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 * Per-runqueue clock, as finegrained as the platform can give us:
323 static unsigned long long __rq_clock(struct rq *rq)
325 u64 prev_raw = rq->prev_clock_raw;
326 u64 now = sched_clock();
327 s64 delta = now - prev_raw;
328 u64 clock = rq->clock;
331 * Protect against sched_clock() occasionally going backwards:
333 if (unlikely(delta < 0)) {
338 * Catch too large forward jumps too:
340 if (unlikely(delta > 2*TICK_NSEC)) {
342 rq->clock_overflows++;
344 if (unlikely(delta > rq->clock_max_delta))
345 rq->clock_max_delta = delta;
350 rq->prev_clock_raw = now;
356 static inline unsigned long long rq_clock(struct rq *rq)
358 int this_cpu = smp_processor_id();
360 if (this_cpu == cpu_of(rq))
361 return __rq_clock(rq);
367 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
368 * See detach_destroy_domains: synchronize_sched for details.
370 * The domain tree of any CPU may only be accessed from within
371 * preempt-disabled sections.
373 #define for_each_domain(cpu, __sd) \
374 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
376 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
377 #define this_rq() (&__get_cpu_var(runqueues))
378 #define task_rq(p) cpu_rq(task_cpu(p))
379 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
382 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
383 * clock constructed from sched_clock():
385 unsigned long long cpu_clock(int cpu)
387 unsigned long long now;
390 local_irq_save(flags);
391 now = rq_clock(cpu_rq(cpu));
392 local_irq_restore(flags);
397 #ifdef CONFIG_FAIR_GROUP_SCHED
398 /* Change a task's ->cfs_rq if it moves across CPUs */
399 static inline void set_task_cfs_rq(struct task_struct *p)
401 p->se.cfs_rq = &task_rq(p)->cfs;
404 static inline void set_task_cfs_rq(struct task_struct *p)
409 #ifndef prepare_arch_switch
410 # define prepare_arch_switch(next) do { } while (0)
412 #ifndef finish_arch_switch
413 # define finish_arch_switch(prev) do { } while (0)
416 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
417 static inline int task_running(struct rq *rq, struct task_struct *p)
419 return rq->curr == p;
422 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
426 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
428 #ifdef CONFIG_DEBUG_SPINLOCK
429 /* this is a valid case when another task releases the spinlock */
430 rq->lock.owner = current;
433 * If we are tracking spinlock dependencies then we have to
434 * fix up the runqueue lock - which gets 'carried over' from
437 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
439 spin_unlock_irq(&rq->lock);
442 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
443 static inline int task_running(struct rq *rq, struct task_struct *p)
448 return rq->curr == p;
452 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
456 * We can optimise this out completely for !SMP, because the
457 * SMP rebalancing from interrupt is the only thing that cares
462 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
463 spin_unlock_irq(&rq->lock);
465 spin_unlock(&rq->lock);
469 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
473 * After ->oncpu is cleared, the task can be moved to a different CPU.
474 * We must ensure this doesn't happen until the switch is completely
480 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
484 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
487 * __task_rq_lock - lock the runqueue a given task resides on.
488 * Must be called interrupts disabled.
490 static inline struct rq *__task_rq_lock(struct task_struct *p)
497 spin_lock(&rq->lock);
498 if (unlikely(rq != task_rq(p))) {
499 spin_unlock(&rq->lock);
500 goto repeat_lock_task;
506 * task_rq_lock - lock the runqueue a given task resides on and disable
507 * interrupts. Note the ordering: we can safely lookup the task_rq without
508 * explicitly disabling preemption.
510 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
516 local_irq_save(*flags);
518 spin_lock(&rq->lock);
519 if (unlikely(rq != task_rq(p))) {
520 spin_unlock_irqrestore(&rq->lock, *flags);
521 goto repeat_lock_task;
526 static inline void __task_rq_unlock(struct rq *rq)
529 spin_unlock(&rq->lock);
532 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
535 spin_unlock_irqrestore(&rq->lock, *flags);
539 * this_rq_lock - lock this runqueue and disable interrupts.
541 static inline struct rq *this_rq_lock(void)
548 spin_lock(&rq->lock);
554 * CPU frequency is/was unstable - start new by setting prev_clock_raw:
556 void sched_clock_unstable_event(void)
561 rq = task_rq_lock(current, &flags);
562 rq->prev_clock_raw = sched_clock();
563 rq->clock_unstable_events++;
564 task_rq_unlock(rq, &flags);
568 * resched_task - mark a task 'to be rescheduled now'.
570 * On UP this means the setting of the need_resched flag, on SMP it
571 * might also involve a cross-CPU call to trigger the scheduler on
576 #ifndef tsk_is_polling
577 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
580 static void resched_task(struct task_struct *p)
584 assert_spin_locked(&task_rq(p)->lock);
586 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
589 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
592 if (cpu == smp_processor_id())
595 /* NEED_RESCHED must be visible before we test polling */
597 if (!tsk_is_polling(p))
598 smp_send_reschedule(cpu);
601 static void resched_cpu(int cpu)
603 struct rq *rq = cpu_rq(cpu);
606 if (!spin_trylock_irqsave(&rq->lock, flags))
608 resched_task(cpu_curr(cpu));
609 spin_unlock_irqrestore(&rq->lock, flags);
612 static inline void resched_task(struct task_struct *p)
614 assert_spin_locked(&task_rq(p)->lock);
615 set_tsk_need_resched(p);
619 static u64 div64_likely32(u64 divident, unsigned long divisor)
621 #if BITS_PER_LONG == 32
622 if (likely(divident <= 0xffffffffULL))
623 return (u32)divident / divisor;
624 do_div(divident, divisor);
628 return divident / divisor;
632 #if BITS_PER_LONG == 32
633 # define WMULT_CONST (~0UL)
635 # define WMULT_CONST (1UL << 32)
638 #define WMULT_SHIFT 32
641 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
642 struct load_weight *lw)
646 if (unlikely(!lw->inv_weight))
647 lw->inv_weight = WMULT_CONST / lw->weight;
649 tmp = (u64)delta_exec * weight;
651 * Check whether we'd overflow the 64-bit multiplication:
653 if (unlikely(tmp > WMULT_CONST)) {
654 tmp = ((tmp >> WMULT_SHIFT/2) * lw->inv_weight)
657 tmp = (tmp * lw->inv_weight) >> WMULT_SHIFT;
660 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
663 static inline unsigned long
664 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
666 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
669 static void update_load_add(struct load_weight *lw, unsigned long inc)
675 static void update_load_sub(struct load_weight *lw, unsigned long dec)
681 static void __update_curr_load(struct rq *rq, struct load_stat *ls)
683 if (rq->curr != rq->idle && ls->load.weight) {
684 ls->delta_exec += ls->delta_stat;
685 ls->delta_fair += calc_delta_fair(ls->delta_stat, &ls->load);
691 * Update delta_exec, delta_fair fields for rq.
693 * delta_fair clock advances at a rate inversely proportional to
694 * total load (rq->ls.load.weight) on the runqueue, while
695 * delta_exec advances at the same rate as wall-clock (provided
698 * delta_exec / delta_fair is a measure of the (smoothened) load on this
699 * runqueue over any given interval. This (smoothened) load is used
700 * during load balance.
702 * This function is called /before/ updating rq->ls.load
703 * and when switching tasks.
705 static void update_curr_load(struct rq *rq, u64 now)
707 struct load_stat *ls = &rq->ls;
710 start = ls->load_update_start;
711 ls->load_update_start = now;
712 ls->delta_stat += now - start;
714 * Stagger updates to ls->delta_fair. Very frequent updates
717 if (ls->delta_stat >= sysctl_sched_stat_granularity)
718 __update_curr_load(rq, ls);
722 * To aid in avoiding the subversion of "niceness" due to uneven distribution
723 * of tasks with abnormal "nice" values across CPUs the contribution that
724 * each task makes to its run queue's load is weighted according to its
725 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
726 * scaled version of the new time slice allocation that they receive on time
730 #define WEIGHT_IDLEPRIO 2
731 #define WMULT_IDLEPRIO (1 << 31)
734 * Nice levels are multiplicative, with a gentle 10% change for every
735 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
736 * nice 1, it will get ~10% less CPU time than another CPU-bound task
737 * that remained on nice 0.
739 * The "10% effect" is relative and cumulative: from _any_ nice level,
740 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
741 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
742 * If a task goes up by ~10% and another task goes down by ~10% then
743 * the relative distance between them is ~25%.)
745 static const int prio_to_weight[40] = {
746 /* -20 */ 88818, 71054, 56843, 45475, 36380, 29104, 23283, 18626, 14901, 11921,
747 /* -10 */ 9537, 7629, 6103, 4883, 3906, 3125, 2500, 2000, 1600, 1280,
748 /* 0 */ NICE_0_LOAD /* 1024 */,
749 /* 1 */ 819, 655, 524, 419, 336, 268, 215, 172, 137,
750 /* 10 */ 110, 87, 70, 56, 45, 36, 29, 23, 18, 15,
754 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
756 * In cases where the weight does not change often, we can use the
757 * precalculated inverse to speed up arithmetics by turning divisions
758 * into multiplications:
760 static const u32 prio_to_wmult[40] = {
761 /* -20 */ 48356, 60446, 75558, 94446, 118058,
762 /* -15 */ 147573, 184467, 230589, 288233, 360285,
763 /* -10 */ 450347, 562979, 703746, 879575, 1099582,
764 /* -5 */ 1374389, 1717986, 2147483, 2684354, 3355443,
765 /* 0 */ 4194304, 5244160, 6557201, 8196502, 10250518,
766 /* 5 */ 12782640, 16025997, 19976592, 24970740, 31350126,
767 /* 10 */ 39045157, 49367440, 61356675, 76695844, 95443717,
768 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
772 inc_load(struct rq *rq, const struct task_struct *p, u64 now)
774 update_curr_load(rq, now);
775 update_load_add(&rq->ls.load, p->se.load.weight);
779 dec_load(struct rq *rq, const struct task_struct *p, u64 now)
781 update_curr_load(rq, now);
782 update_load_sub(&rq->ls.load, p->se.load.weight);
785 static void inc_nr_running(struct task_struct *p, struct rq *rq, u64 now)
788 inc_load(rq, p, now);
791 static void dec_nr_running(struct task_struct *p, struct rq *rq, u64 now)
794 dec_load(rq, p, now);
797 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
800 * runqueue iterator, to support SMP load-balancing between different
801 * scheduling classes, without having to expose their internal data
802 * structures to the load-balancing proper:
806 struct task_struct *(*start)(void *);
807 struct task_struct *(*next)(void *);
810 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
811 unsigned long max_nr_move, unsigned long max_load_move,
812 struct sched_domain *sd, enum cpu_idle_type idle,
813 int *all_pinned, unsigned long *load_moved,
814 int this_best_prio, int best_prio, int best_prio_seen,
815 struct rq_iterator *iterator);
817 #include "sched_stats.h"
818 #include "sched_rt.c"
819 #include "sched_fair.c"
820 #include "sched_idletask.c"
821 #ifdef CONFIG_SCHED_DEBUG
822 # include "sched_debug.c"
825 #define sched_class_highest (&rt_sched_class)
827 static void set_load_weight(struct task_struct *p)
829 task_rq(p)->cfs.wait_runtime -= p->se.wait_runtime;
830 p->se.wait_runtime = 0;
832 if (task_has_rt_policy(p)) {
833 p->se.load.weight = prio_to_weight[0] * 2;
834 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
839 * SCHED_IDLE tasks get minimal weight:
841 if (p->policy == SCHED_IDLE) {
842 p->se.load.weight = WEIGHT_IDLEPRIO;
843 p->se.load.inv_weight = WMULT_IDLEPRIO;
847 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
848 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
852 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, u64 now)
854 sched_info_queued(p);
855 p->sched_class->enqueue_task(rq, p, wakeup, now);
860 dequeue_task(struct rq *rq, struct task_struct *p, int sleep, u64 now)
862 p->sched_class->dequeue_task(rq, p, sleep, now);
867 * __normal_prio - return the priority that is based on the static prio
869 static inline int __normal_prio(struct task_struct *p)
871 return p->static_prio;
875 * Calculate the expected normal priority: i.e. priority
876 * without taking RT-inheritance into account. Might be
877 * boosted by interactivity modifiers. Changes upon fork,
878 * setprio syscalls, and whenever the interactivity
879 * estimator recalculates.
881 static inline int normal_prio(struct task_struct *p)
885 if (task_has_rt_policy(p))
886 prio = MAX_RT_PRIO-1 - p->rt_priority;
888 prio = __normal_prio(p);
893 * Calculate the current priority, i.e. the priority
894 * taken into account by the scheduler. This value might
895 * be boosted by RT tasks, or might be boosted by
896 * interactivity modifiers. Will be RT if the task got
897 * RT-boosted. If not then it returns p->normal_prio.
899 static int effective_prio(struct task_struct *p)
901 p->normal_prio = normal_prio(p);
903 * If we are RT tasks or we were boosted to RT priority,
904 * keep the priority unchanged. Otherwise, update priority
905 * to the normal priority:
907 if (!rt_prio(p->prio))
908 return p->normal_prio;
913 * activate_task - move a task to the runqueue.
915 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
917 u64 now = rq_clock(rq);
919 if (p->state == TASK_UNINTERRUPTIBLE)
920 rq->nr_uninterruptible--;
922 enqueue_task(rq, p, wakeup, now);
923 inc_nr_running(p, rq, now);
927 * activate_idle_task - move idle task to the _front_ of runqueue.
929 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
931 u64 now = rq_clock(rq);
933 if (p->state == TASK_UNINTERRUPTIBLE)
934 rq->nr_uninterruptible--;
936 enqueue_task(rq, p, 0, now);
937 inc_nr_running(p, rq, now);
941 * deactivate_task - remove a task from the runqueue.
943 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
945 u64 now = rq_clock(rq);
947 if (p->state == TASK_UNINTERRUPTIBLE)
948 rq->nr_uninterruptible++;
950 dequeue_task(rq, p, sleep, now);
951 dec_nr_running(p, rq, now);
955 * task_curr - is this task currently executing on a CPU?
956 * @p: the task in question.
958 inline int task_curr(const struct task_struct *p)
960 return cpu_curr(task_cpu(p)) == p;
963 /* Used instead of source_load when we know the type == 0 */
964 unsigned long weighted_cpuload(const int cpu)
966 return cpu_rq(cpu)->ls.load.weight;
969 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
972 task_thread_info(p)->cpu = cpu;
979 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
981 int old_cpu = task_cpu(p);
982 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
983 u64 clock_offset, fair_clock_offset;
985 clock_offset = old_rq->clock - new_rq->clock;
986 fair_clock_offset = old_rq->cfs.fair_clock -
987 new_rq->cfs.fair_clock;
988 if (p->se.wait_start)
989 p->se.wait_start -= clock_offset;
990 if (p->se.wait_start_fair)
991 p->se.wait_start_fair -= fair_clock_offset;
992 if (p->se.sleep_start)
993 p->se.sleep_start -= clock_offset;
994 if (p->se.block_start)
995 p->se.block_start -= clock_offset;
996 if (p->se.sleep_start_fair)
997 p->se.sleep_start_fair -= fair_clock_offset;
999 __set_task_cpu(p, new_cpu);
1002 struct migration_req {
1003 struct list_head list;
1005 struct task_struct *task;
1008 struct completion done;
1012 * The task's runqueue lock must be held.
1013 * Returns true if you have to wait for migration thread.
1016 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1018 struct rq *rq = task_rq(p);
1021 * If the task is not on a runqueue (and not running), then
1022 * it is sufficient to simply update the task's cpu field.
1024 if (!p->se.on_rq && !task_running(rq, p)) {
1025 set_task_cpu(p, dest_cpu);
1029 init_completion(&req->done);
1031 req->dest_cpu = dest_cpu;
1032 list_add(&req->list, &rq->migration_queue);
1038 * wait_task_inactive - wait for a thread to unschedule.
1040 * The caller must ensure that the task *will* unschedule sometime soon,
1041 * else this function might spin for a *long* time. This function can't
1042 * be called with interrupts off, or it may introduce deadlock with
1043 * smp_call_function() if an IPI is sent by the same process we are
1044 * waiting to become inactive.
1046 void wait_task_inactive(struct task_struct *p)
1048 unsigned long flags;
1054 * We do the initial early heuristics without holding
1055 * any task-queue locks at all. We'll only try to get
1056 * the runqueue lock when things look like they will
1062 * If the task is actively running on another CPU
1063 * still, just relax and busy-wait without holding
1066 * NOTE! Since we don't hold any locks, it's not
1067 * even sure that "rq" stays as the right runqueue!
1068 * But we don't care, since "task_running()" will
1069 * return false if the runqueue has changed and p
1070 * is actually now running somewhere else!
1072 while (task_running(rq, p))
1076 * Ok, time to look more closely! We need the rq
1077 * lock now, to be *sure*. If we're wrong, we'll
1078 * just go back and repeat.
1080 rq = task_rq_lock(p, &flags);
1081 running = task_running(rq, p);
1082 on_rq = p->se.on_rq;
1083 task_rq_unlock(rq, &flags);
1086 * Was it really running after all now that we
1087 * checked with the proper locks actually held?
1089 * Oops. Go back and try again..
1091 if (unlikely(running)) {
1097 * It's not enough that it's not actively running,
1098 * it must be off the runqueue _entirely_, and not
1101 * So if it wa still runnable (but just not actively
1102 * running right now), it's preempted, and we should
1103 * yield - it could be a while.
1105 if (unlikely(on_rq)) {
1111 * Ahh, all good. It wasn't running, and it wasn't
1112 * runnable, which means that it will never become
1113 * running in the future either. We're all done!
1118 * kick_process - kick a running thread to enter/exit the kernel
1119 * @p: the to-be-kicked thread
1121 * Cause a process which is running on another CPU to enter
1122 * kernel-mode, without any delay. (to get signals handled.)
1124 * NOTE: this function doesnt have to take the runqueue lock,
1125 * because all it wants to ensure is that the remote task enters
1126 * the kernel. If the IPI races and the task has been migrated
1127 * to another CPU then no harm is done and the purpose has been
1130 void kick_process(struct task_struct *p)
1136 if ((cpu != smp_processor_id()) && task_curr(p))
1137 smp_send_reschedule(cpu);
1142 * Return a low guess at the load of a migration-source cpu weighted
1143 * according to the scheduling class and "nice" value.
