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;
165 /* spinlock to serialize modification to shares */
169 /* Default task group's sched entity on each cpu */
170 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
171 /* Default task group's cfs_rq on each cpu */
172 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
174 static struct sched_entity *init_sched_entity_p[NR_CPUS];
175 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
177 /* Default task group.
178 * Every task in system belong to this group at bootup.
180 struct task_group init_task_group = {
181 .se = init_sched_entity_p,
182 .cfs_rq = init_cfs_rq_p,
185 #ifdef CONFIG_FAIR_USER_SCHED
186 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
188 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
191 static int init_task_group_load = INIT_TASK_GRP_LOAD;
193 /* return group to which a task belongs */
194 static inline struct task_group *task_group(struct task_struct *p)
196 struct task_group *tg;
198 #ifdef CONFIG_FAIR_USER_SCHED
201 tg = &init_task_group;
207 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
208 static inline void set_task_cfs_rq(struct task_struct *p)
210 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
211 p->se.parent = task_group(p)->se[task_cpu(p)];
216 static inline void set_task_cfs_rq(struct task_struct *p) { }
218 #endif /* CONFIG_FAIR_GROUP_SCHED */
220 /* CFS-related fields in a runqueue */
222 struct load_weight load;
223 unsigned long nr_running;
228 struct rb_root tasks_timeline;
229 struct rb_node *rb_leftmost;
230 struct rb_node *rb_load_balance_curr;
231 /* 'curr' points to currently running entity on this cfs_rq.
232 * It is set to NULL otherwise (i.e when none are currently running).
234 struct sched_entity *curr;
236 unsigned long nr_spread_over;
238 #ifdef CONFIG_FAIR_GROUP_SCHED
239 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
241 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
242 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
243 * (like users, containers etc.)
245 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
246 * list is used during load balance.
248 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
249 struct task_group *tg; /* group that "owns" this runqueue */
254 /* Real-Time classes' related field in a runqueue: */
256 struct rt_prio_array active;
257 int rt_load_balance_idx;
258 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
262 * This is the main, per-CPU runqueue data structure.
264 * Locking rule: those places that want to lock multiple runqueues
265 * (such as the load balancing or the thread migration code), lock
266 * acquire operations must be ordered by ascending &runqueue.
269 spinlock_t lock; /* runqueue lock */
272 * nr_running and cpu_load should be in the same cacheline because
273 * remote CPUs use both these fields when doing load calculation.
275 unsigned long nr_running;
276 #define CPU_LOAD_IDX_MAX 5
277 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
278 unsigned char idle_at_tick;
280 unsigned char in_nohz_recently;
282 struct load_weight load; /* capture load from *all* tasks on this cpu */
283 unsigned long nr_load_updates;
287 #ifdef CONFIG_FAIR_GROUP_SCHED
288 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
293 * This is part of a global counter where only the total sum
294 * over all CPUs matters. A task can increase this counter on
295 * one CPU and if it got migrated afterwards it may decrease
296 * it on another CPU. Always updated under the runqueue lock:
298 unsigned long nr_uninterruptible;
300 struct task_struct *curr, *idle;
301 unsigned long next_balance;
302 struct mm_struct *prev_mm;
304 u64 clock, prev_clock_raw;
307 unsigned int clock_warps, clock_overflows;
309 unsigned int clock_deep_idle_events;
315 struct sched_domain *sd;
317 /* For active balancing */
320 int cpu; /* cpu of this runqueue */
322 struct task_struct *migration_thread;
323 struct list_head migration_queue;
326 #ifdef CONFIG_SCHEDSTATS
328 struct sched_info rq_sched_info;
330 /* sys_sched_yield() stats */
331 unsigned long yld_exp_empty;
332 unsigned long yld_act_empty;
333 unsigned long yld_both_empty;
334 unsigned long yld_count;
336 /* schedule() stats */
337 unsigned long sched_switch;
338 unsigned long sched_count;
339 unsigned long sched_goidle;
341 /* try_to_wake_up() stats */
342 unsigned long ttwu_count;
343 unsigned long ttwu_local;
346 unsigned long bkl_count;
348 struct lock_class_key rq_lock_key;
351 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
352 static DEFINE_MUTEX(sched_hotcpu_mutex);
354 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
356 rq->curr->sched_class->check_preempt_curr(rq, p);
359 static inline int cpu_of(struct rq *rq)
369 * Update the per-runqueue clock, as finegrained as the platform can give
370 * us, but without assuming monotonicity, etc.:
372 static void __update_rq_clock(struct rq *rq)
374 u64 prev_raw = rq->prev_clock_raw;
375 u64 now = sched_clock();
376 s64 delta = now - prev_raw;
377 u64 clock = rq->clock;
379 #ifdef CONFIG_SCHED_DEBUG
380 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
383 * Protect against sched_clock() occasionally going backwards:
385 if (unlikely(delta < 0)) {
390 * Catch too large forward jumps too:
392 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
393 if (clock < rq->tick_timestamp + TICK_NSEC)
394 clock = rq->tick_timestamp + TICK_NSEC;
397 rq->clock_overflows++;
399 if (unlikely(delta > rq->clock_max_delta))
400 rq->clock_max_delta = delta;
405 rq->prev_clock_raw = now;
409 static void update_rq_clock(struct rq *rq)
411 if (likely(smp_processor_id() == cpu_of(rq)))
412 __update_rq_clock(rq);
416 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
417 * See detach_destroy_domains: synchronize_sched for details.
419 * The domain tree of any CPU may only be accessed from within
420 * preempt-disabled sections.
422 #define for_each_domain(cpu, __sd) \
423 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
425 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
426 #define this_rq() (&__get_cpu_var(runqueues))
427 #define task_rq(p) cpu_rq(task_cpu(p))
428 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
431 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
433 #ifdef CONFIG_SCHED_DEBUG
434 # define const_debug __read_mostly
436 # define const_debug static const
440 * Debugging: various feature bits
443 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
444 SCHED_FEAT_START_DEBIT = 2,
445 SCHED_FEAT_TREE_AVG = 4,
446 SCHED_FEAT_APPROX_AVG = 8,
447 SCHED_FEAT_WAKEUP_PREEMPT = 16,
448 SCHED_FEAT_PREEMPT_RESTRICT = 32,
451 const_debug unsigned int sysctl_sched_features =
452 SCHED_FEAT_NEW_FAIR_SLEEPERS *1 |
453 SCHED_FEAT_START_DEBIT *1 |
454 SCHED_FEAT_TREE_AVG *0 |
455 SCHED_FEAT_APPROX_AVG *0 |
456 SCHED_FEAT_WAKEUP_PREEMPT *1 |
457 SCHED_FEAT_PREEMPT_RESTRICT *1;
459 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
462 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
463 * clock constructed from sched_clock():
465 unsigned long long cpu_clock(int cpu)
467 unsigned long long now;
471 local_irq_save(flags);
475 local_irq_restore(flags);
479 EXPORT_SYMBOL_GPL(cpu_clock);
481 #ifndef prepare_arch_switch
482 # define prepare_arch_switch(next) do { } while (0)
484 #ifndef finish_arch_switch
485 # define finish_arch_switch(prev) do { } while (0)
488 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
489 static inline int task_running(struct rq *rq, struct task_struct *p)
491 return rq->curr == p;
494 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
498 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
500 #ifdef CONFIG_DEBUG_SPINLOCK
501 /* this is a valid case when another task releases the spinlock */
502 rq->lock.owner = current;
505 * If we are tracking spinlock dependencies then we have to
506 * fix up the runqueue lock - which gets 'carried over' from
509 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
511 spin_unlock_irq(&rq->lock);
514 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
515 static inline int task_running(struct rq *rq, struct task_struct *p)
520 return rq->curr == p;
524 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
528 * We can optimise this out completely for !SMP, because the
529 * SMP rebalancing from interrupt is the only thing that cares
534 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
535 spin_unlock_irq(&rq->lock);
537 spin_unlock(&rq->lock);
541 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
545 * After ->oncpu is cleared, the task can be moved to a different CPU.
546 * We must ensure this doesn't happen until the switch is completely
552 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
556 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
559 * __task_rq_lock - lock the runqueue a given task resides on.
560 * Must be called interrupts disabled.
562 static inline struct rq *__task_rq_lock(struct task_struct *p)
566 struct rq *rq = task_rq(p);
567 spin_lock(&rq->lock);
568 if (likely(rq == task_rq(p)))
570 spin_unlock(&rq->lock);
575 * task_rq_lock - lock the runqueue a given task resides on and disable
576 * interrupts. Note the ordering: we can safely lookup the task_rq without
577 * explicitly disabling preemption.
579 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
585 local_irq_save(*flags);
587 spin_lock(&rq->lock);
588 if (likely(rq == task_rq(p)))
590 spin_unlock_irqrestore(&rq->lock, *flags);
594 static void __task_rq_unlock(struct rq *rq)
597 spin_unlock(&rq->lock);
600 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
603 spin_unlock_irqrestore(&rq->lock, *flags);
607 * this_rq_lock - lock this runqueue and disable interrupts.
609 static struct rq *this_rq_lock(void)
616 spin_lock(&rq->lock);
622 * We are going deep-idle (irqs are disabled):
624 void sched_clock_idle_sleep_event(void)
626 struct rq *rq = cpu_rq(smp_processor_id());
628 spin_lock(&rq->lock);
629 __update_rq_clock(rq);
630 spin_unlock(&rq->lock);
631 rq->clock_deep_idle_events++;
633 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
636 * We just idled delta nanoseconds (called with irqs disabled):
638 void sched_clock_idle_wakeup_event(u64 delta_ns)
640 struct rq *rq = cpu_rq(smp_processor_id());
641 u64 now = sched_clock();
643 rq->idle_clock += delta_ns;
645 * Override the previous timestamp and ignore all
646 * sched_clock() deltas that occured while we idled,
647 * and use the PM-provided delta_ns to advance the
650 spin_lock(&rq->lock);
651 rq->prev_clock_raw = now;
652 rq->clock += delta_ns;
653 spin_unlock(&rq->lock);
655 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
658 * resched_task - mark a task 'to be rescheduled now'.
660 * On UP this means the setting of the need_resched flag, on SMP it
661 * might also involve a cross-CPU call to trigger the scheduler on
666 #ifndef tsk_is_polling
667 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
670 static void resched_task(struct task_struct *p)
674 assert_spin_locked(&task_rq(p)->lock);
676 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
679 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
682 if (cpu == smp_processor_id())
685 /* NEED_RESCHED must be visible before we test polling */
687 if (!tsk_is_polling(p))
688 smp_send_reschedule(cpu);
691 static void resched_cpu(int cpu)
693 struct rq *rq = cpu_rq(cpu);
696 if (!spin_trylock_irqsave(&rq->lock, flags))
698 resched_task(cpu_curr(cpu));
699 spin_unlock_irqrestore(&rq->lock, flags);
702 static inline void resched_task(struct task_struct *p)
704 assert_spin_locked(&task_rq(p)->lock);
705 set_tsk_need_resched(p);
709 #if BITS_PER_LONG == 32
710 # define WMULT_CONST (~0UL)
712 # define WMULT_CONST (1UL << 32)
715 #define WMULT_SHIFT 32
718 * Shift right and round:
720 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
723 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
724 struct load_weight *lw)
728 if (unlikely(!lw->inv_weight))
729 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
731 tmp = (u64)delta_exec * weight;
733 * Check whether we'd overflow the 64-bit multiplication:
735 if (unlikely(tmp > WMULT_CONST))
736 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
739 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
741 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
744 static inline unsigned long
745 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
747 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
750 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
755 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
761 * To aid in avoiding the subversion of "niceness" due to uneven distribution
762 * of tasks with abnormal "nice" values across CPUs the contribution that
763 * each task makes to its run queue's load is weighted according to its
764 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
765 * scaled version of the new time slice allocation that they receive on time
769 #define WEIGHT_IDLEPRIO 2
770 #define WMULT_IDLEPRIO (1 << 31)
773 * Nice levels are multiplicative, with a gentle 10% change for every
774 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
775 * nice 1, it will get ~10% less CPU time than another CPU-bound task
776 * that remained on nice 0.
778 * The "10% effect" is relative and cumulative: from _any_ nice level,
779 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
780 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
781 * If a task goes up by ~10% and another task goes down by ~10% then
782 * the relative distance between them is ~25%.)
784 static const int prio_to_weight[40] = {
785 /* -20 */ 88761, 71755, 56483, 46273, 36291,
786 /* -15 */ 29154, 23254, 18705, 14949, 11916,
787 /* -10 */ 9548, 7620, 6100, 4904, 3906,
788 /* -5 */ 3121, 2501, 1991, 1586, 1277,
789 /* 0 */ 1024, 820, 655, 526, 423,
790 /* 5 */ 335, 272, 215, 172, 137,
791 /* 10 */ 110, 87, 70, 56, 45,
792 /* 15 */ 36, 29, 23, 18, 15,
796 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
798 * In cases where the weight does not change often, we can use the
799 * precalculated inverse to speed up arithmetics by turning divisions
800 * into multiplications:
802 static const u32 prio_to_wmult[40] = {
803 /* -20 */ 48388, 59856, 76040, 92818, 118348,
804 /* -15 */ 147320, 184698, 229616, 287308, 360437,
805 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
806 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
807 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
808 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
809 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
810 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
813 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
816 * runqueue iterator, to support SMP load-balancing between different
817 * scheduling classes, without having to expose their internal data
818 * structures to the load-balancing proper:
822 struct task_struct *(*start)(void *);
823 struct task_struct *(*next)(void *);
826 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
827 unsigned long max_nr_move, unsigned long max_load_move,
828 struct sched_domain *sd, enum cpu_idle_type idle,
829 int *all_pinned, unsigned long *load_moved,
830 int *this_best_prio, struct rq_iterator *iterator);
832 #include "sched_stats.h"
833 #include "sched_idletask.c"
834 #include "sched_fair.c"
835 #include "sched_rt.c"
836 #ifdef CONFIG_SCHED_DEBUG
837 # include "sched_debug.c"
840 #define sched_class_highest (&rt_sched_class)
843 * Update delta_exec, delta_fair fields for rq.
845 * delta_fair clock advances at a rate inversely proportional to
846 * total load (rq->load.weight) on the runqueue, while
847 * delta_exec advances at the same rate as wall-clock (provided
850 * delta_exec / delta_fair is a measure of the (smoothened) load on this
851 * runqueue over any given interval. This (smoothened) load is used
852 * during load balance.
854 * This function is called /before/ updating rq->load
855 * and when switching tasks.
857 static inline void inc_load(struct rq *rq, const struct task_struct *p)
859 update_load_add(&rq->load, p->se.load.weight);
862 static inline void dec_load(struct rq *rq, const struct task_struct *p)
864 update_load_sub(&rq->load, p->se.load.weight);
867 static void inc_nr_running(struct task_struct *p, struct rq *rq)
873 static void dec_nr_running(struct task_struct *p, struct rq *rq)
879 static void set_load_weight(struct task_struct *p)
881 if (task_has_rt_policy(p)) {
882 p->se.load.weight = prio_to_weight[0] * 2;
883 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
888 * SCHED_IDLE tasks get minimal weight:
890 if (p->policy == SCHED_IDLE) {
891 p->se.load.weight = WEIGHT_IDLEPRIO;
892 p->se.load.inv_weight = WMULT_IDLEPRIO;
896 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
897 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
900 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
902 sched_info_queued(p);
903 p->sched_class->enqueue_task(rq, p, wakeup);
907 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
909 p->sched_class->dequeue_task(rq, p, sleep);
914 * __normal_prio - return the priority that is based on the static prio
916 static inline int __normal_prio(struct task_struct *p)
918 return p->static_prio;
922 * Calculate the expected normal priority: i.e. priority
923 * without taking RT-inheritance into account. Might be
924 * boosted by interactivity modifiers. Changes upon fork,
925 * setprio syscalls, and whenever the interactivity
926 * estimator recalculates.
928 static inline int normal_prio(struct task_struct *p)
932 if (task_has_rt_policy(p))
933 prio = MAX_RT_PRIO-1 - p->rt_priority;
935 prio = __normal_prio(p);
940 * Calculate the current priority, i.e. the priority
941 * taken into account by the scheduler. This value might
942 * be boosted by RT tasks, or might be boosted by
943 * interactivity modifiers. Will be RT if the task got
944 * RT-boosted. If not then it returns p->normal_prio.
946 static int effective_prio(struct task_struct *p)
948 p->normal_prio = normal_prio(p);
950 * If we are RT tasks or we were boosted to RT priority,
951 * keep the priority unchanged. Otherwise, update priority
952 * to the normal priority:
954 if (!rt_prio(p->prio))
955 return p->normal_prio;
960 * activate_task - move a task to the runqueue.
962 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
964 if (p->state == TASK_UNINTERRUPTIBLE)
965 rq->nr_uninterruptible--;
967 enqueue_task(rq, p, wakeup);
968 inc_nr_running(p, rq);
972 * deactivate_task - remove a task from the runqueue.
974 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
976 if (p->state == TASK_UNINTERRUPTIBLE)
977 rq->nr_uninterruptible++;
979 dequeue_task(rq, p, sleep);
980 dec_nr_running(p, rq);
984 * task_curr - is this task currently executing on a CPU?
985 * @p: the task in question.
987 inline int task_curr(const struct task_struct *p)
989 return cpu_curr(task_cpu(p)) == p;
992 /* Used instead of source_load when we know the type == 0 */
993 unsigned long weighted_cpuload(const int cpu)
995 return cpu_rq(cpu)->load.weight;
998 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1001 task_thread_info(p)->cpu = cpu;
1009 * Is this task likely cache-hot:
1012 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1016 if (p->sched_class != &fair_sched_class)
1019 if (sysctl_sched_migration_cost == -1)
1021 if (sysctl_sched_migration_cost == 0)
1024 delta = now - p->se.exec_start;
1026 return delta < (s64)sysctl_sched_migration_cost;
1030 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1032 int old_cpu = task_cpu(p);
1033 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1034 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1035 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1038 clock_offset = old_rq->clock - new_rq->clock;
1040 #ifdef CONFIG_SCHEDSTATS
1041 if (p->se.wait_start)
1042 p->se.wait_start -= clock_offset;
1043 if (p->se.sleep_start)
1044 p->se.sleep_start -= clock_offset;
1045 if (p->se.block_start)
1046 p->se.block_start -= clock_offset;
1047 if (old_cpu != new_cpu) {
1048 schedstat_inc(p, se.nr_migrations);
1049 if (task_hot(p, old_rq->clock, NULL))
1050 schedstat_inc(p, se.nr_forced2_migrations);
1053 p->se.vruntime -= old_cfsrq->min_vruntime -
1054 new_cfsrq->min_vruntime;
1056 __set_task_cpu(p, new_cpu);
1059 struct migration_req {
1060 struct list_head list;
1062 struct task_struct *task;
1065 struct completion done;
1069 * The task's runqueue lock must be held.
1070 * Returns true if you have to wait for migration thread.
1073 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1075 struct rq *rq = task_rq(p);
1078 * If the task is not on a runqueue (and not running), then
1079 * it is sufficient to simply update the task's cpu field.
