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
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/reciprocal_div.h>
58 #include <asm/unistd.h>
61 * Scheduler clock - returns current time in nanosec units.
62 * This is default implementation.
63 * Architectures and sub-architectures can override this.
65 unsigned long long __attribute__((weak)) sched_clock(void)
67 return (unsigned long long)jiffies * (1000000000 / HZ);
71 * Convert user-nice values [ -20 ... 0 ... 19 ]
72 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
75 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
76 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
77 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
80 * 'User priority' is the nice value converted to something we
81 * can work with better when scaling various scheduler parameters,
82 * it's a [ 0 ... 39 ] range.
84 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
85 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
86 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
89 * Some helpers for converting nanosecond timing to jiffy resolution
91 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
92 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
94 #define NICE_0_LOAD SCHED_LOAD_SCALE
95 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
98 * These are the 'tuning knobs' of the scheduler:
100 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
101 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
102 * Timeslices get refilled after they expire.
104 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
105 #define DEF_TIMESLICE (100 * HZ / 1000)
106 #define ON_RUNQUEUE_WEIGHT 30
107 #define CHILD_PENALTY 95
108 #define PARENT_PENALTY 100
109 #define EXIT_WEIGHT 3
110 #define PRIO_BONUS_RATIO 25
111 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
112 #define INTERACTIVE_DELTA 2
113 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
114 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
115 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
118 * If a task is 'interactive' then we reinsert it in the active
119 * array after it has expired its current timeslice. (it will not
120 * continue to run immediately, it will still roundrobin with
121 * other interactive tasks.)
123 * This part scales the interactivity limit depending on niceness.
125 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
126 * Here are a few examples of different nice levels:
128 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
129 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
130 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
131 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
132 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
134 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
135 * priority range a task can explore, a value of '1' means the
136 * task is rated interactive.)
138 * Ie. nice +19 tasks can never get 'interactive' enough to be
139 * reinserted into the active array. And only heavily CPU-hog nice -20
140 * tasks will be expired. Default nice 0 tasks are somewhere between,
141 * it takes some effort for them to get interactive, but it's not
145 #define CURRENT_BONUS(p) \
146 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
149 #define GRANULARITY (10 * HZ / 1000 ? : 1)
152 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
153 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
156 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
157 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
160 #define SCALE(v1,v1_max,v2_max) \
161 (v1) * (v2_max) / (v1_max)
164 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
167 #define TASK_INTERACTIVE(p) \
168 ((p)->prio <= (p)->static_prio - DELTA(p))
170 #define INTERACTIVE_SLEEP(p) \
171 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
172 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
174 #define TASK_PREEMPTS_CURR(p, rq) \
175 ((p)->prio < (rq)->curr->prio)
177 #define SCALE_PRIO(x, prio) \
178 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
180 static unsigned int static_prio_timeslice(int static_prio)
182 if (static_prio < NICE_TO_PRIO(0))
183 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
185 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
190 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
191 * Since cpu_power is a 'constant', we can use a reciprocal divide.
193 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
195 return reciprocal_divide(load, sg->reciprocal_cpu_power);
199 * Each time a sched group cpu_power is changed,
200 * we must compute its reciprocal value
202 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
204 sg->__cpu_power += val;
205 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
210 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
211 * to time slice values: [800ms ... 100ms ... 5ms]
213 * The higher a thread's priority, the bigger timeslices
214 * it gets during one round of execution. But even the lowest
215 * priority thread gets MIN_TIMESLICE worth of execution time.
218 static inline unsigned int task_timeslice(struct task_struct *p)
220 return static_prio_timeslice(p->static_prio);
223 static inline int rt_policy(int policy)
225 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
230 static inline int task_has_rt_policy(struct task_struct *p)
232 return rt_policy(p->policy);
236 * This is the priority-queue data structure of the RT scheduling class:
238 struct rt_prio_array {
239 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
240 struct list_head queue[MAX_RT_PRIO];
244 struct load_weight load;
245 u64 load_update_start, load_update_last;
246 unsigned long delta_fair, delta_exec, delta_stat;
249 /* CFS-related fields in a runqueue */
251 struct load_weight load;
252 unsigned long nr_running;
258 unsigned long wait_runtime_overruns, wait_runtime_underruns;
260 struct rb_root tasks_timeline;
261 struct rb_node *rb_leftmost;
262 struct rb_node *rb_load_balance_curr;
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 /* 'curr' points to currently running entity on this cfs_rq.
265 * It is set to NULL otherwise (i.e when none are currently running).
267 struct sched_entity *curr;
268 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
270 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
271 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
272 * (like users, containers etc.)
274 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
275 * list is used during load balance.
277 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
281 /* Real-Time classes' related field in a runqueue: */
283 struct rt_prio_array active;
284 int rt_load_balance_idx;
285 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
289 * The prio-array type of the old scheduler:
292 unsigned int nr_active;
293 DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
294 struct list_head queue[MAX_PRIO];
298 * This is the main, per-CPU runqueue data structure.
300 * Locking rule: those places that want to lock multiple runqueues
301 * (such as the load balancing or the thread migration code), lock
302 * acquire operations must be ordered by ascending &runqueue.
305 spinlock_t lock; /* runqueue lock */
308 * nr_running and cpu_load should be in the same cacheline because
309 * remote CPUs use both these fields when doing load calculation.
311 unsigned long nr_running;
312 unsigned long raw_weighted_load;
313 #define CPU_LOAD_IDX_MAX 5
314 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
315 unsigned char idle_at_tick;
317 unsigned char in_nohz_recently;
319 struct load_stat ls; /* capture load from *all* tasks on this cpu */
320 unsigned long nr_load_updates;
324 #ifdef CONFIG_FAIR_GROUP_SCHED
325 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
330 * This is part of a global counter where only the total sum
331 * over all CPUs matters. A task can increase this counter on
332 * one CPU and if it got migrated afterwards it may decrease
333 * it on another CPU. Always updated under the runqueue lock:
335 unsigned long nr_uninterruptible;
337 unsigned long expired_timestamp;
338 unsigned long long most_recent_timestamp;
340 struct task_struct *curr, *idle;
341 unsigned long next_balance;
342 struct mm_struct *prev_mm;
344 struct prio_array *active, *expired, arrays[2];
345 int best_expired_prio;
347 u64 clock, prev_clock_raw;
350 unsigned int clock_warps, clock_overflows;
351 unsigned int clock_unstable_events;
353 struct sched_class *load_balance_class;
358 struct sched_domain *sd;
360 /* For active balancing */
363 int cpu; /* cpu of this runqueue */
365 struct task_struct *migration_thread;
366 struct list_head migration_queue;
369 #ifdef CONFIG_SCHEDSTATS
371 struct sched_info rq_sched_info;
373 /* sys_sched_yield() stats */
374 unsigned long yld_exp_empty;
375 unsigned long yld_act_empty;
376 unsigned long yld_both_empty;
377 unsigned long yld_cnt;
379 /* schedule() stats */
380 unsigned long sched_switch;
381 unsigned long sched_cnt;
382 unsigned long sched_goidle;
384 /* try_to_wake_up() stats */
385 unsigned long ttwu_cnt;
386 unsigned long ttwu_local;
388 struct lock_class_key rq_lock_key;
391 static DEFINE_PER_CPU(struct rq, runqueues) ____cacheline_aligned_in_smp;
392 static DEFINE_MUTEX(sched_hotcpu_mutex);
394 static inline int cpu_of(struct rq *rq)
404 * Per-runqueue clock, as finegrained as the platform can give us:
406 static unsigned long long __rq_clock(struct rq *rq)
408 u64 prev_raw = rq->prev_clock_raw;
409 u64 now = sched_clock();
410 s64 delta = now - prev_raw;
411 u64 clock = rq->clock;
414 * Protect against sched_clock() occasionally going backwards:
416 if (unlikely(delta < 0)) {
421 * Catch too large forward jumps too:
423 if (unlikely(delta > 2*TICK_NSEC)) {
425 rq->clock_overflows++;
427 if (unlikely(delta > rq->clock_max_delta))
428 rq->clock_max_delta = delta;
433 rq->prev_clock_raw = now;
439 static inline unsigned long long rq_clock(struct rq *rq)
441 int this_cpu = smp_processor_id();
443 if (this_cpu == cpu_of(rq))
444 return __rq_clock(rq);
450 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
451 * See detach_destroy_domains: synchronize_sched for details.
453 * The domain tree of any CPU may only be accessed from within
454 * preempt-disabled sections.
456 #define for_each_domain(cpu, __sd) \
457 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
459 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
460 #define this_rq() (&__get_cpu_var(runqueues))
461 #define task_rq(p) cpu_rq(task_cpu(p))
462 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
464 #ifdef CONFIG_FAIR_GROUP_SCHED
465 /* Change a task's ->cfs_rq if it moves across CPUs */
466 static inline void set_task_cfs_rq(struct task_struct *p)
468 p->se.cfs_rq = &task_rq(p)->cfs;
471 static inline void set_task_cfs_rq(struct task_struct *p)
476 #ifndef prepare_arch_switch
477 # define prepare_arch_switch(next) do { } while (0)
479 #ifndef finish_arch_switch
480 # define finish_arch_switch(prev) do { } while (0)
483 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
484 static inline int task_running(struct rq *rq, struct task_struct *p)
486 return rq->curr == p;
489 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
493 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
495 #ifdef CONFIG_DEBUG_SPINLOCK
496 /* this is a valid case when another task releases the spinlock */
497 rq->lock.owner = current;
500 * If we are tracking spinlock dependencies then we have to
501 * fix up the runqueue lock - which gets 'carried over' from
504 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
506 spin_unlock_irq(&rq->lock);
509 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
510 static inline int task_running(struct rq *rq, struct task_struct *p)
515 return rq->curr == p;
519 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
523 * We can optimise this out completely for !SMP, because the
524 * SMP rebalancing from interrupt is the only thing that cares
529 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
530 spin_unlock_irq(&rq->lock);
532 spin_unlock(&rq->lock);
536 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
540 * After ->oncpu is cleared, the task can be moved to a different CPU.
541 * We must ensure this doesn't happen until the switch is completely
547 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
551 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
554 * __task_rq_lock - lock the runqueue a given task resides on.
555 * Must be called interrupts disabled.
557 static inline struct rq *__task_rq_lock(struct task_struct *p)
564 spin_lock(&rq->lock);
565 if (unlikely(rq != task_rq(p))) {
566 spin_unlock(&rq->lock);
567 goto repeat_lock_task;
573 * task_rq_lock - lock the runqueue a given task resides on and disable
574 * interrupts. Note the ordering: we can safely lookup the task_rq without
575 * explicitly disabling preemption.
577 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
583 local_irq_save(*flags);
585 spin_lock(&rq->lock);
586 if (unlikely(rq != task_rq(p))) {
587 spin_unlock_irqrestore(&rq->lock, *flags);
588 goto repeat_lock_task;
593 static inline void __task_rq_unlock(struct rq *rq)
596 spin_unlock(&rq->lock);
599 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
602 spin_unlock_irqrestore(&rq->lock, *flags);
606 * this_rq_lock - lock this runqueue and disable interrupts.
608 static inline struct rq *this_rq_lock(void)
615 spin_lock(&rq->lock);
621 * resched_task - mark a task 'to be rescheduled now'.
623 * On UP this means the setting of the need_resched flag, on SMP it
624 * might also involve a cross-CPU call to trigger the scheduler on
629 #ifndef tsk_is_polling
630 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
633 static void resched_task(struct task_struct *p)
637 assert_spin_locked(&task_rq(p)->lock);
639 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
642 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
645 if (cpu == smp_processor_id())
648 /* NEED_RESCHED must be visible before we test polling */
650 if (!tsk_is_polling(p))
651 smp_send_reschedule(cpu);
654 static void resched_cpu(int cpu)
656 struct rq *rq = cpu_rq(cpu);
659 if (!spin_trylock_irqsave(&rq->lock, flags))
661 resched_task(cpu_curr(cpu));
662 spin_unlock_irqrestore(&rq->lock, flags);
665 static inline void resched_task(struct task_struct *p)
667 assert_spin_locked(&task_rq(p)->lock);
668 set_tsk_need_resched(p);
672 #include "sched_stats.h"
674 static u64 div64_likely32(u64 divident, unsigned long divisor)
676 #if BITS_PER_LONG == 32
677 if (likely(divident <= 0xffffffffULL))
678 return (u32)divident / divisor;
679 do_div(divident, divisor);
683 return divident / divisor;
687 #if BITS_PER_LONG == 32
688 # define WMULT_CONST (~0UL)
690 # define WMULT_CONST (1UL << 32)
693 #define WMULT_SHIFT 32
695 static inline unsigned long
696 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
697 struct load_weight *lw)
701 if (unlikely(!lw->inv_weight))
702 lw->inv_weight = WMULT_CONST / lw->weight;
704 tmp = (u64)delta_exec * weight;
706 * Check whether we'd overflow the 64-bit multiplication:
708 if (unlikely(tmp > WMULT_CONST)) {
709 tmp = ((tmp >> WMULT_SHIFT/2) * lw->inv_weight)
712 tmp = (tmp * lw->inv_weight) >> WMULT_SHIFT;
715 return (unsigned long)min(tmp, (u64)sysctl_sched_runtime_limit);
718 static inline unsigned long
719 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
721 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
724 static void update_load_add(struct load_weight *lw, unsigned long inc)
730 static void update_load_sub(struct load_weight *lw, unsigned long dec)
736 static void __update_curr_load(struct rq *rq, struct load_stat *ls)
738 if (rq->curr != rq->idle && ls->load.weight) {
739 ls->delta_exec += ls->delta_stat;
740 ls->delta_fair += calc_delta_fair(ls->delta_stat, &ls->load);
746 * Update delta_exec, delta_fair fields for rq.
748 * delta_fair clock advances at a rate inversely proportional to
749 * total load (rq->ls.load.weight) on the runqueue, while
750 * delta_exec advances at the same rate as wall-clock (provided
753 * delta_exec / delta_fair is a measure of the (smoothened) load on this
754 * runqueue over any given interval. This (smoothened) load is used
755 * during load balance.
757 * This function is called /before/ updating rq->ls.load
758 * and when switching tasks.
760 static void update_curr_load(struct rq *rq, u64 now)
762 struct load_stat *ls = &rq->ls;
765 start = ls->load_update_start;
766 ls->load_update_start = now;
767 ls->delta_stat += now - start;
769 * Stagger updates to ls->delta_fair. Very frequent updates
772 if (ls->delta_stat >= sysctl_sched_stat_granularity)
773 __update_curr_load(rq, ls);
777 * To aid in avoiding the subversion of "niceness" due to uneven distribution
778 * of tasks with abnormal "nice" values across CPUs the contribution that
779 * each task makes to its run queue's load is weighted according to its
780 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
781 * scaled version of the new time slice allocation that they receive on time
786 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
787 * If static_prio_timeslice() is ever changed to break this assumption then
788 * this code will need modification
790 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
791 #define LOAD_WEIGHT(lp) \
792 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
793 #define PRIO_TO_LOAD_WEIGHT(prio) \
794 LOAD_WEIGHT(static_prio_timeslice(prio))
795 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
796 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
799 inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
801 rq->raw_weighted_load += p->load_weight;
805 dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
807 rq->raw_weighted_load -= p->load_weight;
810 static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
813 inc_raw_weighted_load(rq, p);
816 static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
819 dec_raw_weighted_load(rq, p);
822 static void set_load_weight(struct task_struct *p)
824 if (task_has_rt_policy(p)) {
826 if (p == task_rq(p)->migration_thread)
828 * The migration thread does the actual balancing.
829 * Giving its load any weight will skew balancing
835 p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
837 p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
841 * Adding/removing a task to/from a priority array:
843 static void dequeue_task(struct task_struct *p, struct prio_array *array)
846 list_del(&p->run_list);
847 if (list_empty(array->queue + p->prio))
848 __clear_bit(p->prio, array->bitmap);
851 static void enqueue_task(struct task_struct *p, struct prio_array *array)
853 sched_info_queued(p);
854 list_add_tail(&p->run_list, array->queue + p->prio);
855 __set_bit(p->prio, array->bitmap);
861 * Put task to the end of the run list without the overhead of dequeue
862 * followed by enqueue.
864 static void requeue_task(struct task_struct *p, struct prio_array *array)
866 list_move_tail(&p->run_list, array->queue + p->prio);
870 enqueue_task_head(struct task_struct *p, struct prio_array *array)
872 list_add(&p->run_list, array->queue + p->prio);
873 __set_bit(p->prio, array->bitmap);
879 * __normal_prio - return the priority that is based on the static
880 * priority but is modified by bonuses/penalties.
882 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
883 * into the -5 ... 0 ... +5 bonus/penalty range.
885 * We use 25% of the full 0...39 priority range so that:
887 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
888 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
890 * Both properties are important to certain workloads.
893 static inline int __normal_prio(struct task_struct *p)
897 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
899 prio = p->static_prio - bonus;
900 if (prio < MAX_RT_PRIO)
902 if (prio > MAX_PRIO-1)
908 * Calculate the expected normal priority: i.e. priority
909 * without taking RT-inheritance into account. Might be
910 * boosted by interactivity modifiers. Changes upon fork,
911 * setprio syscalls, and whenever the interactivity
912 * estimator recalculates.
914 static inline int normal_prio(struct task_struct *p)
918 if (task_has_rt_policy(p))
919 prio = MAX_RT_PRIO-1 - p->rt_priority;
921 prio = __normal_prio(p);
926 * Calculate the current priority, i.e. the priority
927 * taken into account by the scheduler. This value might
928 * be boosted by RT tasks, or might be boosted by
929 * interactivity modifiers. Will be RT if the task got
930 * RT-boosted. If not then it returns p->normal_prio.
932 static int effective_prio(struct task_struct *p)
934 p->normal_prio = normal_prio(p);
936 * If we are RT tasks or we were boosted to RT priority,
937 * keep the priority unchanged. Otherwise, update priority
938 * to the normal priority:
940 if (!rt_prio(p->prio))
941 return p->normal_prio;
946 * __activate_task - move a task to the runqueue.
948 static void __activate_task(struct task_struct *p, struct rq *rq)
950 struct prio_array *target = rq->active;
953 target = rq->expired;
954 enqueue_task(p, target);
955 inc_nr_running(p, rq);
959 * __activate_idle_task - move idle task to the _front_ of runqueue.
961 static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
963 enqueue_task_head(p, rq->active);
964 inc_nr_running(p, rq);
968 * Recalculate p->normal_prio and p->prio after having slept,
969 * updating the sleep-average too:
971 static int recalc_task_prio(struct task_struct *p, unsigned long long now)
973 /* Caller must always ensure 'now >= p->timestamp' */
974 unsigned long sleep_time = now - p->timestamp;
979 if (likely(sleep_time > 0)) {
981 * This ceiling is set to the lowest priority that would allow
982 * a task to be reinserted into the active array on timeslice
985 unsigned long ceiling = INTERACTIVE_SLEEP(p);
987 if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
989 * Prevents user tasks from achieving best priority
990 * with one single large enough sleep.
