IPoIB: Fix unused variable warning
[linux-2.6/linux-2.6-openrd.git] / kernel / sched.c
blob6107a0cd6325f9afe219007e1e1e446fda899b78
1 /*
2 * kernel/sched.c
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
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
27 #include <linux/mm.h>
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
65 #include <asm/tlb.h>
68 * Scheduler clock - returns current time in nanosec units.
69 * This is default implementation.
70 * Architectures and sub-architectures can override this.
72 unsigned long long __attribute__((weak)) sched_clock(void)
74 return (unsigned long long)jiffies * (1000000000 / HZ);
78 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
80 * and back.
82 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
91 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96 * Some helpers for converting nanosecond timing to jiffy resolution
98 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
99 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
108 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
109 * Timeslices get refilled after they expire.
111 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
112 #define DEF_TIMESLICE (100 * HZ / 1000)
114 #ifdef CONFIG_SMP
116 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
117 * Since cpu_power is a 'constant', we can use a reciprocal divide.
119 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
121 return reciprocal_divide(load, sg->reciprocal_cpu_power);
125 * Each time a sched group cpu_power is changed,
126 * we must compute its reciprocal value
128 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
130 sg->__cpu_power += val;
131 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
133 #endif
135 #define SCALE_PRIO(x, prio) \
136 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
139 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
140 * to time slice values: [800ms ... 100ms ... 5ms]
142 static unsigned int static_prio_timeslice(int static_prio)
144 if (static_prio == NICE_TO_PRIO(19))
145 return 1;
147 if (static_prio < NICE_TO_PRIO(0))
148 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
149 else
150 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
153 static inline int rt_policy(int policy)
155 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
156 return 1;
157 return 0;
160 static inline int task_has_rt_policy(struct task_struct *p)
162 return rt_policy(p->policy);
166 * This is the priority-queue data structure of the RT scheduling class:
168 struct rt_prio_array {
169 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
170 struct list_head queue[MAX_RT_PRIO];
173 struct load_stat {
174 struct load_weight load;
175 u64 load_update_start, load_update_last;
176 unsigned long delta_fair, delta_exec, delta_stat;
179 /* CFS-related fields in a runqueue */
180 struct cfs_rq {
181 struct load_weight load;
182 unsigned long nr_running;
184 s64 fair_clock;
185 u64 exec_clock;
186 s64 wait_runtime;
187 u64 sleeper_bonus;
188 unsigned long wait_runtime_overruns, wait_runtime_underruns;
190 struct rb_root tasks_timeline;
191 struct rb_node *rb_leftmost;
192 struct rb_node *rb_load_balance_curr;
193 #ifdef CONFIG_FAIR_GROUP_SCHED
194 /* 'curr' points to currently running entity on this cfs_rq.
195 * It is set to NULL otherwise (i.e when none are currently running).
197 struct sched_entity *curr;
198 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
207 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
208 #endif
211 /* Real-Time classes' related field in a runqueue: */
212 struct rt_rq {
213 struct rt_prio_array active;
214 int rt_load_balance_idx;
215 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
219 * This is the main, per-CPU runqueue data structure.
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
225 struct rq {
226 spinlock_t lock; /* runqueue lock */
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
232 unsigned long nr_running;
233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
235 unsigned char idle_at_tick;
236 #ifdef CONFIG_NO_HZ
237 unsigned char in_nohz_recently;
238 #endif
239 struct load_stat ls; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates;
241 u64 nr_switches;
243 struct cfs_rq cfs;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
246 #endif
247 struct rt_rq rt;
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
255 unsigned long nr_uninterruptible;
257 struct task_struct *curr, *idle;
258 unsigned long next_balance;
259 struct mm_struct *prev_mm;
261 u64 clock, prev_clock_raw;
262 s64 clock_max_delta;
264 unsigned int clock_warps, clock_overflows;
265 u64 idle_clock;
266 unsigned int clock_deep_idle_events;
267 u64 tick_timestamp;
269 atomic_t nr_iowait;
271 #ifdef CONFIG_SMP
272 struct sched_domain *sd;
274 /* For active balancing */
275 int active_balance;
276 int push_cpu;
277 int cpu; /* cpu of this runqueue */
279 struct task_struct *migration_thread;
280 struct list_head migration_queue;
281 #endif
283 #ifdef CONFIG_SCHEDSTATS
284 /* latency stats */
285 struct sched_info rq_sched_info;
287 /* sys_sched_yield() stats */
288 unsigned long yld_exp_empty;
289 unsigned long yld_act_empty;
290 unsigned long yld_both_empty;
291 unsigned long yld_cnt;
293 /* schedule() stats */
294 unsigned long sched_switch;
295 unsigned long sched_cnt;
296 unsigned long sched_goidle;
298 /* try_to_wake_up() stats */
299 unsigned long ttwu_cnt;
300 unsigned long ttwu_local;
301 #endif
302 struct lock_class_key rq_lock_key;
305 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
306 static DEFINE_MUTEX(sched_hotcpu_mutex);
308 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
310 rq->curr->sched_class->check_preempt_curr(rq, p);
313 static inline int cpu_of(struct rq *rq)
315 #ifdef CONFIG_SMP
316 return rq->cpu;
317 #else
318 return 0;
319 #endif
323 * Update the per-runqueue clock, as finegrained as the platform can give
324 * us, but without assuming monotonicity, etc.:
326 static void __update_rq_clock(struct rq *rq)
328 u64 prev_raw = rq->prev_clock_raw;
329 u64 now = sched_clock();
330 s64 delta = now - prev_raw;
331 u64 clock = rq->clock;
333 #ifdef CONFIG_SCHED_DEBUG
334 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
335 #endif
337 * Protect against sched_clock() occasionally going backwards:
339 if (unlikely(delta < 0)) {
340 clock++;
341 rq->clock_warps++;
342 } else {
344 * Catch too large forward jumps too:
346 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
347 if (clock < rq->tick_timestamp + TICK_NSEC)
348 clock = rq->tick_timestamp + TICK_NSEC;
349 else
350 clock++;
351 rq->clock_overflows++;
352 } else {
353 if (unlikely(delta > rq->clock_max_delta))
354 rq->clock_max_delta = delta;
355 clock += delta;
359 rq->prev_clock_raw = now;
360 rq->clock = clock;
363 static void update_rq_clock(struct rq *rq)
365 if (likely(smp_processor_id() == cpu_of(rq)))
366 __update_rq_clock(rq);
370 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
371 * See detach_destroy_domains: synchronize_sched for details.
373 * The domain tree of any CPU may only be accessed from within
374 * preempt-disabled sections.
376 #define for_each_domain(cpu, __sd) \
377 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
379 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
380 #define this_rq() (&__get_cpu_var(runqueues))
381 #define task_rq(p) cpu_rq(task_cpu(p))
382 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
385 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
386 * clock constructed from sched_clock():
388 unsigned long long cpu_clock(int cpu)
390 unsigned long long now;
391 unsigned long flags;
392 struct rq *rq;
394 local_irq_save(flags);
395 rq = cpu_rq(cpu);
396 update_rq_clock(rq);
397 now = rq->clock;
398 local_irq_restore(flags);
400 return now;
403 #ifdef CONFIG_FAIR_GROUP_SCHED
404 /* Change a task's ->cfs_rq if it moves across CPUs */
405 static inline void set_task_cfs_rq(struct task_struct *p)
407 p->se.cfs_rq = &task_rq(p)->cfs;
409 #else
410 static inline void set_task_cfs_rq(struct task_struct *p)
413 #endif
415 #ifndef prepare_arch_switch
416 # define prepare_arch_switch(next) do { } while (0)
417 #endif
418 #ifndef finish_arch_switch
419 # define finish_arch_switch(prev) do { } while (0)
420 #endif
422 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
423 static inline int task_running(struct rq *rq, struct task_struct *p)
425 return rq->curr == p;
428 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
432 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
434 #ifdef CONFIG_DEBUG_SPINLOCK
435 /* this is a valid case when another task releases the spinlock */
436 rq->lock.owner = current;
437 #endif
439 * If we are tracking spinlock dependencies then we have to
440 * fix up the runqueue lock - which gets 'carried over' from
441 * prev into current:
443 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
445 spin_unlock_irq(&rq->lock);
448 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
449 static inline int task_running(struct rq *rq, struct task_struct *p)
451 #ifdef CONFIG_SMP
452 return p->oncpu;
453 #else
454 return rq->curr == p;
455 #endif
458 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
460 #ifdef CONFIG_SMP
462 * We can optimise this out completely for !SMP, because the
463 * SMP rebalancing from interrupt is the only thing that cares
464 * here.
466 next->oncpu = 1;
467 #endif
468 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
469 spin_unlock_irq(&rq->lock);
470 #else
471 spin_unlock(&rq->lock);
472 #endif
475 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
477 #ifdef CONFIG_SMP
479 * After ->oncpu is cleared, the task can be moved to a different CPU.
480 * We must ensure this doesn't happen until the switch is completely
481 * finished.
483 smp_wmb();
484 prev->oncpu = 0;
485 #endif
486 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
487 local_irq_enable();
488 #endif
490 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
493 * __task_rq_lock - lock the runqueue a given task resides on.
494 * Must be called interrupts disabled.
496 static inline struct rq *__task_rq_lock(struct task_struct *p)
497 __acquires(rq->lock)
499 struct rq *rq;
501 repeat_lock_task:
502 rq = task_rq(p);
503 spin_lock(&rq->lock);
504 if (unlikely(rq != task_rq(p))) {
505 spin_unlock(&rq->lock);
506 goto repeat_lock_task;
508 return rq;
512 * task_rq_lock - lock the runqueue a given task resides on and disable
513 * interrupts. Note the ordering: we can safely lookup the task_rq without
514 * explicitly disabling preemption.
516 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
517 __acquires(rq->lock)
519 struct rq *rq;
521 repeat_lock_task:
522 local_irq_save(*flags);
523 rq = task_rq(p);
524 spin_lock(&rq->lock);
525 if (unlikely(rq != task_rq(p))) {
526 spin_unlock_irqrestore(&rq->lock, *flags);
527 goto repeat_lock_task;
529 return rq;
532 static inline void __task_rq_unlock(struct rq *rq)
533 __releases(rq->lock)
535 spin_unlock(&rq->lock);
538 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
539 __releases(rq->lock)
541 spin_unlock_irqrestore(&rq->lock, *flags);
545 * this_rq_lock - lock this runqueue and disable interrupts.
547 static inline struct rq *this_rq_lock(void)
548 __acquires(rq->lock)
550 struct rq *rq;
552 local_irq_disable();
553 rq = this_rq();
554 spin_lock(&rq->lock);
556 return rq;
560 * We are going deep-idle (irqs are disabled):
562 void sched_clock_idle_sleep_event(void)
564 struct rq *rq = cpu_rq(smp_processor_id());
566 spin_lock(&rq->lock);
567 __update_rq_clock(rq);
568 spin_unlock(&rq->lock);
569 rq->clock_deep_idle_events++;
571 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
574 * We just idled delta nanoseconds (called with irqs disabled):
576 void sched_clock_idle_wakeup_event(u64 delta_ns)
578 struct rq *rq = cpu_rq(smp_processor_id());
579 u64 now = sched_clock();
581 rq->idle_clock += delta_ns;
583 * Override the previous timestamp and ignore all
584 * sched_clock() deltas that occured while we idled,
585 * and use the PM-provided delta_ns to advance the
586 * rq clock:
588 spin_lock(&rq->lock);
589 rq->prev_clock_raw = now;
590 rq->clock += delta_ns;
591 spin_unlock(&rq->lock);
593 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
596 * resched_task - mark a task 'to be rescheduled now'.
598 * On UP this means the setting of the need_resched flag, on SMP it
599 * might also involve a cross-CPU call to trigger the scheduler on
600 * the target CPU.
602 #ifdef CONFIG_SMP
604 #ifndef tsk_is_polling
605 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
606 #endif
608 static void resched_task(struct task_struct *p)
610 int cpu;
612 assert_spin_locked(&task_rq(p)->lock);
614 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
615 return;
617 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
619 cpu = task_cpu(p);
620 if (cpu == smp_processor_id())
621 return;
623 /* NEED_RESCHED must be visible before we test polling */
624 smp_mb();
625 if (!tsk_is_polling(p))
626 smp_send_reschedule(cpu);
629 static void resched_cpu(int cpu)
631 struct rq *rq = cpu_rq(cpu);
632 unsigned long flags;
634 if (!spin_trylock_irqsave(&rq->lock, flags))
635 return;
636 resched_task(cpu_curr(cpu));
637 spin_unlock_irqrestore(&rq->lock, flags);
639 #else
640 static inline void resched_task(struct task_struct *p)
642 assert_spin_locked(&task_rq(p)->lock);
643 set_tsk_need_resched(p);
645 #endif
647 static u64 div64_likely32(u64 divident, unsigned long divisor)
649 #if BITS_PER_LONG == 32
650 if (likely(divident <= 0xffffffffULL))
651 return (u32)divident / divisor;
652 do_div(divident, divisor);
654 return divident;
655 #else
656 return divident / divisor;
657 #endif
660 #if BITS_PER_LONG == 32
661 # define WMULT_CONST (~0UL)
662 #else
663 # define WMULT_CONST (1UL << 32)
664 #endif
666 #define WMULT_SHIFT 32
669 * Shift right and round:
671 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
673 static unsigned long
674 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
675 struct load_weight *lw)
677 u64 tmp;
679 if (unlikely(!lw->inv_weight))
680 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
682 tmp = (u64)delta_exec * weight;
684 * Check whether we'd overflow the 64-bit multiplication:
686 if (unlikely(tmp > WMULT_CONST))
687 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
688 WMULT_SHIFT/2);
689 else
690 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
692 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
695 static inline unsigned long
696 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
698 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
701 static void update_load_add(struct load_weight *lw, unsigned long inc)
703 lw->weight += inc;
704 lw->inv_weight = 0;
707 static void update_load_sub(struct load_weight *lw, unsigned long dec)
709 lw->weight -= dec;
710 lw->inv_weight = 0;
714 * To aid in avoiding the subversion of "niceness" due to uneven distribution
715 * of tasks with abnormal "nice" values across CPUs the contribution that
716 * each task makes to its run queue's load is weighted according to its
717 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
718 * scaled version of the new time slice allocation that they receive on time
719 * slice expiry etc.
722 #define WEIGHT_IDLEPRIO 2
723 #define WMULT_IDLEPRIO (1 << 31)
726 * Nice levels are multiplicative, with a gentle 10% change for every
727 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
728 * nice 1, it will get ~10% less CPU time than another CPU-bound task
729 * that remained on nice 0.
731 * The "10% effect" is relative and cumulative: from _any_ nice level,
732 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
733 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
734 * If a task goes up by ~10% and another task goes down by ~10% then
735 * the relative distance between them is ~25%.)
737 static const int prio_to_weight[40] = {
738 /* -20 */ 88761, 71755, 56483, 46273, 36291,
739 /* -15 */ 29154, 23254, 18705, 14949, 11916,
740 /* -10 */ 9548, 7620, 6100, 4904, 3906,
741 /* -5 */ 3121, 2501, 1991, 1586, 1277,
742 /* 0 */ 1024, 820, 655, 526, 423,
743 /* 5 */ 335, 272, 215, 172, 137,
744 /* 10 */ 110, 87, 70, 56, 45,
745 /* 15 */ 36, 29, 23, 18, 15,
749 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
751 * In cases where the weight does not change often, we can use the
752 * precalculated inverse to speed up arithmetics by turning divisions
753 * into multiplications:
755 static const u32 prio_to_wmult[40] = {
756 /* -20 */ 48388, 59856, 76040, 92818, 118348,
757 /* -15 */ 147320, 184698, 229616, 287308, 360437,
758 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
759 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
760 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
761 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
762 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
763 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
766 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
769 * runqueue iterator, to support SMP load-balancing between different
770 * scheduling classes, without having to expose their internal data
771 * structures to the load-balancing proper:
773 struct rq_iterator {
774 void *arg;
775 struct task_struct *(*start)(void *);
776 struct task_struct *(*next)(void *);
779 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
780 unsigned long max_nr_move, unsigned long max_load_move,
781 struct sched_domain *sd, enum cpu_idle_type idle,
782 int *all_pinned, unsigned long *load_moved,
783 int *this_best_prio, struct rq_iterator *iterator);
785 #include "sched_stats.h"
786 #include "sched_rt.c"
787 #include "sched_fair.c"
788 #include "sched_idletask.c"
789 #ifdef CONFIG_SCHED_DEBUG
790 # include "sched_debug.c"
791 #endif
793 #define sched_class_highest (&rt_sched_class)
795 static void __update_curr_load(struct rq *rq, struct load_stat *ls)
797 if (rq->curr != rq->idle && ls->load.weight) {
798 ls->delta_exec += ls->delta_stat;
799 ls->delta_fair += calc_delta_fair(ls->delta_stat, &ls->load);
800 ls->delta_stat = 0;
805 * Update delta_exec, delta_fair fields for rq.
807 * delta_fair clock advances at a rate inversely proportional to
808 * total load (rq->ls.load.weight) on the runqueue, while
809 * delta_exec advances at the same rate as wall-clock (provided
810 * cpu is not idle).
812 * delta_exec / delta_fair is a measure of the (smoothened) load on this
813 * runqueue over any given interval. This (smoothened) load is used
814 * during load balance.
816 * This function is called /before/ updating rq->ls.load
817 * and when switching tasks.
819 static void update_curr_load(struct rq *rq)
821 struct load_stat *ls = &rq->ls;
822 u64 start;
824 start = ls->load_update_start;
825 ls->load_update_start = rq->clock;
826 ls->delta_stat += rq->clock - start;
828 * Stagger updates to ls->delta_fair. Very frequent updates
829 * can be expensive.
831 if (ls->delta_stat >= sysctl_sched_stat_granularity)
832 __update_curr_load(rq, ls);
835 static inline void inc_load(struct rq *rq, const struct task_struct *p)
837 update_curr_load(rq);
838 update_load_add(&rq->ls.load, p->se.load.weight);
841 static inline void dec_load(struct rq *rq, const struct task_struct *p)
843 update_curr_load(rq);
844 update_load_sub(&rq->ls.load, p->se.load.weight);
847 static void inc_nr_running(struct task_struct *p, struct rq *rq)
849 rq->nr_running++;
850 inc_load(rq, p);
853 static void dec_nr_running(struct task_struct *p, struct rq *rq)
855 rq->nr_running--;
856 dec_load(rq, p);
859 static void set_load_weight(struct task_struct *p)
861 p->se.wait_runtime = 0;
863 if (task_has_rt_policy(p)) {
864 p->se.load.weight = prio_to_weight[0] * 2;
865 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
866 return;
870 * SCHED_IDLE tasks get minimal weight:
872 if (p->policy == SCHED_IDLE) {
873 p->se.load.weight = WEIGHT_IDLEPRIO;
874 p->se.load.inv_weight = WMULT_IDLEPRIO;
875 return;
878 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
879 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
882 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
884 sched_info_queued(p);
885 p->sched_class->enqueue_task(rq, p, wakeup);
886 p->se.on_rq = 1;
889 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
891 p->sched_class->dequeue_task(rq, p, sleep);
892 p->se.on_rq = 0;
896 * __normal_prio - return the priority that is based on the static prio
898 static inline int __normal_prio(struct task_struct *p)
900 return p->static_prio;
904 * Calculate the expected normal priority: i.e. priority
905 * without taking RT-inheritance into account. Might be
906 * boosted by interactivity modifiers. Changes upon fork,
907 * setprio syscalls, and whenever the interactivity
908 * estimator recalculates.
