags & SD_LOAD_BALANCE))
2834 if (sd->flags & SD_BALANCE_NEWIDLE)
2835 /* If we've pulled tasks over stop searching: */
2836 pulled_task = load_balance_newidle(this_cpu,
2839 interval = msecs_to_jiffies(sd->balance_interval);
2840 if (time_after(next_balance, sd->last_balance + interval))
2841 next_balance = sd->last_balance + interval;
2845 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2847 * We are going idle. next_balance may be set based on
2848 * a busy processor. So reset next_balance.
2850 this_rq->next_balance = next_balance;
2855 * active_load_balance is run by migration threads. It pushes running tasks
2856 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2857 * running on each physical CPU where possible, and avoids physical /
2858 * logical imbalances.
2860 * Called with busiest_rq locked.
2862 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2864 int target_cpu = busiest_rq->push_cpu;
2865 struct sched_domain *sd;
2866 struct rq *target_rq;
2868 /* Is there any task to move? */
2869 if (busiest_rq->nr_running <= 1)
2872 target_rq = cpu_rq(target_cpu);
2875 * This condition is "impossible", if it occurs
2876 * we need to fix it. Originally reported by
2877 * Bjorn Helgaas on a 128-cpu setup.
2879 BUG_ON(busiest_rq == target_rq);
2881 /* move a task from busiest_rq to target_rq */
2882 double_lock_balance(busiest_rq, target_rq);
2883 update_rq_clock(busiest_rq);
2884 update_rq_clock(target_rq);
2886 /* Search for an sd spanning us and the target CPU. */
2887 for_each_domain(target_cpu, sd) {
2888 if ((sd->flags & SD_LOAD_BALANCE) &&
2889 cpu_isset(busiest_cpu, sd->span))
2894 schedstat_inc(sd, alb_count);
2896 if (move_one_task(target_rq, target_cpu, busiest_rq,
2898 schedstat_inc(sd, alb_pushed);
2900 schedstat_inc(sd, alb_failed);
2902 spin_unlock(&target_rq->lock);
2907 atomic_t load_balancer;
2909 } nohz ____cacheline_aligned = {
2910 .load_balancer = ATOMIC_INIT(-1),
2911 .cpu_mask = CPU_MASK_NONE,
2915 * This routine will try to nominate the ilb (idle load balancing)
2916 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2917 * load balancing on behalf of all those cpus. If all the cpus in the system
2918 * go into this tickless mode, then there will be no ilb owner (as there is
2919 * no need for one) and all the cpus will sleep till the next wakeup event
2922 * For the ilb owner, tick is not stopped. And this tick will be used
2923 * for idle load balancing. ilb owner will still be part of
2926 * While stopping the tick, this cpu will become the ilb owner if there
2927 * is no other owner. And will be the owner till that cpu becomes busy
2928 * or if all cpus in the system stop their ticks at which point
2929 * there is no need for ilb owner.
2931 * When the ilb owner becomes busy, it nominates another owner, during the
2932 * next busy scheduler_tick()
2934 int select_nohz_load_balancer(int stop_tick)
2936 int cpu = smp_processor_id();
2939 cpu_set(cpu, nohz.cpu_mask);
2940 cpu_rq(cpu)->in_nohz_recently = 1;
2943 * If we are going offline and still the leader, give up!
2945 if (cpu_is_offline(cpu) &&
2946 atomic_read(&nohz.load_balancer) == cpu) {
2947 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2952 /* time for ilb owner also to sleep */
2953 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2954 if (atomic_read(&nohz.load_balancer) == cpu)
2955 atomic_set(&nohz.load_balancer, -1);
2959 if (atomic_read(&nohz.load_balancer) == -1) {
2960 /* make me the ilb owner */
2961 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2963 } else if (atomic_read(&nohz.load_balancer) == cpu)
2966 if (!cpu_isset(cpu, nohz.cpu_mask))
2969 cpu_clear(cpu, nohz.cpu_mask);
2971 if (atomic_read(&nohz.load_balancer) == cpu)
2972 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2979 static DEFINE_SPINLOCK(balancing);
2982 * It checks each scheduling domain to see if it is due to be balanced,
2983 * and initiates a balancing operation if so.
2985 * Balancing parameters are set up in arch_init_sched_domains.
2987 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
2990 struct rq *rq = cpu_rq(cpu);
2991 unsigned long interval;
2992 struct sched_domain *sd;
2993 /* Earliest time when we have to do rebalance again */
2994 unsigned long next_balance = jiffies + 60*HZ;
2995 int update_next_balance = 0;
2997 for_each_domain(cpu, sd) {
2998 if (!(sd->flags & SD_LOAD_BALANCE))
3001 interval = sd->balance_interval;
3002 if (idle != CPU_IDLE)
3003 interval *= sd->busy_factor;
3005 /* scale ms to jiffies */
3006 interval = msecs_to_jiffies(interval);
3007 if (unlikely(!interval))
3009 if (interval > HZ*NR_CPUS/10)
3010 interval = HZ*NR_CPUS/10;
3013 if (sd->flags & SD_SERIALIZE) {
3014 if (!spin_trylock(&balancing))
3018 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3019 if (load_balance(cpu, rq, sd, idle, &balance)) {
3021 * We've pulled tasks over so either we're no
3022 * longer idle, or one of our SMT siblings is
3025 idle = CPU_NOT_IDLE;
3027 sd->last_balance = jiffies;
3029 if (sd->flags & SD_SERIALIZE)
3030 spin_unlock(&balancing);
3032 if (time_after(next_balance, sd->last_balance + interval)) {
3033 next_balance = sd->last_balance + interval;
3034 update_next_balance = 1;
3038 * Stop the load balance at this level. There is another
3039 * CPU in our sched group which is doing load balancing more
3047 * next_balance will be updated only when there is a need.
3048 * When the cpu is attached to null domain for ex, it will not be
3051 if (likely(update_next_balance))
3052 rq->next_balance = next_balance;
3056 * run_rebalance_domains is triggered when needed from the scheduler tick.
3057 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3058 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3060 static void run_rebalance_domains(struct softirq_action *h)
3062 int this_cpu = smp_processor_id();
3063 struct rq *this_rq = cpu_rq(this_cpu);
3064 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3065 CPU_IDLE : CPU_NOT_IDLE;
3067 rebalance_domains(this_cpu, idle);
3071 * If this cpu is the owner for idle load balancing, then do the
3072 * balancing on behalf of the other idle cpus whose ticks are
3075 if (this_rq->idle_at_tick &&
3076 atomic_read(&nohz.load_balancer) == this_cpu) {
3077 cpumask_t cpus = nohz.cpu_mask;
3081 cpu_clear(this_cpu, cpus);
3082 for_each_cpu_mask(balance_cpu, cpus) {
3084 * If this cpu gets work to do, stop the load balancing
3085 * work being done for other cpus. Next load
3086 * balancing owner will pick it up.
3091 rebalance_domains(balance_cpu, CPU_IDLE);
3093 rq = cpu_rq(balance_cpu);
3094 if (time_after(this_rq->next_balance, rq->next_balance))
3095 this_rq->next_balance = rq->next_balance;
3102 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3104 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3105 * idle load balancing owner or decide to stop the periodic load balancing,
3106 * if the whole system is idle.
3108 static inline void trigger_load_balance(struct rq *rq, int cpu)
3112 * If we were in the nohz mode recently and busy at the current
3113 * scheduler tick, then check if we need to nominate new idle
3116 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3117 rq->in_nohz_recently = 0;
3119 if (atomic_read(&nohz.load_balancer) == cpu) {
3120 cpu_clear(cpu, nohz.cpu_mask);
3121 atomic_set(&nohz.load_balancer, -1);
3124 if (atomic_read(&nohz.load_balancer) == -1) {
3126 * simple selection for now: Nominate the
3127 * first cpu in the nohz list to be the next
3130 * TBD: Traverse the sched domains and nominate
3131 * the nearest cpu in the nohz.cpu_mask.
3133 int ilb = first_cpu(nohz.cpu_mask);
3141 * If this cpu is idle and doing idle load balancing for all the
3142 * cpus with ticks stopped, is it time for that to stop?
3144 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3145 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3151 * If this cpu is idle and the idle load balancing is done by
3152 * someone else, then no need raise the SCHED_SOFTIRQ
3154 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3155 cpu_isset(cpu, nohz.cpu_mask))
3158 if (time_after_eq(jiffies, rq->next_balance))
3159 raise_softirq(SCHED_SOFTIRQ);
3162 #else /* CONFIG_SMP */
3165 * on UP we do not need to balance between CPUs:
3167 static inline void idle_balance(int cpu, struct rq *rq)
3171 /* Avoid "used but not defined" warning on UP */
3172 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3173 unsigned long max_nr_move, unsigned long max_load_move,
3174 struct sched_domain *sd, enum cpu_idle_type idle,
3175 int *all_pinned, unsigned long *load_moved,
3176 int *this_best_prio, struct rq_iterator *iterator)
3185 DEFINE_PER_CPU(struct kernel_stat, kstat);
3187 EXPORT_PER_CPU_SYMBOL(kstat);
3190 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3191 * that have not yet been banked in case the task is currently running.
3193 unsigned long long task_sched_runtime(struct task_struct *p)
3195 unsigned long flags;
3199 rq = task_rq_lock(p, &flags);
3200 ns = p->se.sum_exec_runtime;
3201 if (rq->curr == p) {
3202 update_rq_clock(rq);
3203 delta_exec = rq->clock - p->se.exec_start;
3204 if ((s64)delta_exec > 0)
3207 task_rq_unlock(rq, &flags);
3213 * Account user cpu time to a process.
3214 * @p: the process that the cpu time gets accounted to
3215 * @hardirq_offset: the offset to subtract from hardirq_count()
3216 * @cputime: the cpu time spent in user space since the last update
3218 void account_user_time(struct task_struct *p, cputime_t cputime)
3220 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3223 p->utime = cputime_add(p->utime, cputime);
3225 /* Add user time to cpustat. */
3226 tmp = cputime_to_cputime64(cputime);
3227 if (TASK_NICE(p) > 0)
3228 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3230 cpustat->user = cputime64_add(cpustat->user, tmp);
3234 * Account system cpu time to a process.
