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>
68 * Scheduler clock - returns current time in nanosec units.
69 * This is default implementation.
70 * Architectures and sub-architectures can override this.
72 unsigned long long __attribute__((weak)) sched_clock(void)
74 return (unsigned long long)jiffies * (1000000000 / HZ);
78 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
82 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
91 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96 * Some helpers for converting nanosecond timing to jiffy resolution
98 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
99 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
108 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
109 * Timeslices get refilled after they expire.
111 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
112 #define DEF_TIMESLICE (100 * HZ / 1000)
116 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
117 * Since cpu_power is a 'constant', we can use a reciprocal divide.
119 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
121 return reciprocal_divide(load, sg->reciprocal_cpu_power);
125 * Each time a sched group cpu_power is changed,
126 * we must compute its reciprocal value
128 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
130 sg->__cpu_power += val;
131 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
135 #define SCALE_PRIO(x, prio) \
136 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
139 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
140 * to time slice values: [800ms ... 100ms ... 5ms]
142 static unsigned int static_prio_timeslice(int static_prio)
144 if (static_prio == NICE_TO_PRIO(19))
147 if (static_prio < NICE_TO_PRIO(0))
148 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
150 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
153 static inline int rt_policy(int policy)
155 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
160 static inline int task_has_rt_policy(struct task_struct *p)
162 return rt_policy(p->policy);
166 * This is the priority-queue data structure of the RT scheduling class:
168 struct rt_prio_array {
169 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
170 struct list_head queue[MAX_RT_PRIO];
174 struct load_weight load;
175 u64 load_update_start, load_update_last;
176 unsigned long delta_fair, delta_exec, delta_stat;
179 /* CFS-related fields in a runqueue */
181 struct load_weight load;
182 unsigned long nr_running;
188 unsigned long wait_runtime_overruns, wait_runtime_underruns;
190 struct rb_root tasks_timeline;
191 struct rb_node *rb_leftmost;
192 struct rb_node *rb_load_balance_curr;
193 #ifdef CONFIG_FAIR_GROUP_SCHED
194 /* 'curr' points to currently running entity on this cfs_rq.
195 * It is set to NULL otherwise (i.e when none are currently running).
197 struct sched_entity *curr;
198 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
207 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
211 /* Real-Time classes' related field in a runqueue: */
213 struct rt_prio_array active;
214 int rt_load_balance_idx;
215 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
219 * This is the main, per-CPU runqueue data structure.
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
226 spinlock_t lock; /* runqueue lock */
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
232 unsigned long nr_running;
233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
235 unsigned char idle_at_tick;
237 unsigned char in_nohz_recently;
239 struct load_stat ls; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
255 unsigned long nr_uninterruptible;
257 struct task_struct *curr, *idle;
258 unsigned long next_balance;
259 struct mm_struct *prev_mm;
261 u64 clock, prev_clock_raw;
264 unsigned int clock_warps, clock_overflows;
265 unsigned int clock_unstable_events;
270 struct sched_domain *sd;
272 /* For active balancing */
275 int cpu; /* cpu of this runqueue */
277 struct task_struct *migration_thread;
278 struct list_head migration_queue;
281 #ifdef CONFIG_SCHEDSTATS
283 struct sched_info rq_sched_info;
285 /* sys_sched_yield() stats */
286 unsigned long yld_exp_empty;
287 unsigned long yld_act_empty;
288 unsigned long yld_both_empty;
289 unsigned long yld_cnt;
291 /* schedule() stats */
292 unsigned long sched_switch;
293 unsigned long sched_cnt;
294 unsigned long sched_goidle;
296 /* try_to_wake_up() stats */
297 unsigned long ttwu_cnt;
298 unsigned long ttwu_local;
300 struct lock_class_key rq_lock_key;
303 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
304 static DEFINE_MUTEX(sched_hotcpu_mutex);
306 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
308 rq->curr->sched_class->check_preempt_curr(rq, p);
311 static inline int cpu_of(struct rq *rq)
321 * Update the per-runqueue clock, as finegrained as the platform can give
322 * us, but without assuming monotonicity, etc.:
324 static void __update_rq_clock(struct rq *rq)
326 u64 prev_raw = rq->prev_clock_raw;
327 u64 now = sched_clock();
328 s64 delta = now - prev_raw;
329 u64 clock = rq->clock;
331 #ifdef CONFIG_SCHED_DEBUG
332 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
335 * Protect against sched_clock() occasionally going backwards:
337 if (unlikely(delta < 0)) {
342 * Catch too large forward jumps too:
344 if (unlikely(delta > 2*TICK_NSEC)) {
346 rq->clock_overflows++;
348 if (unlikely(delta > rq->clock_max_delta))
349 rq->clock_max_delta = delta;
354 rq->prev_clock_raw = now;
358 static void update_rq_clock(struct rq *rq)
360 if (likely(smp_processor_id() == cpu_of(rq)))
361 __update_rq_clock(rq);
364 static u64 __rq_clock(struct rq *rq)
366 __update_rq_clock(rq);
371 static u64 rq_clock(struct rq *rq)
378 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
379 * See detach_destroy_domains: synchronize_sched for details.
381 * The domain tree of any CPU may only be accessed from within
382 * preempt-disabled sections.
384 #define for_each_domain(cpu, __sd) \
385 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
387 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
388 #define this_rq() (&__get_cpu_var(runqueues))
389 #define task_rq(p) cpu_rq(task_cpu(p))
390 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
393 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
394 * clock constructed from sched_clock():
396 unsigned long long cpu_clock(int cpu)
398 unsigned long long now;
402 local_irq_save(flags);
406 local_irq_restore(flags);
411 #ifdef CONFIG_FAIR_GROUP_SCHED
412 /* Change a task's ->cfs_rq if it moves across CPUs */
413 static inline void set_task_cfs_rq(struct task_struct *p)
415 p->se.cfs_rq = &task_rq(p)->cfs;
418 static inline void set_task_cfs_rq(struct task_struct *p)
423 #ifndef prepare_arch_switch
424 # define prepare_arch_switch(next) do { } while (0)
426 #ifndef finish_arch_switch
427 # define finish_arch_switch(prev) do { } while (0)
430 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
431 static inline int task_running(struct rq *rq, struct task_struct *p)
433 return rq->curr == p;
436 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
440 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
442 #ifdef CONFIG_DEBUG_SPINLOCK
443 /* this is a valid case when another task releases the spinlock */
444 rq->lock.owner = current;
447 * If we are tracking spinlock dependencies then we have to
448 * fix up the runqueue lock - which gets 'carried over' from
451 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
453 spin_unlock_irq(&rq->lock);
456 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
457 static inline int task_running(struct rq *rq, struct task_struct *p)
462 return rq->curr == p;
466 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
470 * We can optimise this out completely for !SMP, because the
471 * SMP rebalancing from interrupt is the only thing that cares
476 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
477 spin_unlock_irq(&rq->lock);
479 spin_unlock(&rq->lock);
483 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
487 * After ->oncpu is cleared, the task can be moved to a different CPU.
488 * We must ensure this doesn't happen until the switch is completely
494 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
498 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
501 * __task_rq_lock - lock the runqueue a given task resides on.
502 * Must be called interrupts disabled.
504 static inline struct rq *__task_rq_lock(struct task_struct *p)
511 spin_lock(&rq->lock);
512 if (unlikely(rq != task_rq(p))) {
513 spin_unlock(&rq->lock);
514 goto repeat_lock_task;
520 * task_rq_lock - lock the runqueue a given task resides on and disable
521 * interrupts. Note the ordering: we can safely lookup the task_rq without
522 * explicitly disabling preemption.
524 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
530 local_irq_save(*flags);
532 spin_lock(&rq->lock);
533 if (unlikely(rq != task_rq(p))) {
534 spin_unlock_irqrestore(&rq->lock, *flags);
535 goto repeat_lock_task;
540 static inline void __task_rq_unlock(struct rq *rq)
543 spin_unlock(&rq->lock);
546 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
549 spin_unlock_irqrestore(&rq->lock, *flags);
553 * this_rq_lock - lock this runqueue and disable interrupts.
555 static inline struct rq *this_rq_lock(void)
562 spin_lock(&rq->lock);
568 * CPU frequency is/was unstable - start new by setting prev_clock_raw:
570 void sched_clock_unstable_event(void)
575 rq = task_rq_lock(current, &flags);
576 rq->prev_clock_raw = sched_clock();
577 rq->clock_unstable_events++;
578 task_rq_unlock(rq, &flags);
582 * resched_task - mark a task 'to be rescheduled now'.
584 * On UP this means the setting of the need_resched flag, on SMP it
585 * might also involve a cross-CPU call to trigger the scheduler on
590 #ifndef tsk_is_polling
591 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
594 static void resched_task(struct task_struct *p)
598 assert_spin_locked(&task_rq(p)->lock);
600 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
603 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
606 if (cpu == smp_processor_id())
609 /* NEED_RESCHED must be visible before we test polling */
611 if (!tsk_is_polling(p))
612 smp_send_reschedule(cpu);
615 static void resched_cpu(int cpu)
617 struct rq *rq = cpu_rq(cpu);
620 if (!spin_trylock_irqsave(&rq->lock, flags))
622 resched_task(cpu_curr(cpu));
623 spin_unlock_irqrestore(&rq->lock, flags);
626 static inline void resched_task(struct task_struct *p)
628 assert_spin_locked(&task_rq(p)->lock);
629 set_tsk_need_resched(p);
633 static u64 div64_likely32(u64 divident, unsigned long divisor)
635 #if BITS_PER_LONG == 32
636 if (likely(divident <= 0xffffffffULL))
637 return (u32)divident / divisor;
638 do_div(divident, divisor);
642 return divident / divisor;
646 #if BITS_PER_LONG == 32
647 # define WMULT_CONST (~0UL)
649 # define WMULT_CONST (1UL << 32)
652 #define WMULT_SHIFT 32
655 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
656 struct load_weight *lw)
660 if (unlikely(!lw->inv_weight))
661 lw->inv_weight = WMULT_CONST / lw->weight;
663 tmp = (u64)delta_exec * weight;
665 * Check whether we'd overflow the 64-bit multiplication:
667 if (unlikely(tmp > WMULT_CONST)) {
668 tmp = ((tmp >> WMULT_SHIFT/2) * lw->inv_weight)
671 tmp = (tmp * lw->inv_weight) >> WMULT_SHIFT;
674 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
677 static inline unsigned long
678 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
680 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
683 static void update_load_add(struct load_weight *lw, unsigned long inc)
689 static void update_load_sub(struct load_weight *lw, unsigned long dec)
696 * To aid in avoiding the subversion of "niceness" due to uneven distribution
697 * of tasks with abnormal "nice" values across CPUs the contribution that
698 * each task makes to its run queue's load is weighted according to its
699 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
700 * scaled version of the new time slice allocation that they receive on time
704 #define WEIGHT_IDLEPRIO 2
705 #define WMULT_IDLEPRIO (1 << 31)
708 * Nice levels are multiplicative, with a gentle 10% change for every
709 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
710 * nice 1, it will get ~10% less CPU time than another CPU-bound task
711 * that remained on nice 0.
713 * The "10% effect" is relative and cumulative: from _any_ nice level,
714 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
715 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
716 * If a task goes up by ~10% and another task goes down by ~10% then
717 * the relative distance between them is ~25%.)
719 static const int prio_to_weight[40] = {
720 /* -20 */ 88818, 71054, 56843, 45475, 36380, 29104, 23283, 18626, 14901, 11921,
721 /* -10 */ 9537, 7629, 6103, 4883, 3906, 3125, 2500, 2000, 1600, 1280,
722 /* 0 */ NICE_0_LOAD /* 1024 */,
723 /* 1 */ 819, 655, 524, 419, 336, 268, 215, 172, 137,
724 /* 10 */ 110, 87, 70, 56, 45, 36, 29, 23, 18, 15,
728 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
730 * In cases where the weight does not change often, we can use the
731 * precalculated inverse to speed up arithmetics by turning divisions
732 * into multiplications:
734 static const u32 prio_to_wmult[40] = {
735 /* -20 */ 48356, 60446, 75558, 94446, 118058,
736 /* -15 */ 147573, 184467, 230589, 288233, 360285,
737 /* -10 */ 450347, 562979, 703746, 879575, 1099582,
738 /* -5 */ 1374389, 1717986, 2147483, 2684354, 3355443,
739 /* 0 */ 4194304, 5244160, 6557201, 8196502, 10250518,
740 /* 5 */ 12782640, 16025997, 19976592, 24970740, 31350126,
741 /* 10 */ 39045157, 49367440, 61356675, 76695844, 95443717,
742 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
745 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
748 * runqueue iterator, to support SMP load-balancing between different
749 * scheduling classes, without having to expose their internal data
750 * structures to the load-balancing proper:
754 struct task_struct *(*start)(void *);
755 struct task_struct *(*next)(void *);
758 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
759 unsigned long max_nr_move, unsigned long max_load_move,
760 struct sched_domain *sd, enum cpu_idle_type idle,
761 int *all_pinned, unsigned long *load_moved,
762 int *this_best_prio, struct rq_iterator *iterator);
764 #include "sched_stats.h"
765 #include "sched_rt.c"
766 #include "sched_fair.c"
767 #include "sched_idletask.c"
768 #ifdef CONFIG_SCHED_DEBUG
769 # include "sched_debug.c"
772 #define sched_class_highest (&rt_sched_class)
774 static void __update_curr_load(struct rq *rq, struct load_stat *ls)
776 if (rq->curr != rq->idle && ls->load.weight) {
777 ls->delta_exec += ls->delta_stat;
778 ls->delta_fair += calc_delta_fair(ls->delta_stat, &ls->load);
784 * Update delta_exec, delta_fair fields for rq.
786 * delta_fair clock advances at a rate inversely proportional to
787 * total load (rq->ls.load.weight) on the runqueue, while
788 * delta_exec advances at the same rate as wall-clock (provided
791 * delta_exec / delta_fair is a measure of the (smoothened) load on this
792 * runqueue over any given interval. This (smoothened) load is used
793 * during load balance.
795 * This function is called /before/ updating rq->ls.load
796 * and when switching tasks.
798 static void update_curr_load(struct rq *rq, u64 now)
800 struct load_stat *ls = &rq->ls;
803 start = ls->load_update_start;
804 ls->load_update_start = now;
805 ls->delta_stat += now - start;
807 * Stagger updates to ls->delta_fair. Very frequent updates
810 if (ls->delta_stat >= sysctl_sched_stat_granularity)
811 __update_curr_load(rq, ls);
815 inc_load(struct rq *rq, const struct task_struct *p, u64 now)
817 update_curr_load(rq, now);
818 update_load_add(&rq->ls.load, p->se.load.weight);
822 dec_load(struct rq *rq, const struct task_struct *p, u64 now)
824 update_curr_load(rq, now);
825 update_load_sub(&rq->ls.load, p->se.load.weight);
828 static void inc_nr_running(struct task_struct *p, struct rq *rq, u64 now)
831 inc_load(rq, p, now);
834 static void dec_nr_running(struct task_struct *p, struct rq *rq, u64 now)
837 dec_load(rq, p, now);
840 static void set_load_weight(struct task_struct *p)
842 task_rq(p)->cfs.wait_runtime -= p->se.wait_runtime;
843 p->se.wait_runtime = 0;
845 if (task_has_rt_policy(p)) {
846 p->se.load.weight = prio_to_weight[0] * 2;
847 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
852 * SCHED_IDLE tasks get minimal weight:
854 if (p->policy == SCHED_IDLE) {
855 p->se.load.weight = WEIGHT_IDLEPRIO;
856 p->se.load.inv_weight = WMULT_IDLEPRIO;
860 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
861 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
865 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, u64 now)
867 sched_info_queued(p);
868 p->sched_class->enqueue_task(rq, p, wakeup, now);
873 dequeue_task(struct rq *rq, struct task_struct *p, int sleep, u64 now)
875 p->sched_class->dequeue_task(rq, p, sleep, now);
880 * __normal_prio - return the priority that is based on the static prio
882 static inline int __normal_prio(struct task_struct *p)
884 return p->static_prio;
888 * Calculate the expected normal priority: i.e. priority
889 * without taking RT-inheritance into account. Might be
890 * boosted by interactivity modifiers. Changes upon fork,
891 * setprio syscalls, and whenever the interactivity
892 * estimator recalculates.
894 static inline int normal_prio(struct task_struct *p)
898 if (task_has_rt_policy(p))
899 prio = MAX_RT_PRIO-1 - p->rt_priority;
901 prio = __normal_prio(p);
906 * Calculate the current priority, i.e. the priority
907 * taken into account by the scheduler. This value might
908 * be boosted by RT tasks, or might be boosted by
909 * interactivity modifiers. Will be RT if the task got
910 * RT-boosted. If not then it returns p->normal_prio.
912 static int effective_prio(struct task_struct *p)
914 p->normal_prio = normal_prio(p);
916 * If we are RT tasks or we were boosted to RT priority,
917 * keep the priority unchanged. Otherwise, update priority
918 * to the normal priority:
920 if (!rt_prio(p->prio))
921 return p->normal_prio;
926 * activate_task - move a task to the runqueue.
928 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
930 u64 now = rq_clock(rq);
932 if (p->state == TASK_UNINTERRUPTIBLE)
933 rq->nr_uninterruptible--;
935 enqueue_task(rq, p, wakeup, now);
936 inc_nr_running(p, rq, now);
940 * activate_idle_task - move idle task to the _front_ of runqueue.
942 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
944 u64 now = rq_clock(rq);
946 if (p->state == TASK_UNINTERRUPTIBLE)
947 rq->nr_uninterruptible--;
949 enqueue_task(rq, p, 0, now);
950 inc_nr_running(p, rq, now);
954 * deactivate_task - remove a task from the runqueue.
957 deactivate_task(struct rq *rq, struct task_struct *p, int sleep, u64 now)
959 if (p->state == TASK_UNINTERRUPTIBLE)
960 rq->nr_uninterruptible++;
962 dequeue_task(rq, p, sleep, now);
963 dec_nr_running(p, rq, now);
967 * task_curr - is this task currently executing on a CPU?
968 * @p: the task in question.
970 inline int task_curr(const struct task_struct *p)
972 return cpu_curr(task_cpu(p)) == p;
975 /* Used instead of source_load when we know the type == 0 */
976 unsigned long weighted_cpuload(const int cpu)
978 return cpu_rq(cpu)->ls.load.weight;
981 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
984 task_thread_info(p)->cpu = cpu;
991 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
993 int old_cpu = task_cpu(p);
994 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
995 u64 clock_offset, fair_clock_offset;
997 clock_offset = old_rq->clock - new_rq->clock;
998 fair_clock_offset = old_rq->cfs.fair_clock - new_rq->cfs.fair_clock;
1000 if (p->se.wait_start_fair)
1001 p->se.wait_start_fair -= fair_clock_offset;
1002 if (p->se.sleep_start_fair)
1003 p->se.sleep_start_fair -= fair_clock_offset;
1005 #ifdef CONFIG_SCHEDSTATS
1006 if (p->se.wait_start)
1007 p->se.wait_start -= clock_offset;
1008 if (p->se.sleep_start)
1009 p->se.sleep_start -= clock_offset;
1010 if (p->se.block_start)
1011 p->se.block_start -= clock_offset;
1014 __set_task_cpu(p, new_cpu);
1017 struct migration_req {
1018 struct list_head list;
1020 struct task_struct *task;
1023 struct completion done;
1027 * The task's runqueue lock must be held.