992 p->sleep_avg = ceiling;
994 * Using INTERACTIVE_SLEEP() as a ceiling places a
995 * nice(0) task 1ms sleep away from promotion, and
996 * gives it 700ms to round-robin with no chance of
997 * being demoted. This is more than generous, so
998 * mark this sleep as non-interactive to prevent the
999 * on-runqueue bonus logic from intervening should
1000 * this task not receive cpu immediately.
1002 p->sleep_type = SLEEP_NONINTERACTIVE;
1005 * Tasks waking from uninterruptible sleep are
1006 * limited in their sleep_avg rise as they
1007 * are likely to be waiting on I/O
1009 if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
1010 if (p->sleep_avg >= ceiling)
1012 else if (p->sleep_avg + sleep_time >=
1014 p->sleep_avg = ceiling;
1020 * This code gives a bonus to interactive tasks.
1022 * The boost works by updating the 'average sleep time'
1023 * value here, based on ->timestamp. The more time a
1024 * task spends sleeping, the higher the average gets -
1025 * and the higher the priority boost gets as well.
1027 p->sleep_avg += sleep_time;
1030 if (p->sleep_avg > NS_MAX_SLEEP_AVG)
1031 p->sleep_avg = NS_MAX_SLEEP_AVG;
1034 return effective_prio(p);
1038 * activate_task - move a task to the runqueue and do priority recalculation
1040 * Update all the scheduling statistics stuff. (sleep average
1041 * calculation, priority modifiers, etc.)
1043 static void activate_task(struct task_struct *p, struct rq *rq, int local)
1045 unsigned long long now;
1050 now = sched_clock();
1053 /* Compensate for drifting sched_clock */
1054 struct rq *this_rq = this_rq();
1055 now = (now - this_rq->most_recent_timestamp)
1056 + rq->most_recent_timestamp;
1061 * Sleep time is in units of nanosecs, so shift by 20 to get a
1062 * milliseconds-range estimation of the amount of time that the task
1065 if (unlikely(prof_on == SLEEP_PROFILING)) {
1066 if (p->state == TASK_UNINTERRUPTIBLE)
1067 profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
1068 (now - p->timestamp) >> 20);
1071 p->prio = recalc_task_prio(p, now);
1074 * This checks to make sure it's not an uninterruptible task
1075 * that is now waking up.
1077 if (p->sleep_type == SLEEP_NORMAL) {
1079 * Tasks which were woken up by interrupts (ie. hw events)
1080 * are most likely of interactive nature. So we give them
1081 * the credit of extending their sleep time to the period
1082 * of time they spend on the runqueue, waiting for execution
1083 * on a CPU, first time around:
1086 p->sleep_type = SLEEP_INTERRUPTED;
1089 * Normal first-time wakeups get a credit too for
1090 * on-runqueue time, but it will be weighted down:
1092 p->sleep_type = SLEEP_INTERACTIVE;
1097 __activate_task(p, rq);
1101 * deactivate_task - remove a task from the runqueue.
1103 static void deactivate_task(struct task_struct *p, struct rq *rq)
1105 dec_nr_running(p, rq);
1106 dequeue_task(p, p->array);
1111 * task_curr - is this task currently executing on a CPU?
1112 * @p: the task in question.
1114 inline int task_curr(const struct task_struct *p)
1116 return cpu_curr(task_cpu(p)) == p;
1119 /* Used instead of source_load when we know the type == 0 */
1120 unsigned long weighted_cpuload(const int cpu)
1122 return cpu_rq(cpu)->raw_weighted_load;
1127 void set_task_cpu(struct task_struct *p, unsigned int cpu)
1129 task_thread_info(p)->cpu = cpu;
1132 struct migration_req {
1133 struct list_head list;
1135 struct task_struct *task;
1138 struct completion done;
1142 * The task's runqueue lock must be held.
1143 * Returns true if you have to wait for migration thread.
1146 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1148 struct rq *rq = task_rq(p);
1151 * If the task is not on a runqueue (and not running), then
1152 * it is sufficient to simply update the task's cpu field.
1154 if (!p->array && !task_running(rq, p)) {
1155 set_task_cpu(p, dest_cpu);
1159 init_completion(&req->done);
1161 req->dest_cpu = dest_cpu;
1162 list_add(&req->list, &rq->migration_queue);
1168 * wait_task_inactive - wait for a thread to unschedule.
1170 * The caller must ensure that the task *will* unschedule sometime soon,
1171 * else this function might spin for a *long* time. This function can't
1172 * be called with interrupts off, or it may introduce deadlock with
1173 * smp_call_function() if an IPI is sent by the same process we are
1174 * waiting to become inactive.
1176 void wait_task_inactive(struct task_struct *p)
1178 unsigned long flags;
1180 struct prio_array *array;
1185 * We do the initial early heuristics without holding
1186 * any task-queue locks at all. We'll only try to get
1187 * the runqueue lock when things look like they will
1193 * If the task is actively running on another CPU
1194 * still, just relax and busy-wait without holding
1197 * NOTE! Since we don't hold any locks, it's not
1198 * even sure that "rq" stays as the right runqueue!
1199 * But we don't care, since "task_running()" will
1200 * return false if the runqueue has changed and p
1201 * is actually now running somewhere else!
1203 while (task_running(rq, p))
1207 * Ok, time to look more closely! We need the rq
1208 * lock now, to be *sure*. If we're wrong, we'll
1209 * just go back and repeat.
1211 rq = task_rq_lock(p, &flags);
1212 running = task_running(rq, p);
1214 task_rq_unlock(rq, &flags);
1217 * Was it really running after all now that we
1218 * checked with the proper locks actually held?
1220 * Oops. Go back and try again..
1222 if (unlikely(running)) {
1228 * It's not enough that it's not actively running,
1229 * it must be off the runqueue _entirely_, and not
1232 * So if it wa still runnable (but just not actively
1233 * running right now), it's preempted, and we should
1234 * yield - it could be a while.
1236 if (unlikely(array)) {
1242 * Ahh, all good. It wasn't running, and it wasn't
1243 * runnable, which means that it will never become
1244 * running in the future either. We're all done!
1249 * kick_process - kick a running thread to enter/exit the kernel
1250 * @p: the to-be-kicked thread
1252 * Cause a process which is running on another CPU to enter
1253 * kernel-mode, without any delay. (to get signals handled.)
1255 * NOTE: this function doesnt have to take the runqueue lock,
1256 * because all it wants to ensure is that the remote task enters
1257 * the kernel. If the IPI races and the task has been migrated
1258 * to another CPU then no harm is done and the purpose has been
1261 void kick_process(struct task_struct *p)
1267 if ((cpu != smp_processor_id()) && task_curr(p))
1268 smp_send_reschedule(cpu);
1273 * Return a low guess at the load of a migration-source cpu weighted
1274 * according to the scheduling class and "nice" value.
1276 * We want to under-estimate the load of migration sources, to
1277 * balance conservatively.
1279 static inline unsigned long source_load(int cpu, int type)
1281 struct rq *rq = cpu_rq(cpu);
1284 return rq->raw_weighted_load;
1286 return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1290 * Return a high guess at the load of a migration-target cpu weighted
1291 * according to the scheduling class and "nice" value.
1293 static inline unsigned long target_load(int cpu, int type)
1295 struct rq *rq = cpu_rq(cpu);
1298 return rq->raw_weighted_load;
1300 return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1304 * Return the average load per task on the cpu's run queue
1306 static inline unsigned long cpu_avg_load_per_task(int cpu)
1308 struct rq *rq = cpu_rq(cpu);
1309 unsigned long n = rq->nr_running;
1311 return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1315 * find_idlest_group finds and returns the least busy CPU group within the
1318 static struct sched_group *
1319 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1321 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1322 unsigned long min_load = ULONG_MAX, this_load = 0;
1323 int load_idx = sd->forkexec_idx;
1324 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1327 unsigned long load, avg_load;
1331 /* Skip over this group if it has no CPUs allowed */
1332 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1335 local_group = cpu_isset(this_cpu, group->cpumask);
1337 /* Tally up the load of all CPUs in the group */
1340 for_each_cpu_mask(i, group->cpumask) {
1341 /* Bias balancing toward cpus of our domain */
1343 load = source_load(i, load_idx);
1345 load = target_load(i, load_idx);
1350 /* Adjust by relative CPU power of the group */
1351 avg_load = sg_div_cpu_power(group,
1352 avg_load * SCHED_LOAD_SCALE);
1355 this_load = avg_load;
1357 } else if (avg_load < min_load) {
1358 min_load = avg_load;
1362 group = group->next;
1363 } while (group != sd->groups);
1365 if (!idlest || 100*this_load < imbalance*min_load)
1371 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1374 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1377 unsigned long load, min_load = ULONG_MAX;
1381 /* Traverse only the allowed CPUs */
1382 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1384 for_each_cpu_mask(i, tmp) {
1385 load = weighted_cpuload(i);
1387 if (load < min_load || (load == min_load && i == this_cpu)) {
1397 * sched_balance_self: balance the current task (running on cpu) in domains
1398 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1401 * Balance, ie. select the least loaded group.
1403 * Returns the target CPU number, or the same CPU if no balancing is needed.
1405 * preempt must be disabled.
1407 static int sched_balance_self(int cpu, int flag)
1409 struct task_struct *t = current;
1410 struct sched_domain *tmp, *sd = NULL;
1412 for_each_domain(cpu, tmp) {
1414 * If power savings logic is enabled for a domain, stop there.
1416 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1418 if (tmp->flags & flag)
1424 struct sched_group *group;
1425 int new_cpu, weight;
1427 if (!(sd->flags & flag)) {
1433 group = find_idlest_group(sd, t, cpu);
1439 new_cpu = find_idlest_cpu(group, t, cpu);
1440 if (new_cpu == -1 || new_cpu == cpu) {
1441 /* Now try balancing at a lower domain level of cpu */
1446 /* Now try balancing at a lower domain level of new_cpu */
1449 weight = cpus_weight(span);
1450 for_each_domain(cpu, tmp) {
1451 if (weight <= cpus_weight(tmp->span))
1453 if (tmp->flags & flag)
1456 /* while loop will break here if sd == NULL */
1462 #endif /* CONFIG_SMP */
1465 * wake_idle() will wake a task on an idle cpu if task->cpu is
1466 * not idle and an idle cpu is available. The span of cpus to
1467 * search starts with cpus closest then further out as needed,
1468 * so we always favor a closer, idle cpu.
1470 * Returns the CPU we should wake onto.
1472 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1473 static int wake_idle(int cpu, struct task_struct *p)
1476 struct sched_domain *sd;
1480 * If it is idle, then it is the best cpu to run this task.
1482 * This cpu is also the best, if it has more than one task already.
1483 * Siblings must be also busy(in most cases) as they didn't already
1484 * pickup the extra load from this cpu and hence we need not check
1485 * sibling runqueue info. This will avoid the checks and cache miss
1486 * penalities associated with that.
1488 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1491 for_each_domain(cpu, sd) {
1492 if (sd->flags & SD_WAKE_IDLE) {
1493 cpus_and(tmp, sd->span, p->cpus_allowed);
1494 for_each_cpu_mask(i, tmp) {
1505 static inline int wake_idle(int cpu, struct task_struct *p)
1512 * try_to_wake_up - wake up a thread
1513 * @p: the to-be-woken-up thread
1514 * @state: the mask of task states that can be woken
1515 * @sync: do a synchronous wakeup?
1517 * Put it on the run-queue if it's not already there. The "current"
1518 * thread is always on the run-queue (except when the actual
1519 * re-schedule is in progress), and as such you're allowed to do
1520 * the simpler "current->state = TASK_RUNNING" to mark yourself
1521 * runnable without the overhead of this.
1523 * returns failure only if the task is already active.
1525 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1527 int cpu, this_cpu, success = 0;
1528 unsigned long flags;
1532 struct sched_domain *sd, *this_sd = NULL;
1533 unsigned long load, this_load;
1537 rq = task_rq_lock(p, &flags);
1538 old_state = p->state;
1539 if (!(old_state & state))
1546 this_cpu = smp_processor_id();
1549 if (unlikely(task_running(rq, p)))
1554 schedstat_inc(rq, ttwu_cnt);
1555 if (cpu == this_cpu) {
1556 schedstat_inc(rq, ttwu_local);
1560 for_each_domain(this_cpu, sd) {
1561 if (cpu_isset(cpu, sd->span)) {
1562 schedstat_inc(sd, ttwu_wake_remote);
1568 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1572 * Check for affine wakeup and passive balancing possibilities.
1575 int idx = this_sd->wake_idx;
1576 unsigned int imbalance;
1578 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1580 load = source_load(cpu, idx);
1581 this_load = target_load(this_cpu, idx);
1583 new_cpu = this_cpu; /* Wake to this CPU if we can */
1585 if (this_sd->flags & SD_WAKE_AFFINE) {
1586 unsigned long tl = this_load;
1587 unsigned long tl_per_task;
1589 tl_per_task = cpu_avg_load_per_task(this_cpu);
1592 * If sync wakeup then subtract the (maximum possible)
1593 * effect of the currently running task from the load
1594 * of the current CPU:
1597 tl -= current->load_weight;
1600 tl + target_load(cpu, idx) <= tl_per_task) ||
1601 100*(tl + p->load_weight) <= imbalance*load) {
1603 * This domain has SD_WAKE_AFFINE and
1604 * p is cache cold in this domain, and
1605 * there is no bad imbalance.
1607 schedstat_inc(this_sd, ttwu_move_affine);
1613 * Start passive balancing when half the imbalance_pct
1616 if (this_sd->flags & SD_WAKE_BALANCE) {
1617 if (imbalance*this_load <= 100*load) {
1618 schedstat_inc(this_sd, ttwu_move_balance);
1624 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1626 new_cpu = wake_idle(new_cpu, p);
1627 if (new_cpu != cpu) {
1628 set_task_cpu(p, new_cpu);
1629 task_rq_unlock(rq, &flags);
1630 /* might preempt at this point */
1631 rq = task_rq_lock(p, &flags);
1632 old_state = p->state;
1633 if (!(old_state & state))
1638 this_cpu = smp_processor_id();
1643 #endif /* CONFIG_SMP */
1644 if (old_state == TASK_UNINTERRUPTIBLE) {
1645 rq->nr_uninterruptible--;
1647 * Tasks on involuntary sleep don't earn
1648 * sleep_avg beyond just interactive state.
1650 p->sleep_type = SLEEP_NONINTERACTIVE;
1654 * Tasks that have marked their sleep as noninteractive get
1655 * woken up with their sleep average not weighted in an
1658 if (old_state & TASK_NONINTERACTIVE)
1659 p->sleep_type = SLEEP_NONINTERACTIVE;
1662 activate_task(p, rq, cpu == this_cpu);
1664 * Sync wakeups (i.e. those types of wakeups where the waker
1665 * has indicated that it will leave the CPU in short order)
1666 * don't trigger a preemption, if the woken up task will run on
1667 * this cpu. (in this case the 'I will reschedule' promise of
1668 * the waker guarantees that the freshly woken up task is going
1669 * to be considered on this CPU.)
1671 if (!sync || cpu != this_cpu) {
1672 if (TASK_PREEMPTS_CURR(p, rq))
1673 resched_task(rq->curr);
1678 p->state = TASK_RUNNING;
1680 task_rq_unlock(rq, &flags);
1685 int fastcall wake_up_process(struct task_struct *p)
1687 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1688 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1690 EXPORT_SYMBOL(wake_up_process);
1692 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1694 return try_to_wake_up(p, state, 0);
1697 static void task_running_tick(struct rq *rq, struct task_struct *p);
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1702 void fastcall sched_fork(struct task_struct *p, int clone_flags)
1704 int cpu = get_cpu();
1707 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1709 set_task_cpu(p, cpu);
1712 * We mark the process as running here, but have not actually
1713 * inserted it onto the runqueue yet. This guarantees that
1714 * nobody will actually run it, and a signal or other external
1715 * event cannot wake it up and insert it on the runqueue either.
1717 p->state = TASK_RUNNING;
1720 * Make sure we do not leak PI boosting priority to the child:
1722 p->prio = current->normal_prio;
1724 INIT_LIST_HEAD(&p->run_list);
1726 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1727 if (unlikely(sched_info_on()))
1728 memset(&p->sched_info, 0, sizeof(p->sched_info));
1730 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1733 #ifdef CONFIG_PREEMPT
1734 /* Want to start with kernel preemption disabled. */
1735 task_thread_info(p)->preempt_count = 1;
1738 * Share the timeslice between parent and child, thus the
1739 * total amount of pending timeslices in the system doesn't change,
1740 * resulting in more scheduling fairness.
1742 local_irq_disable();
1743 p->time_slice = (current->time_slice + 1) >> 1;
1745 * The remainder of the first timeslice might be recovered by
1746 * the parent if the child exits early enough.
1748 p->first_time_slice = 1;
1749 current->time_slice >>= 1;
1750 p->timestamp = sched_clock();
1751 if (unlikely(!current->time_slice)) {
1753 * This case is rare, it happens when the parent has only
1754 * a single jiffy left from its timeslice. Taking the
1755 * runqueue lock is not a problem.
1757 current->time_slice = 1;
1758 task_running_tick(cpu_rq(cpu), current);
1765 * wake_up_new_task - wake up a newly created task for the first time.
1767 * This function will do some initial scheduler statistics housekeeping
1768 * that must be done for every newly created context, then puts the task
1769 * on the runqueue and wakes it.
1771 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1773 struct rq *rq, *this_rq;
1774 unsigned long flags;
1777 rq = task_rq_lock(p, &flags);
1778 BUG_ON(p->state != TASK_RUNNING);
1779 this_cpu = smp_processor_id();
1783 * We decrease the sleep average of forking parents
1784 * and children as well, to keep max-interactive tasks
1785 * from forking tasks that are max-interactive. The parent
1786 * (current) is done further down, under its lock.
1788 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1789 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1791 p->prio = effective_prio(p);
1793 if (likely(cpu == this_cpu)) {
1794 if (!(clone_flags & CLONE_VM)) {
1796 * The VM isn't cloned, so we're in a good position to
1797 * do child-runs-first in anticipation of an exec. This
1798 * usually avoids a lot of COW overhead.
1800 if (unlikely(!current->array))
1801 __activate_task(p, rq);
1803 p->prio = current->prio;
1804 p->normal_prio = current->normal_prio;
1805 list_add_tail(&p->run_list, ¤t->run_list);
1806 p->array = current->array;
1807 p->array->nr_active++;
1808 inc_nr_running(p, rq);
1812 /* Run child last */
1813 __activate_task(p, rq);
1815 * We skip the following code due to cpu == this_cpu
1817 * task_rq_unlock(rq, &flags);
1818 * this_rq = task_rq_lock(current, &flags);
1822 this_rq = cpu_rq(this_cpu);
1825 * Not the local CPU - must adjust timestamp. This should
1826 * get optimised away in the !CONFIG_SMP case.