1081 if (!p->se.on_rq && !task_running(rq, p)) {
1082 set_task_cpu(p, dest_cpu);
1086 init_completion(&req->done);
1088 req->dest_cpu = dest_cpu;
1089 list_add(&req->list, &rq->migration_queue);
1095 * wait_task_inactive - wait for a thread to unschedule.
1097 * The caller must ensure that the task *will* unschedule sometime soon,
1098 * else this function might spin for a *long* time. This function can't
1099 * be called with interrupts off, or it may introduce deadlock with
1100 * smp_call_function() if an IPI is sent by the same process we are
1101 * waiting to become inactive.
1103 void wait_task_inactive(struct task_struct *p)
1105 unsigned long flags;
1111 * We do the initial early heuristics without holding
1112 * any task-queue locks at all. We'll only try to get
1113 * the runqueue lock when things look like they will
1119 * If the task is actively running on another CPU
1120 * still, just relax and busy-wait without holding
1123 * NOTE! Since we don't hold any locks, it's not
1124 * even sure that "rq" stays as the right runqueue!
1125 * But we don't care, since "task_running()" will
1126 * return false if the runqueue has changed and p
1127 * is actually now running somewhere else!
1129 while (task_running(rq, p))
1133 * Ok, time to look more closely! We need the rq
1134 * lock now, to be *sure*. If we're wrong, we'll
1135 * just go back and repeat.
1137 rq = task_rq_lock(p, &flags);
1138 running = task_running(rq, p);
1139 on_rq = p->se.on_rq;
1140 task_rq_unlock(rq, &flags);
1143 * Was it really running after all now that we
1144 * checked with the proper locks actually held?
1146 * Oops. Go back and try again..
1148 if (unlikely(running)) {
1154 * It's not enough that it's not actively running,
1155 * it must be off the runqueue _entirely_, and not
1158 * So if it wa still runnable (but just not actively
1159 * running right now), it's preempted, and we should
1160 * yield - it could be a while.
1162 if (unlikely(on_rq)) {
1163 schedule_timeout_uninterruptible(1);
1168 * Ahh, all good. It wasn't running, and it wasn't
1169 * runnable, which means that it will never become
1170 * running in the future either. We're all done!
1177 * kick_process - kick a running thread to enter/exit the kernel
1178 * @p: the to-be-kicked thread
1180 * Cause a process which is running on another CPU to enter
1181 * kernel-mode, without any delay. (to get signals handled.)
1183 * NOTE: this function doesnt have to take the runqueue lock,
1184 * because all it wants to ensure is that the remote task enters
1185 * the kernel. If the IPI races and the task has been migrated
1186 * to another CPU then no harm is done and the purpose has been
1189 void kick_process(struct task_struct *p)
1195 if ((cpu != smp_processor_id()) && task_curr(p))
1196 smp_send_reschedule(cpu);
1201 * Return a low guess at the load of a migration-source cpu weighted
1202 * according to the scheduling class and "nice" value.
1204 * We want to under-estimate the load of migration sources, to
1205 * balance conservatively.
1207 static unsigned long source_load(int cpu, int type)
1209 struct rq *rq = cpu_rq(cpu);
1210 unsigned long total = weighted_cpuload(cpu);
1215 return min(rq->cpu_load[type-1], total);
1219 * Return a high guess at the load of a migration-target cpu weighted
1220 * according to the scheduling class and "nice" value.
1222 static unsigned long target_load(int cpu, int type)
1224 struct rq *rq = cpu_rq(cpu);
1225 unsigned long total = weighted_cpuload(cpu);
1230 return max(rq->cpu_load[type-1], total);
1234 * Return the average load per task on the cpu's run queue
1236 static inline unsigned long cpu_avg_load_per_task(int cpu)
1238 struct rq *rq = cpu_rq(cpu);
1239 unsigned long total = weighted_cpuload(cpu);
1240 unsigned long n = rq->nr_running;
1242 return n ? total / n : SCHED_LOAD_SCALE;
1246 * find_idlest_group finds and returns the least busy CPU group within the
1249 static struct sched_group *
1250 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1252 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1253 unsigned long min_load = ULONG_MAX, this_load = 0;
1254 int load_idx = sd->forkexec_idx;
1255 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1258 unsigned long load, avg_load;
1262 /* Skip over this group if it has no CPUs allowed */
1263 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1266 local_group = cpu_isset(this_cpu, group->cpumask);
1268 /* Tally up the load of all CPUs in the group */
1271 for_each_cpu_mask(i, group->cpumask) {
1272 /* Bias balancing toward cpus of our domain */
1274 load = source_load(i, load_idx);
1276 load = target_load(i, load_idx);
1281 /* Adjust by relative CPU power of the group */
1282 avg_load = sg_div_cpu_power(group,
1283 avg_load * SCHED_LOAD_SCALE);
1286 this_load = avg_load;
1288 } else if (avg_load < min_load) {
1289 min_load = avg_load;
1292 } while (group = group->next, group != sd->groups);
1294 if (!idlest || 100*this_load < imbalance*min_load)
1300 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1303 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1306 unsigned long load, min_load = ULONG_MAX;
1310 /* Traverse only the allowed CPUs */
1311 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1313 for_each_cpu_mask(i, tmp) {
1314 load = weighted_cpuload(i);
1316 if (load < min_load || (load == min_load && i == this_cpu)) {
1326 * sched_balance_self: balance the current task (running on cpu) in domains
1327 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1330 * Balance, ie. select the least loaded group.
1332 * Returns the target CPU number, or the same CPU if no balancing is needed.
1334 * preempt must be disabled.
1336 static int sched_balance_self(int cpu, int flag)
1338 struct task_struct *t = current;
1339 struct sched_domain *tmp, *sd = NULL;
1341 for_each_domain(cpu, tmp) {
1343 * If power savings logic is enabled for a domain, stop there.
1345 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1347 if (tmp->flags & flag)
1353 struct sched_group *group;
1354 int new_cpu, weight;
1356 if (!(sd->flags & flag)) {
1362 group = find_idlest_group(sd, t, cpu);
1368 new_cpu = find_idlest_cpu(group, t, cpu);
1369 if (new_cpu == -1 || new_cpu == cpu) {
1370 /* Now try balancing at a lower domain level of cpu */
1375 /* Now try balancing at a lower domain level of new_cpu */
1378 weight = cpus_weight(span);
1379 for_each_domain(cpu, tmp) {
1380 if (weight <= cpus_weight(tmp->span))
1382 if (tmp->flags & flag)
1385 /* while loop will break here if sd == NULL */
1391 #endif /* CONFIG_SMP */
1394 * wake_idle() will wake a task on an idle cpu if task->cpu is
1395 * not idle and an idle cpu is available. The span of cpus to
1396 * search starts with cpus closest then further out as needed,
1397 * so we always favor a closer, idle cpu.
1399 * Returns the CPU we should wake onto.
1401 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1402 static int wake_idle(int cpu, struct task_struct *p)
1405 struct sched_domain *sd;
1409 * If it is idle, then it is the best cpu to run this task.
1411 * This cpu is also the best, if it has more than one task already.
1412 * Siblings must be also busy(in most cases) as they didn't already
1413 * pickup the extra load from this cpu and hence we need not check
1414 * sibling runqueue info. This will avoid the checks and cache miss
1415 * penalities associated with that.
1417 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1420 for_each_domain(cpu, sd) {
1421 if (sd->flags & SD_WAKE_IDLE) {
1422 cpus_and(tmp, sd->span, p->cpus_allowed);
1423 for_each_cpu_mask(i, tmp) {
1425 if (i != task_cpu(p)) {
1427 se.nr_wakeups_idle);
1439 static inline int wake_idle(int cpu, struct task_struct *p)
1446 * try_to_wake_up - wake up a thread
1447 * @p: the to-be-woken-up thread
1448 * @state: the mask of task states that can be woken
1449 * @sync: do a synchronous wakeup?
1451 * Put it on the run-queue if it's not already there. The "current"
1452 * thread is always on the run-queue (except when the actual
1453 * re-schedule is in progress), and as such you're allowed to do
1454 * the simpler "current->state = TASK_RUNNING" to mark yourself
1455 * runnable without the overhead of this.
1457 * returns failure only if the task is already active.
1459 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1461 int cpu, orig_cpu, this_cpu, success = 0;
1462 unsigned long flags;
1466 struct sched_domain *sd, *this_sd = NULL;
1467 unsigned long load, this_load;
1471 rq = task_rq_lock(p, &flags);
1472 old_state = p->state;
1473 if (!(old_state & state))
1481 this_cpu = smp_processor_id();
1484 if (unlikely(task_running(rq, p)))
1489 schedstat_inc(rq, ttwu_count);
1490 if (cpu == this_cpu) {
1491 schedstat_inc(rq, ttwu_local);
1495 for_each_domain(this_cpu, sd) {
1496 if (cpu_isset(cpu, sd->span)) {
1497 schedstat_inc(sd, ttwu_wake_remote);
1503 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1507 * Check for affine wakeup and passive balancing possibilities.
1510 int idx = this_sd->wake_idx;
1511 unsigned int imbalance;
1513 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1515 load = source_load(cpu, idx);
1516 this_load = target_load(this_cpu, idx);
1518 new_cpu = this_cpu; /* Wake to this CPU if we can */
1520 if (this_sd->flags & SD_WAKE_AFFINE) {
1521 unsigned long tl = this_load;
1522 unsigned long tl_per_task;
1524 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1525 tl_per_task = cpu_avg_load_per_task(this_cpu);
1528 * If sync wakeup then subtract the (maximum possible)
1529 * effect of the currently running task from the load
1530 * of the current CPU:
1533 tl -= current->se.load.weight;
1536 tl + target_load(cpu, idx) <= tl_per_task) ||
1537 100*(tl + p->se.load.weight) <= imbalance*load) {
1539 * This domain has SD_WAKE_AFFINE and
1540 * p is cache cold in this domain, and
1541 * there is no bad imbalance.
1543 schedstat_inc(this_sd, ttwu_move_affine);
1544 schedstat_inc(p, se.nr_wakeups_affine);
1550 * Start passive balancing when half the imbalance_pct
1553 if (this_sd->flags & SD_WAKE_BALANCE) {
1554 if (imbalance*this_load <= 100*load) {
1555 schedstat_inc(this_sd, ttwu_move_balance);
1556 schedstat_inc(p, se.nr_wakeups_passive);
1562 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1564 new_cpu = wake_idle(new_cpu, p);
1565 if (new_cpu != cpu) {
1566 set_task_cpu(p, new_cpu);
1567 task_rq_unlock(rq, &flags);
1568 /* might preempt at this point */
1569 rq = task_rq_lock(p, &flags);
1570 old_state = p->state;
1571 if (!(old_state & state))
1576 this_cpu = smp_processor_id();
1581 #endif /* CONFIG_SMP */
1582 schedstat_inc(p, se.nr_wakeups);
1584 schedstat_inc(p, se.nr_wakeups_sync);
1585 if (orig_cpu != cpu)
1586 schedstat_inc(p, se.nr_wakeups_migrate);
1587 if (cpu == this_cpu)
1588 schedstat_inc(p, se.nr_wakeups_local);
1590 schedstat_inc(p, se.nr_wakeups_remote);
1591 update_rq_clock(rq);
1592 activate_task(rq, p, 1);
1594 * Sync wakeups (i.e. those types of wakeups where the waker
1595 * has indicated that it will leave the CPU in short order)
1596 * don't trigger a preemption, if the woken up task will run on
1597 * this cpu. (in this case the 'I will reschedule' promise of
1598 * the waker guarantees that the freshly woken up task is going
1599 * to be considered on this CPU.)
1601 if (!sync || cpu != this_cpu)
1602 check_preempt_curr(rq, p);
1606 p->state = TASK_RUNNING;
1608 task_rq_unlock(rq, &flags);
1613 int fastcall wake_up_process(struct task_struct *p)
1615 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1616 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1618 EXPORT_SYMBOL(wake_up_process);
1620 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1622 return try_to_wake_up(p, state, 0);
1626 * Perform scheduler related setup for a newly forked process p.
1627 * p is forked by current.
1629 * __sched_fork() is basic setup used by init_idle() too:
1631 static void __sched_fork(struct task_struct *p)
1633 p->se.exec_start = 0;
1634 p->se.sum_exec_runtime = 0;
1635 p->se.prev_sum_exec_runtime = 0;
1637 #ifdef CONFIG_SCHEDSTATS
1638 p->se.wait_start = 0;
1639 p->se.sum_sleep_runtime = 0;
1640 p->se.sleep_start = 0;
1641 p->se.block_start = 0;
1642 p->se.sleep_max = 0;
1643 p->se.block_max = 0;
1645 p->se.slice_max = 0;
1649 INIT_LIST_HEAD(&p->run_list);
1652 #ifdef CONFIG_PREEMPT_NOTIFIERS
1653 INIT_HLIST_HEAD(&p->preempt_notifiers);
1657 * We mark the process as running here, but have not actually
1658 * inserted it onto the runqueue yet. This guarantees that
1659 * nobody will actually run it, and a signal or other external
1660 * event cannot wake it up and insert it on the runqueue either.
1662 p->state = TASK_RUNNING;
1666 * fork()/clone()-time setup:
1668 void sched_fork(struct task_struct *p, int clone_flags)
1670 int cpu = get_cpu();
1675 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1677 set_task_cpu(p, cpu);
1680 * Make sure we do not leak PI boosting priority to the child:
1682 p->prio = current->normal_prio;
1683 if (!rt_prio(p->prio))
1684 p->sched_class = &fair_sched_class;
1686 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1687 if (likely(sched_info_on()))
1688 memset(&p->sched_info, 0, sizeof(p->sched_info));
1690 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1693 #ifdef CONFIG_PREEMPT
1694 /* Want to start with kernel preemption disabled. */
1695 task_thread_info(p)->preempt_count = 1;
1701 * wake_up_new_task - wake up a newly created task for the first time.
1703 * This function will do some initial scheduler statistics housekeeping
1704 * that must be done for every newly created context, then puts the task
1705 * on the runqueue and wakes it.
1707 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1709 unsigned long flags;
1712 rq = task_rq_lock(p, &flags);
1713 BUG_ON(p->state != TASK_RUNNING);
1714 update_rq_clock(rq);
1716 p->prio = effective_prio(p);
1718 if (!p->sched_class->task_new || !current->se.on_rq || !rq->cfs.curr) {
1719 activate_task(rq, p, 0);
1722 * Let the scheduling class do new task startup
1723 * management (if any):
1725 p->sched_class->task_new(rq, p);
1726 inc_nr_running(p, rq);
1728 check_preempt_curr(rq, p);
1729 task_rq_unlock(rq, &flags);
1732 #ifdef CONFIG_PREEMPT_NOTIFIERS
1735 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1736 * @notifier: notifier struct to register
1738 void preempt_notifier_register(struct preempt_notifier *notifier)
1740 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1742 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1745 * preempt_notifier_unregister - no longer interested in preemption notifications
1746 * @notifier: notifier struct to unregister
1748 * This is safe to call from within a preemption notifier.
1750 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1752 hlist_del(¬ifier->link);
1754 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1756 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1758 struct preempt_notifier *notifier;
1759 struct hlist_node *node;
1761 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1762 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1766 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1767 struct task_struct *next)
1769 struct preempt_notifier *notifier;
1770 struct hlist_node *node;
1772 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1773 notifier->ops->sched_out(notifier, next);
1778 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1783 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1784 struct task_struct *next)
1791 * prepare_task_switch - prepare to switch tasks
1792 * @rq: the runqueue preparing to switch
1793 * @prev: the current task that is being switched out
1794 * @next: the task we are going to switch to.
1796 * This is called with the rq lock held and interrupts off. It must
1797 * be paired with a subsequent finish_task_switch after the context
1800 * prepare_task_switch sets up locking and calls architecture specific
1804 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1805 struct task_struct *next)
1807 fire_sched_out_preempt_notifiers(prev, next);
1808 prepare_lock_switch(rq, next);
1809 prepare_arch_switch(next);
1813 * finish_task_switch - clean up after a task-switch
1814 * @rq: runqueue associated with task-switch
1815 * @prev: the thread we just switched away from.
1817 * finish_task_switch must be called after the context switch, paired
1818 * with a prepare_task_switch call before the context switch.
1819 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1820 * and do any other architecture-specific cleanup actions.
1822 * Note that we may have delayed dropping an mm in context_switch(). If
1823 * so, we finish that here outside of the runqueue lock. (Doing it
1824 * with the lock held can cause deadlocks; see schedule() for
1827 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1828 __releases(rq->lock)
1830 struct mm_struct *mm = rq->prev_mm;
1836 * A task struct has one reference for the use as "current".
1837 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1838 * schedule one last time. The schedule call will never return, and
1839 * the scheduled task must drop that reference.
1840 * The test for TASK_DEAD must occur while the runqueue locks are
1841 * still held, otherwise prev could be scheduled on another cpu, die
1842 * there before we look at prev->state, and then the reference would
1844 * Manfred Spraul <manfred@colorfullife.com>
1846 prev_state = prev->state;
1847 finish_arch_switch(prev);
1848 finish_lock_switch(rq, prev);
1849 fire_sched_in_preempt_notifiers(current);
1852 if (unlikely(prev_state == TASK_DEAD)) {
1854 * Remove function-return probe instances associated with this
1855 * task and put them back on the free list.
1857 kprobe_flush_task(prev);
1858 put_task_struct(prev);
1863 * schedule_tail - first thing a freshly forked thread must call.
1864 * @prev: the thread we just switched away from.
1866 asmlinkage void schedule_tail(struct task_struct *prev)
1867 __releases(rq->lock)
1869 struct rq *rq = this_rq();
1871 finish_task_switch(rq, prev);
1872 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1873 /* In this case, finish_task_switch does not reenable preemption */
1876 if (current->set_child_tid)
1877 put_user(current->pid, current->set_child_tid);
1881 * context_switch - switch to the new MM and the new
1882 * thread's register state.
1885 context_switch(struct rq *rq, struct task_struct *prev,
1886 struct task_struct *next)
1888 struct mm_struct *mm, *oldmm;
1890 prepare_task_switch(rq, prev, next);
1892 oldmm = prev->active_mm;
1894 * For paravirt, this is coupled with an exit in switch_to to
1895 * combine the page table reload and the switch backend into
1898 arch_enter_lazy_cpu_mode();
1900 if (unlikely(!mm)) {
1901 next->active_mm = oldmm;
1902 atomic_inc(&oldmm->mm_count);
1903 enter_lazy_tlb(oldmm, next);
1905 switch_mm(oldmm, mm, next);
1907 if (unlikely(!prev->mm)) {
1908 prev->active_mm = NULL;
1909 rq->prev_mm = oldmm;
1912 * Since the runqueue lock will be released by the next
1913 * task (which is an invalid locking op but in the case
1914 * of the scheduler it's an obvious special-case), so we
1915 * do an early lockdep release here:
1917 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1918 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1921 /* Here we just switch the register state and the stack. */
1922 switch_to(prev, next, prev);
1926 * this_rq must be evaluated again because prev may have moved
1927 * CPUs since it called schedule(), thus the 'rq' on its stack
1928 * frame will be invalid.