1028 * Returns true if you have to wait for migration thread.
1031 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1033 struct rq *rq = task_rq(p);
1036 * If the task is not on a runqueue (and not running), then
1037 * it is sufficient to simply update the task's cpu field.
1039 if (!p->se.on_rq && !task_running(rq, p)) {
1040 set_task_cpu(p, dest_cpu);
1044 init_completion(&req->done);
1046 req->dest_cpu = dest_cpu;
1047 list_add(&req->list, &rq->migration_queue);
1053 * wait_task_inactive - wait for a thread to unschedule.
1055 * The caller must ensure that the task *will* unschedule sometime soon,
1056 * else this function might spin for a *long* time. This function can't
1057 * be called with interrupts off, or it may introduce deadlock with
1058 * smp_call_function() if an IPI is sent by the same process we are
1059 * waiting to become inactive.
1061 void wait_task_inactive(struct task_struct *p)
1063 unsigned long flags;
1069 * We do the initial early heuristics without holding
1070 * any task-queue locks at all. We'll only try to get
1071 * the runqueue lock when things look like they will
1077 * If the task is actively running on another CPU
1078 * still, just relax and busy-wait without holding
1081 * NOTE! Since we don't hold any locks, it's not
1082 * even sure that "rq" stays as the right runqueue!
1083 * But we don't care, since "task_running()" will
1084 * return false if the runqueue has changed and p
1085 * is actually now running somewhere else!
1087 while (task_running(rq, p))
1091 * Ok, time to look more closely! We need the rq
1092 * lock now, to be *sure*. If we're wrong, we'll
1093 * just go back and repeat.
1095 rq = task_rq_lock(p, &flags);
1096 running = task_running(rq, p);
1097 on_rq = p->se.on_rq;
1098 task_rq_unlock(rq, &flags);
1101 * Was it really running after all now that we
1102 * checked with the proper locks actually held?
1104 * Oops. Go back and try again..
1106 if (unlikely(running)) {
1112 * It's not enough that it's not actively running,
1113 * it must be off the runqueue _entirely_, and not
1116 * So if it wa still runnable (but just not actively
1117 * running right now), it's preempted, and we should
1118 * yield - it could be a while.
1120 if (unlikely(on_rq)) {
1126 * Ahh, all good. It wasn't running, and it wasn't
1127 * runnable, which means that it will never become
1128 * running in the future either. We're all done!
1133 * kick_process - kick a running thread to enter/exit the kernel
1134 * @p: the to-be-kicked thread
1136 * Cause a process which is running on another CPU to enter
1137 * kernel-mode, without any delay. (to get signals handled.)
1139 * NOTE: this function doesnt have to take the runqueue lock,
1140 * because all it wants to ensure is that the remote task enters
1141 * the kernel. If the IPI races and the task has been migrated
1142 * to another CPU then no harm is done and the purpose has been
1145 void kick_process(struct task_struct *p)
1151 if ((cpu != smp_processor_id()) && task_curr(p))
1152 smp_send_reschedule(cpu);
1157 * Return a low guess at the load of a migration-source cpu weighted
1158 * according to the scheduling class and "nice" value.
1160 * We want to under-estimate the load of migration sources, to
1161 * balance conservatively.
1163 static inline unsigned long source_load(int cpu, int type)
1165 struct rq *rq = cpu_rq(cpu);
1166 unsigned long total = weighted_cpuload(cpu);
1171 return min(rq->cpu_load[type-1], total);
1175 * Return a high guess at the load of a migration-target cpu weighted
1176 * according to the scheduling class and "nice" value.
1178 static inline unsigned long target_load(int cpu, int type)
1180 struct rq *rq = cpu_rq(cpu);
1181 unsigned long total = weighted_cpuload(cpu);
1186 return max(rq->cpu_load[type-1], total);
1190 * Return the average load per task on the cpu's run queue
1192 static inline unsigned long cpu_avg_load_per_task(int cpu)
1194 struct rq *rq = cpu_rq(cpu);
1195 unsigned long total = weighted_cpuload(cpu);
1196 unsigned long n = rq->nr_running;
1198 return n ? total / n : SCHED_LOAD_SCALE;
1202 * find_idlest_group finds and returns the least busy CPU group within the
1205 static struct sched_group *
1206 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1208 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1209 unsigned long min_load = ULONG_MAX, this_load = 0;
1210 int load_idx = sd->forkexec_idx;
1211 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1214 unsigned long load, avg_load;
1218 /* Skip over this group if it has no CPUs allowed */
1219 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1222 local_group = cpu_isset(this_cpu, group->cpumask);
1224 /* Tally up the load of all CPUs in the group */
1227 for_each_cpu_mask(i, group->cpumask) {
1228 /* Bias balancing toward cpus of our domain */
1230 load = source_load(i, load_idx);
1232 load = target_load(i, load_idx);
1237 /* Adjust by relative CPU power of the group */
1238 avg_load = sg_div_cpu_power(group,
1239 avg_load * SCHED_LOAD_SCALE);
1242 this_load = avg_load;
1244 } else if (avg_load < min_load) {
1245 min_load = avg_load;
1249 group = group->next;
1250 } while (group != sd->groups);
1252 if (!idlest || 100*this_load < imbalance*min_load)
1258 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1261 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1264 unsigned long load, min_load = ULONG_MAX;
1268 /* Traverse only the allowed CPUs */
1269 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1271 for_each_cpu_mask(i, tmp) {
1272 load = weighted_cpuload(i);
1274 if (load < min_load || (load == min_load && i == this_cpu)) {
1284 * sched_balance_self: balance the current task (running on cpu) in domains
1285 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1288 * Balance, ie. select the least loaded group.
1290 * Returns the target CPU number, or the same CPU if no balancing is needed.
1292 * preempt must be disabled.
1294 static int sched_balance_self(int cpu, int flag)
1296 struct task_struct *t = current;
1297 struct sched_domain *tmp, *sd = NULL;
1299 for_each_domain(cpu, tmp) {
1301 * If power savings logic is enabled for a domain, stop there.
1303 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1305 if (tmp->flags & flag)
1311 struct sched_group *group;
1312 int new_cpu, weight;
1314 if (!(sd->flags & flag)) {
1320 group = find_idlest_group(sd, t, cpu);
1326 new_cpu = find_idlest_cpu(group, t, cpu);
1327 if (new_cpu == -1 || new_cpu == cpu) {
1328 /* Now try balancing at a lower domain level of cpu */
1333 /* Now try balancing at a lower domain level of new_cpu */
1336 weight = cpus_weight(span);
1337 for_each_domain(cpu, tmp) {
1338 if (weight <= cpus_weight(tmp->span))
1340 if (tmp->flags & flag)
1343 /* while loop will break here if sd == NULL */
1349 #endif /* CONFIG_SMP */
1352 * wake_idle() will wake a task on an idle cpu if task->cpu is
1353 * not idle and an idle cpu is available. The span of cpus to
1354 * search starts with cpus closest then further out as needed,
1355 * so we always favor a closer, idle cpu.
1357 * Returns the CPU we should wake onto.
1359 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1360 static int wake_idle(int cpu, struct task_struct *p)
1363 struct sched_domain *sd;
1367 * If it is idle, then it is the best cpu to run this task.
1369 * This cpu is also the best, if it has more than one task already.
1370 * Siblings must be also busy(in most cases) as they didn't already
1371 * pickup the extra load from this cpu and hence we need not check
1372 * sibling runqueue info. This will avoid the checks and cache miss
1373 * penalities associated with that.
1375 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1378 for_each_domain(cpu, sd) {
1379 if (sd->flags & SD_WAKE_IDLE) {
1380 cpus_and(tmp, sd->span, p->cpus_allowed);
1381 for_each_cpu_mask(i, tmp) {
1392 static inline int wake_idle(int cpu, struct task_struct *p)
1399 * try_to_wake_up - wake up a thread
1400 * @p: the to-be-woken-up thread
1401 * @state: the mask of task states that can be woken
1402 * @sync: do a synchronous wakeup?
1404 * Put it on the run-queue if it's not already there. The "current"
1405 * thread is always on the run-queue (except when the actual
1406 * re-schedule is in progress), and as such you're allowed to do
1407 * the simpler "current->state = TASK_RUNNING" to mark yourself
1408 * runnable without the overhead of this.
1410 * returns failure only if the task is already active.
1412 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1414 int cpu, this_cpu, success = 0;
1415 unsigned long flags;
1419 struct sched_domain *sd, *this_sd = NULL;
1420 unsigned long load, this_load;
1424 rq = task_rq_lock(p, &flags);
1425 old_state = p->state;
1426 if (!(old_state & state))
1433 this_cpu = smp_processor_id();
1436 if (unlikely(task_running(rq, p)))
1441 schedstat_inc(rq, ttwu_cnt);
1442 if (cpu == this_cpu) {
1443 schedstat_inc(rq, ttwu_local);
1447 for_each_domain(this_cpu, sd) {
1448 if (cpu_isset(cpu, sd->span)) {
1449 schedstat_inc(sd, ttwu_wake_remote);
1455 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1459 * Check for affine wakeup and passive balancing possibilities.
1462 int idx = this_sd->wake_idx;
1463 unsigned int imbalance;
1465 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1467 load = source_load(cpu, idx);
1468 this_load = target_load(this_cpu, idx);
1470 new_cpu = this_cpu; /* Wake to this CPU if we can */
1472 if (this_sd->flags & SD_WAKE_AFFINE) {
1473 unsigned long tl = this_load;
1474 unsigned long tl_per_task;
1476 tl_per_task = cpu_avg_load_per_task(this_cpu);
1479 * If sync wakeup then subtract the (maximum possible)
1480 * effect of the currently running task from the load
1481 * of the current CPU:
1484 tl -= current->se.load.weight;
1487 tl + target_load(cpu, idx) <= tl_per_task) ||
1488 100*(tl + p->se.load.weight) <= imbalance*load) {
1490 * This domain has SD_WAKE_AFFINE and
1491 * p is cache cold in this domain, and
1492 * there is no bad imbalance.
1494 schedstat_inc(this_sd, ttwu_move_affine);
1500 * Start passive balancing when half the imbalance_pct
1503 if (this_sd->flags & SD_WAKE_BALANCE) {
1504 if (imbalance*this_load <= 100*load) {
1505 schedstat_inc(this_sd, ttwu_move_balance);
1511 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1513 new_cpu = wake_idle(new_cpu, p);
1514 if (new_cpu != cpu) {
1515 set_task_cpu(p, new_cpu);
1516 task_rq_unlock(rq, &flags);
1517 /* might preempt at this point */
1518 rq = task_rq_lock(p, &flags);
1519 old_state = p->state;
1520 if (!(old_state & state))
1525 this_cpu = smp_processor_id();
1530 #endif /* CONFIG_SMP */
1531 activate_task(rq, p, 1);
1533 * Sync wakeups (i.e. those types of wakeups where the waker
1534 * has indicated that it will leave the CPU in short order)
1535 * don't trigger a preemption, if the woken up task will run on
1536 * this cpu. (in this case the 'I will reschedule' promise of
1537 * the waker guarantees that the freshly woken up task is going
1538 * to be considered on this CPU.)
1540 if (!sync || cpu != this_cpu)
1541 check_preempt_curr(rq, p);
1545 p->state = TASK_RUNNING;
1547 task_rq_unlock(rq, &flags);
1552 int fastcall wake_up_process(struct task_struct *p)
1554 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1555 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1557 EXPORT_SYMBOL(wake_up_process);
1559 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1561 return try_to_wake_up(p, state, 0);
1565 * Perform scheduler related setup for a newly forked process p.
1566 * p is forked by current.
1568 * __sched_fork() is basic setup used by init_idle() too:
1570 static void __sched_fork(struct task_struct *p)
1572 p->se.wait_start_fair = 0;
1573 p->se.exec_start = 0;
1574 p->se.sum_exec_runtime = 0;
1575 p->se.delta_exec = 0;
1576 p->se.delta_fair_run = 0;
1577 p->se.delta_fair_sleep = 0;
1578 p->se.wait_runtime = 0;
1579 p->se.sleep_start_fair = 0;
1581 #ifdef CONFIG_SCHEDSTATS
1582 p->se.wait_start = 0;
1583 p->se.sum_wait_runtime = 0;
1584 p->se.sum_sleep_runtime = 0;
1585 p->se.sleep_start = 0;
1586 p->se.block_start = 0;
1587 p->se.sleep_max = 0;
1588 p->se.block_max = 0;
1591 p->se.wait_runtime_overruns = 0;
1592 p->se.wait_runtime_underruns = 0;
1595 INIT_LIST_HEAD(&p->run_list);
1598 #ifdef CONFIG_PREEMPT_NOTIFIERS
1599 INIT_HLIST_HEAD(&p->preempt_notifiers);
1603 * We mark the process as running here, but have not actually
1604 * inserted it onto the runqueue yet. This guarantees that
1605 * nobody will actually run it, and a signal or other external
1606 * event cannot wake it up and insert it on the runqueue either.
1608 p->state = TASK_RUNNING;
1612 * fork()/clone()-time setup:
1614 void sched_fork(struct task_struct *p, int clone_flags)
1616 int cpu = get_cpu();
1621 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1623 __set_task_cpu(p, cpu);
1626 * Make sure we do not leak PI boosting priority to the child:
1628 p->prio = current->normal_prio;
1630 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1631 if (likely(sched_info_on()))
1632 memset(&p->sched_info, 0, sizeof(p->sched_info));
1634 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1637 #ifdef CONFIG_PREEMPT
1638 /* Want to start with kernel preemption disabled. */
1639 task_thread_info(p)->preempt_count = 1;
1645 * After fork, child runs first. (default) If set to 0 then
1646 * parent will (try to) run first.
1648 unsigned int __read_mostly sysctl_sched_child_runs_first = 1;
1651 * wake_up_new_task - wake up a newly created task for the first time.
1653 * This function will do some initial scheduler statistics housekeeping
1654 * that must be done for every newly created context, then puts the task
1655 * on the runqueue and wakes it.
1657 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1659 unsigned long flags;
1664 rq = task_rq_lock(p, &flags);
1665 BUG_ON(p->state != TASK_RUNNING);
1666 this_cpu = smp_processor_id(); /* parent's CPU */
1669 p->prio = effective_prio(p);
1671 if (!p->sched_class->task_new || !sysctl_sched_child_runs_first ||
1672 (clone_flags & CLONE_VM) || task_cpu(p) != this_cpu ||
1673 !current->se.on_rq) {
1675 activate_task(rq, p, 0);
1678 * Let the scheduling class do new task startup
1679 * management (if any):
1681 p->sched_class->task_new(rq, p, now);
1682 inc_nr_running(p, rq, now);
1684 check_preempt_curr(rq, p);
1685 task_rq_unlock(rq, &flags);
1688 #ifdef CONFIG_PREEMPT_NOTIFIERS
1691 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1692 * @notifier: notifier struct to register
1694 void preempt_notifier_register(struct preempt_notifier *notifier)
1696 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1698 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1701 * preempt_notifier_unregister - no longer interested in preemption notifications
1702 * @notifier: notifier struct to unregister
1704 * This is safe to call from within a preemption notifier.
1706 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1708 hlist_del(¬ifier->link);
1710 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1712 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1714 struct preempt_notifier *notifier;
1715 struct hlist_node *node;
1717 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1718 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1722 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1723 struct task_struct *next)
1725 struct preempt_notifier *notifier;
1726 struct hlist_node *node;
1728 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1729 notifier->ops->sched_out(notifier, next);
1734 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1739 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1740 struct task_struct *next)
1747 * prepare_task_switch - prepare to switch tasks
1748 * @rq: the runqueue preparing to switch
1749 * @prev: the current task that is being switched out
1750 * @next: the task we are going to switch to.
1752 * This is called with the rq lock held and interrupts off. It must
1753 * be paired with a subsequent finish_task_switch after the context
1756 * prepare_task_switch sets up locking and calls architecture specific
1760 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1761 struct task_struct *next)
1763 fire_sched_out_preempt_notifiers(prev, next);
1764 prepare_lock_switch(rq, next);
1765 prepare_arch_switch(next);
1769 * finish_task_switch - clean up after a task-switch
1770 * @rq: runqueue associated with task-switch
1771 * @prev: the thread we just switched away from.
1773 * finish_task_switch must be called after the context switch, paired
1774 * with a prepare_task_switch call before the context switch.
1775 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1776 * and do any other architecture-specific cleanup actions.
1778 * Note that we may have delayed dropping an mm in context_switch(). If
1779 * so, we finish that here outside of the runqueue lock. (Doing it
1780 * with the lock held can cause deadlocks; see schedule() for
1783 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1784 __releases(rq->lock)
1786 struct mm_struct *mm = rq->prev_mm;
1792 * A task struct has one reference for the use as "current".
1793 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1794 * schedule one last time. The schedule call will never return, and
1795 * the scheduled task must drop that reference.
1796 * The test for TASK_DEAD must occur while the runqueue locks are
1797 * still held, otherwise prev could be scheduled on another cpu, die
1798 * there before we look at prev->state, and then the reference would
1800 * Manfred Spraul <manfred@colorfullife.com>
1802 prev_state = prev->state;
1803 finish_arch_switch(prev);
1804 finish_lock_switch(rq, prev);
1805 fire_sched_in_preempt_notifiers(current);
1808 if (unlikely(prev_state == TASK_DEAD)) {
1810 * Remove function-return probe instances associated with this
1811 * task and put them back on the free list.
1813 kprobe_flush_task(prev);
1814 put_task_struct(prev);
1819 * schedule_tail - first thing a freshly forked thread must call.
1820 * @prev: the thread we just switched away from.
1822 asmlinkage void schedule_tail(struct task_struct *prev)
1823 __releases(rq->lock)
1825 struct rq *rq = this_rq();
1827 finish_task_switch(rq, prev);
1828 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1829 /* In this case, finish_task_switch does not reenable preemption */
1832 if (current->set_child_tid)
1833 put_user(current->pid, current->set_child_tid);
1837 * context_switch - switch to the new MM and the new
1838 * thread's register state.