1145 * We want to under-estimate the load of migration sources, to
1146 * balance conservatively.
1148 static inline unsigned long source_load(int cpu, int type)
1150 struct rq *rq = cpu_rq(cpu);
1151 unsigned long total = weighted_cpuload(cpu);
1156 return min(rq->cpu_load[type-1], total);
1160 * Return a high guess at the load of a migration-target cpu weighted
1161 * according to the scheduling class and "nice" value.
1163 static inline unsigned long target_load(int cpu, int type)
1165 struct rq *rq = cpu_rq(cpu);
1166 unsigned long total = weighted_cpuload(cpu);
1171 return max(rq->cpu_load[type-1], total);
1175 * Return the average load per task on the cpu's run queue
1177 static inline unsigned long cpu_avg_load_per_task(int cpu)
1179 struct rq *rq = cpu_rq(cpu);
1180 unsigned long total = weighted_cpuload(cpu);
1181 unsigned long n = rq->nr_running;
1183 return n ? total / n : SCHED_LOAD_SCALE;
1187 * find_idlest_group finds and returns the least busy CPU group within the
1190 static struct sched_group *
1191 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1193 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1194 unsigned long min_load = ULONG_MAX, this_load = 0;
1195 int load_idx = sd->forkexec_idx;
1196 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1199 unsigned long load, avg_load;
1203 /* Skip over this group if it has no CPUs allowed */
1204 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1207 local_group = cpu_isset(this_cpu, group->cpumask);
1209 /* Tally up the load of all CPUs in the group */
1212 for_each_cpu_mask(i, group->cpumask) {
1213 /* Bias balancing toward cpus of our domain */
1215 load = source_load(i, load_idx);
1217 load = target_load(i, load_idx);
1222 /* Adjust by relative CPU power of the group */
1223 avg_load = sg_div_cpu_power(group,
1224 avg_load * SCHED_LOAD_SCALE);
1227 this_load = avg_load;
1229 } else if (avg_load < min_load) {
1230 min_load = avg_load;
1234 group = group->next;
1235 } while (group != sd->groups);
1237 if (!idlest || 100*this_load < imbalance*min_load)
1243 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1246 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1249 unsigned long load, min_load = ULONG_MAX;
1253 /* Traverse only the allowed CPUs */
1254 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1256 for_each_cpu_mask(i, tmp) {
1257 load = weighted_cpuload(i);
1259 if (load < min_load || (load == min_load && i == this_cpu)) {
1269 * sched_balance_self: balance the current task (running on cpu) in domains
1270 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1273 * Balance, ie. select the least loaded group.
1275 * Returns the target CPU number, or the same CPU if no balancing is needed.
1277 * preempt must be disabled.
1279 static int sched_balance_self(int cpu, int flag)
1281 struct task_struct *t = current;
1282 struct sched_domain *tmp, *sd = NULL;
1284 for_each_domain(cpu, tmp) {
1286 * If power savings logic is enabled for a domain, stop there.
1288 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1290 if (tmp->flags & flag)
1296 struct sched_group *group;
1297 int new_cpu, weight;
1299 if (!(sd->flags & flag)) {
1305 group = find_idlest_group(sd, t, cpu);
1311 new_cpu = find_idlest_cpu(group, t, cpu);
1312 if (new_cpu == -1 || new_cpu == cpu) {
1313 /* Now try balancing at a lower domain level of cpu */
1318 /* Now try balancing at a lower domain level of new_cpu */
1321 weight = cpus_weight(span);
1322 for_each_domain(cpu, tmp) {
1323 if (weight <= cpus_weight(tmp->span))
1325 if (tmp->flags & flag)
1328 /* while loop will break here if sd == NULL */
1334 #endif /* CONFIG_SMP */
1337 * wake_idle() will wake a task on an idle cpu if task->cpu is
1338 * not idle and an idle cpu is available. The span of cpus to
1339 * search starts with cpus closest then further out as needed,
1340 * so we always favor a closer, idle cpu.
1342 * Returns the CPU we should wake onto.
1344 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1345 static int wake_idle(int cpu, struct task_struct *p)
1348 struct sched_domain *sd;
1352 * If it is idle, then it is the best cpu to run this task.
1354 * This cpu is also the best, if it has more than one task already.
1355 * Siblings must be also busy(in most cases) as they didn't already
1356 * pickup the extra load from this cpu and hence we need not check
1357 * sibling runqueue info. This will avoid the checks and cache miss
1358 * penalities associated with that.
1360 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1363 for_each_domain(cpu, sd) {
1364 if (sd->flags & SD_WAKE_IDLE) {
1365 cpus_and(tmp, sd->span, p->cpus_allowed);
1366 for_each_cpu_mask(i, tmp) {
1377 static inline int wake_idle(int cpu, struct task_struct *p)
1384 * try_to_wake_up - wake up a thread
1385 * @p: the to-be-woken-up thread
1386 * @state: the mask of task states that can be woken
1387 * @sync: do a synchronous wakeup?
1389 * Put it on the run-queue if it's not already there. The "current"
1390 * thread is always on the run-queue (except when the actual
1391 * re-schedule is in progress), and as such you're allowed to do
1392 * the simpler "current->state = TASK_RUNNING" to mark yourself
1393 * runnable without the overhead of this.
1395 * returns failure only if the task is already active.
1397 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1399 int cpu, this_cpu, success = 0;
1400 unsigned long flags;
1404 struct sched_domain *sd, *this_sd = NULL;
1405 unsigned long load, this_load;
1409 rq = task_rq_lock(p, &flags);
1410 old_state = p->state;
1411 if (!(old_state & state))
1418 this_cpu = smp_processor_id();
1421 if (unlikely(task_running(rq, p)))
1426 schedstat_inc(rq, ttwu_cnt);
1427 if (cpu == this_cpu) {
1428 schedstat_inc(rq, ttwu_local);
1432 for_each_domain(this_cpu, sd) {
1433 if (cpu_isset(cpu, sd->span)) {
1434 schedstat_inc(sd, ttwu_wake_remote);
1440 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1444 * Check for affine wakeup and passive balancing possibilities.
1447 int idx = this_sd->wake_idx;
1448 unsigned int imbalance;
1450 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1452 load = source_load(cpu, idx);
1453 this_load = target_load(this_cpu, idx);
1455 new_cpu = this_cpu; /* Wake to this CPU if we can */
1457 if (this_sd->flags & SD_WAKE_AFFINE) {
1458 unsigned long tl = this_load;
1459 unsigned long tl_per_task;
1461 tl_per_task = cpu_avg_load_per_task(this_cpu);
1464 * If sync wakeup then subtract the (maximum possible)
1465 * effect of the currently running task from the load
1466 * of the current CPU:
1469 tl -= current->se.load.weight;
1472 tl + target_load(cpu, idx) <= tl_per_task) ||
1473 100*(tl + p->se.load.weight) <= imbalance*load) {
1475 * This domain has SD_WAKE_AFFINE and
1476 * p is cache cold in this domain, and
1477 * there is no bad imbalance.
1479 schedstat_inc(this_sd, ttwu_move_affine);
1485 * Start passive balancing when half the imbalance_pct
1488 if (this_sd->flags & SD_WAKE_BALANCE) {
1489 if (imbalance*this_load <= 100*load) {
1490 schedstat_inc(this_sd, ttwu_move_balance);
1496 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1498 new_cpu = wake_idle(new_cpu, p);
1499 if (new_cpu != cpu) {
1500 set_task_cpu(p, new_cpu);
1501 task_rq_unlock(rq, &flags);
1502 /* might preempt at this point */
1503 rq = task_rq_lock(p, &flags);
1504 old_state = p->state;
1505 if (!(old_state & state))
1510 this_cpu = smp_processor_id();
1515 #endif /* CONFIG_SMP */
1516 activate_task(rq, p, 1);
1518 * Sync wakeups (i.e. those types of wakeups where the waker
1519 * has indicated that it will leave the CPU in short order)
1520 * don't trigger a preemption, if the woken up task will run on
1521 * this cpu. (in this case the 'I will reschedule' promise of
1522 * the waker guarantees that the freshly woken up task is going
1523 * to be considered on this CPU.)
1525 if (!sync || cpu != this_cpu)
1526 check_preempt_curr(rq, p);
1530 p->state = TASK_RUNNING;
1532 task_rq_unlock(rq, &flags);
1537 int fastcall wake_up_process(struct task_struct *p)
1539 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1540 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1542 EXPORT_SYMBOL(wake_up_process);
1544 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1546 return try_to_wake_up(p, state, 0);
1550 * Perform scheduler related setup for a newly forked process p.
1551 * p is forked by current.
1553 * __sched_fork() is basic setup used by init_idle() too:
1555 static void __sched_fork(struct task_struct *p)
1557 p->se.wait_start_fair = 0;
1558 p->se.wait_start = 0;
1559 p->se.exec_start = 0;
1560 p->se.sum_exec_runtime = 0;
1561 p->se.delta_exec = 0;
1562 p->se.delta_fair_run = 0;
1563 p->se.delta_fair_sleep = 0;
1564 p->se.wait_runtime = 0;
1565 p->se.sum_wait_runtime = 0;
1566 p->se.sum_sleep_runtime = 0;
1567 p->se.sleep_start = 0;
1568 p->se.sleep_start_fair = 0;
1569 p->se.block_start = 0;
1570 p->se.sleep_max = 0;
1571 p->se.block_max = 0;
1574 p->se.wait_runtime_overruns = 0;
1575 p->se.wait_runtime_underruns = 0;
1577 INIT_LIST_HEAD(&p->run_list);
1580 #ifdef CONFIG_PREEMPT_NOTIFIERS
1581 INIT_HLIST_HEAD(&p->preempt_notifiers);
1585 * We mark the process as running here, but have not actually
1586 * inserted it onto the runqueue yet. This guarantees that
1587 * nobody will actually run it, and a signal or other external
1588 * event cannot wake it up and insert it on the runqueue either.
1590 p->state = TASK_RUNNING;
1594 * fork()/clone()-time setup:
1596 void sched_fork(struct task_struct *p, int clone_flags)
1598 int cpu = get_cpu();
1603 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1605 __set_task_cpu(p, cpu);
1608 * Make sure we do not leak PI boosting priority to the child:
1610 p->prio = current->normal_prio;
1612 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1613 if (likely(sched_info_on()))
1614 memset(&p->sched_info, 0, sizeof(p->sched_info));
1616 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1619 #ifdef CONFIG_PREEMPT
1620 /* Want to start with kernel preemption disabled. */
1621 task_thread_info(p)->preempt_count = 1;
1627 * After fork, child runs first. (default) If set to 0 then
1628 * parent will (try to) run first.
1630 unsigned int __read_mostly sysctl_sched_child_runs_first = 1;
1633 * wake_up_new_task - wake up a newly created task for the first time.
1635 * This function will do some initial scheduler statistics housekeeping
1636 * that must be done for every newly created context, then puts the task
1637 * on the runqueue and wakes it.
1639 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1641 unsigned long flags;
1646 rq = task_rq_lock(p, &flags);
1647 BUG_ON(p->state != TASK_RUNNING);
1648 this_cpu = smp_processor_id(); /* parent's CPU */
1651 p->prio = effective_prio(p);
1653 if (!p->sched_class->task_new || !sysctl_sched_child_runs_first ||
1654 (clone_flags & CLONE_VM) || task_cpu(p) != this_cpu ||
1655 !current->se.on_rq) {
1657 activate_task(rq, p, 0);
1660 * Let the scheduling class do new task startup
1661 * management (if any):
1663 p->sched_class->task_new(rq, p, now);
1664 inc_nr_running(p, rq, now);
1666 check_preempt_curr(rq, p);
1667 task_rq_unlock(rq, &flags);
1670 #ifdef CONFIG_PREEMPT_NOTIFIERS
1673 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1674 * @notifier: notifier struct to register
1676 void preempt_notifier_register(struct preempt_notifier *notifier)
1678 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1680 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1683 * preempt_notifier_unregister - no longer interested in preemption notifications
1684 * @notifier: notifier struct to unregister
1686 * This is safe to call from within a preemption notifier.
1688 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1690 hlist_del(¬ifier->link);
1692 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1694 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1696 struct preempt_notifier *notifier;
1697 struct hlist_node *node;
1699 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1700 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1704 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1705 struct task_struct *next)
1707 struct preempt_notifier *notifier;
1708 struct hlist_node *node;
1710 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1711 notifier->ops->sched_out(notifier, next);
1716 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1721 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1722 struct task_struct *next)
1729 * prepare_task_switch - prepare to switch tasks
1730 * @rq: the runqueue preparing to switch
1731 * @prev: the current task that is being switched out
1732 * @next: the task we are going to switch to.
1734 * This is called with the rq lock held and interrupts off. It must
1735 * be paired with a subsequent finish_task_switch after the context
1738 * prepare_task_switch sets up locking and calls architecture specific
1742 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1743 struct task_struct *next)
1745 fire_sched_out_preempt_notifiers(prev, next);
1746 prepare_lock_switch(rq, next);
1747 prepare_arch_switch(next);
1751 * finish_task_switch - clean up after a task-switch
1752 * @rq: runqueue associated with task-switch
1753 * @prev: the thread we just switched away from.
1755 * finish_task_switch must be called after the context switch, paired
1756 * with a prepare_task_switch call before the context switch.
1757 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1758 * and do any other architecture-specific cleanup actions.
1760 * Note that we may have delayed dropping an mm in context_switch(). If
1761 * so, we finish that here outside of the runqueue lock. (Doing it
1762 * with the lock held can cause deadlocks; see schedule() for
1765 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1766 __releases(rq->lock)
1768 struct mm_struct *mm = rq->prev_mm;
1774 * A task struct has one reference for the use as "current".
1775 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1776 * schedule one last time. The schedule call will never return, and
1777 * the scheduled task must drop that reference.
1778 * The test for TASK_DEAD must occur while the runqueue locks are
1779 * still held, otherwise prev could be scheduled on another cpu, die
1780 * there before we look at prev->state, and then the reference would
1782 * Manfred Spraul <manfred@colorfullife.com>
1784 prev_state = prev->state;
1785 finish_arch_switch(prev);
1786 finish_lock_switch(rq, prev);
1787 fire_sched_in_preempt_notifiers(current);
1790 if (unlikely(prev_state == TASK_DEAD)) {
1792 * Remove function-return probe instances associated with this
1793 * task and put them back on the free list.
1795 kprobe_flush_task(prev);
1796 put_task_struct(prev);
1801 * schedule_tail - first thing a freshly forked thread must call.
1802 * @prev: the thread we just switched away from.
1804 asmlinkage void schedule_tail(struct task_struct *prev)
1805 __releases(rq->lock)
1807 struct rq *rq = this_rq();
1809 finish_task_switch(rq, prev);
1810 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1811 /* In this case, finish_task_switch does not reenable preemption */
1814 if (current->set_child_tid)
1815 put_user(current->pid, current->set_child_tid);
1819 * context_switch - switch to the new MM and the new
1820 * thread's register state.
1823 context_switch(struct rq *rq, struct task_struct *prev,
1824 struct task_struct *next)
1826 struct mm_struct *mm, *oldmm;
1828 prepare_task_switch(rq, prev, next);
1830 oldmm = prev->active_mm;
1832 * For paravirt, this is coupled with an exit in switch_to to
1833 * combine the page table reload and the switch backend into
1836 arch_enter_lazy_cpu_mode();
1838 if (unlikely(!mm)) {
1839 next->active_mm = oldmm;
1840 atomic_inc(&oldmm->mm_count);
1841 enter_lazy_tlb(oldmm, next);
1843 switch_mm(oldmm, mm, next);
1845 if (unlikely(!prev->mm)) {
1846 prev->active_mm = NULL;
1847 rq->prev_mm = oldmm;
1850 * Since the runqueue lock will be released by the next
1851 * task (which is an invalid locking op but in the case
1852 * of the scheduler it's an obvious special-case), so we
1853 * do an early lockdep release here:
1855 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1856 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1859 /* Here we just switch the register state and the stack. */
1860 switch_to(prev, next, prev);
1864 * this_rq must be evaluated again because prev may have moved
1865 * CPUs since it called schedule(), thus the 'rq' on its stack
1866 * frame will be invalid.
1868 finish_task_switch(this_rq(), prev);
1872 * nr_running, nr_uninterruptible and nr_context_switches:
1874 * externally visible scheduler statistics: current number of runnable
1875 * threads, current number of uninterruptible-sleeping threads, total
1876 * number of context switches performed since bootup.
1878 unsigned long nr_running(void)
1880 unsigned long i, sum = 0;
1882 for_each_online_cpu(i)
1883 sum += cpu_rq(i)->nr_running;
1888 unsigned long nr_uninterruptible(void)
1890 unsigned long i, sum = 0;
1892 for_each_possible_cpu(i)
1893 sum += cpu_rq(i)->nr_uninterruptible;
1896 * Since we read the counters lockless, it might be slightly
1897 * inaccurate. Do not allow it to go below zero though:
1899 if (unlikely((long)sum < 0))
1905 unsigned long long nr_context_switches(void)
1908 unsigned long long sum = 0;
1910 for_each_possible_cpu(i)
1911 sum += cpu_rq(i)->nr_switches;
1916 unsigned long nr_iowait(void)
1918 unsigned long i, sum = 0;
1920 for_each_possible_cpu(i)
1921 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1926 unsigned long nr_active(void)
1928 unsigned long i, running = 0, uninterruptible = 0;
1930 for_each_online_cpu(i) {
1931 running += cpu_rq(i)->nr_running;
1932 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1935 if (unlikely((long)uninterruptible < 0))
1936 uninterruptible = 0;
1938 return running + uninterruptible;
1942 * Update rq->cpu_load[] statistics. This function is usually called every
1943 * scheduler tick (TICK_NSEC).