1828 p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
1829 + rq->most_recent_timestamp;
1830 __activate_task(p, rq);
1831 if (TASK_PREEMPTS_CURR(p, rq))
1832 resched_task(rq->curr);
1835 * Parent and child are on different CPUs, now get the
1836 * parent runqueue to update the parent's ->sleep_avg:
1838 task_rq_unlock(rq, &flags);
1839 this_rq = task_rq_lock(current, &flags);
1841 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1842 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1843 task_rq_unlock(this_rq, &flags);
1847 * prepare_task_switch - prepare to switch tasks
1848 * @rq: the runqueue preparing to switch
1849 * @next: the task we are going to switch to.
1851 * This is called with the rq lock held and interrupts off. It must
1852 * be paired with a subsequent finish_task_switch after the context
1855 * prepare_task_switch sets up locking and calls architecture specific
1858 static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1860 prepare_lock_switch(rq, next);
1861 prepare_arch_switch(next);
1865 * finish_task_switch - clean up after a task-switch
1866 * @rq: runqueue associated with task-switch
1867 * @prev: the thread we just switched away from.
1869 * finish_task_switch must be called after the context switch, paired
1870 * with a prepare_task_switch call before the context switch.
1871 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1872 * and do any other architecture-specific cleanup actions.
1874 * Note that we may have delayed dropping an mm in context_switch(). If
1875 * so, we finish that here outside of the runqueue lock. (Doing it
1876 * with the lock held can cause deadlocks; see schedule() for
1879 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1880 __releases(rq->lock)
1882 struct mm_struct *mm = rq->prev_mm;
1888 * A task struct has one reference for the use as "current".
1889 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1890 * schedule one last time. The schedule call will never return, and
1891 * the scheduled task must drop that reference.
1892 * The test for TASK_DEAD must occur while the runqueue locks are
1893 * still held, otherwise prev could be scheduled on another cpu, die
1894 * there before we look at prev->state, and then the reference would
1896 * Manfred Spraul <manfred@colorfullife.com>
1898 prev_state = prev->state;
1899 finish_arch_switch(prev);
1900 finish_lock_switch(rq, prev);
1903 if (unlikely(prev_state == TASK_DEAD)) {
1905 * Remove function-return probe instances associated with this
1906 * task and put them back on the free list.
1908 kprobe_flush_task(prev);
1909 put_task_struct(prev);
1914 * schedule_tail - first thing a freshly forked thread must call.
1915 * @prev: the thread we just switched away from.
1917 asmlinkage void schedule_tail(struct task_struct *prev)
1918 __releases(rq->lock)
1920 struct rq *rq = this_rq();
1922 finish_task_switch(rq, prev);
1923 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1924 /* In this case, finish_task_switch does not reenable preemption */
1927 if (current->set_child_tid)
1928 put_user(current->pid, current->set_child_tid);
1932 * context_switch - switch to the new MM and the new
1933 * thread's register state.
1935 static inline struct task_struct *
1936 context_switch(struct rq *rq, struct task_struct *prev,
1937 struct task_struct *next)
1939 struct mm_struct *mm = next->mm;
1940 struct mm_struct *oldmm = prev->active_mm;
1943 * For paravirt, this is coupled with an exit in switch_to to
1944 * combine the page table reload and the switch backend into
1947 arch_enter_lazy_cpu_mode();
1950 next->active_mm = oldmm;
1951 atomic_inc(&oldmm->mm_count);
1952 enter_lazy_tlb(oldmm, next);
1954 switch_mm(oldmm, mm, next);
1957 prev->active_mm = NULL;
1958 WARN_ON(rq->prev_mm);
1959 rq->prev_mm = oldmm;
1962 * Since the runqueue lock will be released by the next
1963 * task (which is an invalid locking op but in the case
1964 * of the scheduler it's an obvious special-case), so we
1965 * do an early lockdep release here:
1967 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1968 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1971 /* Here we just switch the register state and the stack. */
1972 switch_to(prev, next, prev);
1978 * nr_running, nr_uninterruptible and nr_context_switches:
1980 * externally visible scheduler statistics: current number of runnable
1981 * threads, current number of uninterruptible-sleeping threads, total
1982 * number of context switches performed since bootup.
1984 unsigned long nr_running(void)
1986 unsigned long i, sum = 0;
1988 for_each_online_cpu(i)
1989 sum += cpu_rq(i)->nr_running;
1994 unsigned long nr_uninterruptible(void)
1996 unsigned long i, sum = 0;
1998 for_each_possible_cpu(i)
1999 sum += cpu_rq(i)->nr_uninterruptible;
2002 * Since we read the counters lockless, it might be slightly
2003 * inaccurate. Do not allow it to go below zero though:
2005 if (unlikely((long)sum < 0))
2011 unsigned long long nr_context_switches(void)
2014 unsigned long long sum = 0;
2016 for_each_possible_cpu(i)
2017 sum += cpu_rq(i)->nr_switches;
2022 unsigned long nr_iowait(void)
2024 unsigned long i, sum = 0;
2026 for_each_possible_cpu(i)
2027 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2032 unsigned long nr_active(void)
2034 unsigned long i, running = 0, uninterruptible = 0;
2036 for_each_online_cpu(i) {
2037 running += cpu_rq(i)->nr_running;
2038 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2041 if (unlikely((long)uninterruptible < 0))
2042 uninterruptible = 0;
2044 return running + uninterruptible;
2050 * Is this task likely cache-hot:
2053 task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
2055 return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
2059 * double_rq_lock - safely lock two runqueues
2061 * Note this does not disable interrupts like task_rq_lock,
2062 * you need to do so manually before calling.
2064 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2065 __acquires(rq1->lock)
2066 __acquires(rq2->lock)
2068 BUG_ON(!irqs_disabled());
2070 spin_lock(&rq1->lock);
2071 __acquire(rq2->lock); /* Fake it out ;) */
2074 spin_lock(&rq1->lock);
2075 spin_lock(&rq2->lock);
2077 spin_lock(&rq2->lock);
2078 spin_lock(&rq1->lock);
2084 * double_rq_unlock - safely unlock two runqueues
2086 * Note this does not restore interrupts like task_rq_unlock,
2087 * you need to do so manually after calling.
2089 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2090 __releases(rq1->lock)
2091 __releases(rq2->lock)
2093 spin_unlock(&rq1->lock);
2095 spin_unlock(&rq2->lock);
2097 __release(rq2->lock);
2101 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2103 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2104 __releases(this_rq->lock)
2105 __acquires(busiest->lock)
2106 __acquires(this_rq->lock)
2108 if (unlikely(!irqs_disabled())) {
2109 /* printk() doesn't work good under rq->lock */
2110 spin_unlock(&this_rq->lock);
2113 if (unlikely(!spin_trylock(&busiest->lock))) {
2114 if (busiest < this_rq) {
2115 spin_unlock(&this_rq->lock);
2116 spin_lock(&busiest->lock);
2117 spin_lock(&this_rq->lock);
2119 spin_lock(&busiest->lock);
2124 * If dest_cpu is allowed for this process, migrate the task to it.
2125 * This is accomplished by forcing the cpu_allowed mask to only
2126 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2127 * the cpu_allowed mask is restored.
2129 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2131 struct migration_req req;
2132 unsigned long flags;
2135 rq = task_rq_lock(p, &flags);
2136 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2137 || unlikely(cpu_is_offline(dest_cpu)))
2140 /* force the process onto the specified CPU */
2141 if (migrate_task(p, dest_cpu, &req)) {
2142 /* Need to wait for migration thread (might exit: take ref). */
2143 struct task_struct *mt = rq->migration_thread;
2145 get_task_struct(mt);
2146 task_rq_unlock(rq, &flags);
2147 wake_up_process(mt);
2148 put_task_struct(mt);
2149 wait_for_completion(&req.done);
2154 task_rq_unlock(rq, &flags);
2158 * sched_exec - execve() is a valuable balancing opportunity, because at
2159 * this point the task has the smallest effective memory and cache footprint.
2161 void sched_exec(void)
2163 int new_cpu, this_cpu = get_cpu();
2164 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2166 if (new_cpu != this_cpu)
2167 sched_migrate_task(current, new_cpu);
2171 * pull_task - move a task from a remote runqueue to the local runqueue.
2172 * Both runqueues must be locked.
2174 static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2175 struct task_struct *p, struct rq *this_rq,
2176 struct prio_array *this_array, int this_cpu)
2178 dequeue_task(p, src_array);
2179 dec_nr_running(p, src_rq);
2180 set_task_cpu(p, this_cpu);
2181 inc_nr_running(p, this_rq);
2182 enqueue_task(p, this_array);
2183 p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
2184 + this_rq->most_recent_timestamp;
2186 * Note that idle threads have a prio of MAX_PRIO, for this test
2187 * to be always true for them.
2189 if (TASK_PREEMPTS_CURR(p, this_rq))
2190 resched_task(this_rq->curr);
2194 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2197 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2198 struct sched_domain *sd, enum cpu_idle_type idle,
2202 * We do not migrate tasks that are:
2203 * 1) running (obviously), or
2204 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2205 * 3) are cache-hot on their current CPU.
2207 if (!cpu_isset(this_cpu, p->cpus_allowed))
2211 if (task_running(rq, p))
2215 * Aggressive migration if:
2216 * 1) task is cache cold, or
2217 * 2) too many balance attempts have failed.
2220 if (sd->nr_balance_failed > sd->cache_nice_tries) {
2221 #ifdef CONFIG_SCHEDSTATS
2222 if (task_hot(p, rq->most_recent_timestamp, sd))
2223 schedstat_inc(sd, lb_hot_gained[idle]);
2228 if (task_hot(p, rq->most_recent_timestamp, sd))
2233 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2236 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2237 * load from busiest to this_rq, as part of a balancing operation within
2238 * "domain". Returns the number of tasks moved.
2240 * Called with both runqueues locked.
2242 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2243 unsigned long max_nr_move, unsigned long max_load_move,
2244 struct sched_domain *sd, enum cpu_idle_type idle,
2247 int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2248 best_prio_seen, skip_for_load;
2249 struct prio_array *array, *dst_array;
2250 struct list_head *head, *curr;
2251 struct task_struct *tmp;
2254 if (max_nr_move == 0 || max_load_move == 0)
2257 rem_load_move = max_load_move;
2259 this_best_prio = rq_best_prio(this_rq);
2260 best_prio = rq_best_prio(busiest);
2262 * Enable handling of the case where there is more than one task
2263 * with the best priority. If the current running task is one
2264 * of those with prio==best_prio we know it won't be moved
2265 * and therefore it's safe to override the skip (based on load) of
2266 * any task we find with that prio.
2268 best_prio_seen = best_prio == busiest->curr->prio;
2271 * We first consider expired tasks. Those will likely not be
2272 * executed in the near future, and they are most likely to
2273 * be cache-cold, thus switching CPUs has the least effect
2276 if (busiest->expired->nr_active) {
2277 array = busiest->expired;
2278 dst_array = this_rq->expired;
2280 array = busiest->active;
2281 dst_array = this_rq->active;
2285 /* Start searching at priority 0: */
2289 idx = sched_find_first_bit(array->bitmap);
2291 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2292 if (idx >= MAX_PRIO) {
2293 if (array == busiest->expired && busiest->active->nr_active) {
2294 array = busiest->active;
2295 dst_array = this_rq->active;
2301 head = array->queue + idx;
2304 tmp = list_entry(curr, struct task_struct, run_list);
2309 * To help distribute high priority tasks accross CPUs we don't
2310 * skip a task if it will be the highest priority task (i.e. smallest
2311 * prio value) on its new queue regardless of its load weight
2313 skip_for_load = tmp->load_weight > rem_load_move;
2314 if (skip_for_load && idx < this_best_prio)
2315 skip_for_load = !best_prio_seen && idx == best_prio;
2316 if (skip_for_load ||
2317 !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2319 best_prio_seen |= idx == best_prio;
2326 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2328 rem_load_move -= tmp->load_weight;
2331 * We only want to steal up to the prescribed number of tasks
2332 * and the prescribed amount of weighted load.
2334 if (pulled < max_nr_move && rem_load_move > 0) {
2335 if (idx < this_best_prio)
2336 this_best_prio = idx;
2344 * Right now, this is the only place pull_task() is called,
2345 * so we can safely collect pull_task() stats here rather than
2346 * inside pull_task().
2348 schedstat_add(sd, lb_gained[idle], pulled);
2351 *all_pinned = pinned;
2356 * find_busiest_group finds and returns the busiest CPU group within the
2357 * domain. It calculates and returns the amount of weighted load which
2358 * should be moved to restore balance via the imbalance parameter.
2360 static struct sched_group *
2361 find_busiest_group(struct sched_domain *sd, int this_cpu,
2362 unsigned long *imbalance, enum cpu_idle_type idle, int *sd_idle,
2363 cpumask_t *cpus, int *balance)
2365 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2366 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2367 unsigned long max_pull;
2368 unsigned long busiest_load_per_task, busiest_nr_running;
2369 unsigned long this_load_per_task, this_nr_running;
2371 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2372 int power_savings_balance = 1;
2373 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2374 unsigned long min_nr_running = ULONG_MAX;
2375 struct sched_group *group_min = NULL, *group_leader = NULL;
2378 max_load = this_load = total_load = total_pwr = 0;
2379 busiest_load_per_task = busiest_nr_running = 0;
2380 this_load_per_task = this_nr_running = 0;
2381 if (idle == CPU_NOT_IDLE)
2382 load_idx = sd->busy_idx;
2383 else if (idle == CPU_NEWLY_IDLE)
2384 load_idx = sd->newidle_idx;
2386 load_idx = sd->idle_idx;
2389 unsigned long load, group_capacity;
2392 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2393 unsigned long sum_nr_running, sum_weighted_load;
2395 local_group = cpu_isset(this_cpu, group->cpumask);
2398 balance_cpu = first_cpu(group->cpumask);
2400 /* Tally up the load of all CPUs in the group */
2401 sum_weighted_load = sum_nr_running = avg_load = 0;
2403 for_each_cpu_mask(i, group->cpumask) {
2406 if (!cpu_isset(i, *cpus))
2411 if (*sd_idle && !idle_cpu(i))
2414 /* Bias balancing toward cpus of our domain */
2416 if (idle_cpu(i) && !first_idle_cpu) {
2421 load = target_load(i, load_idx);
2423 load = source_load(i, load_idx);
2426 sum_nr_running += rq->nr_running;
2427 sum_weighted_load += rq->raw_weighted_load;
2431 * First idle cpu or the first cpu(busiest) in this sched group
2432 * is eligible for doing load balancing at this and above
2435 if (local_group && balance_cpu != this_cpu && balance) {
2440 total_load += avg_load;
2441 total_pwr += group->__cpu_power;
2443 /* Adjust by relative CPU power of the group */
2444 avg_load = sg_div_cpu_power(group,
2445 avg_load * SCHED_LOAD_SCALE);
2447 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2450 this_load = avg_load;
2452 this_nr_running = sum_nr_running;
2453 this_load_per_task = sum_weighted_load;
2454 } else if (avg_load > max_load &&
2455 sum_nr_running > group_capacity) {
2456 max_load = avg_load;
2458 busiest_nr_running = sum_nr_running;
2459 busiest_load_per_task = sum_weighted_load;
2462 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2464 * Busy processors will not participate in power savings
2467 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2471 * If the local group is idle or completely loaded
2472 * no need to do power savings balance at this domain
2474 if (local_group && (this_nr_running >= group_capacity ||
2476 power_savings_balance = 0;
2479 * If a group is already running at full capacity or idle,
2480 * don't include that group in power savings calculations
2482 if (!power_savings_balance || sum_nr_running >= group_capacity
2487 * Calculate the group which has the least non-idle load.
2488 * This is the group from where we need to pick up the load
2491 if ((sum_nr_running < min_nr_running) ||
2492 (sum_nr_running == min_nr_running &&
2493 first_cpu(group->cpumask) <
2494 first_cpu(group_min->cpumask))) {
2496 min_nr_running = sum_nr_running;
2497 min_load_per_task = sum_weighted_load /
2502 * Calculate the group which is almost near its
2503 * capacity but still has some space to pick up some load
2504 * from other group and save more power
2506 if (sum_nr_running <= group_capacity - 1) {
2507 if (sum_nr_running > leader_nr_running ||
2508 (sum_nr_running == leader_nr_running &&
2509 first_cpu(group->cpumask) >
2510 first_cpu(group_leader->cpumask))) {
2511 group_leader = group;
2512 leader_nr_running = sum_nr_running;
2517 group = group->next;
2518 } while (group != sd->groups);
2520 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2523 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2525 if (this_load >= avg_load ||
2526 100*max_load <= sd->imbalance_pct*this_load)
2529 busiest_load_per_task /= busiest_nr_running;
2531 * We're trying to get all the cpus to the average_load, so we don't
2532 * want to push ourselves above the average load, nor do we wish to
2533 * reduce the max loaded cpu below the average load, as either of these
2534 * actions would just result in more rebalancing later, and ping-pong
2535 * tasks around. Thus we look for the minimum possible imbalance.
2536 * Negative imbalances (*we* are more loaded than anyone else) will
2537 * be counted as no imbalance for these purposes -- we can't fix that
2538 * by pulling tasks to us. Be careful of negative numbers as they'll
2539 * appear as very large values with unsigned longs.
2541 if (max_load <= busiest_load_per_task)
2545 * In the presence of smp nice balancing, certain scenarios can have
2546 * max load less than avg load(as we skip the groups at or below
2547 * its cpu_power, while calculating max_load..)
2549 if (max_load < avg_load) {
2551 goto small_imbalance;
2554 /* Don't want to pull so many tasks that a group would go idle */
2555 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2557 /* How much load to actually move to equalise the imbalance */
2558 *imbalance = min(max_pull * busiest->__cpu_power,
2559 (avg_load - this_load) * this->__cpu_power)
2563 * if *imbalance is less than the average load per runnable task
2564 * there is no gaurantee that any tasks will be moved so we'll have
2565 * a think about bumping its value to force at least one task to be
2568 if (*imbalance < busiest_load_per_task) {
2569 unsigned long tmp, pwr_now, pwr_move;
2573 pwr_move = pwr_now = 0;
2575 if (this_nr_running) {
2576 this_load_per_task /= this_nr_running;
2577 if (busiest_load_per_task > this_load_per_task)
2580 this_load_per_task = SCHED_LOAD_SCALE;
2582 if (max_load - this_load >= busiest_load_per_task * imbn) {
2583 *imbalance = busiest_load_per_task;
2588 * OK, we don't have enough imbalance to justify moving tasks,
2589 * however we may be able to increase total CPU power used by
2593 pwr_now += busiest->__cpu_power *
2594 min(busiest_load_per_task, max_load);
2595 pwr_now += this->__cpu_power *
2596 min(this_load_per_task, this_load);
2597 pwr_now /= SCHED_LOAD_SCALE;
2599 /* Amount of load we'd subtract */
2600 tmp = sg_div_cpu_power(busiest,
2601 busiest_load_per_task * SCHED_LOAD_SCALE);
2603 pwr_move += busiest->__cpu_power *
2604 min(busiest_load_per_task, max_load - tmp);
2606 /* Amount of load we'd add */
2607 if (max_load * busiest->__cpu_power <
2608 busiest_load_per_task * SCHED_LOAD_SCALE)
2609 tmp = sg_div_cpu_power(this,
2610 max_load * busiest->__cpu_power);
2612 tmp = sg_div_cpu_power(this,
2613 busiest_load_per_task * SCHED_LOAD_SCALE);
2614 pwr_move += this->__cpu_power *
2615 min(this_load_per_task, this_load + tmp);
2616 pwr_move /= SCHED_LOAD_SCALE;
2618 /* Move if we gain throughput */
2619 if (pwr_move <= pwr_now)
2622 *imbalance = busiest_load_per_task;
2628 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2629 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2632 if (this == group_leader && group_leader != group_min) {
2633 *imbalance = min_load_per_task;
2643 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2646 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2647 unsigned long imbalance, cpumask_t *cpus)
2649 struct rq *busiest = NULL, *rq;
2650 unsigned long max_load = 0;
2653 for_each_cpu_mask(i, group->cpumask) {
2655 if (!cpu_isset(i, *cpus))
2660 if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2663 if (rq->raw_weighted_load > max_load) {
2664 max_load = rq->raw_weighted_load;
2673 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2674 * so long as it is large enough.