1841 context_switch(struct rq *rq, struct task_struct *prev,
1842 struct task_struct *next)
1844 struct mm_struct *mm, *oldmm;
1846 prepare_task_switch(rq, prev, next);
1848 oldmm = prev->active_mm;
1850 * For paravirt, this is coupled with an exit in switch_to to
1851 * combine the page table reload and the switch backend into
1854 arch_enter_lazy_cpu_mode();
1856 if (unlikely(!mm)) {
1857 next->active_mm = oldmm;
1858 atomic_inc(&oldmm->mm_count);
1859 enter_lazy_tlb(oldmm, next);
1861 switch_mm(oldmm, mm, next);
1863 if (unlikely(!prev->mm)) {
1864 prev->active_mm = NULL;
1865 rq->prev_mm = oldmm;
1868 * Since the runqueue lock will be released by the next
1869 * task (which is an invalid locking op but in the case
1870 * of the scheduler it's an obvious special-case), so we
1871 * do an early lockdep release here:
1873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1874 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1877 /* Here we just switch the register state and the stack. */
1878 switch_to(prev, next, prev);
1882 * this_rq must be evaluated again because prev may have moved
1883 * CPUs since it called schedule(), thus the 'rq' on its stack
1884 * frame will be invalid.
1886 finish_task_switch(this_rq(), prev);
1890 * nr_running, nr_uninterruptible and nr_context_switches:
1892 * externally visible scheduler statistics: current number of runnable
1893 * threads, current number of uninterruptible-sleeping threads, total
1894 * number of context switches performed since bootup.
1896 unsigned long nr_running(void)
1898 unsigned long i, sum = 0;
1900 for_each_online_cpu(i)
1901 sum += cpu_rq(i)->nr_running;
1906 unsigned long nr_uninterruptible(void)
1908 unsigned long i, sum = 0;
1910 for_each_possible_cpu(i)
1911 sum += cpu_rq(i)->nr_uninterruptible;
1914 * Since we read the counters lockless, it might be slightly
1915 * inaccurate. Do not allow it to go below zero though:
1917 if (unlikely((long)sum < 0))
1923 unsigned long long nr_context_switches(void)
1926 unsigned long long sum = 0;
1928 for_each_possible_cpu(i)
1929 sum += cpu_rq(i)->nr_switches;
1934 unsigned long nr_iowait(void)
1936 unsigned long i, sum = 0;
1938 for_each_possible_cpu(i)
1939 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1944 unsigned long nr_active(void)
1946 unsigned long i, running = 0, uninterruptible = 0;
1948 for_each_online_cpu(i) {
1949 running += cpu_rq(i)->nr_running;
1950 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1953 if (unlikely((long)uninterruptible < 0))
1954 uninterruptible = 0;
1956 return running + uninterruptible;
1960 * Update rq->cpu_load[] statistics. This function is usually called every
1961 * scheduler tick (TICK_NSEC).
1963 static void update_cpu_load(struct rq *this_rq)
1965 u64 fair_delta64, exec_delta64, idle_delta64, sample_interval64, tmp64;
1966 unsigned long total_load = this_rq->ls.load.weight;
1967 unsigned long this_load = total_load;
1968 struct load_stat *ls = &this_rq->ls;
1969 u64 now = __rq_clock(this_rq);
1972 this_rq->nr_load_updates++;
1973 if (unlikely(!(sysctl_sched_features & SCHED_FEAT_PRECISE_CPU_LOAD)))
1976 /* Update delta_fair/delta_exec fields first */
1977 update_curr_load(this_rq, now);
1979 fair_delta64 = ls->delta_fair + 1;
1982 exec_delta64 = ls->delta_exec + 1;
1985 sample_interval64 = now - ls->load_update_last;
1986 ls->load_update_last = now;
1988 if ((s64)sample_interval64 < (s64)TICK_NSEC)
1989 sample_interval64 = TICK_NSEC;
1991 if (exec_delta64 > sample_interval64)
1992 exec_delta64 = sample_interval64;
1994 idle_delta64 = sample_interval64 - exec_delta64;
1996 tmp64 = div64_64(SCHED_LOAD_SCALE * exec_delta64, fair_delta64);
1997 tmp64 = div64_64(tmp64 * exec_delta64, sample_interval64);
1999 this_load = (unsigned long)tmp64;
2003 /* Update our load: */
2004 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2005 unsigned long old_load, new_load;
2007 /* scale is effectively 1 << i now, and >> i divides by scale */
2009 old_load = this_rq->cpu_load[i];
2010 new_load = this_load;
2012 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2019 * double_rq_lock - safely lock two runqueues
2021 * Note this does not disable interrupts like task_rq_lock,
2022 * you need to do so manually before calling.
2024 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2025 __acquires(rq1->lock)
2026 __acquires(rq2->lock)
2028 BUG_ON(!irqs_disabled());
2030 spin_lock(&rq1->lock);
2031 __acquire(rq2->lock); /* Fake it out ;) */
2034 spin_lock(&rq1->lock);
2035 spin_lock(&rq2->lock);
2037 spin_lock(&rq2->lock);
2038 spin_lock(&rq1->lock);
2044 * double_rq_unlock - safely unlock two runqueues
2046 * Note this does not restore interrupts like task_rq_unlock,
2047 * you need to do so manually after calling.
2049 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2050 __releases(rq1->lock)
2051 __releases(rq2->lock)
2053 spin_unlock(&rq1->lock);
2055 spin_unlock(&rq2->lock);
2057 __release(rq2->lock);
2061 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2063 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2064 __releases(this_rq->lock)
2065 __acquires(busiest->lock)
2066 __acquires(this_rq->lock)
2068 if (unlikely(!irqs_disabled())) {
2069 /* printk() doesn't work good under rq->lock */
2070 spin_unlock(&this_rq->lock);
2073 if (unlikely(!spin_trylock(&busiest->lock))) {
2074 if (busiest < this_rq) {
2075 spin_unlock(&this_rq->lock);
2076 spin_lock(&busiest->lock);
2077 spin_lock(&this_rq->lock);
2079 spin_lock(&busiest->lock);
2084 * If dest_cpu is allowed for this process, migrate the task to it.
2085 * This is accomplished by forcing the cpu_allowed mask to only
2086 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2087 * the cpu_allowed mask is restored.
2089 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2091 struct migration_req req;
2092 unsigned long flags;
2095 rq = task_rq_lock(p, &flags);
2096 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2097 || unlikely(cpu_is_offline(dest_cpu)))
2100 /* force the process onto the specified CPU */
2101 if (migrate_task(p, dest_cpu, &req)) {
2102 /* Need to wait for migration thread (might exit: take ref). */
2103 struct task_struct *mt = rq->migration_thread;
2105 get_task_struct(mt);
2106 task_rq_unlock(rq, &flags);
2107 wake_up_process(mt);
2108 put_task_struct(mt);
2109 wait_for_completion(&req.done);
2114 task_rq_unlock(rq, &flags);
2118 * sched_exec - execve() is a valuable balancing opportunity, because at
2119 * this point the task has the smallest effective memory and cache footprint.
2121 void sched_exec(void)
2123 int new_cpu, this_cpu = get_cpu();
2124 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2126 if (new_cpu != this_cpu)
2127 sched_migrate_task(current, new_cpu);
2131 * pull_task - move a task from a remote runqueue to the local runqueue.
2132 * Both runqueues must be locked.
2134 static void pull_task(struct rq *src_rq, struct task_struct *p,
2135 struct rq *this_rq, int this_cpu)
2137 deactivate_task(src_rq, p, 0, rq_clock(src_rq));
2138 set_task_cpu(p, this_cpu);
2139 activate_task(this_rq, p, 0);
2141 * Note that idle threads have a prio of MAX_PRIO, for this test
2142 * to be always true for them.
2144 check_preempt_curr(this_rq, p);
2148 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2151 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2152 struct sched_domain *sd, enum cpu_idle_type idle,
2156 * We do not migrate tasks that are:
2157 * 1) running (obviously), or
2158 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2159 * 3) are cache-hot on their current CPU.
2161 if (!cpu_isset(this_cpu, p->cpus_allowed))
2165 if (task_running(rq, p))
2169 * Aggressive migration if too many balance attempts have failed:
2171 if (sd->nr_balance_failed > sd->cache_nice_tries)
2177 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2178 unsigned long max_nr_move, unsigned long max_load_move,
2179 struct sched_domain *sd, enum cpu_idle_type idle,
2180 int *all_pinned, unsigned long *load_moved,
2181 int *this_best_prio, struct rq_iterator *iterator)
2183 int pulled = 0, pinned = 0, skip_for_load;
2184 struct task_struct *p;
2185 long rem_load_move = max_load_move;
2187 if (max_nr_move == 0 || max_load_move == 0)
2193 * Start the load-balancing iterator:
2195 p = iterator->start(iterator->arg);
2200 * To help distribute high priority tasks accross CPUs we don't
2201 * skip a task if it will be the highest priority task (i.e. smallest
2202 * prio value) on its new queue regardless of its load weight
2204 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2205 SCHED_LOAD_SCALE_FUZZ;
2206 if ((skip_for_load && p->prio >= *this_best_prio) ||
2207 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2208 p = iterator->next(iterator->arg);
2212 pull_task(busiest, p, this_rq, this_cpu);
2214 rem_load_move -= p->se.load.weight;
2217 * We only want to steal up to the prescribed number of tasks
2218 * and the prescribed amount of weighted load.
2220 if (pulled < max_nr_move && rem_load_move > 0) {
2221 if (p->prio < *this_best_prio)
2222 *this_best_prio = p->prio;
2223 p = iterator->next(iterator->arg);
2228 * Right now, this is the only place pull_task() is called,
2229 * so we can safely collect pull_task() stats here rather than
2230 * inside pull_task().
2232 schedstat_add(sd, lb_gained[idle], pulled);
2235 *all_pinned = pinned;
2236 *load_moved = max_load_move - rem_load_move;
2241 * move_tasks tries to move up to max_load_move weighted load from busiest to
2242 * this_rq, as part of a balancing operation within domain "sd".
2243 * Returns 1 if successful and 0 otherwise.
2245 * Called with both runqueues locked.
2247 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2248 unsigned long max_load_move,
2249 struct sched_domain *sd, enum cpu_idle_type idle,
2252 struct sched_class *class = sched_class_highest;
2253 unsigned long total_load_moved = 0;
2254 int this_best_prio = this_rq->curr->prio;
2258 class->load_balance(this_rq, this_cpu, busiest,
2259 ULONG_MAX, max_load_move - total_load_moved,
2260 sd, idle, all_pinned, &this_best_prio);
2261 class = class->next;
2262 } while (class && max_load_move > total_load_moved);
2264 return total_load_moved > 0;
2268 * move_one_task tries to move exactly one task from busiest to this_rq, as
2269 * part of active balancing operations within "domain".
2270 * Returns 1 if successful and 0 otherwise.
2272 * Called with both runqueues locked.
2274 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2275 struct sched_domain *sd, enum cpu_idle_type idle)
2277 struct sched_class *class;
2278 int this_best_prio = MAX_PRIO;
2280 for (class = sched_class_highest; class; class = class->next)
2281 if (class->load_balance(this_rq, this_cpu, busiest,
2282 1, ULONG_MAX, sd, idle, NULL,
2290 * find_busiest_group finds and returns the busiest CPU group within the
2291 * domain. It calculates and returns the amount of weighted load which
2292 * should be moved to restore balance via the imbalance parameter.
2294 static struct sched_group *
2295 find_busiest_group(struct sched_domain *sd, int this_cpu,
2296 unsigned long *imbalance, enum cpu_idle_type idle,
2297 int *sd_idle, cpumask_t *cpus, int *balance)
2299 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2300 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2301 unsigned long max_pull;
2302 unsigned long busiest_load_per_task, busiest_nr_running;
2303 unsigned long this_load_per_task, this_nr_running;
2305 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2306 int power_savings_balance = 1;
2307 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2308 unsigned long min_nr_running = ULONG_MAX;
2309 struct sched_group *group_min = NULL, *group_leader = NULL;
2312 max_load = this_load = total_load = total_pwr = 0;
2313 busiest_load_per_task = busiest_nr_running = 0;
2314 this_load_per_task = this_nr_running = 0;
2315 if (idle == CPU_NOT_IDLE)
2316 load_idx = sd->busy_idx;
2317 else if (idle == CPU_NEWLY_IDLE)
2318 load_idx = sd->newidle_idx;
2320 load_idx = sd->idle_idx;
2323 unsigned long load, group_capacity;
2326 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2327 unsigned long sum_nr_running, sum_weighted_load;
2329 local_group = cpu_isset(this_cpu, group->cpumask);
2332 balance_cpu = first_cpu(group->cpumask);
2334 /* Tally up the load of all CPUs in the group */
2335 sum_weighted_load = sum_nr_running = avg_load = 0;
2337 for_each_cpu_mask(i, group->cpumask) {
2340 if (!cpu_isset(i, *cpus))
2345 if (*sd_idle && rq->nr_running)
2348 /* Bias balancing toward cpus of our domain */
2350 if (idle_cpu(i) && !first_idle_cpu) {
2355 load = target_load(i, load_idx);
2357 load = source_load(i, load_idx);
2360 sum_nr_running += rq->nr_running;
2361 sum_weighted_load += weighted_cpuload(i);
2365 * First idle cpu or the first cpu(busiest) in this sched group
2366 * is eligible for doing load balancing at this and above
2367 * domains. In the newly idle case, we will allow all the cpu's
2368 * to do the newly idle load balance.
2370 if (idle != CPU_NEWLY_IDLE && local_group &&
2371 balance_cpu != this_cpu && balance) {
2376 total_load += avg_load;
2377 total_pwr += group->__cpu_power;
2379 /* Adjust by relative CPU power of the group */
2380 avg_load = sg_div_cpu_power(group,
2381 avg_load * SCHED_LOAD_SCALE);
2383 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2386 this_load = avg_load;
2388 this_nr_running = sum_nr_running;
2389 this_load_per_task = sum_weighted_load;
2390 } else if (avg_load > max_load &&
2391 sum_nr_running > group_capacity) {
2392 max_load = avg_load;
2394 busiest_nr_running = sum_nr_running;
2395 busiest_load_per_task = sum_weighted_load;
2398 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2400 * Busy processors will not participate in power savings
2403 if (idle == CPU_NOT_IDLE ||
2404 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2408 * If the local group is idle or completely loaded
2409 * no need to do power savings balance at this domain
2411 if (local_group && (this_nr_running >= group_capacity ||
2413 power_savings_balance = 0;
2416 * If a group is already running at full capacity or idle,
2417 * don't include that group in power savings calculations
2419 if (!power_savings_balance || sum_nr_running >= group_capacity
2424 * Calculate the group which has the least non-idle load.
2425 * This is the group from where we need to pick up the load
2428 if ((sum_nr_running < min_nr_running) ||
2429 (sum_nr_running == min_nr_running &&
2430 first_cpu(group->cpumask) <
2431 first_cpu(group_min->cpumask))) {
2433 min_nr_running = sum_nr_running;
2434 min_load_per_task = sum_weighted_load /
2439 * Calculate the group which is almost near its
2440 * capacity but still has some space to pick up some load
2441 * from other group and save more power
2443 if (sum_nr_running <= group_capacity - 1) {
2444 if (sum_nr_running > leader_nr_running ||
2445 (sum_nr_running == leader_nr_running &&
2446 first_cpu(group->cpumask) >
2447 first_cpu(group_leader->cpumask))) {
2448 group_leader = group;
2449 leader_nr_running = sum_nr_running;
2454 group = group->next;
2455 } while (group != sd->groups);
2457 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2460 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2462 if (this_load >= avg_load ||
2463 100*max_load <= sd->imbalance_pct*this_load)
2466 busiest_load_per_task /= busiest_nr_running;
2468 * We're trying to get all the cpus to the average_load, so we don't
2469 * want to push ourselves above the average load, nor do we wish to
2470 * reduce the max loaded cpu below the average load, as either of these
2471 * actions would just result in more rebalancing later, and ping-pong
2472 * tasks around. Thus we look for the minimum possible imbalance.
2473 * Negative imbalances (*we* are more loaded than anyone else) will
2474 * be counted as no imbalance for these purposes -- we can't fix that
2475 * by pulling tasks to us. Be careful of negative numbers as they'll
2476 * appear as very large values with unsigned longs.
2478 if (max_load <= busiest_load_per_task)
2482 * In the presence of smp nice balancing, certain scenarios can have
2483 * max load less than avg load(as we skip the groups at or below
2484 * its cpu_power, while calculating max_load..)
2486 if (max_load < avg_load) {
2488 goto small_imbalance;
2491 /* Don't want to pull so many tasks that a group would go idle */
2492 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2494 /* How much load to actually move to equalise the imbalance */
2495 *imbalance = min(max_pull * busiest->__cpu_power,
2496 (avg_load - this_load) * this->__cpu_power)
2500 * if *imbalance is less than the average load per runnable task
2501 * there is no gaurantee that any tasks will be moved so we'll have
2502 * a think about bumping its value to force at least one task to be
2505 if (*imbalance + SCHED_LOAD_SCALE_FUZZ < busiest_load_per_task/2) {
2506 unsigned long tmp, pwr_now, pwr_move;
2510 pwr_move = pwr_now = 0;
2512 if (this_nr_running) {
2513 this_load_per_task /= this_nr_running;
2514 if (busiest_load_per_task > this_load_per_task)
2517 this_load_per_task = SCHED_LOAD_SCALE;
2519 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2520 busiest_load_per_task * imbn) {
2521 *imbalance = busiest_load_per_task;
2526 * OK, we don't have enough imbalance to justify moving tasks,
2527 * however we may be able to increase total CPU power used by
2531 pwr_now += busiest->__cpu_power *
2532 min(busiest_load_per_task, max_load);
2533 pwr_now += this->__cpu_power *
2534 min(this_load_per_task, this_load);
2535 pwr_now /= SCHED_LOAD_SCALE;
2537 /* Amount of load we'd subtract */
2538 tmp = sg_div_cpu_power(busiest,
2539 busiest_load_per_task * SCHED_LOAD_SCALE);
2541 pwr_move += busiest->__cpu_power *
2542 min(busiest_load_per_task, max_load - tmp);
2544 /* Amount of load we'd add */
2545 if (max_load * busiest->__cpu_power <
2546 busiest_load_per_task * SCHED_LOAD_SCALE)
2547 tmp = sg_div_cpu_power(this,
2548 max_load * busiest->__cpu_power);
2550 tmp = sg_div_cpu_power(this,
2551 busiest_load_per_task * SCHED_LOAD_SCALE);
2552 pwr_move += this->__cpu_power *
2553 min(this_load_per_task, this_load + tmp);
2554 pwr_move /= SCHED_LOAD_SCALE;
2556 /* Move if we gain throughput */
2557 if (pwr_move <= pwr_now)
2560 *imbalance = busiest_load_per_task;
2566 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2567 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2570 if (this == group_leader && group_leader != group_min) {
2571 *imbalance = min_load_per_task;
2581 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2584 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2585 unsigned long imbalance, cpumask_t *cpus)
2587 struct rq *busiest = NULL, *rq;
2588 unsigned long max_load = 0;
2591 for_each_cpu_mask(i, group->cpumask) {
2594 if (!cpu_isset(i, *cpus))
2598 wl = weighted_cpuload(i);
2600 if (rq->nr_running == 1 && wl > imbalance)
2603 if (wl > max_load) {
2613 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2614 * so long as it is large enough.