1945 static void update_cpu_load(struct rq *this_rq)
1947 u64 fair_delta64, exec_delta64, idle_delta64, sample_interval64, tmp64;
1948 unsigned long total_load = this_rq->ls.load.weight;
1949 unsigned long this_load = total_load;
1950 struct load_stat *ls = &this_rq->ls;
1951 u64 now = __rq_clock(this_rq);
1954 this_rq->nr_load_updates++;
1955 if (unlikely(!(sysctl_sched_features & SCHED_FEAT_PRECISE_CPU_LOAD)))
1958 /* Update delta_fair/delta_exec fields first */
1959 update_curr_load(this_rq, now);
1961 fair_delta64 = ls->delta_fair + 1;
1964 exec_delta64 = ls->delta_exec + 1;
1967 sample_interval64 = now - ls->load_update_last;
1968 ls->load_update_last = now;
1970 if ((s64)sample_interval64 < (s64)TICK_NSEC)
1971 sample_interval64 = TICK_NSEC;
1973 if (exec_delta64 > sample_interval64)
1974 exec_delta64 = sample_interval64;
1976 idle_delta64 = sample_interval64 - exec_delta64;
1978 tmp64 = div64_64(SCHED_LOAD_SCALE * exec_delta64, fair_delta64);
1979 tmp64 = div64_64(tmp64 * exec_delta64, sample_interval64);
1981 this_load = (unsigned long)tmp64;
1985 /* Update our load: */
1986 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1987 unsigned long old_load, new_load;
1989 /* scale is effectively 1 << i now, and >> i divides by scale */
1991 old_load = this_rq->cpu_load[i];
1992 new_load = this_load;
1994 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2001 * double_rq_lock - safely lock two runqueues
2003 * Note this does not disable interrupts like task_rq_lock,
2004 * you need to do so manually before calling.
2006 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2007 __acquires(rq1->lock)
2008 __acquires(rq2->lock)
2010 BUG_ON(!irqs_disabled());
2012 spin_lock(&rq1->lock);
2013 __acquire(rq2->lock); /* Fake it out ;) */
2016 spin_lock(&rq1->lock);
2017 spin_lock(&rq2->lock);
2019 spin_lock(&rq2->lock);
2020 spin_lock(&rq1->lock);
2026 * double_rq_unlock - safely unlock two runqueues
2028 * Note this does not restore interrupts like task_rq_unlock,
2029 * you need to do so manually after calling.
2031 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2032 __releases(rq1->lock)
2033 __releases(rq2->lock)
2035 spin_unlock(&rq1->lock);
2037 spin_unlock(&rq2->lock);
2039 __release(rq2->lock);
2043 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2045 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2046 __releases(this_rq->lock)
2047 __acquires(busiest->lock)
2048 __acquires(this_rq->lock)
2050 if (unlikely(!irqs_disabled())) {
2051 /* printk() doesn't work good under rq->lock */
2052 spin_unlock(&this_rq->lock);
2055 if (unlikely(!spin_trylock(&busiest->lock))) {
2056 if (busiest < this_rq) {
2057 spin_unlock(&this_rq->lock);
2058 spin_lock(&busiest->lock);
2059 spin_lock(&this_rq->lock);
2061 spin_lock(&busiest->lock);
2066 * If dest_cpu is allowed for this process, migrate the task to it.
2067 * This is accomplished by forcing the cpu_allowed mask to only
2068 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2069 * the cpu_allowed mask is restored.
2071 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2073 struct migration_req req;
2074 unsigned long flags;
2077 rq = task_rq_lock(p, &flags);
2078 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2079 || unlikely(cpu_is_offline(dest_cpu)))
2082 /* force the process onto the specified CPU */
2083 if (migrate_task(p, dest_cpu, &req)) {
2084 /* Need to wait for migration thread (might exit: take ref). */
2085 struct task_struct *mt = rq->migration_thread;
2087 get_task_struct(mt);
2088 task_rq_unlock(rq, &flags);
2089 wake_up_process(mt);
2090 put_task_struct(mt);
2091 wait_for_completion(&req.done);
2096 task_rq_unlock(rq, &flags);
2100 * sched_exec - execve() is a valuable balancing opportunity, because at
2101 * this point the task has the smallest effective memory and cache footprint.
2103 void sched_exec(void)
2105 int new_cpu, this_cpu = get_cpu();
2106 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2108 if (new_cpu != this_cpu)
2109 sched_migrate_task(current, new_cpu);
2113 * pull_task - move a task from a remote runqueue to the local runqueue.
2114 * Both runqueues must be locked.
2116 static void pull_task(struct rq *src_rq, struct task_struct *p,
2117 struct rq *this_rq, int this_cpu)
2119 deactivate_task(src_rq, p, 0);
2120 set_task_cpu(p, this_cpu);
2121 activate_task(this_rq, p, 0);
2123 * Note that idle threads have a prio of MAX_PRIO, for this test
2124 * to be always true for them.
2126 check_preempt_curr(this_rq, p);
2130 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2133 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2134 struct sched_domain *sd, enum cpu_idle_type idle,
2138 * We do not migrate tasks that are:
2139 * 1) running (obviously), or
2140 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2141 * 3) are cache-hot on their current CPU.
2143 if (!cpu_isset(this_cpu, p->cpus_allowed))
2147 if (task_running(rq, p))
2151 * Aggressive migration if too many balance attempts have failed:
2153 if (sd->nr_balance_failed > sd->cache_nice_tries)
2159 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2160 unsigned long max_nr_move, unsigned long max_load_move,
2161 struct sched_domain *sd, enum cpu_idle_type idle,
2162 int *all_pinned, unsigned long *load_moved,
2163 int this_best_prio, int best_prio, int best_prio_seen,
2164 struct rq_iterator *iterator)
2166 int pulled = 0, pinned = 0, skip_for_load;
2167 struct task_struct *p;
2168 long rem_load_move = max_load_move;
2170 if (max_nr_move == 0 || max_load_move == 0)
2176 * Start the load-balancing iterator:
2178 p = iterator->start(iterator->arg);
2183 * To help distribute high priority tasks accross CPUs we don't
2184 * skip a task if it will be the highest priority task (i.e. smallest
2185 * prio value) on its new queue regardless of its load weight
2187 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2188 SCHED_LOAD_SCALE_FUZZ;
2189 if (skip_for_load && p->prio < this_best_prio)
2190 skip_for_load = !best_prio_seen && p->prio == best_prio;
2191 if (skip_for_load ||
2192 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2194 best_prio_seen |= p->prio == best_prio;
2195 p = iterator->next(iterator->arg);
2199 pull_task(busiest, p, this_rq, this_cpu);
2201 rem_load_move -= p->se.load.weight;
2204 * We only want to steal up to the prescribed number of tasks
2205 * and the prescribed amount of weighted load.
2207 if (pulled < max_nr_move && rem_load_move > 0) {
2208 if (p->prio < this_best_prio)
2209 this_best_prio = p->prio;
2210 p = iterator->next(iterator->arg);
2215 * Right now, this is the only place pull_task() is called,
2216 * so we can safely collect pull_task() stats here rather than
2217 * inside pull_task().
2219 schedstat_add(sd, lb_gained[idle], pulled);
2222 *all_pinned = pinned;
2223 *load_moved = max_load_move - rem_load_move;
2228 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2229 * load from busiest to this_rq, as part of a balancing operation within
2230 * "domain". Returns the number of tasks moved.
2232 * Called with both runqueues locked.
2234 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2235 unsigned long max_nr_move, unsigned long max_load_move,
2236 struct sched_domain *sd, enum cpu_idle_type idle,
2239 struct sched_class *class = sched_class_highest;
2240 unsigned long load_moved, total_nr_moved = 0, nr_moved;
2241 long rem_load_move = max_load_move;
2244 nr_moved = class->load_balance(this_rq, this_cpu, busiest,
2245 max_nr_move, (unsigned long)rem_load_move,
2246 sd, idle, all_pinned, &load_moved);
2247 total_nr_moved += nr_moved;
2248 max_nr_move -= nr_moved;
2249 rem_load_move -= load_moved;
2250 class = class->next;
2251 } while (class && max_nr_move && rem_load_move > 0);
2253 return total_nr_moved;
2257 * find_busiest_group finds and returns the busiest CPU group within the
2258 * domain. It calculates and returns the amount of weighted load which
2259 * should be moved to restore balance via the imbalance parameter.
2261 static struct sched_group *
2262 find_busiest_group(struct sched_domain *sd, int this_cpu,
2263 unsigned long *imbalance, enum cpu_idle_type idle,
2264 int *sd_idle, cpumask_t *cpus, int *balance)
2266 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2267 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2268 unsigned long max_pull;
2269 unsigned long busiest_load_per_task, busiest_nr_running;
2270 unsigned long this_load_per_task, this_nr_running;
2272 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2273 int power_savings_balance = 1;
2274 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2275 unsigned long min_nr_running = ULONG_MAX;
2276 struct sched_group *group_min = NULL, *group_leader = NULL;
2279 max_load = this_load = total_load = total_pwr = 0;
2280 busiest_load_per_task = busiest_nr_running = 0;
2281 this_load_per_task = this_nr_running = 0;
2282 if (idle == CPU_NOT_IDLE)
2283 load_idx = sd->busy_idx;
2284 else if (idle == CPU_NEWLY_IDLE)
2285 load_idx = sd->newidle_idx;
2287 load_idx = sd->idle_idx;
2290 unsigned long load, group_capacity;
2293 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2294 unsigned long sum_nr_running, sum_weighted_load;
2296 local_group = cpu_isset(this_cpu, group->cpumask);
2299 balance_cpu = first_cpu(group->cpumask);
2301 /* Tally up the load of all CPUs in the group */
2302 sum_weighted_load = sum_nr_running = avg_load = 0;
2304 for_each_cpu_mask(i, group->cpumask) {
2307 if (!cpu_isset(i, *cpus))
2312 if (*sd_idle && rq->nr_running)
2315 /* Bias balancing toward cpus of our domain */
2317 if (idle_cpu(i) && !first_idle_cpu) {
2322 load = target_load(i, load_idx);
2324 load = source_load(i, load_idx);
2327 sum_nr_running += rq->nr_running;
2328 sum_weighted_load += weighted_cpuload(i);
2332 * First idle cpu or the first cpu(busiest) in this sched group
2333 * is eligible for doing load balancing at this and above
2334 * domains. In the newly idle case, we will allow all the cpu's
2335 * to do the newly idle load balance.
2337 if (idle != CPU_NEWLY_IDLE && local_group &&
2338 balance_cpu != this_cpu && balance) {
2343 total_load += avg_load;
2344 total_pwr += group->__cpu_power;
2346 /* Adjust by relative CPU power of the group */
2347 avg_load = sg_div_cpu_power(group,
2348 avg_load * SCHED_LOAD_SCALE);
2350 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2353 this_load = avg_load;
2355 this_nr_running = sum_nr_running;
2356 this_load_per_task = sum_weighted_load;
2357 } else if (avg_load > max_load &&
2358 sum_nr_running > group_capacity) {
2359 max_load = avg_load;
2361 busiest_nr_running = sum_nr_running;
2362 busiest_load_per_task = sum_weighted_load;
2365 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2367 * Busy processors will not participate in power savings
2370 if (idle == CPU_NOT_IDLE ||
2371 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2375 * If the local group is idle or completely loaded
2376 * no need to do power savings balance at this domain
2378 if (local_group && (this_nr_running >= group_capacity ||
2380 power_savings_balance = 0;
2383 * If a group is already running at full capacity or idle,
2384 * don't include that group in power savings calculations
2386 if (!power_savings_balance || sum_nr_running >= group_capacity
2391 * Calculate the group which has the least non-idle load.
2392 * This is the group from where we need to pick up the load
2395 if ((sum_nr_running < min_nr_running) ||
2396 (sum_nr_running == min_nr_running &&
2397 first_cpu(group->cpumask) <
2398 first_cpu(group_min->cpumask))) {
2400 min_nr_running = sum_nr_running;
2401 min_load_per_task = sum_weighted_load /
2406 * Calculate the group which is almost near its
2407 * capacity but still has some space to pick up some load
2408 * from other group and save more power
2410 if (sum_nr_running <= group_capacity - 1) {
2411 if (sum_nr_running > leader_nr_running ||
2412 (sum_nr_running == leader_nr_running &&
2413 first_cpu(group->cpumask) >
2414 first_cpu(group_leader->cpumask))) {
2415 group_leader = group;
2416 leader_nr_running = sum_nr_running;
2421 group = group->next;
2422 } while (group != sd->groups);
2424 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2427 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2429 if (this_load >= avg_load ||
2430 100*max_load <= sd->imbalance_pct*this_load)
2433 busiest_load_per_task /= busiest_nr_running;
2435 * We're trying to get all the cpus to the average_load, so we don't
2436 * want to push ourselves above the average load, nor do we wish to
2437 * reduce the max loaded cpu below the average load, as either of these
2438 * actions would just result in more rebalancing later, and ping-pong
2439 * tasks around. Thus we look for the minimum possible imbalance.
2440 * Negative imbalances (*we* are more loaded than anyone else) will
2441 * be counted as no imbalance for these purposes -- we can't fix that
2442 * by pulling tasks to us. Be careful of negative numbers as they'll
2443 * appear as very large values with unsigned longs.
2445 if (max_load <= busiest_load_per_task)
2449 * In the presence of smp nice balancing, certain scenarios can have
2450 * max load less than avg load(as we skip the groups at or below
2451 * its cpu_power, while calculating max_load..)
2453 if (max_load < avg_load) {
2455 goto small_imbalance;
2458 /* Don't want to pull so many tasks that a group would go idle */
2459 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2461 /* How much load to actually move to equalise the imbalance */
2462 *imbalance = min(max_pull * busiest->__cpu_power,
2463 (avg_load - this_load) * this->__cpu_power)
2467 * if *imbalance is less than the average load per runnable task
2468 * there is no gaurantee that any tasks will be moved so we'll have
2469 * a think about bumping its value to force at least one task to be
2472 if (*imbalance + SCHED_LOAD_SCALE_FUZZ < busiest_load_per_task/2) {
2473 unsigned long tmp, pwr_now, pwr_move;
2477 pwr_move = pwr_now = 0;
2479 if (this_nr_running) {
2480 this_load_per_task /= this_nr_running;
2481 if (busiest_load_per_task > this_load_per_task)
2484 this_load_per_task = SCHED_LOAD_SCALE;
2486 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2487 busiest_load_per_task * imbn) {
2488 *imbalance = busiest_load_per_task;
2493 * OK, we don't have enough imbalance to justify moving tasks,
2494 * however we may be able to increase total CPU power used by
2498 pwr_now += busiest->__cpu_power *
2499 min(busiest_load_per_task, max_load);
2500 pwr_now += this->__cpu_power *
2501 min(this_load_per_task, this_load);
2502 pwr_now /= SCHED_LOAD_SCALE;
2504 /* Amount of load we'd subtract */
2505 tmp = sg_div_cpu_power(busiest,
2506 busiest_load_per_task * SCHED_LOAD_SCALE);
2508 pwr_move += busiest->__cpu_power *
2509 min(busiest_load_per_task, max_load - tmp);
2511 /* Amount of load we'd add */
2512 if (max_load * busiest->__cpu_power <
2513 busiest_load_per_task * SCHED_LOAD_SCALE)
2514 tmp = sg_div_cpu_power(this,
2515 max_load * busiest->__cpu_power);
2517 tmp = sg_div_cpu_power(this,
2518 busiest_load_per_task * SCHED_LOAD_SCALE);
2519 pwr_move += this->__cpu_power *
2520 min(this_load_per_task, this_load + tmp);
2521 pwr_move /= SCHED_LOAD_SCALE;
2523 /* Move if we gain throughput */
2524 if (pwr_move <= pwr_now)
2527 *imbalance = busiest_load_per_task;
2533 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2534 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2537 if (this == group_leader && group_leader != group_min) {
2538 *imbalance = min_load_per_task;
2548 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2551 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2552 unsigned long imbalance, cpumask_t *cpus)
2554 struct rq *busiest = NULL, *rq;
2555 unsigned long max_load = 0;
2558 for_each_cpu_mask(i, group->cpumask) {
2561 if (!cpu_isset(i, *cpus))
2565 wl = weighted_cpuload(i);
2567 if (rq->nr_running == 1 && wl > imbalance)
2570 if (wl > max_load) {
2580 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2581 * so long as it is large enough.
2583 #define MAX_PINNED_INTERVAL 512
2585 static inline unsigned long minus_1_or_zero(unsigned long n)
2587 return n > 0 ? n - 1 : 0;
2591 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2592 * tasks if there is an imbalance.
2594 static int load_balance(int this_cpu, struct rq *this_rq,
2595 struct sched_domain *sd, enum cpu_idle_type idle,
2598 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2599 struct sched_group *group;
2600 unsigned long imbalance;
2602 cpumask_t cpus = CPU_MASK_ALL;
2603 unsigned long flags;
2606 * When power savings policy is enabled for the parent domain, idle
2607 * sibling can pick up load irrespective of busy siblings. In this case,
2608 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2609 * portraying it as CPU_NOT_IDLE.
2611 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2612 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2615 schedstat_inc(sd, lb_cnt[idle]);
2618 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2625 schedstat_inc(sd, lb_nobusyg[idle]);
2629 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2631 schedstat_inc(sd, lb_nobusyq[idle]);
2635 BUG_ON(busiest == this_rq);
2637 schedstat_add(sd, lb_imbalance[idle], imbalance);
2640 if (busiest->nr_running > 1) {
2642 * Attempt to move tasks. If find_busiest_group has found
2643 * an imbalance but busiest->nr_running <= 1, the group is
2644 * still unbalanced. nr_moved simply stays zero, so it is
2645 * correctly treated as an imbalance.
2647 local_irq_save(flags);
2648 double_rq_lock(this_rq, busiest);
2649 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2650 minus_1_or_zero(busiest->nr_running),
2651 imbalance, sd, idle, &all_pinned);
2652 double_rq_unlock(this_rq, busiest);
2653 local_irq_restore(flags);
2656 * some other cpu did the load balance for us.
2658 if (nr_moved && this_cpu != smp_processor_id())
2659 resched_cpu(this_cpu);
2661 /* All tasks on this runqueue were pinned by CPU affinity */
2662 if (unlikely(all_pinned)) {
2663 cpu_clear(cpu_of(busiest), cpus);
2664 if (!cpus_empty(cpus))
2671 schedstat_inc(sd, lb_failed[idle]);
2672 sd->nr_balance_failed++;
2674 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2676 spin_lock_irqsave(&busiest->lock, flags);
2678 /* don't kick the migration_thread, if the curr
2679 * task on busiest cpu can't be moved to this_cpu
2681 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2682 spin_unlock_irqrestore(&busiest->lock, flags);
2684 goto out_one_pinned;
2687 if (!busiest->active_balance) {
2688 busiest->active_balance = 1;
2689 busiest->push_cpu = this_cpu;
2692 spin_unlock_irqrestore(&busiest->lock, flags);
2694 wake_up_process(busiest->migration_thread);
2697 * We've kicked active balancing, reset the failure
2700 sd->nr_balance_failed = sd->cache_nice_tries+1;
2703 sd->nr_balance_failed = 0;
2705 if (likely(!active_balance)) {
2706 /* We were unbalanced, so reset the balancing interval */
2707 sd->balance_interval = sd->min_interval;
2710 * If we've begun active balancing, start to back off. This
2711 * case may not be covered by the all_pinned logic if there
2712 * is only 1 task on the busy runqueue (because we don't call
2715 if (sd->balance_interval < sd->max_interval)
2716 sd->balance_interval *= 2;
2719 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2720 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2725 schedstat_inc(sd, lb_balanced[idle]);
2727 sd->nr_balance_failed = 0;
2730 /* tune up the balancing interval */
2731 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2732 (sd->balance_interval < sd->max_interval))
2733 sd->balance_interval *= 2;
2735 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2736 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2742 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2743 * tasks if there is an imbalance.