3235 * @p: the process that the cpu time gets accounted to
3236 * @hardirq_offset: the offset to subtract from hardirq_count()
3237 * @cputime: the cpu time spent in kernel space since the last update
3239 void account_system_time(struct task_struct *p, int hardirq_offset,
3242 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3243 struct rq *rq = this_rq();
3246 p->stime = cputime_add(p->stime, cputime);
3248 /* Add system time to cpustat. */
3249 tmp = cputime_to_cputime64(cputime);
3250 if (hardirq_count() - hardirq_offset)
3251 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3252 else if (softirq_count())
3253 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3254 else if (p != rq->idle)
3255 cpustat->system = cputime64_add(cpustat->system, tmp);
3256 else if (atomic_read(&rq->nr_iowait) > 0)
3257 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3259 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3260 /* Account for system time used */
3261 acct_update_integrals(p);
3265 * Account for involuntary wait time.
3266 * @p: the process from which the cpu time has been stolen
3267 * @steal: the cpu time spent in involuntary wait
3269 void account_steal_time(struct task_struct *p, cputime_t steal)
3271 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3272 cputime64_t tmp = cputime_to_cputime64(steal);
3273 struct rq *rq = this_rq();
3275 if (p == rq->idle) {
3276 p->stime = cputime_add(p->stime, steal);
3277 if (atomic_read(&rq->nr_iowait) > 0)
3278 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3280 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3282 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3286 * This function gets called by the timer code, with HZ frequency.
3287 * We call it with interrupts disabled.
3289 * It also gets called by the fork code, when changing the parent's
3292 void scheduler_tick(void)
3294 int cpu = smp_processor_id();
3295 struct rq *rq = cpu_rq(cpu);
3296 struct task_struct *curr = rq->curr;
3297 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3299 spin_lock(&rq->lock);
3300 __update_rq_clock(rq);
3302 * Let rq->clock advance by at least TICK_NSEC:
3304 if (unlikely(rq->clock < next_tick))
3305 rq->clock = next_tick;
3306 rq->tick_timestamp = rq->clock;
3307 update_cpu_load(rq);
3308 if (curr != rq->idle) /* FIXME: needed? */
3309 curr->sched_class->task_tick(rq, curr);
3310 spin_unlock(&rq->lock);
3313 rq->idle_at_tick = idle_cpu(cpu);
3314 trigger_load_balance(rq, cpu);
3318 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3320 void fastcall add_preempt_count(int val)
3325 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3327 preempt_count() += val;
3329 * Spinlock count overflowing soon?
3331 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3334 EXPORT_SYMBOL(add_preempt_count);
3336 void fastcall sub_preempt_count(int val)
3341 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3344 * Is the spinlock portion underflowing?
3346 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3347 !(preempt_count() & PREEMPT_MASK)))
3350 preempt_count() -= val;
3352 EXPORT_SYMBOL(sub_preempt_count);
3357 * Print scheduling while atomic bug:
3359 static noinline void __schedule_bug(struct task_struct *prev)
3361 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3362 prev->comm, preempt_count(), prev->pid);
3363 debug_show_held_locks(prev);
3364 if (irqs_disabled())
3365 print_irqtrace_events(prev);
3370 * Various schedule()-time debugging checks and statistics:
3372 static inline void schedule_debug(struct task_struct *prev)
3375 * Test if we are atomic. Since do_exit() needs to call into
3376 * schedule() atomically, we ignore that path for now.
3377 * Otherwise, whine if we are scheduling when we should not be.
3379 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3380 __schedule_bug(prev);
3382 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3384 schedstat_inc(this_rq(), sched_count);
3385 #ifdef CONFIG_SCHEDSTATS
3386 if (unlikely(prev->lock_depth >= 0)) {
3387 schedstat_inc(this_rq(), bkl_count);
3388 schedstat_inc(prev, sched_info.bkl_count);
3394 * Pick up the highest-prio task:
3396 static inline struct task_struct *
3397 pick_next_task(struct rq *rq, struct task_struct *prev)
3399 const struct sched_class *class;
3400 struct task_struct *p;
3403 * Optimization: we know that if all tasks are in
3404 * the fair class we can call that function directly:
3406 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3407 p = fair_sched_class.pick_next_task(rq);
3412 class = sched_class_highest;
3414 p = class->pick_next_task(rq);
3418 * Will never be NULL as the idle class always
3419 * returns a non-NULL p:
3421 class = class->next;
3426 * schedule() is the main scheduler function.
3428 asmlinkage void __sched schedule(void)
3430 struct task_struct *prev, *next;
3437 cpu = smp_processor_id();
3441 switch_count = &prev->nivcsw;
3443 release_kernel_lock(prev);
3444 need_resched_nonpreemptible:
3446 schedule_debug(prev);
3449 * Do the rq-clock update outside the rq lock:
3451 local_irq_disable();
3452 __update_rq_clock(rq);
3453 spin_lock(&rq->lock);
3454 clear_tsk_need_resched(prev);
3456 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3457 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3458 unlikely(signal_pending(prev)))) {
3459 prev->state = TASK_RUNNING;
3461 deactivate_task(rq, prev, 1);
3463 switch_count = &prev->nvcsw;
3466 if (unlikely(!rq->nr_running))
3467 idle_balance(cpu, rq);
3469 prev->sched_class->put_prev_task(rq, prev);
3470 next = pick_next_task(rq, prev);
3472 sched_info_switch(prev, next);
3474 if (likely(prev != next)) {
3479 context_switch(rq, prev, next); /* unlocks the rq */
3481 spin_unlock_irq(&rq->lock);
3483 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3484 cpu = smp_processor_id();
3486 goto need_resched_nonpreemptible;
3488 preempt_enable_no_resched();
3489 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3492 EXPORT_SYMBOL(schedule);
3494 #ifdef CONFIG_PREEMPT
3496 * this is the entry point to schedule() from in-kernel preemption
3497 * off of preempt_enable. Kernel preemptions off return from interrupt
3498 * occur there and call schedule directly.
3500 asmlinkage void __sched preempt_schedule(void)
3502 struct thread_info *ti = current_thread_info();
3503 #ifdef CONFIG_PREEMPT_BKL
3504 struct task_struct *task = current;
3505 int saved_lock_depth;
3508 * If there is a non-zero preempt_count or interrupts are disabled,
3509 * we do not want to preempt the current task. Just return..
3511 if (likely(ti->preempt_count || irqs_disabled()))
3515 add_preempt_count(PREEMPT_ACTIVE);
3517 * We keep the big kernel semaphore locked, but we
3518 * clear ->lock_depth so that schedule() doesnt
3519 * auto-release the semaphore:
3521 #ifdef CONFIG_PREEMPT_BKL
3522 saved_lock_depth = task->lock_depth;
3523 task->lock_depth = -1;
3526 #ifdef CONFIG_PREEMPT_BKL
3527 task->lock_depth = saved_lock_depth;
3529 sub_preempt_count(PREEMPT_ACTIVE);
3531 /* we could miss a preemption opportunity between schedule and now */
3533 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3536 EXPORT_SYMBOL(preempt_schedule);
3539 * this is the entry point to schedule() from kernel preemption
3540 * off of irq context.
3541 * Note, that this is called and return with irqs disabled. This will
3542 * protect us against recursive calling from irq.
3544 asmlinkage void __sched preempt_schedule_irq(void)
3546 struct thread_info *ti = current_thread_info();
3547 #ifdef CONFIG_PREEMPT_BKL
3548 struct task_struct *task = current;
3549 int saved_lock_depth;
3551 /* Catch callers which need to be fixed */
3552 BUG_ON(ti->preempt_count || !irqs_disabled());
3555 add_preempt_count(PREEMPT_ACTIVE);
3557 * We keep the big kernel semaphore locked, but we
3558 * clear ->lock_depth so that schedule() doesnt
3559 * auto-release the semaphore:
3561 #ifdef CONFIG_PREEMPT_BKL
3562 saved_lock_depth = task->lock_depth;
3563 task->lock_depth = -1;
3567 local_irq_disable();
3568 #ifdef CONFIG_PREEMPT_BKL
3569 task->lock_depth = saved_lock_depth;
3571 sub_preempt_count(PREEMPT_ACTIVE);
3573 /* we could miss a preemption opportunity between schedule and now */
3575 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3579 #endif /* CONFIG_PREEMPT */
3581 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3584 return try_to_wake_up(curr->private, mode, sync);
3586 EXPORT_SYMBOL(default_wake_function);
3589 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3590 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3591 * number) then we wake all the non-exclusive tasks and one exclusive task.
3593 * There are circumstances in which we can try to wake a task which has already
3594 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3595 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3597 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3598 int nr_exclusive, int sync, void *key)
3600 wait_queue_t *curr, *next;
3602 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3603 unsigned flags = curr->flags;
3605 if (curr->func(curr, mode, sync, key) &&
3606 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3612 * __wake_up - wake up threads blocked on a waitqueue.
3614 * @mode: which threads
3615 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3616 * @key: is directly passed to the wakeup function
3618 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3619 int nr_exclusive, void *key)
3621 unsigned long flags;
3623 spin_lock_irqsave(&q->lock, flags);
3624 __wake_up_common(q, mode, nr_exclusive, 0, key);
3625 spin_unlock_irqrestore(&q->lock, flags);
3627 EXPORT_SYMBOL(__wake_up);
3630 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3632 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3634 __wake_up_common(q, mode, 1, 0, NULL);
3638 * __wake_up_sync - wake up threads blocked on a waitqueue.
3640 * @mode: which threads
3641 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3643 * The sync wakeup differs that the waker knows that it will schedule
3644 * away soon, so while the target thread will be woken up, it will not
3645 * be migrated to another CPU - ie. the two threads are 'synchronized'
3646 * with each other. This can prevent needless bouncing between CPUs.
3648 * On UP it can prevent extra preemption.