910 static inline int normal_prio(struct task_struct *p)
912 int prio;
914 if (task_has_rt_policy(p))
915 prio = MAX_RT_PRIO-1 - p->rt_priority;
916 else
917 prio = __normal_prio(p);
918 return prio;
922 * Calculate the current priority, i.e. the priority
923 * taken into account by the scheduler. This value might
924 * be boosted by RT tasks, or might be boosted by
925 * interactivity modifiers. Will be RT if the task got
926 * RT-boosted. If not then it returns p->normal_prio.
928 static int effective_prio(struct task_struct *p)
930 p->normal_prio = normal_prio(p);
932 * If we are RT tasks or we were boosted to RT priority,
933 * keep the priority unchanged. Otherwise, update priority
934 * to the normal priority:
936 if (!rt_prio(p->prio))
937 return p->normal_prio;
938 return p->prio;
942 * activate_task - move a task to the runqueue.
944 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
946 if (p->state == TASK_UNINTERRUPTIBLE)
947 rq->nr_uninterruptible--;
949 enqueue_task(rq, p, wakeup);
950 inc_nr_running(p, rq);
954 * activate_idle_task - move idle task to the _front_ of runqueue.
956 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
958 update_rq_clock(rq);
960 if (p->state == TASK_UNINTERRUPTIBLE)
961 rq->nr_uninterruptible--;
963 enqueue_task(rq, p, 0);
964 inc_nr_running(p, rq);
968 * deactivate_task - remove a task from the runqueue.
970 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
972 if (p->state == TASK_UNINTERRUPTIBLE)
973 rq->nr_uninterruptible++;
975 dequeue_task(rq, p, sleep);
976 dec_nr_running(p, rq);
980 * task_curr - is this task currently executing on a CPU?
981 * @p: the task in question.
983 inline int task_curr(const struct task_struct *p)
985 return cpu_curr(task_cpu(p)) == p;
988 /* Used instead of source_load when we know the type == 0 */
989 unsigned long weighted_cpuload(const int cpu)
991 return cpu_rq(cpu)->ls.load.weight;
994 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
996 #ifdef CONFIG_SMP
997 task_thread_info(p)->cpu = cpu;
998 set_task_cfs_rq(p);
999 #endif
1002 #ifdef CONFIG_SMP
1004 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1006 int old_cpu = task_cpu(p);
1007 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1008 u64 clock_offset, fair_clock_offset;
1010 clock_offset = old_rq->clock - new_rq->clock;
1011 fair_clock_offset = old_rq->cfs.fair_clock - new_rq->cfs.fair_clock;
1013 if (p->se.wait_start_fair)
1014 p->se.wait_start_fair -= fair_clock_offset;
1015 if (p->se.sleep_start_fair)
1016 p->se.sleep_start_fair -= fair_clock_offset;
1018 #ifdef CONFIG_SCHEDSTATS
1019 if (p->se.wait_start)
1020 p->se.wait_start -= clock_offset;
1021 if (p->se.sleep_start)
1022 p->se.sleep_start -= clock_offset;
1023 if (p->se.block_start)
1024 p->se.block_start -= clock_offset;
1025 #endif
1027 __set_task_cpu(p, new_cpu);
1030 struct migration_req {
1031 struct list_head list;
1033 struct task_struct *task;
1034 int dest_cpu;
1036 struct completion done;
1040 * The task's runqueue lock must be held.
1041 * Returns true if you have to wait for migration thread.
1043 static int
1044 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1046 struct rq *rq = task_rq(p);
1049 * If the task is not on a runqueue (and not running), then
1050 * it is sufficient to simply update the task's cpu field.
1052 if (!p->se.on_rq && !task_running(rq, p)) {
1053 set_task_cpu(p, dest_cpu);
1054 return 0;
1057 init_completion(&req->done);
1058 req->task = p;
1059 req->dest_cpu = dest_cpu;
1060 list_add(&req->list, &rq->migration_queue);
1062 return 1;
1066 * wait_task_inactive - wait for a thread to unschedule.
1068 * The caller must ensure that the task *will* unschedule sometime soon,
1069 * else this function might spin for a *long* time. This function can't
1070 * be called with interrupts off, or it may introduce deadlock with
1071 * smp_call_function() if an IPI is sent by the same process we are
1072 * waiting to become inactive.
1074 void wait_task_inactive(struct task_struct *p)
1076 unsigned long flags;
1077 int running, on_rq;
1078 struct rq *rq;
1080 repeat:
1082 * We do the initial early heuristics without holding
1083 * any task-queue locks at all. We'll only try to get
1084 * the runqueue lock when things look like they will
1085 * work out!
1087 rq = task_rq(p);
1090 * If the task is actively running on another CPU
1091 * still, just relax and busy-wait without holding
1092 * any locks.
1094 * NOTE! Since we don't hold any locks, it's not
1095 * even sure that "rq" stays as the right runqueue!
1096 * But we don't care, since "task_running()" will
1097 * return false if the runqueue has changed and p
1098 * is actually now running somewhere else!
1100 while (task_running(rq, p))
1101 cpu_relax();
1104 * Ok, time to look more closely! We need the rq
1105 * lock now, to be *sure*. If we're wrong, we'll
1106 * just go back and repeat.
1108 rq = task_rq_lock(p, &flags);
1109 running = task_running(rq, p);
1110 on_rq = p->se.on_rq;
1111 task_rq_unlock(rq, &flags);
1114 * Was it really running after all now that we
1115 * checked with the proper locks actually held?
1117 * Oops. Go back and try again..
1119 if (unlikely(running)) {
1120 cpu_relax();
1121 goto repeat;
1125 * It's not enough that it's not actively running,
1126 * it must be off the runqueue _entirely_, and not
1127 * preempted!
1129 * So if it wa still runnable (but just not actively
1130 * running right now), it's preempted, and we should
1131 * yield - it could be a while.
1133 if (unlikely(on_rq)) {
1134 yield();
1135 goto repeat;
1139 * Ahh, all good. It wasn't running, and it wasn't
1140 * runnable, which means that it will never become
1141 * running in the future either. We're all done!
1145 /***
1146 * kick_process - kick a running thread to enter/exit the kernel
1147 * @p: the to-be-kicked thread
1149 * Cause a process which is running on another CPU to enter
1150 * kernel-mode, without any delay. (to get signals handled.)
1152 * NOTE: this function doesnt have to take the runqueue lock,
1153 * because all it wants to ensure is that the remote task enters
1154 * the kernel. If the IPI races and the task has been migrated
1155 * to another CPU then no harm is done and the purpose has been
1156 * achieved as well.
1158 void kick_process(struct task_struct *p)
1160 int cpu;
1162 preempt_disable();
1163 cpu = task_cpu(p);
1164 if ((cpu != smp_processor_id()) && task_curr(p))
1165 smp_send_reschedule(cpu);
1166 preempt_enable();
1170 * Return a low guess at the load of a migration-source cpu weighted
1171 * according to the scheduling class and "nice" value.
1173 * We want to under-estimate the load of migration sources, to
1174 * balance conservatively.
1176 static inline unsigned long source_load(int cpu, int type)
1178 struct rq *rq = cpu_rq(cpu);
1179 unsigned long total = weighted_cpuload(cpu);
1181 if (type == 0)
1182 return total;
1184 return min(rq->cpu_load[type-1], total);
1188 * Return a high guess at the load of a migration-target cpu weighted
1189 * according to the scheduling class and "nice" value.
1191 static inline unsigned long target_load(int cpu, int type)
1193 struct rq *rq = cpu_rq(cpu);
1194 unsigned long total = weighted_cpuload(cpu);
1196 if (type == 0)
1197 return total;
1199 return max(rq->cpu_load[type-1], total);
1203 * Return the average load per task on the cpu's run queue
1205 static inline unsigned long cpu_avg_load_per_task(int cpu)
1207 struct rq *rq = cpu_rq(cpu);
1208 unsigned long total = weighted_cpuload(cpu);
1209 unsigned long n = rq->nr_running;
1211 return n ? total / n : SCHED_LOAD_SCALE;
1215 * find_idlest_group finds and returns the least busy CPU group within the
1216 * domain.
1218 static struct sched_group *
1219 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1221 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1222 unsigned long min_load = ULONG_MAX, this_load = 0;
1223 int load_idx = sd->forkexec_idx;
1224 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1226 do {
1227 unsigned long load, avg_load;
1228 int local_group;
1229 int i;
1231 /* Skip over this group if it has no CPUs allowed */
1232 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1233 goto nextgroup;
1235 local_group = cpu_isset(this_cpu, group->cpumask);
1237 /* Tally up the load of all CPUs in the group */
1238 avg_load = 0;
1240 for_each_cpu_mask(i, group->cpumask) {
1241 /* Bias balancing toward cpus of our domain */
1242 if (local_group)
1243 load = source_load(i, load_idx);
1244 else
1245 load = target_load(i, load_idx);
1247 avg_load += load;
1250 /* Adjust by relative CPU power of the group */
1251 avg_load = sg_div_cpu_power(group,
1252 avg_load * SCHED_LOAD_SCALE);
1254 if (local_group) {
1255 this_load = avg_load;
1256 this = group;
1257 } else if (avg_load < min_load) {
1258 min_load = avg_load;
1259 idlest = group;
1261 nextgroup:
1262 group = group->next;
1263 } while (group != sd->groups);
1265 if (!idlest || 100*this_load < imbalance*min_load)
1266 return NULL;
1267 return idlest;
1271 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1273 static int
1274 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1276 cpumask_t tmp;
1277 unsigned long load, min_load = ULONG_MAX;
1278 int idlest = -1;
1279 int i;
1281 /* Traverse only the allowed CPUs */
1282 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1284 for_each_cpu_mask(i, tmp) {
1285 load = weighted_cpuload(i);
1287 if (load < min_load || (load == min_load && i == this_cpu)) {
1288 min_load = load;
1289 idlest = i;
1293 return idlest;
1297 * sched_balance_self: balance the current task (running on cpu) in domains
1298 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1299 * SD_BALANCE_EXEC.
1301 * Balance, ie. select the least loaded group.
1303 * Returns the target CPU number, or the same CPU if no balancing is needed.
1305 * preempt must be disabled.
1307 static int sched_balance_self(int cpu, int flag)
1309 struct task_struct *t = current;
1310 struct sched_domain *tmp, *sd = NULL;
1312 for_each_domain(cpu, tmp) {
1314 * If power savings logic is enabled for a domain, stop there.
1316 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1317 break;
1318 if (tmp->flags & flag)
1319 sd = tmp;
1322 while (sd) {
1323 cpumask_t span;
1324 struct sched_group *group;
1325 int new_cpu, weight;
1327 if (!(sd->flags & flag)) {
1328 sd = sd->child;
1329 continue;
1332 span = sd->span;
1333 group = find_idlest_group(sd, t, cpu);
1334 if (!group) {
1335 sd = sd->child;
1336 continue;
1339 new_cpu = find_idlest_cpu(group, t, cpu);
1340 if (new_cpu == -1 || new_cpu == cpu) {
1341 /* Now try balancing at a lower domain level of cpu */
1342 sd = sd->child;
1343 continue;
1346 /* Now try balancing at a lower domain level of new_cpu */
1347 cpu = new_cpu;
1348 sd = NULL;
1349 weight = cpus_weight(span);
1350 for_each_domain(cpu, tmp) {
1351 if (weight <= cpus_weight(tmp->span))
1352 break;
1353 if (tmp->flags & flag)
1354 sd = tmp;
1356 /* while loop will break here if sd == NULL */
1359 return cpu;
1362 #endif /* CONFIG_SMP */
1365 * wake_idle() will wake a task on an idle cpu if task->cpu is
1366 * not idle and an idle cpu is available. The span of cpus to
1367 * search starts with cpus closest then further out as needed,
1368 * so we always favor a closer, idle cpu.
1370 * Returns the CPU we should wake onto.
1372 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1373 static int wake_idle(int cpu, struct task_struct *p)
1375 cpumask_t tmp;
1376 struct sched_domain *sd;
1377 int i;
1380 * If it is idle, then it is the best cpu to run this task.
1382 * This cpu is also the best, if it has more than one task already.
1383 * Siblings must be also busy(in most cases) as they didn't already
1384 * pickup the extra load from this cpu and hence we need not check
1385 * sibling runqueue info. This will avoid the checks and cache miss
1386 * penalities associated with that.
1388 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1389 return cpu;
1391 for_each_domain(cpu, sd) {
1392 if (sd->flags & SD_WAKE_IDLE) {
1393 cpus_and(tmp, sd->span, p->cpus_allowed);
1394 for_each_cpu_mask(i, tmp) {
1395 if (idle_cpu(i))
1396 return i;
1398 } else {
1399 break;
1402 return cpu;
1404 #else
1405 static inline int wake_idle(int cpu, struct task_struct *p)
1407 return cpu;
1409 #endif
1411 /***
1412 * try_to_wake_up - wake up a thread
1413 * @p: the to-be-woken-up thread
1414 * @state: the mask of task states that can be woken
1415 * @sync: do a synchronous wakeup?
1417 * Put it on the run-queue if it's not already there. The "current"
1418 * thread is always on the run-queue (except when the actual
1419 * re-schedule is in progress), and as such you're allowed to do
1420 * the simpler "current->state = TASK_RUNNING" to mark yourself
1421 * runnable without the overhead of this.
1423 * returns failure only if the task is already active.
1425 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1427 int cpu, this_cpu, success = 0;
1428 unsigned long flags;
1429 long old_state;
1430 struct rq *rq;
1431 #ifdef CONFIG_SMP
1432 struct sched_domain *sd, *this_sd = NULL;
1433 unsigned long load, this_load;
1434 int new_cpu;
1435 #endif
1437 rq = task_rq_lock(p, &flags);
1438 old_state = p->state;
1439 if (!(old_state & state))
1440 goto out;
1442 if (p->se.on_rq)
1443 goto out_running;
1445 cpu = task_cpu(p);
1446 this_cpu = smp_processor_id();
1448 #ifdef CONFIG_SMP
1449 if (unlikely(task_running(rq, p)))
1450 goto out_activate;
1452 new_cpu = cpu;
1454 schedstat_inc(rq, ttwu_cnt);
1455 if (cpu == this_cpu) {
1456 schedstat_inc(rq, ttwu_local);
1457 goto out_set_cpu;
1460 for_each_domain(this_cpu, sd) {
1461 if (cpu_isset(cpu, sd->span)) {
1462 schedstat_inc(sd, ttwu_wake_remote);
1463 this_sd = sd;
1464 break;
1468 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1469 goto out_set_cpu;
1472 * Check for affine wakeup and passive balancing possibilities.
1474 if (this_sd) {
1475 int idx = this_sd->wake_idx;
1476 unsigned int imbalance;
1478 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1480 load = source_load(cpu, idx);
1481 this_load = target_load(this_cpu, idx);
1483 new_cpu = this_cpu; /* Wake to this CPU if we can */
1485 if (this_sd->flags & SD_WAKE_AFFINE) {
1486 unsigned long tl = this_load;
1487 unsigned long tl_per_task;
1489 tl_per_task = cpu_avg_load_per_task(this_cpu);
1492 * If sync wakeup then subtract the (maximum possible)
1493 * effect of the currently running task from the load
1494 * of the current CPU:
1496 if (sync)
1497 tl -= current->se.load.weight;
1499 if ((tl <= load &&
1500 tl + target_load(cpu, idx) <= tl_per_task) ||
1501 100*(tl + p->se.load.weight) <= imbalance*load) {
1503 * This domain has SD_WAKE_AFFINE and
1504 * p is cache cold in this domain, and
1505 * there is no bad imbalance.
1507 schedstat_inc(this_sd, ttwu_move_affine);
1508 goto out_set_cpu;
1513 * Start passive balancing when half the imbalance_pct
1514 * limit is reached.
1516 if (this_sd->flags & SD_WAKE_BALANCE) {
1517 if (imbalance*this_load <= 100*load) {
1518 schedstat_inc(this_sd, ttwu_move_balance);
1519 goto out_set_cpu;
1524 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1525 out_set_cpu:
1526 new_cpu = wake_idle(new_cpu, p);
1527 if (new_cpu != cpu) {
1528 set_task_cpu(p, new_cpu);
1529 task_rq_unlock(rq, &flags);
1530 /* might preempt at this point */
1531 rq = task_rq_lock(p, &flags);
1532 old_state = p->state;
1533 if (!(old_state & state))
1534 goto out;
1535 if (p->se.on_rq)
1536 goto out_running;
1538 this_cpu = smp_processor_id();
1539 cpu = task_cpu(p);
1542 out_activate:
1543 #endif /* CONFIG_SMP */
1544 update_rq_clock(rq);
1545 activate_task(rq, p, 1);
1547 * Sync wakeups (i.e. those types of wakeups where the waker
1548 * has indicated that it will leave the CPU in short order)
1549 * don't trigger a preemption, if the woken up task will run on
1550 * this cpu. (in this case the 'I will reschedule' promise of
1551 * the waker guarantees that the freshly woken up task is going
1552 * to be considered on this CPU.)
1554 if (!sync || cpu != this_cpu)
1555 check_preempt_curr(rq, p);
1556 success = 1;
1558 out_running:
1559 p->state = TASK_RUNNING;
1560 out:
1561 task_rq_unlock(rq, &flags);
1563 return success;
1566 int fastcall wake_up_process(struct task_struct *p)
1568 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1569 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1571 EXPORT_SYMBOL(wake_up_process);
1573 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1575 return try_to_wake_up(p, state, 0);
1579 * Perform scheduler related setup for a newly forked process p.
1580 * p is forked by current.
1582 * __sched_fork() is basic setup used by init_idle() too:
1584 static void __sched_fork(struct task_struct *p)
1586 p->se.wait_start_fair = 0;
1587 p->se.exec_start = 0;
1588 p->se.sum_exec_runtime = 0;
1589 p->se.prev_sum_exec_runtime = 0;
1590 p->se.delta_exec = 0;
1591 p->se.delta_fair_run = 0;
1592 p->se.delta_fair_sleep = 0;
1593 p->se.wait_runtime = 0;
1594 p->se.sleep_start_fair = 0;
1596 #ifdef CONFIG_SCHEDSTATS
1597 p->se.wait_start = 0;
1598 p->se.sum_wait_runtime = 0;
1599 p->se.sum_sleep_runtime = 0;
1600 p->se.sleep_start = 0;
1601 p->se.block_start = 0;
1602 p->se.sleep_max = 0;
1603 p->se.block_max = 0;
1604 p->se.exec_max = 0;
1605 p->se.wait_max = 0;
1606 p->se.wait_runtime_overruns = 0;
1607 p->se.wait_runtime_underruns = 0;
1608 #endif
1610 INIT_LIST_HEAD(&p->run_list);
1611 p->se.on_rq = 0;
1613 #ifdef CONFIG_PREEMPT_NOTIFIERS
1614 INIT_HLIST_HEAD(&p->preempt_notifiers);
1615 #endif
1618 * We mark the process as running here, but have not actually
1619 * inserted it onto the runqueue yet. This guarantees that
1620 * nobody will actually run it, and a signal or other external
1621 * event cannot wake it up and insert it on the runqueue either.
1623 p->state = TASK_RUNNING;
1627 * fork()/clone()-time setup:
1629 void sched_fork(struct task_struct *p, int clone_flags)
1631 int cpu = get_cpu();
1633 __sched_fork(p);
1635 #ifdef CONFIG_SMP
1636 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1637 #endif
1638 __set_task_cpu(p, cpu);
1641 * Make sure we do not leak PI boosting priority to the child:
1643 p->prio = current->normal_prio;
1645 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1646 if (likely(sched_info_on()))
1647 memset(&p->sched_info, 0, sizeof(p->sched_info));
1648 #endif
1649 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1650 p->oncpu = 0;
1651 #endif
1652 #ifdef CONFIG_PREEMPT
1653 /* Want to start with kernel preemption disabled. */
1654 task_thread_info(p)->preempt_count = 1;
1655 #endif
1656 put_cpu();
1660 * After fork, child runs first. (default) If set to 0 then
1661 * parent will (try to) run first.