2745 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2746 * this_rq is locked.
2749 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2751 struct sched_group *group;
2752 struct rq *busiest = NULL;
2753 unsigned long imbalance;
2757 cpumask_t cpus = CPU_MASK_ALL;
2760 * When power savings policy is enabled for the parent domain, idle
2761 * sibling can pick up load irrespective of busy siblings. In this case,
2762 * let the state of idle sibling percolate up as IDLE, instead of
2763 * portraying it as CPU_NOT_IDLE.
2765 if (sd->flags & SD_SHARE_CPUPOWER &&
2766 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2769 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2771 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2772 &sd_idle, &cpus, NULL);
2774 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2778 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2781 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2785 BUG_ON(busiest == this_rq);
2787 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2790 if (busiest->nr_running > 1) {
2791 /* Attempt to move tasks */
2792 double_lock_balance(this_rq, busiest);
2793 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2794 minus_1_or_zero(busiest->nr_running),
2795 imbalance, sd, CPU_NEWLY_IDLE,
2797 spin_unlock(&busiest->lock);
2799 if (unlikely(all_pinned)) {
2800 cpu_clear(cpu_of(busiest), cpus);
2801 if (!cpus_empty(cpus))
2807 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2808 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2809 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2812 sd->nr_balance_failed = 0;
2817 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2818 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2819 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2821 sd->nr_balance_failed = 0;
2827 * idle_balance is called by schedule() if this_cpu is about to become
2828 * idle. Attempts to pull tasks from other CPUs.
2830 static void idle_balance(int this_cpu, struct rq *this_rq)
2832 struct sched_domain *sd;
2833 int pulled_task = -1;
2834 unsigned long next_balance = jiffies + HZ;
2836 for_each_domain(this_cpu, sd) {
2837 unsigned long interval;
2839 if (!(sd->flags & SD_LOAD_BALANCE))
2842 if (sd->flags & SD_BALANCE_NEWIDLE)
2843 /* If we've pulled tasks over stop searching: */
2844 pulled_task = load_balance_newidle(this_cpu,
2847 interval = msecs_to_jiffies(sd->balance_interval);
2848 if (time_after(next_balance, sd->last_balance + interval))
2849 next_balance = sd->last_balance + interval;
2853 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2855 * We are going idle. next_balance may be set based on
2856 * a busy processor. So reset next_balance.
2858 this_rq->next_balance = next_balance;
2863 * active_load_balance is run by migration threads. It pushes running tasks
2864 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2865 * running on each physical CPU where possible, and avoids physical /
2866 * logical imbalances.
2868 * Called with busiest_rq locked.
2870 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2872 int target_cpu = busiest_rq->push_cpu;
2873 struct sched_domain *sd;
2874 struct rq *target_rq;
2876 /* Is there any task to move? */
2877 if (busiest_rq->nr_running <= 1)
2880 target_rq = cpu_rq(target_cpu);
2883 * This condition is "impossible", if it occurs
2884 * we need to fix it. Originally reported by
2885 * Bjorn Helgaas on a 128-cpu setup.
2887 BUG_ON(busiest_rq == target_rq);
2889 /* move a task from busiest_rq to target_rq */
2890 double_lock_balance(busiest_rq, target_rq);
2892 /* Search for an sd spanning us and the target CPU. */
2893 for_each_domain(target_cpu, sd) {
2894 if ((sd->flags & SD_LOAD_BALANCE) &&
2895 cpu_isset(busiest_cpu, sd->span))
2900 schedstat_inc(sd, alb_cnt);
2902 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2903 ULONG_MAX, sd, CPU_IDLE, NULL))
2904 schedstat_inc(sd, alb_pushed);
2906 schedstat_inc(sd, alb_failed);
2908 spin_unlock(&target_rq->lock);
2913 atomic_t load_balancer;
2915 } nohz ____cacheline_aligned = {
2916 .load_balancer = ATOMIC_INIT(-1),
2917 .cpu_mask = CPU_MASK_NONE,
2921 * This routine will try to nominate the ilb (idle load balancing)
2922 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2923 * load balancing on behalf of all those cpus. If all the cpus in the system
2924 * go into this tickless mode, then there will be no ilb owner (as there is
2925 * no need for one) and all the cpus will sleep till the next wakeup event
2928 * For the ilb owner, tick is not stopped. And this tick will be used
2929 * for idle load balancing. ilb owner will still be part of
2932 * While stopping the tick, this cpu will become the ilb owner if there
2933 * is no other owner. And will be the owner till that cpu becomes busy
2934 * or if all cpus in the system stop their ticks at which point
2935 * there is no need for ilb owner.
2937 * When the ilb owner becomes busy, it nominates another owner, during the
2938 * next busy scheduler_tick()
2940 int select_nohz_load_balancer(int stop_tick)
2942 int cpu = smp_processor_id();
2945 cpu_set(cpu, nohz.cpu_mask);
2946 cpu_rq(cpu)->in_nohz_recently = 1;
2949 * If we are going offline and still the leader, give up!
2951 if (cpu_is_offline(cpu) &&
2952 atomic_read(&nohz.load_balancer) == cpu) {
2953 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2958 /* time for ilb owner also to sleep */
2959 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2960 if (atomic_read(&nohz.load_balancer) == cpu)
2961 atomic_set(&nohz.load_balancer, -1);
2965 if (atomic_read(&nohz.load_balancer) == -1) {
2966 /* make me the ilb owner */
2967 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2969 } else if (atomic_read(&nohz.load_balancer) == cpu)
2972 if (!cpu_isset(cpu, nohz.cpu_mask))
2975 cpu_clear(cpu, nohz.cpu_mask);
2977 if (atomic_read(&nohz.load_balancer) == cpu)
2978 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2985 static DEFINE_SPINLOCK(balancing);
2988 * It checks each scheduling domain to see if it is due to be balanced,
2989 * and initiates a balancing operation if so.
2991 * Balancing parameters are set up in arch_init_sched_domains.
2993 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
2996 struct rq *rq = cpu_rq(cpu);
2997 unsigned long interval;
2998 struct sched_domain *sd;
2999 /* Earliest time when we have to do rebalance again */
3000 unsigned long next_balance = jiffies + 60*HZ;
3002 for_each_domain(cpu, sd) {
3003 if (!(sd->flags & SD_LOAD_BALANCE))
3006 interval = sd->balance_interval;
3007 if (idle != CPU_IDLE)
3008 interval *= sd->busy_factor;
3010 /* scale ms to jiffies */
3011 interval = msecs_to_jiffies(interval);
3012 if (unlikely(!interval))
3014 if (interval > HZ*NR_CPUS/10)
3015 interval = HZ*NR_CPUS/10;
3018 if (sd->flags & SD_SERIALIZE) {
3019 if (!spin_trylock(&balancing))
3023 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3024 if (load_balance(cpu, rq, sd, idle, &balance)) {
3026 * We've pulled tasks over so either we're no
3027 * longer idle, or one of our SMT siblings is
3030 idle = CPU_NOT_IDLE;
3032 sd->last_balance = jiffies;
3034 if (sd->flags & SD_SERIALIZE)
3035 spin_unlock(&balancing);
3037 if (time_after(next_balance, sd->last_balance + interval))
3038 next_balance = sd->last_balance + interval;
3041 * Stop the load balance at this level. There is another
3042 * CPU in our sched group which is doing load balancing more
3048 rq->next_balance = next_balance;
3052 * run_rebalance_domains is triggered when needed from the scheduler tick.
3053 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3054 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3056 static void run_rebalance_domains(struct softirq_action *h)
3058 int this_cpu = smp_processor_id();
3059 struct rq *this_rq = cpu_rq(this_cpu);
3060 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3061 CPU_IDLE : CPU_NOT_IDLE;
3063 rebalance_domains(this_cpu, idle);
3067 * If this cpu is the owner for idle load balancing, then do the
3068 * balancing on behalf of the other idle cpus whose ticks are
3071 if (this_rq->idle_at_tick &&
3072 atomic_read(&nohz.load_balancer) == this_cpu) {
3073 cpumask_t cpus = nohz.cpu_mask;
3077 cpu_clear(this_cpu, cpus);
3078 for_each_cpu_mask(balance_cpu, cpus) {
3080 * If this cpu gets work to do, stop the load balancing
3081 * work being done for other cpus. Next load
3082 * balancing owner will pick it up.
3087 rebalance_domains(balance_cpu, SCHED_IDLE);
3089 rq = cpu_rq(balance_cpu);
3090 if (time_after(this_rq->next_balance, rq->next_balance))
3091 this_rq->next_balance = rq->next_balance;
3098 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3100 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3101 * idle load balancing owner or decide to stop the periodic load balancing,
3102 * if the whole system is idle.
3104 static inline void trigger_load_balance(struct rq *rq, int cpu)
3108 * If we were in the nohz mode recently and busy at the current
3109 * scheduler tick, then check if we need to nominate new idle
3112 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3113 rq->in_nohz_recently = 0;
3115 if (atomic_read(&nohz.load_balancer) == cpu) {
3116 cpu_clear(cpu, nohz.cpu_mask);
3117 atomic_set(&nohz.load_balancer, -1);
3120 if (atomic_read(&nohz.load_balancer) == -1) {
3122 * simple selection for now: Nominate the
3123 * first cpu in the nohz list to be the next
3126 * TBD: Traverse the sched domains and nominate
3127 * the nearest cpu in the nohz.cpu_mask.
3129 int ilb = first_cpu(nohz.cpu_mask);
3137 * If this cpu is idle and doing idle load balancing for all the
3138 * cpus with ticks stopped, is it time for that to stop?
3140 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3141 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3147 * If this cpu is idle and the idle load balancing is done by
3148 * someone else, then no need raise the SCHED_SOFTIRQ
3150 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3151 cpu_isset(cpu, nohz.cpu_mask))
3154 if (time_after_eq(jiffies, rq->next_balance))
3155 raise_softirq(SCHED_SOFTIRQ);
3158 #else /* CONFIG_SMP */
3161 * on UP we do not need to balance between CPUs:
3163 static inline void idle_balance(int cpu, struct rq *rq)
3167 /* Avoid "used but not defined" warning on UP */
3168 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3169 unsigned long max_nr_move, unsigned long max_load_move,
3170 struct sched_domain *sd, enum cpu_idle_type idle,
3171 int *all_pinned, unsigned long *load_moved,
3172 int this_best_prio, int best_prio, int best_prio_seen,
3173 struct rq_iterator *iterator)
3182 DEFINE_PER_CPU(struct kernel_stat, kstat);
3184 EXPORT_PER_CPU_SYMBOL(kstat);
3187 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3188 * that have not yet been banked in case the task is currently running.
3190 unsigned long long task_sched_runtime(struct task_struct *p)
3192 unsigned long flags;
3196 rq = task_rq_lock(p, &flags);
3197 ns = p->se.sum_exec_runtime;
3198 if (rq->curr == p) {
3199 delta_exec = rq_clock(rq) - p->se.exec_start;
3200 if ((s64)delta_exec > 0)
3203 task_rq_unlock(rq, &flags);
3209 * Account user cpu time to a process.
3210 * @p: the process that the cpu time gets accounted to
3211 * @hardirq_offset: the offset to subtract from hardirq_count()
3212 * @cputime: the cpu time spent in user space since the last update
3214 void account_user_time(struct task_struct *p, cputime_t cputime)
3216 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3219 p->utime = cputime_add(p->utime, cputime);
3221 /* Add user time to cpustat. */
3222 tmp = cputime_to_cputime64(cputime);
3223 if (TASK_NICE(p) > 0)
3224 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3226 cpustat->user = cputime64_add(cpustat->user, tmp);
3230 * Account system cpu time to a process.
3231 * @p: the process that the cpu time gets accounted to
3232 * @hardirq_offset: the offset to subtract from hardirq_count()
3233 * @cputime: the cpu time spent in kernel space since the last update
3235 void account_system_time(struct task_struct *p, int hardirq_offset,
3238 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3239 struct rq *rq = this_rq();
3242 p->stime = cputime_add(p->stime, cputime);
3244 /* Add system time to cpustat. */
3245 tmp = cputime_to_cputime64(cputime);
3246 if (hardirq_count() - hardirq_offset)
3247 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3248 else if (softirq_count())
3249 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3250 else if (p != rq->idle)
3251 cpustat->system = cputime64_add(cpustat->system, tmp);
3252 else if (atomic_read(&rq->nr_iowait) > 0)
3253 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3255 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3256 /* Account for system time used */
3257 acct_update_integrals(p);
3261 * Account for involuntary wait time.
3262 * @p: the process from which the cpu time has been stolen
3263 * @steal: the cpu time spent in involuntary wait
3265 void account_steal_time(struct task_struct *p, cputime_t steal)
3267 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3268 cputime64_t tmp = cputime_to_cputime64(steal);
3269 struct rq *rq = this_rq();
3271 if (p == rq->idle) {
3272 p->stime = cputime_add(p->stime, steal);
3273 if (atomic_read(&rq->nr_iowait) > 0)
3274 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3276 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3278 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3282 * This function gets called by the timer code, with HZ frequency.
3283 * We call it with interrupts disabled.
3285 * It also gets called by the fork code, when changing the parent's
3288 void scheduler_tick(void)
3290 int cpu = smp_processor_id();
3291 struct rq *rq = cpu_rq(cpu);
3292 struct task_struct *curr = rq->curr;
3294 spin_lock(&rq->lock);
3295 if (curr != rq->idle) /* FIXME: needed? */
3296 curr->sched_class->task_tick(rq, curr);
3297 update_cpu_load(rq);
3298 spin_unlock(&rq->lock);
3301 rq->idle_at_tick = idle_cpu(cpu);
3302 trigger_load_balance(rq, cpu);
3306 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3308 void fastcall add_preempt_count(int val)
3313 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3315 preempt_count() += val;
3317 * Spinlock count overflowing soon?
3319 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3322 EXPORT_SYMBOL(add_preempt_count);
3324 void fastcall sub_preempt_count(int val)
3329 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3332 * Is the spinlock portion underflowing?
3334 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3335 !(preempt_count() & PREEMPT_MASK)))
3338 preempt_count() -= val;
3340 EXPORT_SYMBOL(sub_preempt_count);
3345 * Print scheduling while atomic bug:
3347 static noinline void __schedule_bug(struct task_struct *prev)
3349 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3350 prev->comm, preempt_count(), prev->pid);
3351 debug_show_held_locks(prev);
3352 if (irqs_disabled())
3353 print_irqtrace_events(prev);
3358 * Various schedule()-time debugging checks and statistics:
3360 static inline void schedule_debug(struct task_struct *prev)
3363 * Test if we are atomic. Since do_exit() needs to call into
3364 * schedule() atomically, we ignore that path for now.
3365 * Otherwise, whine if we are scheduling when we should not be.
3367 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3368 __schedule_bug(prev);
3370 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3372 schedstat_inc(this_rq(), sched_cnt);
3376 * Pick up the highest-prio task:
3378 static inline struct task_struct *
3379 pick_next_task(struct rq *rq, struct task_struct *prev, u64 now)
3381 struct sched_class *class;
3382 struct task_struct *p;
3385 * Optimization: we know that if all tasks are in
3386 * the fair class we can call that function directly:
3388 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3389 p = fair_sched_class.pick_next_task(rq, now);
3394 class = sched_class_highest;
3396 p = class->pick_next_task(rq, now);
3400 * Will never be NULL as the idle class always
3401 * returns a non-NULL p:
3403 class = class->next;
3408 * schedule() is the main scheduler function.
3410 asmlinkage void __sched schedule(void)
3412 struct task_struct *prev, *next;
3420 cpu = smp_processor_id();
3424 switch_count = &prev->nivcsw;
3426 release_kernel_lock(prev);
3427 need_resched_nonpreemptible:
3429 schedule_debug(prev);
3431 spin_lock_irq(&rq->lock);
3432 clear_tsk_need_resched(prev);
3434 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3435 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3436 unlikely(signal_pending(prev)))) {
3437 prev->state = TASK_RUNNING;
3439 deactivate_task(rq, prev, 1);
3441 switch_count = &prev->nvcsw;
3444 if (unlikely(!rq->nr_running))
3445 idle_balance(cpu, rq);
3447 now = __rq_clock(rq);
3448 prev->sched_class->put_prev_task(rq, prev, now);
3449 next = pick_next_task(rq, prev, now);
3451 sched_info_switch(prev, next);
3453 if (likely(prev != next)) {
3458 context_switch(rq, prev, next); /* unlocks the rq */
3460 spin_unlock_irq(&rq->lock);
3462 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3463 cpu = smp_processor_id();
3465 goto need_resched_nonpreemptible;
3467 preempt_enable_no_resched();
3468 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3471 EXPORT_SYMBOL(schedule);
3473 #ifdef CONFIG_PREEMPT
3475 * this is the entry point to schedule() from in-kernel preemption
3476 * off of preempt_enable. Kernel preemptions off return from interrupt
3477 * occur there and call schedule directly.
3479 asmlinkage void __sched preempt_schedule(void)
3481 struct thread_info *ti = current_thread_info();
3482 #ifdef CONFIG_PREEMPT_BKL
3483 struct task_struct *task = current;
3484 int saved_lock_depth;
3487 * If there is a non-zero preempt_count or interrupts are disabled,
3488 * we do not want to preempt the current task. Just return..