1930 finish_task_switch(this_rq(), prev);
1934 * nr_running, nr_uninterruptible and nr_context_switches:
1936 * externally visible scheduler statistics: current number of runnable
1937 * threads, current number of uninterruptible-sleeping threads, total
1938 * number of context switches performed since bootup.
1940 unsigned long nr_running(void)
1942 unsigned long i, sum = 0;
1944 for_each_online_cpu(i)
1945 sum += cpu_rq(i)->nr_running;
1950 unsigned long nr_uninterruptible(void)
1952 unsigned long i, sum = 0;
1954 for_each_possible_cpu(i)
1955 sum += cpu_rq(i)->nr_uninterruptible;
1958 * Since we read the counters lockless, it might be slightly
1959 * inaccurate. Do not allow it to go below zero though:
1961 if (unlikely((long)sum < 0))
1967 unsigned long long nr_context_switches(void)
1970 unsigned long long sum = 0;
1972 for_each_possible_cpu(i)
1973 sum += cpu_rq(i)->nr_switches;
1978 unsigned long nr_iowait(void)
1980 unsigned long i, sum = 0;
1982 for_each_possible_cpu(i)
1983 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1988 unsigned long nr_active(void)
1990 unsigned long i, running = 0, uninterruptible = 0;
1992 for_each_online_cpu(i) {
1993 running += cpu_rq(i)->nr_running;
1994 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1997 if (unlikely((long)uninterruptible < 0))
1998 uninterruptible = 0;
2000 return running + uninterruptible;
2004 * Update rq->cpu_load[] statistics. This function is usually called every
2005 * scheduler tick (TICK_NSEC).
2007 static void update_cpu_load(struct rq *this_rq)
2009 unsigned long this_load = this_rq->load.weight;
2012 this_rq->nr_load_updates++;
2014 /* Update our load: */
2015 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2016 unsigned long old_load, new_load;
2018 /* scale is effectively 1 << i now, and >> i divides by scale */
2020 old_load = this_rq->cpu_load[i];
2021 new_load = this_load;
2023 * Round up the averaging division if load is increasing. This
2024 * prevents us from getting stuck on 9 if the load is 10, for
2027 if (new_load > old_load)
2028 new_load += scale-1;
2029 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2036 * double_rq_lock - safely lock two runqueues
2038 * Note this does not disable interrupts like task_rq_lock,
2039 * you need to do so manually before calling.
2041 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2042 __acquires(rq1->lock)
2043 __acquires(rq2->lock)
2045 BUG_ON(!irqs_disabled());
2047 spin_lock(&rq1->lock);
2048 __acquire(rq2->lock); /* Fake it out ;) */
2051 spin_lock(&rq1->lock);
2052 spin_lock(&rq2->lock);
2054 spin_lock(&rq2->lock);
2055 spin_lock(&rq1->lock);
2058 update_rq_clock(rq1);
2059 update_rq_clock(rq2);
2063 * double_rq_unlock - safely unlock two runqueues
2065 * Note this does not restore interrupts like task_rq_unlock,
2066 * you need to do so manually after calling.
2068 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2069 __releases(rq1->lock)
2070 __releases(rq2->lock)
2072 spin_unlock(&rq1->lock);
2074 spin_unlock(&rq2->lock);
2076 __release(rq2->lock);
2080 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2082 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2083 __releases(this_rq->lock)
2084 __acquires(busiest->lock)
2085 __acquires(this_rq->lock)
2087 if (unlikely(!irqs_disabled())) {
2088 /* printk() doesn't work good under rq->lock */
2089 spin_unlock(&this_rq->lock);
2092 if (unlikely(!spin_trylock(&busiest->lock))) {
2093 if (busiest < this_rq) {
2094 spin_unlock(&this_rq->lock);
2095 spin_lock(&busiest->lock);
2096 spin_lock(&this_rq->lock);
2098 spin_lock(&busiest->lock);
2103 * If dest_cpu is allowed for this process, migrate the task to it.
2104 * This is accomplished by forcing the cpu_allowed mask to only
2105 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2106 * the cpu_allowed mask is restored.
2108 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2110 struct migration_req req;
2111 unsigned long flags;
2114 rq = task_rq_lock(p, &flags);
2115 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2116 || unlikely(cpu_is_offline(dest_cpu)))
2119 /* force the process onto the specified CPU */
2120 if (migrate_task(p, dest_cpu, &req)) {
2121 /* Need to wait for migration thread (might exit: take ref). */
2122 struct task_struct *mt = rq->migration_thread;
2124 get_task_struct(mt);
2125 task_rq_unlock(rq, &flags);
2126 wake_up_process(mt);
2127 put_task_struct(mt);
2128 wait_for_completion(&req.done);
2133 task_rq_unlock(rq, &flags);
2137 * sched_exec - execve() is a valuable balancing opportunity, because at
2138 * this point the task has the smallest effective memory and cache footprint.
2140 void sched_exec(void)
2142 int new_cpu, this_cpu = get_cpu();
2143 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2145 if (new_cpu != this_cpu)
2146 sched_migrate_task(current, new_cpu);
2150 * pull_task - move a task from a remote runqueue to the local runqueue.
2151 * Both runqueues must be locked.
2153 static void pull_task(struct rq *src_rq, struct task_struct *p,
2154 struct rq *this_rq, int this_cpu)
2156 deactivate_task(src_rq, p, 0);
2157 set_task_cpu(p, this_cpu);
2158 activate_task(this_rq, p, 0);
2160 * Note that idle threads have a prio of MAX_PRIO, for this test
2161 * to be always true for them.
2163 check_preempt_curr(this_rq, p);
2167 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2170 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2171 struct sched_domain *sd, enum cpu_idle_type idle,
2175 * We do not migrate tasks that are:
2176 * 1) running (obviously), or
2177 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2178 * 3) are cache-hot on their current CPU.
2180 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2181 schedstat_inc(p, se.nr_failed_migrations_affine);
2186 if (task_running(rq, p)) {
2187 schedstat_inc(p, se.nr_failed_migrations_running);
2192 * Aggressive migration if:
2193 * 1) task is cache cold, or
2194 * 2) too many balance attempts have failed.
2197 if (!task_hot(p, rq->clock, sd) ||
2198 sd->nr_balance_failed > sd->cache_nice_tries) {
2199 #ifdef CONFIG_SCHEDSTATS
2200 if (task_hot(p, rq->clock, sd)) {
2201 schedstat_inc(sd, lb_hot_gained[idle]);
2202 schedstat_inc(p, se.nr_forced_migrations);
2208 if (task_hot(p, rq->clock, sd)) {
2209 schedstat_inc(p, se.nr_failed_migrations_hot);
2215 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2216 unsigned long max_nr_move, unsigned long max_load_move,
2217 struct sched_domain *sd, enum cpu_idle_type idle,
2218 int *all_pinned, unsigned long *load_moved,
2219 int *this_best_prio, struct rq_iterator *iterator)
2221 int pulled = 0, pinned = 0, skip_for_load;
2222 struct task_struct *p;
2223 long rem_load_move = max_load_move;
2225 if (max_nr_move == 0 || max_load_move == 0)
2231 * Start the load-balancing iterator:
2233 p = iterator->start(iterator->arg);
2238 * To help distribute high priority tasks accross CPUs we don't
2239 * skip a task if it will be the highest priority task (i.e. smallest
2240 * prio value) on its new queue regardless of its load weight
2242 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2243 SCHED_LOAD_SCALE_FUZZ;
2244 if ((skip_for_load && p->prio >= *this_best_prio) ||
2245 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2246 p = iterator->next(iterator->arg);
2250 pull_task(busiest, p, this_rq, this_cpu);
2252 rem_load_move -= p->se.load.weight;
2255 * We only want to steal up to the prescribed number of tasks
2256 * and the prescribed amount of weighted load.
2258 if (pulled < max_nr_move && rem_load_move > 0) {
2259 if (p->prio < *this_best_prio)
2260 *this_best_prio = p->prio;
2261 p = iterator->next(iterator->arg);
2266 * Right now, this is the only place pull_task() is called,
2267 * so we can safely collect pull_task() stats here rather than
2268 * inside pull_task().
2270 schedstat_add(sd, lb_gained[idle], pulled);
2273 *all_pinned = pinned;
2274 *load_moved = max_load_move - rem_load_move;
2279 * move_tasks tries to move up to max_load_move weighted load from busiest to
2280 * this_rq, as part of a balancing operation within domain "sd".
2281 * Returns 1 if successful and 0 otherwise.
2283 * Called with both runqueues locked.
2285 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2286 unsigned long max_load_move,
2287 struct sched_domain *sd, enum cpu_idle_type idle,
2290 const struct sched_class *class = sched_class_highest;
2291 unsigned long total_load_moved = 0;
2292 int this_best_prio = this_rq->curr->prio;
2296 class->load_balance(this_rq, this_cpu, busiest,
2297 ULONG_MAX, max_load_move - total_load_moved,
2298 sd, idle, all_pinned, &this_best_prio);
2299 class = class->next;
2300 } while (class && max_load_move > total_load_moved);
2302 return total_load_moved > 0;
2306 * move_one_task tries to move exactly one task from busiest to this_rq, as
2307 * part of active balancing operations within "domain".
2308 * Returns 1 if successful and 0 otherwise.
2310 * Called with both runqueues locked.
2312 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2313 struct sched_domain *sd, enum cpu_idle_type idle)
2315 const struct sched_class *class;
2316 int this_best_prio = MAX_PRIO;
2318 for (class = sched_class_highest; class; class = class->next)
2319 if (class->load_balance(this_rq, this_cpu, busiest,
2320 1, ULONG_MAX, sd, idle, NULL,
2328 * find_busiest_group finds and returns the busiest CPU group within the
2329 * domain. It calculates and returns the amount of weighted load which
2330 * should be moved to restore balance via the imbalance parameter.
2332 static struct sched_group *
2333 find_busiest_group(struct sched_domain *sd, int this_cpu,
2334 unsigned long *imbalance, enum cpu_idle_type idle,
2335 int *sd_idle, cpumask_t *cpus, int *balance)
2337 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2338 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2339 unsigned long max_pull;
2340 unsigned long busiest_load_per_task, busiest_nr_running;
2341 unsigned long this_load_per_task, this_nr_running;
2343 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2344 int power_savings_balance = 1;
2345 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2346 unsigned long min_nr_running = ULONG_MAX;
2347 struct sched_group *group_min = NULL, *group_leader = NULL;
2350 max_load = this_load = total_load = total_pwr = 0;
2351 busiest_load_per_task = busiest_nr_running = 0;
2352 this_load_per_task = this_nr_running = 0;
2353 if (idle == CPU_NOT_IDLE)
2354 load_idx = sd->busy_idx;
2355 else if (idle == CPU_NEWLY_IDLE)
2356 load_idx = sd->newidle_idx;
2358 load_idx = sd->idle_idx;
2361 unsigned long load, group_capacity;
2364 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2365 unsigned long sum_nr_running, sum_weighted_load;
2367 local_group = cpu_isset(this_cpu, group->cpumask);
2370 balance_cpu = first_cpu(group->cpumask);
2372 /* Tally up the load of all CPUs in the group */
2373 sum_weighted_load = sum_nr_running = avg_load = 0;
2375 for_each_cpu_mask(i, group->cpumask) {
2378 if (!cpu_isset(i, *cpus))
2383 if (*sd_idle && rq->nr_running)
2386 /* Bias balancing toward cpus of our domain */
2388 if (idle_cpu(i) && !first_idle_cpu) {
2393 load = target_load(i, load_idx);
2395 load = source_load(i, load_idx);
2398 sum_nr_running += rq->nr_running;
2399 sum_weighted_load += weighted_cpuload(i);
2403 * First idle cpu or the first cpu(busiest) in this sched group
2404 * is eligible for doing load balancing at this and above
2405 * domains. In the newly idle case, we will allow all the cpu's
2406 * to do the newly idle load balance.
2408 if (idle != CPU_NEWLY_IDLE && local_group &&
2409 balance_cpu != this_cpu && balance) {
2414 total_load += avg_load;
2415 total_pwr += group->__cpu_power;
2417 /* Adjust by relative CPU power of the group */
2418 avg_load = sg_div_cpu_power(group,
2419 avg_load * SCHED_LOAD_SCALE);
2421 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2424 this_load = avg_load;
2426 this_nr_running = sum_nr_running;
2427 this_load_per_task = sum_weighted_load;
2428 } else if (avg_load > max_load &&
2429 sum_nr_running > group_capacity) {
2430 max_load = avg_load;
2432 busiest_nr_running = sum_nr_running;
2433 busiest_load_per_task = sum_weighted_load;
2436 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2438 * Busy processors will not participate in power savings
2441 if (idle == CPU_NOT_IDLE ||
2442 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2446 * If the local group is idle or completely loaded
2447 * no need to do power savings balance at this domain
2449 if (local_group && (this_nr_running >= group_capacity ||
2451 power_savings_balance = 0;
2454 * If a group is already running at full capacity or idle,
2455 * don't include that group in power savings calculations
2457 if (!power_savings_balance || sum_nr_running >= group_capacity
2462 * Calculate the group which has the least non-idle load.
2463 * This is the group from where we need to pick up the load
2466 if ((sum_nr_running < min_nr_running) ||
2467 (sum_nr_running == min_nr_running &&
2468 first_cpu(group->cpumask) <
2469 first_cpu(group_min->cpumask))) {
2471 min_nr_running = sum_nr_running;
2472 min_load_per_task = sum_weighted_load /
2477 * Calculate the group which is almost near its
2478 * capacity but still has some space to pick up some load
2479 * from other group and save more power
2481 if (sum_nr_running <= group_capacity - 1) {
2482 if (sum_nr_running > leader_nr_running ||
2483 (sum_nr_running == leader_nr_running &&
2484 first_cpu(group->cpumask) >
2485 first_cpu(group_leader->cpumask))) {
2486 group_leader = group;
2487 leader_nr_running = sum_nr_running;
2492 group = group->next;
2493 } while (group != sd->groups);
2495 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2498 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2500 if (this_load >= avg_load ||
2501 100*max_load <= sd->imbalance_pct*this_load)
2504 busiest_load_per_task /= busiest_nr_running;
2506 * We're trying to get all the cpus to the average_load, so we don't
2507 * want to push ourselves above the average load, nor do we wish to
2508 * reduce the max loaded cpu below the average load, as either of these
2509 * actions would just result in more rebalancing later, and ping-pong
2510 * tasks around. Thus we look for the minimum possible imbalance.
2511 * Negative imbalances (*we* are more loaded than anyone else) will
2512 * be counted as no imbalance for these purposes -- we can't fix that
2513 * by pulling tasks to us. Be careful of negative numbers as they'll
2514 * appear as very large values with unsigned longs.
2516 if (max_load <= busiest_load_per_task)
2520 * In the presence of smp nice balancing, certain scenarios can have
2521 * max load less than avg load(as we skip the groups at or below
2522 * its cpu_power, while calculating max_load..)
2524 if (max_load < avg_load) {
2526 goto small_imbalance;
2529 /* Don't want to pull so many tasks that a group would go idle */
2530 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2532 /* How much load to actually move to equalise the imbalance */
2533 *imbalance = min(max_pull * busiest->__cpu_power,
2534 (avg_load - this_load) * this->__cpu_power)
2538 * if *imbalance is less than the average load per runnable task
2539 * there is no gaurantee that any tasks will be moved so we'll have
2540 * a think about bumping its value to force at least one task to be
2543 if (*imbalance < busiest_load_per_task) {
2544 unsigned long tmp, pwr_now, pwr_move;
2548 pwr_move = pwr_now = 0;
2550 if (this_nr_running) {
2551 this_load_per_task /= this_nr_running;
2552 if (busiest_load_per_task > this_load_per_task)
2555 this_load_per_task = SCHED_LOAD_SCALE;
2557 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2558 busiest_load_per_task * imbn) {
2559 *imbalance = busiest_load_per_task;
2564 * OK, we don't have enough imbalance to justify moving tasks,
2565 * however we may be able to increase total CPU power used by
2569 pwr_now += busiest->__cpu_power *
2570 min(busiest_load_per_task, max_load);
2571 pwr_now += this->__cpu_power *
2572 min(this_load_per_task, this_load);
2573 pwr_now /= SCHED_LOAD_SCALE;
2575 /* Amount of load we'd subtract */
2576 tmp = sg_div_cpu_power(busiest,
2577 busiest_load_per_task * SCHED_LOAD_SCALE);
2579 pwr_move += busiest->__cpu_power *
2580 min(busiest_load_per_task, max_load - tmp);
2582 /* Amount of load we'd add */
2583 if (max_load * busiest->__cpu_power <
2584 busiest_load_per_task * SCHED_LOAD_SCALE)
2585 tmp = sg_div_cpu_power(this,
2586 max_load * busiest->__cpu_power);
2588 tmp = sg_div_cpu_power(this,
2589 busiest_load_per_task * SCHED_LOAD_SCALE);
2590 pwr_move += this->__cpu_power *
2591 min(this_load_per_task, this_load + tmp);
2592 pwr_move /= SCHED_LOAD_SCALE;
2594 /* Move if we gain throughput */
2595 if (pwr_move > pwr_now)
2596 *imbalance = busiest_load_per_task;
2602 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2603 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2606 if (this == group_leader && group_leader != group_min) {
2607 *imbalance = min_load_per_task;
2617 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2620 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2621 unsigned long imbalance, cpumask_t *cpus)
2623 struct rq *busiest = NULL, *rq;
2624 unsigned long max_load = 0;
2627 for_each_cpu_mask(i, group->cpumask) {
2630 if (!cpu_isset(i, *cpus))
2634 wl = weighted_cpuload(i);
2636 if (rq->nr_running == 1 && wl > imbalance)
2639 if (wl > max_load) {
2649 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2650 * so long as it is large enough.
2652 #define MAX_PINNED_INTERVAL 512
2655 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2656 * tasks if there is an imbalance.
2658 static int load_balance(int this_cpu, struct rq *this_rq,
2659 struct sched_domain *sd, enum cpu_idle_type idle,
2662 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2663 struct sched_group *group;
2664 unsigned long imbalance;
2666 cpumask_t cpus = CPU_MASK_ALL;
2667 unsigned long flags;
2670 * When power savings policy is enabled for the parent domain, idle
2671 * sibling can pick up load irrespective of busy siblings. In this case,
2672 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2673 * portraying it as CPU_NOT_IDLE.
2675 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2676 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2679 schedstat_inc(sd, lb_count[idle]);
2682 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2689 schedstat_inc(sd, lb_nobusyg[idle]);
2693 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2695 schedstat_inc(sd, lb_nobusyq[idle]);
2699 BUG_ON(busiest == this_rq);
2701 schedstat_add(sd, lb_imbalance[idle], imbalance);
2704 if (busiest->nr_running > 1) {
2706 * Attempt to move tasks. If find_busiest_group has found
2707 * an imbalance but busiest->nr_running <= 1, the group is
2708 * still unbalanced. ld_moved simply stays zero, so it is
2709 * correctly treated as an imbalance.
2711 local_irq_save(flags);
2712 double_rq_lock(this_rq, busiest);
2713 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2714 imbalance, sd, idle, &all_pinned);
2715 double_rq_unlock(this_rq, busiest);
2716 local_irq_restore(flags);
2719 * some other cpu did the load balance for us.