1663 unsigned int __read_mostly sysctl_sched_child_runs_first = 1;
1666 * wake_up_new_task - wake up a newly created task for the first time.
1668 * This function will do some initial scheduler statistics housekeeping
1669 * that must be done for every newly created context, then puts the task
1670 * on the runqueue and wakes it.
1672 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1674 unsigned long flags;
1675 struct rq *rq;
1676 int this_cpu;
1678 rq = task_rq_lock(p, &flags);
1679 BUG_ON(p->state != TASK_RUNNING);
1680 this_cpu = smp_processor_id(); /* parent's CPU */
1681 update_rq_clock(rq);
1683 p->prio = effective_prio(p);
1685 if (rt_prio(p->prio))
1686 p->sched_class = &rt_sched_class;
1687 else
1688 p->sched_class = &fair_sched_class;
1690 if (!p->sched_class->task_new || !sysctl_sched_child_runs_first ||
1691 (clone_flags & CLONE_VM) || task_cpu(p) != this_cpu ||
1692 !current->se.on_rq) {
1694 activate_task(rq, p, 0);
1695 } else {
1697 * Let the scheduling class do new task startup
1698 * management (if any):
1700 p->sched_class->task_new(rq, p);
1701 inc_nr_running(p, rq);
1703 check_preempt_curr(rq, p);
1704 task_rq_unlock(rq, &flags);
1707 #ifdef CONFIG_PREEMPT_NOTIFIERS
1710 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1711 * @notifier: notifier struct to register
1713 void preempt_notifier_register(struct preempt_notifier *notifier)
1715 hlist_add_head(&notifier->link, &current->preempt_notifiers);
1717 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1720 * preempt_notifier_unregister - no longer interested in preemption notifications
1721 * @notifier: notifier struct to unregister
1723 * This is safe to call from within a preemption notifier.
1725 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1727 hlist_del(&notifier->link);
1729 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1731 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1733 struct preempt_notifier *notifier;
1734 struct hlist_node *node;
1736 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1737 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1740 static void
1741 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1742 struct task_struct *next)
1744 struct preempt_notifier *notifier;
1745 struct hlist_node *node;
1747 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1748 notifier->ops->sched_out(notifier, next);
1751 #else
1753 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1757 static void
1758 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1759 struct task_struct *next)
1763 #endif
1766 * prepare_task_switch - prepare to switch tasks
1767 * @rq: the runqueue preparing to switch
1768 * @prev: the current task that is being switched out
1769 * @next: the task we are going to switch to.
1771 * This is called with the rq lock held and interrupts off. It must
1772 * be paired with a subsequent finish_task_switch after the context
1773 * switch.
1775 * prepare_task_switch sets up locking and calls architecture specific
1776 * hooks.
1778 static inline void
1779 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1780 struct task_struct *next)
1782 fire_sched_out_preempt_notifiers(prev, next);
1783 prepare_lock_switch(rq, next);
1784 prepare_arch_switch(next);
1788 * finish_task_switch - clean up after a task-switch
1789 * @rq: runqueue associated with task-switch
1790 * @prev: the thread we just switched away from.
1792 * finish_task_switch must be called after the context switch, paired
1793 * with a prepare_task_switch call before the context switch.
1794 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1795 * and do any other architecture-specific cleanup actions.
1797 * Note that we may have delayed dropping an mm in context_switch(). If
1798 * so, we finish that here outside of the runqueue lock. (Doing it
1799 * with the lock held can cause deadlocks; see schedule() for
1800 * details.)
1802 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1803 __releases(rq->lock)
1805 struct mm_struct *mm = rq->prev_mm;
1806 long prev_state;
1808 rq->prev_mm = NULL;
1811 * A task struct has one reference for the use as "current".
1812 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1813 * schedule one last time. The schedule call will never return, and
1814 * the scheduled task must drop that reference.
1815 * The test for TASK_DEAD must occur while the runqueue locks are
1816 * still held, otherwise prev could be scheduled on another cpu, die
1817 * there before we look at prev->state, and then the reference would
1818 * be dropped twice.
1819 * Manfred Spraul <manfred@colorfullife.com>
1821 prev_state = prev->state;
1822 finish_arch_switch(prev);
1823 finish_lock_switch(rq, prev);
1824 fire_sched_in_preempt_notifiers(current);
1825 if (mm)
1826 mmdrop(mm);
1827 if (unlikely(prev_state == TASK_DEAD)) {
1829 * Remove function-return probe instances associated with this
1830 * task and put them back on the free list.
1832 kprobe_flush_task(prev);
1833 put_task_struct(prev);
1838 * schedule_tail - first thing a freshly forked thread must call.
1839 * @prev: the thread we just switched away from.
1841 asmlinkage void schedule_tail(struct task_struct *prev)
1842 __releases(rq->lock)
1844 struct rq *rq = this_rq();
1846 finish_task_switch(rq, prev);
1847 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1848 /* In this case, finish_task_switch does not reenable preemption */
1849 preempt_enable();
1850 #endif
1851 if (current->set_child_tid)
1852 put_user(current->pid, current->set_child_tid);
1856 * context_switch - switch to the new MM and the new
1857 * thread's register state.
1859 static inline void
1860 context_switch(struct rq *rq, struct task_struct *prev,
1861 struct task_struct *next)
1863 struct mm_struct *mm, *oldmm;
1865 prepare_task_switch(rq, prev, next);
1866 mm = next->mm;
1867 oldmm = prev->active_mm;
1869 * For paravirt, this is coupled with an exit in switch_to to
1870 * combine the page table reload and the switch backend into
1871 * one hypercall.
1873 arch_enter_lazy_cpu_mode();
1875 if (unlikely(!mm)) {
1876 next->active_mm = oldmm;
1877 atomic_inc(&oldmm->mm_count);
1878 enter_lazy_tlb(oldmm, next);
1879 } else
1880 switch_mm(oldmm, mm, next);
1882 if (unlikely(!prev->mm)) {
1883 prev->active_mm = NULL;
1884 rq->prev_mm = oldmm;
1887 * Since the runqueue lock will be released by the next
1888 * task (which is an invalid locking op but in the case
1889 * of the scheduler it's an obvious special-case), so we
1890 * do an early lockdep release here:
1892 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1893 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1894 #endif
1896 /* Here we just switch the register state and the stack. */
1897 switch_to(prev, next, prev);
1899 barrier();
1901 * this_rq must be evaluated again because prev may have moved
1902 * CPUs since it called schedule(), thus the 'rq' on its stack
1903 * frame will be invalid.
1905 finish_task_switch(this_rq(), prev);
1909 * nr_running, nr_uninterruptible and nr_context_switches:
1911 * externally visible scheduler statistics: current number of runnable
1912 * threads, current number of uninterruptible-sleeping threads, total
1913 * number of context switches performed since bootup.
1915 unsigned long nr_running(void)
1917 unsigned long i, sum = 0;
1919 for_each_online_cpu(i)
1920 sum += cpu_rq(i)->nr_running;
1922 return sum;
1925 unsigned long nr_uninterruptible(void)
1927 unsigned long i, sum = 0;
1929 for_each_possible_cpu(i)
1930 sum += cpu_rq(i)->nr_uninterruptible;
1933 * Since we read the counters lockless, it might be slightly
1934 * inaccurate. Do not allow it to go below zero though:
1936 if (unlikely((long)sum < 0))
1937 sum = 0;
1939 return sum;
1942 unsigned long long nr_context_switches(void)
1944 int i;
1945 unsigned long long sum = 0;
1947 for_each_possible_cpu(i)
1948 sum += cpu_rq(i)->nr_switches;
1950 return sum;
1953 unsigned long nr_iowait(void)
1955 unsigned long i, sum = 0;
1957 for_each_possible_cpu(i)
1958 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1960 return sum;
1963 unsigned long nr_active(void)
1965 unsigned long i, running = 0, uninterruptible = 0;
1967 for_each_online_cpu(i) {
1968 running += cpu_rq(i)->nr_running;
1969 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1972 if (unlikely((long)uninterruptible < 0))
1973 uninterruptible = 0;
1975 return running + uninterruptible;
1979 * Update rq->cpu_load[] statistics. This function is usually called every
1980 * scheduler tick (TICK_NSEC).
1982 static void update_cpu_load(struct rq *this_rq)
1984 u64 fair_delta64, exec_delta64, idle_delta64, sample_interval64, tmp64;
1985 unsigned long total_load = this_rq->ls.load.weight;
1986 unsigned long this_load = total_load;
1987 struct load_stat *ls = &this_rq->ls;
1988 int i, scale;
1990 this_rq->nr_load_updates++;
1991 if (unlikely(!(sysctl_sched_features & SCHED_FEAT_PRECISE_CPU_LOAD)))
1992 goto do_avg;
1994 /* Update delta_fair/delta_exec fields first */
1995 update_curr_load(this_rq);
1997 fair_delta64 = ls->delta_fair + 1;
1998 ls->delta_fair = 0;
2000 exec_delta64 = ls->delta_exec + 1;
2001 ls->delta_exec = 0;
2003 sample_interval64 = this_rq->clock - ls->load_update_last;
2004 ls->load_update_last = this_rq->clock;
2006 if ((s64)sample_interval64 < (s64)TICK_NSEC)
2007 sample_interval64 = TICK_NSEC;
2009 if (exec_delta64 > sample_interval64)
2010 exec_delta64 = sample_interval64;
2012 idle_delta64 = sample_interval64 - exec_delta64;
2014 tmp64 = div64_64(SCHED_LOAD_SCALE * exec_delta64, fair_delta64);
2015 tmp64 = div64_64(tmp64 * exec_delta64, sample_interval64);
2017 this_load = (unsigned long)tmp64;
2019 do_avg:
2021 /* Update our load: */
2022 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2023 unsigned long old_load, new_load;
2025 /* scale is effectively 1 << i now, and >> i divides by scale */
2027 old_load = this_rq->cpu_load[i];
2028 new_load = this_load;
2030 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2034 #ifdef CONFIG_SMP
2037 * double_rq_lock - safely lock two runqueues
2039 * Note this does not disable interrupts like task_rq_lock,
2040 * you need to do so manually before calling.
2042 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2043 __acquires(rq1->lock)
2044 __acquires(rq2->lock)
2046 BUG_ON(!irqs_disabled());
2047 if (rq1 == rq2) {
2048 spin_lock(&rq1->lock);
2049 __acquire(rq2->lock); /* Fake it out ;) */
2050 } else {
2051 if (rq1 < rq2) {
2052 spin_lock(&rq1->lock);
2053 spin_lock(&rq2->lock);
2054 } else {
2055 spin_lock(&rq2->lock);
2056 spin_lock(&rq1->lock);
2059 update_rq_clock(rq1);
2060 update_rq_clock(rq2);
2064 * double_rq_unlock - safely unlock two runqueues
2066 * Note this does not restore interrupts like task_rq_unlock,
2067 * you need to do so manually after calling.
2069 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2070 __releases(rq1->lock)
2071 __releases(rq2->lock)
2073 spin_unlock(&rq1->lock);
2074 if (rq1 != rq2)
2075 spin_unlock(&rq2->lock);
2076 else
2077 __release(rq2->lock);
2081 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2083 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2084 __releases(this_rq->lock)
2085 __acquires(busiest->lock)
2086 __acquires(this_rq->lock)
2088 if (unlikely(!irqs_disabled())) {
2089 /* printk() doesn't work good under rq->lock */
2090 spin_unlock(&this_rq->lock);
2091 BUG_ON(1);
2093 if (unlikely(!spin_trylock(&busiest->lock))) {
2094 if (busiest < this_rq) {
2095 spin_unlock(&this_rq->lock);
2096 spin_lock(&busiest->lock);
2097 spin_lock(&this_rq->lock);
2098 } else
2099 spin_lock(&busiest->lock);
2104 * If dest_cpu is allowed for this process, migrate the task to it.
2105 * This is accomplished by forcing the cpu_allowed mask to only
2106 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2107 * the cpu_allowed mask is restored.
2109 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2111 struct migration_req req;
2112 unsigned long flags;
2113 struct rq *rq;
2115 rq = task_rq_lock(p, &flags);
2116 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2117 || unlikely(cpu_is_offline(dest_cpu)))
2118 goto out;
2120 /* force the process onto the specified CPU */
2121 if (migrate_task(p, dest_cpu, &req)) {
2122 /* Need to wait for migration thread (might exit: take ref). */
2123 struct task_struct *mt = rq->migration_thread;
2125 get_task_struct(mt);
2126 task_rq_unlock(rq, &flags);
2127 wake_up_process(mt);
2128 put_task_struct(mt);
2129 wait_for_completion(&req.done);
2131 return;
2133 out:
2134 task_rq_unlock(rq, &flags);
2138 * sched_exec - execve() is a valuable balancing opportunity, because at
2139 * this point the task has the smallest effective memory and cache footprint.
2141 void sched_exec(void)
2143 int new_cpu, this_cpu = get_cpu();
2144 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2145 put_cpu();
2146 if (new_cpu != this_cpu)
2147 sched_migrate_task(current, new_cpu);
2151 * pull_task - move a task from a remote runqueue to the local runqueue.
2152 * Both runqueues must be locked.
2154 static void pull_task(struct rq *src_rq, struct task_struct *p,
2155 struct rq *this_rq, int this_cpu)
2157 deactivate_task(src_rq, p, 0);
2158 set_task_cpu(p, this_cpu);
2159 activate_task(this_rq, p, 0);
2161 * Note that idle threads have a prio of MAX_PRIO, for this test
2162 * to be always true for them.
2164 check_preempt_curr(this_rq, p);
2168 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2170 static
2171 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2172 struct sched_domain *sd, enum cpu_idle_type idle,
2173 int *all_pinned)
2176 * We do not migrate tasks that are:
2177 * 1) running (obviously), or
2178 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2179 * 3) are cache-hot on their current CPU.
2181 if (!cpu_isset(this_cpu, p->cpus_allowed))
2182 return 0;
2183 *all_pinned = 0;
2185 if (task_running(rq, p))
2186 return 0;
2188 return 1;
2191 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2192 unsigned long max_nr_move, unsigned long max_load_move,
2193 struct sched_domain *sd, enum cpu_idle_type idle,
2194 int *all_pinned, unsigned long *load_moved,
2195 int *this_best_prio, struct rq_iterator *iterator)
2197 int pulled = 0, pinned = 0, skip_for_load;
2198 struct task_struct *p;
2199 long rem_load_move = max_load_move;
2201 if (max_nr_move == 0 || max_load_move == 0)
2202 goto out;
2204 pinned = 1;
2207 * Start the load-balancing iterator:
2209 p = iterator->start(iterator->arg);
2210 next:
2211 if (!p)
2212 goto out;
2214 * To help distribute high priority tasks accross CPUs we don't
2215 * skip a task if it will be the highest priority task (i.e. smallest
2216 * prio value) on its new queue regardless of its load weight
2218 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2219 SCHED_LOAD_SCALE_FUZZ;
2220 if ((skip_for_load && p->prio >= *this_best_prio) ||
2221 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2222 p = iterator->next(iterator->arg);
2223 goto next;
2226 pull_task(busiest, p, this_rq, this_cpu);
2227 pulled++;
2228 rem_load_move -= p->se.load.weight;
2231 * We only want to steal up to the prescribed number of tasks
2232 * and the prescribed amount of weighted load.
2234 if (pulled < max_nr_move && rem_load_move > 0) {
2235 if (p->prio < *this_best_prio)
2236 *this_best_prio = p->prio;
2237 p = iterator->next(iterator->arg);
2238 goto next;
2240 out:
2242 * Right now, this is the only place pull_task() is called,
2243 * so we can safely collect pull_task() stats here rather than
2244 * inside pull_task().
2246 schedstat_add(sd, lb_gained[idle], pulled);
2248 if (all_pinned)
2249 *all_pinned = pinned;
2250 *load_moved = max_load_move - rem_load_move;
2251 return pulled;
2255 * move_tasks tries to move up to max_load_move weighted load from busiest to
2256 * this_rq, as part of a balancing operation within domain "sd".
2257 * Returns 1 if successful and 0 otherwise.
2259 * Called with both runqueues locked.
2261 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2262 unsigned long max_load_move,
2263 struct sched_domain *sd, enum cpu_idle_type idle,
2264 int *all_pinned)
2266 struct sched_class *class = sched_class_highest;
2267 unsigned long total_load_moved = 0;
2268 int this_best_prio = this_rq->curr->prio;
2270 do {
2271 total_load_moved +=
2272 class->load_balance(this_rq, this_cpu, busiest,
2273 ULONG_MAX, max_load_move - total_load_moved,
2274 sd, idle, all_pinned, &this_best_prio);
2275 class = class->next;
2276 } while (class && max_load_move > total_load_moved);
2278 return total_load_moved > 0;
2282 * move_one_task tries to move exactly one task from busiest to this_rq, as
2283 * part of active balancing operations within "domain".
2284 * Returns 1 if successful and 0 otherwise.
2286 * Called with both runqueues locked.
2288 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2289 struct sched_domain *sd, enum cpu_idle_type idle)
2291 struct sched_class *class;
2292 int this_best_prio = MAX_PRIO;
2294 for (class = sched_class_highest; class; class = class->next)
2295 if (class->load_balance(this_rq, this_cpu, busiest,
2296 1, ULONG_MAX, sd, idle, NULL,
2297 &this_best_prio))
2298 return 1;
2300 return 0;
2304 * find_busiest_group finds and returns the busiest CPU group within the
2305 * domain. It calculates and returns the amount of weighted load which
2306 * should be moved to restore balance via the imbalance parameter.
2308 static struct sched_group *
2309 find_busiest_group(struct sched_domain *sd, int this_cpu,
2310 unsigned long *imbalance, enum cpu_idle_type idle,
2311 int *sd_idle, cpumask_t *cpus, int *balance)
2313 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2314 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2315 unsigned long max_pull;
2316 unsigned long busiest_load_per_task, busiest_nr_running;
2317 unsigned long this_load_per_task, this_nr_running;
2318 int load_idx;
2319 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2320 int power_savings_balance = 1;
2321 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2322 unsigned long min_nr_running = ULONG_MAX;
2323 struct sched_group *group_min = NULL, *group_leader = NULL;
2324 #endif
2326 max_load = this_load = total_load = total_pwr = 0;
2327 busiest_load_per_task = busiest_nr_running = 0;
2328 this_load_per_task = this_nr_running = 0;
2329 if (idle == CPU_NOT_IDLE)
2330 load_idx = sd->busy_idx;
2331 else if (idle == CPU_NEWLY_IDLE)
2332 load_idx = sd->newidle_idx;
2333 else
2334 load_idx = sd->idle_idx;
2336 do {
2337 unsigned long load, group_capacity;
2338 int local_group;
2339 int i;
2340 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2341 unsigned long sum_nr_running, sum_weighted_load;
2343 local_group = cpu_isset(this_cpu, group->cpumask);
2345 if (local_group)
2346 balance_cpu = first_cpu(group->cpumask);
2348 /* Tally up the load of all CPUs in the group */
2349 sum_weighted_load = sum_nr_running = avg_load = 0;
2351 for_each_cpu_mask(i, group->cpumask) {
2352 struct rq *rq;
2354 if (!cpu_isset(i, *cpus))
2355 continue;
2357 rq = cpu_rq(i);
2359 if (*sd_idle && rq->nr_running)
2360 *sd_idle = 0;
2362 /* Bias balancing toward cpus of our domain */
2363 if (local_group) {
2364 if (idle_cpu(i) && !first_idle_cpu) {
2365 first_idle_cpu = 1;
2366 balance_cpu = i;
2369 load = target_load(i, load_idx);
2370 } else
2371 load = source_load(i, load_idx);
2373 avg_load += load;
2374 sum_nr_running += rq->nr_running;
2375 sum_weighted_load += weighted_cpuload(i);
2379 * First idle cpu or the first cpu(busiest) in this sched group
2380 * is eligible for doing load balancing at this and above
2381 * domains. In the newly idle case, we will allow all the cpu's
2382 * to do the newly idle load balance.