2616 #define MAX_PINNED_INTERVAL 512
2619 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2620 * tasks if there is an imbalance.
2622 static int load_balance(int this_cpu, struct rq *this_rq,
2623 struct sched_domain *sd, enum cpu_idle_type idle,
2626 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2627 struct sched_group *group;
2628 unsigned long imbalance;
2630 cpumask_t cpus = CPU_MASK_ALL;
2631 unsigned long flags;
2634 * When power savings policy is enabled for the parent domain, idle
2635 * sibling can pick up load irrespective of busy siblings. In this case,
2636 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2637 * portraying it as CPU_NOT_IDLE.
2639 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2640 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2643 schedstat_inc(sd, lb_cnt[idle]);
2646 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2653 schedstat_inc(sd, lb_nobusyg[idle]);
2657 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2659 schedstat_inc(sd, lb_nobusyq[idle]);
2663 BUG_ON(busiest == this_rq);
2665 schedstat_add(sd, lb_imbalance[idle], imbalance);
2668 if (busiest->nr_running > 1) {
2670 * Attempt to move tasks. If find_busiest_group has found
2671 * an imbalance but busiest->nr_running <= 1, the group is
2672 * still unbalanced. ld_moved simply stays zero, so it is
2673 * correctly treated as an imbalance.
2675 local_irq_save(flags);
2676 double_rq_lock(this_rq, busiest);
2677 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2678 imbalance, sd, idle, &all_pinned);
2679 double_rq_unlock(this_rq, busiest);
2680 local_irq_restore(flags);
2683 * some other cpu did the load balance for us.
2685 if (ld_moved && this_cpu != smp_processor_id())
2686 resched_cpu(this_cpu);
2688 /* All tasks on this runqueue were pinned by CPU affinity */
2689 if (unlikely(all_pinned)) {
2690 cpu_clear(cpu_of(busiest), cpus);
2691 if (!cpus_empty(cpus))
2698 schedstat_inc(sd, lb_failed[idle]);
2699 sd->nr_balance_failed++;
2701 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2703 spin_lock_irqsave(&busiest->lock, flags);
2705 /* don't kick the migration_thread, if the curr
2706 * task on busiest cpu can't be moved to this_cpu
2708 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2709 spin_unlock_irqrestore(&busiest->lock, flags);
2711 goto out_one_pinned;
2714 if (!busiest->active_balance) {
2715 busiest->active_balance = 1;
2716 busiest->push_cpu = this_cpu;
2719 spin_unlock_irqrestore(&busiest->lock, flags);
2721 wake_up_process(busiest->migration_thread);
2724 * We've kicked active balancing, reset the failure
2727 sd->nr_balance_failed = sd->cache_nice_tries+1;
2730 sd->nr_balance_failed = 0;
2732 if (likely(!active_balance)) {
2733 /* We were unbalanced, so reset the balancing interval */
2734 sd->balance_interval = sd->min_interval;
2737 * If we've begun active balancing, start to back off. This
2738 * case may not be covered by the all_pinned logic if there
2739 * is only 1 task on the busy runqueue (because we don't call
2742 if (sd->balance_interval < sd->max_interval)
2743 sd->balance_interval *= 2;
2746 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2747 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2752 schedstat_inc(sd, lb_balanced[idle]);
2754 sd->nr_balance_failed = 0;
2757 /* tune up the balancing interval */
2758 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2759 (sd->balance_interval < sd->max_interval))
2760 sd->balance_interval *= 2;
2762 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2763 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2769 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2770 * tasks if there is an imbalance.
2772 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2773 * this_rq is locked.
2776 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2778 struct sched_group *group;
2779 struct rq *busiest = NULL;
2780 unsigned long imbalance;
2784 cpumask_t cpus = CPU_MASK_ALL;
2787 * When power savings policy is enabled for the parent domain, idle
2788 * sibling can pick up load irrespective of busy siblings. In this case,
2789 * let the state of idle sibling percolate up as IDLE, instead of
2790 * portraying it as CPU_NOT_IDLE.
2792 if (sd->flags & SD_SHARE_CPUPOWER &&
2793 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2796 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2798 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2799 &sd_idle, &cpus, NULL);
2801 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2805 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2808 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2812 BUG_ON(busiest == this_rq);
2814 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2817 if (busiest->nr_running > 1) {
2818 /* Attempt to move tasks */
2819 double_lock_balance(this_rq, busiest);
2820 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2821 imbalance, sd, CPU_NEWLY_IDLE,
2823 spin_unlock(&busiest->lock);
2825 if (unlikely(all_pinned)) {
2826 cpu_clear(cpu_of(busiest), cpus);
2827 if (!cpus_empty(cpus))
2833 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2834 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2835 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2838 sd->nr_balance_failed = 0;
2843 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2844 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2845 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2847 sd->nr_balance_failed = 0;
2853 * idle_balance is called by schedule() if this_cpu is about to become
2854 * idle. Attempts to pull tasks from other CPUs.
2856 static void idle_balance(int this_cpu, struct rq *this_rq)
2858 struct sched_domain *sd;
2859 int pulled_task = -1;
2860 unsigned long next_balance = jiffies + HZ;
2862 for_each_domain(this_cpu, sd) {
2863 unsigned long interval;
2865 if (!(sd->flags & SD_LOAD_BALANCE))
2868 if (sd->flags & SD_BALANCE_NEWIDLE)
2869 /* If we've pulled tasks over stop searching: */
2870 pulled_task = load_balance_newidle(this_cpu,
2873 interval = msecs_to_jiffies(sd->balance_interval);
2874 if (time_after(next_balance, sd->last_balance + interval))
2875 next_balance = sd->last_balance + interval;
2879 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2881 * We are going idle. next_balance may be set based on
2882 * a busy processor. So reset next_balance.
2884 this_rq->next_balance = next_balance;
2889 * active_load_balance is run by migration threads. It pushes running tasks
2890 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2891 * running on each physical CPU where possible, and avoids physical /
2892 * logical imbalances.
2894 * Called with busiest_rq locked.
2896 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2898 int target_cpu = busiest_rq->push_cpu;
2899 struct sched_domain *sd;
2900 struct rq *target_rq;
2902 /* Is there any task to move? */
2903 if (busiest_rq->nr_running <= 1)
2906 target_rq = cpu_rq(target_cpu);
2909 * This condition is "impossible", if it occurs
2910 * we need to fix it. Originally reported by
2911 * Bjorn Helgaas on a 128-cpu setup.
2913 BUG_ON(busiest_rq == target_rq);
2915 /* move a task from busiest_rq to target_rq */
2916 double_lock_balance(busiest_rq, target_rq);
2918 /* Search for an sd spanning us and the target CPU. */
2919 for_each_domain(target_cpu, sd) {
2920 if ((sd->flags & SD_LOAD_BALANCE) &&
2921 cpu_isset(busiest_cpu, sd->span))
2926 schedstat_inc(sd, alb_cnt);
2928 if (move_one_task(target_rq, target_cpu, busiest_rq,
2930 schedstat_inc(sd, alb_pushed);
2932 schedstat_inc(sd, alb_failed);
2934 spin_unlock(&target_rq->lock);
2939 atomic_t load_balancer;
2941 } nohz ____cacheline_aligned = {
2942 .load_balancer = ATOMIC_INIT(-1),
2943 .cpu_mask = CPU_MASK_NONE,
2947 * This routine will try to nominate the ilb (idle load balancing)
2948 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2949 * load balancing on behalf of all those cpus. If all the cpus in the system
2950 * go into this tickless mode, then there will be no ilb owner (as there is
2951 * no need for one) and all the cpus will sleep till the next wakeup event
2954 * For the ilb owner, tick is not stopped. And this tick will be used
2955 * for idle load balancing. ilb owner will still be part of
2958 * While stopping the tick, this cpu will become the ilb owner if there
2959 * is no other owner. And will be the owner till that cpu becomes busy
2960 * or if all cpus in the system stop their ticks at which point
2961 * there is no need for ilb owner.
2963 * When the ilb owner becomes busy, it nominates another owner, during the
2964 * next busy scheduler_tick()
2966 int select_nohz_load_balancer(int stop_tick)
2968 int cpu = smp_processor_id();
2971 cpu_set(cpu, nohz.cpu_mask);
2972 cpu_rq(cpu)->in_nohz_recently = 1;
2975 * If we are going offline and still the leader, give up!
2977 if (cpu_is_offline(cpu) &&
2978 atomic_read(&nohz.load_balancer) == cpu) {
2979 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2984 /* time for ilb owner also to sleep */
2985 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2986 if (atomic_read(&nohz.load_balancer) == cpu)
2987 atomic_set(&nohz.load_balancer, -1);
2991 if (atomic_read(&nohz.load_balancer) == -1) {
2992 /* make me the ilb owner */
2993 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2995 } else if (atomic_read(&nohz.load_balancer) == cpu)
2998 if (!cpu_isset(cpu, nohz.cpu_mask))
3001 cpu_clear(cpu, nohz.cpu_mask);
3003 if (atomic_read(&nohz.load_balancer) == cpu)
3004 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3011 static DEFINE_SPINLOCK(balancing);
3014 * It checks each scheduling domain to see if it is due to be balanced,
3015 * and initiates a balancing operation if so.
3017 * Balancing parameters are set up in arch_init_sched_domains.
3019 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3022 struct rq *rq = cpu_rq(cpu);
3023 unsigned long interval;
3024 struct sched_domain *sd;
3025 /* Earliest time when we have to do rebalance again */
3026 unsigned long next_balance = jiffies + 60*HZ;
3028 for_each_domain(cpu, sd) {
3029 if (!(sd->flags & SD_LOAD_BALANCE))
3032 interval = sd->balance_interval;
3033 if (idle != CPU_IDLE)
3034 interval *= sd->busy_factor;
3036 /* scale ms to jiffies */
3037 interval = msecs_to_jiffies(interval);
3038 if (unlikely(!interval))
3040 if (interval > HZ*NR_CPUS/10)
3041 interval = HZ*NR_CPUS/10;
3044 if (sd->flags & SD_SERIALIZE) {
3045 if (!spin_trylock(&balancing))
3049 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3050 if (load_balance(cpu, rq, sd, idle, &balance)) {
3052 * We've pulled tasks over so either we're no
3053 * longer idle, or one of our SMT siblings is
3056 idle = CPU_NOT_IDLE;
3058 sd->last_balance = jiffies;
3060 if (sd->flags & SD_SERIALIZE)
3061 spin_unlock(&balancing);
3063 if (time_after(next_balance, sd->last_balance + interval))
3064 next_balance = sd->last_balance + interval;
3067 * Stop the load balance at this level. There is another
3068 * CPU in our sched group which is doing load balancing more
3074 rq->next_balance = next_balance;
3078 * run_rebalance_domains is triggered when needed from the scheduler tick.
3079 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3080 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3082 static void run_rebalance_domains(struct softirq_action *h)
3084 int this_cpu = smp_processor_id();
3085 struct rq *this_rq = cpu_rq(this_cpu);
3086 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3087 CPU_IDLE : CPU_NOT_IDLE;
3089 rebalance_domains(this_cpu, idle);
3093 * If this cpu is the owner for idle load balancing, then do the
3094 * balancing on behalf of the other idle cpus whose ticks are
3097 if (this_rq->idle_at_tick &&
3098 atomic_read(&nohz.load_balancer) == this_cpu) {
3099 cpumask_t cpus = nohz.cpu_mask;
3103 cpu_clear(this_cpu, cpus);
3104 for_each_cpu_mask(balance_cpu, cpus) {
3106 * If this cpu gets work to do, stop the load balancing
3107 * work being done for other cpus. Next load
3108 * balancing owner will pick it up.
3113 rebalance_domains(balance_cpu, SCHED_IDLE);
3115 rq = cpu_rq(balance_cpu);
3116 if (time_after(this_rq->next_balance, rq->next_balance))
3117 this_rq->next_balance = rq->next_balance;
3124 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3126 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3127 * idle load balancing owner or decide to stop the periodic load balancing,
3128 * if the whole system is idle.
3130 static inline void trigger_load_balance(struct rq *rq, int cpu)
3134 * If we were in the nohz mode recently and busy at the current
3135 * scheduler tick, then check if we need to nominate new idle
3138 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3139 rq->in_nohz_recently = 0;
3141 if (atomic_read(&nohz.load_balancer) == cpu) {
3142 cpu_clear(cpu, nohz.cpu_mask);
3143 atomic_set(&nohz.load_balancer, -1);
3146 if (atomic_read(&nohz.load_balancer) == -1) {
3148 * simple selection for now: Nominate the
3149 * first cpu in the nohz list to be the next
3152 * TBD: Traverse the sched domains and nominate
3153 * the nearest cpu in the nohz.cpu_mask.
3155 int ilb = first_cpu(nohz.cpu_mask);
3163 * If this cpu is idle and doing idle load balancing for all the
3164 * cpus with ticks stopped, is it time for that to stop?
3166 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3167 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3173 * If this cpu is idle and the idle load balancing is done by
3174 * someone else, then no need raise the SCHED_SOFTIRQ
3176 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3177 cpu_isset(cpu, nohz.cpu_mask))
3180 if (time_after_eq(jiffies, rq->next_balance))
3181 raise_softirq(SCHED_SOFTIRQ);
3184 #else /* CONFIG_SMP */
3187 * on UP we do not need to balance between CPUs:
3189 static inline void idle_balance(int cpu, struct rq *rq)
3193 /* Avoid "used but not defined" warning on UP */
3194 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3195 unsigned long max_nr_move, unsigned long max_load_move,
3196 struct sched_domain *sd, enum cpu_idle_type idle,
3197 int *all_pinned, unsigned long *load_moved,
3198 int *this_best_prio, struct rq_iterator *iterator)
3207 DEFINE_PER_CPU(struct kernel_stat, kstat);
3209 EXPORT_PER_CPU_SYMBOL(kstat);
3212 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3213 * that have not yet been banked in case the task is currently running.
3215 unsigned long long task_sched_runtime(struct task_struct *p)
3217 unsigned long flags;
3221 rq = task_rq_lock(p, &flags);
3222 ns = p->se.sum_exec_runtime;
3223 if (rq->curr == p) {
3224 delta_exec = rq_clock(rq) - p->se.exec_start;
3225 if ((s64)delta_exec > 0)
3228 task_rq_unlock(rq, &flags);
3234 * Account user cpu time to a process.
3235 * @p: the process that the cpu time gets accounted to
3236 * @hardirq_offset: the offset to subtract from hardirq_count()
3237 * @cputime: the cpu time spent in user space since the last update
3239 void account_user_time(struct task_struct *p, cputime_t cputime)
3241 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3244 p->utime = cputime_add(p->utime, cputime);
3246 /* Add user time to cpustat. */
3247 tmp = cputime_to_cputime64(cputime);
3248 if (TASK_NICE(p) > 0)
3249 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3251 cpustat->user = cputime64_add(cpustat->user, tmp);
3255 * Account system cpu time to a process.
3256 * @p: the process that the cpu time gets accounted to
3257 * @hardirq_offset: the offset to subtract from hardirq_count()
3258 * @cputime: the cpu time spent in kernel space since the last update
3260 void account_system_time(struct task_struct *p, int hardirq_offset,
3263 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3264 struct rq *rq = this_rq();
3267 p->stime = cputime_add(p->stime, cputime);
3269 /* Add system time to cpustat. */
3270 tmp = cputime_to_cputime64(cputime);
3271 if (hardirq_count() - hardirq_offset)
3272 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3273 else if (softirq_count())
3274 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3275 else if (p != rq->idle)
3276 cpustat->system = cputime64_add(cpustat->system, tmp);
3277 else if (atomic_read(&rq->nr_iowait) > 0)
3278 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3280 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3281 /* Account for system time used */
3282 acct_update_integrals(p);
3286 * Account for involuntary wait time.
3287 * @p: the process from which the cpu time has been stolen
3288 * @steal: the cpu time spent in involuntary wait
3290 void account_steal_time(struct task_struct *p, cputime_t steal)
3292 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3293 cputime64_t tmp = cputime_to_cputime64(steal);
3294 struct rq *rq = this_rq();
3296 if (p == rq->idle) {
3297 p->stime = cputime_add(p->stime, steal);
3298 if (atomic_read(&rq->nr_iowait) > 0)
3299 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3301 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3303 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3307 * This function gets called by the timer code, with HZ frequency.
3308 * We call it with interrupts disabled.
3310 * It also gets called by the fork code, when changing the parent's
3313 void scheduler_tick(void)
3315 int cpu = smp_processor_id();
3316 struct rq *rq = cpu_rq(cpu);
3317 struct task_struct *curr = rq->curr;
3319 spin_lock(&rq->lock);
3320 update_cpu_load(rq);
3321 if (curr != rq->idle) /* FIXME: needed? */
3322 curr->sched_class->task_tick(rq, curr);
3323 spin_unlock(&rq->lock);
3326 rq->idle_at_tick = idle_cpu(cpu);
3327 trigger_load_balance(rq, cpu);
3331 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3333 void fastcall add_preempt_count(int val)
3338 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3340 preempt_count() += val;
3342 * Spinlock count overflowing soon?
3344 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3347 EXPORT_SYMBOL(add_preempt_count);
3349 void fastcall sub_preempt_count(int val)
3354 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3357 * Is the spinlock portion underflowing?
3359 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3360 !(preempt_count() & PREEMPT_MASK)))
3363 preempt_count() -= val;
3365 EXPORT_SYMBOL(sub_preempt_count);
3370 * Print scheduling while atomic bug:
3372 static noinline void __schedule_bug(struct task_struct *prev)
3374 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3375 prev->comm, preempt_count(), prev->pid);
3376 debug_show_held_locks(prev);
3377 if (irqs_disabled())
3378 print_irqtrace_events(prev);
3383 * Various schedule()-time debugging checks and statistics:
3385 static inline void schedule_debug(struct task_struct *prev)
3388 * Test if we are atomic. Since do_exit() needs to call into
3389 * schedule() atomically, we ignore that path for now.
3390 * Otherwise, whine if we are scheduling when we should not be.
3392 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3393 __schedule_bug(prev);
3395 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3397 schedstat_inc(this_rq(), sched_cnt);
3401 * Pick up the highest-prio task:
3403 static inline struct task_struct *
3404 pick_next_task(struct rq *rq, struct task_struct *prev, u64 now)
3406 struct sched_class *class;
3407 struct task_struct *p;
3410 * Optimization: we know that if all tasks are in
3411 * the fair class we can call that function directly:
3413 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3414 p = fair_sched_class.pick_next_task(rq, now);
3419 class = sched_class_highest;
3421 p = class->pick_next_task(rq, now);
3425 * Will never be NULL as the idle class always
3426 * returns a non-NULL p:
3428 class = class->next;
3433 * schedule() is the main scheduler function.