2721 if (ld_moved && this_cpu != smp_processor_id())
2722 resched_cpu(this_cpu);
2724 /* All tasks on this runqueue were pinned by CPU affinity */
2725 if (unlikely(all_pinned)) {
2726 cpu_clear(cpu_of(busiest), cpus);
2727 if (!cpus_empty(cpus))
2734 schedstat_inc(sd, lb_failed[idle]);
2735 sd->nr_balance_failed++;
2737 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2739 spin_lock_irqsave(&busiest->lock, flags);
2741 /* don't kick the migration_thread, if the curr
2742 * task on busiest cpu can't be moved to this_cpu
2744 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2745 spin_unlock_irqrestore(&busiest->lock, flags);
2747 goto out_one_pinned;
2750 if (!busiest->active_balance) {
2751 busiest->active_balance = 1;
2752 busiest->push_cpu = this_cpu;
2755 spin_unlock_irqrestore(&busiest->lock, flags);
2757 wake_up_process(busiest->migration_thread);
2760 * We've kicked active balancing, reset the failure
2763 sd->nr_balance_failed = sd->cache_nice_tries+1;
2766 sd->nr_balance_failed = 0;
2768 if (likely(!active_balance)) {
2769 /* We were unbalanced, so reset the balancing interval */
2770 sd->balance_interval = sd->min_interval;
2773 * If we've begun active balancing, start to back off. This
2774 * case may not be covered by the all_pinned logic if there
2775 * is only 1 task on the busy runqueue (because we don't call
2778 if (sd->balance_interval < sd->max_interval)
2779 sd->balance_interval *= 2;
2782 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2783 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2788 schedstat_inc(sd, lb_balanced[idle]);
2790 sd->nr_balance_failed = 0;
2793 /* tune up the balancing interval */
2794 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2795 (sd->balance_interval < sd->max_interval))
2796 sd->balance_interval *= 2;
2798 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2799 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2805 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2806 * tasks if there is an imbalance.
2808 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2809 * this_rq is locked.
2812 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2814 struct sched_group *group;
2815 struct rq *busiest = NULL;
2816 unsigned long imbalance;
2820 cpumask_t cpus = CPU_MASK_ALL;
2823 * When power savings policy is enabled for the parent domain, idle
2824 * sibling can pick up load irrespective of busy siblings. In this case,
2825 * let the state of idle sibling percolate up as IDLE, instead of
2826 * portraying it as CPU_NOT_IDLE.
2828 if (sd->flags & SD_SHARE_CPUPOWER &&
2829 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2832 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2834 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2835 &sd_idle, &cpus, NULL);
2837 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2841 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2844 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2848 BUG_ON(busiest == this_rq);
2850 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2853 if (busiest->nr_running > 1) {
2854 /* Attempt to move tasks */
2855 double_lock_balance(this_rq, busiest);
2856 /* this_rq->clock is already updated */
2857 update_rq_clock(busiest);
2858 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2859 imbalance, sd, CPU_NEWLY_IDLE,
2861 spin_unlock(&busiest->lock);
2863 if (unlikely(all_pinned)) {
2864 cpu_clear(cpu_of(busiest), cpus);
2865 if (!cpus_empty(cpus))
2871 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2872 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2873 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2876 sd->nr_balance_failed = 0;
2881 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2882 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2883 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2885 sd->nr_balance_failed = 0;
2891 * idle_balance is called by schedule() if this_cpu is about to become
2892 * idle. Attempts to pull tasks from other CPUs.
2894 static void idle_balance(int this_cpu, struct rq *this_rq)
2896 struct sched_domain *sd;
2897 int pulled_task = -1;
2898 unsigned long next_balance = jiffies + HZ;
2900 for_each_domain(this_cpu, sd) {
2901 unsigned long interval;
2903 if (!(sd->flags & SD_LOAD_BALANCE))
2906 if (sd->flags & SD_BALANCE_NEWIDLE)
2907 /* If we've pulled tasks over stop searching: */
2908 pulled_task = load_balance_newidle(this_cpu,
2911 interval = msecs_to_jiffies(sd->balance_interval);
2912 if (time_after(next_balance, sd->last_balance + interval))
2913 next_balance = sd->last_balance + interval;
2917 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2919 * We are going idle. next_balance may be set based on
2920 * a busy processor. So reset next_balance.
2922 this_rq->next_balance = next_balance;
2927 * active_load_balance is run by migration threads. It pushes running tasks
2928 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2929 * running on each physical CPU where possible, and avoids physical /
2930 * logical imbalances.
2932 * Called with busiest_rq locked.
2934 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2936 int target_cpu = busiest_rq->push_cpu;
2937 struct sched_domain *sd;
2938 struct rq *target_rq;
2940 /* Is there any task to move? */
2941 if (busiest_rq->nr_running <= 1)
2944 target_rq = cpu_rq(target_cpu);
2947 * This condition is "impossible", if it occurs
2948 * we need to fix it. Originally reported by
2949 * Bjorn Helgaas on a 128-cpu setup.
2951 BUG_ON(busiest_rq == target_rq);
2953 /* move a task from busiest_rq to target_rq */
2954 double_lock_balance(busiest_rq, target_rq);
2955 update_rq_clock(busiest_rq);
2956 update_rq_clock(target_rq);
2958 /* Search for an sd spanning us and the target CPU. */
2959 for_each_domain(target_cpu, sd) {
2960 if ((sd->flags & SD_LOAD_BALANCE) &&
2961 cpu_isset(busiest_cpu, sd->span))
2966 schedstat_inc(sd, alb_count);
2968 if (move_one_task(target_rq, target_cpu, busiest_rq,
2970 schedstat_inc(sd, alb_pushed);
2972 schedstat_inc(sd, alb_failed);
2974 spin_unlock(&target_rq->lock);
2979 atomic_t load_balancer;
2981 } nohz ____cacheline_aligned = {
2982 .load_balancer = ATOMIC_INIT(-1),
2983 .cpu_mask = CPU_MASK_NONE,
2987 * This routine will try to nominate the ilb (idle load balancing)
2988 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2989 * load balancing on behalf of all those cpus. If all the cpus in the system
2990 * go into this tickless mode, then there will be no ilb owner (as there is
2991 * no need for one) and all the cpus will sleep till the next wakeup event
2994 * For the ilb owner, tick is not stopped. And this tick will be used
2995 * for idle load balancing. ilb owner will still be part of
2998 * While stopping the tick, this cpu will become the ilb owner if there
2999 * is no other owner. And will be the owner till that cpu becomes busy
3000 * or if all cpus in the system stop their ticks at which point
3001 * there is no need for ilb owner.
3003 * When the ilb owner becomes busy, it nominates another owner, during the
3004 * next busy scheduler_tick()
3006 int select_nohz_load_balancer(int stop_tick)
3008 int cpu = smp_processor_id();
3011 cpu_set(cpu, nohz.cpu_mask);
3012 cpu_rq(cpu)->in_nohz_recently = 1;
3015 * If we are going offline and still the leader, give up!
3017 if (cpu_is_offline(cpu) &&
3018 atomic_read(&nohz.load_balancer) == cpu) {
3019 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3024 /* time for ilb owner also to sleep */
3025 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3026 if (atomic_read(&nohz.load_balancer) == cpu)
3027 atomic_set(&nohz.load_balancer, -1);
3031 if (atomic_read(&nohz.load_balancer) == -1) {
3032 /* make me the ilb owner */
3033 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3035 } else if (atomic_read(&nohz.load_balancer) == cpu)
3038 if (!cpu_isset(cpu, nohz.cpu_mask))
3041 cpu_clear(cpu, nohz.cpu_mask);
3043 if (atomic_read(&nohz.load_balancer) == cpu)
3044 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3051 static DEFINE_SPINLOCK(balancing);
3054 * It checks each scheduling domain to see if it is due to be balanced,
3055 * and initiates a balancing operation if so.
3057 * Balancing parameters are set up in arch_init_sched_domains.
3059 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3062 struct rq *rq = cpu_rq(cpu);
3063 unsigned long interval;
3064 struct sched_domain *sd;
3065 /* Earliest time when we have to do rebalance again */
3066 unsigned long next_balance = jiffies + 60*HZ;
3067 int update_next_balance = 0;
3069 for_each_domain(cpu, sd) {
3070 if (!(sd->flags & SD_LOAD_BALANCE))
3073 interval = sd->balance_interval;
3074 if (idle != CPU_IDLE)
3075 interval *= sd->busy_factor;
3077 /* scale ms to jiffies */
3078 interval = msecs_to_jiffies(interval);
3079 if (unlikely(!interval))
3081 if (interval > HZ*NR_CPUS/10)
3082 interval = HZ*NR_CPUS/10;
3085 if (sd->flags & SD_SERIALIZE) {
3086 if (!spin_trylock(&balancing))
3090 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3091 if (load_balance(cpu, rq, sd, idle, &balance)) {
3093 * We've pulled tasks over so either we're no
3094 * longer idle, or one of our SMT siblings is
3097 idle = CPU_NOT_IDLE;
3099 sd->last_balance = jiffies;
3101 if (sd->flags & SD_SERIALIZE)
3102 spin_unlock(&balancing);
3104 if (time_after(next_balance, sd->last_balance + interval)) {
3105 next_balance = sd->last_balance + interval;
3106 update_next_balance = 1;
3110 * Stop the load balance at this level. There is another
3111 * CPU in our sched group which is doing load balancing more
3119 * next_balance will be updated only when there is a need.
3120 * When the cpu is attached to null domain for ex, it will not be
3123 if (likely(update_next_balance))
3124 rq->next_balance = next_balance;
3128 * run_rebalance_domains is triggered when needed from the scheduler tick.
3129 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3130 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3132 static void run_rebalance_domains(struct softirq_action *h)
3134 int this_cpu = smp_processor_id();
3135 struct rq *this_rq = cpu_rq(this_cpu);
3136 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3137 CPU_IDLE : CPU_NOT_IDLE;
3139 rebalance_domains(this_cpu, idle);
3143 * If this cpu is the owner for idle load balancing, then do the
3144 * balancing on behalf of the other idle cpus whose ticks are
3147 if (this_rq->idle_at_tick &&
3148 atomic_read(&nohz.load_balancer) == this_cpu) {
3149 cpumask_t cpus = nohz.cpu_mask;
3153 cpu_clear(this_cpu, cpus);
3154 for_each_cpu_mask(balance_cpu, cpus) {
3156 * If this cpu gets work to do, stop the load balancing
3157 * work being done for other cpus. Next load
3158 * balancing owner will pick it up.
3163 rebalance_domains(balance_cpu, CPU_IDLE);
3165 rq = cpu_rq(balance_cpu);
3166 if (time_after(this_rq->next_balance, rq->next_balance))
3167 this_rq->next_balance = rq->next_balance;
3174 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3176 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3177 * idle load balancing owner or decide to stop the periodic load balancing,
3178 * if the whole system is idle.
3180 static inline void trigger_load_balance(struct rq *rq, int cpu)
3184 * If we were in the nohz mode recently and busy at the current
3185 * scheduler tick, then check if we need to nominate new idle
3188 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3189 rq->in_nohz_recently = 0;
3191 if (atomic_read(&nohz.load_balancer) == cpu) {
3192 cpu_clear(cpu, nohz.cpu_mask);
3193 atomic_set(&nohz.load_balancer, -1);
3196 if (atomic_read(&nohz.load_balancer) == -1) {
3198 * simple selection for now: Nominate the
3199 * first cpu in the nohz list to be the next
3202 * TBD: Traverse the sched domains and nominate
3203 * the nearest cpu in the nohz.cpu_mask.
3205 int ilb = first_cpu(nohz.cpu_mask);
3213 * If this cpu is idle and doing idle load balancing for all the
3214 * cpus with ticks stopped, is it time for that to stop?
3216 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3217 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3223 * If this cpu is idle and the idle load balancing is done by
3224 * someone else, then no need raise the SCHED_SOFTIRQ
3226 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3227 cpu_isset(cpu, nohz.cpu_mask))
3230 if (time_after_eq(jiffies, rq->next_balance))
3231 raise_softirq(SCHED_SOFTIRQ);
3234 #else /* CONFIG_SMP */
3237 * on UP we do not need to balance between CPUs:
3239 static inline void idle_balance(int cpu, struct rq *rq)
3243 /* Avoid "used but not defined" warning on UP */
3244 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3245 unsigned long max_nr_move, unsigned long max_load_move,
3246 struct sched_domain *sd, enum cpu_idle_type idle,
3247 int *all_pinned, unsigned long *load_moved,
3248 int *this_best_prio, struct rq_iterator *iterator)
3257 DEFINE_PER_CPU(struct kernel_stat, kstat);
3259 EXPORT_PER_CPU_SYMBOL(kstat);
3262 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3263 * that have not yet been banked in case the task is currently running.
3265 unsigned long long task_sched_runtime(struct task_struct *p)
3267 unsigned long flags;
3271 rq = task_rq_lock(p, &flags);
3272 ns = p->se.sum_exec_runtime;
3273 if (rq->curr == p) {
3274 update_rq_clock(rq);
3275 delta_exec = rq->clock - p->se.exec_start;
3276 if ((s64)delta_exec > 0)
3279 task_rq_unlock(rq, &flags);
3285 * Account user cpu time to a process.
3286 * @p: the process that the cpu time gets accounted to
3287 * @hardirq_offset: the offset to subtract from hardirq_count()
3288 * @cputime: the cpu time spent in user space since the last update
3290 void account_user_time(struct task_struct *p, cputime_t cputime)
3292 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3295 p->utime = cputime_add(p->utime, cputime);
3297 /* Add user time to cpustat. */
3298 tmp = cputime_to_cputime64(cputime);
3299 if (TASK_NICE(p) > 0)
3300 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3302 cpustat->user = cputime64_add(cpustat->user, tmp);
3306 * Account guest cpu time to a process.
3307 * @p: the process that the cpu time gets accounted to
3308 * @cputime: the cpu time spent in virtual machine since the last update
3310 void account_guest_time(struct task_struct *p, cputime_t cputime)
3313 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3315 tmp = cputime_to_cputime64(cputime);
3317 p->utime = cputime_add(p->utime, cputime);
3318 p->gtime = cputime_add(p->gtime, cputime);
3320 cpustat->user = cputime64_add(cpustat->user, tmp);
3321 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3325 * Account system cpu time to a process.
3326 * @p: the process that the cpu time gets accounted to
3327 * @hardirq_offset: the offset to subtract from hardirq_count()
3328 * @cputime: the cpu time spent in kernel space since the last update
3330 void account_system_time(struct task_struct *p, int hardirq_offset,
3333 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3334 struct rq *rq = this_rq();
3337 if (p->flags & PF_VCPU) {
3338 account_guest_time(p, cputime);
3339 p->flags &= ~PF_VCPU;
3343 p->stime = cputime_add(p->stime, cputime);
3345 /* Add system time to cpustat. */
3346 tmp = cputime_to_cputime64(cputime);
3347 if (hardirq_count() - hardirq_offset)
3348 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3349 else if (softirq_count())
3350 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3351 else if (p != rq->idle)
3352 cpustat->system = cputime64_add(cpustat->system, tmp);
3353 else if (atomic_read(&rq->nr_iowait) > 0)
3354 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3356 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3357 /* Account for system time used */
3358 acct_update_integrals(p);
3362 * Account for involuntary wait time.
3363 * @p: the process from which the cpu time has been stolen
3364 * @steal: the cpu time spent in involuntary wait
3366 void account_steal_time(struct task_struct *p, cputime_t steal)
3368 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3369 cputime64_t tmp = cputime_to_cputime64(steal);
3370 struct rq *rq = this_rq();
3372 if (p == rq->idle) {
3373 p->stime = cputime_add(p->stime, steal);
3374 if (atomic_read(&rq->nr_iowait) > 0)
3375 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3377 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3379 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3383 * This function gets called by the timer code, with HZ frequency.
3384 * We call it with interrupts disabled.
3386 * It also gets called by the fork code, when changing the parent's
3389 void scheduler_tick(void)
3391 int cpu = smp_processor_id();
3392 struct rq *rq = cpu_rq(cpu);
3393 struct task_struct *curr = rq->curr;
3394 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3396 spin_lock(&rq->lock);
3397 __update_rq_clock(rq);
3399 * Let rq->clock advance by at least TICK_NSEC:
3401 if (unlikely(rq->clock < next_tick))
3402 rq->clock = next_tick;
3403 rq->tick_timestamp = rq->clock;
3404 update_cpu_load(rq);
3405 if (curr != rq->idle) /* FIXME: needed? */
3406 curr->sched_class->task_tick(rq, curr);
3407 spin_unlock(&rq->lock);
3410 rq->idle_at_tick = idle_cpu(cpu);
3411 trigger_load_balance(rq, cpu);
3415 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3417 void fastcall add_preempt_count(int val)
3422 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3424 preempt_count() += val;
3426 * Spinlock count overflowing soon?
3428 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3431 EXPORT_SYMBOL(add_preempt_count);
3433 void fastcall sub_preempt_count(int val)
3438 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3441 * Is the spinlock portion underflowing?
3443 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3444 !(preempt_count() & PREEMPT_MASK)))
3447 preempt_count() -= val;
3449 EXPORT_SYMBOL(sub_preempt_count);
3454 * Print scheduling while atomic bug:
3456 static noinline void __schedule_bug(struct task_struct *prev)
3458 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3459 prev->comm, preempt_count(), prev->pid);
3460 debug_show_held_locks(prev);
3461 if (irqs_disabled())
3462 print_irqtrace_events(prev);
3467 * Various schedule()-time debugging checks and statistics:
3469 static inline void schedule_debug(struct task_struct *prev)
3472 * Test if we are atomic. Since do_exit() needs to call into
3473 * schedule() atomically, we ignore that path for now.
3474 * Otherwise, whine if we are scheduling when we should not be.
3476 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3477 __schedule_bug(prev);
3479 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3481 schedstat_inc(this_rq(), sched_count);
3482 #ifdef CONFIG_SCHEDSTATS
3483 if (unlikely(prev->lock_depth >= 0)) {
3484 schedstat_inc(this_rq(), bkl_count);
3485 schedstat_inc(prev, sched_info.bkl_count);
3491 * Pick up the highest-prio task:
3493 static inline struct task_struct *
3494 pick_next_task(struct rq *rq, struct task_struct *prev)
3496 const struct sched_class *class;
3497 struct task_struct *p;
3500 * Optimization: we know that if all tasks are in
3501 * the fair class we can call that function directly:
3503 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3504 p = fair_sched_class.pick_next_task(rq);
3509 class = sched_class_highest;
3511 p = class->pick_next_task(rq);
3515 * Will never be NULL as the idle class always
3516 * returns a non-NULL p:
3518 class = class->next;
3523 * schedule() is the main scheduler function.