2384 if (idle != CPU_NEWLY_IDLE && local_group &&
2385 balance_cpu != this_cpu && balance) {
2386 *balance = 0;
2387 goto ret;
2390 total_load += avg_load;
2391 total_pwr += group->__cpu_power;
2393 /* Adjust by relative CPU power of the group */
2394 avg_load = sg_div_cpu_power(group,
2395 avg_load * SCHED_LOAD_SCALE);
2397 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2399 if (local_group) {
2400 this_load = avg_load;
2401 this = group;
2402 this_nr_running = sum_nr_running;
2403 this_load_per_task = sum_weighted_load;
2404 } else if (avg_load > max_load &&
2405 sum_nr_running > group_capacity) {
2406 max_load = avg_load;
2407 busiest = group;
2408 busiest_nr_running = sum_nr_running;
2409 busiest_load_per_task = sum_weighted_load;
2412 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2414 * Busy processors will not participate in power savings
2415 * balance.
2417 if (idle == CPU_NOT_IDLE ||
2418 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2419 goto group_next;
2422 * If the local group is idle or completely loaded
2423 * no need to do power savings balance at this domain
2425 if (local_group && (this_nr_running >= group_capacity ||
2426 !this_nr_running))
2427 power_savings_balance = 0;
2430 * If a group is already running at full capacity or idle,
2431 * don't include that group in power savings calculations
2433 if (!power_savings_balance || sum_nr_running >= group_capacity
2434 || !sum_nr_running)
2435 goto group_next;
2438 * Calculate the group which has the least non-idle load.
2439 * This is the group from where we need to pick up the load
2440 * for saving power
2442 if ((sum_nr_running < min_nr_running) ||
2443 (sum_nr_running == min_nr_running &&
2444 first_cpu(group->cpumask) <
2445 first_cpu(group_min->cpumask))) {
2446 group_min = group;
2447 min_nr_running = sum_nr_running;
2448 min_load_per_task = sum_weighted_load /
2449 sum_nr_running;
2453 * Calculate the group which is almost near its
2454 * capacity but still has some space to pick up some load
2455 * from other group and save more power
2457 if (sum_nr_running <= group_capacity - 1) {
2458 if (sum_nr_running > leader_nr_running ||
2459 (sum_nr_running == leader_nr_running &&
2460 first_cpu(group->cpumask) >
2461 first_cpu(group_leader->cpumask))) {
2462 group_leader = group;
2463 leader_nr_running = sum_nr_running;
2466 group_next:
2467 #endif
2468 group = group->next;
2469 } while (group != sd->groups);
2471 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2472 goto out_balanced;
2474 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2476 if (this_load >= avg_load ||
2477 100*max_load <= sd->imbalance_pct*this_load)
2478 goto out_balanced;
2480 busiest_load_per_task /= busiest_nr_running;
2482 * We're trying to get all the cpus to the average_load, so we don't
2483 * want to push ourselves above the average load, nor do we wish to
2484 * reduce the max loaded cpu below the average load, as either of these
2485 * actions would just result in more rebalancing later, and ping-pong
2486 * tasks around. Thus we look for the minimum possible imbalance.
2487 * Negative imbalances (*we* are more loaded than anyone else) will
2488 * be counted as no imbalance for these purposes -- we can't fix that
2489 * by pulling tasks to us. Be careful of negative numbers as they'll
2490 * appear as very large values with unsigned longs.
2492 if (max_load <= busiest_load_per_task)
2493 goto out_balanced;
2496 * In the presence of smp nice balancing, certain scenarios can have
2497 * max load less than avg load(as we skip the groups at or below
2498 * its cpu_power, while calculating max_load..)
2500 if (max_load < avg_load) {
2501 *imbalance = 0;
2502 goto small_imbalance;
2505 /* Don't want to pull so many tasks that a group would go idle */
2506 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2508 /* How much load to actually move to equalise the imbalance */
2509 *imbalance = min(max_pull * busiest->__cpu_power,
2510 (avg_load - this_load) * this->__cpu_power)
2511 / SCHED_LOAD_SCALE;
2514 * if *imbalance is less than the average load per runnable task
2515 * there is no gaurantee that any tasks will be moved so we'll have
2516 * a think about bumping its value to force at least one task to be
2517 * moved
2519 if (*imbalance < busiest_load_per_task) {
2520 unsigned long tmp, pwr_now, pwr_move;
2521 unsigned int imbn;
2523 small_imbalance:
2524 pwr_move = pwr_now = 0;
2525 imbn = 2;
2526 if (this_nr_running) {
2527 this_load_per_task /= this_nr_running;
2528 if (busiest_load_per_task > this_load_per_task)
2529 imbn = 1;
2530 } else
2531 this_load_per_task = SCHED_LOAD_SCALE;
2533 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2534 busiest_load_per_task * imbn) {
2535 *imbalance = busiest_load_per_task;
2536 return busiest;
2540 * OK, we don't have enough imbalance to justify moving tasks,
2541 * however we may be able to increase total CPU power used by
2542 * moving them.
2545 pwr_now += busiest->__cpu_power *
2546 min(busiest_load_per_task, max_load);
2547 pwr_now += this->__cpu_power *
2548 min(this_load_per_task, this_load);
2549 pwr_now /= SCHED_LOAD_SCALE;
2551 /* Amount of load we'd subtract */
2552 tmp = sg_div_cpu_power(busiest,
2553 busiest_load_per_task * SCHED_LOAD_SCALE);
2554 if (max_load > tmp)
2555 pwr_move += busiest->__cpu_power *
2556 min(busiest_load_per_task, max_load - tmp);
2558 /* Amount of load we'd add */
2559 if (max_load * busiest->__cpu_power <
2560 busiest_load_per_task * SCHED_LOAD_SCALE)
2561 tmp = sg_div_cpu_power(this,
2562 max_load * busiest->__cpu_power);
2563 else
2564 tmp = sg_div_cpu_power(this,
2565 busiest_load_per_task * SCHED_LOAD_SCALE);
2566 pwr_move += this->__cpu_power *
2567 min(this_load_per_task, this_load + tmp);
2568 pwr_move /= SCHED_LOAD_SCALE;
2570 /* Move if we gain throughput */
2571 if (pwr_move > pwr_now)
2572 *imbalance = busiest_load_per_task;
2575 return busiest;
2577 out_balanced:
2578 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2579 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2580 goto ret;
2582 if (this == group_leader && group_leader != group_min) {
2583 *imbalance = min_load_per_task;
2584 return group_min;
2586 #endif
2587 ret:
2588 *imbalance = 0;
2589 return NULL;
2593 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2595 static struct rq *
2596 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2597 unsigned long imbalance, cpumask_t *cpus)
2599 struct rq *busiest = NULL, *rq;
2600 unsigned long max_load = 0;
2601 int i;
2603 for_each_cpu_mask(i, group->cpumask) {
2604 unsigned long wl;
2606 if (!cpu_isset(i, *cpus))
2607 continue;
2609 rq = cpu_rq(i);
2610 wl = weighted_cpuload(i);
2612 if (rq->nr_running == 1 && wl > imbalance)
2613 continue;
2615 if (wl > max_load) {
2616 max_load = wl;
2617 busiest = rq;
2621 return busiest;
2625 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2626 * so long as it is large enough.
2628 #define MAX_PINNED_INTERVAL 512
2631 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2632 * tasks if there is an imbalance.
2634 static int load_balance(int this_cpu, struct rq *this_rq,
2635 struct sched_domain *sd, enum cpu_idle_type idle,
2636 int *balance)
2638 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2639 struct sched_group *group;
2640 unsigned long imbalance;
2641 struct rq *busiest;
2642 cpumask_t cpus = CPU_MASK_ALL;
2643 unsigned long flags;
2646 * When power savings policy is enabled for the parent domain, idle
2647 * sibling can pick up load irrespective of busy siblings. In this case,
2648 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2649 * portraying it as CPU_NOT_IDLE.
2651 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2652 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2653 sd_idle = 1;
2655 schedstat_inc(sd, lb_cnt[idle]);
2657 redo:
2658 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2659 &cpus, balance);
2661 if (*balance == 0)
2662 goto out_balanced;
2664 if (!group) {
2665 schedstat_inc(sd, lb_nobusyg[idle]);
2666 goto out_balanced;
2669 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2670 if (!busiest) {
2671 schedstat_inc(sd, lb_nobusyq[idle]);
2672 goto out_balanced;
2675 BUG_ON(busiest == this_rq);
2677 schedstat_add(sd, lb_imbalance[idle], imbalance);
2679 ld_moved = 0;
2680 if (busiest->nr_running > 1) {
2682 * Attempt to move tasks. If find_busiest_group has found
2683 * an imbalance but busiest->nr_running <= 1, the group is
2684 * still unbalanced. ld_moved simply stays zero, so it is
2685 * correctly treated as an imbalance.
2687 local_irq_save(flags);
2688 double_rq_lock(this_rq, busiest);
2689 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2690 imbalance, sd, idle, &all_pinned);
2691 double_rq_unlock(this_rq, busiest);
2692 local_irq_restore(flags);
2695 * some other cpu did the load balance for us.
2697 if (ld_moved && this_cpu != smp_processor_id())
2698 resched_cpu(this_cpu);
2700 /* All tasks on this runqueue were pinned by CPU affinity */
2701 if (unlikely(all_pinned)) {
2702 cpu_clear(cpu_of(busiest), cpus);
2703 if (!cpus_empty(cpus))
2704 goto redo;
2705 goto out_balanced;
2709 if (!ld_moved) {
2710 schedstat_inc(sd, lb_failed[idle]);
2711 sd->nr_balance_failed++;
2713 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2715 spin_lock_irqsave(&busiest->lock, flags);
2717 /* don't kick the migration_thread, if the curr
2718 * task on busiest cpu can't be moved to this_cpu
2720 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2721 spin_unlock_irqrestore(&busiest->lock, flags);
2722 all_pinned = 1;
2723 goto out_one_pinned;
2726 if (!busiest->active_balance) {
2727 busiest->active_balance = 1;
2728 busiest->push_cpu = this_cpu;
2729 active_balance = 1;
2731 spin_unlock_irqrestore(&busiest->lock, flags);
2732 if (active_balance)
2733 wake_up_process(busiest->migration_thread);
2736 * We've kicked active balancing, reset the failure
2737 * counter.
2739 sd->nr_balance_failed = sd->cache_nice_tries+1;
2741 } else
2742 sd->nr_balance_failed = 0;
2744 if (likely(!active_balance)) {
2745 /* We were unbalanced, so reset the balancing interval */
2746 sd->balance_interval = sd->min_interval;
2747 } else {
2749 * If we've begun active balancing, start to back off. This
2750 * case may not be covered by the all_pinned logic if there
2751 * is only 1 task on the busy runqueue (because we don't call
2752 * move_tasks).
2754 if (sd->balance_interval < sd->max_interval)
2755 sd->balance_interval *= 2;
2758 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2759 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2760 return -1;
2761 return ld_moved;
2763 out_balanced:
2764 schedstat_inc(sd, lb_balanced[idle]);
2766 sd->nr_balance_failed = 0;
2768 out_one_pinned:
2769 /* tune up the balancing interval */
2770 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2771 (sd->balance_interval < sd->max_interval))
2772 sd->balance_interval *= 2;
2774 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2775 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2776 return -1;
2777 return 0;
2781 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2782 * tasks if there is an imbalance.
2784 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2785 * this_rq is locked.
2787 static int
2788 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2790 struct sched_group *group;
2791 struct rq *busiest = NULL;
2792 unsigned long imbalance;
2793 int ld_moved = 0;
2794 int sd_idle = 0;
2795 int all_pinned = 0;
2796 cpumask_t cpus = CPU_MASK_ALL;
2799 * When power savings policy is enabled for the parent domain, idle
2800 * sibling can pick up load irrespective of busy siblings. In this case,
2801 * let the state of idle sibling percolate up as IDLE, instead of
2802 * portraying it as CPU_NOT_IDLE.
2804 if (sd->flags & SD_SHARE_CPUPOWER &&
2805 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2806 sd_idle = 1;
2808 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2809 redo:
2810 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2811 &sd_idle, &cpus, NULL);
2812 if (!group) {
2813 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2814 goto out_balanced;
2817 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2818 &cpus);
2819 if (!busiest) {
2820 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2821 goto out_balanced;
2824 BUG_ON(busiest == this_rq);
2826 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2828 ld_moved = 0;
2829 if (busiest->nr_running > 1) {
2830 /* Attempt to move tasks */
2831 double_lock_balance(this_rq, busiest);
2832 /* this_rq->clock is already updated */
2833 update_rq_clock(busiest);
2834 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2835 imbalance, sd, CPU_NEWLY_IDLE,
2836 &all_pinned);
2837 spin_unlock(&busiest->lock);
2839 if (unlikely(all_pinned)) {
2840 cpu_clear(cpu_of(busiest), cpus);
2841 if (!cpus_empty(cpus))
2842 goto redo;
2846 if (!ld_moved) {
2847 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2848 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2849 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2850 return -1;
2851 } else
2852 sd->nr_balance_failed = 0;
2854 return ld_moved;
2856 out_balanced:
2857 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2858 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2859 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2860 return -1;
2861 sd->nr_balance_failed = 0;
2863 return 0;
2867 * idle_balance is called by schedule() if this_cpu is about to become
2868 * idle. Attempts to pull tasks from other CPUs.
2870 static void idle_balance(int this_cpu, struct rq *this_rq)
2872 struct sched_domain *sd;
2873 int pulled_task = -1;
2874 unsigned long next_balance = jiffies + HZ;
2876 for_each_domain(this_cpu, sd) {
2877 unsigned long interval;
2879 if (!(sd->flags & SD_LOAD_BALANCE))
2880 continue;
2882 if (sd->flags & SD_BALANCE_NEWIDLE)
2883 /* If we've pulled tasks over stop searching: */
2884 pulled_task = load_balance_newidle(this_cpu,
2885 this_rq, sd);
2887 interval = msecs_to_jiffies(sd->balance_interval);
2888 if (time_after(next_balance, sd->last_balance + interval))
2889 next_balance = sd->last_balance + interval;
2890 if (pulled_task)
2891 break;
2893 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2895 * We are going idle. next_balance may be set based on
2896 * a busy processor. So reset next_balance.
2898 this_rq->next_balance = next_balance;
2903 * active_load_balance is run by migration threads. It pushes running tasks
2904 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2905 * running on each physical CPU where possible, and avoids physical /
2906 * logical imbalances.
2908 * Called with busiest_rq locked.
2910 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2912 int target_cpu = busiest_rq->push_cpu;
2913 struct sched_domain *sd;
2914 struct rq *target_rq;
2916 /* Is there any task to move? */
2917 if (busiest_rq->nr_running <= 1)
2918 return;
2920 target_rq = cpu_rq(target_cpu);
2923 * This condition is "impossible", if it occurs
2924 * we need to fix it. Originally reported by
2925 * Bjorn Helgaas on a 128-cpu setup.
2927 BUG_ON(busiest_rq == target_rq);
2929 /* move a task from busiest_rq to target_rq */
2930 double_lock_balance(busiest_rq, target_rq);
2931 update_rq_clock(busiest_rq);
2932 update_rq_clock(target_rq);
2934 /* Search for an sd spanning us and the target CPU. */
2935 for_each_domain(target_cpu, sd) {
2936 if ((sd->flags & SD_LOAD_BALANCE) &&
2937 cpu_isset(busiest_cpu, sd->span))
2938 break;
2941 if (likely(sd)) {
2942 schedstat_inc(sd, alb_cnt);
2944 if (move_one_task(target_rq, target_cpu, busiest_rq,
2945 sd, CPU_IDLE))
2946 schedstat_inc(sd, alb_pushed);
2947 else
2948 schedstat_inc(sd, alb_failed);
2950 spin_unlock(&target_rq->lock);
2953 #ifdef CONFIG_NO_HZ
2954 static struct {
2955 atomic_t load_balancer;
2956 cpumask_t cpu_mask;
2957 } nohz ____cacheline_aligned = {
2958 .load_balancer = ATOMIC_INIT(-1),
2959 .cpu_mask = CPU_MASK_NONE,
2963 * This routine will try to nominate the ilb (idle load balancing)
2964 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2965 * load balancing on behalf of all those cpus. If all the cpus in the system
2966 * go into this tickless mode, then there will be no ilb owner (as there is
2967 * no need for one) and all the cpus will sleep till the next wakeup event
2968 * arrives...
2970 * For the ilb owner, tick is not stopped. And this tick will be used
2971 * for idle load balancing. ilb owner will still be part of
2972 * nohz.cpu_mask..
2974 * While stopping the tick, this cpu will become the ilb owner if there
2975 * is no other owner. And will be the owner till that cpu becomes busy
2976 * or if all cpus in the system stop their ticks at which point
2977 * there is no need for ilb owner.
2979 * When the ilb owner becomes busy, it nominates another owner, during the
2980 * next busy scheduler_tick()
2982 int select_nohz_load_balancer(int stop_tick)
2984 int cpu = smp_processor_id();
2986 if (stop_tick) {
2987 cpu_set(cpu, nohz.cpu_mask);
2988 cpu_rq(cpu)->in_nohz_recently = 1;
2991 * If we are going offline and still the leader, give up!
2993 if (cpu_is_offline(cpu) &&
2994 atomic_read(&nohz.load_balancer) == cpu) {
2995 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2996 BUG();
2997 return 0;
3000 /* time for ilb owner also to sleep */
3001 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3002 if (atomic_read(&nohz.load_balancer) == cpu)
3003 atomic_set(&nohz.load_balancer, -1);
3004 return 0;
3007 if (atomic_read(&nohz.load_balancer) == -1) {
3008 /* make me the ilb owner */
3009 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3010 return 1;
3011 } else if (atomic_read(&nohz.load_balancer) == cpu)
3012 return 1;
3013 } else {
3014 if (!cpu_isset(cpu, nohz.cpu_mask))
3015 return 0;
3017 cpu_clear(cpu, nohz.cpu_mask);
3019 if (atomic_read(&nohz.load_balancer) == cpu)
3020 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3021 BUG();
3023 return 0;
3025 #endif
3027 static DEFINE_SPINLOCK(balancing);
3030 * It checks each scheduling domain to see if it is due to be balanced,
3031 * and initiates a balancing operation if so.
3033 * Balancing parameters are set up in arch_init_sched_domains.