3490 if (likely(ti->preempt_count || irqs_disabled()))
3494 add_preempt_count(PREEMPT_ACTIVE);
3496 * We keep the big kernel semaphore locked, but we
3497 * clear ->lock_depth so that schedule() doesnt
3498 * auto-release the semaphore:
3500 #ifdef CONFIG_PREEMPT_BKL
3501 saved_lock_depth = task->lock_depth;
3502 task->lock_depth = -1;
3505 #ifdef CONFIG_PREEMPT_BKL
3506 task->lock_depth = saved_lock_depth;
3508 sub_preempt_count(PREEMPT_ACTIVE);
3510 /* we could miss a preemption opportunity between schedule and now */
3512 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3515 EXPORT_SYMBOL(preempt_schedule);
3518 * this is the entry point to schedule() from kernel preemption
3519 * off of irq context.
3520 * Note, that this is called and return with irqs disabled. This will
3521 * protect us against recursive calling from irq.
3523 asmlinkage void __sched preempt_schedule_irq(void)
3525 struct thread_info *ti = current_thread_info();
3526 #ifdef CONFIG_PREEMPT_BKL
3527 struct task_struct *task = current;
3528 int saved_lock_depth;
3530 /* Catch callers which need to be fixed */
3531 BUG_ON(ti->preempt_count || !irqs_disabled());
3534 add_preempt_count(PREEMPT_ACTIVE);
3536 * We keep the big kernel semaphore locked, but we
3537 * clear ->lock_depth so that schedule() doesnt
3538 * auto-release the semaphore:
3540 #ifdef CONFIG_PREEMPT_BKL
3541 saved_lock_depth = task->lock_depth;
3542 task->lock_depth = -1;
3546 local_irq_disable();
3547 #ifdef CONFIG_PREEMPT_BKL
3548 task->lock_depth = saved_lock_depth;
3550 sub_preempt_count(PREEMPT_ACTIVE);
3552 /* we could miss a preemption opportunity between schedule and now */
3554 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3558 #endif /* CONFIG_PREEMPT */
3560 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3563 return try_to_wake_up(curr->private, mode, sync);
3565 EXPORT_SYMBOL(default_wake_function);
3568 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3569 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3570 * number) then we wake all the non-exclusive tasks and one exclusive task.
3572 * There are circumstances in which we can try to wake a task which has already
3573 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3574 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3576 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3577 int nr_exclusive, int sync, void *key)
3579 struct list_head *tmp, *next;
3581 list_for_each_safe(tmp, next, &q->task_list) {
3582 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3583 unsigned flags = curr->flags;
3585 if (curr->func(curr, mode, sync, key) &&
3586 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3592 * __wake_up - wake up threads blocked on a waitqueue.
3594 * @mode: which threads
3595 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3596 * @key: is directly passed to the wakeup function
3598 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3599 int nr_exclusive, void *key)
3601 unsigned long flags;
3603 spin_lock_irqsave(&q->lock, flags);
3604 __wake_up_common(q, mode, nr_exclusive, 0, key);
3605 spin_unlock_irqrestore(&q->lock, flags);
3607 EXPORT_SYMBOL(__wake_up);
3610 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3612 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3614 __wake_up_common(q, mode, 1, 0, NULL);
3618 * __wake_up_sync - wake up threads blocked on a waitqueue.
3620 * @mode: which threads
3621 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3623 * The sync wakeup differs that the waker knows that it will schedule
3624 * away soon, so while the target thread will be woken up, it will not
3625 * be migrated to another CPU - ie. the two threads are 'synchronized'
3626 * with each other. This can prevent needless bouncing between CPUs.
3628 * On UP it can prevent extra preemption.
3631 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3633 unsigned long flags;
3639 if (unlikely(!nr_exclusive))
3642 spin_lock_irqsave(&q->lock, flags);
3643 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3644 spin_unlock_irqrestore(&q->lock, flags);
3646 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3648 void fastcall complete(struct completion *x)
3650 unsigned long flags;
3652 spin_lock_irqsave(&x->wait.lock, flags);
3654 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3656 spin_unlock_irqrestore(&x->wait.lock, flags);
3658 EXPORT_SYMBOL(complete);
3660 void fastcall complete_all(struct completion *x)
3662 unsigned long flags;
3664 spin_lock_irqsave(&x->wait.lock, flags);
3665 x->done += UINT_MAX/2;
3666 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3668 spin_unlock_irqrestore(&x->wait.lock, flags);
3670 EXPORT_SYMBOL(complete_all);
3672 void fastcall __sched wait_for_completion(struct completion *x)
3676 spin_lock_irq(&x->wait.lock);
3678 DECLARE_WAITQUEUE(wait, current);
3680 wait.flags |= WQ_FLAG_EXCLUSIVE;
3681 __add_wait_queue_tail(&x->wait, &wait);
3683 __set_current_state(TASK_UNINTERRUPTIBLE);
3684 spin_unlock_irq(&x->wait.lock);
3686 spin_lock_irq(&x->wait.lock);
3688 __remove_wait_queue(&x->wait, &wait);
3691 spin_unlock_irq(&x->wait.lock);
3693 EXPORT_SYMBOL(wait_for_completion);
3695 unsigned long fastcall __sched
3696 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3700 spin_lock_irq(&x->wait.lock);
3702 DECLARE_WAITQUEUE(wait, current);
3704 wait.flags |= WQ_FLAG_EXCLUSIVE;
3705 __add_wait_queue_tail(&x->wait, &wait);
3707 __set_current_state(TASK_UNINTERRUPTIBLE);
3708 spin_unlock_irq(&x->wait.lock);
3709 timeout = schedule_timeout(timeout);
3710 spin_lock_irq(&x->wait.lock);
3712 __remove_wait_queue(&x->wait, &wait);
3716 __remove_wait_queue(&x->wait, &wait);
3720 spin_unlock_irq(&x->wait.lock);
3723 EXPORT_SYMBOL(wait_for_completion_timeout);
3725 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3731 spin_lock_irq(&x->wait.lock);
3733 DECLARE_WAITQUEUE(wait, current);
3735 wait.flags |= WQ_FLAG_EXCLUSIVE;
3736 __add_wait_queue_tail(&x->wait, &wait);
3738 if (signal_pending(current)) {
3740 __remove_wait_queue(&x->wait, &wait);
3743 __set_current_state(TASK_INTERRUPTIBLE);
3744 spin_unlock_irq(&x->wait.lock);
3746 spin_lock_irq(&x->wait.lock);
3748 __remove_wait_queue(&x->wait, &wait);
3752 spin_unlock_irq(&x->wait.lock);
3756 EXPORT_SYMBOL(wait_for_completion_interruptible);
3758 unsigned long fastcall __sched
3759 wait_for_completion_interruptible_timeout(struct completion *x,
3760 unsigned long timeout)
3764 spin_lock_irq(&x->wait.lock);
3766 DECLARE_WAITQUEUE(wait, current);
3768 wait.flags |= WQ_FLAG_EXCLUSIVE;
3769 __add_wait_queue_tail(&x->wait, &wait);
3771 if (signal_pending(current)) {
3772 timeout = -ERESTARTSYS;
3773 __remove_wait_queue(&x->wait, &wait);
3776 __set_current_state(TASK_INTERRUPTIBLE);
3777 spin_unlock_irq(&x->wait.lock);
3778 timeout = schedule_timeout(timeout);
3779 spin_lock_irq(&x->wait.lock);
3781 __remove_wait_queue(&x->wait, &wait);
3785 __remove_wait_queue(&x->wait, &wait);
3789 spin_unlock_irq(&x->wait.lock);
3792 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3795 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3797 spin_lock_irqsave(&q->lock, *flags);
3798 __add_wait_queue(q, wait);
3799 spin_unlock(&q->lock);
3803 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3805 spin_lock_irq(&q->lock);
3806 __remove_wait_queue(q, wait);
3807 spin_unlock_irqrestore(&q->lock, *flags);
3810 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3812 unsigned long flags;
3815 init_waitqueue_entry(&wait, current);
3817 current->state = TASK_INTERRUPTIBLE;
3819 sleep_on_head(q, &wait, &flags);
3821 sleep_on_tail(q, &wait, &flags);
3823 EXPORT_SYMBOL(interruptible_sleep_on);
3826 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3828 unsigned long flags;
3831 init_waitqueue_entry(&wait, current);
3833 current->state = TASK_INTERRUPTIBLE;
3835 sleep_on_head(q, &wait, &flags);
3836 timeout = schedule_timeout(timeout);
3837 sleep_on_tail(q, &wait, &flags);
3841 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3843 void __sched sleep_on(wait_queue_head_t *q)
3845 unsigned long flags;
3848 init_waitqueue_entry(&wait, current);
3850 current->state = TASK_UNINTERRUPTIBLE;
3852 sleep_on_head(q, &wait, &flags);
3854 sleep_on_tail(q, &wait, &flags);
3856 EXPORT_SYMBOL(sleep_on);
3858 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3860 unsigned long flags;
3863 init_waitqueue_entry(&wait, current);
3865 current->state = TASK_UNINTERRUPTIBLE;
3867 sleep_on_head(q, &wait, &flags);
3868 timeout = schedule_timeout(timeout);
3869 sleep_on_tail(q, &wait, &flags);
3873 EXPORT_SYMBOL(sleep_on_timeout);
3875 #ifdef CONFIG_RT_MUTEXES
3878 * rt_mutex_setprio - set the current priority of a task
3880 * @prio: prio value (kernel-internal form)
3882 * This function changes the 'effective' priority of a task. It does
3883 * not touch ->normal_prio like __setscheduler().
3885 * Used by the rt_mutex code to implement priority inheritance logic.
3887 void rt_mutex_setprio(struct task_struct *p, int prio)
3889 unsigned long flags;
3894 BUG_ON(prio < 0 || prio > MAX_PRIO);
3896 rq = task_rq_lock(p, &flags);
3900 on_rq = p->se.on_rq;
3902 dequeue_task(rq, p, 0, now);
3905 p->sched_class = &rt_sched_class;
3907 p->sched_class = &fair_sched_class;
3912 enqueue_task(rq, p, 0, now);
3914 * Reschedule if we are currently running on this runqueue and
3915 * our priority decreased, or if we are not currently running on
3916 * this runqueue and our priority is higher than the current's
3918 if (task_running(rq, p)) {
3919 if (p->prio > oldprio)
3920 resched_task(rq->curr);
3922 check_preempt_curr(rq, p);
3925 task_rq_unlock(rq, &flags);
3930 void set_user_nice(struct task_struct *p, long nice)
3932 int old_prio, delta, on_rq;
3933 unsigned long flags;
3937 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3940 * We have to be careful, if called from sys_setpriority(),
3941 * the task might be in the middle of scheduling on another CPU.
3943 rq = task_rq_lock(p, &flags);
3946 * The RT priorities are set via sched_setscheduler(), but we still
3947 * allow the 'normal' nice value to be set - but as expected
3948 * it wont have any effect on scheduling until the task is
3949 * SCHED_FIFO/SCHED_RR:
3951 if (task_has_rt_policy(p)) {
3952 p->static_prio = NICE_TO_PRIO(nice);
3955 on_rq = p->se.on_rq;
3957 dequeue_task(rq, p, 0, now);
3958 dec_load(rq, p, now);
3961 p->static_prio = NICE_TO_PRIO(nice);
3964 p->prio = effective_prio(p);
3965 delta = p->prio - old_prio;
3968 enqueue_task(rq, p, 0, now);
3969 inc_load(rq, p, now);
3971 * If the task increased its priority or is running and
3972 * lowered its priority, then reschedule its CPU:
3974 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3975 resched_task(rq->curr);
3978 task_rq_unlock(rq, &flags);
3980 EXPORT_SYMBOL(set_user_nice);
3983 * can_nice - check if a task can reduce its nice value
3987 int can_nice(const struct task_struct *p, const int nice)
3989 /* convert nice value [19,-20] to rlimit style value [1,40] */
3990 int nice_rlim = 20 - nice;
3992 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3993 capable(CAP_SYS_NICE));
3996 #ifdef __ARCH_WANT_SYS_NICE
3999 * sys_nice - change the priority of the current process.
4000 * @increment: priority increment
4002 * sys_setpriority is a more generic, but much slower function that
4003 * does similar things.
4005 asmlinkage long sys_nice(int increment)
4010 * Setpriority might change our priority at the same moment.
4011 * We don't have to worry. Conceptually one call occurs first
4012 * and we have a single winner.
4014 if (increment < -40)
4019 nice = PRIO_TO_NICE(current->static_prio) + increment;
4025 if (increment < 0 && !can_nice(current, nice))
4028 retval = security_task_setnice(current, nice);
4032 set_user_nice(current, nice);
4039 * task_prio - return the priority value of a given task.
4040 * @p: the task in question.
4042 * This is the priority value as seen by users in /proc.
4043 * RT tasks are offset by -200. Normal tasks are centered
4044 * around 0, value goes from -16 to +15.
4046 int task_prio(const struct task_struct *p)
4048 return p->prio - MAX_RT_PRIO;
4052 * task_nice - return the nice value of a given task.
4053 * @p: the task in question.
4055 int task_nice(const struct task_struct *p)
4057 return TASK_NICE(p);
4059 EXPORT_SYMBOL_GPL(task_nice);
4062 * idle_cpu - is a given cpu idle currently?
4063 * @cpu: the processor in question.
4065 int idle_cpu(int cpu)
4067 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4071 * idle_task - return the idle task for a given cpu.
4072 * @cpu: the processor in question.
4074 struct task_struct *idle_task(int cpu)
4076 return cpu_rq(cpu)->idle;
4080 * find_process_by_pid - find a process with a matching PID value.
4081 * @pid: the pid in question.
4083 static inline struct task_struct *find_process_by_pid(pid_t pid)
4085 return pid ? find_task_by_pid(pid) : current;
4088 /* Actually do priority change: must hold rq lock. */
4090 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4092 BUG_ON(p->se.on_rq);
4095 switch (p->policy) {
4099 p->sched_class = &fair_sched_class;
4103 p->sched_class = &rt_sched_class;
4107 p->rt_priority = prio;
4108 p->normal_prio = normal_prio(p);
4109 /* we are holding p->pi_lock already */
4110 p->prio = rt_mutex_getprio(p);
4115 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4116 * @p: the task in question.
4117 * @policy: new policy.
4118 * @param: structure containing the new RT priority.
4120 * NOTE that the task may be already dead.
4122 int sched_setscheduler(struct task_struct *p, int policy,
4123 struct sched_param *param)
4125 int retval, oldprio, oldpolicy = -1, on_rq;
4126 unsigned long flags;
4129 /* may grab non-irq protected spin_locks */
4130 BUG_ON(in_interrupt());
4132 /* double check policy once rq lock held */
4134 policy = oldpolicy = p->policy;
4135 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4136 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4137 policy != SCHED_IDLE)
4140 * Valid priorities for SCHED_FIFO and SCHED_RR are
4141 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4142 * SCHED_BATCH and SCHED_IDLE is 0.
4144 if (param->sched_priority < 0 ||
4145 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4146 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4148 if (rt_policy(policy) != (param->sched_priority != 0))
4152 * Allow unprivileged RT tasks to decrease priority:
4154 if (!capable(CAP_SYS_NICE)) {
4155 if (rt_policy(policy)) {
4156 unsigned long rlim_rtprio;
4158 if (!lock_task_sighand(p, &flags))
4160 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4161 unlock_task_sighand(p, &flags);
4163 /* can't set/change the rt policy */
4164 if (policy != p->policy && !rlim_rtprio)
4167 /* can't increase priority */
4168 if (param->sched_priority > p->rt_priority &&
4169 param->sched_priority > rlim_rtprio)
4173 * Like positive nice levels, dont allow tasks to
4174 * move out of SCHED_IDLE either:
4176 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4179 /* can't change other user's priorities */
4180 if ((current->euid != p->euid) &&
4181 (current->euid != p->uid))
4185 retval = security_task_setscheduler(p, policy, param);
4189 * make sure no PI-waiters arrive (or leave) while we are
4190 * changing the priority of the task:
4192 spin_lock_irqsave(&p->pi_lock, flags);
4194 * To be able to change p->policy safely, the apropriate
4195 * runqueue lock must be held.
4197 rq = __task_rq_lock(p);
4198 /* recheck policy now with rq lock held */
4199 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4200 policy = oldpolicy = -1;
4201 __task_rq_unlock(rq);
4202 spin_unlock_irqrestore(&p->pi_lock, flags);
4205 on_rq = p->se.on_rq;
4207 deactivate_task(rq, p, 0);
4209 __setscheduler(rq, p, policy, param->sched_priority);
4211 activate_task(rq, p, 0);
4213 * Reschedule if we are currently running on this runqueue and
4214 * our priority decreased, or if we are not currently running on
4215 * this runqueue and our priority is higher than the current's
4217 if (task_running(rq, p)) {
4218 if (p->prio > oldprio)
4219 resched_task(rq->curr);
4221 check_preempt_curr(rq, p);
4224 __task_rq_unlock(rq);
4225 spin_unlock_irqrestore(&p->pi_lock, flags);
4227 rt_mutex_adjust_pi(p);
4231 EXPORT_SYMBOL_GPL(sched_setscheduler);
4234 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4236 struct sched_param lparam;
4237 struct task_struct *p;
4240 if (!param || pid < 0)
4242 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4247 p = find_process_by_pid(pid);
4249 retval = sched_setscheduler(p, policy, &lparam);
4256 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4257 * @pid: the pid in question.
4258 * @policy: new policy.
4259 * @param: structure containing the new RT priority.
4261 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4262 struct sched_param __user *param)
4264 /* negative values for policy are not valid */
4268 return do_sched_setscheduler(pid, policy, param);
4272 * sys_sched_setparam - set/change the RT priority of a thread
4273 * @pid: the pid in question.
4274 * @param: structure containing the new RT priority.
4276 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4278 return do_sched_setscheduler(pid, -1, param);
4282 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4283 * @pid: the pid in question.
4285 asmlinkage long sys_sched_getscheduler(pid_t pid)
4287 struct task_struct *p;
4288 int retval = -EINVAL;
4294 read_lock(&tasklist_lock);
4295 p = find_process_by_pid(pid);
4297 retval = security_task_getscheduler(p);
4301 read_unlock(&tasklist_lock);
4308 * sys_sched_getscheduler - get the RT priority of a thread
4309 * @pid: the pid in question.
4310 * @param: structure containing the RT priority.
4312 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4314 struct sched_param lp;
4315 struct task_struct *p;
4316 int retval = -EINVAL;
4318 if (!param || pid < 0)
4321 read_lock(&tasklist_lock);
4322 p = find_process_by_pid(pid);
4327 retval = security_task_getscheduler(p);
4331 lp.sched_priority = p->rt_priority;
4332 read_unlock(&tasklist_lock);
4335 * This one might sleep, we cannot do it with a spinlock held ...