3525 asmlinkage void __sched schedule(void)
3527 struct task_struct *prev, *next;
3534 cpu = smp_processor_id();
3538 switch_count = &prev->nivcsw;
3540 release_kernel_lock(prev);
3541 need_resched_nonpreemptible:
3543 schedule_debug(prev);
3546 * Do the rq-clock update outside the rq lock:
3548 local_irq_disable();
3549 __update_rq_clock(rq);
3550 spin_lock(&rq->lock);
3551 clear_tsk_need_resched(prev);
3553 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3554 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3555 unlikely(signal_pending(prev)))) {
3556 prev->state = TASK_RUNNING;
3558 deactivate_task(rq, prev, 1);
3560 switch_count = &prev->nvcsw;
3563 if (unlikely(!rq->nr_running))
3564 idle_balance(cpu, rq);
3566 prev->sched_class->put_prev_task(rq, prev);
3567 next = pick_next_task(rq, prev);
3569 sched_info_switch(prev, next);
3571 if (likely(prev != next)) {
3576 context_switch(rq, prev, next); /* unlocks the rq */
3578 spin_unlock_irq(&rq->lock);
3580 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3581 cpu = smp_processor_id();
3583 goto need_resched_nonpreemptible;
3585 preempt_enable_no_resched();
3586 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3589 EXPORT_SYMBOL(schedule);
3591 #ifdef CONFIG_PREEMPT
3593 * this is the entry point to schedule() from in-kernel preemption
3594 * off of preempt_enable. Kernel preemptions off return from interrupt
3595 * occur there and call schedule directly.
3597 asmlinkage void __sched preempt_schedule(void)
3599 struct thread_info *ti = current_thread_info();
3600 #ifdef CONFIG_PREEMPT_BKL
3601 struct task_struct *task = current;
3602 int saved_lock_depth;
3605 * If there is a non-zero preempt_count or interrupts are disabled,
3606 * we do not want to preempt the current task. Just return..
3608 if (likely(ti->preempt_count || irqs_disabled()))
3612 add_preempt_count(PREEMPT_ACTIVE);
3615 * We keep the big kernel semaphore locked, but we
3616 * clear ->lock_depth so that schedule() doesnt
3617 * auto-release the semaphore:
3619 #ifdef CONFIG_PREEMPT_BKL
3620 saved_lock_depth = task->lock_depth;
3621 task->lock_depth = -1;
3624 #ifdef CONFIG_PREEMPT_BKL
3625 task->lock_depth = saved_lock_depth;
3627 sub_preempt_count(PREEMPT_ACTIVE);
3630 * Check again in case we missed a preemption opportunity
3631 * between schedule and now.
3634 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3636 EXPORT_SYMBOL(preempt_schedule);
3639 * this is the entry point to schedule() from kernel preemption
3640 * off of irq context.
3641 * Note, that this is called and return with irqs disabled. This will
3642 * protect us against recursive calling from irq.
3644 asmlinkage void __sched preempt_schedule_irq(void)
3646 struct thread_info *ti = current_thread_info();
3647 #ifdef CONFIG_PREEMPT_BKL
3648 struct task_struct *task = current;
3649 int saved_lock_depth;
3651 /* Catch callers which need to be fixed */
3652 BUG_ON(ti->preempt_count || !irqs_disabled());
3655 add_preempt_count(PREEMPT_ACTIVE);
3658 * We keep the big kernel semaphore locked, but we
3659 * clear ->lock_depth so that schedule() doesnt
3660 * auto-release the semaphore:
3662 #ifdef CONFIG_PREEMPT_BKL
3663 saved_lock_depth = task->lock_depth;
3664 task->lock_depth = -1;
3668 local_irq_disable();
3669 #ifdef CONFIG_PREEMPT_BKL
3670 task->lock_depth = saved_lock_depth;
3672 sub_preempt_count(PREEMPT_ACTIVE);
3675 * Check again in case we missed a preemption opportunity
3676 * between schedule and now.
3679 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3682 #endif /* CONFIG_PREEMPT */
3684 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3687 return try_to_wake_up(curr->private, mode, sync);
3689 EXPORT_SYMBOL(default_wake_function);
3692 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3693 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3694 * number) then we wake all the non-exclusive tasks and one exclusive task.
3696 * There are circumstances in which we can try to wake a task which has already
3697 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3698 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3700 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3701 int nr_exclusive, int sync, void *key)
3703 wait_queue_t *curr, *next;
3705 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3706 unsigned flags = curr->flags;
3708 if (curr->func(curr, mode, sync, key) &&
3709 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3715 * __wake_up - wake up threads blocked on a waitqueue.
3717 * @mode: which threads
3718 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3719 * @key: is directly passed to the wakeup function
3721 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3722 int nr_exclusive, void *key)
3724 unsigned long flags;
3726 spin_lock_irqsave(&q->lock, flags);
3727 __wake_up_common(q, mode, nr_exclusive, 0, key);
3728 spin_unlock_irqrestore(&q->lock, flags);
3730 EXPORT_SYMBOL(__wake_up);
3733 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3735 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3737 __wake_up_common(q, mode, 1, 0, NULL);
3741 * __wake_up_sync - wake up threads blocked on a waitqueue.
3743 * @mode: which threads
3744 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3746 * The sync wakeup differs that the waker knows that it will schedule
3747 * away soon, so while the target thread will be woken up, it will not
3748 * be migrated to another CPU - ie. the two threads are 'synchronized'
3749 * with each other. This can prevent needless bouncing between CPUs.
3751 * On UP it can prevent extra preemption.
3754 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3756 unsigned long flags;
3762 if (unlikely(!nr_exclusive))
3765 spin_lock_irqsave(&q->lock, flags);
3766 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3767 spin_unlock_irqrestore(&q->lock, flags);
3769 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3771 void fastcall complete(struct completion *x)
3773 unsigned long flags;
3775 spin_lock_irqsave(&x->wait.lock, flags);
3777 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3779 spin_unlock_irqrestore(&x->wait.lock, flags);
3781 EXPORT_SYMBOL(complete);
3783 void fastcall complete_all(struct completion *x)
3785 unsigned long flags;
3787 spin_lock_irqsave(&x->wait.lock, flags);
3788 x->done += UINT_MAX/2;
3789 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3791 spin_unlock_irqrestore(&x->wait.lock, flags);
3793 EXPORT_SYMBOL(complete_all);
3795 static inline long __sched
3796 do_wait_for_common(struct completion *x, long timeout, int state)
3799 DECLARE_WAITQUEUE(wait, current);
3801 wait.flags |= WQ_FLAG_EXCLUSIVE;
3802 __add_wait_queue_tail(&x->wait, &wait);
3804 if (state == TASK_INTERRUPTIBLE &&
3805 signal_pending(current)) {
3806 __remove_wait_queue(&x->wait, &wait);
3807 return -ERESTARTSYS;
3809 __set_current_state(state);
3810 spin_unlock_irq(&x->wait.lock);
3811 timeout = schedule_timeout(timeout);
3812 spin_lock_irq(&x->wait.lock);
3814 __remove_wait_queue(&x->wait, &wait);
3818 __remove_wait_queue(&x->wait, &wait);
3825 wait_for_common(struct completion *x, long timeout, int state)
3829 spin_lock_irq(&x->wait.lock);
3830 timeout = do_wait_for_common(x, timeout, state);
3831 spin_unlock_irq(&x->wait.lock);
3835 void fastcall __sched wait_for_completion(struct completion *x)
3837 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3839 EXPORT_SYMBOL(wait_for_completion);
3841 unsigned long fastcall __sched
3842 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3844 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3846 EXPORT_SYMBOL(wait_for_completion_timeout);
3848 int __sched wait_for_completion_interruptible(struct completion *x)
3850 return wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3852 EXPORT_SYMBOL(wait_for_completion_interruptible);
3854 unsigned long fastcall __sched
3855 wait_for_completion_interruptible_timeout(struct completion *x,
3856 unsigned long timeout)
3858 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3860 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3863 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3865 unsigned long flags;
3868 init_waitqueue_entry(&wait, current);
3870 __set_current_state(state);
3872 spin_lock_irqsave(&q->lock, flags);
3873 __add_wait_queue(q, &wait);
3874 spin_unlock(&q->lock);
3875 timeout = schedule_timeout(timeout);
3876 spin_lock_irq(&q->lock);
3877 __remove_wait_queue(q, &wait);
3878 spin_unlock_irqrestore(&q->lock, flags);
3883 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3885 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3887 EXPORT_SYMBOL(interruptible_sleep_on);
3890 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3892 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3894 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3896 void __sched sleep_on(wait_queue_head_t *q)
3898 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3900 EXPORT_SYMBOL(sleep_on);
3902 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3904 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3906 EXPORT_SYMBOL(sleep_on_timeout);
3908 #ifdef CONFIG_RT_MUTEXES
3911 * rt_mutex_setprio - set the current priority of a task
3913 * @prio: prio value (kernel-internal form)
3915 * This function changes the 'effective' priority of a task. It does
3916 * not touch ->normal_prio like __setscheduler().
3918 * Used by the rt_mutex code to implement priority inheritance logic.
3920 void rt_mutex_setprio(struct task_struct *p, int prio)
3922 unsigned long flags;
3923 int oldprio, on_rq, running;
3926 BUG_ON(prio < 0 || prio > MAX_PRIO);
3928 rq = task_rq_lock(p, &flags);
3929 update_rq_clock(rq);
3932 on_rq = p->se.on_rq;
3933 running = task_running(rq, p);
3935 dequeue_task(rq, p, 0);
3937 p->sched_class->put_prev_task(rq, p);
3941 p->sched_class = &rt_sched_class;
3943 p->sched_class = &fair_sched_class;
3949 p->sched_class->set_curr_task(rq);
3950 enqueue_task(rq, p, 0);
3952 * Reschedule if we are currently running on this runqueue and
3953 * our priority decreased, or if we are not currently running on
3954 * this runqueue and our priority is higher than the current's
3957 if (p->prio > oldprio)
3958 resched_task(rq->curr);
3960 check_preempt_curr(rq, p);
3963 task_rq_unlock(rq, &flags);
3968 void set_user_nice(struct task_struct *p, long nice)
3970 int old_prio, delta, on_rq;
3971 unsigned long flags;
3974 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3977 * We have to be careful, if called from sys_setpriority(),
3978 * the task might be in the middle of scheduling on another CPU.
3980 rq = task_rq_lock(p, &flags);
3981 update_rq_clock(rq);
3983 * The RT priorities are set via sched_setscheduler(), but we still
3984 * allow the 'normal' nice value to be set - but as expected
3985 * it wont have any effect on scheduling until the task is
3986 * SCHED_FIFO/SCHED_RR:
3988 if (task_has_rt_policy(p)) {
3989 p->static_prio = NICE_TO_PRIO(nice);
3992 on_rq = p->se.on_rq;
3994 dequeue_task(rq, p, 0);
3998 p->static_prio = NICE_TO_PRIO(nice);
4001 p->prio = effective_prio(p);
4002 delta = p->prio - old_prio;
4005 enqueue_task(rq, p, 0);
4008 * If the task increased its priority or is running and
4009 * lowered its priority, then reschedule its CPU:
4011 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4012 resched_task(rq->curr);
4015 task_rq_unlock(rq, &flags);
4017 EXPORT_SYMBOL(set_user_nice);
4020 * can_nice - check if a task can reduce its nice value
4024 int can_nice(const struct task_struct *p, const int nice)
4026 /* convert nice value [19,-20] to rlimit style value [1,40] */
4027 int nice_rlim = 20 - nice;
4029 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4030 capable(CAP_SYS_NICE));
4033 #ifdef __ARCH_WANT_SYS_NICE
4036 * sys_nice - change the priority of the current process.
4037 * @increment: priority increment
4039 * sys_setpriority is a more generic, but much slower function that
4040 * does similar things.
4042 asmlinkage long sys_nice(int increment)
4047 * Setpriority might change our priority at the same moment.
4048 * We don't have to worry. Conceptually one call occurs first
4049 * and we have a single winner.
4051 if (increment < -40)
4056 nice = PRIO_TO_NICE(current->static_prio) + increment;
4062 if (increment < 0 && !can_nice(current, nice))
4065 retval = security_task_setnice(current, nice);
4069 set_user_nice(current, nice);
4076 * task_prio - return the priority value of a given task.
4077 * @p: the task in question.
4079 * This is the priority value as seen by users in /proc.
4080 * RT tasks are offset by -200. Normal tasks are centered
4081 * around 0, value goes from -16 to +15.
4083 int task_prio(const struct task_struct *p)
4085 return p->prio - MAX_RT_PRIO;
4089 * task_nice - return the nice value of a given task.
4090 * @p: the task in question.
4092 int task_nice(const struct task_struct *p)
4094 return TASK_NICE(p);
4096 EXPORT_SYMBOL_GPL(task_nice);
4099 * idle_cpu - is a given cpu idle currently?
4100 * @cpu: the processor in question.
4102 int idle_cpu(int cpu)
4104 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4108 * idle_task - return the idle task for a given cpu.
4109 * @cpu: the processor in question.
4111 struct task_struct *idle_task(int cpu)
4113 return cpu_rq(cpu)->idle;
4117 * find_process_by_pid - find a process with a matching PID value.
4118 * @pid: the pid in question.
4120 static struct task_struct *find_process_by_pid(pid_t pid)
4122 return pid ? find_task_by_pid(pid) : current;
4125 /* Actually do priority change: must hold rq lock. */
4127 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4129 BUG_ON(p->se.on_rq);
4132 switch (p->policy) {
4136 p->sched_class = &fair_sched_class;
4140 p->sched_class = &rt_sched_class;
4144 p->rt_priority = prio;
4145 p->normal_prio = normal_prio(p);
4146 /* we are holding p->pi_lock already */
4147 p->prio = rt_mutex_getprio(p);
4152 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4153 * @p: the task in question.
4154 * @policy: new policy.
4155 * @param: structure containing the new RT priority.
4157 * NOTE that the task may be already dead.
4159 int sched_setscheduler(struct task_struct *p, int policy,
4160 struct sched_param *param)
4162 int retval, oldprio, oldpolicy = -1, on_rq, running;
4163 unsigned long flags;
4166 /* may grab non-irq protected spin_locks */
4167 BUG_ON(in_interrupt());
4169 /* double check policy once rq lock held */
4171 policy = oldpolicy = p->policy;
4172 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4173 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4174 policy != SCHED_IDLE)
4177 * Valid priorities for SCHED_FIFO and SCHED_RR are
4178 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4179 * SCHED_BATCH and SCHED_IDLE is 0.
4181 if (param->sched_priority < 0 ||
4182 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4183 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4185 if (rt_policy(policy) != (param->sched_priority != 0))
4189 * Allow unprivileged RT tasks to decrease priority:
4191 if (!capable(CAP_SYS_NICE)) {
4192 if (rt_policy(policy)) {
4193 unsigned long rlim_rtprio;
4195 if (!lock_task_sighand(p, &flags))
4197 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4198 unlock_task_sighand(p, &flags);
4200 /* can't set/change the rt policy */
4201 if (policy != p->policy && !rlim_rtprio)
4204 /* can't increase priority */
4205 if (param->sched_priority > p->rt_priority &&
4206 param->sched_priority > rlim_rtprio)
4210 * Like positive nice levels, dont allow tasks to
4211 * move out of SCHED_IDLE either:
4213 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4216 /* can't change other user's priorities */
4217 if ((current->euid != p->euid) &&
4218 (current->euid != p->uid))
4222 retval = security_task_setscheduler(p, policy, param);
4226 * make sure no PI-waiters arrive (or leave) while we are
4227 * changing the priority of the task:
4229 spin_lock_irqsave(&p->pi_lock, flags);
4231 * To be able to change p->policy safely, the apropriate
4232 * runqueue lock must be held.
4234 rq = __task_rq_lock(p);
4235 /* recheck policy now with rq lock held */
4236 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4237 policy = oldpolicy = -1;
4238 __task_rq_unlock(rq);
4239 spin_unlock_irqrestore(&p->pi_lock, flags);
4242 update_rq_clock(rq);
4243 on_rq = p->se.on_rq;
4244 running = task_running(rq, p);
4246 deactivate_task(rq, p, 0);
4248 p->sched_class->put_prev_task(rq, p);
4252 __setscheduler(rq, p, policy, param->sched_priority);
4256 p->sched_class->set_curr_task(rq);
4257 activate_task(rq, p, 0);
4259 * Reschedule if we are currently running on this runqueue and
4260 * our priority decreased, or if we are not currently running on
4261 * this runqueue and our priority is higher than the current's
4264 if (p->prio > oldprio)
4265 resched_task(rq->curr);
4267 check_preempt_curr(rq, p);
4270 __task_rq_unlock(rq);
4271 spin_unlock_irqrestore(&p->pi_lock, flags);
4273 rt_mutex_adjust_pi(p);
4277 EXPORT_SYMBOL_GPL(sched_setscheduler);
4280 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4282 struct sched_param lparam;
4283 struct task_struct *p;
4286 if (!param || pid < 0)
4288 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4293 p = find_process_by_pid(pid);
4295 retval = sched_setscheduler(p, policy, &lparam);
4302 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4303 * @pid: the pid in question.
4304 * @policy: new policy.
4305 * @param: structure containing the new RT priority.
4307 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4308 struct sched_param __user *param)
4310 /* negative values for policy are not valid */
4314 return do_sched_setscheduler(pid, policy, param);
4318 * sys_sched_setparam - set/change the RT priority of a thread
4319 * @pid: the pid in question.
4320 * @param: structure containing the new RT priority.
4322 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4324 return do_sched_setscheduler(pid, -1, param);
4328 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4329 * @pid: the pid in question.
4331 asmlinkage long sys_sched_getscheduler(pid_t pid)
4333 struct task_struct *p;
4340 read_lock(&tasklist_lock);
4341 p = find_process_by_pid(pid);
4343 retval = security_task_getscheduler(p);
4347 read_unlock(&tasklist_lock);
4352 * sys_sched_getscheduler - get the RT priority of a thread
4353 * @pid: the pid in question.
4354 * @param: structure containing the RT priority.
4356 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4358 struct sched_param lp;
4359 struct task_struct *p;
4362 if (!param || pid < 0)
4365 read_lock(&tasklist_lock);
4366 p = find_process_by_pid(pid);
4371 retval = security_task_getscheduler(p);
4375 lp.sched_priority = p->rt_priority;
4376 read_unlock(&tasklist_lock);
4379 * This one might sleep, we cannot do it with a spinlock held ...
4381 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4386 read_unlock(&tasklist_lock);
4390 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4392 cpumask_t cpus_allowed;
4393 struct task_struct *p;
4396 mutex_lock(&sched_hotcpu_mutex);
4397 read_lock(&tasklist_lock);
4399 p = find_process_by_pid(pid);
4401 read_unlock(&tasklist_lock);
4402 mutex_unlock(&sched_hotcpu_mutex);
4407 * It is not safe to call set_cpus_allowed with the
4408 * tasklist_lock held. We will bump the task_struct's
4409 * usage count and then drop tasklist_lock.