3035 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3037 int balance = 1;
3038 struct rq *rq = cpu_rq(cpu);
3039 unsigned long interval;
3040 struct sched_domain *sd;
3041 /* Earliest time when we have to do rebalance again */
3042 unsigned long next_balance = jiffies + 60*HZ;
3043 int update_next_balance = 0;
3045 for_each_domain(cpu, sd) {
3046 if (!(sd->flags & SD_LOAD_BALANCE))
3047 continue;
3049 interval = sd->balance_interval;
3050 if (idle != CPU_IDLE)
3051 interval *= sd->busy_factor;
3053 /* scale ms to jiffies */
3054 interval = msecs_to_jiffies(interval);
3055 if (unlikely(!interval))
3056 interval = 1;
3057 if (interval > HZ*NR_CPUS/10)
3058 interval = HZ*NR_CPUS/10;
3061 if (sd->flags & SD_SERIALIZE) {
3062 if (!spin_trylock(&balancing))
3063 goto out;
3066 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3067 if (load_balance(cpu, rq, sd, idle, &balance)) {
3069 * We've pulled tasks over so either we're no
3070 * longer idle, or one of our SMT siblings is
3071 * not idle.
3073 idle = CPU_NOT_IDLE;
3075 sd->last_balance = jiffies;
3077 if (sd->flags & SD_SERIALIZE)
3078 spin_unlock(&balancing);
3079 out:
3080 if (time_after(next_balance, sd->last_balance + interval)) {
3081 next_balance = sd->last_balance + interval;
3082 update_next_balance = 1;
3086 * Stop the load balance at this level. There is another
3087 * CPU in our sched group which is doing load balancing more
3088 * actively.
3090 if (!balance)
3091 break;
3095 * next_balance will be updated only when there is a need.
3096 * When the cpu is attached to null domain for ex, it will not be
3097 * updated.
3099 if (likely(update_next_balance))
3100 rq->next_balance = next_balance;
3104 * run_rebalance_domains is triggered when needed from the scheduler tick.
3105 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3106 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3108 static void run_rebalance_domains(struct softirq_action *h)
3110 int this_cpu = smp_processor_id();
3111 struct rq *this_rq = cpu_rq(this_cpu);
3112 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3113 CPU_IDLE : CPU_NOT_IDLE;
3115 rebalance_domains(this_cpu, idle);
3117 #ifdef CONFIG_NO_HZ
3119 * If this cpu is the owner for idle load balancing, then do the
3120 * balancing on behalf of the other idle cpus whose ticks are
3121 * stopped.
3123 if (this_rq->idle_at_tick &&
3124 atomic_read(&nohz.load_balancer) == this_cpu) {
3125 cpumask_t cpus = nohz.cpu_mask;
3126 struct rq *rq;
3127 int balance_cpu;
3129 cpu_clear(this_cpu, cpus);
3130 for_each_cpu_mask(balance_cpu, cpus) {
3132 * If this cpu gets work to do, stop the load balancing
3133 * work being done for other cpus. Next load
3134 * balancing owner will pick it up.
3136 if (need_resched())
3137 break;
3139 rebalance_domains(balance_cpu, CPU_IDLE);
3141 rq = cpu_rq(balance_cpu);
3142 if (time_after(this_rq->next_balance, rq->next_balance))
3143 this_rq->next_balance = rq->next_balance;
3146 #endif
3150 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3152 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3153 * idle load balancing owner or decide to stop the periodic load balancing,
3154 * if the whole system is idle.
3156 static inline void trigger_load_balance(struct rq *rq, int cpu)
3158 #ifdef CONFIG_NO_HZ
3160 * If we were in the nohz mode recently and busy at the current
3161 * scheduler tick, then check if we need to nominate new idle
3162 * load balancer.
3164 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3165 rq->in_nohz_recently = 0;
3167 if (atomic_read(&nohz.load_balancer) == cpu) {
3168 cpu_clear(cpu, nohz.cpu_mask);
3169 atomic_set(&nohz.load_balancer, -1);
3172 if (atomic_read(&nohz.load_balancer) == -1) {
3174 * simple selection for now: Nominate the
3175 * first cpu in the nohz list to be the next
3176 * ilb owner.
3178 * TBD: Traverse the sched domains and nominate
3179 * the nearest cpu in the nohz.cpu_mask.
3181 int ilb = first_cpu(nohz.cpu_mask);
3183 if (ilb != NR_CPUS)
3184 resched_cpu(ilb);
3189 * If this cpu is idle and doing idle load balancing for all the
3190 * cpus with ticks stopped, is it time for that to stop?
3192 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3193 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3194 resched_cpu(cpu);
3195 return;
3199 * If this cpu is idle and the idle load balancing is done by
3200 * someone else, then no need raise the SCHED_SOFTIRQ
3202 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3203 cpu_isset(cpu, nohz.cpu_mask))
3204 return;
3205 #endif
3206 if (time_after_eq(jiffies, rq->next_balance))
3207 raise_softirq(SCHED_SOFTIRQ);
3210 #else /* CONFIG_SMP */
3213 * on UP we do not need to balance between CPUs:
3215 static inline void idle_balance(int cpu, struct rq *rq)
3219 /* Avoid "used but not defined" warning on UP */
3220 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3221 unsigned long max_nr_move, unsigned long max_load_move,
3222 struct sched_domain *sd, enum cpu_idle_type idle,
3223 int *all_pinned, unsigned long *load_moved,
3224 int *this_best_prio, struct rq_iterator *iterator)
3226 *load_moved = 0;
3228 return 0;
3231 #endif
3233 DEFINE_PER_CPU(struct kernel_stat, kstat);
3235 EXPORT_PER_CPU_SYMBOL(kstat);
3238 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3239 * that have not yet been banked in case the task is currently running.
3241 unsigned long long task_sched_runtime(struct task_struct *p)
3243 unsigned long flags;
3244 u64 ns, delta_exec;
3245 struct rq *rq;
3247 rq = task_rq_lock(p, &flags);
3248 ns = p->se.sum_exec_runtime;
3249 if (rq->curr == p) {
3250 update_rq_clock(rq);
3251 delta_exec = rq->clock - p->se.exec_start;
3252 if ((s64)delta_exec > 0)
3253 ns += delta_exec;
3255 task_rq_unlock(rq, &flags);
3257 return ns;
3261 * Account user cpu time to a process.
3262 * @p: the process that the cpu time gets accounted to
3263 * @hardirq_offset: the offset to subtract from hardirq_count()
3264 * @cputime: the cpu time spent in user space since the last update
3266 void account_user_time(struct task_struct *p, cputime_t cputime)
3268 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3269 cputime64_t tmp;
3271 p->utime = cputime_add(p->utime, cputime);
3273 /* Add user time to cpustat. */
3274 tmp = cputime_to_cputime64(cputime);
3275 if (TASK_NICE(p) > 0)
3276 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3277 else
3278 cpustat->user = cputime64_add(cpustat->user, tmp);
3282 * Account system cpu time to a process.
3283 * @p: the process that the cpu time gets accounted to
3284 * @hardirq_offset: the offset to subtract from hardirq_count()
3285 * @cputime: the cpu time spent in kernel space since the last update
3287 void account_system_time(struct task_struct *p, int hardirq_offset,
3288 cputime_t cputime)
3290 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3291 struct rq *rq = this_rq();
3292 cputime64_t tmp;
3294 p->stime = cputime_add(p->stime, cputime);
3296 /* Add system time to cpustat. */
3297 tmp = cputime_to_cputime64(cputime);
3298 if (hardirq_count() - hardirq_offset)
3299 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3300 else if (softirq_count())
3301 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3302 else if (p != rq->idle)
3303 cpustat->system = cputime64_add(cpustat->system, tmp);
3304 else if (atomic_read(&rq->nr_iowait) > 0)
3305 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3306 else
3307 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3308 /* Account for system time used */
3309 acct_update_integrals(p);
3313 * Account for involuntary wait time.
3314 * @p: the process from which the cpu time has been stolen
3315 * @steal: the cpu time spent in involuntary wait
3317 void account_steal_time(struct task_struct *p, cputime_t steal)
3319 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3320 cputime64_t tmp = cputime_to_cputime64(steal);
3321 struct rq *rq = this_rq();
3323 if (p == rq->idle) {
3324 p->stime = cputime_add(p->stime, steal);
3325 if (atomic_read(&rq->nr_iowait) > 0)
3326 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3327 else
3328 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3329 } else
3330 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3334 * This function gets called by the timer code, with HZ frequency.
3335 * We call it with interrupts disabled.
3337 * It also gets called by the fork code, when changing the parent's
3338 * timeslices.
3340 void scheduler_tick(void)
3342 int cpu = smp_processor_id();
3343 struct rq *rq = cpu_rq(cpu);
3344 struct task_struct *curr = rq->curr;
3345 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3347 spin_lock(&rq->lock);
3348 __update_rq_clock(rq);
3350 * Let rq->clock advance by at least TICK_NSEC:
3352 if (unlikely(rq->clock < next_tick))
3353 rq->clock = next_tick;
3354 rq->tick_timestamp = rq->clock;
3355 update_cpu_load(rq);
3356 if (curr != rq->idle) /* FIXME: needed? */
3357 curr->sched_class->task_tick(rq, curr);
3358 spin_unlock(&rq->lock);
3360 #ifdef CONFIG_SMP
3361 rq->idle_at_tick = idle_cpu(cpu);
3362 trigger_load_balance(rq, cpu);
3363 #endif
3366 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3368 void fastcall add_preempt_count(int val)
3371 * Underflow?
3373 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3374 return;
3375 preempt_count() += val;
3377 * Spinlock count overflowing soon?
3379 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3380 PREEMPT_MASK - 10);
3382 EXPORT_SYMBOL(add_preempt_count);
3384 void fastcall sub_preempt_count(int val)
3387 * Underflow?
3389 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3390 return;
3392 * Is the spinlock portion underflowing?
3394 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3395 !(preempt_count() & PREEMPT_MASK)))
3396 return;
3398 preempt_count() -= val;
3400 EXPORT_SYMBOL(sub_preempt_count);
3402 #endif
3405 * Print scheduling while atomic bug:
3407 static noinline void __schedule_bug(struct task_struct *prev)
3409 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3410 prev->comm, preempt_count(), prev->pid);
3411 debug_show_held_locks(prev);
3412 if (irqs_disabled())
3413 print_irqtrace_events(prev);
3414 dump_stack();
3418 * Various schedule()-time debugging checks and statistics:
3420 static inline void schedule_debug(struct task_struct *prev)
3423 * Test if we are atomic. Since do_exit() needs to call into
3424 * schedule() atomically, we ignore that path for now.
3425 * Otherwise, whine if we are scheduling when we should not be.
3427 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3428 __schedule_bug(prev);
3430 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3432 schedstat_inc(this_rq(), sched_cnt);
3436 * Pick up the highest-prio task:
3438 static inline struct task_struct *
3439 pick_next_task(struct rq *rq, struct task_struct *prev)
3441 struct sched_class *class;
3442 struct task_struct *p;
3445 * Optimization: we know that if all tasks are in
3446 * the fair class we can call that function directly:
3448 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3449 p = fair_sched_class.pick_next_task(rq);
3450 if (likely(p))
3451 return p;
3454 class = sched_class_highest;
3455 for ( ; ; ) {
3456 p = class->pick_next_task(rq);
3457 if (p)
3458 return p;
3460 * Will never be NULL as the idle class always
3461 * returns a non-NULL p:
3463 class = class->next;
3468 * schedule() is the main scheduler function.
3470 asmlinkage void __sched schedule(void)
3472 struct task_struct *prev, *next;
3473 long *switch_count;
3474 struct rq *rq;
3475 int cpu;
3477 need_resched:
3478 preempt_disable();
3479 cpu = smp_processor_id();
3480 rq = cpu_rq(cpu);
3481 rcu_qsctr_inc(cpu);
3482 prev = rq->curr;
3483 switch_count = &prev->nivcsw;
3485 release_kernel_lock(prev);
3486 need_resched_nonpreemptible:
3488 schedule_debug(prev);
3490 spin_lock_irq(&rq->lock);
3491 clear_tsk_need_resched(prev);
3492 __update_rq_clock(rq);
3494 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3495 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3496 unlikely(signal_pending(prev)))) {
3497 prev->state = TASK_RUNNING;
3498 } else {
3499 deactivate_task(rq, prev, 1);
3501 switch_count = &prev->nvcsw;
3504 if (unlikely(!rq->nr_running))
3505 idle_balance(cpu, rq);
3507 prev->sched_class->put_prev_task(rq, prev);
3508 next = pick_next_task(rq, prev);
3510 sched_info_switch(prev, next);
3512 if (likely(prev != next)) {
3513 rq->nr_switches++;
3514 rq->curr = next;
3515 ++*switch_count;
3517 context_switch(rq, prev, next); /* unlocks the rq */
3518 } else
3519 spin_unlock_irq(&rq->lock);
3521 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3522 cpu = smp_processor_id();
3523 rq = cpu_rq(cpu);
3524 goto need_resched_nonpreemptible;
3526 preempt_enable_no_resched();
3527 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3528 goto need_resched;
3530 EXPORT_SYMBOL(schedule);
3532 #ifdef CONFIG_PREEMPT
3534 * this is the entry point to schedule() from in-kernel preemption
3535 * off of preempt_enable. Kernel preemptions off return from interrupt
3536 * occur there and call schedule directly.
3538 asmlinkage void __sched preempt_schedule(void)
3540 struct thread_info *ti = current_thread_info();
3541 #ifdef CONFIG_PREEMPT_BKL
3542 struct task_struct *task = current;
3543 int saved_lock_depth;
3544 #endif
3546 * If there is a non-zero preempt_count or interrupts are disabled,
3547 * we do not want to preempt the current task. Just return..
3549 if (likely(ti->preempt_count || irqs_disabled()))
3550 return;
3552 need_resched:
3553 add_preempt_count(PREEMPT_ACTIVE);
3555 * We keep the big kernel semaphore locked, but we
3556 * clear ->lock_depth so that schedule() doesnt
3557 * auto-release the semaphore:
3559 #ifdef CONFIG_PREEMPT_BKL
3560 saved_lock_depth = task->lock_depth;
3561 task->lock_depth = -1;
3562 #endif
3563 schedule();
3564 #ifdef CONFIG_PREEMPT_BKL
3565 task->lock_depth = saved_lock_depth;
3566 #endif
3567 sub_preempt_count(PREEMPT_ACTIVE);
3569 /* we could miss a preemption opportunity between schedule and now */
3570 barrier();
3571 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3572 goto need_resched;
3574 EXPORT_SYMBOL(preempt_schedule);
3577 * this is the entry point to schedule() from kernel preemption
3578 * off of irq context.
3579 * Note, that this is called and return with irqs disabled. This will
3580 * protect us against recursive calling from irq.
3582 asmlinkage void __sched preempt_schedule_irq(void)
3584 struct thread_info *ti = current_thread_info();
3585 #ifdef CONFIG_PREEMPT_BKL
3586 struct task_struct *task = current;
3587 int saved_lock_depth;
3588 #endif
3589 /* Catch callers which need to be fixed */
3590 BUG_ON(ti->preempt_count || !irqs_disabled());
3592 need_resched:
3593 add_preempt_count(PREEMPT_ACTIVE);
3595 * We keep the big kernel semaphore locked, but we
3596 * clear ->lock_depth so that schedule() doesnt
3597 * auto-release the semaphore:
3599 #ifdef CONFIG_PREEMPT_BKL
3600 saved_lock_depth = task->lock_depth;
3601 task->lock_depth = -1;
3602 #endif
3603 local_irq_enable();
3604 schedule();
3605 local_irq_disable();
3606 #ifdef CONFIG_PREEMPT_BKL
3607 task->lock_depth = saved_lock_depth;
3608 #endif
3609 sub_preempt_count(PREEMPT_ACTIVE);
3611 /* we could miss a preemption opportunity between schedule and now */
3612 barrier();
3613 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3614 goto need_resched;
3617 #endif /* CONFIG_PREEMPT */
3619 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3620 void *key)
3622 return try_to_wake_up(curr->private, mode, sync);
3624 EXPORT_SYMBOL(default_wake_function);
3627 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3628 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3629 * number) then we wake all the non-exclusive tasks and one exclusive task.
3631 * There are circumstances in which we can try to wake a task which has already
3632 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3633 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3635 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3636 int nr_exclusive, int sync, void *key)
3638 struct list_head *tmp, *next;
3640 list_for_each_safe(tmp, next, &q->task_list) {
3641 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3642 unsigned flags = curr->flags;
3644 if (curr->func(curr, mode, sync, key) &&
3645 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3646 break;
3651 * __wake_up - wake up threads blocked on a waitqueue.
3652 * @q: the waitqueue
3653 * @mode: which threads
3654 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3655 * @key: is directly passed to the wakeup function
3657 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3658 int nr_exclusive, void *key)
3660 unsigned long flags;
3662 spin_lock_irqsave(&q->lock, flags);
3663 __wake_up_common(q, mode, nr_exclusive, 0, key);
3664 spin_unlock_irqrestore(&q->lock, flags);
3666 EXPORT_SYMBOL(__wake_up);
3669 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3671 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3673 __wake_up_common(q, mode, 1, 0, NULL);
3677 * __wake_up_sync - wake up threads blocked on a waitqueue.
3678 * @q: the waitqueue
3679 * @mode: which threads
3680 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3682 * The sync wakeup differs that the waker knows that it will schedule
3683 * away soon, so while the target thread will be woken up, it will not
3684 * be migrated to another CPU - ie. the two threads are 'synchronized'
3685 * with each other. This can prevent needless bouncing between CPUs.
3687 * On UP it can prevent extra preemption.