3651 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3653 unsigned long flags;
3659 if (unlikely(!nr_exclusive))
3662 spin_lock_irqsave(&q->lock, flags);
3663 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3664 spin_unlock_irqrestore(&q->lock, flags);
3666 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3668 void fastcall complete(struct completion *x)
3670 unsigned long flags;
3672 spin_lock_irqsave(&x->wait.lock, flags);
3674 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3676 spin_unlock_irqrestore(&x->wait.lock, flags);
3678 EXPORT_SYMBOL(complete);
3680 void fastcall complete_all(struct completion *x)
3682 unsigned long flags;
3684 spin_lock_irqsave(&x->wait.lock, flags);
3685 x->done += UINT_MAX/2;
3686 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3688 spin_unlock_irqrestore(&x->wait.lock, flags);
3690 EXPORT_SYMBOL(complete_all);
3692 void fastcall __sched wait_for_completion(struct completion *x)
3696 spin_lock_irq(&x->wait.lock);
3698 DECLARE_WAITQUEUE(wait, current);
3700 wait.flags |= WQ_FLAG_EXCLUSIVE;
3701 __add_wait_queue_tail(&x->wait, &wait);
3703 __set_current_state(TASK_UNINTERRUPTIBLE);
3704 spin_unlock_irq(&x->wait.lock);
3706 spin_lock_irq(&x->wait.lock);
3708 __remove_wait_queue(&x->wait, &wait);
3711 spin_unlock_irq(&x->wait.lock);
3713 EXPORT_SYMBOL(wait_for_completion);
3715 unsigned long fastcall __sched
3716 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3720 spin_lock_irq(&x->wait.lock);
3722 DECLARE_WAITQUEUE(wait, current);
3724 wait.flags |= WQ_FLAG_EXCLUSIVE;
3725 __add_wait_queue_tail(&x->wait, &wait);
3727 __set_current_state(TASK_UNINTERRUPTIBLE);
3728 spin_unlock_irq(&x->wait.lock);
3729 timeout = schedule_timeout(timeout);
3730 spin_lock_irq(&x->wait.lock);
3732 __remove_wait_queue(&x->wait, &wait);
3736 __remove_wait_queue(&x->wait, &wait);
3740 spin_unlock_irq(&x->wait.lock);
3743 EXPORT_SYMBOL(wait_for_completion_timeout);
3745 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3751 spin_lock_irq(&x->wait.lock);
3753 DECLARE_WAITQUEUE(wait, current);
3755 wait.flags |= WQ_FLAG_EXCLUSIVE;
3756 __add_wait_queue_tail(&x->wait, &wait);
3758 if (signal_pending(current)) {
3760 __remove_wait_queue(&x->wait, &wait);
3763 __set_current_state(TASK_INTERRUPTIBLE);
3764 spin_unlock_irq(&x->wait.lock);
3766 spin_lock_irq(&x->wait.lock);
3768 __remove_wait_queue(&x->wait, &wait);
3772 spin_unlock_irq(&x->wait.lock);
3776 EXPORT_SYMBOL(wait_for_completion_interruptible);
3778 unsigned long fastcall __sched
3779 wait_for_completion_interruptible_timeout(struct completion *x,
3780 unsigned long timeout)
3784 spin_lock_irq(&x->wait.lock);
3786 DECLARE_WAITQUEUE(wait, current);
3788 wait.flags |= WQ_FLAG_EXCLUSIVE;
3789 __add_wait_queue_tail(&x->wait, &wait);
3791 if (signal_pending(current)) {
3792 timeout = -ERESTARTSYS;
3793 __remove_wait_queue(&x->wait, &wait);
3796 __set_current_state(TASK_INTERRUPTIBLE);
3797 spin_unlock_irq(&x->wait.lock);
3798 timeout = schedule_timeout(timeout);
3799 spin_lock_irq(&x->wait.lock);
3801 __remove_wait_queue(&x->wait, &wait);
3805 __remove_wait_queue(&x->wait, &wait);
3809 spin_unlock_irq(&x->wait.lock);
3812 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3815 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3817 spin_lock_irqsave(&q->lock, *flags);
3818 __add_wait_queue(q, wait);
3819 spin_unlock(&q->lock);
3823 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3825 spin_lock_irq(&q->lock);
3826 __remove_wait_queue(q, wait);
3827 spin_unlock_irqrestore(&q->lock, *flags);
3830 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3832 unsigned long flags;
3835 init_waitqueue_entry(&wait, current);
3837 current->state = TASK_INTERRUPTIBLE;
3839 sleep_on_head(q, &wait, &flags);
3841 sleep_on_tail(q, &wait, &flags);
3843 EXPORT_SYMBOL(interruptible_sleep_on);
3846 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3848 unsigned long flags;
3851 init_waitqueue_entry(&wait, current);
3853 current->state = TASK_INTERRUPTIBLE;
3855 sleep_on_head(q, &wait, &flags);
3856 timeout = schedule_timeout(timeout);
3857 sleep_on_tail(q, &wait, &flags);
3861 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3863 void __sched sleep_on(wait_queue_head_t *q)
3865 unsigned long flags;
3868 init_waitqueue_entry(&wait, current);
3870 current->state = TASK_UNINTERRUPTIBLE;
3872 sleep_on_head(q, &wait, &flags);
3874 sleep_on_tail(q, &wait, &flags);
3876 EXPORT_SYMBOL(sleep_on);
3878 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3880 unsigned long flags;
3883 init_waitqueue_entry(&wait, current);
3885 current->state = TASK_UNINTERRUPTIBLE;
3887 sleep_on_head(q, &wait, &flags);
3888 timeout = schedule_timeout(timeout);
3889 sleep_on_tail(q, &wait, &flags);
3893 EXPORT_SYMBOL(sleep_on_timeout);
3895 #ifdef CONFIG_RT_MUTEXES
3898 * rt_mutex_setprio - set the current priority of a task
3900 * @prio: prio value (kernel-internal form)
3902 * This function changes the 'effective' priority of a task. It does
3903 * not touch ->normal_prio like __setscheduler().
3905 * Used by the rt_mutex code to implement priority inheritance logic.
3907 void rt_mutex_setprio(struct task_struct *p, int prio)
3909 unsigned long flags;
3910 int oldprio, on_rq, running;
3913 BUG_ON(prio < 0 || prio > MAX_PRIO);
3915 rq = task_rq_lock(p, &flags);
3916 update_rq_clock(rq);
3919 on_rq = p->se.on_rq;
3920 running = task_running(rq, p);
3922 dequeue_task(rq, p, 0);
3924 p->sched_class->put_prev_task(rq, p);
3928 p->sched_class = &rt_sched_class;
3930 p->sched_class = &fair_sched_class;
3936 p->sched_class->set_curr_task(rq);
3937 enqueue_task(rq, p, 0);
3939 * Reschedule if we are currently running on this runqueue and
3940 * our priority decreased, or if we are not currently running on
3941 * this runqueue and our priority is higher than the current's
3944 if (p->prio > oldprio)
3945 resched_task(rq->curr);
3947 check_preempt_curr(rq, p);
3950 task_rq_unlock(rq, &flags);
3955 void set_user_nice(struct task_struct *p, long nice)
3957 int old_prio, delta, on_rq;
3958 unsigned long flags;
3961 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3964 * We have to be careful, if called from sys_setpriority(),
3965 * the task might be in the middle of scheduling on another CPU.
3967 rq = task_rq_lock(p, &flags);
3968 update_rq_clock(rq);
3970 * The RT priorities are set via sched_setscheduler(), but we still
3971 * allow the 'normal' nice value to be set - but as expected
3972 * it wont have any effect on scheduling until the task is
3973 * SCHED_FIFO/SCHED_RR:
3975 if (task_has_rt_policy(p)) {
3976 p->static_prio = NICE_TO_PRIO(nice);
3979 on_rq = p->se.on_rq;
3981 dequeue_task(rq, p, 0);
3985 p->static_prio = NICE_TO_PRIO(nice);
3988 p->prio = effective_prio(p);
3989 delta = p->prio - old_prio;
3992 enqueue_task(rq, p, 0);
3995 * If the task increased its priority or is running and
3996 * lowered its priority, then reschedule its CPU:
3998 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3999 resched_task(rq->curr);
4002 task_rq_unlock(rq, &flags);
4004 EXPORT_SYMBOL(set_user_nice);
4007 * can_nice - check if a task can reduce its nice value
4011 int can_nice(const struct task_struct *p, const int nice)
4013 /* convert nice value [19,-20] to rlimit style value [1,40] */
4014 int nice_rlim = 20 - nice;
4016 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4017 capable(CAP_SYS_NICE));
4020 #ifdef __ARCH_WANT_SYS_NICE
4023 * sys_nice - change the priority of the current process.
4024 * @increment: priority increment
4026 * sys_setpriority is a more generic, but much slower function that
4027 * does similar things.
4029 asmlinkage long sys_nice(int increment)
4034 * Setpriority might change our priority at the same moment.
4035 * We don't have to worry. Conceptually one call occurs first
4036 * and we have a single winner.
4038 if (increment < -40)
4043 nice = PRIO_TO_NICE(current->static_prio) + increment;
4049 if (increment < 0 && !can_nice(current, nice))
4052 retval = security_task_setnice(current, nice);
4056 set_user_nice(current, nice);
4063 * task_prio - return the priority value of a given task.
4064 * @p: the task in question.
4066 * This is the priority value as seen by users in /proc.
4067 * RT tasks are offset by -200. Normal tasks are centered
4068 * around 0, value goes from -16 to +15.
4070 int task_prio(const struct task_struct *p)
4072 return p->prio - MAX_RT_PRIO;
4076 * task_nice - return the nice value of a given task.
4077 * @p: the task in question.
4079 int task_nice(const struct task_struct *p)
4081 return TASK_NICE(p);
4083 EXPORT_SYMBOL_GPL(task_nice);
4086 * idle_cpu - is a given cpu idle currently?
4087 * @cpu: the processor in question.
4089 int idle_cpu(int cpu)
4091 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4095 * idle_task - return the idle task for a given cpu.
4096 * @cpu: the processor in question.
4098 struct task_struct *idle_task(int cpu)
4100 return cpu_rq(cpu)->idle;
4104 * find_process_by_pid - find a process with a matching PID value.
4105 * @pid: the pid in question.
4107 static struct task_struct *find_process_by_pid(pid_t pid)
4109 return pid ? find_task_by_pid(pid) : current;
4112 /* Actually do priority change: must hold rq lock. */
4114 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4116 BUG_ON(p->se.on_rq);
4119 switch (p->policy) {
4123 p->sched_class = &fair_sched_class;
4127 p->sched_class = &rt_sched_class;
4131 p->rt_priority = prio;
4132 p->normal_prio = normal_prio(p);
4133 /* we are holding p->pi_lock already */
4134 p->prio = rt_mutex_getprio(p);
4139 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4140 * @p: the task in question.
4141 * @policy: new policy.
4142 * @param: structure containing the new RT priority.
4144 * NOTE that the task may be already dead.
4146 int sched_setscheduler(struct task_struct *p, int policy,
4147 struct sched_param *param)
4149 int retval, oldprio, oldpolicy = -1, on_rq, running;
4150 unsigned long flags;
4153 /* may grab non-irq protected spin_locks */
4154 BUG_ON(in_interrupt());
4156 /* double check policy once rq lock held */
4158 policy = oldpolicy = p->policy;
4159 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4160 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4161 policy != SCHED_IDLE)
4164 * Valid priorities for SCHED_FIFO and SCHED_RR are
4165 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4166 * SCHED_BATCH and SCHED_IDLE is 0.