3435 asmlinkage void __sched schedule(void)
3437 struct task_struct *prev, *next;
3445 cpu = smp_processor_id();
3449 switch_count = &prev->nivcsw;
3451 release_kernel_lock(prev);
3452 need_resched_nonpreemptible:
3454 schedule_debug(prev);
3456 spin_lock_irq(&rq->lock);
3457 clear_tsk_need_resched(prev);
3458 now = __rq_clock(rq);
3460 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3461 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3462 unlikely(signal_pending(prev)))) {
3463 prev->state = TASK_RUNNING;
3465 deactivate_task(rq, prev, 1, now);
3467 switch_count = &prev->nvcsw;
3470 if (unlikely(!rq->nr_running))
3471 idle_balance(cpu, rq);
3473 prev->sched_class->put_prev_task(rq, prev, now);
3474 next = pick_next_task(rq, prev, now);
3476 sched_info_switch(prev, next);
3478 if (likely(prev != next)) {
3483 context_switch(rq, prev, next); /* unlocks the rq */
3485 spin_unlock_irq(&rq->lock);
3487 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3488 cpu = smp_processor_id();
3490 goto need_resched_nonpreemptible;
3492 preempt_enable_no_resched();
3493 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3496 EXPORT_SYMBOL(schedule);
3498 #ifdef CONFIG_PREEMPT
3500 * this is the entry point to schedule() from in-kernel preemption
3501 * off of preempt_enable. Kernel preemptions off return from interrupt
3502 * occur there and call schedule directly.
3504 asmlinkage void __sched preempt_schedule(void)
3506 struct thread_info *ti = current_thread_info();
3507 #ifdef CONFIG_PREEMPT_BKL
3508 struct task_struct *task = current;
3509 int saved_lock_depth;
3512 * If there is a non-zero preempt_count or interrupts are disabled,
3513 * we do not want to preempt the current task. Just return..
3515 if (likely(ti->preempt_count || irqs_disabled()))
3519 add_preempt_count(PREEMPT_ACTIVE);
3521 * We keep the big kernel semaphore locked, but we
3522 * clear ->lock_depth so that schedule() doesnt
3523 * auto-release the semaphore:
3525 #ifdef CONFIG_PREEMPT_BKL
3526 saved_lock_depth = task->lock_depth;
3527 task->lock_depth = -1;
3530 #ifdef CONFIG_PREEMPT_BKL
3531 task->lock_depth = saved_lock_depth;
3533 sub_preempt_count(PREEMPT_ACTIVE);
3535 /* we could miss a preemption opportunity between schedule and now */
3537 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3540 EXPORT_SYMBOL(preempt_schedule);
3543 * this is the entry point to schedule() from kernel preemption
3544 * off of irq context.
3545 * Note, that this is called and return with irqs disabled. This will
3546 * protect us against recursive calling from irq.
3548 asmlinkage void __sched preempt_schedule_irq(void)
3550 struct thread_info *ti = current_thread_info();
3551 #ifdef CONFIG_PREEMPT_BKL
3552 struct task_struct *task = current;
3553 int saved_lock_depth;
3555 /* Catch callers which need to be fixed */
3556 BUG_ON(ti->preempt_count || !irqs_disabled());
3559 add_preempt_count(PREEMPT_ACTIVE);
3561 * We keep the big kernel semaphore locked, but we
3562 * clear ->lock_depth so that schedule() doesnt
3563 * auto-release the semaphore:
3565 #ifdef CONFIG_PREEMPT_BKL
3566 saved_lock_depth = task->lock_depth;
3567 task->lock_depth = -1;
3571 local_irq_disable();
3572 #ifdef CONFIG_PREEMPT_BKL
3573 task->lock_depth = saved_lock_depth;
3575 sub_preempt_count(PREEMPT_ACTIVE);
3577 /* we could miss a preemption opportunity between schedule and now */
3579 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3583 #endif /* CONFIG_PREEMPT */
3585 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3588 return try_to_wake_up(curr->private, mode, sync);
3590 EXPORT_SYMBOL(default_wake_function);
3593 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3594 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3595 * number) then we wake all the non-exclusive tasks and one exclusive task.
3597 * There are circumstances in which we can try to wake a task which has already
3598 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3599 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3601 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3602 int nr_exclusive, int sync, void *key)
3604 struct list_head *tmp, *next;
3606 list_for_each_safe(tmp, next, &q->task_list) {
3607 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3608 unsigned flags = curr->flags;
3610 if (curr->func(curr, mode, sync, key) &&
3611 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3617 * __wake_up - wake up threads blocked on a waitqueue.
3619 * @mode: which threads
3620 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3621 * @key: is directly passed to the wakeup function
3623 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3624 int nr_exclusive, void *key)
3626 unsigned long flags;
3628 spin_lock_irqsave(&q->lock, flags);
3629 __wake_up_common(q, mode, nr_exclusive, 0, key);
3630 spin_unlock_irqrestore(&q->lock, flags);
3632 EXPORT_SYMBOL(__wake_up);
3635 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3637 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3639 __wake_up_common(q, mode, 1, 0, NULL);
3643 * __wake_up_sync - wake up threads blocked on a waitqueue.
3645 * @mode: which threads
3646 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3648 * The sync wakeup differs that the waker knows that it will schedule
3649 * away soon, so while the target thread will be woken up, it will not
3650 * be migrated to another CPU - ie. the two threads are 'synchronized'
3651 * with each other. This can prevent needless bouncing between CPUs.
3653 * On UP it can prevent extra preemption.
3656 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3658 unsigned long flags;
3664 if (unlikely(!nr_exclusive))
3667 spin_lock_irqsave(&q->lock, flags);
3668 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3669 spin_unlock_irqrestore(&q->lock, flags);
3671 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3673 void fastcall complete(struct completion *x)
3675 unsigned long flags;
3677 spin_lock_irqsave(&x->wait.lock, flags);
3679 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3681 spin_unlock_irqrestore(&x->wait.lock, flags);
3683 EXPORT_SYMBOL(complete);
3685 void fastcall complete_all(struct completion *x)
3687 unsigned long flags;
3689 spin_lock_irqsave(&x->wait.lock, flags);
3690 x->done += UINT_MAX/2;
3691 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3693 spin_unlock_irqrestore(&x->wait.lock, flags);
3695 EXPORT_SYMBOL(complete_all);
3697 void fastcall __sched wait_for_completion(struct completion *x)
3701 spin_lock_irq(&x->wait.lock);
3703 DECLARE_WAITQUEUE(wait, current);
3705 wait.flags |= WQ_FLAG_EXCLUSIVE;
3706 __add_wait_queue_tail(&x->wait, &wait);
3708 __set_current_state(TASK_UNINTERRUPTIBLE);
3709 spin_unlock_irq(&x->wait.lock);
3711 spin_lock_irq(&x->wait.lock);
3713 __remove_wait_queue(&x->wait, &wait);
3716 spin_unlock_irq(&x->wait.lock);
3718 EXPORT_SYMBOL(wait_for_completion);
3720 unsigned long fastcall __sched
3721 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3725 spin_lock_irq(&x->wait.lock);
3727 DECLARE_WAITQUEUE(wait, current);
3729 wait.flags |= WQ_FLAG_EXCLUSIVE;
3730 __add_wait_queue_tail(&x->wait, &wait);
3732 __set_current_state(TASK_UNINTERRUPTIBLE);
3733 spin_unlock_irq(&x->wait.lock);
3734 timeout = schedule_timeout(timeout);
3735 spin_lock_irq(&x->wait.lock);
3737 __remove_wait_queue(&x->wait, &wait);
3741 __remove_wait_queue(&x->wait, &wait);
3745 spin_unlock_irq(&x->wait.lock);
3748 EXPORT_SYMBOL(wait_for_completion_timeout);
3750 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3756 spin_lock_irq(&x->wait.lock);
3758 DECLARE_WAITQUEUE(wait, current);
3760 wait.flags |= WQ_FLAG_EXCLUSIVE;
3761 __add_wait_queue_tail(&x->wait, &wait);
3763 if (signal_pending(current)) {
3765 __remove_wait_queue(&x->wait, &wait);
3768 __set_current_state(TASK_INTERRUPTIBLE);
3769 spin_unlock_irq(&x->wait.lock);
3771 spin_lock_irq(&x->wait.lock);
3773 __remove_wait_queue(&x->wait, &wait);
3777 spin_unlock_irq(&x->wait.lock);
3781 EXPORT_SYMBOL(wait_for_completion_interruptible);
3783 unsigned long fastcall __sched
3784 wait_for_completion_interruptible_timeout(struct completion *x,
3785 unsigned long timeout)
3789 spin_lock_irq(&x->wait.lock);
3791 DECLARE_WAITQUEUE(wait, current);
3793 wait.flags |= WQ_FLAG_EXCLUSIVE;
3794 __add_wait_queue_tail(&x->wait, &wait);
3796 if (signal_pending(current)) {
3797 timeout = -ERESTARTSYS;
3798 __remove_wait_queue(&x->wait, &wait);
3801 __set_current_state(TASK_INTERRUPTIBLE);
3802 spin_unlock_irq(&x->wait.lock);
3803 timeout = schedule_timeout(timeout);
3804 spin_lock_irq(&x->wait.lock);
3806 __remove_wait_queue(&x->wait, &wait);
3810 __remove_wait_queue(&x->wait, &wait);
3814 spin_unlock_irq(&x->wait.lock);
3817 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3820 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3822 spin_lock_irqsave(&q->lock, *flags);
3823 __add_wait_queue(q, wait);
3824 spin_unlock(&q->lock);
3828 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3830 spin_lock_irq(&q->lock);
3831 __remove_wait_queue(q, wait);
3832 spin_unlock_irqrestore(&q->lock, *flags);
3835 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3837 unsigned long flags;
3840 init_waitqueue_entry(&wait, current);
3842 current->state = TASK_INTERRUPTIBLE;
3844 sleep_on_head(q, &wait, &flags);
3846 sleep_on_tail(q, &wait, &flags);
3848 EXPORT_SYMBOL(interruptible_sleep_on);
3851 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3853 unsigned long flags;
3856 init_waitqueue_entry(&wait, current);
3858 current->state = TASK_INTERRUPTIBLE;
3860 sleep_on_head(q, &wait, &flags);
3861 timeout = schedule_timeout(timeout);
3862 sleep_on_tail(q, &wait, &flags);
3866 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3868 void __sched sleep_on(wait_queue_head_t *q)
3870 unsigned long flags;
3873 init_waitqueue_entry(&wait, current);
3875 current->state = TASK_UNINTERRUPTIBLE;
3877 sleep_on_head(q, &wait, &flags);
3879 sleep_on_tail(q, &wait, &flags);
3881 EXPORT_SYMBOL(sleep_on);
3883 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3885 unsigned long flags;
3888 init_waitqueue_entry(&wait, current);
3890 current->state = TASK_UNINTERRUPTIBLE;
3892 sleep_on_head(q, &wait, &flags);
3893 timeout = schedule_timeout(timeout);
3894 sleep_on_tail(q, &wait, &flags);
3898 EXPORT_SYMBOL(sleep_on_timeout);
3900 #ifdef CONFIG_RT_MUTEXES
3903 * rt_mutex_setprio - set the current priority of a task
3905 * @prio: prio value (kernel-internal form)
3907 * This function changes the 'effective' priority of a task. It does
3908 * not touch ->normal_prio like __setscheduler().
3910 * Used by the rt_mutex code to implement priority inheritance logic.
3912 void rt_mutex_setprio(struct task_struct *p, int prio)
3914 unsigned long flags;
3919 BUG_ON(prio < 0 || prio > MAX_PRIO);
3921 rq = task_rq_lock(p, &flags);
3925 on_rq = p->se.on_rq;
3927 dequeue_task(rq, p, 0, now);
3930 p->sched_class = &rt_sched_class;
3932 p->sched_class = &fair_sched_class;
3937 enqueue_task(rq, p, 0, now);
3939 * Reschedule if we are currently running on this runqueue and
3940 * our priority decreased, or if we are not currently running on
3941 * this runqueue and our priority is higher than the current's
3943 if (task_running(rq, p)) {
3944 if (p->prio > oldprio)
3945 resched_task(rq->curr);
3947 check_preempt_curr(rq, p);
3950 task_rq_unlock(rq, &flags);
3955 void set_user_nice(struct task_struct *p, long nice)
3957 int old_prio, delta, on_rq;
3958 unsigned long flags;
3962 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3965 * We have to be careful, if called from sys_setpriority(),
3966 * the task might be in the middle of scheduling on another CPU.
3968 rq = task_rq_lock(p, &flags);
3971 * The RT priorities are set via sched_setscheduler(), but we still
3972 * allow the 'normal' nice value to be set - but as expected
3973 * it wont have any effect on scheduling until the task is
3974 * SCHED_FIFO/SCHED_RR:
3976 if (task_has_rt_policy(p)) {
3977 p->static_prio = NICE_TO_PRIO(nice);
3980 on_rq = p->se.on_rq;
3982 dequeue_task(rq, p, 0, now);
3983 dec_load(rq, p, now);
3986 p->static_prio = NICE_TO_PRIO(nice);
3989 p->prio = effective_prio(p);
3990 delta = p->prio - old_prio;
3993 enqueue_task(rq, p, 0, now);
3994 inc_load(rq, p, now);
3996 * If the task increased its priority or is running and
3997 * lowered its priority, then reschedule its CPU:
3999 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4000 resched_task(rq->curr);
4003 task_rq_unlock(rq, &flags);
4005 EXPORT_SYMBOL(set_user_nice);
4008 * can_nice - check if a task can reduce its nice value
4012 int can_nice(const struct task_struct *p, const int nice)
4014 /* convert nice value [19,-20] to rlimit style value [1,40] */
4015 int nice_rlim = 20 - nice;
4017 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4018 capable(CAP_SYS_NICE));
4021 #ifdef __ARCH_WANT_SYS_NICE
4024 * sys_nice - change the priority of the current process.
4025 * @increment: priority increment
4027 * sys_setpriority is a more generic, but much slower function that
4028 * does similar things.
4030 asmlinkage long sys_nice(int increment)
4035 * Setpriority might change our priority at the same moment.
4036 * We don't have to worry. Conceptually one call occurs first
4037 * and we have a single winner.
4039 if (increment < -40)
4044 nice = PRIO_TO_NICE(current->static_prio) + increment;
4050 if (increment < 0 && !can_nice(current, nice))
4053 retval = security_task_setnice(current, nice);
4057 set_user_nice(current, nice);
4064 * task_prio - return the priority value of a given task.
4065 * @p: the task in question.
4067 * This is the priority value as seen by users in /proc.
4068 * RT tasks are offset by -200. Normal tasks are centered
4069 * around 0, value goes from -16 to +15.
4071 int task_prio(const struct task_struct *p)
4073 return p->prio - MAX_RT_PRIO;
4077 * task_nice - return the nice value of a given task.
4078 * @p: the task in question.
4080 int task_nice(const struct task_struct *p)
4082 return TASK_NICE(p);
4084 EXPORT_SYMBOL_GPL(task_nice);
4087 * idle_cpu - is a given cpu idle currently?
4088 * @cpu: the processor in question.
4090 int idle_cpu(int cpu)
4092 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4096 * idle_task - return the idle task for a given cpu.
4097 * @cpu: the processor in question.
4099 struct task_struct *idle_task(int cpu)
4101 return cpu_rq(cpu)->idle;
4105 * find_process_by_pid - find a process with a matching PID value.
4106 * @pid: the pid in question.
4108 static inline struct task_struct *find_process_by_pid(pid_t pid)
4110 return pid ? find_task_by_pid(pid) : current;
4113 /* Actually do priority change: must hold rq lock. */
4115 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4117 BUG_ON(p->se.on_rq);
4120 switch (p->policy) {
4124 p->sched_class = &fair_sched_class;
4128 p->sched_class = &rt_sched_class;
4132 p->rt_priority = prio;
4133 p->normal_prio = normal_prio(p);
4134 /* we are holding p->pi_lock already */
4135 p->prio = rt_mutex_getprio(p);
4140 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4141 * @p: the task in question.
4142 * @policy: new policy.
4143 * @param: structure containing the new RT priority.
4145 * NOTE that the task may be already dead.
4147 int sched_setscheduler(struct task_struct *p, int policy,
4148 struct sched_param *param)
4150 int retval, oldprio, oldpolicy = -1, on_rq;
4151 unsigned long flags;
4154 /* may grab non-irq protected spin_locks */
4155 BUG_ON(in_interrupt());
4157 /* double check policy once rq lock held */
4159 policy = oldpolicy = p->policy;
4160 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4161 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4162 policy != SCHED_IDLE)
4165 * Valid priorities for SCHED_FIFO and SCHED_RR are
4166 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4167 * SCHED_BATCH and SCHED_IDLE is 0.
4169 if (param->sched_priority < 0 ||
4170 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4171 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4173 if (rt_policy(policy) != (param->sched_priority != 0))
4177 * Allow unprivileged RT tasks to decrease priority:
4179 if (!capable(CAP_SYS_NICE)) {
4180 if (rt_policy(policy)) {
4181 unsigned long rlim_rtprio;
4183 if (!lock_task_sighand(p, &flags))
4185 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4186 unlock_task_sighand(p, &flags);
4188 /* can't set/change the rt policy */
4189 if (policy != p->policy && !rlim_rtprio)
4192 /* can't increase priority */
4193 if (param->sched_priority > p->rt_priority &&
4194 param->sched_priority > rlim_rtprio)
4198 * Like positive nice levels, dont allow tasks to
4199 * move out of SCHED_IDLE either:
4201 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4204 /* can't change other user's priorities */
4205 if ((current->euid != p->euid) &&
4206 (current->euid != p->uid))
4210 retval = security_task_setscheduler(p, policy, param);
4214 * make sure no PI-waiters arrive (or leave) while we are
4215 * changing the priority of the task:
4217 spin_lock_irqsave(&p->pi_lock, flags);
4219 * To be able to change p->policy safely, the apropriate
4220 * runqueue lock must be held.