3689 void fastcall
3690 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3692 unsigned long flags;
3693 int sync = 1;
3695 if (unlikely(!q))
3696 return;
3698 if (unlikely(!nr_exclusive))
3699 sync = 0;
3701 spin_lock_irqsave(&q->lock, flags);
3702 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3703 spin_unlock_irqrestore(&q->lock, flags);
3705 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3707 void fastcall complete(struct completion *x)
3709 unsigned long flags;
3711 spin_lock_irqsave(&x->wait.lock, flags);
3712 x->done++;
3713 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3714 1, 0, NULL);
3715 spin_unlock_irqrestore(&x->wait.lock, flags);
3717 EXPORT_SYMBOL(complete);
3719 void fastcall complete_all(struct completion *x)
3721 unsigned long flags;
3723 spin_lock_irqsave(&x->wait.lock, flags);
3724 x->done += UINT_MAX/2;
3725 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3726 0, 0, NULL);
3727 spin_unlock_irqrestore(&x->wait.lock, flags);
3729 EXPORT_SYMBOL(complete_all);
3731 void fastcall __sched wait_for_completion(struct completion *x)
3733 might_sleep();
3735 spin_lock_irq(&x->wait.lock);
3736 if (!x->done) {
3737 DECLARE_WAITQUEUE(wait, current);
3739 wait.flags |= WQ_FLAG_EXCLUSIVE;
3740 __add_wait_queue_tail(&x->wait, &wait);
3741 do {
3742 __set_current_state(TASK_UNINTERRUPTIBLE);
3743 spin_unlock_irq(&x->wait.lock);
3744 schedule();
3745 spin_lock_irq(&x->wait.lock);
3746 } while (!x->done);
3747 __remove_wait_queue(&x->wait, &wait);
3749 x->done--;
3750 spin_unlock_irq(&x->wait.lock);
3752 EXPORT_SYMBOL(wait_for_completion);
3754 unsigned long fastcall __sched
3755 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3757 might_sleep();
3759 spin_lock_irq(&x->wait.lock);
3760 if (!x->done) {
3761 DECLARE_WAITQUEUE(wait, current);
3763 wait.flags |= WQ_FLAG_EXCLUSIVE;
3764 __add_wait_queue_tail(&x->wait, &wait);
3765 do {
3766 __set_current_state(TASK_UNINTERRUPTIBLE);
3767 spin_unlock_irq(&x->wait.lock);
3768 timeout = schedule_timeout(timeout);
3769 spin_lock_irq(&x->wait.lock);
3770 if (!timeout) {
3771 __remove_wait_queue(&x->wait, &wait);
3772 goto out;
3774 } while (!x->done);
3775 __remove_wait_queue(&x->wait, &wait);
3777 x->done--;
3778 out:
3779 spin_unlock_irq(&x->wait.lock);
3780 return timeout;
3782 EXPORT_SYMBOL(wait_for_completion_timeout);
3784 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3786 int ret = 0;
3788 might_sleep();
3790 spin_lock_irq(&x->wait.lock);
3791 if (!x->done) {
3792 DECLARE_WAITQUEUE(wait, current);
3794 wait.flags |= WQ_FLAG_EXCLUSIVE;
3795 __add_wait_queue_tail(&x->wait, &wait);
3796 do {
3797 if (signal_pending(current)) {
3798 ret = -ERESTARTSYS;
3799 __remove_wait_queue(&x->wait, &wait);
3800 goto out;
3802 __set_current_state(TASK_INTERRUPTIBLE);
3803 spin_unlock_irq(&x->wait.lock);
3804 schedule();
3805 spin_lock_irq(&x->wait.lock);
3806 } while (!x->done);
3807 __remove_wait_queue(&x->wait, &wait);
3809 x->done--;
3810 out:
3811 spin_unlock_irq(&x->wait.lock);
3813 return ret;
3815 EXPORT_SYMBOL(wait_for_completion_interruptible);
3817 unsigned long fastcall __sched
3818 wait_for_completion_interruptible_timeout(struct completion *x,
3819 unsigned long timeout)
3821 might_sleep();
3823 spin_lock_irq(&x->wait.lock);
3824 if (!x->done) {
3825 DECLARE_WAITQUEUE(wait, current);
3827 wait.flags |= WQ_FLAG_EXCLUSIVE;
3828 __add_wait_queue_tail(&x->wait, &wait);
3829 do {
3830 if (signal_pending(current)) {
3831 timeout = -ERESTARTSYS;
3832 __remove_wait_queue(&x->wait, &wait);
3833 goto out;
3835 __set_current_state(TASK_INTERRUPTIBLE);
3836 spin_unlock_irq(&x->wait.lock);
3837 timeout = schedule_timeout(timeout);
3838 spin_lock_irq(&x->wait.lock);
3839 if (!timeout) {
3840 __remove_wait_queue(&x->wait, &wait);
3841 goto out;
3843 } while (!x->done);
3844 __remove_wait_queue(&x->wait, &wait);
3846 x->done--;
3847 out:
3848 spin_unlock_irq(&x->wait.lock);
3849 return timeout;
3851 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3853 static inline void
3854 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3856 spin_lock_irqsave(&q->lock, *flags);
3857 __add_wait_queue(q, wait);
3858 spin_unlock(&q->lock);
3861 static inline void
3862 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3864 spin_lock_irq(&q->lock);
3865 __remove_wait_queue(q, wait);
3866 spin_unlock_irqrestore(&q->lock, *flags);
3869 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3871 unsigned long flags;
3872 wait_queue_t wait;
3874 init_waitqueue_entry(&wait, current);
3876 current->state = TASK_INTERRUPTIBLE;
3878 sleep_on_head(q, &wait, &flags);
3879 schedule();
3880 sleep_on_tail(q, &wait, &flags);
3882 EXPORT_SYMBOL(interruptible_sleep_on);
3884 long __sched
3885 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3887 unsigned long flags;
3888 wait_queue_t wait;
3890 init_waitqueue_entry(&wait, current);
3892 current->state = TASK_INTERRUPTIBLE;
3894 sleep_on_head(q, &wait, &flags);
3895 timeout = schedule_timeout(timeout);
3896 sleep_on_tail(q, &wait, &flags);
3898 return timeout;
3900 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3902 void __sched sleep_on(wait_queue_head_t *q)
3904 unsigned long flags;
3905 wait_queue_t wait;
3907 init_waitqueue_entry(&wait, current);
3909 current->state = TASK_UNINTERRUPTIBLE;
3911 sleep_on_head(q, &wait, &flags);
3912 schedule();
3913 sleep_on_tail(q, &wait, &flags);
3915 EXPORT_SYMBOL(sleep_on);
3917 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3919 unsigned long flags;
3920 wait_queue_t wait;
3922 init_waitqueue_entry(&wait, current);
3924 current->state = TASK_UNINTERRUPTIBLE;
3926 sleep_on_head(q, &wait, &flags);
3927 timeout = schedule_timeout(timeout);
3928 sleep_on_tail(q, &wait, &flags);
3930 return timeout;
3932 EXPORT_SYMBOL(sleep_on_timeout);
3934 #ifdef CONFIG_RT_MUTEXES
3937 * rt_mutex_setprio - set the current priority of a task
3938 * @p: task
3939 * @prio: prio value (kernel-internal form)
3941 * This function changes the 'effective' priority of a task. It does
3942 * not touch ->normal_prio like __setscheduler().
3944 * Used by the rt_mutex code to implement priority inheritance logic.
3946 void rt_mutex_setprio(struct task_struct *p, int prio)
3948 unsigned long flags;
3949 int oldprio, on_rq;
3950 struct rq *rq;
3952 BUG_ON(prio < 0 || prio > MAX_PRIO);
3954 rq = task_rq_lock(p, &flags);
3955 update_rq_clock(rq);
3957 oldprio = p->prio;
3958 on_rq = p->se.on_rq;
3959 if (on_rq)
3960 dequeue_task(rq, p, 0);
3962 if (rt_prio(prio))
3963 p->sched_class = &rt_sched_class;
3964 else
3965 p->sched_class = &fair_sched_class;
3967 p->prio = prio;
3969 if (on_rq) {
3970 enqueue_task(rq, p, 0);
3972 * Reschedule if we are currently running on this runqueue and
3973 * our priority decreased, or if we are not currently running on
3974 * this runqueue and our priority is higher than the current's
3976 if (task_running(rq, p)) {
3977 if (p->prio > oldprio)
3978 resched_task(rq->curr);
3979 } else {
3980 check_preempt_curr(rq, p);
3983 task_rq_unlock(rq, &flags);
3986 #endif
3988 void set_user_nice(struct task_struct *p, long nice)
3990 int old_prio, delta, on_rq;
3991 unsigned long flags;
3992 struct rq *rq;
3994 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3995 return;
3997 * We have to be careful, if called from sys_setpriority(),
3998 * the task might be in the middle of scheduling on another CPU.
4000 rq = task_rq_lock(p, &flags);
4001 update_rq_clock(rq);
4003 * The RT priorities are set via sched_setscheduler(), but we still
4004 * allow the 'normal' nice value to be set - but as expected
4005 * it wont have any effect on scheduling until the task is
4006 * SCHED_FIFO/SCHED_RR:
4008 if (task_has_rt_policy(p)) {
4009 p->static_prio = NICE_TO_PRIO(nice);
4010 goto out_unlock;
4012 on_rq = p->se.on_rq;
4013 if (on_rq) {
4014 dequeue_task(rq, p, 0);
4015 dec_load(rq, p);
4018 p->static_prio = NICE_TO_PRIO(nice);
4019 set_load_weight(p);
4020 old_prio = p->prio;
4021 p->prio = effective_prio(p);
4022 delta = p->prio - old_prio;
4024 if (on_rq) {
4025 enqueue_task(rq, p, 0);
4026 inc_load(rq, p);
4028 * If the task increased its priority or is running and
4029 * lowered its priority, then reschedule its CPU:
4031 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4032 resched_task(rq->curr);
4034 out_unlock:
4035 task_rq_unlock(rq, &flags);
4037 EXPORT_SYMBOL(set_user_nice);
4040 * can_nice - check if a task can reduce its nice value
4041 * @p: task
4042 * @nice: nice value
4044 int can_nice(const struct task_struct *p, const int nice)
4046 /* convert nice value [19,-20] to rlimit style value [1,40] */
4047 int nice_rlim = 20 - nice;
4049 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4050 capable(CAP_SYS_NICE));
4053 #ifdef __ARCH_WANT_SYS_NICE
4056 * sys_nice - change the priority of the current process.
4057 * @increment: priority increment
4059 * sys_setpriority is a more generic, but much slower function that
4060 * does similar things.
4062 asmlinkage long sys_nice(int increment)
4064 long nice, retval;
4067 * Setpriority might change our priority at the same moment.
4068 * We don't have to worry. Conceptually one call occurs first
4069 * and we have a single winner.
4071 if (increment < -40)
4072 increment = -40;
4073 if (increment > 40)
4074 increment = 40;
4076 nice = PRIO_TO_NICE(current->static_prio) + increment;
4077 if (nice < -20)
4078 nice = -20;
4079 if (nice > 19)
4080 nice = 19;
4082 if (increment < 0 && !can_nice(current, nice))
4083 return -EPERM;
4085 retval = security_task_setnice(current, nice);
4086 if (retval)
4087 return retval;
4089 set_user_nice(current, nice);
4090 return 0;
4093 #endif
4096 * task_prio - return the priority value of a given task.
4097 * @p: the task in question.
4099 * This is the priority value as seen by users in /proc.
4100 * RT tasks are offset by -200. Normal tasks are centered
4101 * around 0, value goes from -16 to +15.
4103 int task_prio(const struct task_struct *p)
4105 return p->prio - MAX_RT_PRIO;
4109 * task_nice - return the nice value of a given task.
4110 * @p: the task in question.
4112 int task_nice(const struct task_struct *p)
4114 return TASK_NICE(p);
4116 EXPORT_SYMBOL_GPL(task_nice);
4119 * idle_cpu - is a given cpu idle currently?
4120 * @cpu: the processor in question.
4122 int idle_cpu(int cpu)
4124 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4128 * idle_task - return the idle task for a given cpu.
4129 * @cpu: the processor in question.
4131 struct task_struct *idle_task(int cpu)
4133 return cpu_rq(cpu)->idle;
4137 * find_process_by_pid - find a process with a matching PID value.
4138 * @pid: the pid in question.
4140 static inline struct task_struct *find_process_by_pid(pid_t pid)
4142 return pid ? find_task_by_pid(pid) : current;
4145 /* Actually do priority change: must hold rq lock. */
4146 static void
4147 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4149 BUG_ON(p->se.on_rq);
4151 p->policy = policy;
4152 switch (p->policy) {
4153 case SCHED_NORMAL:
4154 case SCHED_BATCH:
4155 case SCHED_IDLE:
4156 p->sched_class = &fair_sched_class;
4157 break;
4158 case SCHED_FIFO:
4159 case SCHED_RR:
4160 p->sched_class = &rt_sched_class;
4161 break;
4164 p->rt_priority = prio;
4165 p->normal_prio = normal_prio(p);
4166 /* we are holding p->pi_lock already */
4167 p->prio = rt_mutex_getprio(p);
4168 set_load_weight(p);
4172 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4173 * @p: the task in question.
4174 * @policy: new policy.
4175 * @param: structure containing the new RT priority.
4177 * NOTE that the task may be already dead.
4179 int sched_setscheduler(struct task_struct *p, int policy,
4180 struct sched_param *param)
4182 int retval, oldprio, oldpolicy = -1, on_rq;
4183 unsigned long flags;
4184 struct rq *rq;
4186 /* may grab non-irq protected spin_locks */
4187 BUG_ON(in_interrupt());
4188 recheck:
4189 /* double check policy once rq lock held */
4190 if (policy < 0)
4191 policy = oldpolicy = p->policy;
4192 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4193 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4194 policy != SCHED_IDLE)
4195 return -EINVAL;
4197 * Valid priorities for SCHED_FIFO and SCHED_RR are
4198 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4199 * SCHED_BATCH and SCHED_IDLE is 0.
4201 if (param->sched_priority < 0 ||
4202 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4203 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4204 return -EINVAL;
4205 if (rt_policy(policy) != (param->sched_priority != 0))
4206 return -EINVAL;
4209 * Allow unprivileged RT tasks to decrease priority:
4211 if (!capable(CAP_SYS_NICE)) {
4212 if (rt_policy(policy)) {
4213 unsigned long rlim_rtprio;
4215 if (!lock_task_sighand(p, &flags))
4216 return -ESRCH;
4217 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4218 unlock_task_sighand(p, &flags);
4220 /* can't set/change the rt policy */
4221 if (policy != p->policy && !rlim_rtprio)
4222 return -EPERM;
4224 /* can't increase priority */
4225 if (param->sched_priority > p->rt_priority &&
4226 param->sched_priority > rlim_rtprio)
4227 return -EPERM;
4230 * Like positive nice levels, dont allow tasks to
4231 * move out of SCHED_IDLE either:
4233 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4234 return -EPERM;
4236 /* can't change other user's priorities */
4237 if ((current->euid != p->euid) &&
4238 (current->euid != p->uid))
4239 return -EPERM;
4242 retval = security_task_setscheduler(p, policy, param);
4243 if (retval)
4244 return retval;
4246 * make sure no PI-waiters arrive (or leave) while we are
4247 * changing the priority of the task:
4249 spin_lock_irqsave(&p->pi_lock, flags);
4251 * To be able to change p->policy safely, the apropriate
4252 * runqueue lock must be held.
4254 rq = __task_rq_lock(p);
4255 /* recheck policy now with rq lock held */
4256 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4257 policy = oldpolicy = -1;
4258 __task_rq_unlock(rq);
4259 spin_unlock_irqrestore(&p->pi_lock, flags);
4260 goto recheck;
4262 update_rq_clock(rq);
4263 on_rq = p->se.on_rq;
4264 if (on_rq)
4265 deactivate_task(rq, p, 0);
4266 oldprio = p->prio;
4267 __setscheduler(rq, p, policy, param->sched_priority);
4268 if (on_rq) {
4269 activate_task(rq, p, 0);
4271 * Reschedule if we are currently running on this runqueue and
4272 * our priority decreased, or if we are not currently running on
4273 * this runqueue and our priority is higher than the current's
4275 if (task_running(rq, p)) {
4276 if (p->prio > oldprio)
4277 resched_task(rq->curr);
4278 } else {
4279 check_preempt_curr(rq, p);
4282 __task_rq_unlock(rq);
4283 spin_unlock_irqrestore(&p->pi_lock, flags);
4285 rt_mutex_adjust_pi(p);
4287 return 0;
4289 EXPORT_SYMBOL_GPL(sched_setscheduler);
4291 static int
4292 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4294 struct sched_param lparam;
4295 struct task_struct *p;
4296 int retval;
4298 if (!param || pid < 0)
4299 return -EINVAL;
4300 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4301 return -EFAULT;
4303 rcu_read_lock();
4304 retval = -ESRCH;
4305 p = find_process_by_pid(pid);
4306 if (p != NULL)
4307 retval = sched_setscheduler(p, policy, &lparam);
4308 rcu_read_unlock();
4310 return retval;
4314 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4315 * @pid: the pid in question.
4316 * @policy: new policy.
4317 * @param: structure containing the new RT priority.
4319 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4320 struct sched_param __user *param)
4322 /* negative values for policy are not valid */
4323 if (policy < 0)
4324 return -EINVAL;
4326 return do_sched_setscheduler(pid, policy, param);
4330 * sys_sched_setparam - set/change the RT priority of a thread
4331 * @pid: the pid in question.
4332 * @param: structure containing the new RT priority.
4334 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4336 return do_sched_setscheduler(pid, -1, param);
4340 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4341 * @pid: the pid in question.
4343 asmlinkage long sys_sched_getscheduler(pid_t pid)
4345 struct task_struct *p;
4346 int retval = -EINVAL;
4348 if (pid < 0)
4349 goto out_nounlock;
4351 retval = -ESRCH;
4352 read_lock(&tasklist_lock);
4353 p = find_process_by_pid(pid);
4354 if (p) {
4355 retval = security_task_getscheduler(p);
4356 if (!retval)
4357 retval = p->policy;
4359 read_unlock(&tasklist_lock);
4361 out_nounlock:
4362 return retval;
4366 * sys_sched_getscheduler - get the RT priority of a thread
4367 * @pid: the pid in question.
4368 * @param: structure containing the RT priority.
4370 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4372 struct sched_param lp;
4373 struct task_struct *p;
4374 int retval = -EINVAL;
4376 if (!param || pid < 0)
4377 goto out_nounlock;
4379 read_lock(&tasklist_lock);
4380 p = find_process_by_pid(pid);
4381 retval = -ESRCH;
4382 if (!p)
4383 goto out_unlock;
4385 retval = security_task_getscheduler(p);
4386 if (retval)
4387 goto out_unlock;
4389 lp.sched_priority = p->rt_priority;
4390 read_unlock(&tasklist_lock);
4393 * This one might sleep, we cannot do it with a spinlock held ...
4395 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4397 out_nounlock:
4398 return retval;
4400 out_unlock:
4401 read_unlock(&tasklist_lock);
4402 return retval;
4405 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4407 cpumask_t cpus_allowed;
4408 struct task_struct *p;
4409 int retval;
4411 mutex_lock(&sched_hotcpu_mutex);
4412 read_lock(&tasklist_lock);
4414 p = find_process_by_pid(pid);
4415 if (!p) {
4416 read_unlock(&tasklist_lock);
4417 mutex_unlock(&sched_hotcpu_mutex);
4418 return -ESRCH;
4422 * It is not safe to call set_cpus_allowed with the
4423 * tasklist_lock held. We will bump the task_struct's
4424 * usage count and then drop tasklist_lock.
4426 get_task_struct(p);
4427 read_unlock(&tasklist_lock);
4429 retval = -EPERM;
4430 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4431 !capable(CAP_SYS_NICE))
4432 goto out_unlock;
4434 retval = security_task_setscheduler(p, 0, NULL);
4435 if (retval)
4436 goto out_unlock;
4438 cpus_allowed = cpuset_cpus_allowed(p);
4439 cpus_and(new_mask, new_mask, cpus_allowed);
4440 retval = set_cpus_allowed(p, new_mask);
4442 out_unlock:
4443 put_task_struct(p);
4444 mutex_unlock(&sched_hotcpu_mutex);
4445 return retval;
4448 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4449 cpumask_t *new_mask)
4451 if (len < sizeof(cpumask_t)) {
4452 memset(new_mask, 0, sizeof(cpumask_t));
4453 } else if (len > sizeof(cpumask_t)) {
4454 len = sizeof(cpumask_t);
4456 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4460 * sys_sched_setaffinity - set the cpu affinity of a process
4461 * @pid: pid of the process
4462 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4463 * @user_mask_ptr: user-space pointer to the new cpu mask
4465 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4466 unsigned long __user *user_mask_ptr)
4468 cpumask_t new_mask;
4469 int retval;
4471 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4472 if (retval)
4473 return retval;
4475 return sched_setaffinity(pid, new_mask);
4479 * Represents all cpu's present in the system
4480 * In systems capable of hotplug, this map could dynamically grow
4481 * as new cpu's are detected in the system via any platform specific
4482 * method, such as ACPI for e.g.