2676 #define MAX_PINNED_INTERVAL 512
2678 static inline unsigned long minus_1_or_zero(unsigned long n)
2680 return n > 0 ? n - 1 : 0;
2684 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2685 * tasks if there is an imbalance.
2687 static int load_balance(int this_cpu, struct rq *this_rq,
2688 struct sched_domain *sd, enum cpu_idle_type idle,
2691 int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2692 struct sched_group *group;
2693 unsigned long imbalance;
2695 cpumask_t cpus = CPU_MASK_ALL;
2696 unsigned long flags;
2699 * When power savings policy is enabled for the parent domain, idle
2700 * sibling can pick up load irrespective of busy siblings. In this case,
2701 * let the state of idle sibling percolate up as IDLE, instead of
2702 * portraying it as CPU_NOT_IDLE.
2704 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2705 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2708 schedstat_inc(sd, lb_cnt[idle]);
2711 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2718 schedstat_inc(sd, lb_nobusyg[idle]);
2722 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2724 schedstat_inc(sd, lb_nobusyq[idle]);
2728 BUG_ON(busiest == this_rq);
2730 schedstat_add(sd, lb_imbalance[idle], imbalance);
2733 if (busiest->nr_running > 1) {
2735 * Attempt to move tasks. If find_busiest_group has found
2736 * an imbalance but busiest->nr_running <= 1, the group is
2737 * still unbalanced. nr_moved simply stays zero, so it is
2738 * correctly treated as an imbalance.
2740 local_irq_save(flags);
2741 double_rq_lock(this_rq, busiest);
2742 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2743 minus_1_or_zero(busiest->nr_running),
2744 imbalance, sd, idle, &all_pinned);
2745 double_rq_unlock(this_rq, busiest);
2746 local_irq_restore(flags);
2749 * some other cpu did the load balance for us.
2751 if (nr_moved && this_cpu != smp_processor_id())
2752 resched_cpu(this_cpu);
2754 /* All tasks on this runqueue were pinned by CPU affinity */
2755 if (unlikely(all_pinned)) {
2756 cpu_clear(cpu_of(busiest), cpus);
2757 if (!cpus_empty(cpus))
2764 schedstat_inc(sd, lb_failed[idle]);
2765 sd->nr_balance_failed++;
2767 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2769 spin_lock_irqsave(&busiest->lock, flags);
2771 /* don't kick the migration_thread, if the curr
2772 * task on busiest cpu can't be moved to this_cpu
2774 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2775 spin_unlock_irqrestore(&busiest->lock, flags);
2777 goto out_one_pinned;
2780 if (!busiest->active_balance) {
2781 busiest->active_balance = 1;
2782 busiest->push_cpu = this_cpu;
2785 spin_unlock_irqrestore(&busiest->lock, flags);
2787 wake_up_process(busiest->migration_thread);
2790 * We've kicked active balancing, reset the failure
2793 sd->nr_balance_failed = sd->cache_nice_tries+1;
2796 sd->nr_balance_failed = 0;
2798 if (likely(!active_balance)) {
2799 /* We were unbalanced, so reset the balancing interval */
2800 sd->balance_interval = sd->min_interval;
2803 * If we've begun active balancing, start to back off. This
2804 * case may not be covered by the all_pinned logic if there
2805 * is only 1 task on the busy runqueue (because we don't call
2808 if (sd->balance_interval < sd->max_interval)
2809 sd->balance_interval *= 2;
2812 if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2813 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2818 schedstat_inc(sd, lb_balanced[idle]);
2820 sd->nr_balance_failed = 0;
2823 /* tune up the balancing interval */
2824 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2825 (sd->balance_interval < sd->max_interval))
2826 sd->balance_interval *= 2;
2828 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2829 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2835 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2836 * tasks if there is an imbalance.
2838 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2839 * this_rq is locked.
2842 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2844 struct sched_group *group;
2845 struct rq *busiest = NULL;
2846 unsigned long imbalance;
2849 cpumask_t cpus = CPU_MASK_ALL;
2852 * When power savings policy is enabled for the parent domain, idle
2853 * sibling can pick up load irrespective of busy siblings. In this case,
2854 * let the state of idle sibling percolate up as IDLE, instead of
2855 * portraying it as CPU_NOT_IDLE.
2857 if (sd->flags & SD_SHARE_CPUPOWER &&
2858 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2861 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2863 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2864 &sd_idle, &cpus, NULL);
2866 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2870 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2873 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2877 BUG_ON(busiest == this_rq);
2879 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2882 if (busiest->nr_running > 1) {
2883 /* Attempt to move tasks */
2884 double_lock_balance(this_rq, busiest);
2885 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2886 minus_1_or_zero(busiest->nr_running),
2887 imbalance, sd, CPU_NEWLY_IDLE, NULL);
2888 spin_unlock(&busiest->lock);
2891 cpu_clear(cpu_of(busiest), cpus);
2892 if (!cpus_empty(cpus))
2898 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2899 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2900 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2903 sd->nr_balance_failed = 0;
2908 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2909 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2910 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2912 sd->nr_balance_failed = 0;
2918 * idle_balance is called by schedule() if this_cpu is about to become
2919 * idle. Attempts to pull tasks from other CPUs.
2921 static void idle_balance(int this_cpu, struct rq *this_rq)
2923 struct sched_domain *sd;
2924 int pulled_task = 0;
2925 unsigned long next_balance = jiffies + 60 * HZ;
2927 for_each_domain(this_cpu, sd) {
2928 unsigned long interval;
2930 if (!(sd->flags & SD_LOAD_BALANCE))
2933 if (sd->flags & SD_BALANCE_NEWIDLE)
2934 /* If we've pulled tasks over stop searching: */
2935 pulled_task = load_balance_newidle(this_cpu,
2938 interval = msecs_to_jiffies(sd->balance_interval);
2939 if (time_after(next_balance, sd->last_balance + interval))
2940 next_balance = sd->last_balance + interval;
2946 * We are going idle. next_balance may be set based on
2947 * a busy processor. So reset next_balance.
2949 this_rq->next_balance = next_balance;
2953 * active_load_balance is run by migration threads. It pushes running tasks
2954 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2955 * running on each physical CPU where possible, and avoids physical /
2956 * logical imbalances.
2958 * Called with busiest_rq locked.
2960 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2962 int target_cpu = busiest_rq->push_cpu;
2963 struct sched_domain *sd;
2964 struct rq *target_rq;
2966 /* Is there any task to move? */
2967 if (busiest_rq->nr_running <= 1)
2970 target_rq = cpu_rq(target_cpu);
2973 * This condition is "impossible", if it occurs
2974 * we need to fix it. Originally reported by
2975 * Bjorn Helgaas on a 128-cpu setup.
2977 BUG_ON(busiest_rq == target_rq);
2979 /* move a task from busiest_rq to target_rq */
2980 double_lock_balance(busiest_rq, target_rq);
2982 /* Search for an sd spanning us and the target CPU. */
2983 for_each_domain(target_cpu, sd) {
2984 if ((sd->flags & SD_LOAD_BALANCE) &&
2985 cpu_isset(busiest_cpu, sd->span))
2990 schedstat_inc(sd, alb_cnt);
2992 if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2993 RTPRIO_TO_LOAD_WEIGHT(100), sd, CPU_IDLE,
2995 schedstat_inc(sd, alb_pushed);
2997 schedstat_inc(sd, alb_failed);
2999 spin_unlock(&target_rq->lock);
3002 static void update_load(struct rq *this_rq)
3004 unsigned long this_load;
3005 unsigned int i, scale;
3007 this_load = this_rq->raw_weighted_load;
3009 /* Update our load: */
3010 for (i = 0, scale = 1; i < 3; i++, scale += scale) {
3011 unsigned long old_load, new_load;
3013 /* scale is effectively 1 << i now, and >> i divides by scale */
3015 old_load = this_rq->cpu_load[i];
3016 new_load = this_load;
3018 * Round up the averaging division if load is increasing. This
3019 * prevents us from getting stuck on 9 if the load is 10, for
3022 if (new_load > old_load)
3023 new_load += scale-1;
3024 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3030 atomic_t load_balancer;
3032 } nohz ____cacheline_aligned = {
3033 .load_balancer = ATOMIC_INIT(-1),
3034 .cpu_mask = CPU_MASK_NONE,
3038 * This routine will try to nominate the ilb (idle load balancing)
3039 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3040 * load balancing on behalf of all those cpus. If all the cpus in the system
3041 * go into this tickless mode, then there will be no ilb owner (as there is
3042 * no need for one) and all the cpus will sleep till the next wakeup event
3045 * For the ilb owner, tick is not stopped. And this tick will be used
3046 * for idle load balancing. ilb owner will still be part of
3049 * While stopping the tick, this cpu will become the ilb owner if there
3050 * is no other owner. And will be the owner till that cpu becomes busy
3051 * or if all cpus in the system stop their ticks at which point
3052 * there is no need for ilb owner.
3054 * When the ilb owner becomes busy, it nominates another owner, during the
3055 * next busy scheduler_tick()
3057 int select_nohz_load_balancer(int stop_tick)
3059 int cpu = smp_processor_id();
3062 cpu_set(cpu, nohz.cpu_mask);
3063 cpu_rq(cpu)->in_nohz_recently = 1;
3066 * If we are going offline and still the leader, give up!
3068 if (cpu_is_offline(cpu) &&
3069 atomic_read(&nohz.load_balancer) == cpu) {
3070 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3075 /* time for ilb owner also to sleep */
3076 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3077 if (atomic_read(&nohz.load_balancer) == cpu)
3078 atomic_set(&nohz.load_balancer, -1);
3082 if (atomic_read(&nohz.load_balancer) == -1) {
3083 /* make me the ilb owner */
3084 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3086 } else if (atomic_read(&nohz.load_balancer) == cpu)
3089 if (!cpu_isset(cpu, nohz.cpu_mask))
3092 cpu_clear(cpu, nohz.cpu_mask);
3094 if (atomic_read(&nohz.load_balancer) == cpu)
3095 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3102 static DEFINE_SPINLOCK(balancing);
3105 * It checks each scheduling domain to see if it is due to be balanced,
3106 * and initiates a balancing operation if so.
3108 * Balancing parameters are set up in arch_init_sched_domains.
3110 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3113 struct rq *rq = cpu_rq(cpu);
3114 unsigned long interval;
3115 struct sched_domain *sd;
3116 /* Earliest time when we have to do rebalance again */
3117 unsigned long next_balance = jiffies + 60*HZ;
3119 for_each_domain(cpu, sd) {
3120 if (!(sd->flags & SD_LOAD_BALANCE))
3123 interval = sd->balance_interval;
3124 if (idle != CPU_IDLE)
3125 interval *= sd->busy_factor;
3127 /* scale ms to jiffies */
3128 interval = msecs_to_jiffies(interval);
3129 if (unlikely(!interval))
3132 if (sd->flags & SD_SERIALIZE) {
3133 if (!spin_trylock(&balancing))
3137 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3138 if (load_balance(cpu, rq, sd, idle, &balance)) {
3140 * We've pulled tasks over so either we're no
3141 * longer idle, or one of our SMT siblings is
3144 idle = CPU_NOT_IDLE;
3146 sd->last_balance = jiffies;
3148 if (sd->flags & SD_SERIALIZE)
3149 spin_unlock(&balancing);
3151 if (time_after(next_balance, sd->last_balance + interval))
3152 next_balance = sd->last_balance + interval;
3155 * Stop the load balance at this level. There is another
3156 * CPU in our sched group which is doing load balancing more
3162 rq->next_balance = next_balance;
3166 * run_rebalance_domains is triggered when needed from the scheduler tick.
3167 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3168 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3170 static void run_rebalance_domains(struct softirq_action *h)
3172 int local_cpu = smp_processor_id();
3173 struct rq *local_rq = cpu_rq(local_cpu);
3174 enum cpu_idle_type idle = local_rq->idle_at_tick ? CPU_IDLE : CPU_NOT_IDLE;
3176 rebalance_domains(local_cpu, idle);
3180 * If this cpu is the owner for idle load balancing, then do the
3181 * balancing on behalf of the other idle cpus whose ticks are
3184 if (local_rq->idle_at_tick &&
3185 atomic_read(&nohz.load_balancer) == local_cpu) {
3186 cpumask_t cpus = nohz.cpu_mask;
3190 cpu_clear(local_cpu, cpus);
3191 for_each_cpu_mask(balance_cpu, cpus) {
3193 * If this cpu gets work to do, stop the load balancing
3194 * work being done for other cpus. Next load
3195 * balancing owner will pick it up.
3200 rebalance_domains(balance_cpu, CPU_IDLE);
3202 rq = cpu_rq(balance_cpu);
3203 if (time_after(local_rq->next_balance, rq->next_balance))
3204 local_rq->next_balance = rq->next_balance;
3211 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3213 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3214 * idle load balancing owner or decide to stop the periodic load balancing,
3215 * if the whole system is idle.
3217 static inline void trigger_load_balance(int cpu)
3219 struct rq *rq = cpu_rq(cpu);
3222 * If we were in the nohz mode recently and busy at the current
3223 * scheduler tick, then check if we need to nominate new idle
3226 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3227 rq->in_nohz_recently = 0;
3229 if (atomic_read(&nohz.load_balancer) == cpu) {
3230 cpu_clear(cpu, nohz.cpu_mask);
3231 atomic_set(&nohz.load_balancer, -1);
3234 if (atomic_read(&nohz.load_balancer) == -1) {
3236 * simple selection for now: Nominate the
3237 * first cpu in the nohz list to be the next
3240 * TBD: Traverse the sched domains and nominate
3241 * the nearest cpu in the nohz.cpu_mask.
3243 int ilb = first_cpu(nohz.cpu_mask);
3251 * If this cpu is idle and doing idle load balancing for all the
3252 * cpus with ticks stopped, is it time for that to stop?
3254 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3255 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3261 * If this cpu is idle and the idle load balancing is done by
3262 * someone else, then no need raise the SCHED_SOFTIRQ
3264 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3265 cpu_isset(cpu, nohz.cpu_mask))
3268 if (time_after_eq(jiffies, rq->next_balance))
3269 raise_softirq(SCHED_SOFTIRQ);
3273 * on UP we do not need to balance between CPUs:
3275 static inline void idle_balance(int cpu, struct rq *rq)
3280 DEFINE_PER_CPU(struct kernel_stat, kstat);
3282 EXPORT_PER_CPU_SYMBOL(kstat);
3285 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3286 * that have not yet been banked in case the task is currently running.
3288 unsigned long long task_sched_runtime(struct task_struct *p)
3290 unsigned long flags;
3294 rq = task_rq_lock(p, &flags);
3295 ns = p->se.sum_exec_runtime;
3296 if (rq->curr == p) {
3297 delta_exec = rq_clock(rq) - p->se.exec_start;
3298 if ((s64)delta_exec > 0)
3301 task_rq_unlock(rq, &flags);
3307 * We place interactive tasks back into the active array, if possible.
3309 * To guarantee that this does not starve expired tasks we ignore the
3310 * interactivity of a task if the first expired task had to wait more
3311 * than a 'reasonable' amount of time. This deadline timeout is
3312 * load-dependent, as the frequency of array switched decreases with
3313 * increasing number of running tasks. We also ignore the interactivity
3314 * if a better static_prio task has expired:
3316 static inline int expired_starving(struct rq *rq)
3318 if (rq->curr->static_prio > rq->best_expired_prio)
3320 if (!STARVATION_LIMIT || !rq->expired_timestamp)
3322 if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
3328 * Account user cpu time to a process.
3329 * @p: the process that the cpu time gets accounted to
3330 * @hardirq_offset: the offset to subtract from hardirq_count()
3331 * @cputime: the cpu time spent in user space since the last update
3333 void account_user_time(struct task_struct *p, cputime_t cputime)
3335 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3338 p->utime = cputime_add(p->utime, cputime);
3340 /* Add user time to cpustat. */
3341 tmp = cputime_to_cputime64(cputime);
3342 if (TASK_NICE(p) > 0)
3343 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3345 cpustat->user = cputime64_add(cpustat->user, tmp);
3349 * Account system cpu time to a process.
3350 * @p: the process that the cpu time gets accounted to
3351 * @hardirq_offset: the offset to subtract from hardirq_count()
3352 * @cputime: the cpu time spent in kernel space since the last update
3354 void account_system_time(struct task_struct *p, int hardirq_offset,
3357 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3358 struct rq *rq = this_rq();
3361 p->stime = cputime_add(p->stime, cputime);
3363 /* Add system time to cpustat. */
3364 tmp = cputime_to_cputime64(cputime);
3365 if (hardirq_count() - hardirq_offset)
3366 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3367 else if (softirq_count())
3368 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3369 else if (p != rq->idle)
3370 cpustat->system = cputime64_add(cpustat->system, tmp);
3371 else if (atomic_read(&rq->nr_iowait) > 0)
3372 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3374 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3375 /* Account for system time used */
3376 acct_update_integrals(p);
3380 * Account for involuntary wait time.
3381 * @p: the process from which the cpu time has been stolen
3382 * @steal: the cpu time spent in involuntary wait
3384 void account_steal_time(struct task_struct *p, cputime_t steal)
3386 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3387 cputime64_t tmp = cputime_to_cputime64(steal);
3388 struct rq *rq = this_rq();
3390 if (p == rq->idle) {
3391 p->stime = cputime_add(p->stime, steal);
3392 if (atomic_read(&rq->nr_iowait) > 0)
3393 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3395 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3397 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3400 static void task_running_tick(struct rq *rq, struct task_struct *p)
3402 if (p->array != rq->active) {
3403 /* Task has expired but was not scheduled yet */
3404 set_tsk_need_resched(p);
3407 spin_lock(&rq->lock);
3409 * The task was running during this tick - update the
3410 * time slice counter. Note: we do not update a thread's
3411 * priority until it either goes to sleep or uses up its
3412 * timeslice. This makes it possible for interactive tasks
3413 * to use up their timeslices at their highest priority levels.