4412 read_unlock(&tasklist_lock);
4415 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4416 !capable(CAP_SYS_NICE))
4419 retval = security_task_setscheduler(p, 0, NULL);
4423 cpus_allowed = cpuset_cpus_allowed(p);
4424 cpus_and(new_mask, new_mask, cpus_allowed);
4425 retval = set_cpus_allowed(p, new_mask);
4429 mutex_unlock(&sched_hotcpu_mutex);
4433 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4434 cpumask_t *new_mask)
4436 if (len < sizeof(cpumask_t)) {
4437 memset(new_mask, 0, sizeof(cpumask_t));
4438 } else if (len > sizeof(cpumask_t)) {
4439 len = sizeof(cpumask_t);
4441 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4445 * sys_sched_setaffinity - set the cpu affinity of a process
4446 * @pid: pid of the process
4447 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4448 * @user_mask_ptr: user-space pointer to the new cpu mask
4450 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4451 unsigned long __user *user_mask_ptr)
4456 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4460 return sched_setaffinity(pid, new_mask);
4464 * Represents all cpu's present in the system
4465 * In systems capable of hotplug, this map could dynamically grow
4466 * as new cpu's are detected in the system via any platform specific
4467 * method, such as ACPI for e.g.
4470 cpumask_t cpu_present_map __read_mostly;
4471 EXPORT_SYMBOL(cpu_present_map);
4474 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4475 EXPORT_SYMBOL(cpu_online_map);
4477 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4478 EXPORT_SYMBOL(cpu_possible_map);
4481 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4483 struct task_struct *p;
4486 mutex_lock(&sched_hotcpu_mutex);
4487 read_lock(&tasklist_lock);
4490 p = find_process_by_pid(pid);
4494 retval = security_task_getscheduler(p);
4498 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4501 read_unlock(&tasklist_lock);
4502 mutex_unlock(&sched_hotcpu_mutex);
4508 * sys_sched_getaffinity - get the cpu affinity of a process
4509 * @pid: pid of the process
4510 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4511 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4513 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4514 unsigned long __user *user_mask_ptr)
4519 if (len < sizeof(cpumask_t))
4522 ret = sched_getaffinity(pid, &mask);
4526 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4529 return sizeof(cpumask_t);
4533 * sys_sched_yield - yield the current processor to other threads.
4535 * This function yields the current CPU to other tasks. If there are no
4536 * other threads running on this CPU then this function will return.
4538 asmlinkage long sys_sched_yield(void)
4540 struct rq *rq = this_rq_lock();
4542 schedstat_inc(rq, yld_count);
4543 current->sched_class->yield_task(rq);
4546 * Since we are going to call schedule() anyway, there's
4547 * no need to preempt or enable interrupts:
4549 __release(rq->lock);
4550 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4551 _raw_spin_unlock(&rq->lock);
4552 preempt_enable_no_resched();
4559 static void __cond_resched(void)
4561 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4562 __might_sleep(__FILE__, __LINE__);
4565 * The BKS might be reacquired before we have dropped
4566 * PREEMPT_ACTIVE, which could trigger a second
4567 * cond_resched() call.
4570 add_preempt_count(PREEMPT_ACTIVE);
4572 sub_preempt_count(PREEMPT_ACTIVE);
4573 } while (need_resched());
4576 int __sched cond_resched(void)
4578 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4579 system_state == SYSTEM_RUNNING) {
4585 EXPORT_SYMBOL(cond_resched);
4588 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4589 * call schedule, and on return reacquire the lock.
4591 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4592 * operations here to prevent schedule() from being called twice (once via
4593 * spin_unlock(), once by hand).
4595 int cond_resched_lock(spinlock_t *lock)
4599 if (need_lockbreak(lock)) {
4605 if (need_resched() && system_state == SYSTEM_RUNNING) {
4606 spin_release(&lock->dep_map, 1, _THIS_IP_);
4607 _raw_spin_unlock(lock);
4608 preempt_enable_no_resched();
4615 EXPORT_SYMBOL(cond_resched_lock);
4617 int __sched cond_resched_softirq(void)
4619 BUG_ON(!in_softirq());
4621 if (need_resched() && system_state == SYSTEM_RUNNING) {
4629 EXPORT_SYMBOL(cond_resched_softirq);
4632 * yield - yield the current processor to other threads.
4634 * This is a shortcut for kernel-space yielding - it marks the
4635 * thread runnable and calls sys_sched_yield().
4637 void __sched yield(void)
4639 set_current_state(TASK_RUNNING);
4642 EXPORT_SYMBOL(yield);
4645 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4646 * that process accounting knows that this is a task in IO wait state.
4648 * But don't do that if it is a deliberate, throttling IO wait (this task
4649 * has set its backing_dev_info: the queue against which it should throttle)
4651 void __sched io_schedule(void)
4653 struct rq *rq = &__raw_get_cpu_var(runqueues);
4655 delayacct_blkio_start();
4656 atomic_inc(&rq->nr_iowait);
4658 atomic_dec(&rq->nr_iowait);
4659 delayacct_blkio_end();
4661 EXPORT_SYMBOL(io_schedule);
4663 long __sched io_schedule_timeout(long timeout)
4665 struct rq *rq = &__raw_get_cpu_var(runqueues);
4668 delayacct_blkio_start();
4669 atomic_inc(&rq->nr_iowait);
4670 ret = schedule_timeout(timeout);
4671 atomic_dec(&rq->nr_iowait);
4672 delayacct_blkio_end();
4677 * sys_sched_get_priority_max - return maximum RT priority.
4678 * @policy: scheduling class.
4680 * this syscall returns the maximum rt_priority that can be used
4681 * by a given scheduling class.
4683 asmlinkage long sys_sched_get_priority_max(int policy)
4690 ret = MAX_USER_RT_PRIO-1;
4702 * sys_sched_get_priority_min - return minimum RT priority.
4703 * @policy: scheduling class.
4705 * this syscall returns the minimum rt_priority that can be used
4706 * by a given scheduling class.
4708 asmlinkage long sys_sched_get_priority_min(int policy)
4726 * sys_sched_rr_get_interval - return the default timeslice of a process.
4727 * @pid: pid of the process.
4728 * @interval: userspace pointer to the timeslice value.
4730 * this syscall writes the default timeslice value of a given process
4731 * into the user-space timespec buffer. A value of '0' means infinity.
4734 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4736 struct task_struct *p;
4737 unsigned int time_slice;
4745 read_lock(&tasklist_lock);
4746 p = find_process_by_pid(pid);
4750 retval = security_task_getscheduler(p);
4754 if (p->policy == SCHED_FIFO)
4756 else if (p->policy == SCHED_RR)
4757 time_slice = DEF_TIMESLICE;
4759 struct sched_entity *se = &p->se;
4760 unsigned long flags;
4763 rq = task_rq_lock(p, &flags);
4764 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4765 task_rq_unlock(rq, &flags);
4767 read_unlock(&tasklist_lock);
4768 jiffies_to_timespec(time_slice, &t);
4769 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4773 read_unlock(&tasklist_lock);
4777 static const char stat_nam[] = "RSDTtZX";
4779 static void show_task(struct task_struct *p)
4781 unsigned long free = 0;
4784 state = p->state ? __ffs(p->state) + 1 : 0;
4785 printk("%-13.13s %c", p->comm,
4786 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4787 #if BITS_PER_LONG == 32
4788 if (state == TASK_RUNNING)
4789 printk(" running ");
4791 printk(" %08lx ", thread_saved_pc(p));
4793 if (state == TASK_RUNNING)
4794 printk(" running task ");
4796 printk(" %016lx ", thread_saved_pc(p));
4798 #ifdef CONFIG_DEBUG_STACK_USAGE
4800 unsigned long *n = end_of_stack(p);
4803 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4806 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4808 if (state != TASK_RUNNING)
4809 show_stack(p, NULL);
4812 void show_state_filter(unsigned long state_filter)
4814 struct task_struct *g, *p;
4816 #if BITS_PER_LONG == 32
4818 " task PC stack pid father\n");
4821 " task PC stack pid father\n");
4823 read_lock(&tasklist_lock);
4824 do_each_thread(g, p) {
4826 * reset the NMI-timeout, listing all files on a slow
4827 * console might take alot of time:
4829 touch_nmi_watchdog();
4830 if (!state_filter || (p->state & state_filter))
4832 } while_each_thread(g, p);
4834 touch_all_softlockup_watchdogs();
4836 #ifdef CONFIG_SCHED_DEBUG
4837 sysrq_sched_debug_show();
4839 read_unlock(&tasklist_lock);
4841 * Only show locks if all tasks are dumped:
4843 if (state_filter == -1)
4844 debug_show_all_locks();
4847 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4849 idle->sched_class = &idle_sched_class;
4853 * init_idle - set up an idle thread for a given CPU
4854 * @idle: task in question
4855 * @cpu: cpu the idle task belongs to
4857 * NOTE: this function does not set the idle thread's NEED_RESCHED
4858 * flag, to make booting more robust.
4860 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4862 struct rq *rq = cpu_rq(cpu);
4863 unsigned long flags;
4866 idle->se.exec_start = sched_clock();
4868 idle->prio = idle->normal_prio = MAX_PRIO;
4869 idle->cpus_allowed = cpumask_of_cpu(cpu);
4870 __set_task_cpu(idle, cpu);
4872 spin_lock_irqsave(&rq->lock, flags);
4873 rq->curr = rq->idle = idle;
4874 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4877 spin_unlock_irqrestore(&rq->lock, flags);
4879 /* Set the preempt count _outside_ the spinlocks! */
4880 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4881 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4883 task_thread_info(idle)->preempt_count = 0;
4886 * The idle tasks have their own, simple scheduling class:
4888 idle->sched_class = &idle_sched_class;
4892 * In a system that switches off the HZ timer nohz_cpu_mask
4893 * indicates which cpus entered this state. This is used
4894 * in the rcu update to wait only for active cpus. For system
4895 * which do not switch off the HZ timer nohz_cpu_mask should
4896 * always be CPU_MASK_NONE.
4898 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4902 * This is how migration works:
4904 * 1) we queue a struct migration_req structure in the source CPU's
4905 * runqueue and wake up that CPU's migration thread.
4906 * 2) we down() the locked semaphore => thread blocks.
4907 * 3) migration thread wakes up (implicitly it forces the migrated
4908 * thread off the CPU)
4909 * 4) it gets the migration request and checks whether the migrated
4910 * task is still in the wrong runqueue.
4911 * 5) if it's in the wrong runqueue then the migration thread removes
4912 * it and puts it into the right queue.
4913 * 6) migration thread up()s the semaphore.
4914 * 7) we wake up and the migration is done.
4918 * Change a given task's CPU affinity. Migrate the thread to a
4919 * proper CPU and schedule it away if the CPU it's executing on
4920 * is removed from the allowed bitmask.
4922 * NOTE: the caller must have a valid reference to the task, the
4923 * task must not exit() & deallocate itself prematurely. The
4924 * call is not atomic; no spinlocks may be held.
4926 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4928 struct migration_req req;
4929 unsigned long flags;
4933 rq = task_rq_lock(p, &flags);
4934 if (!cpus_intersects(new_mask, cpu_online_map)) {
4939 p->cpus_allowed = new_mask;
4940 /* Can the task run on the task's current CPU? If so, we're done */
4941 if (cpu_isset(task_cpu(p), new_mask))
4944 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4945 /* Need help from migration thread: drop lock and wait. */
4946 task_rq_unlock(rq, &flags);
4947 wake_up_process(rq->migration_thread);
4948 wait_for_completion(&req.done);
4949 tlb_migrate_finish(p->mm);
4953 task_rq_unlock(rq, &flags);
4957 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4960 * Move (not current) task off this cpu, onto dest cpu. We're doing
4961 * this because either it can't run here any more (set_cpus_allowed()
4962 * away from this CPU, or CPU going down), or because we're
4963 * attempting to rebalance this task on exec (sched_exec).
4965 * So we race with normal scheduler movements, but that's OK, as long
4966 * as the task is no longer on this CPU.
4968 * Returns non-zero if task was successfully migrated.
4970 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4972 struct rq *rq_dest, *rq_src;
4975 if (unlikely(cpu_is_offline(dest_cpu)))
4978 rq_src = cpu_rq(src_cpu);
4979 rq_dest = cpu_rq(dest_cpu);
4981 double_rq_lock(rq_src, rq_dest);
4982 /* Already moved. */
4983 if (task_cpu(p) != src_cpu)
4985 /* Affinity changed (again). */
4986 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4989 on_rq = p->se.on_rq;
4991 deactivate_task(rq_src, p, 0);
4993 set_task_cpu(p, dest_cpu);
4995 activate_task(rq_dest, p, 0);
4996 check_preempt_curr(rq_dest, p);
5000 double_rq_unlock(rq_src, rq_dest);
5005 * migration_thread - this is a highprio system thread that performs
5006 * thread migration by bumping thread off CPU then 'pushing' onto
5009 static int migration_thread(void *data)
5011 int cpu = (long)data;
5015 BUG_ON(rq->migration_thread != current);
5017 set_current_state(TASK_INTERRUPTIBLE);
5018 while (!kthread_should_stop()) {
5019 struct migration_req *req;
5020 struct list_head *head;
5022 spin_lock_irq(&rq->lock);
5024 if (cpu_is_offline(cpu)) {
5025 spin_unlock_irq(&rq->lock);
5029 if (rq->active_balance) {
5030 active_load_balance(rq, cpu);
5031 rq->active_balance = 0;
5034 head = &rq->migration_queue;
5036 if (list_empty(head)) {
5037 spin_unlock_irq(&rq->lock);
5039 set_current_state(TASK_INTERRUPTIBLE);
5042 req = list_entry(head->next, struct migration_req, list);
5043 list_del_init(head->next);
5045 spin_unlock(&rq->lock);
5046 __migrate_task(req->task, cpu, req->dest_cpu);
5049 complete(&req->done);
5051 __set_current_state(TASK_RUNNING);
5055 /* Wait for kthread_stop */
5056 set_current_state(TASK_INTERRUPTIBLE);
5057 while (!kthread_should_stop()) {
5059 set_current_state(TASK_INTERRUPTIBLE);
5061 __set_current_state(TASK_RUNNING);
5065 #ifdef CONFIG_HOTPLUG_CPU
5067 * Figure out where task on dead CPU should go, use force if neccessary.
5068 * NOTE: interrupts should be disabled by the caller
5070 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5072 unsigned long flags;
5079 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5080 cpus_and(mask, mask, p->cpus_allowed);
5081 dest_cpu = any_online_cpu(mask);
5083 /* On any allowed CPU? */
5084 if (dest_cpu == NR_CPUS)
5085 dest_cpu = any_online_cpu(p->cpus_allowed);
5087 /* No more Mr. Nice Guy. */
5088 if (dest_cpu == NR_CPUS) {
5089 rq = task_rq_lock(p, &flags);
5090 cpus_setall(p->cpus_allowed);
5091 dest_cpu = any_online_cpu(p->cpus_allowed);
5092 task_rq_unlock(rq, &flags);
5095 * Don't tell them about moving exiting tasks or
5096 * kernel threads (both mm NULL), since they never
5099 if (p->mm && printk_ratelimit())
5100 printk(KERN_INFO "process %d (%s) no "
5101 "longer affine to cpu%d\n",
5102 p->pid, p->comm, dead_cpu);
5104 } while (!__migrate_task(p, dead_cpu, dest_cpu));
5108 * While a dead CPU has no uninterruptible tasks queued at this point,
5109 * it might still have a nonzero ->nr_uninterruptible counter, because
5110 * for performance reasons the counter is not stricly tracking tasks to
5111 * their home CPUs. So we just add the counter to another CPU's counter,
5112 * to keep the global sum constant after CPU-down:
5114 static void migrate_nr_uninterruptible(struct rq *rq_src)
5116 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5117 unsigned long flags;
5119 local_irq_save(flags);
5120 double_rq_lock(rq_src, rq_dest);
5121 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5122 rq_src->nr_uninterruptible = 0;
5123 double_rq_unlock(rq_src, rq_dest);
5124 local_irq_restore(flags);
5127 /* Run through task list and migrate tasks from the dead cpu. */
5128 static void migrate_live_tasks(int src_cpu)
5130 struct task_struct *p, *t;
5132 write_lock_irq(&tasklist_lock);
5134 do_each_thread(t, p) {
5138 if (task_cpu(p) == src_cpu)
5139 move_task_off_dead_cpu(src_cpu, p);
5140 } while_each_thread(t, p);
5142 write_unlock_irq(&tasklist_lock);
5146 * activate_idle_task - move idle task to the _front_ of runqueue.
5148 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5150 update_rq_clock(rq);
5152 if (p->state == TASK_UNINTERRUPTIBLE)
5153 rq->nr_uninterruptible--;
5155 enqueue_task(rq, p, 0);
5156 inc_nr_running(p, rq);
5160 * Schedules idle task to be the next runnable task on current CPU.
5161 * It does so by boosting its priority to highest possible and adding it to
5162 * the _front_ of the runqueue. Used by CPU offline code.
5164 void sched_idle_next(void)
5166 int this_cpu = smp_processor_id();
5167 struct rq *rq = cpu_rq(this_cpu);
5168 struct task_struct *p = rq->idle;
5169 unsigned long flags;
5171 /* cpu has to be offline */
5172 BUG_ON(cpu_online(this_cpu));
5175 * Strictly not necessary since rest of the CPUs are stopped by now
5176 * and interrupts disabled on the current cpu.