4168 if (param->sched_priority < 0 ||
4169 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4170 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4172 if (rt_policy(policy) != (param->sched_priority != 0))
4176 * Allow unprivileged RT tasks to decrease priority:
4178 if (!capable(CAP_SYS_NICE)) {
4179 if (rt_policy(policy)) {
4180 unsigned long rlim_rtprio;
4182 if (!lock_task_sighand(p, &flags))
4184 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4185 unlock_task_sighand(p, &flags);
4187 /* can't set/change the rt policy */
4188 if (policy != p->policy && !rlim_rtprio)
4191 /* can't increase priority */
4192 if (param->sched_priority > p->rt_priority &&
4193 param->sched_priority > rlim_rtprio)
4197 * Like positive nice levels, dont allow tasks to
4198 * move out of SCHED_IDLE either:
4200 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4203 /* can't change other user's priorities */
4204 if ((current->euid != p->euid) &&
4205 (current->euid != p->uid))
4209 retval = security_task_setscheduler(p, policy, param);
4213 * make sure no PI-waiters arrive (or leave) while we are
4214 * changing the priority of the task:
4216 spin_lock_irqsave(&p->pi_lock, flags);
4218 * To be able to change p->policy safely, the apropriate
4219 * runqueue lock must be held.
4221 rq = __task_rq_lock(p);
4222 /* recheck policy now with rq lock held */
4223 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4224 policy = oldpolicy = -1;
4225 __task_rq_unlock(rq);
4226 spin_unlock_irqrestore(&p->pi_lock, flags);
4229 update_rq_clock(rq);
4230 on_rq = p->se.on_rq;
4231 running = task_running(rq, p);
4233 deactivate_task(rq, p, 0);
4235 p->sched_class->put_prev_task(rq, p);
4239 __setscheduler(rq, p, policy, param->sched_priority);
4243 p->sched_class->set_curr_task(rq);
4244 activate_task(rq, p, 0);
4246 * Reschedule if we are currently running on this runqueue and
4247 * our priority decreased, or if we are not currently running on
4248 * this runqueue and our priority is higher than the current's
4251 if (p->prio > oldprio)
4252 resched_task(rq->curr);
4254 check_preempt_curr(rq, p);
4257 __task_rq_unlock(rq);
4258 spin_unlock_irqrestore(&p->pi_lock, flags);
4260 rt_mutex_adjust_pi(p);
4264 EXPORT_SYMBOL_GPL(sched_setscheduler);
4267 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4269 struct sched_param lparam;
4270 struct task_struct *p;
4273 if (!param || pid < 0)
4275 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4280 p = find_process_by_pid(pid);
4282 retval = sched_setscheduler(p, policy, &lparam);
4289 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4290 * @pid: the pid in question.
4291 * @policy: new policy.
4292 * @param: structure containing the new RT priority.
4294 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4295 struct sched_param __user *param)
4297 /* negative values for policy are not valid */
4301 return do_sched_setscheduler(pid, policy, param);
4305 * sys_sched_setparam - set/change the RT priority of a thread
4306 * @pid: the pid in question.
4307 * @param: structure containing the new RT priority.
4309 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4311 return do_sched_setscheduler(pid, -1, param);
4315 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4316 * @pid: the pid in question.
4318 asmlinkage long sys_sched_getscheduler(pid_t pid)
4320 struct task_struct *p;
4321 int retval = -EINVAL;
4327 read_lock(&tasklist_lock);
4328 p = find_process_by_pid(pid);
4330 retval = security_task_getscheduler(p);
4334 read_unlock(&tasklist_lock);
4341 * sys_sched_getscheduler - get the RT priority of a thread
4342 * @pid: the pid in question.
4343 * @param: structure containing the RT priority.
4345 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4347 struct sched_param lp;
4348 struct task_struct *p;
4349 int retval = -EINVAL;
4351 if (!param || pid < 0)
4354 read_lock(&tasklist_lock);
4355 p = find_process_by_pid(pid);
4360 retval = security_task_getscheduler(p);
4364 lp.sched_priority = p->rt_priority;
4365 read_unlock(&tasklist_lock);
4368 * This one might sleep, we cannot do it with a spinlock held ...
4370 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4376 read_unlock(&tasklist_lock);
4380 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4382 cpumask_t cpus_allowed;
4383 struct task_struct *p;
4386 mutex_lock(&sched_hotcpu_mutex);
4387 read_lock(&tasklist_lock);
4389 p = find_process_by_pid(pid);
4391 read_unlock(&tasklist_lock);
4392 mutex_unlock(&sched_hotcpu_mutex);
4397 * It is not safe to call set_cpus_allowed with the
4398 * tasklist_lock held. We will bump the task_struct's
4399 * usage count and then drop tasklist_lock.
4402 read_unlock(&tasklist_lock);
4405 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4406 !capable(CAP_SYS_NICE))
4409 retval = security_task_setscheduler(p, 0, NULL);
4413 cpus_allowed = cpuset_cpus_allowed(p);
4414 cpus_and(new_mask, new_mask, cpus_allowed);
4415 retval = set_cpus_allowed(p, new_mask);
4419 mutex_unlock(&sched_hotcpu_mutex);
4423 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4424 cpumask_t *new_mask)
4426 if (len < sizeof(cpumask_t)) {
4427 memset(new_mask, 0, sizeof(cpumask_t));
4428 } else if (len > sizeof(cpumask_t)) {
4429 len = sizeof(cpumask_t);
4431 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4435 * sys_sched_setaffinity - set the cpu affinity of a process
4436 * @pid: pid of the process
4437 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4438 * @user_mask_ptr: user-space pointer to the new cpu mask
4440 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4441 unsigned long __user *user_mask_ptr)
4446 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4450 return sched_setaffinity(pid, new_mask);
4454 * Represents all cpu's present in the system
4455 * In systems capable of hotplug, this map could dynamically grow
4456 * as new cpu's are detected in the system via any platform specific
4457 * method, such as ACPI for e.g.
4460 cpumask_t cpu_present_map __read_mostly;
4461 EXPORT_SYMBOL(cpu_present_map);
4464 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4465 EXPORT_SYMBOL(cpu_online_map);
4467 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4468 EXPORT_SYMBOL(cpu_possible_map);
4471 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4473 struct task_struct *p;
4476 mutex_lock(&sched_hotcpu_mutex);
4477 read_lock(&tasklist_lock);
4480 p = find_process_by_pid(pid);
4484 retval = security_task_getscheduler(p);
4488 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4491 read_unlock(&tasklist_lock);
4492 mutex_unlock(&sched_hotcpu_mutex);
4498 * sys_sched_getaffinity - get the cpu affinity of a process
4499 * @pid: pid of the process
4500 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4501 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4503 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4504 unsigned long __user *user_mask_ptr)
4509 if (len < sizeof(cpumask_t))
4512 ret = sched_getaffinity(pid, &mask);
4516 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4519 return sizeof(cpumask_t);
4523 * sys_sched_yield - yield the current processor to other threads.
4525 * This function yields the current CPU to other tasks. If there are no
4526 * other threads running on this CPU then this function will return.
4528 asmlinkage long sys_sched_yield(void)
4530 struct rq *rq = this_rq_lock();
4532 schedstat_inc(rq, yld_count);
4533 current->sched_class->yield_task(rq);
4536 * Since we are going to call schedule() anyway, there's
4537 * no need to preempt or enable interrupts:
4539 __release(rq->lock);
4540 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4541 _raw_spin_unlock(&rq->lock);
4542 preempt_enable_no_resched();
4549 static void __cond_resched(void)
4551 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4552 __might_sleep(__FILE__, __LINE__);
4555 * The BKS might be reacquired before we have dropped
4556 * PREEMPT_ACTIVE, which could trigger a second
4557 * cond_resched() call.
4560 add_preempt_count(PREEMPT_ACTIVE);
4562 sub_preempt_count(PREEMPT_ACTIVE);
4563 } while (need_resched());
4566 int __sched cond_resched(void)
4568 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4569 system_state == SYSTEM_RUNNING) {
4575 EXPORT_SYMBOL(cond_resched);
4578 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4579 * call schedule, and on return reacquire the lock.
4581 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4582 * operations here to prevent schedule() from being called twice (once via
4583 * spin_unlock(), once by hand).
4585 int cond_resched_lock(spinlock_t *lock)
4589 if (need_lockbreak(lock)) {
4595 if (need_resched() && system_state == SYSTEM_RUNNING) {
4596 spin_release(&lock->dep_map, 1, _THIS_IP_);
4597 _raw_spin_unlock(lock);
4598 preempt_enable_no_resched();
4605 EXPORT_SYMBOL(cond_resched_lock);
4607 int __sched cond_resched_softirq(void)
4609 BUG_ON(!in_softirq());
4611 if (need_resched() && system_state == SYSTEM_RUNNING) {
4619 EXPORT_SYMBOL(cond_resched_softirq);
4622 * yield - yield the current processor to other threads.
4624 * This is a shortcut for kernel-space yielding - it marks the
4625 * thread runnable and calls sys_sched_yield().
4627 void __sched yield(void)
4629 set_current_state(TASK_RUNNING);
4632 EXPORT_SYMBOL(yield);
4635 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4636 * that process accounting knows that this is a task in IO wait state.
4638 * But don't do that if it is a deliberate, throttling IO wait (this task
4639 * has set its backing_dev_info: the queue against which it should throttle)
4641 void __sched io_schedule(void)
4643 struct rq *rq = &__raw_get_cpu_var(runqueues);
4645 delayacct_blkio_start();
4646 atomic_inc(&rq->nr_iowait);
4648 atomic_dec(&rq->nr_iowait);
4649 delayacct_blkio_end();
4651 EXPORT_SYMBOL(io_schedule);
4653 long __sched io_schedule_timeout(long timeout)
4655 struct rq *rq = &__raw_get_cpu_var(runqueues);
4658 delayacct_blkio_start();
4659 atomic_inc(&rq->nr_iowait);
4660 ret = schedule_timeout(timeout);
4661 atomic_dec(&rq->nr_iowait);
4662 delayacct_blkio_end();
4667 * sys_sched_get_priority_max - return maximum RT priority.
4668 * @policy: scheduling class.
4670 * this syscall returns the maximum rt_priority that can be used
4671 * by a given scheduling class.
4673 asmlinkage long sys_sched_get_priority_max(int policy)
4680 ret = MAX_USER_RT_PRIO-1;
4692 * sys_sched_get_priority_min - return minimum RT priority.
4693 * @policy: scheduling class.