4222 rq = __task_rq_lock(p);
4223 /* recheck policy now with rq lock held */
4224 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4225 policy = oldpolicy = -1;
4226 __task_rq_unlock(rq);
4227 spin_unlock_irqrestore(&p->pi_lock, flags);
4230 on_rq = p->se.on_rq;
4232 deactivate_task(rq, p, 0, rq_clock(rq));
4234 __setscheduler(rq, p, policy, param->sched_priority);
4236 activate_task(rq, p, 0);
4238 * Reschedule if we are currently running on this runqueue and
4239 * our priority decreased, or if we are not currently running on
4240 * this runqueue and our priority is higher than the current's
4242 if (task_running(rq, p)) {
4243 if (p->prio > oldprio)
4244 resched_task(rq->curr);
4246 check_preempt_curr(rq, p);
4249 __task_rq_unlock(rq);
4250 spin_unlock_irqrestore(&p->pi_lock, flags);
4252 rt_mutex_adjust_pi(p);
4256 EXPORT_SYMBOL_GPL(sched_setscheduler);
4259 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4261 struct sched_param lparam;
4262 struct task_struct *p;
4265 if (!param || pid < 0)
4267 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4272 p = find_process_by_pid(pid);
4274 retval = sched_setscheduler(p, policy, &lparam);
4281 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4282 * @pid: the pid in question.
4283 * @policy: new policy.
4284 * @param: structure containing the new RT priority.
4286 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4287 struct sched_param __user *param)
4289 /* negative values for policy are not valid */
4293 return do_sched_setscheduler(pid, policy, param);
4297 * sys_sched_setparam - set/change the RT priority of a thread
4298 * @pid: the pid in question.
4299 * @param: structure containing the new RT priority.
4301 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4303 return do_sched_setscheduler(pid, -1, param);
4307 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4308 * @pid: the pid in question.
4310 asmlinkage long sys_sched_getscheduler(pid_t pid)
4312 struct task_struct *p;
4313 int retval = -EINVAL;
4319 read_lock(&tasklist_lock);
4320 p = find_process_by_pid(pid);
4322 retval = security_task_getscheduler(p);
4326 read_unlock(&tasklist_lock);
4333 * sys_sched_getscheduler - get the RT priority of a thread
4334 * @pid: the pid in question.
4335 * @param: structure containing the RT priority.
4337 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4339 struct sched_param lp;
4340 struct task_struct *p;
4341 int retval = -EINVAL;
4343 if (!param || pid < 0)
4346 read_lock(&tasklist_lock);
4347 p = find_process_by_pid(pid);
4352 retval = security_task_getscheduler(p);
4356 lp.sched_priority = p->rt_priority;
4357 read_unlock(&tasklist_lock);
4360 * This one might sleep, we cannot do it with a spinlock held ...
4362 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4368 read_unlock(&tasklist_lock);
4372 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4374 cpumask_t cpus_allowed;
4375 struct task_struct *p;
4378 mutex_lock(&sched_hotcpu_mutex);
4379 read_lock(&tasklist_lock);
4381 p = find_process_by_pid(pid);
4383 read_unlock(&tasklist_lock);
4384 mutex_unlock(&sched_hotcpu_mutex);
4389 * It is not safe to call set_cpus_allowed with the
4390 * tasklist_lock held. We will bump the task_struct's
4391 * usage count and then drop tasklist_lock.
4394 read_unlock(&tasklist_lock);
4397 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4398 !capable(CAP_SYS_NICE))
4401 retval = security_task_setscheduler(p, 0, NULL);
4405 cpus_allowed = cpuset_cpus_allowed(p);
4406 cpus_and(new_mask, new_mask, cpus_allowed);
4407 retval = set_cpus_allowed(p, new_mask);
4411 mutex_unlock(&sched_hotcpu_mutex);
4415 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4416 cpumask_t *new_mask)
4418 if (len < sizeof(cpumask_t)) {
4419 memset(new_mask, 0, sizeof(cpumask_t));
4420 } else if (len > sizeof(cpumask_t)) {
4421 len = sizeof(cpumask_t);
4423 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4427 * sys_sched_setaffinity - set the cpu affinity of a process
4428 * @pid: pid of the process
4429 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4430 * @user_mask_ptr: user-space pointer to the new cpu mask
4432 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4433 unsigned long __user *user_mask_ptr)
4438 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4442 return sched_setaffinity(pid, new_mask);
4446 * Represents all cpu's present in the system
4447 * In systems capable of hotplug, this map could dynamically grow
4448 * as new cpu's are detected in the system via any platform specific
4449 * method, such as ACPI for e.g.
4452 cpumask_t cpu_present_map __read_mostly;
4453 EXPORT_SYMBOL(cpu_present_map);
4456 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4457 EXPORT_SYMBOL(cpu_online_map);
4459 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4460 EXPORT_SYMBOL(cpu_possible_map);
4463 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4465 struct task_struct *p;
4468 mutex_lock(&sched_hotcpu_mutex);
4469 read_lock(&tasklist_lock);
4472 p = find_process_by_pid(pid);
4476 retval = security_task_getscheduler(p);
4480 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4483 read_unlock(&tasklist_lock);
4484 mutex_unlock(&sched_hotcpu_mutex);
4490 * sys_sched_getaffinity - get the cpu affinity of a process
4491 * @pid: pid of the process
4492 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4493 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4495 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4496 unsigned long __user *user_mask_ptr)
4501 if (len < sizeof(cpumask_t))
4504 ret = sched_getaffinity(pid, &mask);
4508 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4511 return sizeof(cpumask_t);
4515 * sys_sched_yield - yield the current processor to other threads.
4517 * This function yields the current CPU to other tasks. If there are no
4518 * other threads running on this CPU then this function will return.
4520 asmlinkage long sys_sched_yield(void)
4522 struct rq *rq = this_rq_lock();
4524 schedstat_inc(rq, yld_cnt);
4525 if (unlikely(rq->nr_running == 1))
4526 schedstat_inc(rq, yld_act_empty);
4528 current->sched_class->yield_task(rq, current);
4531 * Since we are going to call schedule() anyway, there's
4532 * no need to preempt or enable interrupts:
4534 __release(rq->lock);
4535 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4536 _raw_spin_unlock(&rq->lock);
4537 preempt_enable_no_resched();
4544 static void __cond_resched(void)
4546 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4547 __might_sleep(__FILE__, __LINE__);
4550 * The BKS might be reacquired before we have dropped
4551 * PREEMPT_ACTIVE, which could trigger a second
4552 * cond_resched() call.
4555 add_preempt_count(PREEMPT_ACTIVE);
4557 sub_preempt_count(PREEMPT_ACTIVE);
4558 } while (need_resched());
4561 int __sched cond_resched(void)
4563 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4564 system_state == SYSTEM_RUNNING) {
4570 EXPORT_SYMBOL(cond_resched);
4573 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4574 * call schedule, and on return reacquire the lock.
4576 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4577 * operations here to prevent schedule() from being called twice (once via
4578 * spin_unlock(), once by hand).
4580 int cond_resched_lock(spinlock_t *lock)
4584 if (need_lockbreak(lock)) {
4590 if (need_resched() && system_state == SYSTEM_RUNNING) {
4591 spin_release(&lock->dep_map, 1, _THIS_IP_);
4592 _raw_spin_unlock(lock);
4593 preempt_enable_no_resched();
4600 EXPORT_SYMBOL(cond_resched_lock);
4602 int __sched cond_resched_softirq(void)
4604 BUG_ON(!in_softirq());
4606 if (need_resched() && system_state == SYSTEM_RUNNING) {
4614 EXPORT_SYMBOL(cond_resched_softirq);
4617 * yield - yield the current processor to other threads.
4619 * This is a shortcut for kernel-space yielding - it marks the
4620 * thread runnable and calls sys_sched_yield().
4622 void __sched yield(void)
4624 set_current_state(TASK_RUNNING);
4627 EXPORT_SYMBOL(yield);
4630 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4631 * that process accounting knows that this is a task in IO wait state.
4633 * But don't do that if it is a deliberate, throttling IO wait (this task
4634 * has set its backing_dev_info: the queue against which it should throttle)
4636 void __sched io_schedule(void)
4638 struct rq *rq = &__raw_get_cpu_var(runqueues);
4640 delayacct_blkio_start();
4641 atomic_inc(&rq->nr_iowait);
4643 atomic_dec(&rq->nr_iowait);
4644 delayacct_blkio_end();
4646 EXPORT_SYMBOL(io_schedule);
4648 long __sched io_schedule_timeout(long timeout)
4650 struct rq *rq = &__raw_get_cpu_var(runqueues);
4653 delayacct_blkio_start();
4654 atomic_inc(&rq->nr_iowait);
4655 ret = schedule_timeout(timeout);
4656 atomic_dec(&rq->nr_iowait);
4657 delayacct_blkio_end();
4662 * sys_sched_get_priority_max - return maximum RT priority.
4663 * @policy: scheduling class.
4665 * this syscall returns the maximum rt_priority that can be used
4666 * by a given scheduling class.
4668 asmlinkage long sys_sched_get_priority_max(int policy)
4675 ret = MAX_USER_RT_PRIO-1;
4687 * sys_sched_get_priority_min - return minimum RT priority.
4688 * @policy: scheduling class.
4690 * this syscall returns the minimum rt_priority that can be used
4691 * by a given scheduling class.
4693 asmlinkage long sys_sched_get_priority_min(int policy)
4711 * sys_sched_rr_get_interval - return the default timeslice of a process.
4712 * @pid: pid of the process.
4713 * @interval: userspace pointer to the timeslice value.
4715 * this syscall writes the default timeslice value of a given process
4716 * into the user-space timespec buffer. A value of '0' means infinity.
4719 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4721 struct task_struct *p;
4722 int retval = -EINVAL;
4729 read_lock(&tasklist_lock);
4730 p = find_process_by_pid(pid);
4734 retval = security_task_getscheduler(p);
4738 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4739 0 : static_prio_timeslice(p->static_prio), &t);
4740 read_unlock(&tasklist_lock);
4741 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4745 read_unlock(&tasklist_lock);
4749 static const char stat_nam[] = "RSDTtZX";
4751 static void show_task(struct task_struct *p)
4753 unsigned long free = 0;
4756 state = p->state ? __ffs(p->state) + 1 : 0;
4757 printk("%-13.13s %c", p->comm,
4758 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4759 #if BITS_PER_LONG == 32
4760 if (state == TASK_RUNNING)
4761 printk(" running ");
4763 printk(" %08lx ", thread_saved_pc(p));
4765 if (state == TASK_RUNNING)
4766 printk(" running task ");
4768 printk(" %016lx ", thread_saved_pc(p));
4770 #ifdef CONFIG_DEBUG_STACK_USAGE
4772 unsigned long *n = end_of_stack(p);
4775 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4778 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4780 if (state != TASK_RUNNING)
4781 show_stack(p, NULL);
4784 void show_state_filter(unsigned long state_filter)
4786 struct task_struct *g, *p;
4788 #if BITS_PER_LONG == 32
4790 " task PC stack pid father\n");
4793 " task PC stack pid father\n");
4795 read_lock(&tasklist_lock);
4796 do_each_thread(g, p) {
4798 * reset the NMI-timeout, listing all files on a slow
4799 * console might take alot of time:
4801 touch_nmi_watchdog();
4802 if (!state_filter || (p->state & state_filter))
4804 } while_each_thread(g, p);
4806 touch_all_softlockup_watchdogs();
4808 #ifdef CONFIG_SCHED_DEBUG
4809 sysrq_sched_debug_show();
4811 read_unlock(&tasklist_lock);
4813 * Only show locks if all tasks are dumped:
4815 if (state_filter == -1)
4816 debug_show_all_locks();
4819 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4821 idle->sched_class = &idle_sched_class;
4825 * init_idle - set up an idle thread for a given CPU
4826 * @idle: task in question
4827 * @cpu: cpu the idle task belongs to
4829 * NOTE: this function does not set the idle thread's NEED_RESCHED
4830 * flag, to make booting more robust.
4832 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4834 struct rq *rq = cpu_rq(cpu);
4835 unsigned long flags;
4838 idle->se.exec_start = sched_clock();
4840 idle->prio = idle->normal_prio = MAX_PRIO;
4841 idle->cpus_allowed = cpumask_of_cpu(cpu);
4842 __set_task_cpu(idle, cpu);
4844 spin_lock_irqsave(&rq->lock, flags);
4845 rq->curr = rq->idle = idle;
4846 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4849 spin_unlock_irqrestore(&rq->lock, flags);
4851 /* Set the preempt count _outside_ the spinlocks! */
4852 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4853 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4855 task_thread_info(idle)->preempt_count = 0;
4858 * The idle tasks have their own, simple scheduling class:
4860 idle->sched_class = &idle_sched_class;
4864 * In a system that switches off the HZ timer nohz_cpu_mask
4865 * indicates which cpus entered this state. This is used
4866 * in the rcu update to wait only for active cpus. For system
4867 * which do not switch off the HZ timer nohz_cpu_mask should
4868 * always be CPU_MASK_NONE.
4870 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4873 * Increase the granularity value when there are more CPUs,
4874 * because with more CPUs the 'effective latency' as visible
4875 * to users decreases. But the relationship is not linear,
4876 * so pick a second-best guess by going with the log2 of the
4879 * This idea comes from the SD scheduler of Con Kolivas:
4881 static inline void sched_init_granularity(void)
4883 unsigned int factor = 1 + ilog2(num_online_cpus());
4884 const unsigned long gran_limit = 100000000;
4886 sysctl_sched_granularity *= factor;
4887 if (sysctl_sched_granularity > gran_limit)
4888 sysctl_sched_granularity = gran_limit;
4890 sysctl_sched_runtime_limit = sysctl_sched_granularity * 4;
4891 sysctl_sched_wakeup_granularity = sysctl_sched_granularity / 2;
4896 * This is how migration works:
4898 * 1) we queue a struct migration_req structure in the source CPU's
4899 * runqueue and wake up that CPU's migration thread.
4900 * 2) we down() the locked semaphore => thread blocks.
4901 * 3) migration thread wakes up (implicitly it forces the migrated
4902 * thread off the CPU)
4903 * 4) it gets the migration request and checks whether the migrated
4904 * task is still in the wrong runqueue.
4905 * 5) if it's in the wrong runqueue then the migration thread removes
4906 * it and puts it into the right queue.
4907 * 6) migration thread up()s the semaphore.
4908 * 7) we wake up and the migration is done.
4912 * Change a given task's CPU affinity. Migrate the thread to a
4913 * proper CPU and schedule it away if the CPU it's executing on
4914 * is removed from the allowed bitmask.
4916 * NOTE: the caller must have a valid reference to the task, the
4917 * task must not exit() & deallocate itself prematurely. The
4918 * call is not atomic; no spinlocks may be held.
4920 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4922 struct migration_req req;
4923 unsigned long flags;
4927 rq = task_rq_lock(p, &flags);
4928 if (!cpus_intersects(new_mask, cpu_online_map)) {
4933 p->cpus_allowed = new_mask;
4934 /* Can the task run on the task's current CPU? If so, we're done */
4935 if (cpu_isset(task_cpu(p), new_mask))
4938 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4939 /* Need help from migration thread: drop lock and wait. */
4940 task_rq_unlock(rq, &flags);
4941 wake_up_process(rq->migration_thread);
4942 wait_for_completion(&req.done);
4943 tlb_migrate_finish(p->mm);
4947 task_rq_unlock(rq, &flags);
4951 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4954 * Move (not current) task off this cpu, onto dest cpu. We're doing
4955 * this because either it can't run here any more (set_cpus_allowed()
4956 * away from this CPU, or CPU going down), or because we're
4957 * attempting to rebalance this task on exec (sched_exec).
4959 * So we race with normal scheduler movements, but that's OK, as long
4960 * as the task is no longer on this CPU.
4962 * Returns non-zero if task was successfully migrated.
4964 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4966 struct rq *rq_dest, *rq_src;
4969 if (unlikely(cpu_is_offline(dest_cpu)))
4972 rq_src = cpu_rq(src_cpu);
4973 rq_dest = cpu_rq(dest_cpu);
4975 double_rq_lock(rq_src, rq_dest);
4976 /* Already moved. */
4977 if (task_cpu(p) != src_cpu)
4979 /* Affinity changed (again). */
4980 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4983 on_rq = p->se.on_rq;
4985 deactivate_task(rq_src, p, 0, rq_clock(rq_src));
4986 set_task_cpu(p, dest_cpu);
4988 activate_task(rq_dest, p, 0);
4989 check_preempt_curr(rq_dest, p);
4993 double_rq_unlock(rq_src, rq_dest);
4998 * migration_thread - this is a highprio system thread that performs
4999 * thread migration by bumping thread off CPU then 'pushing' onto
5002 static int migration_thread(void *data)
5004 int cpu = (long)data;
5008 BUG_ON(rq->migration_thread != current);
5010 set_current_state(TASK_INTERRUPTIBLE);
5011 while (!kthread_should_stop()) {
5012 struct migration_req *req;
5013 struct list_head *head;
5015 spin_lock_irq(&rq->lock);
5017 if (cpu_is_offline(cpu)) {
5018 spin_unlock_irq(&rq->lock);
5022 if (rq->active_balance) {
5023 active_load_balance(rq, cpu);
5024 rq->active_balance = 0;
5027 head = &rq->migration_queue;
5029 if (list_empty(head)) {
5030 spin_unlock_irq(&rq->lock);
5032 set_current_state(TASK_INTERRUPTIBLE);
5035 req = list_entry(head->next, struct migration_req, list);
5036 list_del_init(head->next);
5038 spin_unlock(&rq->lock);
5039 __migrate_task(req->task, cpu, req->dest_cpu);
5042 complete(&req->done);
5044 __set_current_state(TASK_RUNNING);
5048 /* Wait for kthread_stop */
5049 set_current_state(TASK_INTERRUPTIBLE);
5050 while (!kthread_should_stop()) {
5052 set_current_state(TASK_INTERRUPTIBLE);
5054 __set_current_state(TASK_RUNNING);
5058 #ifdef CONFIG_HOTPLUG_CPU
5060 * Figure out where task on dead CPU should go, use force if neccessary.
5061 * NOTE: interrupts should be disabled by the caller
5063 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5065 unsigned long flags;
5072 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5073 cpus_and(mask, mask, p->cpus_allowed);
5074 dest_cpu = any_online_cpu(mask);
5076 /* On any allowed CPU? */
5077 if (dest_cpu == NR_CPUS)
5078 dest_cpu = any_online_cpu(p->cpus_allowed);
5080 /* No more Mr. Nice Guy. */
5081 if (dest_cpu == NR_CPUS) {
5082 rq = task_rq_lock(p, &flags);
5083 cpus_setall(p->cpus_allowed);
5084 dest_cpu = any_online_cpu(p->cpus_allowed);
5085 task_rq_unlock(rq, &flags);
5088 * Don't tell them about moving exiting tasks or
5089 * kernel threads (both mm NULL), since they never
5092 if (p->mm && printk_ratelimit())
5093 printk(KERN_INFO "process %d (%s) no "
5094 "longer affine to cpu%d\n",
5095 p->pid, p->comm, dead_cpu);
5097 if (!__migrate_task(p, dead_cpu, dest_cpu))
5102 * While a dead CPU has no uninterruptible tasks queued at this point,
5103 * it might still have a nonzero ->nr_uninterruptible counter, because
5104 * for performance reasons the counter is not stricly tracking tasks to
5105 * their home CPUs. So we just add the counter to another CPU's counter,
5106 * to keep the global sum constant after CPU-down:
5108 static void migrate_nr_uninterruptible(struct rq *rq_src)
5110 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5111 unsigned long flags;
5113 local_irq_save(flags);
5114 double_rq_lock(rq_src, rq_dest);
5115 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5116 rq_src->nr_uninterruptible = 0;
5117 double_rq_unlock(rq_src, rq_dest);
5118 local_irq_restore(flags);
5121 /* Run through task list and migrate tasks from the dead cpu. */
5122 static void migrate_live_tasks(int src_cpu)
5124 struct task_struct *p, *t;
5126 write_lock_irq(&tasklist_lock);
5128 do_each_thread(t, p) {
5132 if (task_cpu(p) == src_cpu)
5133 move_task_off_dead_cpu(src_cpu, p);
5134 } while_each_thread(t, p);
5136 write_unlock_irq(&tasklist_lock);
5140 * Schedules idle task to be the next runnable task on current CPU.