5178 spin_lock_irqsave(&rq->lock, flags);
5180 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5182 /* Add idle task to the _front_ of its priority queue: */
5183 activate_idle_task(p, rq);
5185 spin_unlock_irqrestore(&rq->lock, flags);
5189 * Ensures that the idle task is using init_mm right before its cpu goes
5192 void idle_task_exit(void)
5194 struct mm_struct *mm = current->active_mm;
5196 BUG_ON(cpu_online(smp_processor_id()));
5199 switch_mm(mm, &init_mm, current);
5203 /* called under rq->lock with disabled interrupts */
5204 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5206 struct rq *rq = cpu_rq(dead_cpu);
5208 /* Must be exiting, otherwise would be on tasklist. */
5209 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5211 /* Cannot have done final schedule yet: would have vanished. */
5212 BUG_ON(p->state == TASK_DEAD);
5217 * Drop lock around migration; if someone else moves it,
5218 * that's OK. No task can be added to this CPU, so iteration is
5220 * NOTE: interrupts should be left disabled --dev@
5222 spin_unlock(&rq->lock);
5223 move_task_off_dead_cpu(dead_cpu, p);
5224 spin_lock(&rq->lock);
5229 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5230 static void migrate_dead_tasks(unsigned int dead_cpu)
5232 struct rq *rq = cpu_rq(dead_cpu);
5233 struct task_struct *next;
5236 if (!rq->nr_running)
5238 update_rq_clock(rq);
5239 next = pick_next_task(rq, rq->curr);
5242 migrate_dead(dead_cpu, next);
5246 #endif /* CONFIG_HOTPLUG_CPU */
5248 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5250 static struct ctl_table sd_ctl_dir[] = {
5252 .procname = "sched_domain",
5258 static struct ctl_table sd_ctl_root[] = {
5260 .ctl_name = CTL_KERN,
5261 .procname = "kernel",
5263 .child = sd_ctl_dir,
5268 static struct ctl_table *sd_alloc_ctl_entry(int n)
5270 struct ctl_table *entry =
5271 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5276 static void sd_free_ctl_entry(struct ctl_table **tablep)
5278 struct ctl_table *entry = *tablep;
5280 for (entry = *tablep; entry->procname; entry++)
5282 sd_free_ctl_entry(&entry->child);
5289 set_table_entry(struct ctl_table *entry,
5290 const char *procname, void *data, int maxlen,
5291 mode_t mode, proc_handler *proc_handler)
5293 entry->procname = procname;
5295 entry->maxlen = maxlen;
5297 entry->proc_handler = proc_handler;
5300 static struct ctl_table *
5301 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5303 struct ctl_table *table = sd_alloc_ctl_entry(12);
5308 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5309 sizeof(long), 0644, proc_doulongvec_minmax);
5310 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5311 sizeof(long), 0644, proc_doulongvec_minmax);
5312 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5313 sizeof(int), 0644, proc_dointvec_minmax);
5314 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5315 sizeof(int), 0644, proc_dointvec_minmax);
5316 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5317 sizeof(int), 0644, proc_dointvec_minmax);
5318 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5319 sizeof(int), 0644, proc_dointvec_minmax);
5320 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5321 sizeof(int), 0644, proc_dointvec_minmax);
5322 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5323 sizeof(int), 0644, proc_dointvec_minmax);
5324 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5325 sizeof(int), 0644, proc_dointvec_minmax);
5326 set_table_entry(&table[9], "cache_nice_tries",
5327 &sd->cache_nice_tries,
5328 sizeof(int), 0644, proc_dointvec_minmax);
5329 set_table_entry(&table[10], "flags", &sd->flags,
5330 sizeof(int), 0644, proc_dointvec_minmax);
5331 /* &table[11] is terminator */
5336 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5338 struct ctl_table *entry, *table;
5339 struct sched_domain *sd;
5340 int domain_num = 0, i;
5343 for_each_domain(cpu, sd)
5345 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5350 for_each_domain(cpu, sd) {
5351 snprintf(buf, 32, "domain%d", i);
5352 entry->procname = kstrdup(buf, GFP_KERNEL);
5354 entry->child = sd_alloc_ctl_domain_table(sd);
5361 static struct ctl_table_header *sd_sysctl_header;
5362 static void register_sched_domain_sysctl(void)
5364 int i, cpu_num = num_online_cpus();
5365 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5371 sd_ctl_dir[0].child = entry;
5373 for_each_online_cpu(i) {
5374 snprintf(buf, 32, "cpu%d", i);
5375 entry->procname = kstrdup(buf, GFP_KERNEL);
5377 entry->child = sd_alloc_ctl_cpu_table(i);
5380 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5383 static void unregister_sched_domain_sysctl(void)
5385 unregister_sysctl_table(sd_sysctl_header);
5386 sd_sysctl_header = NULL;
5387 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5390 static void register_sched_domain_sysctl(void)
5393 static void unregister_sched_domain_sysctl(void)
5399 * migration_call - callback that gets triggered when a CPU is added.
5400 * Here we can start up the necessary migration thread for the new CPU.
5402 static int __cpuinit
5403 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5405 struct task_struct *p;
5406 int cpu = (long)hcpu;
5407 unsigned long flags;
5411 case CPU_LOCK_ACQUIRE:
5412 mutex_lock(&sched_hotcpu_mutex);
5415 case CPU_UP_PREPARE:
5416 case CPU_UP_PREPARE_FROZEN:
5417 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5420 kthread_bind(p, cpu);
5421 /* Must be high prio: stop_machine expects to yield to it. */
5422 rq = task_rq_lock(p, &flags);
5423 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5424 task_rq_unlock(rq, &flags);
5425 cpu_rq(cpu)->migration_thread = p;
5429 case CPU_ONLINE_FROZEN:
5430 /* Strictly unneccessary, as first user will wake it. */
5431 wake_up_process(cpu_rq(cpu)->migration_thread);
5434 #ifdef CONFIG_HOTPLUG_CPU
5435 case CPU_UP_CANCELED:
5436 case CPU_UP_CANCELED_FROZEN:
5437 if (!cpu_rq(cpu)->migration_thread)
5439 /* Unbind it from offline cpu so it can run. Fall thru. */
5440 kthread_bind(cpu_rq(cpu)->migration_thread,
5441 any_online_cpu(cpu_online_map));
5442 kthread_stop(cpu_rq(cpu)->migration_thread);
5443 cpu_rq(cpu)->migration_thread = NULL;
5447 case CPU_DEAD_FROZEN:
5448 migrate_live_tasks(cpu);
5450 kthread_stop(rq->migration_thread);
5451 rq->migration_thread = NULL;
5452 /* Idle task back to normal (off runqueue, low prio) */
5453 rq = task_rq_lock(rq->idle, &flags);
5454 update_rq_clock(rq);
5455 deactivate_task(rq, rq->idle, 0);
5456 rq->idle->static_prio = MAX_PRIO;
5457 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5458 rq->idle->sched_class = &idle_sched_class;
5459 migrate_dead_tasks(cpu);
5460 task_rq_unlock(rq, &flags);
5461 migrate_nr_uninterruptible(rq);
5462 BUG_ON(rq->nr_running != 0);
5464 /* No need to migrate the tasks: it was best-effort if
5465 * they didn't take sched_hotcpu_mutex. Just wake up
5466 * the requestors. */
5467 spin_lock_irq(&rq->lock);
5468 while (!list_empty(&rq->migration_queue)) {
5469 struct migration_req *req;
5471 req = list_entry(rq->migration_queue.next,
5472 struct migration_req, list);
5473 list_del_init(&req->list);
5474 complete(&req->done);
5476 spin_unlock_irq(&rq->lock);
5479 case CPU_LOCK_RELEASE:
5480 mutex_unlock(&sched_hotcpu_mutex);
5486 /* Register at highest priority so that task migration (migrate_all_tasks)
5487 * happens before everything else.
5489 static struct notifier_block __cpuinitdata migration_notifier = {
5490 .notifier_call = migration_call,
5494 int __init migration_init(void)
5496 void *cpu = (void *)(long)smp_processor_id();
5499 /* Start one for the boot CPU: */
5500 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5501 BUG_ON(err == NOTIFY_BAD);
5502 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5503 register_cpu_notifier(&migration_notifier);
5511 /* Number of possible processor ids */
5512 int nr_cpu_ids __read_mostly = NR_CPUS;
5513 EXPORT_SYMBOL(nr_cpu_ids);
5515 #ifdef CONFIG_SCHED_DEBUG
5516 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5521 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5525 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5530 struct sched_group *group = sd->groups;
5531 cpumask_t groupmask;
5533 cpumask_scnprintf(str, NR_CPUS, sd->span);
5534 cpus_clear(groupmask);
5537 for (i = 0; i < level + 1; i++)
5539 printk("domain %d: ", level);
5541 if (!(sd->flags & SD_LOAD_BALANCE)) {
5542 printk("does not load-balance\n");
5544 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5549 printk("span %s\n", str);
5551 if (!cpu_isset(cpu, sd->span))
5552 printk(KERN_ERR "ERROR: domain->span does not contain "
5554 if (!cpu_isset(cpu, group->cpumask))
5555 printk(KERN_ERR "ERROR: domain->groups does not contain"
5559 for (i = 0; i < level + 2; i++)
5565 printk(KERN_ERR "ERROR: group is NULL\n");
5569 if (!group->__cpu_power) {
5571 printk(KERN_ERR "ERROR: domain->cpu_power not "
5576 if (!cpus_weight(group->cpumask)) {
5578 printk(KERN_ERR "ERROR: empty group\n");
5582 if (cpus_intersects(groupmask, group->cpumask)) {
5584 printk(KERN_ERR "ERROR: repeated CPUs\n");
5588 cpus_or(groupmask, groupmask, group->cpumask);
5590 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5593 group = group->next;
5594 } while (group != sd->groups);
5597 if (!cpus_equal(sd->span, groupmask))
5598 printk(KERN_ERR "ERROR: groups don't span "
5606 if (!cpus_subset(groupmask, sd->span))
5607 printk(KERN_ERR "ERROR: parent span is not a superset "
5608 "of domain->span\n");
5613 # define sched_domain_debug(sd, cpu) do { } while (0)
5616 static int sd_degenerate(struct sched_domain *sd)
5618 if (cpus_weight(sd->span) == 1)
5621 /* Following flags need at least 2 groups */
5622 if (sd->flags & (SD_LOAD_BALANCE |
5623 SD_BALANCE_NEWIDLE |
5627 SD_SHARE_PKG_RESOURCES)) {
5628 if (sd->groups != sd->groups->next)
5632 /* Following flags don't use groups */
5633 if (sd->flags & (SD_WAKE_IDLE |
5642 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5644 unsigned long cflags = sd->flags, pflags = parent->flags;
5646 if (sd_degenerate(parent))
5649 if (!cpus_equal(sd->span, parent->span))
5652 /* Does parent contain flags not in child? */
5653 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5654 if (cflags & SD_WAKE_AFFINE)
5655 pflags &= ~SD_WAKE_BALANCE;
5656 /* Flags needing groups don't count if only 1 group in parent */
5657 if (parent->groups == parent->groups->next) {
5658 pflags &= ~(SD_LOAD_BALANCE |
5659 SD_BALANCE_NEWIDLE |
5663 SD_SHARE_PKG_RESOURCES);
5665 if (~cflags & pflags)
5672 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5673 * hold the hotplug lock.
5675 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5677 struct rq *rq = cpu_rq(cpu);
5678 struct sched_domain *tmp;
5680 /* Remove the sched domains which do not contribute to scheduling. */
5681 for (tmp = sd; tmp; tmp = tmp->parent) {
5682 struct sched_domain *parent = tmp->parent;
5685 if (sd_parent_degenerate(tmp, parent)) {
5686 tmp->parent = parent->parent;
5688 parent->parent->child = tmp;
5692 if (sd && sd_degenerate(sd)) {
5698 sched_domain_debug(sd, cpu);
5700 rcu_assign_pointer(rq->sd, sd);
5703 /* cpus with isolated domains */
5704 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5706 /* Setup the mask of cpus configured for isolated domains */
5707 static int __init isolated_cpu_setup(char *str)
5709 int ints[NR_CPUS], i;
5711 str = get_options(str, ARRAY_SIZE(ints), ints);
5712 cpus_clear(cpu_isolated_map);
5713 for (i = 1; i <= ints[0]; i++)
5714 if (ints[i] < NR_CPUS)
5715 cpu_set(ints[i], cpu_isolated_map);
5719 __setup("isolcpus=", isolated_cpu_setup);
5722 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5723 * to a function which identifies what group(along with sched group) a CPU
5724 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5725 * (due to the fact that we keep track of groups covered with a cpumask_t).
5727 * init_sched_build_groups will build a circular linked list of the groups
5728 * covered by the given span, and will set each group's ->cpumask correctly,
5729 * and ->cpu_power to 0.
5732 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5733 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5734 struct sched_group **sg))
5736 struct sched_group *first = NULL, *last = NULL;
5737 cpumask_t covered = CPU_MASK_NONE;
5740 for_each_cpu_mask(i, span) {
5741 struct sched_group *sg;
5742 int group = group_fn(i, cpu_map, &sg);
5745 if (cpu_isset(i, covered))
5748 sg->cpumask = CPU_MASK_NONE;
5749 sg->__cpu_power = 0;
5751 for_each_cpu_mask(j, span) {
5752 if (group_fn(j, cpu_map, NULL) != group)
5755 cpu_set(j, covered);
5756 cpu_set(j, sg->cpumask);
5767 #define SD_NODES_PER_DOMAIN 16
5772 * find_next_best_node - find the next node to include in a sched_domain
5773 * @node: node whose sched_domain we're building
5774 * @used_nodes: nodes already in the sched_domain
5776 * Find the next node to include in a given scheduling domain. Simply
5777 * finds the closest node not already in the @used_nodes map.
5779 * Should use nodemask_t.
5781 static int find_next_best_node(int node, unsigned long *used_nodes)
5783 int i, n, val, min_val, best_node = 0;
5787 for (i = 0; i < MAX_NUMNODES; i++) {
5788 /* Start at @node */
5789 n = (node + i) % MAX_NUMNODES;
5791 if (!nr_cpus_node(n))
5794 /* Skip already used nodes */
5795 if (test_bit(n, used_nodes))
5798 /* Simple min distance search */
5799 val = node_distance(node, n);
5801 if (val < min_val) {
5807 set_bit(best_node, used_nodes);
5812 * sched_domain_node_span - get a cpumask for a node's sched_domain
5813 * @node: node whose cpumask we're constructing
5814 * @size: number of nodes to include in this span
5816 * Given a node, construct a good cpumask for its sched_domain to span. It
5817 * should be one that prevents unnecessary balancing, but also spreads tasks
5820 static cpumask_t sched_domain_node_span(int node)
5822 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5823 cpumask_t span, nodemask;
5827 bitmap_zero(used_nodes, MAX_NUMNODES);
5829 nodemask = node_to_cpumask(node);
5830 cpus_or(span, span, nodemask);
5831 set_bit(node, used_nodes);
5833 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5834 int next_node = find_next_best_node(node, used_nodes);
5836 nodemask = node_to_cpumask(next_node);
5837 cpus_or(span, span, nodemask);
5844 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5847 * SMT sched-domains:
5849 #ifdef CONFIG_SCHED_SMT
5850 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5851 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5853 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5854 struct sched_group **sg)
5857 *sg = &per_cpu(sched_group_cpus, cpu);
5863 * multi-core sched-domains:
5865 #ifdef CONFIG_SCHED_MC
5866 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5867 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5870 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5871 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5872 struct sched_group **sg)
5875 cpumask_t mask = cpu_sibling_map[cpu];
5876 cpus_and(mask, mask, *cpu_map);
5877 group = first_cpu(mask);
5879 *sg = &per_cpu(sched_group_core, group);
5882 #elif defined(CONFIG_SCHED_MC)
5883 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5884 struct sched_group **sg)
5887 *sg = &per_cpu(sched_group_core, cpu);
5892 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5893 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5895 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5896 struct sched_group **sg)
5899 #ifdef CONFIG_SCHED_MC
5900 cpumask_t mask = cpu_coregroup_map(cpu);
5901 cpus_and(mask, mask, *cpu_map);
5902 group = first_cpu(mask);
5903 #elif defined(CONFIG_SCHED_SMT)
5904 cpumask_t mask = cpu_sibling_map[cpu];
5905 cpus_and(mask, mask, *cpu_map);
5906 group = first_cpu(mask);
5911 *sg = &per_cpu(sched_group_phys, group);
5917 * The init_sched_build_groups can't handle what we want to do with node
5918 * groups, so roll our own. Now each node has its own list of groups which
5919 * gets dynamically allocated.
5921 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5922 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5924 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5925 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5927 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5928 struct sched_group **sg)
5930 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5933 cpus_and(nodemask, nodemask, *cpu_map);
5934 group = first_cpu(nodemask);
5937 *sg = &per_cpu(sched_group_allnodes, group);
5941 static void init_numa_sched_groups_power(struct sched_group *group_head)
5943 struct sched_group *sg = group_head;
5949 for_each_cpu_mask(j, sg->cpumask) {
5950 struct sched_domain *sd;
5952 sd = &per_cpu(phys_domains, j);
5953 if (j != first_cpu(sd->groups->cpumask)) {
5955 * Only add "power" once for each
5961 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5964 } while (sg != group_head);
5969 /* Free memory allocated for various sched_group structures */
5970 static void free_sched_groups(const cpumask_t *cpu_map)
5974 for_each_cpu_mask(cpu, *cpu_map) {
5975 struct sched_group **sched_group_nodes
5976 = sched_group_nodes_bycpu[cpu];
5978 if (!sched_group_nodes)
5981 for (i = 0; i < MAX_NUMNODES; i++) {
5982 cpumask_t nodemask = node_to_cpumask(i);
5983 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5985 cpus_and(nodemask, nodemask, *cpu_map);
5986 if (cpus_empty(nodemask))
5996 if (oldsg != sched_group_nodes[i])
5999 kfree(sched_group_nodes);
6000 sched_group_nodes_bycpu[cpu] = NULL;
6004 static void free_sched_groups(const cpumask_t *cpu_map)
6010 * Initialize sched groups cpu_power.
6012 * cpu_power indicates the capacity of sched group, which is used while
6013 * distributing the load between different sched groups in a sched domain.
6014 * Typically cpu_power for all the groups in a sched domain will be same unless
6015 * there are asymmetries in the topology. If there are asymmetries, group
6016 * having more cpu_power will pickup more load compared to the group having
6019 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6020 * the maximum number of tasks a group can handle in the presence of other idle
6021 * or lightly loaded groups in the same sched domain.
6023 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6025 struct sched_domain *child;
6026 struct sched_group *group;
6028 WARN_ON(!sd || !sd->groups);
6030 if (cpu != first_cpu(sd->groups->cpumask))
6035 sd->groups->__cpu_power = 0;
6038 * For perf policy, if the groups in child domain share resources
6039 * (for example cores sharing some portions of the cache hierarchy
6040 * or SMT), then set this domain groups cpu_power such that each group
6041 * can handle only one task, when there are other idle groups in the
6042 * same sched domain.
6044 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6046 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6047 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6052 * add cpu_power of each child group to this groups cpu_power
6054 group = child->groups;
6056 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6057 group = group->next;
6058 } while (group != child->groups);
6062 * Build sched domains for a given set of cpus and attach the sched domains
6063 * to the individual cpus
6065 static int build_sched_domains(const cpumask_t *cpu_map)
6069 struct sched_group **sched_group_nodes = NULL;
6070 int sd_allnodes = 0;
6073 * Allocate the per-node list of sched groups
6075 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6077 if (!sched_group_nodes) {
6078 printk(KERN_WARNING "Can not alloc sched group node list\n");
6081 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6085 * Set up domains for cpus specified by the cpu_map.