4485 cpumask_t cpu_present_map __read_mostly;
4486 EXPORT_SYMBOL(cpu_present_map);
4488 #ifndef CONFIG_SMP
4489 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4490 EXPORT_SYMBOL(cpu_online_map);
4492 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4493 EXPORT_SYMBOL(cpu_possible_map);
4494 #endif
4496 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4498 struct task_struct *p;
4499 int retval;
4501 mutex_lock(&sched_hotcpu_mutex);
4502 read_lock(&tasklist_lock);
4504 retval = -ESRCH;
4505 p = find_process_by_pid(pid);
4506 if (!p)
4507 goto out_unlock;
4509 retval = security_task_getscheduler(p);
4510 if (retval)
4511 goto out_unlock;
4513 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4515 out_unlock:
4516 read_unlock(&tasklist_lock);
4517 mutex_unlock(&sched_hotcpu_mutex);
4519 return retval;
4523 * sys_sched_getaffinity - get the cpu affinity of a process
4524 * @pid: pid of the process
4525 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4526 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4528 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4529 unsigned long __user *user_mask_ptr)
4531 int ret;
4532 cpumask_t mask;
4534 if (len < sizeof(cpumask_t))
4535 return -EINVAL;
4537 ret = sched_getaffinity(pid, &mask);
4538 if (ret < 0)
4539 return ret;
4541 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4542 return -EFAULT;
4544 return sizeof(cpumask_t);
4548 * sys_sched_yield - yield the current processor to other threads.
4550 * This function yields the current CPU to other tasks. If there are no
4551 * other threads running on this CPU then this function will return.
4553 asmlinkage long sys_sched_yield(void)
4555 struct rq *rq = this_rq_lock();
4557 schedstat_inc(rq, yld_cnt);
4558 current->sched_class->yield_task(rq, current);
4561 * Since we are going to call schedule() anyway, there's
4562 * no need to preempt or enable interrupts:
4564 __release(rq->lock);
4565 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4566 _raw_spin_unlock(&rq->lock);
4567 preempt_enable_no_resched();
4569 schedule();
4571 return 0;
4574 static void __cond_resched(void)
4576 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4577 __might_sleep(__FILE__, __LINE__);
4578 #endif
4580 * The BKS might be reacquired before we have dropped
4581 * PREEMPT_ACTIVE, which could trigger a second
4582 * cond_resched() call.
4584 do {
4585 add_preempt_count(PREEMPT_ACTIVE);
4586 schedule();
4587 sub_preempt_count(PREEMPT_ACTIVE);
4588 } while (need_resched());
4591 int __sched cond_resched(void)
4593 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4594 system_state == SYSTEM_RUNNING) {
4595 __cond_resched();
4596 return 1;
4598 return 0;
4600 EXPORT_SYMBOL(cond_resched);
4603 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4604 * call schedule, and on return reacquire the lock.
4606 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4607 * operations here to prevent schedule() from being called twice (once via
4608 * spin_unlock(), once by hand).
4610 int cond_resched_lock(spinlock_t *lock)
4612 int ret = 0;
4614 if (need_lockbreak(lock)) {
4615 spin_unlock(lock);
4616 cpu_relax();
4617 ret = 1;
4618 spin_lock(lock);
4620 if (need_resched() && system_state == SYSTEM_RUNNING) {
4621 spin_release(&lock->dep_map, 1, _THIS_IP_);
4622 _raw_spin_unlock(lock);
4623 preempt_enable_no_resched();
4624 __cond_resched();
4625 ret = 1;
4626 spin_lock(lock);
4628 return ret;
4630 EXPORT_SYMBOL(cond_resched_lock);
4632 int __sched cond_resched_softirq(void)
4634 BUG_ON(!in_softirq());
4636 if (need_resched() && system_state == SYSTEM_RUNNING) {
4637 local_bh_enable();
4638 __cond_resched();
4639 local_bh_disable();
4640 return 1;
4642 return 0;
4644 EXPORT_SYMBOL(cond_resched_softirq);
4647 * yield - yield the current processor to other threads.
4649 * This is a shortcut for kernel-space yielding - it marks the
4650 * thread runnable and calls sys_sched_yield().
4652 void __sched yield(void)
4654 set_current_state(TASK_RUNNING);
4655 sys_sched_yield();
4657 EXPORT_SYMBOL(yield);
4660 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4661 * that process accounting knows that this is a task in IO wait state.
4663 * But don't do that if it is a deliberate, throttling IO wait (this task
4664 * has set its backing_dev_info: the queue against which it should throttle)
4666 void __sched io_schedule(void)
4668 struct rq *rq = &__raw_get_cpu_var(runqueues);
4670 delayacct_blkio_start();
4671 atomic_inc(&rq->nr_iowait);
4672 schedule();
4673 atomic_dec(&rq->nr_iowait);
4674 delayacct_blkio_end();
4676 EXPORT_SYMBOL(io_schedule);
4678 long __sched io_schedule_timeout(long timeout)
4680 struct rq *rq = &__raw_get_cpu_var(runqueues);
4681 long ret;
4683 delayacct_blkio_start();
4684 atomic_inc(&rq->nr_iowait);
4685 ret = schedule_timeout(timeout);
4686 atomic_dec(&rq->nr_iowait);
4687 delayacct_blkio_end();
4688 return ret;
4692 * sys_sched_get_priority_max - return maximum RT priority.
4693 * @policy: scheduling class.
4695 * this syscall returns the maximum rt_priority that can be used
4696 * by a given scheduling class.
4698 asmlinkage long sys_sched_get_priority_max(int policy)
4700 int ret = -EINVAL;
4702 switch (policy) {
4703 case SCHED_FIFO:
4704 case SCHED_RR:
4705 ret = MAX_USER_RT_PRIO-1;
4706 break;
4707 case SCHED_NORMAL:
4708 case SCHED_BATCH:
4709 case SCHED_IDLE:
4710 ret = 0;
4711 break;
4713 return ret;
4717 * sys_sched_get_priority_min - return minimum RT priority.
4718 * @policy: scheduling class.
4720 * this syscall returns the minimum rt_priority that can be used
4721 * by a given scheduling class.
4723 asmlinkage long sys_sched_get_priority_min(int policy)
4725 int ret = -EINVAL;
4727 switch (policy) {
4728 case SCHED_FIFO:
4729 case SCHED_RR:
4730 ret = 1;
4731 break;
4732 case SCHED_NORMAL:
4733 case SCHED_BATCH:
4734 case SCHED_IDLE:
4735 ret = 0;
4737 return ret;
4741 * sys_sched_rr_get_interval - return the default timeslice of a process.
4742 * @pid: pid of the process.
4743 * @interval: userspace pointer to the timeslice value.
4745 * this syscall writes the default timeslice value of a given process
4746 * into the user-space timespec buffer. A value of '0' means infinity.
4748 asmlinkage
4749 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4751 struct task_struct *p;
4752 int retval = -EINVAL;
4753 struct timespec t;
4755 if (pid < 0)
4756 goto out_nounlock;
4758 retval = -ESRCH;
4759 read_lock(&tasklist_lock);
4760 p = find_process_by_pid(pid);
4761 if (!p)
4762 goto out_unlock;
4764 retval = security_task_getscheduler(p);
4765 if (retval)
4766 goto out_unlock;
4768 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4769 0 : static_prio_timeslice(p->static_prio), &t);
4770 read_unlock(&tasklist_lock);
4771 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4772 out_nounlock:
4773 return retval;
4774 out_unlock:
4775 read_unlock(&tasklist_lock);
4776 return retval;
4779 static const char stat_nam[] = "RSDTtZX";
4781 static void show_task(struct task_struct *p)
4783 unsigned long free = 0;
4784 unsigned state;
4786 state = p->state ? __ffs(p->state) + 1 : 0;
4787 printk("%-13.13s %c", p->comm,
4788 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4789 #if BITS_PER_LONG == 32
4790 if (state == TASK_RUNNING)
4791 printk(" running ");
4792 else
4793 printk(" %08lx ", thread_saved_pc(p));
4794 #else
4795 if (state == TASK_RUNNING)
4796 printk(" running task ");
4797 else
4798 printk(" %016lx ", thread_saved_pc(p));
4799 #endif
4800 #ifdef CONFIG_DEBUG_STACK_USAGE
4802 unsigned long *n = end_of_stack(p);
4803 while (!*n)
4804 n++;
4805 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4807 #endif
4808 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4810 if (state != TASK_RUNNING)
4811 show_stack(p, NULL);
4814 void show_state_filter(unsigned long state_filter)
4816 struct task_struct *g, *p;
4818 #if BITS_PER_LONG == 32
4819 printk(KERN_INFO
4820 " task PC stack pid father\n");
4821 #else
4822 printk(KERN_INFO
4823 " task PC stack pid father\n");
4824 #endif
4825 read_lock(&tasklist_lock);
4826 do_each_thread(g, p) {
4828 * reset the NMI-timeout, listing all files on a slow
4829 * console might take alot of time:
4831 touch_nmi_watchdog();
4832 if (!state_filter || (p->state & state_filter))
4833 show_task(p);
4834 } while_each_thread(g, p);
4836 touch_all_softlockup_watchdogs();
4838 #ifdef CONFIG_SCHED_DEBUG
4839 sysrq_sched_debug_show();
4840 #endif
4841 read_unlock(&tasklist_lock);
4843 * Only show locks if all tasks are dumped:
4845 if (state_filter == -1)
4846 debug_show_all_locks();
4849 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4851 idle->sched_class = &idle_sched_class;
4855 * init_idle - set up an idle thread for a given CPU
4856 * @idle: task in question
4857 * @cpu: cpu the idle task belongs to
4859 * NOTE: this function does not set the idle thread's NEED_RESCHED
4860 * flag, to make booting more robust.
4862 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4864 struct rq *rq = cpu_rq(cpu);
4865 unsigned long flags;
4867 __sched_fork(idle);
4868 idle->se.exec_start = sched_clock();
4870 idle->prio = idle->normal_prio = MAX_PRIO;
4871 idle->cpus_allowed = cpumask_of_cpu(cpu);
4872 __set_task_cpu(idle, cpu);
4874 spin_lock_irqsave(&rq->lock, flags);
4875 rq->curr = rq->idle = idle;
4876 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4877 idle->oncpu = 1;
4878 #endif
4879 spin_unlock_irqrestore(&rq->lock, flags);
4881 /* Set the preempt count _outside_ the spinlocks! */
4882 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4883 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4884 #else
4885 task_thread_info(idle)->preempt_count = 0;
4886 #endif
4888 * The idle tasks have their own, simple scheduling class:
4890 idle->sched_class = &idle_sched_class;
4894 * In a system that switches off the HZ timer nohz_cpu_mask
4895 * indicates which cpus entered this state. This is used
4896 * in the rcu update to wait only for active cpus. For system
4897 * which do not switch off the HZ timer nohz_cpu_mask should
4898 * always be CPU_MASK_NONE.
4900 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4903 * Increase the granularity value when there are more CPUs,
4904 * because with more CPUs the 'effective latency' as visible
4905 * to users decreases. But the relationship is not linear,
4906 * so pick a second-best guess by going with the log2 of the
4907 * number of CPUs.
4909 * This idea comes from the SD scheduler of Con Kolivas:
4911 static inline void sched_init_granularity(void)
4913 unsigned int factor = 1 + ilog2(num_online_cpus());
4914 const unsigned long limit = 100000000;
4916 sysctl_sched_min_granularity *= factor;
4917 if (sysctl_sched_min_granularity > limit)
4918 sysctl_sched_min_granularity = limit;
4920 sysctl_sched_latency *= factor;
4921 if (sysctl_sched_latency > limit)
4922 sysctl_sched_latency = limit;
4924 sysctl_sched_runtime_limit = sysctl_sched_latency;
4925 sysctl_sched_wakeup_granularity = sysctl_sched_min_granularity / 2;
4928 #ifdef CONFIG_SMP
4930 * This is how migration works:
4932 * 1) we queue a struct migration_req structure in the source CPU's
4933 * runqueue and wake up that CPU's migration thread.
4934 * 2) we down() the locked semaphore => thread blocks.
4935 * 3) migration thread wakes up (implicitly it forces the migrated
4936 * thread off the CPU)
4937 * 4) it gets the migration request and checks whether the migrated
4938 * task is still in the wrong runqueue.
4939 * 5) if it's in the wrong runqueue then the migration thread removes
4940 * it and puts it into the right queue.
4941 * 6) migration thread up()s the semaphore.
4942 * 7) we wake up and the migration is done.
4946 * Change a given task's CPU affinity. Migrate the thread to a
4947 * proper CPU and schedule it away if the CPU it's executing on
4948 * is removed from the allowed bitmask.
4950 * NOTE: the caller must have a valid reference to the task, the
4951 * task must not exit() & deallocate itself prematurely. The
4952 * call is not atomic; no spinlocks may be held.
4954 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4956 struct migration_req req;
4957 unsigned long flags;
4958 struct rq *rq;
4959 int ret = 0;
4961 rq = task_rq_lock(p, &flags);
4962 if (!cpus_intersects(new_mask, cpu_online_map)) {
4963 ret = -EINVAL;
4964 goto out;
4967 p->cpus_allowed = new_mask;
4968 /* Can the task run on the task's current CPU? If so, we're done */
4969 if (cpu_isset(task_cpu(p), new_mask))
4970 goto out;
4972 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4973 /* Need help from migration thread: drop lock and wait. */
4974 task_rq_unlock(rq, &flags);
4975 wake_up_process(rq->migration_thread);
4976 wait_for_completion(&req.done);
4977 tlb_migrate_finish(p->mm);
4978 return 0;
4980 out:
4981 task_rq_unlock(rq, &flags);
4983 return ret;
4985 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4988 * Move (not current) task off this cpu, onto dest cpu. We're doing
4989 * this because either it can't run here any more (set_cpus_allowed()
4990 * away from this CPU, or CPU going down), or because we're
4991 * attempting to rebalance this task on exec (sched_exec).
4993 * So we race with normal scheduler movements, but that's OK, as long
4994 * as the task is no longer on this CPU.
4996 * Returns non-zero if task was successfully migrated.
4998 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5000 struct rq *rq_dest, *rq_src;
5001 int ret = 0, on_rq;
5003 if (unlikely(cpu_is_offline(dest_cpu)))
5004 return ret;
5006 rq_src = cpu_rq(src_cpu);
5007 rq_dest = cpu_rq(dest_cpu);
5009 double_rq_lock(rq_src, rq_dest);
5010 /* Already moved. */
5011 if (task_cpu(p) != src_cpu)
5012 goto out;
5013 /* Affinity changed (again). */
5014 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5015 goto out;
5017 on_rq = p->se.on_rq;
5018 if (on_rq)
5019 deactivate_task(rq_src, p, 0);
5021 set_task_cpu(p, dest_cpu);
5022 if (on_rq) {
5023 activate_task(rq_dest, p, 0);
5024 check_preempt_curr(rq_dest, p);
5026 ret = 1;
5027 out:
5028 double_rq_unlock(rq_src, rq_dest);
5029 return ret;
5033 * migration_thread - this is a highprio system thread that performs
5034 * thread migration by bumping thread off CPU then 'pushing' onto
5035 * another runqueue.
5037 static int migration_thread(void *data)
5039 int cpu = (long)data;
5040 struct rq *rq;
5042 rq = cpu_rq(cpu);
5043 BUG_ON(rq->migration_thread != current);
5045 set_current_state(TASK_INTERRUPTIBLE);
5046 while (!kthread_should_stop()) {
5047 struct migration_req *req;
5048 struct list_head *head;
5050 spin_lock_irq(&rq->lock);
5052 if (cpu_is_offline(cpu)) {
5053 spin_unlock_irq(&rq->lock);
5054 goto wait_to_die;
5057 if (rq->active_balance) {
5058 active_load_balance(rq, cpu);
5059 rq->active_balance = 0;
5062 head = &rq->migration_queue;
5064 if (list_empty(head)) {
5065 spin_unlock_irq(&rq->lock);
5066 schedule();
5067 set_current_state(TASK_INTERRUPTIBLE);
5068 continue;
5070 req = list_entry(head->next, struct migration_req, list);
5071 list_del_init(head->next);
5073 spin_unlock(&rq->lock);
5074 __migrate_task(req->task, cpu, req->dest_cpu);
5075 local_irq_enable();
5077 complete(&req->done);
5079 __set_current_state(TASK_RUNNING);
5080 return 0;
5082 wait_to_die:
5083 /* Wait for kthread_stop */
5084 set_current_state(TASK_INTERRUPTIBLE);
5085 while (!kthread_should_stop()) {
5086 schedule();
5087 set_current_state(TASK_INTERRUPTIBLE);
5089 __set_current_state(TASK_RUNNING);
5090 return 0;
5093 #ifdef CONFIG_HOTPLUG_CPU
5095 * Figure out where task on dead CPU should go, use force if neccessary.
5096 * NOTE: interrupts should be disabled by the caller
5098 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5100 unsigned long flags;
5101 cpumask_t mask;
5102 struct rq *rq;
5103 int dest_cpu;
5105 restart:
5106 /* On same node? */
5107 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5108 cpus_and(mask, mask, p->cpus_allowed);
5109 dest_cpu = any_online_cpu(mask);
5111 /* On any allowed CPU? */
5112 if (dest_cpu == NR_CPUS)
5113 dest_cpu = any_online_cpu(p->cpus_allowed);
5115 /* No more Mr. Nice Guy. */
5116 if (dest_cpu == NR_CPUS) {
5117 rq = task_rq_lock(p, &flags);
5118 cpus_setall(p->cpus_allowed);
5119 dest_cpu = any_online_cpu(p->cpus_allowed);
5120 task_rq_unlock(rq, &flags);
5123 * Don't tell them about moving exiting tasks or
5124 * kernel threads (both mm NULL), since they never
5125 * leave kernel.
5127 if (p->mm && printk_ratelimit())
5128 printk(KERN_INFO "process %d (%s) no "
5129 "longer affine to cpu%d\n",
5130 p->pid, p->comm, dead_cpu);
5132 if (!__migrate_task(p, dead_cpu, dest_cpu))
5133 goto restart;
5137 * While a dead CPU has no uninterruptible tasks queued at this point,
5138 * it might still have a nonzero ->nr_uninterruptible counter, because
5139 * for performance reasons the counter is not stricly tracking tasks to
5140 * their home CPUs. So we just add the counter to another CPU's counter,
5141 * to keep the global sum constant after CPU-down:
5143 static void migrate_nr_uninterruptible(struct rq *rq_src)
5145 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5146 unsigned long flags;
5148 local_irq_save(flags);
5149 double_rq_lock(rq_src, rq_dest);
5150 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5151 rq_src->nr_uninterruptible = 0;
5152 double_rq_unlock(rq_src, rq_dest);
5153 local_irq_restore(flags);
5156 /* Run through task list and migrate tasks from the dead cpu. */
5157 static void migrate_live_tasks(int src_cpu)
5159 struct task_struct *p, *t;
5161 write_lock_irq(&tasklist_lock);
5163 do_each_thread(t, p) {
5164 if (p == current)
5165 continue;
5167 if (task_cpu(p) == src_cpu)
5168 move_task_off_dead_cpu(src_cpu, p);
5169 } while_each_thread(t, p);
5171 write_unlock_irq(&tasklist_lock);
5175 * Schedules idle task to be the next runnable task on current CPU.
5176 * It does so by boosting its priority to highest possible and adding it to
5177 * the _front_ of the runqueue. Used by CPU offline code.
5179 void sched_idle_next(void)
5181 int this_cpu = smp_processor_id();
5182 struct rq *rq = cpu_rq(this_cpu);
5183 struct task_struct *p = rq->idle;
5184 unsigned long flags;
5186 /* cpu has to be offline */
5187 BUG_ON(cpu_online(this_cpu));
5190 * Strictly not necessary since rest of the CPUs are stopped by now
5191 * and interrupts disabled on the current cpu.
5193 spin_lock_irqsave(&rq->lock, flags);
5195 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5197 /* Add idle task to the _front_ of its priority queue: */
5198 activate_idle_task(p, rq);
5200 spin_unlock_irqrestore(&rq->lock, flags);
5204 * Ensures that the idle task is using init_mm right before its cpu goes
5205 * offline.