3417 * RR tasks need a special form of timeslice management.
3418 * FIFO tasks have no timeslices.
3420 if ((p->policy == SCHED_RR) && !--p->time_slice) {
3421 p->time_slice = task_timeslice(p);
3422 p->first_time_slice = 0;
3423 set_tsk_need_resched(p);
3425 /* put it at the end of the queue: */
3426 requeue_task(p, rq->active);
3430 if (!--p->time_slice) {
3431 dequeue_task(p, rq->active);
3432 set_tsk_need_resched(p);
3433 p->prio = effective_prio(p);
3434 p->time_slice = task_timeslice(p);
3435 p->first_time_slice = 0;
3437 if (!rq->expired_timestamp)
3438 rq->expired_timestamp = jiffies;
3439 if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3440 enqueue_task(p, rq->expired);
3441 if (p->static_prio < rq->best_expired_prio)
3442 rq->best_expired_prio = p->static_prio;
3444 enqueue_task(p, rq->active);
3447 * Prevent a too long timeslice allowing a task to monopolize
3448 * the CPU. We do this by splitting up the timeslice into
3451 * Note: this does not mean the task's timeslices expire or
3452 * get lost in any way, they just might be preempted by
3453 * another task of equal priority. (one with higher
3454 * priority would have preempted this task already.) We
3455 * requeue this task to the end of the list on this priority
3456 * level, which is in essence a round-robin of tasks with
3459 * This only applies to tasks in the interactive
3460 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3462 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3463 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3464 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3465 (p->array == rq->active)) {
3467 requeue_task(p, rq->active);
3468 set_tsk_need_resched(p);
3472 spin_unlock(&rq->lock);
3476 * This function gets called by the timer code, with HZ frequency.
3477 * We call it with interrupts disabled.
3479 * It also gets called by the fork code, when changing the parent's
3482 void scheduler_tick(void)
3484 struct task_struct *p = current;
3485 int cpu = smp_processor_id();
3486 int idle_at_tick = idle_cpu(cpu);
3487 struct rq *rq = cpu_rq(cpu);
3490 task_running_tick(rq, p);
3493 rq->idle_at_tick = idle_at_tick;
3494 trigger_load_balance(cpu);
3498 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3500 void fastcall add_preempt_count(int val)
3505 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3507 preempt_count() += val;
3509 * Spinlock count overflowing soon?
3511 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3514 EXPORT_SYMBOL(add_preempt_count);
3516 void fastcall sub_preempt_count(int val)
3521 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3524 * Is the spinlock portion underflowing?
3526 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3527 !(preempt_count() & PREEMPT_MASK)))
3530 preempt_count() -= val;
3532 EXPORT_SYMBOL(sub_preempt_count);
3536 static inline int interactive_sleep(enum sleep_type sleep_type)
3538 return (sleep_type == SLEEP_INTERACTIVE ||
3539 sleep_type == SLEEP_INTERRUPTED);
3543 * schedule() is the main scheduler function.
3545 asmlinkage void __sched schedule(void)
3547 struct task_struct *prev, *next;
3548 struct prio_array *array;
3549 struct list_head *queue;
3550 unsigned long long now;
3551 unsigned long run_time;
3552 int cpu, idx, new_prio;
3557 * Test if we are atomic. Since do_exit() needs to call into
3558 * schedule() atomically, we ignore that path for now.
3559 * Otherwise, whine if we are scheduling when we should not be.
3561 if (unlikely(in_atomic() && !current->exit_state)) {
3562 printk(KERN_ERR "BUG: scheduling while atomic: "
3564 current->comm, preempt_count(), current->pid);
3565 debug_show_held_locks(current);
3566 if (irqs_disabled())
3567 print_irqtrace_events(current);
3570 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3575 release_kernel_lock(prev);
3576 need_resched_nonpreemptible:
3580 * The idle thread is not allowed to schedule!
3581 * Remove this check after it has been exercised a bit.
3583 if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3584 printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3588 schedstat_inc(rq, sched_cnt);
3589 now = sched_clock();
3590 if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3591 run_time = now - prev->timestamp;
3592 if (unlikely((long long)(now - prev->timestamp) < 0))
3595 run_time = NS_MAX_SLEEP_AVG;
3598 * Tasks charged proportionately less run_time at high sleep_avg to
3599 * delay them losing their interactive status
3601 run_time /= (CURRENT_BONUS(prev) ? : 1);
3603 spin_lock_irq(&rq->lock);
3605 switch_count = &prev->nivcsw;
3606 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3607 switch_count = &prev->nvcsw;
3608 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3609 unlikely(signal_pending(prev))))
3610 prev->state = TASK_RUNNING;
3612 if (prev->state == TASK_UNINTERRUPTIBLE)
3613 rq->nr_uninterruptible++;
3614 deactivate_task(prev, rq);
3618 cpu = smp_processor_id();
3619 if (unlikely(!rq->nr_running)) {
3620 idle_balance(cpu, rq);
3621 if (!rq->nr_running) {
3623 rq->expired_timestamp = 0;
3629 if (unlikely(!array->nr_active)) {
3631 * Switch the active and expired arrays.
3633 schedstat_inc(rq, sched_switch);
3634 rq->active = rq->expired;
3635 rq->expired = array;
3637 rq->expired_timestamp = 0;
3638 rq->best_expired_prio = MAX_PRIO;
3641 idx = sched_find_first_bit(array->bitmap);
3642 queue = array->queue + idx;
3643 next = list_entry(queue->next, struct task_struct, run_list);
3645 if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3646 unsigned long long delta = now - next->timestamp;
3647 if (unlikely((long long)(now - next->timestamp) < 0))
3650 if (next->sleep_type == SLEEP_INTERACTIVE)
3651 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3653 array = next->array;
3654 new_prio = recalc_task_prio(next, next->timestamp + delta);
3656 if (unlikely(next->prio != new_prio)) {
3657 dequeue_task(next, array);
3658 next->prio = new_prio;
3659 enqueue_task(next, array);
3662 next->sleep_type = SLEEP_NORMAL;
3664 if (next == rq->idle)
3665 schedstat_inc(rq, sched_goidle);
3667 prefetch_stack(next);
3668 clear_tsk_need_resched(prev);
3669 rcu_qsctr_inc(task_cpu(prev));
3671 prev->sleep_avg -= run_time;
3672 if ((long)prev->sleep_avg <= 0)
3673 prev->sleep_avg = 0;
3674 prev->timestamp = prev->last_ran = now;
3676 sched_info_switch(prev, next);
3677 if (likely(prev != next)) {
3678 next->timestamp = next->last_ran = now;
3683 prepare_task_switch(rq, next);
3684 prev = context_switch(rq, prev, next);
3687 * this_rq must be evaluated again because prev may have moved
3688 * CPUs since it called schedule(), thus the 'rq' on its stack
3689 * frame will be invalid.
3691 finish_task_switch(this_rq(), prev);
3693 spin_unlock_irq(&rq->lock);
3696 if (unlikely(reacquire_kernel_lock(prev) < 0))
3697 goto need_resched_nonpreemptible;
3698 preempt_enable_no_resched();
3699 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3702 EXPORT_SYMBOL(schedule);
3704 #ifdef CONFIG_PREEMPT
3706 * this is the entry point to schedule() from in-kernel preemption
3707 * off of preempt_enable. Kernel preemptions off return from interrupt
3708 * occur there and call schedule directly.
3710 asmlinkage void __sched preempt_schedule(void)
3712 struct thread_info *ti = current_thread_info();
3713 #ifdef CONFIG_PREEMPT_BKL
3714 struct task_struct *task = current;
3715 int saved_lock_depth;
3718 * If there is a non-zero preempt_count or interrupts are disabled,
3719 * we do not want to preempt the current task. Just return..
3721 if (likely(ti->preempt_count || irqs_disabled()))
3725 add_preempt_count(PREEMPT_ACTIVE);
3727 * We keep the big kernel semaphore locked, but we
3728 * clear ->lock_depth so that schedule() doesnt
3729 * auto-release the semaphore:
3731 #ifdef CONFIG_PREEMPT_BKL
3732 saved_lock_depth = task->lock_depth;
3733 task->lock_depth = -1;
3736 #ifdef CONFIG_PREEMPT_BKL
3737 task->lock_depth = saved_lock_depth;
3739 sub_preempt_count(PREEMPT_ACTIVE);
3741 /* we could miss a preemption opportunity between schedule and now */
3743 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3746 EXPORT_SYMBOL(preempt_schedule);
3749 * this is the entry point to schedule() from kernel preemption
3750 * off of irq context.
3751 * Note, that this is called and return with irqs disabled. This will
3752 * protect us against recursive calling from irq.
3754 asmlinkage void __sched preempt_schedule_irq(void)
3756 struct thread_info *ti = current_thread_info();
3757 #ifdef CONFIG_PREEMPT_BKL
3758 struct task_struct *task = current;
3759 int saved_lock_depth;
3761 /* Catch callers which need to be fixed */
3762 BUG_ON(ti->preempt_count || !irqs_disabled());
3765 add_preempt_count(PREEMPT_ACTIVE);
3767 * We keep the big kernel semaphore locked, but we
3768 * clear ->lock_depth so that schedule() doesnt
3769 * auto-release the semaphore:
3771 #ifdef CONFIG_PREEMPT_BKL
3772 saved_lock_depth = task->lock_depth;
3773 task->lock_depth = -1;
3777 local_irq_disable();
3778 #ifdef CONFIG_PREEMPT_BKL
3779 task->lock_depth = saved_lock_depth;
3781 sub_preempt_count(PREEMPT_ACTIVE);
3783 /* we could miss a preemption opportunity between schedule and now */
3785 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3789 #endif /* CONFIG_PREEMPT */
3791 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3794 return try_to_wake_up(curr->private, mode, sync);
3796 EXPORT_SYMBOL(default_wake_function);
3799 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3800 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3801 * number) then we wake all the non-exclusive tasks and one exclusive task.
3803 * There are circumstances in which we can try to wake a task which has already
3804 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3805 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3807 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3808 int nr_exclusive, int sync, void *key)
3810 struct list_head *tmp, *next;
3812 list_for_each_safe(tmp, next, &q->task_list) {
3813 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3814 unsigned flags = curr->flags;
3816 if (curr->func(curr, mode, sync, key) &&
3817 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3823 * __wake_up - wake up threads blocked on a waitqueue.
3825 * @mode: which threads
3826 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3827 * @key: is directly passed to the wakeup function
3829 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3830 int nr_exclusive, void *key)
3832 unsigned long flags;
3834 spin_lock_irqsave(&q->lock, flags);
3835 __wake_up_common(q, mode, nr_exclusive, 0, key);
3836 spin_unlock_irqrestore(&q->lock, flags);
3838 EXPORT_SYMBOL(__wake_up);
3841 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3843 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3845 __wake_up_common(q, mode, 1, 0, NULL);
3849 * __wake_up_sync - wake up threads blocked on a waitqueue.
3851 * @mode: which threads
3852 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3854 * The sync wakeup differs that the waker knows that it will schedule
3855 * away soon, so while the target thread will be woken up, it will not
3856 * be migrated to another CPU - ie. the two threads are 'synchronized'
3857 * with each other. This can prevent needless bouncing between CPUs.
3859 * On UP it can prevent extra preemption.
3862 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3864 unsigned long flags;
3870 if (unlikely(!nr_exclusive))
3873 spin_lock_irqsave(&q->lock, flags);
3874 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3875 spin_unlock_irqrestore(&q->lock, flags);
3877 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3879 void fastcall complete(struct completion *x)
3881 unsigned long flags;
3883 spin_lock_irqsave(&x->wait.lock, flags);
3885 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3887 spin_unlock_irqrestore(&x->wait.lock, flags);
3889 EXPORT_SYMBOL(complete);
3891 void fastcall complete_all(struct completion *x)
3893 unsigned long flags;
3895 spin_lock_irqsave(&x->wait.lock, flags);
3896 x->done += UINT_MAX/2;
3897 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3899 spin_unlock_irqrestore(&x->wait.lock, flags);
3901 EXPORT_SYMBOL(complete_all);
3903 void fastcall __sched wait_for_completion(struct completion *x)
3907 spin_lock_irq(&x->wait.lock);
3909 DECLARE_WAITQUEUE(wait, current);
3911 wait.flags |= WQ_FLAG_EXCLUSIVE;
3912 __add_wait_queue_tail(&x->wait, &wait);
3914 __set_current_state(TASK_UNINTERRUPTIBLE);
3915 spin_unlock_irq(&x->wait.lock);
3917 spin_lock_irq(&x->wait.lock);
3919 __remove_wait_queue(&x->wait, &wait);
3922 spin_unlock_irq(&x->wait.lock);
3924 EXPORT_SYMBOL(wait_for_completion);
3926 unsigned long fastcall __sched
3927 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3931 spin_lock_irq(&x->wait.lock);
3933 DECLARE_WAITQUEUE(wait, current);
3935 wait.flags |= WQ_FLAG_EXCLUSIVE;
3936 __add_wait_queue_tail(&x->wait, &wait);
3938 __set_current_state(TASK_UNINTERRUPTIBLE);
3939 spin_unlock_irq(&x->wait.lock);
3940 timeout = schedule_timeout(timeout);
3941 spin_lock_irq(&x->wait.lock);
3943 __remove_wait_queue(&x->wait, &wait);
3947 __remove_wait_queue(&x->wait, &wait);
3951 spin_unlock_irq(&x->wait.lock);
3954 EXPORT_SYMBOL(wait_for_completion_timeout);
3956 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3962 spin_lock_irq(&x->wait.lock);
3964 DECLARE_WAITQUEUE(wait, current);
3966 wait.flags |= WQ_FLAG_EXCLUSIVE;
3967 __add_wait_queue_tail(&x->wait, &wait);
3969 if (signal_pending(current)) {
3971 __remove_wait_queue(&x->wait, &wait);
3974 __set_current_state(TASK_INTERRUPTIBLE);
3975 spin_unlock_irq(&x->wait.lock);
3977 spin_lock_irq(&x->wait.lock);
3979 __remove_wait_queue(&x->wait, &wait);
3983 spin_unlock_irq(&x->wait.lock);
3987 EXPORT_SYMBOL(wait_for_completion_interruptible);
3989 unsigned long fastcall __sched
3990 wait_for_completion_interruptible_timeout(struct completion *x,
3991 unsigned long timeout)
3995 spin_lock_irq(&x->wait.lock);
3997 DECLARE_WAITQUEUE(wait, current);
3999 wait.flags |= WQ_FLAG_EXCLUSIVE;
4000 __add_wait_queue_tail(&x->wait, &wait);
4002 if (signal_pending(current)) {
4003 timeout = -ERESTARTSYS;
4004 __remove_wait_queue(&x->wait, &wait);
4007 __set_current_state(TASK_INTERRUPTIBLE);
4008 spin_unlock_irq(&x->wait.lock);
4009 timeout = schedule_timeout(timeout);
4010 spin_lock_irq(&x->wait.lock);
4012 __remove_wait_queue(&x->wait, &wait);
4016 __remove_wait_queue(&x->wait, &wait);
4020 spin_unlock_irq(&x->wait.lock);
4023 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4026 #define SLEEP_ON_VAR \
4027 unsigned long flags; \
4028 wait_queue_t wait; \
4029 init_waitqueue_entry(&wait, current);
4031 #define SLEEP_ON_HEAD \
4032 spin_lock_irqsave(&q->lock,flags); \
4033 __add_wait_queue(q, &wait); \
4034 spin_unlock(&q->lock);
4036 #define SLEEP_ON_TAIL \
4037 spin_lock_irq(&q->lock); \
4038 __remove_wait_queue(q, &wait); \
4039 spin_unlock_irqrestore(&q->lock, flags);
4041 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
4045 current->state = TASK_INTERRUPTIBLE;
4051 EXPORT_SYMBOL(interruptible_sleep_on);
4053 long fastcall __sched
4054 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4058 current->state = TASK_INTERRUPTIBLE;
4061 timeout = schedule_timeout(timeout);
4066 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4068 void fastcall __sched sleep_on(wait_queue_head_t *q)
4072 current->state = TASK_UNINTERRUPTIBLE;
4078 EXPORT_SYMBOL(sleep_on);
4080 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4084 current->state = TASK_UNINTERRUPTIBLE;
4087 timeout = schedule_timeout(timeout);
4093 EXPORT_SYMBOL(sleep_on_timeout);
4095 #ifdef CONFIG_RT_MUTEXES
4098 * rt_mutex_setprio - set the current priority of a task
4100 * @prio: prio value (kernel-internal form)
4102 * This function changes the 'effective' priority of a task. It does
4103 * not touch ->normal_prio like __setscheduler().
4105 * Used by the rt_mutex code to implement priority inheritance logic.
4107 void rt_mutex_setprio(struct task_struct *p, int prio)
4109 struct prio_array *array;
4110 unsigned long flags;
4114 BUG_ON(prio < 0 || prio > MAX_PRIO);
4116 rq = task_rq_lock(p, &flags);
4121 dequeue_task(p, array);
4126 * If changing to an RT priority then queue it
4127 * in the active array!
4131 enqueue_task(p, array);
4133 * Reschedule if we are currently running on this runqueue and
4134 * our priority decreased, or if we are not currently running on
4135 * this runqueue and our priority is higher than the current's
4137 if (task_running(rq, p)) {
4138 if (p->prio > oldprio)
4139 resched_task(rq->curr);
4140 } else if (TASK_PREEMPTS_CURR(p, rq))
4141 resched_task(rq->curr);
4143 task_rq_unlock(rq, &flags);
4148 void set_user_nice(struct task_struct *p, long nice)
4150 struct prio_array *array;
4151 int old_prio, delta;
4152 unsigned long flags;
4155 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4158 * We have to be careful, if called from sys_setpriority(),
4159 * the task might be in the middle of scheduling on another CPU.
4161 rq = task_rq_lock(p, &flags);
4163 * The RT priorities are set via sched_setscheduler(), but we still
4164 * allow the 'normal' nice value to be set - but as expected
4165 * it wont have any effect on scheduling until the task is
4166 * not SCHED_NORMAL/SCHED_BATCH:
4168 if (task_has_rt_policy(p)) {
4169 p->static_prio = NICE_TO_PRIO(nice);
4174 dequeue_task(p, array);
4175 dec_raw_weighted_load(rq, p);
4178 p->static_prio = NICE_TO_PRIO(nice);
4181 p->prio = effective_prio(p);
4182 delta = p->prio - old_prio;
4185 enqueue_task(p, array);
4186 inc_raw_weighted_load(rq, p);
4188 * If the task increased its priority or is running and
4189 * lowered its priority, then reschedule its CPU:
4191 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4192 resched_task(rq->curr);
4195 task_rq_unlock(rq, &flags);
4197 EXPORT_SYMBOL(set_user_nice);
4200 * can_nice - check if a task can reduce its nice value
4204 int can_nice(const struct task_struct *p, const int nice)
4206 /* convert nice value [19,-20] to rlimit style value [1,40] */
4207 int nice_rlim = 20 - nice;
4209 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4210 capable(CAP_SYS_NICE));
4213 #ifdef __ARCH_WANT_SYS_NICE
4216 * sys_nice - change the priority of the current process.