5141 * It does so by boosting its priority to highest possible and adding it to
5142 * the _front_ of the runqueue. Used by CPU offline code.
5144 void sched_idle_next(void)
5146 int this_cpu = smp_processor_id();
5147 struct rq *rq = cpu_rq(this_cpu);
5148 struct task_struct *p = rq->idle;
5149 unsigned long flags;
5151 /* cpu has to be offline */
5152 BUG_ON(cpu_online(this_cpu));
5155 * Strictly not necessary since rest of the CPUs are stopped by now
5156 * and interrupts disabled on the current cpu.
5158 spin_lock_irqsave(&rq->lock, flags);
5160 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5162 /* Add idle task to the _front_ of its priority queue: */
5163 activate_idle_task(p, rq);
5165 spin_unlock_irqrestore(&rq->lock, flags);
5169 * Ensures that the idle task is using init_mm right before its cpu goes
5172 void idle_task_exit(void)
5174 struct mm_struct *mm = current->active_mm;
5176 BUG_ON(cpu_online(smp_processor_id()));
5179 switch_mm(mm, &init_mm, current);
5183 /* called under rq->lock with disabled interrupts */
5184 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5186 struct rq *rq = cpu_rq(dead_cpu);
5188 /* Must be exiting, otherwise would be on tasklist. */
5189 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5191 /* Cannot have done final schedule yet: would have vanished. */
5192 BUG_ON(p->state == TASK_DEAD);
5197 * Drop lock around migration; if someone else moves it,
5198 * that's OK. No task can be added to this CPU, so iteration is
5200 * NOTE: interrupts should be left disabled --dev@
5202 spin_unlock(&rq->lock);
5203 move_task_off_dead_cpu(dead_cpu, p);
5204 spin_lock(&rq->lock);
5209 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5210 static void migrate_dead_tasks(unsigned int dead_cpu)
5212 struct rq *rq = cpu_rq(dead_cpu);
5213 struct task_struct *next;
5216 if (!rq->nr_running)
5218 next = pick_next_task(rq, rq->curr, rq_clock(rq));
5221 migrate_dead(dead_cpu, next);
5225 #endif /* CONFIG_HOTPLUG_CPU */
5227 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5229 static struct ctl_table sd_ctl_dir[] = {
5231 .procname = "sched_domain",
5237 static struct ctl_table sd_ctl_root[] = {
5239 .procname = "kernel",
5241 .child = sd_ctl_dir,
5246 static struct ctl_table *sd_alloc_ctl_entry(int n)
5248 struct ctl_table *entry =
5249 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5252 memset(entry, 0, n * sizeof(struct ctl_table));
5258 set_table_entry(struct ctl_table *entry,
5259 const char *procname, void *data, int maxlen,
5260 mode_t mode, proc_handler *proc_handler)
5262 entry->procname = procname;
5264 entry->maxlen = maxlen;
5266 entry->proc_handler = proc_handler;
5269 static struct ctl_table *
5270 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5272 struct ctl_table *table = sd_alloc_ctl_entry(14);
5274 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5275 sizeof(long), 0644, proc_doulongvec_minmax);
5276 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5277 sizeof(long), 0644, proc_doulongvec_minmax);
5278 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5279 sizeof(int), 0644, proc_dointvec_minmax);
5280 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5281 sizeof(int), 0644, proc_dointvec_minmax);
5282 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5283 sizeof(int), 0644, proc_dointvec_minmax);
5284 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5285 sizeof(int), 0644, proc_dointvec_minmax);
5286 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5287 sizeof(int), 0644, proc_dointvec_minmax);
5288 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5289 sizeof(int), 0644, proc_dointvec_minmax);
5290 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5291 sizeof(int), 0644, proc_dointvec_minmax);
5292 set_table_entry(&table[10], "cache_nice_tries",
5293 &sd->cache_nice_tries,
5294 sizeof(int), 0644, proc_dointvec_minmax);
5295 set_table_entry(&table[12], "flags", &sd->flags,
5296 sizeof(int), 0644, proc_dointvec_minmax);
5301 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5303 struct ctl_table *entry, *table;
5304 struct sched_domain *sd;
5305 int domain_num = 0, i;
5308 for_each_domain(cpu, sd)
5310 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5313 for_each_domain(cpu, sd) {
5314 snprintf(buf, 32, "domain%d", i);
5315 entry->procname = kstrdup(buf, GFP_KERNEL);
5317 entry->child = sd_alloc_ctl_domain_table(sd);
5324 static struct ctl_table_header *sd_sysctl_header;
5325 static void init_sched_domain_sysctl(void)
5327 int i, cpu_num = num_online_cpus();
5328 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5331 sd_ctl_dir[0].child = entry;
5333 for (i = 0; i < cpu_num; i++, entry++) {
5334 snprintf(buf, 32, "cpu%d", i);
5335 entry->procname = kstrdup(buf, GFP_KERNEL);
5337 entry->child = sd_alloc_ctl_cpu_table(i);
5339 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5342 static void init_sched_domain_sysctl(void)
5348 * migration_call - callback that gets triggered when a CPU is added.
5349 * Here we can start up the necessary migration thread for the new CPU.
5351 static int __cpuinit
5352 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5354 struct task_struct *p;
5355 int cpu = (long)hcpu;
5356 unsigned long flags;
5360 case CPU_LOCK_ACQUIRE:
5361 mutex_lock(&sched_hotcpu_mutex);
5364 case CPU_UP_PREPARE:
5365 case CPU_UP_PREPARE_FROZEN:
5366 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5369 kthread_bind(p, cpu);
5370 /* Must be high prio: stop_machine expects to yield to it. */
5371 rq = task_rq_lock(p, &flags);
5372 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5373 task_rq_unlock(rq, &flags);
5374 cpu_rq(cpu)->migration_thread = p;
5378 case CPU_ONLINE_FROZEN:
5379 /* Strictly unneccessary, as first user will wake it. */
5380 wake_up_process(cpu_rq(cpu)->migration_thread);
5383 #ifdef CONFIG_HOTPLUG_CPU
5384 case CPU_UP_CANCELED:
5385 case CPU_UP_CANCELED_FROZEN:
5386 if (!cpu_rq(cpu)->migration_thread)
5388 /* Unbind it from offline cpu so it can run. Fall thru. */
5389 kthread_bind(cpu_rq(cpu)->migration_thread,
5390 any_online_cpu(cpu_online_map));
5391 kthread_stop(cpu_rq(cpu)->migration_thread);
5392 cpu_rq(cpu)->migration_thread = NULL;
5396 case CPU_DEAD_FROZEN:
5397 migrate_live_tasks(cpu);
5399 kthread_stop(rq->migration_thread);
5400 rq->migration_thread = NULL;
5401 /* Idle task back to normal (off runqueue, low prio) */
5402 rq = task_rq_lock(rq->idle, &flags);
5403 deactivate_task(rq, rq->idle, 0, rq_clock(rq));
5404 rq->idle->static_prio = MAX_PRIO;
5405 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5406 rq->idle->sched_class = &idle_sched_class;
5407 migrate_dead_tasks(cpu);
5408 task_rq_unlock(rq, &flags);
5409 migrate_nr_uninterruptible(rq);
5410 BUG_ON(rq->nr_running != 0);
5412 /* No need to migrate the tasks: it was best-effort if
5413 * they didn't take sched_hotcpu_mutex. Just wake up
5414 * the requestors. */
5415 spin_lock_irq(&rq->lock);
5416 while (!list_empty(&rq->migration_queue)) {
5417 struct migration_req *req;
5419 req = list_entry(rq->migration_queue.next,
5420 struct migration_req, list);
5421 list_del_init(&req->list);
5422 complete(&req->done);
5424 spin_unlock_irq(&rq->lock);
5427 case CPU_LOCK_RELEASE:
5428 mutex_unlock(&sched_hotcpu_mutex);
5434 /* Register at highest priority so that task migration (migrate_all_tasks)
5435 * happens before everything else.
5437 static struct notifier_block __cpuinitdata migration_notifier = {
5438 .notifier_call = migration_call,
5442 int __init migration_init(void)
5444 void *cpu = (void *)(long)smp_processor_id();
5447 /* Start one for the boot CPU: */
5448 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5449 BUG_ON(err == NOTIFY_BAD);
5450 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5451 register_cpu_notifier(&migration_notifier);
5459 /* Number of possible processor ids */
5460 int nr_cpu_ids __read_mostly = NR_CPUS;
5461 EXPORT_SYMBOL(nr_cpu_ids);
5463 #undef SCHED_DOMAIN_DEBUG
5464 #ifdef SCHED_DOMAIN_DEBUG
5465 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5470 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5474 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5479 struct sched_group *group = sd->groups;
5480 cpumask_t groupmask;
5482 cpumask_scnprintf(str, NR_CPUS, sd->span);
5483 cpus_clear(groupmask);
5486 for (i = 0; i < level + 1; i++)
5488 printk("domain %d: ", level);
5490 if (!(sd->flags & SD_LOAD_BALANCE)) {
5491 printk("does not load-balance\n");
5493 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5498 printk("span %s\n", str);
5500 if (!cpu_isset(cpu, sd->span))
5501 printk(KERN_ERR "ERROR: domain->span does not contain "
5503 if (!cpu_isset(cpu, group->cpumask))
5504 printk(KERN_ERR "ERROR: domain->groups does not contain"
5508 for (i = 0; i < level + 2; i++)
5514 printk(KERN_ERR "ERROR: group is NULL\n");
5518 if (!group->__cpu_power) {
5520 printk(KERN_ERR "ERROR: domain->cpu_power not "
5524 if (!cpus_weight(group->cpumask)) {
5526 printk(KERN_ERR "ERROR: empty group\n");
5529 if (cpus_intersects(groupmask, group->cpumask)) {
5531 printk(KERN_ERR "ERROR: repeated CPUs\n");
5534 cpus_or(groupmask, groupmask, group->cpumask);
5536 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5539 group = group->next;
5540 } while (group != sd->groups);
5543 if (!cpus_equal(sd->span, groupmask))
5544 printk(KERN_ERR "ERROR: groups don't span "
5552 if (!cpus_subset(groupmask, sd->span))
5553 printk(KERN_ERR "ERROR: parent span is not a superset "
5554 "of domain->span\n");
5559 # define sched_domain_debug(sd, cpu) do { } while (0)
5562 static int sd_degenerate(struct sched_domain *sd)
5564 if (cpus_weight(sd->span) == 1)
5567 /* Following flags need at least 2 groups */
5568 if (sd->flags & (SD_LOAD_BALANCE |
5569 SD_BALANCE_NEWIDLE |
5573 SD_SHARE_PKG_RESOURCES)) {
5574 if (sd->groups != sd->groups->next)
5578 /* Following flags don't use groups */
5579 if (sd->flags & (SD_WAKE_IDLE |
5588 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5590 unsigned long cflags = sd->flags, pflags = parent->flags;
5592 if (sd_degenerate(parent))
5595 if (!cpus_equal(sd->span, parent->span))
5598 /* Does parent contain flags not in child? */
5599 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5600 if (cflags & SD_WAKE_AFFINE)
5601 pflags &= ~SD_WAKE_BALANCE;
5602 /* Flags needing groups don't count if only 1 group in parent */
5603 if (parent->groups == parent->groups->next) {
5604 pflags &= ~(SD_LOAD_BALANCE |
5605 SD_BALANCE_NEWIDLE |
5609 SD_SHARE_PKG_RESOURCES);
5611 if (~cflags & pflags)
5618 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5619 * hold the hotplug lock.
5621 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5623 struct rq *rq = cpu_rq(cpu);
5624 struct sched_domain *tmp;
5626 /* Remove the sched domains which do not contribute to scheduling. */
5627 for (tmp = sd; tmp; tmp = tmp->parent) {
5628 struct sched_domain *parent = tmp->parent;
5631 if (sd_parent_degenerate(tmp, parent)) {
5632 tmp->parent = parent->parent;
5634 parent->parent->child = tmp;
5638 if (sd && sd_degenerate(sd)) {
5644 sched_domain_debug(sd, cpu);
5646 rcu_assign_pointer(rq->sd, sd);
5649 /* cpus with isolated domains */
5650 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5652 /* Setup the mask of cpus configured for isolated domains */
5653 static int __init isolated_cpu_setup(char *str)
5655 int ints[NR_CPUS], i;
5657 str = get_options(str, ARRAY_SIZE(ints), ints);
5658 cpus_clear(cpu_isolated_map);
5659 for (i = 1; i <= ints[0]; i++)
5660 if (ints[i] < NR_CPUS)
5661 cpu_set(ints[i], cpu_isolated_map);
5665 __setup ("isolcpus=", isolated_cpu_setup);
5668 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5669 * to a function which identifies what group(along with sched group) a CPU
5670 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5671 * (due to the fact that we keep track of groups covered with a cpumask_t).
5673 * init_sched_build_groups will build a circular linked list of the groups
5674 * covered by the given span, and will set each group's ->cpumask correctly,
5675 * and ->cpu_power to 0.
5678 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5679 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5680 struct sched_group **sg))
5682 struct sched_group *first = NULL, *last = NULL;
5683 cpumask_t covered = CPU_MASK_NONE;
5686 for_each_cpu_mask(i, span) {
5687 struct sched_group *sg;
5688 int group = group_fn(i, cpu_map, &sg);
5691 if (cpu_isset(i, covered))
5694 sg->cpumask = CPU_MASK_NONE;
5695 sg->__cpu_power = 0;
5697 for_each_cpu_mask(j, span) {
5698 if (group_fn(j, cpu_map, NULL) != group)
5701 cpu_set(j, covered);
5702 cpu_set(j, sg->cpumask);
5713 #define SD_NODES_PER_DOMAIN 16
5718 * find_next_best_node - find the next node to include in a sched_domain
5719 * @node: node whose sched_domain we're building
5720 * @used_nodes: nodes already in the sched_domain
5722 * Find the next node to include in a given scheduling domain. Simply
5723 * finds the closest node not already in the @used_nodes map.
5725 * Should use nodemask_t.
5727 static int find_next_best_node(int node, unsigned long *used_nodes)
5729 int i, n, val, min_val, best_node = 0;
5733 for (i = 0; i < MAX_NUMNODES; i++) {
5734 /* Start at @node */
5735 n = (node + i) % MAX_NUMNODES;
5737 if (!nr_cpus_node(n))
5740 /* Skip already used nodes */
5741 if (test_bit(n, used_nodes))
5744 /* Simple min distance search */
5745 val = node_distance(node, n);
5747 if (val < min_val) {
5753 set_bit(best_node, used_nodes);
5758 * sched_domain_node_span - get a cpumask for a node's sched_domain
5759 * @node: node whose cpumask we're constructing
5760 * @size: number of nodes to include in this span
5762 * Given a node, construct a good cpumask for its sched_domain to span. It
5763 * should be one that prevents unnecessary balancing, but also spreads tasks
5766 static cpumask_t sched_domain_node_span(int node)
5768 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5769 cpumask_t span, nodemask;
5773 bitmap_zero(used_nodes, MAX_NUMNODES);
5775 nodemask = node_to_cpumask(node);
5776 cpus_or(span, span, nodemask);
5777 set_bit(node, used_nodes);
5779 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5780 int next_node = find_next_best_node(node, used_nodes);
5782 nodemask = node_to_cpumask(next_node);
5783 cpus_or(span, span, nodemask);
5790 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5793 * SMT sched-domains:
5795 #ifdef CONFIG_SCHED_SMT
5796 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5797 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5799 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5800 struct sched_group **sg)
5803 *sg = &per_cpu(sched_group_cpus, cpu);
5809 * multi-core sched-domains:
5811 #ifdef CONFIG_SCHED_MC
5812 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5813 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5816 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5817 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5818 struct sched_group **sg)
5821 cpumask_t mask = cpu_sibling_map[cpu];
5822 cpus_and(mask, mask, *cpu_map);
5823 group = first_cpu(mask);
5825 *sg = &per_cpu(sched_group_core, group);
5828 #elif defined(CONFIG_SCHED_MC)
5829 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5830 struct sched_group **sg)
5833 *sg = &per_cpu(sched_group_core, cpu);
5838 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5839 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5841 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5842 struct sched_group **sg)
5845 #ifdef CONFIG_SCHED_MC
5846 cpumask_t mask = cpu_coregroup_map(cpu);
5847 cpus_and(mask, mask, *cpu_map);
5848 group = first_cpu(mask);
5849 #elif defined(CONFIG_SCHED_SMT)
5850 cpumask_t mask = cpu_sibling_map[cpu];
5851 cpus_and(mask, mask, *cpu_map);
5852 group = first_cpu(mask);
5857 *sg = &per_cpu(sched_group_phys, group);
5863 * The init_sched_build_groups can't handle what we want to do with node
5864 * groups, so roll our own. Now each node has its own list of groups which
5865 * gets dynamically allocated.
5867 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5868 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5870 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5871 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5873 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5874 struct sched_group **sg)
5876 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5879 cpus_and(nodemask, nodemask, *cpu_map);
5880 group = first_cpu(nodemask);
5883 *sg = &per_cpu(sched_group_allnodes, group);
5887 static void init_numa_sched_groups_power(struct sched_group *group_head)
5889 struct sched_group *sg = group_head;
5895 for_each_cpu_mask(j, sg->cpumask) {
5896 struct sched_domain *sd;
5898 sd = &per_cpu(phys_domains, j);
5899 if (j != first_cpu(sd->groups->cpumask)) {
5901 * Only add "power" once for each
5907 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5910 if (sg != group_head)
5916 /* Free memory allocated for various sched_group structures */
5917 static void free_sched_groups(const cpumask_t *cpu_map)
5921 for_each_cpu_mask(cpu, *cpu_map) {
5922 struct sched_group **sched_group_nodes
5923 = sched_group_nodes_bycpu[cpu];
5925 if (!sched_group_nodes)
5928 for (i = 0; i < MAX_NUMNODES; i++) {
5929 cpumask_t nodemask = node_to_cpumask(i);
5930 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5932 cpus_and(nodemask, nodemask, *cpu_map);
5933 if (cpus_empty(nodemask))
5943 if (oldsg != sched_group_nodes[i])
5946 kfree(sched_group_nodes);
5947 sched_group_nodes_bycpu[cpu] = NULL;
5951 static void free_sched_groups(const cpumask_t *cpu_map)
5957 * Initialize sched groups cpu_power.