4337 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4343 read_unlock(&tasklist_lock);
4347 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4349 cpumask_t cpus_allowed;
4350 struct task_struct *p;
4353 mutex_lock(&sched_hotcpu_mutex);
4354 read_lock(&tasklist_lock);
4356 p = find_process_by_pid(pid);
4358 read_unlock(&tasklist_lock);
4359 mutex_unlock(&sched_hotcpu_mutex);
4364 * It is not safe to call set_cpus_allowed with the
4365 * tasklist_lock held. We will bump the task_struct's
4366 * usage count and then drop tasklist_lock.
4369 read_unlock(&tasklist_lock);
4372 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4373 !capable(CAP_SYS_NICE))
4376 retval = security_task_setscheduler(p, 0, NULL);
4380 cpus_allowed = cpuset_cpus_allowed(p);
4381 cpus_and(new_mask, new_mask, cpus_allowed);
4382 retval = set_cpus_allowed(p, new_mask);
4386 mutex_unlock(&sched_hotcpu_mutex);
4390 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4391 cpumask_t *new_mask)
4393 if (len < sizeof(cpumask_t)) {
4394 memset(new_mask, 0, sizeof(cpumask_t));
4395 } else if (len > sizeof(cpumask_t)) {
4396 len = sizeof(cpumask_t);
4398 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4402 * sys_sched_setaffinity - set the cpu affinity of a process
4403 * @pid: pid of the process
4404 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4405 * @user_mask_ptr: user-space pointer to the new cpu mask
4407 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4408 unsigned long __user *user_mask_ptr)
4413 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4417 return sched_setaffinity(pid, new_mask);
4421 * Represents all cpu's present in the system
4422 * In systems capable of hotplug, this map could dynamically grow
4423 * as new cpu's are detected in the system via any platform specific
4424 * method, such as ACPI for e.g.
4427 cpumask_t cpu_present_map __read_mostly;
4428 EXPORT_SYMBOL(cpu_present_map);
4431 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4432 EXPORT_SYMBOL(cpu_online_map);
4434 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4435 EXPORT_SYMBOL(cpu_possible_map);
4438 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4440 struct task_struct *p;
4443 mutex_lock(&sched_hotcpu_mutex);
4444 read_lock(&tasklist_lock);
4447 p = find_process_by_pid(pid);
4451 retval = security_task_getscheduler(p);
4455 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4458 read_unlock(&tasklist_lock);
4459 mutex_unlock(&sched_hotcpu_mutex);
4467 * sys_sched_getaffinity - get the cpu affinity of a process
4468 * @pid: pid of the process
4469 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4470 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4472 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4473 unsigned long __user *user_mask_ptr)
4478 if (len < sizeof(cpumask_t))
4481 ret = sched_getaffinity(pid, &mask);
4485 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4488 return sizeof(cpumask_t);
4492 * sys_sched_yield - yield the current processor to other threads.
4494 * This function yields the current CPU to other tasks. If there are no
4495 * other threads running on this CPU then this function will return.
4497 asmlinkage long sys_sched_yield(void)
4499 struct rq *rq = this_rq_lock();
4501 schedstat_inc(rq, yld_cnt);
4502 if (unlikely(rq->nr_running == 1))
4503 schedstat_inc(rq, yld_act_empty);
4505 current->sched_class->yield_task(rq, current);
4508 * Since we are going to call schedule() anyway, there's
4509 * no need to preempt or enable interrupts:
4511 __release(rq->lock);
4512 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4513 _raw_spin_unlock(&rq->lock);
4514 preempt_enable_no_resched();
4521 static void __cond_resched(void)
4523 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4524 __might_sleep(__FILE__, __LINE__);
4527 * The BKS might be reacquired before we have dropped
4528 * PREEMPT_ACTIVE, which could trigger a second
4529 * cond_resched() call.
4532 add_preempt_count(PREEMPT_ACTIVE);
4534 sub_preempt_count(PREEMPT_ACTIVE);
4535 } while (need_resched());
4538 int __sched cond_resched(void)
4540 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4541 system_state == SYSTEM_RUNNING) {
4547 EXPORT_SYMBOL(cond_resched);
4550 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4551 * call schedule, and on return reacquire the lock.
4553 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4554 * operations here to prevent schedule() from being called twice (once via
4555 * spin_unlock(), once by hand).
4557 int cond_resched_lock(spinlock_t *lock)
4561 if (need_lockbreak(lock)) {
4567 if (need_resched() && system_state == SYSTEM_RUNNING) {
4568 spin_release(&lock->dep_map, 1, _THIS_IP_);
4569 _raw_spin_unlock(lock);
4570 preempt_enable_no_resched();
4577 EXPORT_SYMBOL(cond_resched_lock);
4579 int __sched cond_resched_softirq(void)
4581 BUG_ON(!in_softirq());
4583 if (need_resched() && system_state == SYSTEM_RUNNING) {
4591 EXPORT_SYMBOL(cond_resched_softirq);
4594 * yield - yield the current processor to other threads.
4596 * This is a shortcut for kernel-space yielding - it marks the
4597 * thread runnable and calls sys_sched_yield().
4599 void __sched yield(void)
4601 set_current_state(TASK_RUNNING);
4604 EXPORT_SYMBOL(yield);
4607 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4608 * that process accounting knows that this is a task in IO wait state.
4610 * But don't do that if it is a deliberate, throttling IO wait (this task
4611 * has set its backing_dev_info: the queue against which it should throttle)
4613 void __sched io_schedule(void)
4615 struct rq *rq = &__raw_get_cpu_var(runqueues);
4617 delayacct_blkio_start();
4618 atomic_inc(&rq->nr_iowait);
4620 atomic_dec(&rq->nr_iowait);
4621 delayacct_blkio_end();
4623 EXPORT_SYMBOL(io_schedule);
4625 long __sched io_schedule_timeout(long timeout)
4627 struct rq *rq = &__raw_get_cpu_var(runqueues);
4630 delayacct_blkio_start();
4631 atomic_inc(&rq->nr_iowait);
4632 ret = schedule_timeout(timeout);
4633 atomic_dec(&rq->nr_iowait);
4634 delayacct_blkio_end();
4639 * sys_sched_get_priority_max - return maximum RT priority.
4640 * @policy: scheduling class.
4642 * this syscall returns the maximum rt_priority that can be used
4643 * by a given scheduling class.
4645 asmlinkage long sys_sched_get_priority_max(int policy)
4652 ret = MAX_USER_RT_PRIO-1;
4664 * sys_sched_get_priority_min - return minimum RT priority.
4665 * @policy: scheduling class.
4667 * this syscall returns the minimum rt_priority that can be used
4668 * by a given scheduling class.
4670 asmlinkage long sys_sched_get_priority_min(int policy)
4688 * sys_sched_rr_get_interval - return the default timeslice of a process.
4689 * @pid: pid of the process.
4690 * @interval: userspace pointer to the timeslice value.
4692 * this syscall writes the default timeslice value of a given process
4693 * into the user-space timespec buffer. A value of '0' means infinity.
4696 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4698 struct task_struct *p;
4699 int retval = -EINVAL;
4706 read_lock(&tasklist_lock);
4707 p = find_process_by_pid(pid);
4711 retval = security_task_getscheduler(p);
4715 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4716 0 : static_prio_timeslice(p->static_prio), &t);
4717 read_unlock(&tasklist_lock);
4718 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4722 read_unlock(&tasklist_lock);
4726 static const char stat_nam[] = "RSDTtZX";
4728 static void show_task(struct task_struct *p)
4730 unsigned long free = 0;
4733 state = p->state ? __ffs(p->state) + 1 : 0;
4734 printk("%-13.13s %c", p->comm,
4735 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4736 #if BITS_PER_LONG == 32
4737 if (state == TASK_RUNNING)
4738 printk(" running ");
4740 printk(" %08lx ", thread_saved_pc(p));
4742 if (state == TASK_RUNNING)
4743 printk(" running task ");
4745 printk(" %016lx ", thread_saved_pc(p));
4747 #ifdef CONFIG_DEBUG_STACK_USAGE
4749 unsigned long *n = end_of_stack(p);
4752 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4755 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4757 if (state != TASK_RUNNING)
4758 show_stack(p, NULL);
4761 void show_state_filter(unsigned long state_filter)
4763 struct task_struct *g, *p;
4765 #if BITS_PER_LONG == 32
4767 " task PC stack pid father\n");
4770 " task PC stack pid father\n");
4772 read_lock(&tasklist_lock);
4773 do_each_thread(g, p) {
4775 * reset the NMI-timeout, listing all files on a slow
4776 * console might take alot of time:
4778 touch_nmi_watchdog();
4779 if (!state_filter || (p->state & state_filter))
4781 } while_each_thread(g, p);
4783 touch_all_softlockup_watchdogs();
4785 #ifdef CONFIG_SCHED_DEBUG
4786 sysrq_sched_debug_show();
4788 read_unlock(&tasklist_lock);
4790 * Only show locks if all tasks are dumped:
4792 if (state_filter == -1)
4793 debug_show_all_locks();
4796 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4798 idle->sched_class = &idle_sched_class;
4802 * init_idle - set up an idle thread for a given CPU
4803 * @idle: task in question
4804 * @cpu: cpu the idle task belongs to
4806 * NOTE: this function does not set the idle thread's NEED_RESCHED
4807 * flag, to make booting more robust.
4809 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4811 struct rq *rq = cpu_rq(cpu);
4812 unsigned long flags;
4815 idle->se.exec_start = sched_clock();
4817 idle->prio = idle->normal_prio = MAX_PRIO;
4818 idle->cpus_allowed = cpumask_of_cpu(cpu);
4819 __set_task_cpu(idle, cpu);
4821 spin_lock_irqsave(&rq->lock, flags);
4822 rq->curr = rq->idle = idle;
4823 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4826 spin_unlock_irqrestore(&rq->lock, flags);
4828 /* Set the preempt count _outside_ the spinlocks! */
4829 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4830 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4832 task_thread_info(idle)->preempt_count = 0;
4835 * The idle tasks have their own, simple scheduling class:
4837 idle->sched_class = &idle_sched_class;
4841 * In a system that switches off the HZ timer nohz_cpu_mask
4842 * indicates which cpus entered this state. This is used
4843 * in the rcu update to wait only for active cpus. For system
4844 * which do not switch off the HZ timer nohz_cpu_mask should
4845 * always be CPU_MASK_NONE.
4847 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4850 * Increase the granularity value when there are more CPUs,
4851 * because with more CPUs the 'effective latency' as visible
4852 * to users decreases. But the relationship is not linear,
4853 * so pick a second-best guess by going with the log2 of the
4856 * This idea comes from the SD scheduler of Con Kolivas:
4858 static inline void sched_init_granularity(void)
4860 unsigned int factor = 1 + ilog2(num_online_cpus());
4861 const unsigned long gran_limit = 100000000;
4863 sysctl_sched_granularity *= factor;
4864 if (sysctl_sched_granularity > gran_limit)
4865 sysctl_sched_granularity = gran_limit;
4867 sysctl_sched_runtime_limit = sysctl_sched_granularity * 4;
4868 sysctl_sched_wakeup_granularity = sysctl_sched_granularity / 2;
4873 * This is how migration works:
4875 * 1) we queue a struct migration_req structure in the source CPU's
4876 * runqueue and wake up that CPU's migration thread.
4877 * 2) we down() the locked semaphore => thread blocks.
4878 * 3) migration thread wakes up (implicitly it forces the migrated
4879 * thread off the CPU)
4880 * 4) it gets the migration request and checks whether the migrated
4881 * task is still in the wrong runqueue.
4882 * 5) if it's in the wrong runqueue then the migration thread removes
4883 * it and puts it into the right queue.
4884 * 6) migration thread up()s the semaphore.
4885 * 7) we wake up and the migration is done.
4889 * Change a given task's CPU affinity. Migrate the thread to a
4890 * proper CPU and schedule it away if the CPU it's executing on
4891 * is removed from the allowed bitmask.
4893 * NOTE: the caller must have a valid reference to the task, the
4894 * task must not exit() & deallocate itself prematurely. The
4895 * call is not atomic; no spinlocks may be held.
4897 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4899 struct migration_req req;
4900 unsigned long flags;
4904 rq = task_rq_lock(p, &flags);
4905 if (!cpus_intersects(new_mask, cpu_online_map)) {
4910 p->cpus_allowed = new_mask;
4911 /* Can the task run on the task's current CPU? If so, we're done */
4912 if (cpu_isset(task_cpu(p), new_mask))
4915 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4916 /* Need help from migration thread: drop lock and wait. */
4917 task_rq_unlock(rq, &flags);
4918 wake_up_process(rq->migration_thread);
4919 wait_for_completion(&req.done);
4920 tlb_migrate_finish(p->mm);
4924 task_rq_unlock(rq, &flags);
4928 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4931 * Move (not current) task off this cpu, onto dest cpu. We're doing
4932 * this because either it can't run here any more (set_cpus_allowed()
4933 * away from this CPU, or CPU going down), or because we're
4934 * attempting to rebalance this task on exec (sched_exec).
4936 * So we race with normal scheduler movements, but that's OK, as long
4937 * as the task is no longer on this CPU.
4939 * Returns non-zero if task was successfully migrated.
4941 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4943 struct rq *rq_dest, *rq_src;
4946 if (unlikely(cpu_is_offline(dest_cpu)))
4949 rq_src = cpu_rq(src_cpu);
4950 rq_dest = cpu_rq(dest_cpu);
4952 double_rq_lock(rq_src, rq_dest);
4953 /* Already moved. */
4954 if (task_cpu(p) != src_cpu)
4956 /* Affinity changed (again). */
4957 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4960 on_rq = p->se.on_rq;
4962 deactivate_task(rq_src, p, 0);
4963 set_task_cpu(p, dest_cpu);
4965 activate_task(rq_dest, p, 0);
4966 check_preempt_curr(rq_dest, p);
4970 double_rq_unlock(rq_src, rq_dest);
4975 * migration_thread - this is a highprio system thread that performs
4976 * thread migration by bumping thread off CPU then 'pushing' onto
4979 static int migration_thread(void *data)
4981 int cpu = (long)data;
4985 BUG_ON(rq->migration_thread != current);
4987 set_current_state(TASK_INTERRUPTIBLE);
4988 while (!kthread_should_stop()) {
4989 struct migration_req *req;
4990 struct list_head *head;
4992 spin_lock_irq(&rq->lock);
4994 if (cpu_is_offline(cpu)) {
4995 spin_unlock_irq(&rq->lock);
4999 if (rq->active_balance) {
5000 active_load_balance(rq, cpu);
5001 rq->active_balance = 0;
5004 head = &rq->migration_queue;
5006 if (list_empty(head)) {
5007 spin_unlock_irq(&rq->lock);
5009 set_current_state(TASK_INTERRUPTIBLE);
5012 req = list_entry(head->next, struct migration_req, list);
5013 list_del_init(head->next);
5015 spin_unlock(&rq->lock);
5016 __migrate_task(req->task, cpu, req->dest_cpu);
5019 complete(&req->done);
5021 __set_current_state(TASK_RUNNING);
5025 /* Wait for kthread_stop */
5026 set_current_state(TASK_INTERRUPTIBLE);
5027 while (!kthread_should_stop()) {
5029 set_current_state(TASK_INTERRUPTIBLE);
5031 __set_current_state(TASK_RUNNING);
5035 #ifdef CONFIG_HOTPLUG_CPU
5037 * Figure out where task on dead CPU should go, use force if neccessary.
5038 * NOTE: interrupts should be disabled by the caller
5040 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5042 unsigned long flags;
5049 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5050 cpus_and(mask, mask, p->cpus_allowed);
5051 dest_cpu = any_online_cpu(mask);
5053 /* On any allowed CPU? */
5054 if (dest_cpu == NR_CPUS)
5055 dest_cpu = any_online_cpu(p->cpus_allowed);
5057 /* No more Mr. Nice Guy. */
5058 if (dest_cpu == NR_CPUS) {
5059 rq = task_rq_lock(p, &flags);
5060 cpus_setall(p->cpus_allowed);
5061 dest_cpu = any_online_cpu(p->cpus_allowed);
5062 task_rq_unlock(rq, &flags);
5065 * Don't tell them about moving exiting tasks or
5066 * kernel threads (both mm NULL), since they never
5069 if (p->mm && printk_ratelimit())
5070 printk(KERN_INFO "process %d (%s) no "
5071 "longer affine to cpu%d\n",
5072 p->pid, p->comm, dead_cpu);
5074 if (!__migrate_task(p, dead_cpu, dest_cpu))
5079 * While a dead CPU has no uninterruptible tasks queued at this point,
5080 * it might still have a nonzero ->nr_uninterruptible counter, because
5081 * for performance reasons the counter is not stricly tracking tasks to
5082 * their home CPUs. So we just add the counter to another CPU's counter,
5083 * to keep the global sum constant after CPU-down:
5085 static void migrate_nr_uninterruptible(struct rq *rq_src)
5087 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5088 unsigned long flags;
5090 local_irq_save(flags);
5091 double_rq_lock(rq_src, rq_dest);
5092 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5093 rq_src->nr_uninterruptible = 0;
5094 double_rq_unlock(rq_src, rq_dest);
5095 local_irq_restore(flags);
5098 /* Run through task list and migrate tasks from the dead cpu. */
5099 static void migrate_live_tasks(int src_cpu)
5101 struct task_struct *p, *t;
5103 write_lock_irq(&tasklist_lock);
5105 do_each_thread(t, p) {
5109 if (task_cpu(p) == src_cpu)
5110 move_task_off_dead_cpu(src_cpu, p);
5111 } while_each_thread(t, p);
5113 write_unlock_irq(&tasklist_lock);
5117 * Schedules idle task to be the next runnable task on current CPU.
5118 * It does so by boosting its priority to highest possible and adding it to
5119 * the _front_ of the runqueue. Used by CPU offline code.
5121 void sched_idle_next(void)
5123 int this_cpu = smp_processor_id();
5124 struct rq *rq = cpu_rq(this_cpu);
5125 struct task_struct *p = rq->idle;
5126 unsigned long flags;
5128 /* cpu has to be offline */
5129 BUG_ON(cpu_online(this_cpu));
5132 * Strictly not necessary since rest of the CPUs are stopped by now
5133 * and interrupts disabled on the current cpu.