4217 * @increment: priority increment
4219 * sys_setpriority is a more generic, but much slower function that
4220 * does similar things.
4222 asmlinkage long sys_nice(int increment)
4227 * Setpriority might change our priority at the same moment.
4228 * We don't have to worry. Conceptually one call occurs first
4229 * and we have a single winner.
4231 if (increment < -40)
4236 nice = PRIO_TO_NICE(current->static_prio) + increment;
4242 if (increment < 0 && !can_nice(current, nice))
4245 retval = security_task_setnice(current, nice);
4249 set_user_nice(current, nice);
4256 * task_prio - return the priority value of a given task.
4257 * @p: the task in question.
4259 * This is the priority value as seen by users in /proc.
4260 * RT tasks are offset by -200. Normal tasks are centered
4261 * around 0, value goes from -16 to +15.
4263 int task_prio(const struct task_struct *p)
4265 return p->prio - MAX_RT_PRIO;
4269 * task_nice - return the nice value of a given task.
4270 * @p: the task in question.
4272 int task_nice(const struct task_struct *p)
4274 return TASK_NICE(p);
4276 EXPORT_SYMBOL_GPL(task_nice);
4279 * idle_cpu - is a given cpu idle currently?
4280 * @cpu: the processor in question.
4282 int idle_cpu(int cpu)
4284 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4288 * idle_task - return the idle task for a given cpu.
4289 * @cpu: the processor in question.
4291 struct task_struct *idle_task(int cpu)
4293 return cpu_rq(cpu)->idle;
4297 * find_process_by_pid - find a process with a matching PID value.
4298 * @pid: the pid in question.
4300 static inline struct task_struct *find_process_by_pid(pid_t pid)
4302 return pid ? find_task_by_pid(pid) : current;
4305 /* Actually do priority change: must hold rq lock. */
4306 static void __setscheduler(struct task_struct *p, int policy, int prio)
4311 p->rt_priority = prio;
4312 p->normal_prio = normal_prio(p);
4313 /* we are holding p->pi_lock already */
4314 p->prio = rt_mutex_getprio(p);
4316 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4318 if (policy == SCHED_BATCH)
4324 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4325 * @p: the task in question.
4326 * @policy: new policy.
4327 * @param: structure containing the new RT priority.
4329 * NOTE that the task may be already dead.
4331 int sched_setscheduler(struct task_struct *p, int policy,
4332 struct sched_param *param)
4334 int retval, oldprio, oldpolicy = -1;
4335 struct prio_array *array;
4336 unsigned long flags;
4339 /* may grab non-irq protected spin_locks */
4340 BUG_ON(in_interrupt());
4342 /* double check policy once rq lock held */
4344 policy = oldpolicy = p->policy;
4345 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4346 policy != SCHED_NORMAL && policy != SCHED_BATCH)
4349 * Valid priorities for SCHED_FIFO and SCHED_RR are
4350 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4353 if (param->sched_priority < 0 ||
4354 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4355 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4357 if (rt_policy(policy) != (param->sched_priority != 0))
4361 * Allow unprivileged RT tasks to decrease priority:
4363 if (!capable(CAP_SYS_NICE)) {
4364 if (rt_policy(policy)) {
4365 unsigned long rlim_rtprio;
4366 unsigned long flags;
4368 if (!lock_task_sighand(p, &flags))
4370 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4371 unlock_task_sighand(p, &flags);
4373 /* can't set/change the rt policy */
4374 if (policy != p->policy && !rlim_rtprio)
4377 /* can't increase priority */
4378 if (param->sched_priority > p->rt_priority &&
4379 param->sched_priority > rlim_rtprio)
4383 /* can't change other user's priorities */
4384 if ((current->euid != p->euid) &&
4385 (current->euid != p->uid))
4389 retval = security_task_setscheduler(p, policy, param);
4393 * make sure no PI-waiters arrive (or leave) while we are
4394 * changing the priority of the task:
4396 spin_lock_irqsave(&p->pi_lock, flags);
4398 * To be able to change p->policy safely, the apropriate
4399 * runqueue lock must be held.
4401 rq = __task_rq_lock(p);
4402 /* recheck policy now with rq lock held */
4403 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4404 policy = oldpolicy = -1;
4405 __task_rq_unlock(rq);
4406 spin_unlock_irqrestore(&p->pi_lock, flags);
4411 deactivate_task(p, rq);
4413 __setscheduler(p, policy, param->sched_priority);
4415 __activate_task(p, rq);
4417 * Reschedule if we are currently running on this runqueue and
4418 * our priority decreased, or if we are not currently running on
4419 * this runqueue and our priority is higher than the current's
4421 if (task_running(rq, p)) {
4422 if (p->prio > oldprio)
4423 resched_task(rq->curr);
4424 } else if (TASK_PREEMPTS_CURR(p, rq))
4425 resched_task(rq->curr);
4427 __task_rq_unlock(rq);
4428 spin_unlock_irqrestore(&p->pi_lock, flags);
4430 rt_mutex_adjust_pi(p);
4434 EXPORT_SYMBOL_GPL(sched_setscheduler);
4437 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4439 struct sched_param lparam;
4440 struct task_struct *p;
4443 if (!param || pid < 0)
4445 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4450 p = find_process_by_pid(pid);
4452 retval = sched_setscheduler(p, policy, &lparam);
4459 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4460 * @pid: the pid in question.
4461 * @policy: new policy.
4462 * @param: structure containing the new RT priority.
4464 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4465 struct sched_param __user *param)
4467 /* negative values for policy are not valid */
4471 return do_sched_setscheduler(pid, policy, param);
4475 * sys_sched_setparam - set/change the RT priority of a thread
4476 * @pid: the pid in question.
4477 * @param: structure containing the new RT priority.
4479 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4481 return do_sched_setscheduler(pid, -1, param);
4485 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4486 * @pid: the pid in question.
4488 asmlinkage long sys_sched_getscheduler(pid_t pid)
4490 struct task_struct *p;
4491 int retval = -EINVAL;
4497 read_lock(&tasklist_lock);
4498 p = find_process_by_pid(pid);
4500 retval = security_task_getscheduler(p);
4504 read_unlock(&tasklist_lock);
4511 * sys_sched_getscheduler - get the RT priority of a thread
4512 * @pid: the pid in question.
4513 * @param: structure containing the RT priority.
4515 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4517 struct sched_param lp;
4518 struct task_struct *p;
4519 int retval = -EINVAL;
4521 if (!param || pid < 0)
4524 read_lock(&tasklist_lock);
4525 p = find_process_by_pid(pid);
4530 retval = security_task_getscheduler(p);
4534 lp.sched_priority = p->rt_priority;
4535 read_unlock(&tasklist_lock);
4538 * This one might sleep, we cannot do it with a spinlock held ...
4540 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4546 read_unlock(&tasklist_lock);
4550 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4552 cpumask_t cpus_allowed;
4553 struct task_struct *p;
4556 mutex_lock(&sched_hotcpu_mutex);
4557 read_lock(&tasklist_lock);
4559 p = find_process_by_pid(pid);
4561 read_unlock(&tasklist_lock);
4562 mutex_unlock(&sched_hotcpu_mutex);
4567 * It is not safe to call set_cpus_allowed with the
4568 * tasklist_lock held. We will bump the task_struct's
4569 * usage count and then drop tasklist_lock.
4572 read_unlock(&tasklist_lock);
4575 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4576 !capable(CAP_SYS_NICE))
4579 retval = security_task_setscheduler(p, 0, NULL);
4583 cpus_allowed = cpuset_cpus_allowed(p);
4584 cpus_and(new_mask, new_mask, cpus_allowed);
4585 retval = set_cpus_allowed(p, new_mask);
4589 mutex_unlock(&sched_hotcpu_mutex);
4593 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4594 cpumask_t *new_mask)
4596 if (len < sizeof(cpumask_t)) {
4597 memset(new_mask, 0, sizeof(cpumask_t));
4598 } else if (len > sizeof(cpumask_t)) {
4599 len = sizeof(cpumask_t);
4601 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4605 * sys_sched_setaffinity - set the cpu affinity of a process
4606 * @pid: pid of the process
4607 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4608 * @user_mask_ptr: user-space pointer to the new cpu mask
4610 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4611 unsigned long __user *user_mask_ptr)
4616 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4620 return sched_setaffinity(pid, new_mask);
4624 * Represents all cpu's present in the system
4625 * In systems capable of hotplug, this map could dynamically grow
4626 * as new cpu's are detected in the system via any platform specific
4627 * method, such as ACPI for e.g.
4630 cpumask_t cpu_present_map __read_mostly;
4631 EXPORT_SYMBOL(cpu_present_map);
4634 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4635 EXPORT_SYMBOL(cpu_online_map);
4637 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4638 EXPORT_SYMBOL(cpu_possible_map);
4641 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4643 struct task_struct *p;
4646 mutex_lock(&sched_hotcpu_mutex);
4647 read_lock(&tasklist_lock);
4650 p = find_process_by_pid(pid);
4654 retval = security_task_getscheduler(p);
4658 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4661 read_unlock(&tasklist_lock);
4662 mutex_unlock(&sched_hotcpu_mutex);
4670 * sys_sched_getaffinity - get the cpu affinity of a process
4671 * @pid: pid of the process
4672 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4673 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4675 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4676 unsigned long __user *user_mask_ptr)
4681 if (len < sizeof(cpumask_t))
4684 ret = sched_getaffinity(pid, &mask);
4688 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4691 return sizeof(cpumask_t);
4695 * sys_sched_yield - yield the current processor to other threads.
4697 * This function yields the current CPU by moving the calling thread
4698 * to the expired array. If there are no other threads running on this
4699 * CPU then this function will return.
4701 asmlinkage long sys_sched_yield(void)
4703 struct rq *rq = this_rq_lock();
4704 struct prio_array *array = current->array, *target = rq->expired;
4706 schedstat_inc(rq, yld_cnt);
4708 * We implement yielding by moving the task into the expired
4711 * (special rule: RT tasks will just roundrobin in the active
4714 if (rt_task(current))
4715 target = rq->active;
4717 if (array->nr_active == 1) {
4718 schedstat_inc(rq, yld_act_empty);
4719 if (!rq->expired->nr_active)
4720 schedstat_inc(rq, yld_both_empty);
4721 } else if (!rq->expired->nr_active)
4722 schedstat_inc(rq, yld_exp_empty);
4724 if (array != target) {
4725 dequeue_task(current, array);
4726 enqueue_task(current, target);
4729 * requeue_task is cheaper so perform that if possible.
4731 requeue_task(current, array);
4734 * Since we are going to call schedule() anyway, there's
4735 * no need to preempt or enable interrupts:
4737 __release(rq->lock);
4738 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4739 _raw_spin_unlock(&rq->lock);
4740 preempt_enable_no_resched();
4747 static void __cond_resched(void)
4749 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4750 __might_sleep(__FILE__, __LINE__);
4753 * The BKS might be reacquired before we have dropped
4754 * PREEMPT_ACTIVE, which could trigger a second
4755 * cond_resched() call.
4758 add_preempt_count(PREEMPT_ACTIVE);
4760 sub_preempt_count(PREEMPT_ACTIVE);
4761 } while (need_resched());
4764 int __sched cond_resched(void)
4766 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4767 system_state == SYSTEM_RUNNING) {
4773 EXPORT_SYMBOL(cond_resched);
4776 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4777 * call schedule, and on return reacquire the lock.
4779 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4780 * operations here to prevent schedule() from being called twice (once via
4781 * spin_unlock(), once by hand).
4783 int cond_resched_lock(spinlock_t *lock)
4787 if (need_lockbreak(lock)) {
4793 if (need_resched() && system_state == SYSTEM_RUNNING) {
4794 spin_release(&lock->dep_map, 1, _THIS_IP_);
4795 _raw_spin_unlock(lock);
4796 preempt_enable_no_resched();
4803 EXPORT_SYMBOL(cond_resched_lock);
4805 int __sched cond_resched_softirq(void)
4807 BUG_ON(!in_softirq());
4809 if (need_resched() && system_state == SYSTEM_RUNNING) {
4817 EXPORT_SYMBOL(cond_resched_softirq);
4820 * yield - yield the current processor to other threads.
4822 * This is a shortcut for kernel-space yielding - it marks the
4823 * thread runnable and calls sys_sched_yield().
4825 void __sched yield(void)
4827 set_current_state(TASK_RUNNING);
4830 EXPORT_SYMBOL(yield);
4833 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4834 * that process accounting knows that this is a task in IO wait state.
4836 * But don't do that if it is a deliberate, throttling IO wait (this task
4837 * has set its backing_dev_info: the queue against which it should throttle)
4839 void __sched io_schedule(void)
4841 struct rq *rq = &__raw_get_cpu_var(runqueues);
4843 delayacct_blkio_start();
4844 atomic_inc(&rq->nr_iowait);
4846 atomic_dec(&rq->nr_iowait);
4847 delayacct_blkio_end();
4849 EXPORT_SYMBOL(io_schedule);
4851 long __sched io_schedule_timeout(long timeout)
4853 struct rq *rq = &__raw_get_cpu_var(runqueues);
4856 delayacct_blkio_start();
4857 atomic_inc(&rq->nr_iowait);
4858 ret = schedule_timeout(timeout);
4859 atomic_dec(&rq->nr_iowait);
4860 delayacct_blkio_end();
4865 * sys_sched_get_priority_max - return maximum RT priority.
4866 * @policy: scheduling class.
4868 * this syscall returns the maximum rt_priority that can be used
4869 * by a given scheduling class.
4871 asmlinkage long sys_sched_get_priority_max(int policy)
4878 ret = MAX_USER_RT_PRIO-1;
4889 * sys_sched_get_priority_min - return minimum RT priority.
4890 * @policy: scheduling class.
4892 * this syscall returns the minimum rt_priority that can be used
4893 * by a given scheduling class.
4895 asmlinkage long sys_sched_get_priority_min(int policy)
4912 * sys_sched_rr_get_interval - return the default timeslice of a process.
4913 * @pid: pid of the process.
4914 * @interval: userspace pointer to the timeslice value.
4916 * this syscall writes the default timeslice value of a given process
4917 * into the user-space timespec buffer. A value of '0' means infinity.
4920 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4922 struct task_struct *p;
4923 int retval = -EINVAL;
4930 read_lock(&tasklist_lock);
4931 p = find_process_by_pid(pid);
4935 retval = security_task_getscheduler(p);
4939 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4940 0 : task_timeslice(p), &t);
4941 read_unlock(&tasklist_lock);
4942 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4946 read_unlock(&tasklist_lock);
4950 static const char stat_nam[] = "RSDTtZX";
4952 static void show_task(struct task_struct *p)
4954 unsigned long free = 0;
4957 state = p->state ? __ffs(p->state) + 1 : 0;
4958 printk("%-13.13s %c", p->comm,
4959 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4960 #if (BITS_PER_LONG == 32)
4961 if (state == TASK_RUNNING)
4962 printk(" running ");
4964 printk(" %08lX ", thread_saved_pc(p));
4966 if (state == TASK_RUNNING)
4967 printk(" running task ");
4969 printk(" %016lx ", thread_saved_pc(p));
4971 #ifdef CONFIG_DEBUG_STACK_USAGE
4973 unsigned long *n = end_of_stack(p);
4976 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4979 printk("%5lu %5d %6d", free, p->pid, p->parent->pid);
4981 printk(" (L-TLB)\n");
4983 printk(" (NOTLB)\n");
4985 if (state != TASK_RUNNING)
4986 show_stack(p, NULL);
4989 void show_state_filter(unsigned long state_filter)
4991 struct task_struct *g, *p;
4993 #if (BITS_PER_LONG == 32)
4996 printk(" task PC stack pid father child younger older\n");
5000 printk(" task PC stack pid father child younger older\n");
5002 read_lock(&tasklist_lock);
5003 do_each_thread(g, p) {
5005 * reset the NMI-timeout, listing all files on a slow
5006 * console might take alot of time:
5008 touch_nmi_watchdog();
5009 if (!state_filter || (p->state & state_filter))
5011 } while_each_thread(g, p);
5013 touch_all_softlockup_watchdogs();
5015 read_unlock(&tasklist_lock);
5017 * Only show locks if all tasks are dumped:
5019 if (state_filter == -1)
5020 debug_show_all_locks();
5023 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5029 * init_idle - set up an idle thread for a given CPU
5030 * @idle: task in question
5031 * @cpu: cpu the idle task belongs to
5033 * NOTE: this function does not set the idle thread's NEED_RESCHED
5034 * flag, to make booting more robust.
5036 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5038 struct rq *rq = cpu_rq(cpu);
5039 unsigned long flags;
5041 idle->timestamp = sched_clock();
5042 idle->sleep_avg = 0;
5044 idle->prio = idle->normal_prio = MAX_PRIO;
5045 idle->state = TASK_RUNNING;
5046 idle->cpus_allowed = cpumask_of_cpu(cpu);
5047 set_task_cpu(idle, cpu);
5049 spin_lock_irqsave(&rq->lock, flags);
5050 rq->curr = rq->idle = idle;
5051 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5054 spin_unlock_irqrestore(&rq->lock, flags);
5056 /* Set the preempt count _outside_ the spinlocks! */
5057 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5058 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5060 task_thread_info(idle)->preempt_count = 0;
5065 * In a system that switches off the HZ timer nohz_cpu_mask
5066 * indicates which cpus entered this state. This is used
5067 * in the rcu update to wait only for active cpus. For system
5068 * which do not switch off the HZ timer nohz_cpu_mask should
5069 * always be CPU_MASK_NONE.
5071 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5075 * This is how migration works:
5077 * 1) we queue a struct migration_req structure in the source CPU's
5078 * runqueue and wake up that CPU's migration thread.
5079 * 2) we down() the locked semaphore => thread blocks.
5080 * 3) migration thread wakes up (implicitly it forces the migrated
5081 * thread off the CPU)
5082 * 4) it gets the migration request and checks whether the migrated
5083 * task is still in the wrong runqueue.
5084 * 5) if it's in the wrong runqueue then the migration thread removes
5085 * it and puts it into the right queue.
5086 * 6) migration thread up()s the semaphore.
5087 * 7) we wake up and the migration is done.
5091 * Change a given task's CPU affinity. Migrate the thread to a
5092 * proper CPU and schedule it away if the CPU it's executing on
5093 * is removed from the allowed bitmask.
5095 * NOTE: the caller must have a valid reference to the task, the
5096 * task must not exit() & deallocate itself prematurely. The
5097 * call is not atomic; no spinlocks may be held.
5099 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5101 struct migration_req req;
5102 unsigned long flags;
5106 rq = task_rq_lock(p, &flags);
5107 if (!cpus_intersects(new_mask, cpu_online_map)) {
5112 p->cpus_allowed = new_mask;
5113 /* Can the task run on the task's current CPU? If so, we're done */
5114 if (cpu_isset(task_cpu(p), new_mask))
5117 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5118 /* Need help from migration thread: drop lock and wait. */
5119 task_rq_unlock(rq, &flags);
5120 wake_up_process(rq->migration_thread);
5121 wait_for_completion(&req.done);
5122 tlb_migrate_finish(p->mm);
5126 task_rq_unlock(rq, &flags);
5130 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5133 * Move (not current) task off this cpu, onto dest cpu. We're doing
5134 * this because either it can't run here any more (set_cpus_allowed()
5135 * away from this CPU, or CPU going down), or because we're
5136 * attempting to rebalance this task on exec (sched_exec).