4695 * this syscall returns the minimum rt_priority that can be used
4696 * by a given scheduling class.
4698 asmlinkage long sys_sched_get_priority_min(int policy)
4716 * sys_sched_rr_get_interval - return the default timeslice of a process.
4717 * @pid: pid of the process.
4718 * @interval: userspace pointer to the timeslice value.
4720 * this syscall writes the default timeslice value of a given process
4721 * into the user-space timespec buffer. A value of '0' means infinity.
4724 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4726 struct task_struct *p;
4727 unsigned int time_slice;
4728 int retval = -EINVAL;
4735 read_lock(&tasklist_lock);
4736 p = find_process_by_pid(pid);
4740 retval = security_task_getscheduler(p);
4744 if (p->policy == SCHED_FIFO)
4746 else if (p->policy == SCHED_RR)
4747 time_slice = DEF_TIMESLICE;
4749 struct sched_entity *se = &p->se;
4750 unsigned long flags;
4753 rq = task_rq_lock(p, &flags);
4754 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4755 task_rq_unlock(rq, &flags);
4757 read_unlock(&tasklist_lock);
4758 jiffies_to_timespec(time_slice, &t);
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 * activate_idle_task - move idle task to the _front_ of runqueue.
5139 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5141 update_rq_clock(rq);
5143 if (p->state == TASK_UNINTERRUPTIBLE)
5144 rq->nr_uninterruptible--;
5146 enqueue_task(rq, p, 0);
5147 inc_nr_running(p, rq);
5151 * Schedules idle task to be the next runnable task on current CPU.
5152 * It does so by boosting its priority to highest possible and adding it to
5153 * the _front_ of the runqueue. Used by CPU offline code.
5155 void sched_idle_next(void)
5157 int this_cpu = smp_processor_id();
5158 struct rq *rq = cpu_rq(this_cpu);
5159 struct task_struct *p = rq->idle;
5160 unsigned long flags;
5162 /* cpu has to be offline */
5163 BUG_ON(cpu_online(this_cpu));
5166 * Strictly not necessary since rest of the CPUs are stopped by now
5167 * and interrupts disabled on the current cpu.
5169 spin_lock_irqsave(&rq->lock, flags);
5171 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5173 /* Add idle task to the _front_ of its priority queue: */
5174 activate_idle_task(p, rq);
5176 spin_unlock_irqrestore(&rq->lock, flags);
5180 * Ensures that the idle task is using init_mm right before its cpu goes
5183 void idle_task_exit(void)
5185 struct mm_struct *mm = current->active_mm;
5187 BUG_ON(cpu_online(smp_processor_id()));
5190 switch_mm(mm, &init_mm, current);
5194 /* called under rq->lock with disabled interrupts */
5195 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5197 struct rq *rq = cpu_rq(dead_cpu);
5199 /* Must be exiting, otherwise would be on tasklist. */
5200 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5202 /* Cannot have done final schedule yet: would have vanished. */
5203 BUG_ON(p->state == TASK_DEAD);
5208 * Drop lock around migration; if someone else moves it,
5209 * that's OK. No task can be added to this CPU, so iteration is
5211 * NOTE: interrupts should be left disabled --dev@
5213 spin_unlock(&rq->lock);
5214 move_task_off_dead_cpu(dead_cpu, p);
5215 spin_lock(&rq->lock);
5220 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5221 static void migrate_dead_tasks(unsigned int dead_cpu)
5223 struct rq *rq = cpu_rq(dead_cpu);
5224 struct task_struct *next;
5227 if (!rq->nr_running)
5229 update_rq_clock(rq);
5230 next = pick_next_task(rq, rq->curr);
5233 migrate_dead(dead_cpu, next);
5237 #endif /* CONFIG_HOTPLUG_CPU */
5239 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5241 static struct ctl_table sd_ctl_dir[] = {
5243 .procname = "sched_domain",
5249 static struct ctl_table sd_ctl_root[] = {
5251 .ctl_name = CTL_KERN,
5252 .procname = "kernel",
5254 .child = sd_ctl_dir,
5259 static struct ctl_table *sd_alloc_ctl_entry(int n)
5261 struct ctl_table *entry =
5262 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5265 memset(entry, 0, n * sizeof(struct ctl_table));
5271 set_table_entry(struct ctl_table *entry,
5272 const char *procname, void *data, int maxlen,
5273 mode_t mode, proc_handler *proc_handler)
5275 entry->procname = procname;
5277 entry->maxlen = maxlen;
5279 entry->proc_handler = proc_handler;
5282 static struct ctl_table *
5283 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5285 struct ctl_table *table = sd_alloc_ctl_entry(14);
5287 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5288 sizeof(long), 0644, proc_doulongvec_minmax);
5289 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5290 sizeof(long), 0644, proc_doulongvec_minmax);
5291 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5292 sizeof(int), 0644, proc_dointvec_minmax);
5293 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5294 sizeof(int), 0644, proc_dointvec_minmax);
5295 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5296 sizeof(int), 0644, proc_dointvec_minmax);
5297 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5298 sizeof(int), 0644, proc_dointvec_minmax);
5299 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5300 sizeof(int), 0644, proc_dointvec_minmax);
5301 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5302 sizeof(int), 0644, proc_dointvec_minmax);
5303 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5304 sizeof(int), 0644, proc_dointvec_minmax);
5305 set_table_entry(&table[10], "cache_nice_tries",
5306 &sd->cache_nice_tries,
5307 sizeof(int), 0644, proc_dointvec_minmax);
5308 set_table_entry(&table[12], "flags", &sd->flags,
5309 sizeof(int), 0644, proc_dointvec_minmax);
5314 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5316 struct ctl_table *entry, *table;
5317 struct sched_domain *sd;
5318 int domain_num = 0, i;
5321 for_each_domain(cpu, sd)
5323 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5326 for_each_domain(cpu, sd) {
5327 snprintf(buf, 32, "domain%d", i);
5328 entry->procname = kstrdup(buf, GFP_KERNEL);
5330 entry->child = sd_alloc_ctl_domain_table(sd);
5337 static struct ctl_table_header *sd_sysctl_header;
5338 static void init_sched_domain_sysctl(void)
5340 int i, cpu_num = num_online_cpus();
5341 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5344 sd_ctl_dir[0].child = entry;
5346 for (i = 0; i < cpu_num; i++, entry++) {
5347 snprintf(buf, 32, "cpu%d", i);
5348 entry->procname = kstrdup(buf, GFP_KERNEL);
5350 entry->child = sd_alloc_ctl_cpu_table(i);
5352 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5355 static void init_sched_domain_sysctl(void)
5361 * migration_call - callback that gets triggered when a CPU is added.
5362 * Here we can start up the necessary migration thread for the new CPU.
5364 static int __cpuinit
5365 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5367 struct task_struct *p;
5368 int cpu = (long)hcpu;
5369 unsigned long flags;
5373 case CPU_LOCK_ACQUIRE:
5374 mutex_lock(&sched_hotcpu_mutex);
5377 case CPU_UP_PREPARE:
5378 case CPU_UP_PREPARE_FROZEN:
5379 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5382 kthread_bind(p, cpu);
5383 /* Must be high prio: stop_machine expects to yield to it. */
5384 rq = task_rq_lock(p, &flags);
5385 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5386 task_rq_unlock(rq, &flags);
5387 cpu_rq(cpu)->migration_thread = p;
5391 case CPU_ONLINE_FROZEN:
5392 /* Strictly unneccessary, as first user will wake it. */
5393 wake_up_process(cpu_rq(cpu)->migration_thread);
5396 #ifdef CONFIG_HOTPLUG_CPU
5397 case CPU_UP_CANCELED:
5398 case CPU_UP_CANCELED_FROZEN:
5399 if (!cpu_rq(cpu)->migration_thread)
5401 /* Unbind it from offline cpu so it can run. Fall thru. */
5402 kthread_bind(cpu_rq(cpu)->migration_thread,
5403 any_online_cpu(cpu_online_map));
5404 kthread_stop(cpu_rq(cpu)->migration_thread);
5405 cpu_rq(cpu)->migration_thread = NULL;
5409 case CPU_DEAD_FROZEN:
5410 migrate_live_tasks(cpu);
5412 kthread_stop(rq->migration_thread);
5413 rq->migration_thread = NULL;
5414 /* Idle task back to normal (off runqueue, low prio) */
5415 rq = task_rq_lock(rq->idle, &flags);
5416 update_rq_clock(rq);
5417 deactivate_task(rq, rq->idle, 0);
5418 rq->idle->static_prio = MAX_PRIO;
5419 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5420 rq->idle->sched_class = &idle_sched_class;
5421 migrate_dead_tasks(cpu);
5422 task_rq_unlock(rq, &flags);
5423 migrate_nr_uninterruptible(rq);
5424 BUG_ON(rq->nr_running != 0);
5426 /* No need to migrate the tasks: it was best-effort if
5427 * they didn't take sched_hotcpu_mutex. Just wake up
5428 * the requestors. */
5429 spin_lock_irq(&rq->lock);
5430 while (!list_empty(&rq->migration_queue)) {
5431 struct migration_req *req;
5433 req = list_entry(rq->migration_queue.next,
5434 struct migration_req, list);
5435 list_del_init(&req->list);
5436 complete(&req->done);
5438 spin_unlock_irq(&rq->lock);
5441 case CPU_LOCK_RELEASE:
5442 mutex_unlock(&sched_hotcpu_mutex);
5448 /* Register at highest priority so that task migration (migrate_all_tasks)
5449 * happens before everything else.