5135 spin_lock_irqsave(&rq->lock, flags);
5137 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5139 /* Add idle task to the _front_ of its priority queue: */
5140 activate_idle_task(p, rq);
5142 spin_unlock_irqrestore(&rq->lock, flags);
5146 * Ensures that the idle task is using init_mm right before its cpu goes
5149 void idle_task_exit(void)
5151 struct mm_struct *mm = current->active_mm;
5153 BUG_ON(cpu_online(smp_processor_id()));
5156 switch_mm(mm, &init_mm, current);
5160 /* called under rq->lock with disabled interrupts */
5161 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5163 struct rq *rq = cpu_rq(dead_cpu);
5165 /* Must be exiting, otherwise would be on tasklist. */
5166 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5168 /* Cannot have done final schedule yet: would have vanished. */
5169 BUG_ON(p->state == TASK_DEAD);
5174 * Drop lock around migration; if someone else moves it,
5175 * that's OK. No task can be added to this CPU, so iteration is
5177 * NOTE: interrupts should be left disabled --dev@
5179 spin_unlock(&rq->lock);
5180 move_task_off_dead_cpu(dead_cpu, p);
5181 spin_lock(&rq->lock);
5186 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5187 static void migrate_dead_tasks(unsigned int dead_cpu)
5189 struct rq *rq = cpu_rq(dead_cpu);
5190 struct task_struct *next;
5193 if (!rq->nr_running)
5195 next = pick_next_task(rq, rq->curr, rq_clock(rq));
5198 migrate_dead(dead_cpu, next);
5202 #endif /* CONFIG_HOTPLUG_CPU */
5204 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5206 static struct ctl_table sd_ctl_dir[] = {
5207 {CTL_UNNUMBERED, "sched_domain", NULL, 0, 0755, NULL, },
5211 static struct ctl_table sd_ctl_root[] = {
5212 {CTL_UNNUMBERED, "kernel", NULL, 0, 0755, sd_ctl_dir, },
5216 static struct ctl_table *sd_alloc_ctl_entry(int n)
5218 struct ctl_table *entry =
5219 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5222 memset(entry, 0, n * sizeof(struct ctl_table));
5228 set_table_entry(struct ctl_table *entry, int ctl_name,
5229 const char *procname, void *data, int maxlen,
5230 mode_t mode, proc_handler *proc_handler)
5232 entry->ctl_name = ctl_name;
5233 entry->procname = procname;
5235 entry->maxlen = maxlen;
5237 entry->proc_handler = proc_handler;
5240 static struct ctl_table *
5241 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5243 struct ctl_table *table = sd_alloc_ctl_entry(14);
5245 set_table_entry(&table[0], 1, "min_interval", &sd->min_interval,
5246 sizeof(long), 0644, proc_doulongvec_minmax);
5247 set_table_entry(&table[1], 2, "max_interval", &sd->max_interval,
5248 sizeof(long), 0644, proc_doulongvec_minmax);
5249 set_table_entry(&table[2], 3, "busy_idx", &sd->busy_idx,
5250 sizeof(int), 0644, proc_dointvec_minmax);
5251 set_table_entry(&table[3], 4, "idle_idx", &sd->idle_idx,
5252 sizeof(int), 0644, proc_dointvec_minmax);
5253 set_table_entry(&table[4], 5, "newidle_idx", &sd->newidle_idx,
5254 sizeof(int), 0644, proc_dointvec_minmax);
5255 set_table_entry(&table[5], 6, "wake_idx", &sd->wake_idx,
5256 sizeof(int), 0644, proc_dointvec_minmax);
5257 set_table_entry(&table[6], 7, "forkexec_idx", &sd->forkexec_idx,
5258 sizeof(int), 0644, proc_dointvec_minmax);
5259 set_table_entry(&table[7], 8, "busy_factor", &sd->busy_factor,
5260 sizeof(int), 0644, proc_dointvec_minmax);
5261 set_table_entry(&table[8], 9, "imbalance_pct", &sd->imbalance_pct,
5262 sizeof(int), 0644, proc_dointvec_minmax);
5263 set_table_entry(&table[10], 11, "cache_nice_tries",
5264 &sd->cache_nice_tries,
5265 sizeof(int), 0644, proc_dointvec_minmax);
5266 set_table_entry(&table[12], 13, "flags", &sd->flags,
5267 sizeof(int), 0644, proc_dointvec_minmax);
5272 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5274 struct ctl_table *entry, *table;
5275 struct sched_domain *sd;
5276 int domain_num = 0, i;
5279 for_each_domain(cpu, sd)
5281 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5284 for_each_domain(cpu, sd) {
5285 snprintf(buf, 32, "domain%d", i);
5286 entry->ctl_name = i + 1;
5287 entry->procname = kstrdup(buf, GFP_KERNEL);
5289 entry->child = sd_alloc_ctl_domain_table(sd);
5296 static struct ctl_table_header *sd_sysctl_header;
5297 static void init_sched_domain_sysctl(void)
5299 int i, cpu_num = num_online_cpus();
5300 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5303 sd_ctl_dir[0].child = entry;
5305 for (i = 0; i < cpu_num; i++, entry++) {
5306 snprintf(buf, 32, "cpu%d", i);
5307 entry->ctl_name = i + 1;
5308 entry->procname = kstrdup(buf, GFP_KERNEL);
5310 entry->child = sd_alloc_ctl_cpu_table(i);
5312 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5315 static void init_sched_domain_sysctl(void)
5321 * migration_call - callback that gets triggered when a CPU is added.
5322 * Here we can start up the necessary migration thread for the new CPU.
5324 static int __cpuinit
5325 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5327 struct task_struct *p;
5328 int cpu = (long)hcpu;
5329 unsigned long flags;
5333 case CPU_LOCK_ACQUIRE:
5334 mutex_lock(&sched_hotcpu_mutex);
5337 case CPU_UP_PREPARE:
5338 case CPU_UP_PREPARE_FROZEN:
5339 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5342 kthread_bind(p, cpu);
5343 /* Must be high prio: stop_machine expects to yield to it. */
5344 rq = task_rq_lock(p, &flags);
5345 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5346 task_rq_unlock(rq, &flags);
5347 cpu_rq(cpu)->migration_thread = p;
5351 case CPU_ONLINE_FROZEN:
5352 /* Strictly unneccessary, as first user will wake it. */
5353 wake_up_process(cpu_rq(cpu)->migration_thread);
5356 #ifdef CONFIG_HOTPLUG_CPU
5357 case CPU_UP_CANCELED:
5358 case CPU_UP_CANCELED_FROZEN:
5359 if (!cpu_rq(cpu)->migration_thread)
5361 /* Unbind it from offline cpu so it can run. Fall thru. */
5362 kthread_bind(cpu_rq(cpu)->migration_thread,
5363 any_online_cpu(cpu_online_map));
5364 kthread_stop(cpu_rq(cpu)->migration_thread);
5365 cpu_rq(cpu)->migration_thread = NULL;
5369 case CPU_DEAD_FROZEN:
5370 migrate_live_tasks(cpu);
5372 kthread_stop(rq->migration_thread);
5373 rq->migration_thread = NULL;
5374 /* Idle task back to normal (off runqueue, low prio) */
5375 rq = task_rq_lock(rq->idle, &flags);
5376 deactivate_task(rq, rq->idle, 0);
5377 rq->idle->static_prio = MAX_PRIO;
5378 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5379 rq->idle->sched_class = &idle_sched_class;
5380 migrate_dead_tasks(cpu);
5381 task_rq_unlock(rq, &flags);
5382 migrate_nr_uninterruptible(rq);
5383 BUG_ON(rq->nr_running != 0);
5385 /* No need to migrate the tasks: it was best-effort if
5386 * they didn't take sched_hotcpu_mutex. Just wake up
5387 * the requestors. */
5388 spin_lock_irq(&rq->lock);
5389 while (!list_empty(&rq->migration_queue)) {
5390 struct migration_req *req;
5392 req = list_entry(rq->migration_queue.next,
5393 struct migration_req, list);
5394 list_del_init(&req->list);
5395 complete(&req->done);
5397 spin_unlock_irq(&rq->lock);
5400 case CPU_LOCK_RELEASE:
5401 mutex_unlock(&sched_hotcpu_mutex);
5407 /* Register at highest priority so that task migration (migrate_all_tasks)
5408 * happens before everything else.
5410 static struct notifier_block __cpuinitdata migration_notifier = {
5411 .notifier_call = migration_call,
5415 int __init migration_init(void)
5417 void *cpu = (void *)(long)smp_processor_id();
5420 /* Start one for the boot CPU: */
5421 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5422 BUG_ON(err == NOTIFY_BAD);
5423 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5424 register_cpu_notifier(&migration_notifier);
5432 /* Number of possible processor ids */
5433 int nr_cpu_ids __read_mostly = NR_CPUS;
5434 EXPORT_SYMBOL(nr_cpu_ids);
5436 #undef SCHED_DOMAIN_DEBUG
5437 #ifdef SCHED_DOMAIN_DEBUG
5438 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5443 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5447 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5452 struct sched_group *group = sd->groups;
5453 cpumask_t groupmask;
5455 cpumask_scnprintf(str, NR_CPUS, sd->span);
5456 cpus_clear(groupmask);
5459 for (i = 0; i < level + 1; i++)
5461 printk("domain %d: ", level);
5463 if (!(sd->flags & SD_LOAD_BALANCE)) {
5464 printk("does not load-balance\n");
5466 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5471 printk("span %s\n", str);
5473 if (!cpu_isset(cpu, sd->span))
5474 printk(KERN_ERR "ERROR: domain->span does not contain "
5476 if (!cpu_isset(cpu, group->cpumask))
5477 printk(KERN_ERR "ERROR: domain->groups does not contain"
5481 for (i = 0; i < level + 2; i++)
5487 printk(KERN_ERR "ERROR: group is NULL\n");
5491 if (!group->__cpu_power) {
5493 printk(KERN_ERR "ERROR: domain->cpu_power not "
5497 if (!cpus_weight(group->cpumask)) {
5499 printk(KERN_ERR "ERROR: empty group\n");
5502 if (cpus_intersects(groupmask, group->cpumask)) {
5504 printk(KERN_ERR "ERROR: repeated CPUs\n");
5507 cpus_or(groupmask, groupmask, group->cpumask);
5509 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5512 group = group->next;
5513 } while (group != sd->groups);
5516 if (!cpus_equal(sd->span, groupmask))
5517 printk(KERN_ERR "ERROR: groups don't span "
5525 if (!cpus_subset(groupmask, sd->span))
5526 printk(KERN_ERR "ERROR: parent span is not a superset "
5527 "of domain->span\n");
5532 # define sched_domain_debug(sd, cpu) do { } while (0)
5535 static int sd_degenerate(struct sched_domain *sd)
5537 if (cpus_weight(sd->span) == 1)
5540 /* Following flags need at least 2 groups */
5541 if (sd->flags & (SD_LOAD_BALANCE |
5542 SD_BALANCE_NEWIDLE |
5546 SD_SHARE_PKG_RESOURCES)) {
5547 if (sd->groups != sd->groups->next)
5551 /* Following flags don't use groups */
5552 if (sd->flags & (SD_WAKE_IDLE |
5561 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5563 unsigned long cflags = sd->flags, pflags = parent->flags;
5565 if (sd_degenerate(parent))
5568 if (!cpus_equal(sd->span, parent->span))
5571 /* Does parent contain flags not in child? */
5572 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5573 if (cflags & SD_WAKE_AFFINE)
5574 pflags &= ~SD_WAKE_BALANCE;
5575 /* Flags needing groups don't count if only 1 group in parent */
5576 if (parent->groups == parent->groups->next) {
5577 pflags &= ~(SD_LOAD_BALANCE |
5578 SD_BALANCE_NEWIDLE |
5582 SD_SHARE_PKG_RESOURCES);
5584 if (~cflags & pflags)
5591 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5592 * hold the hotplug lock.
5594 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5596 struct rq *rq = cpu_rq(cpu);
5597 struct sched_domain *tmp;
5599 /* Remove the sched domains which do not contribute to scheduling. */
5600 for (tmp = sd; tmp; tmp = tmp->parent) {
5601 struct sched_domain *parent = tmp->parent;
5604 if (sd_parent_degenerate(tmp, parent)) {
5605 tmp->parent = parent->parent;
5607 parent->parent->child = tmp;
5611 if (sd && sd_degenerate(sd)) {
5617 sched_domain_debug(sd, cpu);
5619 rcu_assign_pointer(rq->sd, sd);
5622 /* cpus with isolated domains */
5623 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5625 /* Setup the mask of cpus configured for isolated domains */
5626 static int __init isolated_cpu_setup(char *str)
5628 int ints[NR_CPUS], i;
5630 str = get_options(str, ARRAY_SIZE(ints), ints);
5631 cpus_clear(cpu_isolated_map);
5632 for (i = 1; i <= ints[0]; i++)
5633 if (ints[i] < NR_CPUS)
5634 cpu_set(ints[i], cpu_isolated_map);
5638 __setup ("isolcpus=", isolated_cpu_setup);
5641 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5642 * to a function which identifies what group(along with sched group) a CPU
5643 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5644 * (due to the fact that we keep track of groups covered with a cpumask_t).
5646 * init_sched_build_groups will build a circular linked list of the groups
5647 * covered by the given span, and will set each group's ->cpumask correctly,
5648 * and ->cpu_power to 0.
5651 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5652 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5653 struct sched_group **sg))
5655 struct sched_group *first = NULL, *last = NULL;
5656 cpumask_t covered = CPU_MASK_NONE;
5659 for_each_cpu_mask(i, span) {
5660 struct sched_group *sg;
5661 int group = group_fn(i, cpu_map, &sg);
5664 if (cpu_isset(i, covered))
5667 sg->cpumask = CPU_MASK_NONE;
5668 sg->__cpu_power = 0;
5670 for_each_cpu_mask(j, span) {
5671 if (group_fn(j, cpu_map, NULL) != group)
5674 cpu_set(j, covered);
5675 cpu_set(j, sg->cpumask);
5686 #define SD_NODES_PER_DOMAIN 16
5691 * find_next_best_node - find the next node to include in a sched_domain
5692 * @node: node whose sched_domain we're building
5693 * @used_nodes: nodes already in the sched_domain
5695 * Find the next node to include in a given scheduling domain. Simply
5696 * finds the closest node not already in the @used_nodes map.
5698 * Should use nodemask_t.
5700 static int find_next_best_node(int node, unsigned long *used_nodes)
5702 int i, n, val, min_val, best_node = 0;
5706 for (i = 0; i < MAX_NUMNODES; i++) {
5707 /* Start at @node */
5708 n = (node + i) % MAX_NUMNODES;
5710 if (!nr_cpus_node(n))
5713 /* Skip already used nodes */
5714 if (test_bit(n, used_nodes))
5717 /* Simple min distance search */
5718 val = node_distance(node, n);
5720 if (val < min_val) {
5726 set_bit(best_node, used_nodes);
5731 * sched_domain_node_span - get a cpumask for a node's sched_domain
5732 * @node: node whose cpumask we're constructing
5733 * @size: number of nodes to include in this span
5735 * Given a node, construct a good cpumask for its sched_domain to span. It
5736 * should be one that prevents unnecessary balancing, but also spreads tasks
5739 static cpumask_t sched_domain_node_span(int node)
5741 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5742 cpumask_t span, nodemask;
5746 bitmap_zero(used_nodes, MAX_NUMNODES);
5748 nodemask = node_to_cpumask(node);
5749 cpus_or(span, span, nodemask);
5750 set_bit(node, used_nodes);
5752 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5753 int next_node = find_next_best_node(node, used_nodes);
5755 nodemask = node_to_cpumask(next_node);
5756 cpus_or(span, span, nodemask);
5763 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5766 * SMT sched-domains:
5768 #ifdef CONFIG_SCHED_SMT
5769 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5770 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5772 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5773 struct sched_group **sg)
5776 *sg = &per_cpu(sched_group_cpus, cpu);
5782 * multi-core sched-domains:
5784 #ifdef CONFIG_SCHED_MC
5785 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5786 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5789 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5790 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5791 struct sched_group **sg)
5794 cpumask_t mask = cpu_sibling_map[cpu];
5795 cpus_and(mask, mask, *cpu_map);
5796 group = first_cpu(mask);
5798 *sg = &per_cpu(sched_group_core, group);
5801 #elif defined(CONFIG_SCHED_MC)
5802 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5803 struct sched_group **sg)
5806 *sg = &per_cpu(sched_group_core, cpu);
5811 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5812 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5814 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5815 struct sched_group **sg)
5818 #ifdef CONFIG_SCHED_MC
5819 cpumask_t mask = cpu_coregroup_map(cpu);
5820 cpus_and(mask, mask, *cpu_map);
5821 group = first_cpu(mask);
5822 #elif defined(CONFIG_SCHED_SMT)
5823 cpumask_t mask = cpu_sibling_map[cpu];
5824 cpus_and(mask, mask, *cpu_map);
5825 group = first_cpu(mask);
5830 *sg = &per_cpu(sched_group_phys, group);
5836 * The init_sched_build_groups can't handle what we want to do with node
5837 * groups, so roll our own. Now each node has its own list of groups which
5838 * gets dynamically allocated.
5840 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5841 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5843 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5844 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5846 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5847 struct sched_group **sg)
5849 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5852 cpus_and(nodemask, nodemask, *cpu_map);
5853 group = first_cpu(nodemask);
5856 *sg = &per_cpu(sched_group_allnodes, group);
5860 static void init_numa_sched_groups_power(struct sched_group *group_head)
5862 struct sched_group *sg = group_head;
5868 for_each_cpu_mask(j, sg->cpumask) {
5869 struct sched_domain *sd;
5871 sd = &per_cpu(phys_domains, j);
5872 if (j != first_cpu(sd->groups->cpumask)) {
5874 * Only add "power" once for each
5880 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5883 if (sg != group_head)
5889 /* Free memory allocated for various sched_group structures */
5890 static void free_sched_groups(const cpumask_t *cpu_map)
5894 for_each_cpu_mask(cpu, *cpu_map) {
5895 struct sched_group **sched_group_nodes
5896 = sched_group_nodes_bycpu[cpu];
5898 if (!sched_group_nodes)
5901 for (i = 0; i < MAX_NUMNODES; i++) {
5902 cpumask_t nodemask = node_to_cpumask(i);
5903 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5905 cpus_and(nodemask, nodemask, *cpu_map);
5906 if (cpus_empty(nodemask))
5916 if (oldsg != sched_group_nodes[i])
5919 kfree(sched_group_nodes);
5920 sched_group_nodes_bycpu[cpu] = NULL;
5924 static void free_sched_groups(const cpumask_t *cpu_map)
5930 * Initialize sched groups cpu_power.