5138 * So we race with normal scheduler movements, but that's OK, as long
5139 * as the task is no longer on this CPU.
5141 * Returns non-zero if task was successfully migrated.
5143 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5145 struct rq *rq_dest, *rq_src;
5148 if (unlikely(cpu_is_offline(dest_cpu)))
5151 rq_src = cpu_rq(src_cpu);
5152 rq_dest = cpu_rq(dest_cpu);
5154 double_rq_lock(rq_src, rq_dest);
5155 /* Already moved. */
5156 if (task_cpu(p) != src_cpu)
5158 /* Affinity changed (again). */
5159 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5162 set_task_cpu(p, dest_cpu);
5165 * Sync timestamp with rq_dest's before activating.
5166 * The same thing could be achieved by doing this step
5167 * afterwards, and pretending it was a local activate.
5168 * This way is cleaner and logically correct.
5170 p->timestamp = p->timestamp - rq_src->most_recent_timestamp
5171 + rq_dest->most_recent_timestamp;
5172 deactivate_task(p, rq_src);
5173 __activate_task(p, rq_dest);
5174 if (TASK_PREEMPTS_CURR(p, rq_dest))
5175 resched_task(rq_dest->curr);
5179 double_rq_unlock(rq_src, rq_dest);
5184 * migration_thread - this is a highprio system thread that performs
5185 * thread migration by bumping thread off CPU then 'pushing' onto
5188 static int migration_thread(void *data)
5190 int cpu = (long)data;
5194 BUG_ON(rq->migration_thread != current);
5196 set_current_state(TASK_INTERRUPTIBLE);
5197 while (!kthread_should_stop()) {
5198 struct migration_req *req;
5199 struct list_head *head;
5203 spin_lock_irq(&rq->lock);
5205 if (cpu_is_offline(cpu)) {
5206 spin_unlock_irq(&rq->lock);
5210 if (rq->active_balance) {
5211 active_load_balance(rq, cpu);
5212 rq->active_balance = 0;
5215 head = &rq->migration_queue;
5217 if (list_empty(head)) {
5218 spin_unlock_irq(&rq->lock);
5220 set_current_state(TASK_INTERRUPTIBLE);
5223 req = list_entry(head->next, struct migration_req, list);
5224 list_del_init(head->next);
5226 spin_unlock(&rq->lock);
5227 __migrate_task(req->task, cpu, req->dest_cpu);
5230 complete(&req->done);
5232 __set_current_state(TASK_RUNNING);
5236 /* Wait for kthread_stop */
5237 set_current_state(TASK_INTERRUPTIBLE);
5238 while (!kthread_should_stop()) {
5240 set_current_state(TASK_INTERRUPTIBLE);
5242 __set_current_state(TASK_RUNNING);
5246 #ifdef CONFIG_HOTPLUG_CPU
5248 * Figure out where task on dead CPU should go, use force if neccessary.
5249 * NOTE: interrupts should be disabled by the caller
5251 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5253 unsigned long flags;
5260 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5261 cpus_and(mask, mask, p->cpus_allowed);
5262 dest_cpu = any_online_cpu(mask);
5264 /* On any allowed CPU? */
5265 if (dest_cpu == NR_CPUS)
5266 dest_cpu = any_online_cpu(p->cpus_allowed);
5268 /* No more Mr. Nice Guy. */
5269 if (dest_cpu == NR_CPUS) {
5270 rq = task_rq_lock(p, &flags);
5271 cpus_setall(p->cpus_allowed);
5272 dest_cpu = any_online_cpu(p->cpus_allowed);
5273 task_rq_unlock(rq, &flags);
5276 * Don't tell them about moving exiting tasks or
5277 * kernel threads (both mm NULL), since they never
5280 if (p->mm && printk_ratelimit())
5281 printk(KERN_INFO "process %d (%s) no "
5282 "longer affine to cpu%d\n",
5283 p->pid, p->comm, dead_cpu);
5285 if (!__migrate_task(p, dead_cpu, dest_cpu))
5290 * While a dead CPU has no uninterruptible tasks queued at this point,
5291 * it might still have a nonzero ->nr_uninterruptible counter, because
5292 * for performance reasons the counter is not stricly tracking tasks to
5293 * their home CPUs. So we just add the counter to another CPU's counter,
5294 * to keep the global sum constant after CPU-down:
5296 static void migrate_nr_uninterruptible(struct rq *rq_src)
5298 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5299 unsigned long flags;
5301 local_irq_save(flags);
5302 double_rq_lock(rq_src, rq_dest);
5303 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5304 rq_src->nr_uninterruptible = 0;
5305 double_rq_unlock(rq_src, rq_dest);
5306 local_irq_restore(flags);
5309 /* Run through task list and migrate tasks from the dead cpu. */
5310 static void migrate_live_tasks(int src_cpu)
5312 struct task_struct *p, *t;
5314 write_lock_irq(&tasklist_lock);
5316 do_each_thread(t, p) {
5320 if (task_cpu(p) == src_cpu)
5321 move_task_off_dead_cpu(src_cpu, p);
5322 } while_each_thread(t, p);
5324 write_unlock_irq(&tasklist_lock);
5327 /* Schedules idle task to be the next runnable task on current CPU.
5328 * It does so by boosting its priority to highest possible and adding it to
5329 * the _front_ of the runqueue. Used by CPU offline code.
5331 void sched_idle_next(void)
5333 int this_cpu = smp_processor_id();
5334 struct rq *rq = cpu_rq(this_cpu);
5335 struct task_struct *p = rq->idle;
5336 unsigned long flags;
5338 /* cpu has to be offline */
5339 BUG_ON(cpu_online(this_cpu));
5342 * Strictly not necessary since rest of the CPUs are stopped by now
5343 * and interrupts disabled on the current cpu.
5345 spin_lock_irqsave(&rq->lock, flags);
5347 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5349 /* Add idle task to the _front_ of its priority queue: */
5350 __activate_idle_task(p, rq);
5352 spin_unlock_irqrestore(&rq->lock, flags);
5356 * Ensures that the idle task is using init_mm right before its cpu goes
5359 void idle_task_exit(void)
5361 struct mm_struct *mm = current->active_mm;
5363 BUG_ON(cpu_online(smp_processor_id()));
5366 switch_mm(mm, &init_mm, current);
5370 /* called under rq->lock with disabled interrupts */
5371 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5373 struct rq *rq = cpu_rq(dead_cpu);
5375 /* Must be exiting, otherwise would be on tasklist. */
5376 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5378 /* Cannot have done final schedule yet: would have vanished. */
5379 BUG_ON(p->state == TASK_DEAD);
5384 * Drop lock around migration; if someone else moves it,
5385 * that's OK. No task can be added to this CPU, so iteration is
5387 * NOTE: interrupts should be left disabled --dev@
5389 spin_unlock(&rq->lock);
5390 move_task_off_dead_cpu(dead_cpu, p);
5391 spin_lock(&rq->lock);
5396 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5397 static void migrate_dead_tasks(unsigned int dead_cpu)
5399 struct rq *rq = cpu_rq(dead_cpu);
5400 unsigned int arr, i;
5402 for (arr = 0; arr < 2; arr++) {
5403 for (i = 0; i < MAX_PRIO; i++) {
5404 struct list_head *list = &rq->arrays[arr].queue[i];
5406 while (!list_empty(list))
5407 migrate_dead(dead_cpu, list_entry(list->next,
5408 struct task_struct, run_list));
5412 #endif /* CONFIG_HOTPLUG_CPU */
5415 * migration_call - callback that gets triggered when a CPU is added.
5416 * Here we can start up the necessary migration thread for the new CPU.
5418 static int __cpuinit
5419 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5421 struct task_struct *p;
5422 int cpu = (long)hcpu;
5423 unsigned long flags;
5427 case CPU_LOCK_ACQUIRE:
5428 mutex_lock(&sched_hotcpu_mutex);
5431 case CPU_UP_PREPARE:
5432 case CPU_UP_PREPARE_FROZEN:
5433 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
5436 p->flags |= PF_NOFREEZE;
5437 kthread_bind(p, cpu);
5438 /* Must be high prio: stop_machine expects to yield to it. */
5439 rq = task_rq_lock(p, &flags);
5440 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
5441 task_rq_unlock(rq, &flags);
5442 cpu_rq(cpu)->migration_thread = p;
5446 case CPU_ONLINE_FROZEN:
5447 /* Strictly unneccessary, as first user will wake it. */
5448 wake_up_process(cpu_rq(cpu)->migration_thread);
5451 #ifdef CONFIG_HOTPLUG_CPU
5452 case CPU_UP_CANCELED:
5453 case CPU_UP_CANCELED_FROZEN:
5454 if (!cpu_rq(cpu)->migration_thread)
5456 /* Unbind it from offline cpu so it can run. Fall thru. */
5457 kthread_bind(cpu_rq(cpu)->migration_thread,
5458 any_online_cpu(cpu_online_map));
5459 kthread_stop(cpu_rq(cpu)->migration_thread);
5460 cpu_rq(cpu)->migration_thread = NULL;
5464 case CPU_DEAD_FROZEN:
5465 migrate_live_tasks(cpu);
5467 kthread_stop(rq->migration_thread);
5468 rq->migration_thread = NULL;
5469 /* Idle task back to normal (off runqueue, low prio) */
5470 rq = task_rq_lock(rq->idle, &flags);
5471 deactivate_task(rq->idle, rq);
5472 rq->idle->static_prio = MAX_PRIO;
5473 __setscheduler(rq->idle, SCHED_NORMAL, 0);
5474 migrate_dead_tasks(cpu);
5475 task_rq_unlock(rq, &flags);
5476 migrate_nr_uninterruptible(rq);
5477 BUG_ON(rq->nr_running != 0);
5479 /* No need to migrate the tasks: it was best-effort if
5480 * they didn't take sched_hotcpu_mutex. Just wake up
5481 * the requestors. */
5482 spin_lock_irq(&rq->lock);
5483 while (!list_empty(&rq->migration_queue)) {
5484 struct migration_req *req;
5486 req = list_entry(rq->migration_queue.next,
5487 struct migration_req, list);
5488 list_del_init(&req->list);
5489 complete(&req->done);
5491 spin_unlock_irq(&rq->lock);
5494 case CPU_LOCK_RELEASE:
5495 mutex_unlock(&sched_hotcpu_mutex);
5501 /* Register at highest priority so that task migration (migrate_all_tasks)
5502 * happens before everything else.
5504 static struct notifier_block __cpuinitdata migration_notifier = {
5505 .notifier_call = migration_call,
5509 int __init migration_init(void)
5511 void *cpu = (void *)(long)smp_processor_id();
5514 /* Start one for the boot CPU: */
5515 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5516 BUG_ON(err == NOTIFY_BAD);
5517 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5518 register_cpu_notifier(&migration_notifier);
5526 /* Number of possible processor ids */
5527 int nr_cpu_ids __read_mostly = NR_CPUS;
5528 EXPORT_SYMBOL(nr_cpu_ids);
5530 #undef SCHED_DOMAIN_DEBUG
5531 #ifdef SCHED_DOMAIN_DEBUG
5532 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5537 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5541 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5546 struct sched_group *group = sd->groups;
5547 cpumask_t groupmask;
5549 cpumask_scnprintf(str, NR_CPUS, sd->span);
5550 cpus_clear(groupmask);
5553 for (i = 0; i < level + 1; i++)
5555 printk("domain %d: ", level);
5557 if (!(sd->flags & SD_LOAD_BALANCE)) {
5558 printk("does not load-balance\n");
5560 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5565 printk("span %s\n", str);
5567 if (!cpu_isset(cpu, sd->span))
5568 printk(KERN_ERR "ERROR: domain->span does not contain "
5570 if (!cpu_isset(cpu, group->cpumask))
5571 printk(KERN_ERR "ERROR: domain->groups does not contain"
5575 for (i = 0; i < level + 2; i++)
5581 printk(KERN_ERR "ERROR: group is NULL\n");
5585 if (!group->__cpu_power) {
5587 printk(KERN_ERR "ERROR: domain->cpu_power not "
5591 if (!cpus_weight(group->cpumask)) {
5593 printk(KERN_ERR "ERROR: empty group\n");
5596 if (cpus_intersects(groupmask, group->cpumask)) {
5598 printk(KERN_ERR "ERROR: repeated CPUs\n");
5601 cpus_or(groupmask, groupmask, group->cpumask);
5603 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5606 group = group->next;
5607 } while (group != sd->groups);
5610 if (!cpus_equal(sd->span, groupmask))
5611 printk(KERN_ERR "ERROR: groups don't span "
5619 if (!cpus_subset(groupmask, sd->span))
5620 printk(KERN_ERR "ERROR: parent span is not a superset "
5621 "of domain->span\n");
5626 # define sched_domain_debug(sd, cpu) do { } while (0)
5629 static int sd_degenerate(struct sched_domain *sd)
5631 if (cpus_weight(sd->span) == 1)
5634 /* Following flags need at least 2 groups */
5635 if (sd->flags & (SD_LOAD_BALANCE |
5636 SD_BALANCE_NEWIDLE |
5640 SD_SHARE_PKG_RESOURCES)) {
5641 if (sd->groups != sd->groups->next)
5645 /* Following flags don't use groups */
5646 if (sd->flags & (SD_WAKE_IDLE |
5655 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5657 unsigned long cflags = sd->flags, pflags = parent->flags;
5659 if (sd_degenerate(parent))
5662 if (!cpus_equal(sd->span, parent->span))
5665 /* Does parent contain flags not in child? */
5666 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5667 if (cflags & SD_WAKE_AFFINE)
5668 pflags &= ~SD_WAKE_BALANCE;
5669 /* Flags needing groups don't count if only 1 group in parent */
5670 if (parent->groups == parent->groups->next) {
5671 pflags &= ~(SD_LOAD_BALANCE |
5672 SD_BALANCE_NEWIDLE |
5676 SD_SHARE_PKG_RESOURCES);
5678 if (~cflags & pflags)
5685 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5686 * hold the hotplug lock.
5688 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5690 struct rq *rq = cpu_rq(cpu);
5691 struct sched_domain *tmp;
5693 /* Remove the sched domains which do not contribute to scheduling. */
5694 for (tmp = sd; tmp; tmp = tmp->parent) {
5695 struct sched_domain *parent = tmp->parent;
5698 if (sd_parent_degenerate(tmp, parent)) {
5699 tmp->parent = parent->parent;
5701 parent->parent->child = tmp;
5705 if (sd && sd_degenerate(sd)) {
5711 sched_domain_debug(sd, cpu);
5713 rcu_assign_pointer(rq->sd, sd);
5716 /* cpus with isolated domains */
5717 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5719 /* Setup the mask of cpus configured for isolated domains */
5720 static int __init isolated_cpu_setup(char *str)
5722 int ints[NR_CPUS], i;
5724 str = get_options(str, ARRAY_SIZE(ints), ints);
5725 cpus_clear(cpu_isolated_map);
5726 for (i = 1; i <= ints[0]; i++)
5727 if (ints[i] < NR_CPUS)
5728 cpu_set(ints[i], cpu_isolated_map);
5732 __setup ("isolcpus=", isolated_cpu_setup);
5735 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5736 * to a function which identifies what group(along with sched group) a CPU
5737 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5738 * (due to the fact that we keep track of groups covered with a cpumask_t).
5740 * init_sched_build_groups will build a circular linked list of the groups
5741 * covered by the given span, and will set each group's ->cpumask correctly,
5742 * and ->cpu_power to 0.
5745 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5746 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5747 struct sched_group **sg))
5749 struct sched_group *first = NULL, *last = NULL;
5750 cpumask_t covered = CPU_MASK_NONE;
5753 for_each_cpu_mask(i, span) {
5754 struct sched_group *sg;
5755 int group = group_fn(i, cpu_map, &sg);
5758 if (cpu_isset(i, covered))
5761 sg->cpumask = CPU_MASK_NONE;
5762 sg->__cpu_power = 0;
5764 for_each_cpu_mask(j, span) {
5765 if (group_fn(j, cpu_map, NULL) != group)
5768 cpu_set(j, covered);
5769 cpu_set(j, sg->cpumask);
5780 #define SD_NODES_PER_DOMAIN 16
5785 * find_next_best_node - find the next node to include in a sched_domain
5786 * @node: node whose sched_domain we're building
5787 * @used_nodes: nodes already in the sched_domain
5789 * Find the next node to include in a given scheduling domain. Simply
5790 * finds the closest node not already in the @used_nodes map.
5792 * Should use nodemask_t.
5794 static int find_next_best_node(int node, unsigned long *used_nodes)
5796 int i, n, val, min_val, best_node = 0;
5800 for (i = 0; i < MAX_NUMNODES; i++) {
5801 /* Start at @node */
5802 n = (node + i) % MAX_NUMNODES;
5804 if (!nr_cpus_node(n))
5807 /* Skip already used nodes */
5808 if (test_bit(n, used_nodes))
5811 /* Simple min distance search */
5812 val = node_distance(node, n);
5814 if (val < min_val) {
5820 set_bit(best_node, used_nodes);
5825 * sched_domain_node_span - get a cpumask for a node's sched_domain
5826 * @node: node whose cpumask we're constructing
5827 * @size: number of nodes to include in this span
5829 * Given a node, construct a good cpumask for its sched_domain to span. It
5830 * should be one that prevents unnecessary balancing, but also spreads tasks
5833 static cpumask_t sched_domain_node_span(int node)
5835 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5836 cpumask_t span, nodemask;
5840 bitmap_zero(used_nodes, MAX_NUMNODES);
5842 nodemask = node_to_cpumask(node);
5843 cpus_or(span, span, nodemask);
5844 set_bit(node, used_nodes);
5846 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5847 int next_node = find_next_best_node(node, used_nodes);
5849 nodemask = node_to_cpumask(next_node);
5850 cpus_or(span, span, nodemask);
5857 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5860 * SMT sched-domains:
5862 #ifdef CONFIG_SCHED_SMT
5863 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5864 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5866 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5867 struct sched_group **sg)
5870 *sg = &per_cpu(sched_group_cpus, cpu);
5876 * multi-core sched-domains:
5878 #ifdef CONFIG_SCHED_MC
5879 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5880 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5883 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5884 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5885 struct sched_group **sg)
5888 cpumask_t mask = cpu_sibling_map[cpu];
5889 cpus_and(mask, mask, *cpu_map);
5890 group = first_cpu(mask);
5892 *sg = &per_cpu(sched_group_core, group);
5895 #elif defined(CONFIG_SCHED_MC)
5896 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5897 struct sched_group **sg)
5900 *sg = &per_cpu(sched_group_core, cpu);
5905 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5906 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5908 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5909 struct sched_group **sg)
5912 #ifdef CONFIG_SCHED_MC
5913 cpumask_t mask = cpu_coregroup_map(cpu);
5914 cpus_and(mask, mask, *cpu_map);
5915 group = first_cpu(mask);
5916 #elif defined(CONFIG_SCHED_SMT)
5917 cpumask_t mask = cpu_sibling_map[cpu];
5918 cpus_and(mask, mask, *cpu_map);
5919 group = first_cpu(mask);
5924 *sg = &per_cpu(sched_group_phys, group);
5930 * The init_sched_build_groups can't handle what we want to do with node
5931 * groups, so roll our own. Now each node has its own list of groups which
5932 * gets dynamically allocated.