5451 static struct notifier_block __cpuinitdata migration_notifier = {
5452 .notifier_call = migration_call,
5456 int __init migration_init(void)
5458 void *cpu = (void *)(long)smp_processor_id();
5461 /* Start one for the boot CPU: */
5462 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5463 BUG_ON(err == NOTIFY_BAD);
5464 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5465 register_cpu_notifier(&migration_notifier);
5473 /* Number of possible processor ids */
5474 int nr_cpu_ids __read_mostly = NR_CPUS;
5475 EXPORT_SYMBOL(nr_cpu_ids);
5477 #ifdef CONFIG_SCHED_DEBUG
5478 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5483 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5487 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5492 struct sched_group *group = sd->groups;
5493 cpumask_t groupmask;
5495 cpumask_scnprintf(str, NR_CPUS, sd->span);
5496 cpus_clear(groupmask);
5499 for (i = 0; i < level + 1; i++)
5501 printk("domain %d: ", level);
5503 if (!(sd->flags & SD_LOAD_BALANCE)) {
5504 printk("does not load-balance\n");
5506 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5511 printk("span %s\n", str);
5513 if (!cpu_isset(cpu, sd->span))
5514 printk(KERN_ERR "ERROR: domain->span does not contain "
5516 if (!cpu_isset(cpu, group->cpumask))
5517 printk(KERN_ERR "ERROR: domain->groups does not contain"
5521 for (i = 0; i < level + 2; i++)
5527 printk(KERN_ERR "ERROR: group is NULL\n");
5531 if (!group->__cpu_power) {
5533 printk(KERN_ERR "ERROR: domain->cpu_power not "
5538 if (!cpus_weight(group->cpumask)) {
5540 printk(KERN_ERR "ERROR: empty group\n");
5544 if (cpus_intersects(groupmask, group->cpumask)) {
5546 printk(KERN_ERR "ERROR: repeated CPUs\n");
5550 cpus_or(groupmask, groupmask, group->cpumask);
5552 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5555 group = group->next;
5556 } while (group != sd->groups);
5559 if (!cpus_equal(sd->span, groupmask))
5560 printk(KERN_ERR "ERROR: groups don't span "
5568 if (!cpus_subset(groupmask, sd->span))
5569 printk(KERN_ERR "ERROR: parent span is not a superset "
5570 "of domain->span\n");
5575 # define sched_domain_debug(sd, cpu) do { } while (0)
5578 static int sd_degenerate(struct sched_domain *sd)
5580 if (cpus_weight(sd->span) == 1)
5583 /* Following flags need at least 2 groups */
5584 if (sd->flags & (SD_LOAD_BALANCE |
5585 SD_BALANCE_NEWIDLE |
5589 SD_SHARE_PKG_RESOURCES)) {
5590 if (sd->groups != sd->groups->next)
5594 /* Following flags don't use groups */
5595 if (sd->flags & (SD_WAKE_IDLE |
5604 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5606 unsigned long cflags = sd->flags, pflags = parent->flags;
5608 if (sd_degenerate(parent))
5611 if (!cpus_equal(sd->span, parent->span))
5614 /* Does parent contain flags not in child? */
5615 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5616 if (cflags & SD_WAKE_AFFINE)
5617 pflags &= ~SD_WAKE_BALANCE;
5618 /* Flags needing groups don't count if only 1 group in parent */
5619 if (parent->groups == parent->groups->next) {
5620 pflags &= ~(SD_LOAD_BALANCE |
5621 SD_BALANCE_NEWIDLE |
5625 SD_SHARE_PKG_RESOURCES);
5627 if (~cflags & pflags)
5634 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5635 * hold the hotplug lock.
5637 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5639 struct rq *rq = cpu_rq(cpu);
5640 struct sched_domain *tmp;
5642 /* Remove the sched domains which do not contribute to scheduling. */
5643 for (tmp = sd; tmp; tmp = tmp->parent) {
5644 struct sched_domain *parent = tmp->parent;
5647 if (sd_parent_degenerate(tmp, parent)) {
5648 tmp->parent = parent->parent;
5650 parent->parent->child = tmp;
5654 if (sd && sd_degenerate(sd)) {
5660 sched_domain_debug(sd, cpu);
5662 rcu_assign_pointer(rq->sd, sd);
5665 /* cpus with isolated domains */
5666 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5668 /* Setup the mask of cpus configured for isolated domains */
5669 static int __init isolated_cpu_setup(char *str)
5671 int ints[NR_CPUS], i;
5673 str = get_options(str, ARRAY_SIZE(ints), ints);
5674 cpus_clear(cpu_isolated_map);
5675 for (i = 1; i <= ints[0]; i++)
5676 if (ints[i] < NR_CPUS)
5677 cpu_set(ints[i], cpu_isolated_map);
5681 __setup("isolcpus=", isolated_cpu_setup);
5684 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5685 * to a function which identifies what group(along with sched group) a CPU
5686 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5687 * (due to the fact that we keep track of groups covered with a cpumask_t).
5689 * init_sched_build_groups will build a circular linked list of the groups
5690 * covered by the given span, and will set each group's ->cpumask correctly,
5691 * and ->cpu_power to 0.
5694 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5695 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5696 struct sched_group **sg))
5698 struct sched_group *first = NULL, *last = NULL;
5699 cpumask_t covered = CPU_MASK_NONE;
5702 for_each_cpu_mask(i, span) {
5703 struct sched_group *sg;
5704 int group = group_fn(i, cpu_map, &sg);
5707 if (cpu_isset(i, covered))
5710 sg->cpumask = CPU_MASK_NONE;
5711 sg->__cpu_power = 0;
5713 for_each_cpu_mask(j, span) {
5714 if (group_fn(j, cpu_map, NULL) != group)
5717 cpu_set(j, covered);
5718 cpu_set(j, sg->cpumask);
5729 #define SD_NODES_PER_DOMAIN 16
5734 * find_next_best_node - find the next node to include in a sched_domain
5735 * @node: node whose sched_domain we're building
5736 * @used_nodes: nodes already in the sched_domain
5738 * Find the next node to include in a given scheduling domain. Simply
5739 * finds the closest node not already in the @used_nodes map.
5741 * Should use nodemask_t.
5743 static int find_next_best_node(int node, unsigned long *used_nodes)
5745 int i, n, val, min_val, best_node = 0;
5749 for (i = 0; i < MAX_NUMNODES; i++) {
5750 /* Start at @node */
5751 n = (node + i) % MAX_NUMNODES;
5753 if (!nr_cpus_node(n))
5756 /* Skip already used nodes */
5757 if (test_bit(n, used_nodes))
5760 /* Simple min distance search */
5761 val = node_distance(node, n);
5763 if (val < min_val) {
5769 set_bit(best_node, used_nodes);
5774 * sched_domain_node_span - get a cpumask for a node's sched_domain
5775 * @node: node whose cpumask we're constructing
5776 * @size: number of nodes to include in this span
5778 * Given a node, construct a good cpumask for its sched_domain to span. It
5779 * should be one that prevents unnecessary balancing, but also spreads tasks
5782 static cpumask_t sched_domain_node_span(int node)
5784 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5785 cpumask_t span, nodemask;
5789 bitmap_zero(used_nodes, MAX_NUMNODES);
5791 nodemask = node_to_cpumask(node);
5792 cpus_or(span, span, nodemask);
5793 set_bit(node, used_nodes);
5795 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5796 int next_node = find_next_best_node(node, used_nodes);
5798 nodemask = node_to_cpumask(next_node);
5799 cpus_or(span, span, nodemask);
5806 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5809 * SMT sched-domains:
5811 #ifdef CONFIG_SCHED_SMT
5812 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5813 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5815 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5816 struct sched_group **sg)
5819 *sg = &per_cpu(sched_group_cpus, cpu);
5825 * multi-core sched-domains:
5827 #ifdef CONFIG_SCHED_MC
5828 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5829 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5832 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5833 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5834 struct sched_group **sg)
5837 cpumask_t mask = cpu_sibling_map[cpu];
5838 cpus_and(mask, mask, *cpu_map);
5839 group = first_cpu(mask);
5841 *sg = &per_cpu(sched_group_core, group);
5844 #elif defined(CONFIG_SCHED_MC)
5845 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5846 struct sched_group **sg)
5849 *sg = &per_cpu(sched_group_core, cpu);
5854 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5855 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5857 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5858 struct sched_group **sg)
5861 #ifdef CONFIG_SCHED_MC
5862 cpumask_t mask = cpu_coregroup_map(cpu);
5863 cpus_and(mask, mask, *cpu_map);
5864 group = first_cpu(mask);
5865 #elif defined(CONFIG_SCHED_SMT)
5866 cpumask_t mask = cpu_sibling_map[cpu];
5867 cpus_and(mask, mask, *cpu_map);
5868 group = first_cpu(mask);
5873 *sg = &per_cpu(sched_group_phys, group);
5879 * The init_sched_build_groups can't handle what we want to do with node
5880 * groups, so roll our own. Now each node has its own list of groups which
5881 * gets dynamically allocated.
5883 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5884 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5886 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5887 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5889 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5890 struct sched_group **sg)
5892 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5895 cpus_and(nodemask, nodemask, *cpu_map);
5896 group = first_cpu(nodemask);
5899 *sg = &per_cpu(sched_group_allnodes, group);
5903 static void init_numa_sched_groups_power(struct sched_group *group_head)
5905 struct sched_group *sg = group_head;
5911 for_each_cpu_mask(j, sg->cpumask) {
5912 struct sched_domain *sd;
5914 sd = &per_cpu(phys_domains, j);
5915 if (j != first_cpu(sd->groups->cpumask)) {
5917 * Only add "power" once for each
5923 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5926 if (sg != group_head)
5932 /* Free memory allocated for various sched_group structures */
5933 static void free_sched_groups(const cpumask_t *cpu_map)
5937 for_each_cpu_mask(cpu, *cpu_map) {
5938 struct sched_group **sched_group_nodes
5939 = sched_group_nodes_bycpu[cpu];
5941 if (!sched_group_nodes)
5944 for (i = 0; i < MAX_NUMNODES; i++) {
5945 cpumask_t nodemask = node_to_cpumask(i);
5946 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5948 cpus_and(nodemask, nodemask, *cpu_map);
5949 if (cpus_empty(nodemask))
5959 if (oldsg != sched_group_nodes[i])
5962 kfree(sched_group_nodes);
5963 sched_group_nodes_bycpu[cpu] = NULL;
5967 static void free_sched_groups(const cpumask_t *cpu_map)
5973 * Initialize sched groups cpu_power.
5975 * cpu_power indicates the capacity of sched group, which is used while
5976 * distributing the load between different sched groups in a sched domain.
5977 * Typically cpu_power for all the groups in a sched domain will be same unless
5978 * there are asymmetries in the topology. If there are asymmetries, group
5979 * having more cpu_power will pickup more load compared to the group having
5982 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5983 * the maximum number of tasks a group can handle in the presence of other idle
5984 * or lightly loaded groups in the same sched domain.
5986 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5988 struct sched_domain *child;
5989 struct sched_group *group;
5991 WARN_ON(!sd || !sd->groups);
5993 if (cpu != first_cpu(sd->groups->cpumask))
5998 sd->groups->__cpu_power = 0;
6001 * For perf policy, if the groups in child domain share resources
6002 * (for example cores sharing some portions of the cache hierarchy
6003 * or SMT), then set this domain groups cpu_power such that each group
6004 * can handle only one task, when there are other idle groups in the
6005 * same sched domain.