5959 * cpu_power indicates the capacity of sched group, which is used while
5960 * distributing the load between different sched groups in a sched domain.
5961 * Typically cpu_power for all the groups in a sched domain will be same unless
5962 * there are asymmetries in the topology. If there are asymmetries, group
5963 * having more cpu_power will pickup more load compared to the group having
5966 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5967 * the maximum number of tasks a group can handle in the presence of other idle
5968 * or lightly loaded groups in the same sched domain.
5970 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5972 struct sched_domain *child;
5973 struct sched_group *group;
5975 WARN_ON(!sd || !sd->groups);
5977 if (cpu != first_cpu(sd->groups->cpumask))
5982 sd->groups->__cpu_power = 0;
5985 * For perf policy, if the groups in child domain share resources
5986 * (for example cores sharing some portions of the cache hierarchy
5987 * or SMT), then set this domain groups cpu_power such that each group
5988 * can handle only one task, when there are other idle groups in the
5989 * same sched domain.
5991 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5993 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5994 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5999 * add cpu_power of each child group to this groups cpu_power
6001 group = child->groups;
6003 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6004 group = group->next;
6005 } while (group != child->groups);
6009 * Build sched domains for a given set of cpus and attach the sched domains
6010 * to the individual cpus
6012 static int build_sched_domains(const cpumask_t *cpu_map)
6016 struct sched_group **sched_group_nodes = NULL;
6017 int sd_allnodes = 0;
6020 * Allocate the per-node list of sched groups
6022 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
6024 if (!sched_group_nodes) {
6025 printk(KERN_WARNING "Can not alloc sched group node list\n");
6028 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6032 * Set up domains for cpus specified by the cpu_map.
6034 for_each_cpu_mask(i, *cpu_map) {
6035 struct sched_domain *sd = NULL, *p;
6036 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6038 cpus_and(nodemask, nodemask, *cpu_map);
6041 if (cpus_weight(*cpu_map) >
6042 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6043 sd = &per_cpu(allnodes_domains, i);
6044 *sd = SD_ALLNODES_INIT;
6045 sd->span = *cpu_map;
6046 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6052 sd = &per_cpu(node_domains, i);
6054 sd->span = sched_domain_node_span(cpu_to_node(i));
6058 cpus_and(sd->span, sd->span, *cpu_map);
6062 sd = &per_cpu(phys_domains, i);
6064 sd->span = nodemask;
6068 cpu_to_phys_group(i, cpu_map, &sd->groups);
6070 #ifdef CONFIG_SCHED_MC
6072 sd = &per_cpu(core_domains, i);
6074 sd->span = cpu_coregroup_map(i);
6075 cpus_and(sd->span, sd->span, *cpu_map);
6078 cpu_to_core_group(i, cpu_map, &sd->groups);
6081 #ifdef CONFIG_SCHED_SMT
6083 sd = &per_cpu(cpu_domains, i);
6084 *sd = SD_SIBLING_INIT;
6085 sd->span = cpu_sibling_map[i];
6086 cpus_and(sd->span, sd->span, *cpu_map);
6089 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6093 #ifdef CONFIG_SCHED_SMT
6094 /* Set up CPU (sibling) groups */
6095 for_each_cpu_mask(i, *cpu_map) {
6096 cpumask_t this_sibling_map = cpu_sibling_map[i];
6097 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6098 if (i != first_cpu(this_sibling_map))
6101 init_sched_build_groups(this_sibling_map, cpu_map,
6106 #ifdef CONFIG_SCHED_MC
6107 /* Set up multi-core groups */
6108 for_each_cpu_mask(i, *cpu_map) {
6109 cpumask_t this_core_map = cpu_coregroup_map(i);
6110 cpus_and(this_core_map, this_core_map, *cpu_map);
6111 if (i != first_cpu(this_core_map))
6113 init_sched_build_groups(this_core_map, cpu_map,
6114 &cpu_to_core_group);
6118 /* Set up physical groups */
6119 for (i = 0; i < MAX_NUMNODES; i++) {
6120 cpumask_t nodemask = node_to_cpumask(i);
6122 cpus_and(nodemask, nodemask, *cpu_map);
6123 if (cpus_empty(nodemask))
6126 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6130 /* Set up node groups */
6132 init_sched_build_groups(*cpu_map, cpu_map,
6133 &cpu_to_allnodes_group);
6135 for (i = 0; i < MAX_NUMNODES; i++) {
6136 /* Set up node groups */
6137 struct sched_group *sg, *prev;
6138 cpumask_t nodemask = node_to_cpumask(i);
6139 cpumask_t domainspan;
6140 cpumask_t covered = CPU_MASK_NONE;
6143 cpus_and(nodemask, nodemask, *cpu_map);
6144 if (cpus_empty(nodemask)) {
6145 sched_group_nodes[i] = NULL;
6149 domainspan = sched_domain_node_span(i);
6150 cpus_and(domainspan, domainspan, *cpu_map);
6152 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6154 printk(KERN_WARNING "Can not alloc domain group for "
6158 sched_group_nodes[i] = sg;
6159 for_each_cpu_mask(j, nodemask) {
6160 struct sched_domain *sd;
6162 sd = &per_cpu(node_domains, j);
6165 sg->__cpu_power = 0;
6166 sg->cpumask = nodemask;
6168 cpus_or(covered, covered, nodemask);
6171 for (j = 0; j < MAX_NUMNODES; j++) {
6172 cpumask_t tmp, notcovered;
6173 int n = (i + j) % MAX_NUMNODES;
6175 cpus_complement(notcovered, covered);
6176 cpus_and(tmp, notcovered, *cpu_map);
6177 cpus_and(tmp, tmp, domainspan);
6178 if (cpus_empty(tmp))
6181 nodemask = node_to_cpumask(n);
6182 cpus_and(tmp, tmp, nodemask);
6183 if (cpus_empty(tmp))
6186 sg = kmalloc_node(sizeof(struct sched_group),
6190 "Can not alloc domain group for node %d\n", j);
6193 sg->__cpu_power = 0;
6195 sg->next = prev->next;
6196 cpus_or(covered, covered, tmp);
6203 /* Calculate CPU power for physical packages and nodes */
6204 #ifdef CONFIG_SCHED_SMT
6205 for_each_cpu_mask(i, *cpu_map) {
6206 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6208 init_sched_groups_power(i, sd);
6211 #ifdef CONFIG_SCHED_MC
6212 for_each_cpu_mask(i, *cpu_map) {
6213 struct sched_domain *sd = &per_cpu(core_domains, i);
6215 init_sched_groups_power(i, sd);
6219 for_each_cpu_mask(i, *cpu_map) {
6220 struct sched_domain *sd = &per_cpu(phys_domains, i);
6222 init_sched_groups_power(i, sd);
6226 for (i = 0; i < MAX_NUMNODES; i++)
6227 init_numa_sched_groups_power(sched_group_nodes[i]);
6230 struct sched_group *sg;
6232 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6233 init_numa_sched_groups_power(sg);
6237 /* Attach the domains */
6238 for_each_cpu_mask(i, *cpu_map) {
6239 struct sched_domain *sd;
6240 #ifdef CONFIG_SCHED_SMT
6241 sd = &per_cpu(cpu_domains, i);
6242 #elif defined(CONFIG_SCHED_MC)
6243 sd = &per_cpu(core_domains, i);
6245 sd = &per_cpu(phys_domains, i);
6247 cpu_attach_domain(sd, i);
6254 free_sched_groups(cpu_map);
6259 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6261 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6263 cpumask_t cpu_default_map;
6267 * Setup mask for cpus without special case scheduling requirements.
6268 * For now this just excludes isolated cpus, but could be used to
6269 * exclude other special cases in the future.
6271 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6273 err = build_sched_domains(&cpu_default_map);
6278 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6280 free_sched_groups(cpu_map);
6284 * Detach sched domains from a group of cpus specified in cpu_map
6285 * These cpus will now be attached to the NULL domain
6287 static void detach_destroy_domains(const cpumask_t *cpu_map)
6291 for_each_cpu_mask(i, *cpu_map)
6292 cpu_attach_domain(NULL, i);
6293 synchronize_sched();
6294 arch_destroy_sched_domains(cpu_map);
6298 * Partition sched domains as specified by the cpumasks below.
6299 * This attaches all cpus from the cpumasks to the NULL domain,
6300 * waits for a RCU quiescent period, recalculates sched
6301 * domain information and then attaches them back to the
6302 * correct sched domains
6303 * Call with hotplug lock held
6305 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6307 cpumask_t change_map;
6310 cpus_and(*partition1, *partition1, cpu_online_map);
6311 cpus_and(*partition2, *partition2, cpu_online_map);
6312 cpus_or(change_map, *partition1, *partition2);
6314 /* Detach sched domains from all of the affected cpus */
6315 detach_destroy_domains(&change_map);
6316 if (!cpus_empty(*partition1))
6317 err = build_sched_domains(partition1);
6318 if (!err && !cpus_empty(*partition2))
6319 err = build_sched_domains(partition2);
6324 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6325 int arch_reinit_sched_domains(void)
6329 mutex_lock(&sched_hotcpu_mutex);
6330 detach_destroy_domains(&cpu_online_map);
6331 err = arch_init_sched_domains(&cpu_online_map);
6332 mutex_unlock(&sched_hotcpu_mutex);
6337 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6341 if (buf[0] != '0' && buf[0] != '1')
6345 sched_smt_power_savings = (buf[0] == '1');
6347 sched_mc_power_savings = (buf[0] == '1');
6349 ret = arch_reinit_sched_domains();
6351 return ret ? ret : count;
6354 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6358 #ifdef CONFIG_SCHED_SMT
6360 err = sysfs_create_file(&cls->kset.kobj,
6361 &attr_sched_smt_power_savings.attr);
6363 #ifdef CONFIG_SCHED_MC
6364 if (!err && mc_capable())
6365 err = sysfs_create_file(&cls->kset.kobj,
6366 &attr_sched_mc_power_savings.attr);
6372 #ifdef CONFIG_SCHED_MC
6373 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6375 return sprintf(page, "%u\n", sched_mc_power_savings);
6377 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6378 const char *buf, size_t count)
6380 return sched_power_savings_store(buf, count, 0);
6382 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6383 sched_mc_power_savings_store);
6386 #ifdef CONFIG_SCHED_SMT
6387 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6389 return sprintf(page, "%u\n", sched_smt_power_savings);
6391 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6392 const char *buf, size_t count)
6394 return sched_power_savings_store(buf, count, 1);
6396 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6397 sched_smt_power_savings_store);
6401 * Force a reinitialization of the sched domains hierarchy. The domains
6402 * and groups cannot be updated in place without racing with the balancing
6403 * code, so we temporarily attach all running cpus to the NULL domain
6404 * which will prevent rebalancing while the sched domains are recalculated.
6406 static int update_sched_domains(struct notifier_block *nfb,
6407 unsigned long action, void *hcpu)
6410 case CPU_UP_PREPARE:
6411 case CPU_UP_PREPARE_FROZEN:
6412 case CPU_DOWN_PREPARE:
6413 case CPU_DOWN_PREPARE_FROZEN:
6414 detach_destroy_domains(&cpu_online_map);
6417 case CPU_UP_CANCELED:
6418 case CPU_UP_CANCELED_FROZEN:
6419 case CPU_DOWN_FAILED:
6420 case CPU_DOWN_FAILED_FROZEN:
6422 case CPU_ONLINE_FROZEN:
6424 case CPU_DEAD_FROZEN:
6426 * Fall through and re-initialise the domains.
6433 /* The hotplug lock is already held by cpu_up/cpu_down */
6434 arch_init_sched_domains(&cpu_online_map);
6439 void __init sched_init_smp(void)
6441 cpumask_t non_isolated_cpus;
6443 mutex_lock(&sched_hotcpu_mutex);
6444 arch_init_sched_domains(&cpu_online_map);
6445 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6446 if (cpus_empty(non_isolated_cpus))
6447 cpu_set(smp_processor_id(), non_isolated_cpus);
6448 mutex_unlock(&sched_hotcpu_mutex);
6449 /* XXX: Theoretical race here - CPU may be hotplugged now */
6450 hotcpu_notifier(update_sched_domains, 0);
6452 init_sched_domain_sysctl();
6454 /* Move init over to a non-isolated CPU */
6455 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6457 sched_init_granularity();
6460 void __init sched_init_smp(void)
6462 sched_init_granularity();
6464 #endif /* CONFIG_SMP */
6466 int in_sched_functions(unsigned long addr)
6468 /* Linker adds these: start and end of __sched functions */
6469 extern char __sched_text_start[], __sched_text_end[];
6471 return in_lock_functions(addr) ||
6472 (addr >= (unsigned long)__sched_text_start
6473 && addr < (unsigned long)__sched_text_end);
6476 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6478 cfs_rq->tasks_timeline = RB_ROOT;
6479 cfs_rq->fair_clock = 1;
6480 #ifdef CONFIG_FAIR_GROUP_SCHED
6485 void __init sched_init(void)
6487 u64 now = sched_clock();
6488 int highest_cpu = 0;
6492 * Link up the scheduling class hierarchy:
6494 rt_sched_class.next = &fair_sched_class;
6495 fair_sched_class.next = &idle_sched_class;
6496 idle_sched_class.next = NULL;
6498 for_each_possible_cpu(i) {
6499 struct rt_prio_array *array;
6503 spin_lock_init(&rq->lock);
6504 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6507 init_cfs_rq(&rq->cfs, rq);
6508 #ifdef CONFIG_FAIR_GROUP_SCHED
6509 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6510 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6512 rq->ls.load_update_last = now;
6513 rq->ls.load_update_start = now;
6515 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6516 rq->cpu_load[j] = 0;
6519 rq->active_balance = 0;
6520 rq->next_balance = jiffies;
6523 rq->migration_thread = NULL;
6524 INIT_LIST_HEAD(&rq->migration_queue);
6526 atomic_set(&rq->nr_iowait, 0);
6528 array = &rq->rt.active;
6529 for (j = 0; j < MAX_RT_PRIO; j++) {
6530 INIT_LIST_HEAD(array->queue + j);
6531 __clear_bit(j, array->bitmap);
6534 /* delimiter for bitsearch: */
6535 __set_bit(MAX_RT_PRIO, array->bitmap);
6538 set_load_weight(&init_task);
6540 #ifdef CONFIG_PREEMPT_NOTIFIERS
6541 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6545 nr_cpu_ids = highest_cpu + 1;
6546 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6549 #ifdef CONFIG_RT_MUTEXES
6550 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6554 * The boot idle thread does lazy MMU switching as well:
6556 atomic_inc(&init_mm.mm_count);
6557 enter_lazy_tlb(&init_mm, current);
6560 * Make us the idle thread. Technically, schedule() should not be
6561 * called from this thread, however somewhere below it might be,
6562 * but because we are the idle thread, we just pick up running again
6563 * when this runqueue becomes "idle".
6565 init_idle(current, smp_processor_id());
6567 * During early bootup we pretend to be a normal task:
6569 current->sched_class = &fair_sched_class;
6572 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6573 void __might_sleep(char *file, int line)
6576 static unsigned long prev_jiffy; /* ratelimiting */
6578 if ((in_atomic() || irqs_disabled()) &&
6579 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6580 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6582 prev_jiffy = jiffies;
6583 printk(KERN_ERR "BUG: sleeping function called from invalid"
6584 " context at %s:%d\n", file, line);
6585 printk("in_atomic():%d, irqs_disabled():%d\n",
6586 in_atomic(), irqs_disabled());
6587 debug_show_held_locks(current);
6588 if (irqs_disabled())
6589 print_irqtrace_events(current);
6594 EXPORT_SYMBOL(__might_sleep);
6597 #ifdef CONFIG_MAGIC_SYSRQ
6598 void normalize_rt_tasks(void)
6600 struct task_struct *g, *p;
6601 unsigned long flags;
6605 read_lock_irq(&tasklist_lock);
6606 do_each_thread(g, p) {
6608 p->se.wait_runtime = 0;
6609 p->se.exec_start = 0;
6610 p->se.wait_start_fair = 0;
6611 p->se.sleep_start_fair = 0;
6612 #ifdef CONFIG_SCHEDSTATS
6613 p->se.wait_start = 0;
6614 p->se.sleep_start = 0;
6615 p->se.block_start = 0;
6617 task_rq(p)->cfs.fair_clock = 0;
6618 task_rq(p)->clock = 0;
6622 * Renice negative nice level userspace
6625 if (TASK_NICE(p) < 0 && p->mm)
6626 set_user_nice(p, 0);
6630 spin_lock_irqsave(&p->pi_lock, flags);
6631 rq = __task_rq_lock(p);
6634 * Do not touch the migration thread:
6636 if (p == rq->migration_thread)
6640 on_rq = p->se.on_rq;
6642 deactivate_task(task_rq(p), p, 0, rq_clock(task_rq(p)));
6643 __setscheduler(rq, p, SCHED_NORMAL, 0);
6645 activate_task(task_rq(p), p, 0);
6646 resched_task(rq->curr);
6651 __task_rq_unlock(rq);
6652 spin_unlock_irqrestore(&p->pi_lock, flags);
6653 } while_each_thread(g, p);
6655 read_unlock_irq(&tasklist_lock);
6658 #endif /* CONFIG_MAGIC_SYSRQ */
6662 * These functions are only useful for the IA64 MCA handling.
6664 * They can only be called when the whole system has been
6665 * stopped - every CPU needs to be quiescent, and no scheduling
6666 * activity can take place. Using them for anything else would
6667 * be a serious bug, and as a result, they aren't even visible
6668 * under any other configuration.
6672 * curr_task - return the current task for a given cpu.
6673 * @cpu: the processor in question.
6675 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6677 struct task_struct *curr_task(int cpu)
6679 return cpu_curr(cpu);
6683 * set_curr_task - set the current task for a given cpu.
6684 * @cpu: the processor in question.
6685 * @p: the task pointer to set.
6687 * Description: This function must only be used when non-maskable interrupts
6688 * are serviced on a separate stack. It allows the architecture to switch the
6689 * notion of the current task on a cpu in a non-blocking manner. This function
6690 * must be called with all CPU's synchronized, and interrupts disabled, the
6691 * and caller must save the original value of the current task (see
6692 * curr_task() above) and restore that value before reenabling interrupts and
6693 * re-starting the system.
6695 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6697 void set_curr_task(int cpu, struct task_struct *p)