5934 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5935 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5937 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5938 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5940 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5941 struct sched_group **sg)
5943 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5946 cpus_and(nodemask, nodemask, *cpu_map);
5947 group = first_cpu(nodemask);
5950 *sg = &per_cpu(sched_group_allnodes, group);
5954 static void init_numa_sched_groups_power(struct sched_group *group_head)
5956 struct sched_group *sg = group_head;
5962 for_each_cpu_mask(j, sg->cpumask) {
5963 struct sched_domain *sd;
5965 sd = &per_cpu(phys_domains, j);
5966 if (j != first_cpu(sd->groups->cpumask)) {
5968 * Only add "power" once for each
5974 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5977 if (sg != group_head)
5983 /* Free memory allocated for various sched_group structures */
5984 static void free_sched_groups(const cpumask_t *cpu_map)
5988 for_each_cpu_mask(cpu, *cpu_map) {
5989 struct sched_group **sched_group_nodes
5990 = sched_group_nodes_bycpu[cpu];
5992 if (!sched_group_nodes)
5995 for (i = 0; i < MAX_NUMNODES; i++) {
5996 cpumask_t nodemask = node_to_cpumask(i);
5997 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5999 cpus_and(nodemask, nodemask, *cpu_map);
6000 if (cpus_empty(nodemask))
6010 if (oldsg != sched_group_nodes[i])
6013 kfree(sched_group_nodes);
6014 sched_group_nodes_bycpu[cpu] = NULL;
6018 static void free_sched_groups(const cpumask_t *cpu_map)
6024 * Initialize sched groups cpu_power.
6026 * cpu_power indicates the capacity of sched group, which is used while
6027 * distributing the load between different sched groups in a sched domain.
6028 * Typically cpu_power for all the groups in a sched domain will be same unless
6029 * there are asymmetries in the topology. If there are asymmetries, group
6030 * having more cpu_power will pickup more load compared to the group having
6033 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6034 * the maximum number of tasks a group can handle in the presence of other idle
6035 * or lightly loaded groups in the same sched domain.
6037 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6039 struct sched_domain *child;
6040 struct sched_group *group;
6042 WARN_ON(!sd || !sd->groups);
6044 if (cpu != first_cpu(sd->groups->cpumask))
6049 sd->groups->__cpu_power = 0;
6052 * For perf policy, if the groups in child domain share resources
6053 * (for example cores sharing some portions of the cache hierarchy
6054 * or SMT), then set this domain groups cpu_power such that each group
6055 * can handle only one task, when there are other idle groups in the
6056 * same sched domain.
6058 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6060 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6061 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6066 * add cpu_power of each child group to this groups cpu_power
6068 group = child->groups;
6070 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6071 group = group->next;
6072 } while (group != child->groups);
6076 * Build sched domains for a given set of cpus and attach the sched domains
6077 * to the individual cpus
6079 static int build_sched_domains(const cpumask_t *cpu_map)
6082 struct sched_domain *sd;
6084 struct sched_group **sched_group_nodes = NULL;
6085 int sd_allnodes = 0;
6088 * Allocate the per-node list of sched groups
6090 sched_group_nodes = kzalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
6092 if (!sched_group_nodes) {
6093 printk(KERN_WARNING "Can not alloc sched group node list\n");
6096 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6100 * Set up domains for cpus specified by the cpu_map.
6102 for_each_cpu_mask(i, *cpu_map) {
6103 struct sched_domain *sd = NULL, *p;
6104 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6106 cpus_and(nodemask, nodemask, *cpu_map);
6109 if (cpus_weight(*cpu_map)
6110 > SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6111 sd = &per_cpu(allnodes_domains, i);
6112 *sd = SD_ALLNODES_INIT;
6113 sd->span = *cpu_map;
6114 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6120 sd = &per_cpu(node_domains, i);
6122 sd->span = sched_domain_node_span(cpu_to_node(i));
6126 cpus_and(sd->span, sd->span, *cpu_map);
6130 sd = &per_cpu(phys_domains, i);
6132 sd->span = nodemask;
6136 cpu_to_phys_group(i, cpu_map, &sd->groups);
6138 #ifdef CONFIG_SCHED_MC
6140 sd = &per_cpu(core_domains, i);
6142 sd->span = cpu_coregroup_map(i);
6143 cpus_and(sd->span, sd->span, *cpu_map);
6146 cpu_to_core_group(i, cpu_map, &sd->groups);
6149 #ifdef CONFIG_SCHED_SMT
6151 sd = &per_cpu(cpu_domains, i);
6152 *sd = SD_SIBLING_INIT;
6153 sd->span = cpu_sibling_map[i];
6154 cpus_and(sd->span, sd->span, *cpu_map);
6157 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6161 #ifdef CONFIG_SCHED_SMT
6162 /* Set up CPU (sibling) groups */
6163 for_each_cpu_mask(i, *cpu_map) {
6164 cpumask_t this_sibling_map = cpu_sibling_map[i];
6165 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6166 if (i != first_cpu(this_sibling_map))
6169 init_sched_build_groups(this_sibling_map, cpu_map, &cpu_to_cpu_group);
6173 #ifdef CONFIG_SCHED_MC
6174 /* Set up multi-core groups */
6175 for_each_cpu_mask(i, *cpu_map) {
6176 cpumask_t this_core_map = cpu_coregroup_map(i);
6177 cpus_and(this_core_map, this_core_map, *cpu_map);
6178 if (i != first_cpu(this_core_map))
6180 init_sched_build_groups(this_core_map, cpu_map, &cpu_to_core_group);
6185 /* Set up physical groups */
6186 for (i = 0; i < MAX_NUMNODES; i++) {
6187 cpumask_t nodemask = node_to_cpumask(i);
6189 cpus_and(nodemask, nodemask, *cpu_map);
6190 if (cpus_empty(nodemask))
6193 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6197 /* Set up node groups */
6199 init_sched_build_groups(*cpu_map, cpu_map, &cpu_to_allnodes_group);
6201 for (i = 0; i < MAX_NUMNODES; i++) {
6202 /* Set up node groups */
6203 struct sched_group *sg, *prev;
6204 cpumask_t nodemask = node_to_cpumask(i);
6205 cpumask_t domainspan;
6206 cpumask_t covered = CPU_MASK_NONE;
6209 cpus_and(nodemask, nodemask, *cpu_map);
6210 if (cpus_empty(nodemask)) {
6211 sched_group_nodes[i] = NULL;
6215 domainspan = sched_domain_node_span(i);
6216 cpus_and(domainspan, domainspan, *cpu_map);
6218 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6220 printk(KERN_WARNING "Can not alloc domain group for "
6224 sched_group_nodes[i] = sg;
6225 for_each_cpu_mask(j, nodemask) {
6226 struct sched_domain *sd;
6227 sd = &per_cpu(node_domains, j);
6230 sg->__cpu_power = 0;
6231 sg->cpumask = nodemask;
6233 cpus_or(covered, covered, nodemask);
6236 for (j = 0; j < MAX_NUMNODES; j++) {
6237 cpumask_t tmp, notcovered;
6238 int n = (i + j) % MAX_NUMNODES;
6240 cpus_complement(notcovered, covered);
6241 cpus_and(tmp, notcovered, *cpu_map);
6242 cpus_and(tmp, tmp, domainspan);
6243 if (cpus_empty(tmp))
6246 nodemask = node_to_cpumask(n);
6247 cpus_and(tmp, tmp, nodemask);
6248 if (cpus_empty(tmp))
6251 sg = kmalloc_node(sizeof(struct sched_group),
6255 "Can not alloc domain group for node %d\n", j);
6258 sg->__cpu_power = 0;
6260 sg->next = prev->next;
6261 cpus_or(covered, covered, tmp);
6268 /* Calculate CPU power for physical packages and nodes */
6269 #ifdef CONFIG_SCHED_SMT
6270 for_each_cpu_mask(i, *cpu_map) {
6271 sd = &per_cpu(cpu_domains, i);
6272 init_sched_groups_power(i, sd);
6275 #ifdef CONFIG_SCHED_MC
6276 for_each_cpu_mask(i, *cpu_map) {
6277 sd = &per_cpu(core_domains, i);
6278 init_sched_groups_power(i, sd);
6282 for_each_cpu_mask(i, *cpu_map) {
6283 sd = &per_cpu(phys_domains, i);
6284 init_sched_groups_power(i, sd);
6288 for (i = 0; i < MAX_NUMNODES; i++)
6289 init_numa_sched_groups_power(sched_group_nodes[i]);
6292 struct sched_group *sg;
6294 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6295 init_numa_sched_groups_power(sg);
6299 /* Attach the domains */
6300 for_each_cpu_mask(i, *cpu_map) {
6301 struct sched_domain *sd;
6302 #ifdef CONFIG_SCHED_SMT
6303 sd = &per_cpu(cpu_domains, i);
6304 #elif defined(CONFIG_SCHED_MC)
6305 sd = &per_cpu(core_domains, i);
6307 sd = &per_cpu(phys_domains, i);
6309 cpu_attach_domain(sd, i);
6316 free_sched_groups(cpu_map);
6321 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6323 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6325 cpumask_t cpu_default_map;
6329 * Setup mask for cpus without special case scheduling requirements.
6330 * For now this just excludes isolated cpus, but could be used to
6331 * exclude other special cases in the future.
6333 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6335 err = build_sched_domains(&cpu_default_map);
6340 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6342 free_sched_groups(cpu_map);
6346 * Detach sched domains from a group of cpus specified in cpu_map
6347 * These cpus will now be attached to the NULL domain
6349 static void detach_destroy_domains(const cpumask_t *cpu_map)
6353 for_each_cpu_mask(i, *cpu_map)
6354 cpu_attach_domain(NULL, i);
6355 synchronize_sched();
6356 arch_destroy_sched_domains(cpu_map);
6360 * Partition sched domains as specified by the cpumasks below.
6361 * This attaches all cpus from the cpumasks to the NULL domain,
6362 * waits for a RCU quiescent period, recalculates sched
6363 * domain information and then attaches them back to the
6364 * correct sched domains
6365 * Call with hotplug lock held
6367 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6369 cpumask_t change_map;
6372 cpus_and(*partition1, *partition1, cpu_online_map);
6373 cpus_and(*partition2, *partition2, cpu_online_map);
6374 cpus_or(change_map, *partition1, *partition2);
6376 /* Detach sched domains from all of the affected cpus */
6377 detach_destroy_domains(&change_map);
6378 if (!cpus_empty(*partition1))
6379 err = build_sched_domains(partition1);
6380 if (!err && !cpus_empty(*partition2))
6381 err = build_sched_domains(partition2);
6386 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6387 int arch_reinit_sched_domains(void)
6391 mutex_lock(&sched_hotcpu_mutex);
6392 detach_destroy_domains(&cpu_online_map);
6393 err = arch_init_sched_domains(&cpu_online_map);
6394 mutex_unlock(&sched_hotcpu_mutex);
6399 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6403 if (buf[0] != '0' && buf[0] != '1')
6407 sched_smt_power_savings = (buf[0] == '1');
6409 sched_mc_power_savings = (buf[0] == '1');
6411 ret = arch_reinit_sched_domains();
6413 return ret ? ret : count;
6416 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6420 #ifdef CONFIG_SCHED_SMT
6422 err = sysfs_create_file(&cls->kset.kobj,
6423 &attr_sched_smt_power_savings.attr);
6425 #ifdef CONFIG_SCHED_MC
6426 if (!err && mc_capable())
6427 err = sysfs_create_file(&cls->kset.kobj,
6428 &attr_sched_mc_power_savings.attr);
6434 #ifdef CONFIG_SCHED_MC
6435 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6437 return sprintf(page, "%u\n", sched_mc_power_savings);
6439 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6440 const char *buf, size_t count)
6442 return sched_power_savings_store(buf, count, 0);
6444 SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6445 sched_mc_power_savings_store);
6448 #ifdef CONFIG_SCHED_SMT
6449 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6451 return sprintf(page, "%u\n", sched_smt_power_savings);
6453 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6454 const char *buf, size_t count)
6456 return sched_power_savings_store(buf, count, 1);
6458 SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6459 sched_smt_power_savings_store);
6463 * Force a reinitialization of the sched domains hierarchy. The domains
6464 * and groups cannot be updated in place without racing with the balancing
6465 * code, so we temporarily attach all running cpus to the NULL domain
6466 * which will prevent rebalancing while the sched domains are recalculated.
6468 static int update_sched_domains(struct notifier_block *nfb,
6469 unsigned long action, void *hcpu)
6472 case CPU_UP_PREPARE:
6473 case CPU_UP_PREPARE_FROZEN:
6474 case CPU_DOWN_PREPARE:
6475 case CPU_DOWN_PREPARE_FROZEN:
6476 detach_destroy_domains(&cpu_online_map);
6479 case CPU_UP_CANCELED:
6480 case CPU_UP_CANCELED_FROZEN:
6481 case CPU_DOWN_FAILED:
6482 case CPU_DOWN_FAILED_FROZEN:
6484 case CPU_ONLINE_FROZEN:
6486 case CPU_DEAD_FROZEN:
6488 * Fall through and re-initialise the domains.
6495 /* The hotplug lock is already held by cpu_up/cpu_down */
6496 arch_init_sched_domains(&cpu_online_map);
6501 void __init sched_init_smp(void)
6503 cpumask_t non_isolated_cpus;
6505 mutex_lock(&sched_hotcpu_mutex);
6506 arch_init_sched_domains(&cpu_online_map);
6507 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6508 if (cpus_empty(non_isolated_cpus))
6509 cpu_set(smp_processor_id(), non_isolated_cpus);
6510 mutex_unlock(&sched_hotcpu_mutex);
6511 /* XXX: Theoretical race here - CPU may be hotplugged now */
6512 hotcpu_notifier(update_sched_domains, 0);
6514 /* Move init over to a non-isolated CPU */
6515 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6519 void __init sched_init_smp(void)
6522 #endif /* CONFIG_SMP */
6524 int in_sched_functions(unsigned long addr)
6526 /* Linker adds these: start and end of __sched functions */
6527 extern char __sched_text_start[], __sched_text_end[];
6529 return in_lock_functions(addr) ||
6530 (addr >= (unsigned long)__sched_text_start
6531 && addr < (unsigned long)__sched_text_end);
6534 void __init sched_init(void)
6537 int highest_cpu = 0;
6539 for_each_possible_cpu(i) {
6540 struct prio_array *array;
6544 spin_lock_init(&rq->lock);
6545 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6547 rq->active = rq->arrays;
6548 rq->expired = rq->arrays + 1;
6549 rq->best_expired_prio = MAX_PRIO;
6553 for (j = 1; j < 3; j++)
6554 rq->cpu_load[j] = 0;
6555 rq->active_balance = 0;
6558 rq->migration_thread = NULL;
6559 INIT_LIST_HEAD(&rq->migration_queue);
6561 atomic_set(&rq->nr_iowait, 0);
6563 for (j = 0; j < 2; j++) {
6564 array = rq->arrays + j;
6565 for (k = 0; k < MAX_PRIO; k++) {
6566 INIT_LIST_HEAD(array->queue + k);
6567 __clear_bit(k, array->bitmap);
6569 // delimiter for bitsearch
6570 __set_bit(MAX_PRIO, array->bitmap);
6575 set_load_weight(&init_task);
6578 nr_cpu_ids = highest_cpu + 1;
6579 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6582 #ifdef CONFIG_RT_MUTEXES
6583 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6587 * The boot idle thread does lazy MMU switching as well:
6589 atomic_inc(&init_mm.mm_count);
6590 enter_lazy_tlb(&init_mm, current);
6593 * Make us the idle thread. Technically, schedule() should not be
6594 * called from this thread, however somewhere below it might be,
6595 * but because we are the idle thread, we just pick up running again
6596 * when this runqueue becomes "idle".
6598 init_idle(current, smp_processor_id());
6601 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6602 void __might_sleep(char *file, int line)
6605 static unsigned long prev_jiffy; /* ratelimiting */
6607 if ((in_atomic() || irqs_disabled()) &&
6608 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6609 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6611 prev_jiffy = jiffies;
6612 printk(KERN_ERR "BUG: sleeping function called from invalid"
6613 " context at %s:%d\n", file, line);
6614 printk("in_atomic():%d, irqs_disabled():%d\n",
6615 in_atomic(), irqs_disabled());
6616 debug_show_held_locks(current);
6617 if (irqs_disabled())
6618 print_irqtrace_events(current);
6623 EXPORT_SYMBOL(__might_sleep);
6626 #ifdef CONFIG_MAGIC_SYSRQ
6627 void normalize_rt_tasks(void)
6629 struct prio_array *array;
6630 struct task_struct *g, *p;
6631 unsigned long flags;
6634 read_lock_irq(&tasklist_lock);
6636 do_each_thread(g, p) {
6640 spin_lock_irqsave(&p->pi_lock, flags);
6641 rq = __task_rq_lock(p);
6645 deactivate_task(p, task_rq(p));
6646 __setscheduler(p, SCHED_NORMAL, 0);
6648 __activate_task(p, task_rq(p));
6649 resched_task(rq->curr);
6652 __task_rq_unlock(rq);
6653 spin_unlock_irqrestore(&p->pi_lock, flags);
6654 } while_each_thread(g, p);
6656 read_unlock_irq(&tasklist_lock);
6659 #endif /* CONFIG_MAGIC_SYSRQ */
6663 * These functions are only useful for the IA64 MCA handling.
6665 * They can only be called when the whole system has been
6666 * stopped - every CPU needs to be quiescent, and no scheduling
6667 * activity can take place. Using them for anything else would
6668 * be a serious bug, and as a result, they aren't even visible
6669 * under any other configuration.
6673 * curr_task - return the current task for a given cpu.
6674 * @cpu: the processor in question.
6676 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6678 struct task_struct *curr_task(int cpu)
6680 return cpu_curr(cpu);
6684 * set_curr_task - set the current task for a given cpu.
6685 * @cpu: the processor in question.
6686 * @p: the task pointer to set.
6688 * Description: This function must only be used when non-maskable interrupts
6689 * are serviced on a separate stack. It allows the architecture to switch the
6690 * notion of the current task on a cpu in a non-blocking manner. This function
6691 * must be called with all CPU's synchronized, and interrupts disabled, the
6692 * and caller must save the original value of the current task (see
6693 * curr_task() above) and restore that value before reenabling interrupts and
6694 * re-starting the system.
6696 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6698 void set_curr_task(int cpu, struct task_struct *p)