6007 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6009 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6010 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6015 * add cpu_power of each child group to this groups cpu_power
6017 group = child->groups;
6019 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6020 group = group->next;
6021 } while (group != child->groups);
6025 * Build sched domains for a given set of cpus and attach the sched domains
6026 * to the individual cpus
6028 static int build_sched_domains(const cpumask_t *cpu_map)
6032 struct sched_group **sched_group_nodes = NULL;
6033 int sd_allnodes = 0;
6036 * Allocate the per-node list of sched groups
6038 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
6040 if (!sched_group_nodes) {
6041 printk(KERN_WARNING "Can not alloc sched group node list\n");
6044 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6048 * Set up domains for cpus specified by the cpu_map.
6050 for_each_cpu_mask(i, *cpu_map) {
6051 struct sched_domain *sd = NULL, *p;
6052 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6054 cpus_and(nodemask, nodemask, *cpu_map);
6057 if (cpus_weight(*cpu_map) >
6058 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6059 sd = &per_cpu(allnodes_domains, i);
6060 *sd = SD_ALLNODES_INIT;
6061 sd->span = *cpu_map;
6062 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6068 sd = &per_cpu(node_domains, i);
6070 sd->span = sched_domain_node_span(cpu_to_node(i));
6074 cpus_and(sd->span, sd->span, *cpu_map);
6078 sd = &per_cpu(phys_domains, i);
6080 sd->span = nodemask;
6084 cpu_to_phys_group(i, cpu_map, &sd->groups);
6086 #ifdef CONFIG_SCHED_MC
6088 sd = &per_cpu(core_domains, i);
6090 sd->span = cpu_coregroup_map(i);
6091 cpus_and(sd->span, sd->span, *cpu_map);
6094 cpu_to_core_group(i, cpu_map, &sd->groups);
6097 #ifdef CONFIG_SCHED_SMT
6099 sd = &per_cpu(cpu_domains, i);
6100 *sd = SD_SIBLING_INIT;
6101 sd->span = cpu_sibling_map[i];
6102 cpus_and(sd->span, sd->span, *cpu_map);
6105 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6109 #ifdef CONFIG_SCHED_SMT
6110 /* Set up CPU (sibling) groups */
6111 for_each_cpu_mask(i, *cpu_map) {
6112 cpumask_t this_sibling_map = cpu_sibling_map[i];
6113 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6114 if (i != first_cpu(this_sibling_map))
6117 init_sched_build_groups(this_sibling_map, cpu_map,
6122 #ifdef CONFIG_SCHED_MC
6123 /* Set up multi-core groups */
6124 for_each_cpu_mask(i, *cpu_map) {
6125 cpumask_t this_core_map = cpu_coregroup_map(i);
6126 cpus_and(this_core_map, this_core_map, *cpu_map);
6127 if (i != first_cpu(this_core_map))
6129 init_sched_build_groups(this_core_map, cpu_map,
6130 &cpu_to_core_group);
6134 /* Set up physical groups */
6135 for (i = 0; i < MAX_NUMNODES; i++) {
6136 cpumask_t nodemask = node_to_cpumask(i);
6138 cpus_and(nodemask, nodemask, *cpu_map);
6139 if (cpus_empty(nodemask))
6142 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6146 /* Set up node groups */
6148 init_sched_build_groups(*cpu_map, cpu_map,
6149 &cpu_to_allnodes_group);
6151 for (i = 0; i < MAX_NUMNODES; i++) {
6152 /* Set up node groups */
6153 struct sched_group *sg, *prev;
6154 cpumask_t nodemask = node_to_cpumask(i);
6155 cpumask_t domainspan;
6156 cpumask_t covered = CPU_MASK_NONE;
6159 cpus_and(nodemask, nodemask, *cpu_map);
6160 if (cpus_empty(nodemask)) {
6161 sched_group_nodes[i] = NULL;
6165 domainspan = sched_domain_node_span(i);
6166 cpus_and(domainspan, domainspan, *cpu_map);
6168 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6170 printk(KERN_WARNING "Can not alloc domain group for "
6174 sched_group_nodes[i] = sg;
6175 for_each_cpu_mask(j, nodemask) {
6176 struct sched_domain *sd;
6178 sd = &per_cpu(node_domains, j);
6181 sg->__cpu_power = 0;
6182 sg->cpumask = nodemask;
6184 cpus_or(covered, covered, nodemask);
6187 for (j = 0; j < MAX_NUMNODES; j++) {
6188 cpumask_t tmp, notcovered;
6189 int n = (i + j) % MAX_NUMNODES;
6191 cpus_complement(notcovered, covered);
6192 cpus_and(tmp, notcovered, *cpu_map);
6193 cpus_and(tmp, tmp, domainspan);
6194 if (cpus_empty(tmp))
6197 nodemask = node_to_cpumask(n);
6198 cpus_and(tmp, tmp, nodemask);
6199 if (cpus_empty(tmp))
6202 sg = kmalloc_node(sizeof(struct sched_group),
6206 "Can not alloc domain group for node %d\n", j);
6209 sg->__cpu_power = 0;
6211 sg->next = prev->next;
6212 cpus_or(covered, covered, tmp);
6219 /* Calculate CPU power for physical packages and nodes */
6220 #ifdef CONFIG_SCHED_SMT
6221 for_each_cpu_mask(i, *cpu_map) {
6222 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6224 init_sched_groups_power(i, sd);
6227 #ifdef CONFIG_SCHED_MC
6228 for_each_cpu_mask(i, *cpu_map) {
6229 struct sched_domain *sd = &per_cpu(core_domains, i);
6231 init_sched_groups_power(i, sd);
6235 for_each_cpu_mask(i, *cpu_map) {
6236 struct sched_domain *sd = &per_cpu(phys_domains, i);
6238 init_sched_groups_power(i, sd);
6242 for (i = 0; i < MAX_NUMNODES; i++)
6243 init_numa_sched_groups_power(sched_group_nodes[i]);
6246 struct sched_group *sg;
6248 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6249 init_numa_sched_groups_power(sg);
6253 /* Attach the domains */
6254 for_each_cpu_mask(i, *cpu_map) {
6255 struct sched_domain *sd;
6256 #ifdef CONFIG_SCHED_SMT
6257 sd = &per_cpu(cpu_domains, i);
6258 #elif defined(CONFIG_SCHED_MC)
6259 sd = &per_cpu(core_domains, i);
6261 sd = &per_cpu(phys_domains, i);
6263 cpu_attach_domain(sd, i);
6270 free_sched_groups(cpu_map);
6275 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6277 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6279 cpumask_t cpu_default_map;
6283 * Setup mask for cpus without special case scheduling requirements.
6284 * For now this just excludes isolated cpus, but could be used to
6285 * exclude other special cases in the future.
6287 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6289 err = build_sched_domains(&cpu_default_map);
6294 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6296 free_sched_groups(cpu_map);
6300 * Detach sched domains from a group of cpus specified in cpu_map
6301 * These cpus will now be attached to the NULL domain
6303 static void detach_destroy_domains(const cpumask_t *cpu_map)
6307 for_each_cpu_mask(i, *cpu_map)
6308 cpu_attach_domain(NULL, i);
6309 synchronize_sched();
6310 arch_destroy_sched_domains(cpu_map);
6314 * Partition sched domains as specified by the cpumasks below.
6315 * This attaches all cpus from the cpumasks to the NULL domain,
6316 * waits for a RCU quiescent period, recalculates sched
6317 * domain information and then attaches them back to the
6318 * correct sched domains
6319 * Call with hotplug lock held
6321 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6323 cpumask_t change_map;
6326 cpus_and(*partition1, *partition1, cpu_online_map);
6327 cpus_and(*partition2, *partition2, cpu_online_map);
6328 cpus_or(change_map, *partition1, *partition2);
6330 /* Detach sched domains from all of the affected cpus */
6331 detach_destroy_domains(&change_map);
6332 if (!cpus_empty(*partition1))
6333 err = build_sched_domains(partition1);
6334 if (!err && !cpus_empty(*partition2))
6335 err = build_sched_domains(partition2);
6340 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6341 static int arch_reinit_sched_domains(void)
6345 mutex_lock(&sched_hotcpu_mutex);
6346 detach_destroy_domains(&cpu_online_map);
6347 err = arch_init_sched_domains(&cpu_online_map);
6348 mutex_unlock(&sched_hotcpu_mutex);
6353 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6357 if (buf[0] != '0' && buf[0] != '1')
6361 sched_smt_power_savings = (buf[0] == '1');
6363 sched_mc_power_savings = (buf[0] == '1');
6365 ret = arch_reinit_sched_domains();
6367 return ret ? ret : count;
6370 #ifdef CONFIG_SCHED_MC
6371 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6373 return sprintf(page, "%u\n", sched_mc_power_savings);
6375 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6376 const char *buf, size_t count)
6378 return sched_power_savings_store(buf, count, 0);
6380 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6381 sched_mc_power_savings_store);
6384 #ifdef CONFIG_SCHED_SMT
6385 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6387 return sprintf(page, "%u\n", sched_smt_power_savings);
6389 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6390 const char *buf, size_t count)
6392 return sched_power_savings_store(buf, count, 1);
6394 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6395 sched_smt_power_savings_store);
6398 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6402 #ifdef CONFIG_SCHED_SMT
6404 err = sysfs_create_file(&cls->kset.kobj,
6405 &attr_sched_smt_power_savings.attr);
6407 #ifdef CONFIG_SCHED_MC
6408 if (!err && mc_capable())
6409 err = sysfs_create_file(&cls->kset.kobj,
6410 &attr_sched_mc_power_savings.attr);
6417 * Force a reinitialization of the sched domains hierarchy. The domains
6418 * and groups cannot be updated in place without racing with the balancing
6419 * code, so we temporarily attach all running cpus to the NULL domain
6420 * which will prevent rebalancing while the sched domains are recalculated.
6422 static int update_sched_domains(struct notifier_block *nfb,
6423 unsigned long action, void *hcpu)
6426 case CPU_UP_PREPARE:
6427 case CPU_UP_PREPARE_FROZEN:
6428 case CPU_DOWN_PREPARE:
6429 case CPU_DOWN_PREPARE_FROZEN:
6430 detach_destroy_domains(&cpu_online_map);
6433 case CPU_UP_CANCELED:
6434 case CPU_UP_CANCELED_FROZEN:
6435 case CPU_DOWN_FAILED:
6436 case CPU_DOWN_FAILED_FROZEN:
6438 case CPU_ONLINE_FROZEN:
6440 case CPU_DEAD_FROZEN:
6442 * Fall through and re-initialise the domains.