6087 for_each_cpu_mask(i, *cpu_map) {
6088 struct sched_domain *sd = NULL, *p;
6089 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6091 cpus_and(nodemask, nodemask, *cpu_map);
6094 if (cpus_weight(*cpu_map) >
6095 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6096 sd = &per_cpu(allnodes_domains, i);
6097 *sd = SD_ALLNODES_INIT;
6098 sd->span = *cpu_map;
6099 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6105 sd = &per_cpu(node_domains, i);
6107 sd->span = sched_domain_node_span(cpu_to_node(i));
6111 cpus_and(sd->span, sd->span, *cpu_map);
6115 sd = &per_cpu(phys_domains, i);
6117 sd->span = nodemask;
6121 cpu_to_phys_group(i, cpu_map, &sd->groups);
6123 #ifdef CONFIG_SCHED_MC
6125 sd = &per_cpu(core_domains, i);
6127 sd->span = cpu_coregroup_map(i);
6128 cpus_and(sd->span, sd->span, *cpu_map);
6131 cpu_to_core_group(i, cpu_map, &sd->groups);
6134 #ifdef CONFIG_SCHED_SMT
6136 sd = &per_cpu(cpu_domains, i);
6137 *sd = SD_SIBLING_INIT;
6138 sd->span = cpu_sibling_map[i];
6139 cpus_and(sd->span, sd->span, *cpu_map);
6142 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6146 #ifdef CONFIG_SCHED_SMT
6147 /* Set up CPU (sibling) groups */
6148 for_each_cpu_mask(i, *cpu_map) {
6149 cpumask_t this_sibling_map = cpu_sibling_map[i];
6150 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6151 if (i != first_cpu(this_sibling_map))
6154 init_sched_build_groups(this_sibling_map, cpu_map,
6159 #ifdef CONFIG_SCHED_MC
6160 /* Set up multi-core groups */
6161 for_each_cpu_mask(i, *cpu_map) {
6162 cpumask_t this_core_map = cpu_coregroup_map(i);
6163 cpus_and(this_core_map, this_core_map, *cpu_map);
6164 if (i != first_cpu(this_core_map))
6166 init_sched_build_groups(this_core_map, cpu_map,
6167 &cpu_to_core_group);
6171 /* Set up physical groups */
6172 for (i = 0; i < MAX_NUMNODES; i++) {
6173 cpumask_t nodemask = node_to_cpumask(i);
6175 cpus_and(nodemask, nodemask, *cpu_map);
6176 if (cpus_empty(nodemask))
6179 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6183 /* Set up node groups */
6185 init_sched_build_groups(*cpu_map, cpu_map,
6186 &cpu_to_allnodes_group);
6188 for (i = 0; i < MAX_NUMNODES; i++) {
6189 /* Set up node groups */
6190 struct sched_group *sg, *prev;
6191 cpumask_t nodemask = node_to_cpumask(i);
6192 cpumask_t domainspan;
6193 cpumask_t covered = CPU_MASK_NONE;
6196 cpus_and(nodemask, nodemask, *cpu_map);
6197 if (cpus_empty(nodemask)) {
6198 sched_group_nodes[i] = NULL;
6202 domainspan = sched_domain_node_span(i);
6203 cpus_and(domainspan, domainspan, *cpu_map);
6205 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6207 printk(KERN_WARNING "Can not alloc domain group for "
6211 sched_group_nodes[i] = sg;
6212 for_each_cpu_mask(j, nodemask) {
6213 struct sched_domain *sd;
6215 sd = &per_cpu(node_domains, j);
6218 sg->__cpu_power = 0;
6219 sg->cpumask = nodemask;
6221 cpus_or(covered, covered, nodemask);
6224 for (j = 0; j < MAX_NUMNODES; j++) {
6225 cpumask_t tmp, notcovered;
6226 int n = (i + j) % MAX_NUMNODES;
6228 cpus_complement(notcovered, covered);
6229 cpus_and(tmp, notcovered, *cpu_map);
6230 cpus_and(tmp, tmp, domainspan);
6231 if (cpus_empty(tmp))
6234 nodemask = node_to_cpumask(n);
6235 cpus_and(tmp, tmp, nodemask);
6236 if (cpus_empty(tmp))
6239 sg = kmalloc_node(sizeof(struct sched_group),
6243 "Can not alloc domain group for node %d\n", j);
6246 sg->__cpu_power = 0;
6248 sg->next = prev->next;
6249 cpus_or(covered, covered, tmp);
6256 /* Calculate CPU power for physical packages and nodes */
6257 #ifdef CONFIG_SCHED_SMT
6258 for_each_cpu_mask(i, *cpu_map) {
6259 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6261 init_sched_groups_power(i, sd);
6264 #ifdef CONFIG_SCHED_MC
6265 for_each_cpu_mask(i, *cpu_map) {
6266 struct sched_domain *sd = &per_cpu(core_domains, i);
6268 init_sched_groups_power(i, sd);
6272 for_each_cpu_mask(i, *cpu_map) {
6273 struct sched_domain *sd = &per_cpu(phys_domains, i);
6275 init_sched_groups_power(i, sd);
6279 for (i = 0; i < MAX_NUMNODES; i++)
6280 init_numa_sched_groups_power(sched_group_nodes[i]);
6283 struct sched_group *sg;
6285 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6286 init_numa_sched_groups_power(sg);
6290 /* Attach the domains */
6291 for_each_cpu_mask(i, *cpu_map) {
6292 struct sched_domain *sd;
6293 #ifdef CONFIG_SCHED_SMT
6294 sd = &per_cpu(cpu_domains, i);
6295 #elif defined(CONFIG_SCHED_MC)
6296 sd = &per_cpu(core_domains, i);
6298 sd = &per_cpu(phys_domains, i);
6300 cpu_attach_domain(sd, i);
6307 free_sched_groups(cpu_map);
6312 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6314 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6316 cpumask_t cpu_default_map;
6320 * Setup mask for cpus without special case scheduling requirements.
6321 * For now this just excludes isolated cpus, but could be used to
6322 * exclude other special cases in the future.
6324 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6326 err = build_sched_domains(&cpu_default_map);
6328 register_sched_domain_sysctl();
6333 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6335 free_sched_groups(cpu_map);
6339 * Detach sched domains from a group of cpus specified in cpu_map
6340 * These cpus will now be attached to the NULL domain
6342 static void detach_destroy_domains(const cpumask_t *cpu_map)
6346 unregister_sched_domain_sysctl();
6348 for_each_cpu_mask(i, *cpu_map)
6349 cpu_attach_domain(NULL, i);
6350 synchronize_sched();
6351 arch_destroy_sched_domains(cpu_map);
6355 * Partition sched domains as specified by the cpumasks below.
6356 * This attaches all cpus from the cpumasks to the NULL domain,
6357 * waits for a RCU quiescent period, recalculates sched
6358 * domain information and then attaches them back to the
6359 * correct sched domains
6360 * Call with hotplug lock held
6362 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6364 cpumask_t change_map;
6367 cpus_and(*partition1, *partition1, cpu_online_map);
6368 cpus_and(*partition2, *partition2, cpu_online_map);
6369 cpus_or(change_map, *partition1, *partition2);
6371 /* Detach sched domains from all of the affected cpus */
6372 detach_destroy_domains(&change_map);
6373 if (!cpus_empty(*partition1))
6374 err = build_sched_domains(partition1);
6375 if (!err && !cpus_empty(*partition2))
6376 err = build_sched_domains(partition2);
6378 register_sched_domain_sysctl();
6383 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6384 static int arch_reinit_sched_domains(void)
6388 mutex_lock(&sched_hotcpu_mutex);
6389 detach_destroy_domains(&cpu_online_map);
6390 err = arch_init_sched_domains(&cpu_online_map);
6391 mutex_unlock(&sched_hotcpu_mutex);
6396 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6400 if (buf[0] != '0' && buf[0] != '1')
6404 sched_smt_power_savings = (buf[0] == '1');
6406 sched_mc_power_savings = (buf[0] == '1');
6408 ret = arch_reinit_sched_domains();
6410 return ret ? ret : count;
6413 #ifdef CONFIG_SCHED_MC
6414 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6416 return sprintf(page, "%u\n", sched_mc_power_savings);
6418 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6419 const char *buf, size_t count)
6421 return sched_power_savings_store(buf, count, 0);
6423 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6424 sched_mc_power_savings_store);
6427 #ifdef CONFIG_SCHED_SMT
6428 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6430 return sprintf(page, "%u\n", sched_smt_power_savings);
6432 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6433 const char *buf, size_t count)
6435 return sched_power_savings_store(buf, count, 1);
6437 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6438 sched_smt_power_savings_store);
6441 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6445 #ifdef CONFIG_SCHED_SMT
6447 err = sysfs_create_file(&cls->kset.kobj,
6448 &attr_sched_smt_power_savings.attr);
6450 #ifdef CONFIG_SCHED_MC
6451 if (!err && mc_capable())
6452 err = sysfs_create_file(&cls->kset.kobj,
6453 &attr_sched_mc_power_savings.attr);
6460 * Force a reinitialization of the sched domains hierarchy. The domains
6461 * and groups cannot be updated in place without racing with the balancing
6462 * code, so we temporarily attach all running cpus to the NULL domain
6463 * which will prevent rebalancing while the sched domains are recalculated.
6465 static int update_sched_domains(struct notifier_block *nfb,
6466 unsigned long action, void *hcpu)
6469 case CPU_UP_PREPARE:
6470 case CPU_UP_PREPARE_FROZEN:
6471 case CPU_DOWN_PREPARE:
6472 case CPU_DOWN_PREPARE_FROZEN:
6473 detach_destroy_domains(&cpu_online_map);
6476 case CPU_UP_CANCELED:
6477 case CPU_UP_CANCELED_FROZEN:
6478 case CPU_DOWN_FAILED:
6479 case CPU_DOWN_FAILED_FROZEN:
6481 case CPU_ONLINE_FROZEN:
6483 case CPU_DEAD_FROZEN:
6485 * Fall through and re-initialise the domains.
6492 /* The hotplug lock is already held by cpu_up/cpu_down */
6493 arch_init_sched_domains(&cpu_online_map);
6498 void __init sched_init_smp(void)
6500 cpumask_t non_isolated_cpus;
6502 mutex_lock(&sched_hotcpu_mutex);
6503 arch_init_sched_domains(&cpu_online_map);
6504 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6505 if (cpus_empty(non_isolated_cpus))
6506 cpu_set(smp_processor_id(), non_isolated_cpus);
6507 mutex_unlock(&sched_hotcpu_mutex);
6508 /* XXX: Theoretical race here - CPU may be hotplugged now */
6509 hotcpu_notifier(update_sched_domains, 0);
6511 /* Move init over to a non-isolated CPU */
6512 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6516 void __init sched_init_smp(void)
6519 #endif /* CONFIG_SMP */
6521 int in_sched_functions(unsigned long addr)
6523 /* Linker adds these: start and end of __sched functions */
6524 extern char __sched_text_start[], __sched_text_end[];
6526 return in_lock_functions(addr) ||
6527 (addr >= (unsigned long)__sched_text_start
6528 && addr < (unsigned long)__sched_text_end);
6531 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6533 cfs_rq->tasks_timeline = RB_ROOT;
6534 #ifdef CONFIG_FAIR_GROUP_SCHED
6537 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6540 void __init sched_init(void)
6542 int highest_cpu = 0;
6545 for_each_possible_cpu(i) {
6546 struct rt_prio_array *array;
6550 spin_lock_init(&rq->lock);
6551 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6554 init_cfs_rq(&rq->cfs, rq);
6555 #ifdef CONFIG_FAIR_GROUP_SCHED
6556 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6558 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6559 struct sched_entity *se =
6560 &per_cpu(init_sched_entity, i);
6562 init_cfs_rq_p[i] = cfs_rq;
6563 init_cfs_rq(cfs_rq, rq);
6564 cfs_rq->tg = &init_task_group;
6565 list_add(&cfs_rq->leaf_cfs_rq_list,
6566 &rq->leaf_cfs_rq_list);
6568 init_sched_entity_p[i] = se;
6569 se->cfs_rq = &rq->cfs;
6571 se->load.weight = init_task_group_load;
6572 se->load.inv_weight =
6573 div64_64(1ULL<<32, init_task_group_load);
6576 init_task_group.shares = init_task_group_load;
6577 spin_lock_init(&init_task_group.lock);
6580 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6581 rq->cpu_load[j] = 0;
6584 rq->active_balance = 0;
6585 rq->next_balance = jiffies;
6588 rq->migration_thread = NULL;
6589 INIT_LIST_HEAD(&rq->migration_queue);
6591 atomic_set(&rq->nr_iowait, 0);
6593 array = &rq->rt.active;
6594 for (j = 0; j < MAX_RT_PRIO; j++) {
6595 INIT_LIST_HEAD(array->queue + j);
6596 __clear_bit(j, array->bitmap);
6599 /* delimiter for bitsearch: */
6600 __set_bit(MAX_RT_PRIO, array->bitmap);
6603 set_load_weight(&init_task);
6605 #ifdef CONFIG_PREEMPT_NOTIFIERS
6606 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6610 nr_cpu_ids = highest_cpu + 1;
6611 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6614 #ifdef CONFIG_RT_MUTEXES
6615 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6619 * The boot idle thread does lazy MMU switching as well:
6621 atomic_inc(&init_mm.mm_count);
6622 enter_lazy_tlb(&init_mm, current);
6625 * Make us the idle thread. Technically, schedule() should not be
6626 * called from this thread, however somewhere below it might be,
6627 * but because we are the idle thread, we just pick up running again
6628 * when this runqueue becomes "idle".
6630 init_idle(current, smp_processor_id());
6632 * During early bootup we pretend to be a normal task:
6634 current->sched_class = &fair_sched_class;
6637 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6638 void __might_sleep(char *file, int line)
6641 static unsigned long prev_jiffy; /* ratelimiting */
6643 if ((in_atomic() || irqs_disabled()) &&
6644 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6645 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6647 prev_jiffy = jiffies;
6648 printk(KERN_ERR "BUG: sleeping function called from invalid"
6649 " context at %s:%d\n", file, line);
6650 printk("in_atomic():%d, irqs_disabled():%d\n",
6651 in_atomic(), irqs_disabled());
6652 debug_show_held_locks(current);
6653 if (irqs_disabled())
6654 print_irqtrace_events(current);
6659 EXPORT_SYMBOL(__might_sleep);
6662 #ifdef CONFIG_MAGIC_SYSRQ
6663 static void normalize_task(struct rq *rq, struct task_struct *p)
6666 update_rq_clock(rq);
6667 on_rq = p->se.on_rq;
6669 deactivate_task(rq, p, 0);
6670 __setscheduler(rq, p, SCHED_NORMAL, 0);
6672 activate_task(rq, p, 0);
6673 resched_task(rq->curr);
6677 void normalize_rt_tasks(void)
6679 struct task_struct *g, *p;
6680 unsigned long flags;
6683 read_lock_irq(&tasklist_lock);
6684 do_each_thread(g, p) {
6686 * Only normalize user tasks:
6691 p->se.exec_start = 0;
6692 #ifdef CONFIG_SCHEDSTATS
6693 p->se.wait_start = 0;
6694 p->se.sleep_start = 0;
6695 p->se.block_start = 0;
6697 task_rq(p)->clock = 0;
6701 * Renice negative nice level userspace
6704 if (TASK_NICE(p) < 0 && p->mm)
6705 set_user_nice(p, 0);
6709 spin_lock_irqsave(&p->pi_lock, flags);
6710 rq = __task_rq_lock(p);
6712 normalize_task(rq, p);
6714 __task_rq_unlock(rq);
6715 spin_unlock_irqrestore(&p->pi_lock, flags);
6716 } while_each_thread(g, p);
6718 read_unlock_irq(&tasklist_lock);
6721 #endif /* CONFIG_MAGIC_SYSRQ */
6725 * These functions are only useful for the IA64 MCA handling.
6727 * They can only be called when the whole system has been
6728 * stopped - every CPU needs to be quiescent, and no scheduling
6729 * activity can take place. Using them for anything else would
6730 * be a serious bug, and as a result, they aren't even visible
6731 * under any other configuration.
6735 * curr_task - return the current task for a given cpu.
6736 * @cpu: the processor in question.
6738 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6740 struct task_struct *curr_task(int cpu)
6742 return cpu_curr(cpu);
6746 * set_curr_task - set the current task for a given cpu.
6747 * @cpu: the processor in question.
6748 * @p: the task pointer to set.
6750 * Description: This function must only be used when non-maskable interrupts
6751 * are serviced on a separate stack. It allows the architecture to switch the
6752 * notion of the current task on a cpu in a non-blocking manner. This function
6753 * must be called with all CPU's synchronized, and interrupts disabled, the
6754 * and caller must save the original value of the current task (see
6755 * curr_task() above) and restore that value before reenabling interrupts and
6756 * re-starting the system.
6758 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6760 void set_curr_task(int cpu, struct task_struct *p)
6767 #ifdef CONFIG_FAIR_GROUP_SCHED
6769 /* allocate runqueue etc for a new task group */
6770 struct task_group *sched_create_group(void)
6772 struct task_group *tg;
6773 struct cfs_rq *cfs_rq;
6774 struct sched_entity *se;
6778 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6780 return ERR_PTR(-ENOMEM);
6782 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6785 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6789 for_each_possible_cpu(i) {
6792 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6797 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6802 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6803 memset(se, 0, sizeof(struct sched_entity));
6805 tg->cfs_rq[i] = cfs_rq;
6806 init_cfs_rq(cfs_rq, rq);
6810 se->cfs_rq = &rq->cfs;
6812 se->load.weight = NICE_0_LOAD;
6813 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
6817 for_each_possible_cpu(i) {
6819 cfs_rq = tg->cfs_rq[i];
6820 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6823 tg->shares = NICE_0_LOAD;
6824 spin_lock_init(&tg->lock);
6829 for_each_possible_cpu(i) {
6831 kfree(tg->cfs_rq[i]);
6839 return ERR_PTR(-ENOMEM);
6842 /* rcu callback to free various structures associated with a task group */
6843 static void free_sched_group(struct rcu_head *rhp)
6845 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
6846 struct task_group *tg = cfs_rq->tg;
6847 struct sched_entity *se;
6850 /* now it should be safe to free those cfs_rqs */
6851 for_each_possible_cpu(i) {
6852 cfs_rq = tg->cfs_rq[i];
6864 /* Destroy runqueue etc associated with a task group */
6865 void sched_destroy_group(struct task_group *tg)
6867 struct cfs_rq *cfs_rq;
6870 for_each_possible_cpu(i) {
6871 cfs_rq = tg->cfs_rq[i];
6872 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
6875 cfs_rq = tg->cfs_rq[0];
6877 /* wait for possible concurrent references to cfs_rqs complete */
6878 call_rcu(&cfs_rq->rcu, free_sched_group);
6881 /* change task's runqueue when it moves between groups.
6882 * The caller of this function should have put the task in its new group
6883 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6884 * reflect its new group.
6886 void sched_move_task(struct task_struct *tsk)
6889 unsigned long flags;
6892 rq = task_rq_lock(tsk, &flags);
6894 if (tsk->sched_class != &fair_sched_class)
6897 update_rq_clock(rq);
6899 running = task_running(rq, tsk);
6900 on_rq = tsk->se.on_rq;
6903 dequeue_task(rq, tsk, 0);
6904 if (unlikely(running))
6905 tsk->sched_class->put_prev_task(rq, tsk);
6908 set_task_cfs_rq(tsk);
6911 if (unlikely(running))
6912 tsk->sched_class->set_curr_task(rq);
6913 enqueue_task(rq, tsk, 0);
6917 task_rq_unlock(rq, &flags);
6920 static void set_se_shares(struct sched_entity *se, unsigned long shares)
6922 struct cfs_rq *cfs_rq = se->cfs_rq;
6923 struct rq *rq = cfs_rq->rq;
6926 spin_lock_irq(&rq->lock);
6930 dequeue_entity(cfs_rq, se, 0);
6932 se->load.weight = shares;
6933 se->load.inv_weight = div64_64((1ULL<<32), shares);
6936 enqueue_entity(cfs_rq, se, 0);
6938 spin_unlock_irq(&rq->lock);
6941 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6945 spin_lock(&tg->lock);
6946 if (tg->shares == shares)
6949 tg->shares = shares;
6950 for_each_possible_cpu(i)
6951 set_se_shares(tg->se[i], shares);
6954 spin_unlock(&tg->lock);
6958 unsigned long sched_group_shares(struct task_group *tg)
6963 #endif /* CONFIG_FAIR_GROUP_SCHED */