5932 * cpu_power indicates the capacity of sched group, which is used while
5933 * distributing the load between different sched groups in a sched domain.
5934 * Typically cpu_power for all the groups in a sched domain will be same unless
5935 * there are asymmetries in the topology. If there are asymmetries, group
5936 * having more cpu_power will pickup more load compared to the group having
5939 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5940 * the maximum number of tasks a group can handle in the presence of other idle
5941 * or lightly loaded groups in the same sched domain.
5943 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5945 struct sched_domain *child;
5946 struct sched_group *group;
5948 WARN_ON(!sd || !sd->groups);
5950 if (cpu != first_cpu(sd->groups->cpumask))
5955 sd->groups->__cpu_power = 0;
5958 * For perf policy, if the groups in child domain share resources
5959 * (for example cores sharing some portions of the cache hierarchy
5960 * or SMT), then set this domain groups cpu_power such that each group
5961 * can handle only one task, when there are other idle groups in the
5962 * same sched domain.
5964 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5966 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5967 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5972 * add cpu_power of each child group to this groups cpu_power
5974 group = child->groups;
5976 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5977 group = group->next;
5978 } while (group != child->groups);
5982 * Build sched domains for a given set of cpus and attach the sched domains
5983 * to the individual cpus
5985 static int build_sched_domains(const cpumask_t *cpu_map)
5989 struct sched_group **sched_group_nodes = NULL;
5990 int sd_allnodes = 0;
5993 * Allocate the per-node list of sched groups
5995 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
5997 if (!sched_group_nodes) {
5998 printk(KERN_WARNING "Can not alloc sched group node list\n");
6001 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6005 * Set up domains for cpus specified by the cpu_map.
6007 for_each_cpu_mask(i, *cpu_map) {
6008 struct sched_domain *sd = NULL, *p;
6009 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6011 cpus_and(nodemask, nodemask, *cpu_map);
6014 if (cpus_weight(*cpu_map) >
6015 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6016 sd = &per_cpu(allnodes_domains, i);
6017 *sd = SD_ALLNODES_INIT;
6018 sd->span = *cpu_map;
6019 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6025 sd = &per_cpu(node_domains, i);
6027 sd->span = sched_domain_node_span(cpu_to_node(i));
6031 cpus_and(sd->span, sd->span, *cpu_map);
6035 sd = &per_cpu(phys_domains, i);
6037 sd->span = nodemask;
6041 cpu_to_phys_group(i, cpu_map, &sd->groups);
6043 #ifdef CONFIG_SCHED_MC
6045 sd = &per_cpu(core_domains, i);
6047 sd->span = cpu_coregroup_map(i);
6048 cpus_and(sd->span, sd->span, *cpu_map);
6051 cpu_to_core_group(i, cpu_map, &sd->groups);
6054 #ifdef CONFIG_SCHED_SMT
6056 sd = &per_cpu(cpu_domains, i);
6057 *sd = SD_SIBLING_INIT;
6058 sd->span = cpu_sibling_map[i];
6059 cpus_and(sd->span, sd->span, *cpu_map);
6062 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6066 #ifdef CONFIG_SCHED_SMT
6067 /* Set up CPU (sibling) groups */
6068 for_each_cpu_mask(i, *cpu_map) {
6069 cpumask_t this_sibling_map = cpu_sibling_map[i];
6070 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6071 if (i != first_cpu(this_sibling_map))
6074 init_sched_build_groups(this_sibling_map, cpu_map,
6079 #ifdef CONFIG_SCHED_MC
6080 /* Set up multi-core groups */
6081 for_each_cpu_mask(i, *cpu_map) {
6082 cpumask_t this_core_map = cpu_coregroup_map(i);
6083 cpus_and(this_core_map, this_core_map, *cpu_map);
6084 if (i != first_cpu(this_core_map))
6086 init_sched_build_groups(this_core_map, cpu_map,
6087 &cpu_to_core_group);
6091 /* Set up physical groups */
6092 for (i = 0; i < MAX_NUMNODES; i++) {
6093 cpumask_t nodemask = node_to_cpumask(i);
6095 cpus_and(nodemask, nodemask, *cpu_map);
6096 if (cpus_empty(nodemask))
6099 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6103 /* Set up node groups */
6105 init_sched_build_groups(*cpu_map, cpu_map,
6106 &cpu_to_allnodes_group);
6108 for (i = 0; i < MAX_NUMNODES; i++) {
6109 /* Set up node groups */
6110 struct sched_group *sg, *prev;
6111 cpumask_t nodemask = node_to_cpumask(i);
6112 cpumask_t domainspan;
6113 cpumask_t covered = CPU_MASK_NONE;
6116 cpus_and(nodemask, nodemask, *cpu_map);
6117 if (cpus_empty(nodemask)) {
6118 sched_group_nodes[i] = NULL;
6122 domainspan = sched_domain_node_span(i);
6123 cpus_and(domainspan, domainspan, *cpu_map);
6125 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6127 printk(KERN_WARNING "Can not alloc domain group for "
6131 sched_group_nodes[i] = sg;
6132 for_each_cpu_mask(j, nodemask) {
6133 struct sched_domain *sd;
6135 sd = &per_cpu(node_domains, j);
6138 sg->__cpu_power = 0;
6139 sg->cpumask = nodemask;
6141 cpus_or(covered, covered, nodemask);
6144 for (j = 0; j < MAX_NUMNODES; j++) {
6145 cpumask_t tmp, notcovered;
6146 int n = (i + j) % MAX_NUMNODES;
6148 cpus_complement(notcovered, covered);
6149 cpus_and(tmp, notcovered, *cpu_map);
6150 cpus_and(tmp, tmp, domainspan);
6151 if (cpus_empty(tmp))
6154 nodemask = node_to_cpumask(n);
6155 cpus_and(tmp, tmp, nodemask);
6156 if (cpus_empty(tmp))
6159 sg = kmalloc_node(sizeof(struct sched_group),
6163 "Can not alloc domain group for node %d\n", j);
6166 sg->__cpu_power = 0;
6168 sg->next = prev->next;
6169 cpus_or(covered, covered, tmp);
6176 /* Calculate CPU power for physical packages and nodes */
6177 #ifdef CONFIG_SCHED_SMT
6178 for_each_cpu_mask(i, *cpu_map) {
6179 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6181 init_sched_groups_power(i, sd);
6184 #ifdef CONFIG_SCHED_MC
6185 for_each_cpu_mask(i, *cpu_map) {
6186 struct sched_domain *sd = &per_cpu(core_domains, i);
6188 init_sched_groups_power(i, sd);
6192 for_each_cpu_mask(i, *cpu_map) {
6193 struct sched_domain *sd = &per_cpu(phys_domains, i);
6195 init_sched_groups_power(i, sd);
6199 for (i = 0; i < MAX_NUMNODES; i++)
6200 init_numa_sched_groups_power(sched_group_nodes[i]);
6203 struct sched_group *sg;
6205 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6206 init_numa_sched_groups_power(sg);
6210 /* Attach the domains */
6211 for_each_cpu_mask(i, *cpu_map) {
6212 struct sched_domain *sd;
6213 #ifdef CONFIG_SCHED_SMT
6214 sd = &per_cpu(cpu_domains, i);
6215 #elif defined(CONFIG_SCHED_MC)
6216 sd = &per_cpu(core_domains, i);
6218 sd = &per_cpu(phys_domains, i);
6220 cpu_attach_domain(sd, i);
6227 free_sched_groups(cpu_map);
6232 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6234 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6236 cpumask_t cpu_default_map;
6240 * Setup mask for cpus without special case scheduling requirements.
6241 * For now this just excludes isolated cpus, but could be used to
6242 * exclude other special cases in the future.
6244 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6246 err = build_sched_domains(&cpu_default_map);
6251 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6253 free_sched_groups(cpu_map);
6257 * Detach sched domains from a group of cpus specified in cpu_map
6258 * These cpus will now be attached to the NULL domain
6260 static void detach_destroy_domains(const cpumask_t *cpu_map)
6264 for_each_cpu_mask(i, *cpu_map)
6265 cpu_attach_domain(NULL, i);
6266 synchronize_sched();
6267 arch_destroy_sched_domains(cpu_map);
6271 * Partition sched domains as specified by the cpumasks below.
6272 * This attaches all cpus from the cpumasks to the NULL domain,
6273 * waits for a RCU quiescent period, recalculates sched
6274 * domain information and then attaches them back to the
6275 * correct sched domains
6276 * Call with hotplug lock held
6278 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6280 cpumask_t change_map;
6283 cpus_and(*partition1, *partition1, cpu_online_map);
6284 cpus_and(*partition2, *partition2, cpu_online_map);
6285 cpus_or(change_map, *partition1, *partition2);
6287 /* Detach sched domains from all of the affected cpus */
6288 detach_destroy_domains(&change_map);
6289 if (!cpus_empty(*partition1))
6290 err = build_sched_domains(partition1);
6291 if (!err && !cpus_empty(*partition2))
6292 err = build_sched_domains(partition2);
6297 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6298 int arch_reinit_sched_domains(void)
6302 mutex_lock(&sched_hotcpu_mutex);
6303 detach_destroy_domains(&cpu_online_map);
6304 err = arch_init_sched_domains(&cpu_online_map);
6305 mutex_unlock(&sched_hotcpu_mutex);
6310 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6314 if (buf[0] != '0' && buf[0] != '1')
6318 sched_smt_power_savings = (buf[0] == '1');
6320 sched_mc_power_savings = (buf[0] == '1');
6322 ret = arch_reinit_sched_domains();
6324 return ret ? ret : count;
6327 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6331 #ifdef CONFIG_SCHED_SMT
6333 err = sysfs_create_file(&cls->kset.kobj,
6334 &attr_sched_smt_power_savings.attr);
6336 #ifdef CONFIG_SCHED_MC
6337 if (!err && mc_capable())
6338 err = sysfs_create_file(&cls->kset.kobj,
6339 &attr_sched_mc_power_savings.attr);
6345 #ifdef CONFIG_SCHED_MC
6346 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6348 return sprintf(page, "%u\n", sched_mc_power_savings);
6350 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6351 const char *buf, size_t count)
6353 return sched_power_savings_store(buf, count, 0);
6355 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6356 sched_mc_power_savings_store);
6359 #ifdef CONFIG_SCHED_SMT
6360 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6362 return sprintf(page, "%u\n", sched_smt_power_savings);
6364 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6365 const char *buf, size_t count)
6367 return sched_power_savings_store(buf, count, 1);
6369 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6370 sched_smt_power_savings_store);
6374 * Force a reinitialization of the sched domains hierarchy. The domains
6375 * and groups cannot be updated in place without racing with the balancing
6376 * code, so we temporarily attach all running cpus to the NULL domain
6377 * which will prevent rebalancing while the sched domains are recalculated.
6379 static int update_sched_domains(struct notifier_block *nfb,
6380 unsigned long action, void *hcpu)
6383 case CPU_UP_PREPARE:
6384 case CPU_UP_PREPARE_FROZEN:
6385 case CPU_DOWN_PREPARE:
6386 case CPU_DOWN_PREPARE_FROZEN:
6387 detach_destroy_domains(&cpu_online_map);
6390 case CPU_UP_CANCELED:
6391 case CPU_UP_CANCELED_FROZEN:
6392 case CPU_DOWN_FAILED:
6393 case CPU_DOWN_FAILED_FROZEN:
6395 case CPU_ONLINE_FROZEN:
6397 case CPU_DEAD_FROZEN:
6399 * Fall through and re-initialise the domains.
6406 /* The hotplug lock is already held by cpu_up/cpu_down */
6407 arch_init_sched_domains(&cpu_online_map);
6412 void __init sched_init_smp(void)
6414 cpumask_t non_isolated_cpus;
6416 mutex_lock(&sched_hotcpu_mutex);
6417 arch_init_sched_domains(&cpu_online_map);
6418 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6419 if (cpus_empty(non_isolated_cpus))
6420 cpu_set(smp_processor_id(), non_isolated_cpus);
6421 mutex_unlock(&sched_hotcpu_mutex);
6422 /* XXX: Theoretical race here - CPU may be hotplugged now */
6423 hotcpu_notifier(update_sched_domains, 0);
6425 init_sched_domain_sysctl();
6427 /* Move init over to a non-isolated CPU */
6428 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6430 sched_init_granularity();
6433 void __init sched_init_smp(void)
6435 sched_init_granularity();
6437 #endif /* CONFIG_SMP */
6439 int in_sched_functions(unsigned long addr)
6441 /* Linker adds these: start and end of __sched functions */
6442 extern char __sched_text_start[], __sched_text_end[];
6444 return in_lock_functions(addr) ||
6445 (addr >= (unsigned long)__sched_text_start
6446 && addr < (unsigned long)__sched_text_end);
6449 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6451 cfs_rq->tasks_timeline = RB_ROOT;
6452 cfs_rq->fair_clock = 1;
6453 #ifdef CONFIG_FAIR_GROUP_SCHED
6458 void __init sched_init(void)
6460 u64 now = sched_clock();
6461 int highest_cpu = 0;
6465 * Link up the scheduling class hierarchy:
6467 rt_sched_class.next = &fair_sched_class;
6468 fair_sched_class.next = &idle_sched_class;
6469 idle_sched_class.next = NULL;
6471 for_each_possible_cpu(i) {
6472 struct rt_prio_array *array;
6476 spin_lock_init(&rq->lock);
6477 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6480 init_cfs_rq(&rq->cfs, rq);
6481 #ifdef CONFIG_FAIR_GROUP_SCHED
6482 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6483 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6485 rq->ls.load_update_last = now;
6486 rq->ls.load_update_start = now;
6488 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6489 rq->cpu_load[j] = 0;
6492 rq->active_balance = 0;
6493 rq->next_balance = jiffies;
6496 rq->migration_thread = NULL;
6497 INIT_LIST_HEAD(&rq->migration_queue);
6499 atomic_set(&rq->nr_iowait, 0);
6501 array = &rq->rt.active;
6502 for (j = 0; j < MAX_RT_PRIO; j++) {
6503 INIT_LIST_HEAD(array->queue + j);
6504 __clear_bit(j, array->bitmap);
6507 /* delimiter for bitsearch: */
6508 __set_bit(MAX_RT_PRIO, array->bitmap);
6511 set_load_weight(&init_task);
6513 #ifdef CONFIG_PREEMPT_NOTIFIERS
6514 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6518 nr_cpu_ids = highest_cpu + 1;
6519 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6522 #ifdef CONFIG_RT_MUTEXES
6523 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6527 * The boot idle thread does lazy MMU switching as well:
6529 atomic_inc(&init_mm.mm_count);
6530 enter_lazy_tlb(&init_mm, current);
6533 * Make us the idle thread. Technically, schedule() should not be
6534 * called from this thread, however somewhere below it might be,
6535 * but because we are the idle thread, we just pick up running again
6536 * when this runqueue becomes "idle".
6538 init_idle(current, smp_processor_id());
6540 * During early bootup we pretend to be a normal task:
6542 current->sched_class = &fair_sched_class;
6545 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6546 void __might_sleep(char *file, int line)
6549 static unsigned long prev_jiffy; /* ratelimiting */
6551 if ((in_atomic() || irqs_disabled()) &&
6552 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6553 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6555 prev_jiffy = jiffies;
6556 printk(KERN_ERR "BUG: sleeping function called from invalid"
6557 " context at %s:%d\n", file, line);
6558 printk("in_atomic():%d, irqs_disabled():%d\n",
6559 in_atomic(), irqs_disabled());
6560 debug_show_held_locks(current);
6561 if (irqs_disabled())
6562 print_irqtrace_events(current);
6567 EXPORT_SYMBOL(__might_sleep);
6570 #ifdef CONFIG_MAGIC_SYSRQ
6571 void normalize_rt_tasks(void)
6573 struct task_struct *g, *p;
6574 unsigned long flags;
6578 read_lock_irq(&tasklist_lock);
6579 do_each_thread(g, p) {
6581 p->se.wait_runtime = 0;
6582 p->se.wait_start_fair = 0;
6583 p->se.wait_start = 0;
6584 p->se.exec_start = 0;
6585 p->se.sleep_start = 0;
6586 p->se.sleep_start_fair = 0;
6587 p->se.block_start = 0;
6588 task_rq(p)->cfs.fair_clock = 0;
6589 task_rq(p)->clock = 0;
6593 * Renice negative nice level userspace
6596 if (TASK_NICE(p) < 0 && p->mm)
6597 set_user_nice(p, 0);
6601 spin_lock_irqsave(&p->pi_lock, flags);
6602 rq = __task_rq_lock(p);
6605 * Do not touch the migration thread:
6607 if (p == rq->migration_thread)
6611 on_rq = p->se.on_rq;
6613 deactivate_task(task_rq(p), p, 0);
6614 __setscheduler(rq, p, SCHED_NORMAL, 0);
6616 activate_task(task_rq(p), p, 0);
6617 resched_task(rq->curr);
6622 __task_rq_unlock(rq);
6623 spin_unlock_irqrestore(&p->pi_lock, flags);
6624 } while_each_thread(g, p);
6626 read_unlock_irq(&tasklist_lock);
6629 #endif /* CONFIG_MAGIC_SYSRQ */
6633 * These functions are only useful for the IA64 MCA handling.
6635 * They can only be called when the whole system has been
6636 * stopped - every CPU needs to be quiescent, and no scheduling
6637 * activity can take place. Using them for anything else would
6638 * be a serious bug, and as a result, they aren't even visible
6639 * under any other configuration.
6643 * curr_task - return the current task for a given cpu.
6644 * @cpu: the processor in question.
6646 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6648 struct task_struct *curr_task(int cpu)
6650 return cpu_curr(cpu);
6654 * set_curr_task - set the current task for a given cpu.
6655 * @cpu: the processor in question.
6656 * @p: the task pointer to set.
6658 * Description: This function must only be used when non-maskable interrupts
6659 * are serviced on a separate stack. It allows the architecture to switch the
6660 * notion of the current task on a cpu in a non-blocking manner. This function
6661 * must be called with all CPU's synchronized, and interrupts disabled, the
6662 * and caller must save the original value of the current task (see
6663 * curr_task() above) and restore that value before reenabling interrupts and
6664 * re-starting the system.
6666 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6668 void set_curr_task(int cpu, struct task_struct *p)