6449 /* The hotplug lock is already held by cpu_up/cpu_down */
6450 arch_init_sched_domains(&cpu_online_map);
6455 void __init sched_init_smp(void)
6457 cpumask_t non_isolated_cpus;
6459 mutex_lock(&sched_hotcpu_mutex);
6460 arch_init_sched_domains(&cpu_online_map);
6461 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6462 if (cpus_empty(non_isolated_cpus))
6463 cpu_set(smp_processor_id(), non_isolated_cpus);
6464 mutex_unlock(&sched_hotcpu_mutex);
6465 /* XXX: Theoretical race here - CPU may be hotplugged now */
6466 hotcpu_notifier(update_sched_domains, 0);
6468 init_sched_domain_sysctl();
6470 /* Move init over to a non-isolated CPU */
6471 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6475 void __init sched_init_smp(void)
6478 #endif /* CONFIG_SMP */
6480 int in_sched_functions(unsigned long addr)
6482 /* Linker adds these: start and end of __sched functions */
6483 extern char __sched_text_start[], __sched_text_end[];
6485 return in_lock_functions(addr) ||
6486 (addr >= (unsigned long)__sched_text_start
6487 && addr < (unsigned long)__sched_text_end);
6490 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6492 cfs_rq->tasks_timeline = RB_ROOT;
6493 #ifdef CONFIG_FAIR_GROUP_SCHED
6496 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6499 void __init sched_init(void)
6501 int highest_cpu = 0;
6504 for_each_possible_cpu(i) {
6505 struct rt_prio_array *array;
6509 spin_lock_init(&rq->lock);
6510 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6513 init_cfs_rq(&rq->cfs, rq);
6514 #ifdef CONFIG_FAIR_GROUP_SCHED
6515 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6517 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6518 struct sched_entity *se =
6519 &per_cpu(init_sched_entity, i);
6521 init_cfs_rq_p[i] = cfs_rq;
6522 init_cfs_rq(cfs_rq, rq);
6523 cfs_rq->tg = &init_task_group;
6524 list_add(&cfs_rq->leaf_cfs_rq_list,
6525 &rq->leaf_cfs_rq_list);
6527 init_sched_entity_p[i] = se;
6528 se->cfs_rq = &rq->cfs;
6530 se->load.weight = init_task_group_load;
6531 se->load.inv_weight =
6532 div64_64(1ULL<<32, init_task_group_load);
6535 init_task_group.shares = init_task_group_load;
6538 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6539 rq->cpu_load[j] = 0;
6542 rq->active_balance = 0;
6543 rq->next_balance = jiffies;
6546 rq->migration_thread = NULL;
6547 INIT_LIST_HEAD(&rq->migration_queue);
6549 atomic_set(&rq->nr_iowait, 0);
6551 array = &rq->rt.active;
6552 for (j = 0; j < MAX_RT_PRIO; j++) {
6553 INIT_LIST_HEAD(array->queue + j);
6554 __clear_bit(j, array->bitmap);
6557 /* delimiter for bitsearch: */
6558 __set_bit(MAX_RT_PRIO, array->bitmap);
6561 set_load_weight(&init_task);
6563 #ifdef CONFIG_PREEMPT_NOTIFIERS
6564 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6568 nr_cpu_ids = highest_cpu + 1;
6569 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6572 #ifdef CONFIG_RT_MUTEXES
6573 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6577 * The boot idle thread does lazy MMU switching as well:
6579 atomic_inc(&init_mm.mm_count);
6580 enter_lazy_tlb(&init_mm, current);
6583 * Make us the idle thread. Technically, schedule() should not be
6584 * called from this thread, however somewhere below it might be,
6585 * but because we are the idle thread, we just pick up running again
6586 * when this runqueue becomes "idle".
6588 init_idle(current, smp_processor_id());
6590 * During early bootup we pretend to be a normal task:
6592 current->sched_class = &fair_sched_class;
6595 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6596 void __might_sleep(char *file, int line)
6599 static unsigned long prev_jiffy; /* ratelimiting */
6601 if ((in_atomic() || irqs_disabled()) &&
6602 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6603 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6605 prev_jiffy = jiffies;
6606 printk(KERN_ERR "BUG: sleeping function called from invalid"
6607 " context at %s:%d\n", file, line);
6608 printk("in_atomic():%d, irqs_disabled():%d\n",
6609 in_atomic(), irqs_disabled());
6610 debug_show_held_locks(current);
6611 if (irqs_disabled())
6612 print_irqtrace_events(current);
6617 EXPORT_SYMBOL(__might_sleep);
6620 #ifdef CONFIG_MAGIC_SYSRQ
6621 void normalize_rt_tasks(void)
6623 struct task_struct *g, *p;
6624 unsigned long flags;
6628 read_lock_irq(&tasklist_lock);
6629 do_each_thread(g, p) {
6630 p->se.exec_start = 0;
6631 #ifdef CONFIG_SCHEDSTATS
6632 p->se.wait_start = 0;
6633 p->se.sleep_start = 0;
6634 p->se.block_start = 0;
6636 task_rq(p)->clock = 0;
6640 * Renice negative nice level userspace
6643 if (TASK_NICE(p) < 0 && p->mm)
6644 set_user_nice(p, 0);
6648 spin_lock_irqsave(&p->pi_lock, flags);
6649 rq = __task_rq_lock(p);
6652 * Do not touch the migration thread:
6654 if (p == rq->migration_thread)
6658 update_rq_clock(rq);
6659 on_rq = p->se.on_rq;
6661 deactivate_task(rq, p, 0);
6662 __setscheduler(rq, p, SCHED_NORMAL, 0);
6664 activate_task(rq, p, 0);
6665 resched_task(rq->curr);
6670 __task_rq_unlock(rq);
6671 spin_unlock_irqrestore(&p->pi_lock, flags);
6672 } while_each_thread(g, p);
6674 read_unlock_irq(&tasklist_lock);
6677 #endif /* CONFIG_MAGIC_SYSRQ */
6681 * These functions are only useful for the IA64 MCA handling.
6683 * They can only be called when the whole system has been
6684 * stopped - every CPU needs to be quiescent, and no scheduling
6685 * activity can take place. Using them for anything else would
6686 * be a serious bug, and as a result, they aren't even visible
6687 * under any other configuration.
6691 * curr_task - return the current task for a given cpu.
6692 * @cpu: the processor in question.
6694 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6696 struct task_struct *curr_task(int cpu)
6698 return cpu_curr(cpu);
6702 * set_curr_task - set the current task for a given cpu.
6703 * @cpu: the processor in question.
6704 * @p: the task pointer to set.
6706 * Description: This function must only be used when non-maskable interrupts
6707 * are serviced on a separate stack. It allows the architecture to switch the
6708 * notion of the current task on a cpu in a non-blocking manner. This function
6709 * must be called with all CPU's synchronized, and interrupts disabled, the
6710 * and caller must save the original value of the current task (see
6711 * curr_task() above) and restore that value before reenabling interrupts and
6712 * re-starting the system.
6714 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6716 void set_curr_task(int cpu, struct task_struct *p)
6723 #ifdef CONFIG_FAIR_GROUP_SCHED
6725 /* allocate runqueue etc for a new task group */
6726 struct task_group *sched_create_group(void)
6728 struct task_group *tg;
6729 struct cfs_rq *cfs_rq;
6730 struct sched_entity *se;
6734 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6736 return ERR_PTR(-ENOMEM);
6738 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6741 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6745 for_each_possible_cpu(i) {
6748 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6753 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6758 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6759 memset(se, 0, sizeof(struct sched_entity));
6761 tg->cfs_rq[i] = cfs_rq;
6762 init_cfs_rq(cfs_rq, rq);
6766 se->cfs_rq = &rq->cfs;
6768 se->load.weight = NICE_0_LOAD;
6769 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
6773 for_each_possible_cpu(i) {
6775 cfs_rq = tg->cfs_rq[i];
6776 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6779 tg->shares = NICE_0_LOAD;
6784 for_each_possible_cpu(i) {
6786 kfree(tg->cfs_rq[i]);
6794 return ERR_PTR(-ENOMEM);
6797 /* rcu callback to free various structures associated with a task group */
6798 static void free_sched_group(struct rcu_head *rhp)
6800 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
6801 struct task_group *tg = cfs_rq->tg;
6802 struct sched_entity *se;
6805 /* now it should be safe to free those cfs_rqs */
6806 for_each_possible_cpu(i) {
6807 cfs_rq = tg->cfs_rq[i];
6819 /* Destroy runqueue etc associated with a task group */
6820 void sched_destroy_group(struct task_group *tg)
6822 struct cfs_rq *cfs_rq;
6825 for_each_possible_cpu(i) {
6826 cfs_rq = tg->cfs_rq[i];
6827 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
6830 cfs_rq = tg->cfs_rq[0];
6832 /* wait for possible concurrent references to cfs_rqs complete */
6833 call_rcu(&cfs_rq->rcu, free_sched_group);
6836 /* change task's runqueue when it moves between groups.
6837 * The caller of this function should have put the task in its new group
6838 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6839 * reflect its new group.
6841 void sched_move_task(struct task_struct *tsk)
6844 unsigned long flags;
6847 rq = task_rq_lock(tsk, &flags);
6849 if (tsk->sched_class != &fair_sched_class)
6852 update_rq_clock(rq);
6854 running = task_running(rq, tsk);
6855 on_rq = tsk->se.on_rq;
6858 dequeue_task(rq, tsk, 0);
6859 if (unlikely(running))
6860 tsk->sched_class->put_prev_task(rq, tsk);
6863 set_task_cfs_rq(tsk);
6866 if (unlikely(running))
6867 tsk->sched_class->set_curr_task(rq);
6868 enqueue_task(rq, tsk, 0);
6872 task_rq_unlock(rq, &flags);
6875 static void set_se_shares(struct sched_entity *se, unsigned long shares)
6877 struct cfs_rq *cfs_rq = se->cfs_rq;
6878 struct rq *rq = cfs_rq->rq;
6881 spin_lock_irq(&rq->lock);
6885 dequeue_entity(cfs_rq, se, 0);
6887 se->load.weight = shares;
6888 se->load.inv_weight = div64_64((1ULL<<32), shares);
6891 enqueue_entity(cfs_rq, se, 0);
6893 spin_unlock_irq(&rq->lock);
6896 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6900 if (tg->shares == shares)
6903 /* return -EINVAL if the new value is not sane */
6905 tg->shares = shares;
6906 for_each_possible_cpu(i)
6907 set_se_shares(tg->se[i], shares);
6912 #endif /* CONFIG_FAIR_GROUP_SCHED */