5207 void idle_task_exit(void)
5209 struct mm_struct *mm = current->active_mm;
5211 BUG_ON(cpu_online(smp_processor_id()));
5213 if (mm != &init_mm)
5214 switch_mm(mm, &init_mm, current);
5215 mmdrop(mm);
5218 /* called under rq->lock with disabled interrupts */
5219 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5221 struct rq *rq = cpu_rq(dead_cpu);
5223 /* Must be exiting, otherwise would be on tasklist. */
5224 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5226 /* Cannot have done final schedule yet: would have vanished. */
5227 BUG_ON(p->state == TASK_DEAD);
5229 get_task_struct(p);
5232 * Drop lock around migration; if someone else moves it,
5233 * that's OK. No task can be added to this CPU, so iteration is
5234 * fine.
5235 * NOTE: interrupts should be left disabled --dev@
5237 spin_unlock(&rq->lock);
5238 move_task_off_dead_cpu(dead_cpu, p);
5239 spin_lock(&rq->lock);
5241 put_task_struct(p);
5244 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5245 static void migrate_dead_tasks(unsigned int dead_cpu)
5247 struct rq *rq = cpu_rq(dead_cpu);
5248 struct task_struct *next;
5250 for ( ; ; ) {
5251 if (!rq->nr_running)
5252 break;
5253 update_rq_clock(rq);
5254 next = pick_next_task(rq, rq->curr);
5255 if (!next)
5256 break;
5257 migrate_dead(dead_cpu, next);
5261 #endif /* CONFIG_HOTPLUG_CPU */
5263 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5265 static struct ctl_table sd_ctl_dir[] = {
5267 .procname = "sched_domain",
5268 .mode = 0555,
5270 {0,},
5273 static struct ctl_table sd_ctl_root[] = {
5275 .ctl_name = CTL_KERN,
5276 .procname = "kernel",
5277 .mode = 0555,
5278 .child = sd_ctl_dir,
5280 {0,},
5283 static struct ctl_table *sd_alloc_ctl_entry(int n)
5285 struct ctl_table *entry =
5286 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5288 BUG_ON(!entry);
5289 memset(entry, 0, n * sizeof(struct ctl_table));
5291 return entry;
5294 static void
5295 set_table_entry(struct ctl_table *entry,
5296 const char *procname, void *data, int maxlen,
5297 mode_t mode, proc_handler *proc_handler)
5299 entry->procname = procname;
5300 entry->data = data;
5301 entry->maxlen = maxlen;
5302 entry->mode = mode;
5303 entry->proc_handler = proc_handler;
5306 static struct ctl_table *
5307 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5309 struct ctl_table *table = sd_alloc_ctl_entry(14);
5311 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5312 sizeof(long), 0644, proc_doulongvec_minmax);
5313 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5314 sizeof(long), 0644, proc_doulongvec_minmax);
5315 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5316 sizeof(int), 0644, proc_dointvec_minmax);
5317 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5318 sizeof(int), 0644, proc_dointvec_minmax);
5319 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5320 sizeof(int), 0644, proc_dointvec_minmax);
5321 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5322 sizeof(int), 0644, proc_dointvec_minmax);
5323 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5324 sizeof(int), 0644, proc_dointvec_minmax);
5325 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5326 sizeof(int), 0644, proc_dointvec_minmax);
5327 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5328 sizeof(int), 0644, proc_dointvec_minmax);
5329 set_table_entry(&table[10], "cache_nice_tries",
5330 &sd->cache_nice_tries,
5331 sizeof(int), 0644, proc_dointvec_minmax);
5332 set_table_entry(&table[12], "flags", &sd->flags,
5333 sizeof(int), 0644, proc_dointvec_minmax);
5335 return table;
5338 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5340 struct ctl_table *entry, *table;
5341 struct sched_domain *sd;
5342 int domain_num = 0, i;
5343 char buf[32];
5345 for_each_domain(cpu, sd)
5346 domain_num++;
5347 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5349 i = 0;
5350 for_each_domain(cpu, sd) {
5351 snprintf(buf, 32, "domain%d", i);
5352 entry->procname = kstrdup(buf, GFP_KERNEL);
5353 entry->mode = 0555;
5354 entry->child = sd_alloc_ctl_domain_table(sd);
5355 entry++;
5356 i++;
5358 return table;
5361 static struct ctl_table_header *sd_sysctl_header;
5362 static void init_sched_domain_sysctl(void)
5364 int i, cpu_num = num_online_cpus();
5365 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5366 char buf[32];
5368 sd_ctl_dir[0].child = entry;
5370 for (i = 0; i < cpu_num; i++, entry++) {
5371 snprintf(buf, 32, "cpu%d", i);
5372 entry->procname = kstrdup(buf, GFP_KERNEL);
5373 entry->mode = 0555;
5374 entry->child = sd_alloc_ctl_cpu_table(i);
5376 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5378 #else
5379 static void init_sched_domain_sysctl(void)
5382 #endif
5385 * migration_call - callback that gets triggered when a CPU is added.
5386 * Here we can start up the necessary migration thread for the new CPU.
5388 static int __cpuinit
5389 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5391 struct task_struct *p;
5392 int cpu = (long)hcpu;
5393 unsigned long flags;
5394 struct rq *rq;
5396 switch (action) {
5397 case CPU_LOCK_ACQUIRE:
5398 mutex_lock(&sched_hotcpu_mutex);
5399 break;
5401 case CPU_UP_PREPARE:
5402 case CPU_UP_PREPARE_FROZEN:
5403 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5404 if (IS_ERR(p))
5405 return NOTIFY_BAD;
5406 kthread_bind(p, cpu);
5407 /* Must be high prio: stop_machine expects to yield to it. */
5408 rq = task_rq_lock(p, &flags);
5409 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5410 task_rq_unlock(rq, &flags);
5411 cpu_rq(cpu)->migration_thread = p;
5412 break;
5414 case CPU_ONLINE:
5415 case CPU_ONLINE_FROZEN:
5416 /* Strictly unneccessary, as first user will wake it. */
5417 wake_up_process(cpu_rq(cpu)->migration_thread);
5418 break;
5420 #ifdef CONFIG_HOTPLUG_CPU
5421 case CPU_UP_CANCELED:
5422 case CPU_UP_CANCELED_FROZEN:
5423 if (!cpu_rq(cpu)->migration_thread)
5424 break;
5425 /* Unbind it from offline cpu so it can run. Fall thru. */
5426 kthread_bind(cpu_rq(cpu)->migration_thread,
5427 any_online_cpu(cpu_online_map));
5428 kthread_stop(cpu_rq(cpu)->migration_thread);
5429 cpu_rq(cpu)->migration_thread = NULL;
5430 break;
5432 case CPU_DEAD:
5433 case CPU_DEAD_FROZEN:
5434 migrate_live_tasks(cpu);
5435 rq = cpu_rq(cpu);
5436 kthread_stop(rq->migration_thread);
5437 rq->migration_thread = NULL;
5438 /* Idle task back to normal (off runqueue, low prio) */
5439 rq = task_rq_lock(rq->idle, &flags);
5440 update_rq_clock(rq);
5441 deactivate_task(rq, rq->idle, 0);
5442 rq->idle->static_prio = MAX_PRIO;
5443 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5444 rq->idle->sched_class = &idle_sched_class;
5445 migrate_dead_tasks(cpu);
5446 task_rq_unlock(rq, &flags);
5447 migrate_nr_uninterruptible(rq);
5448 BUG_ON(rq->nr_running != 0);
5450 /* No need to migrate the tasks: it was best-effort if
5451 * they didn't take sched_hotcpu_mutex. Just wake up
5452 * the requestors. */
5453 spin_lock_irq(&rq->lock);
5454 while (!list_empty(&rq->migration_queue)) {
5455 struct migration_req *req;
5457 req = list_entry(rq->migration_queue.next,
5458 struct migration_req, list);
5459 list_del_init(&req->list);
5460 complete(&req->done);
5462 spin_unlock_irq(&rq->lock);
5463 break;
5464 #endif
5465 case CPU_LOCK_RELEASE:
5466 mutex_unlock(&sched_hotcpu_mutex);
5467 break;
5469 return NOTIFY_OK;
5472 /* Register at highest priority so that task migration (migrate_all_tasks)
5473 * happens before everything else.
5475 static struct notifier_block __cpuinitdata migration_notifier = {
5476 .notifier_call = migration_call,
5477 .priority = 10
5480 int __init migration_init(void)
5482 void *cpu = (void *)(long)smp_processor_id();
5483 int err;
5485 /* Start one for the boot CPU: */
5486 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5487 BUG_ON(err == NOTIFY_BAD);
5488 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5489 register_cpu_notifier(&migration_notifier);
5491 return 0;
5493 #endif
5495 #ifdef CONFIG_SMP
5497 /* Number of possible processor ids */
5498 int nr_cpu_ids __read_mostly = NR_CPUS;
5499 EXPORT_SYMBOL(nr_cpu_ids);
5501 #undef SCHED_DOMAIN_DEBUG
5502 #ifdef SCHED_DOMAIN_DEBUG
5503 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5505 int level = 0;
5507 if (!sd) {
5508 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5509 return;
5512 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5514 do {
5515 int i;
5516 char str[NR_CPUS];
5517 struct sched_group *group = sd->groups;
5518 cpumask_t groupmask;
5520 cpumask_scnprintf(str, NR_CPUS, sd->span);
5521 cpus_clear(groupmask);
5523 printk(KERN_DEBUG);
5524 for (i = 0; i < level + 1; i++)
5525 printk(" ");
5526 printk("domain %d: ", level);
5528 if (!(sd->flags & SD_LOAD_BALANCE)) {
5529 printk("does not load-balance\n");
5530 if (sd->parent)
5531 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5532 " has parent");
5533 break;
5536 printk("span %s\n", str);
5538 if (!cpu_isset(cpu, sd->span))
5539 printk(KERN_ERR "ERROR: domain->span does not contain "
5540 "CPU%d\n", cpu);
5541 if (!cpu_isset(cpu, group->cpumask))
5542 printk(KERN_ERR "ERROR: domain->groups does not contain"
5543 " CPU%d\n", cpu);
5545 printk(KERN_DEBUG);
5546 for (i = 0; i < level + 2; i++)
5547 printk(" ");
5548 printk("groups:");
5549 do {
5550 if (!group) {
5551 printk("\n");
5552 printk(KERN_ERR "ERROR: group is NULL\n");
5553 break;
5556 if (!group->__cpu_power) {
5557 printk("\n");
5558 printk(KERN_ERR "ERROR: domain->cpu_power not "
5559 "set\n");
5562 if (!cpus_weight(group->cpumask)) {
5563 printk("\n");
5564 printk(KERN_ERR "ERROR: empty group\n");
5567 if (cpus_intersects(groupmask, group->cpumask)) {
5568 printk("\n");
5569 printk(KERN_ERR "ERROR: repeated CPUs\n");
5572 cpus_or(groupmask, groupmask, group->cpumask);
5574 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5575 printk(" %s", str);
5577 group = group->next;
5578 } while (group != sd->groups);
5579 printk("\n");
5581 if (!cpus_equal(sd->span, groupmask))
5582 printk(KERN_ERR "ERROR: groups don't span "
5583 "domain->span\n");
5585 level++;
5586 sd = sd->parent;
5587 if (!sd)
5588 continue;
5590 if (!cpus_subset(groupmask, sd->span))
5591 printk(KERN_ERR "ERROR: parent span is not a superset "
5592 "of domain->span\n");
5594 } while (sd);
5596 #else
5597 # define sched_domain_debug(sd, cpu) do { } while (0)
5598 #endif
5600 static int sd_degenerate(struct sched_domain *sd)
5602 if (cpus_weight(sd->span) == 1)
5603 return 1;
5605 /* Following flags need at least 2 groups */
5606 if (sd->flags & (SD_LOAD_BALANCE |
5607 SD_BALANCE_NEWIDLE |
5608 SD_BALANCE_FORK |
5609 SD_BALANCE_EXEC |
5610 SD_SHARE_CPUPOWER |
5611 SD_SHARE_PKG_RESOURCES)) {
5612 if (sd->groups != sd->groups->next)
5613 return 0;
5616 /* Following flags don't use groups */
5617 if (sd->flags & (SD_WAKE_IDLE |
5618 SD_WAKE_AFFINE |
5619 SD_WAKE_BALANCE))
5620 return 0;
5622 return 1;
5625 static int
5626 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5628 unsigned long cflags = sd->flags, pflags = parent->flags;
5630 if (sd_degenerate(parent))
5631 return 1;
5633 if (!cpus_equal(sd->span, parent->span))
5634 return 0;
5636 /* Does parent contain flags not in child? */
5637 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5638 if (cflags & SD_WAKE_AFFINE)
5639 pflags &= ~SD_WAKE_BALANCE;
5640 /* Flags needing groups don't count if only 1 group in parent */
5641 if (parent->groups == parent->groups->next) {
5642 pflags &= ~(SD_LOAD_BALANCE |
5643 SD_BALANCE_NEWIDLE |
5644 SD_BALANCE_FORK |
5645 SD_BALANCE_EXEC |
5646 SD_SHARE_CPUPOWER |
5647 SD_SHARE_PKG_RESOURCES);
5649 if (~cflags & pflags)
5650 return 0;
5652 return 1;
5656 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5657 * hold the hotplug lock.
5659 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5661 struct rq *rq = cpu_rq(cpu);
5662 struct sched_domain *tmp;
5664 /* Remove the sched domains which do not contribute to scheduling. */
5665 for (tmp = sd; tmp; tmp = tmp->parent) {
5666 struct sched_domain *parent = tmp->parent;
5667 if (!parent)
5668 break;
5669 if (sd_parent_degenerate(tmp, parent)) {
5670 tmp->parent = parent->parent;
5671 if (parent->parent)
5672 parent->parent->child = tmp;
5676 if (sd && sd_degenerate(sd)) {
5677 sd = sd->parent;
5678 if (sd)
5679 sd->child = NULL;
5682 sched_domain_debug(sd, cpu);
5684 rcu_assign_pointer(rq->sd, sd);
5687 /* cpus with isolated domains */
5688 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5690 /* Setup the mask of cpus configured for isolated domains */
5691 static int __init isolated_cpu_setup(char *str)
5693 int ints[NR_CPUS], i;
5695 str = get_options(str, ARRAY_SIZE(ints), ints);
5696 cpus_clear(cpu_isolated_map);
5697 for (i = 1; i <= ints[0]; i++)
5698 if (ints[i] < NR_CPUS)
5699 cpu_set(ints[i], cpu_isolated_map);
5700 return 1;
5703 __setup ("isolcpus=", isolated_cpu_setup);
5706 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5707 * to a function which identifies what group(along with sched group) a CPU
5708 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5709 * (due to the fact that we keep track of groups covered with a cpumask_t).
5711 * init_sched_build_groups will build a circular linked list of the groups
5712 * covered by the given span, and will set each group's ->cpumask correctly,
5713 * and ->cpu_power to 0.
5715 static void
5716 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5717 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5718 struct sched_group **sg))
5720 struct sched_group *first = NULL, *last = NULL;
5721 cpumask_t covered = CPU_MASK_NONE;
5722 int i;
5724 for_each_cpu_mask(i, span) {
5725 struct sched_group *sg;
5726 int group = group_fn(i, cpu_map, &sg);
5727 int j;
5729 if (cpu_isset(i, covered))
5730 continue;
5732 sg->cpumask = CPU_MASK_NONE;
5733 sg->__cpu_power = 0;
5735 for_each_cpu_mask(j, span) {
5736 if (group_fn(j, cpu_map, NULL) != group)
5737 continue;
5739 cpu_set(j, covered);
5740 cpu_set(j, sg->cpumask);
5742 if (!first)
5743 first = sg;
5744 if (last)
5745 last->next = sg;
5746 last = sg;
5748 last->next = first;
5751 #define SD_NODES_PER_DOMAIN 16
5753 #ifdef CONFIG_NUMA
5756 * find_next_best_node - find the next node to include in a sched_domain
5757 * @node: node whose sched_domain we're building
5758 * @used_nodes: nodes already in the sched_domain
5760 * Find the next node to include in a given scheduling domain. Simply
5761 * finds the closest node not already in the @used_nodes map.
5763 * Should use nodemask_t.
5765 static int find_next_best_node(int node, unsigned long *used_nodes)
5767 int i, n, val, min_val, best_node = 0;
5769 min_val = INT_MAX;
5771 for (i = 0; i < MAX_NUMNODES; i++) {
5772 /* Start at @node */
5773 n = (node + i) % MAX_NUMNODES;
5775 if (!nr_cpus_node(n))
5776 continue;
5778 /* Skip already used nodes */
5779 if (test_bit(n, used_nodes))
5780 continue;
5782 /* Simple min distance search */
5783 val = node_distance(node, n);
5785 if (val < min_val) {
5786 min_val = val;
5787 best_node = n;
5791 set_bit(best_node, used_nodes);
5792 return best_node;
5796 * sched_domain_node_span - get a cpumask for a node's sched_domain
5797 * @node: node whose cpumask we're constructing
5798 * @size: number of nodes to include in this span
5800 * Given a node, construct a good cpumask for its sched_domain to span. It
5801 * should be one that prevents unnecessary balancing, but also spreads tasks
5802 * out optimally.
5804 static cpumask_t sched_domain_node_span(int node)
5806 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5807 cpumask_t span, nodemask;
5808 int i;
5810 cpus_clear(span);
5811 bitmap_zero(used_nodes, MAX_NUMNODES);
5813 nodemask = node_to_cpumask(node);
5814 cpus_or(span, span, nodemask);
5815 set_bit(node, used_nodes);
5817 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5818 int next_node = find_next_best_node(node, used_nodes);
5820 nodemask = node_to_cpumask(next_node);
5821 cpus_or(span, span, nodemask);
5824 return span;
5826 #endif
5828 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5831 * SMT sched-domains:
5833 #ifdef CONFIG_SCHED_SMT
5834 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5835 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5837 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5838 struct sched_group **sg)
5840 if (sg)
5841 *sg = &per_cpu(sched_group_cpus, cpu);
5842 return cpu;
5844 #endif
5847 * multi-core sched-domains:
5849 #ifdef CONFIG_SCHED_MC
5850 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5851 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5852 #endif
5854 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5855 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5856 struct sched_group **sg)
5858 int group;
5859 cpumask_t mask = cpu_sibling_map[cpu];
5860 cpus_and(mask, mask, *cpu_map);
5861 group = first_cpu(mask);
5862 if (sg)
5863 *sg = &per_cpu(sched_group_core, group);
5864 return group;
5866 #elif defined(CONFIG_SCHED_MC)
5867 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5868 struct sched_group **sg)
5870 if (sg)
5871 *sg = &per_cpu(sched_group_core, cpu);
5872 return cpu;
5874 #endif
5876 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5877 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5879 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5880 struct sched_group **sg)
5882 int group;
5883 #ifdef CONFIG_SCHED_MC
5884 cpumask_t mask = cpu_coregroup_map(cpu);
5885 cpus_and(mask, mask, *cpu_map);
5886 group = first_cpu(mask);
5887 #elif defined(CONFIG_SCHED_SMT)
5888 cpumask_t mask = cpu_sibling_map[cpu];
5889 cpus_and(mask, mask, *cpu_map);
5890 group = first_cpu(mask);
5891 #else
5892 group = cpu;
5893 #endif
5894 if (sg)
5895 *sg = &per_cpu(sched_group_phys, group);
5896 return group;
5899 #ifdef CONFIG_NUMA
5901 * The init_sched_build_groups can't handle what we want to do with node
5902 * groups, so roll our own. Now each node has its own list of groups which
5903 * gets dynamically allocated.
5905 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5906 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5908 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5909 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5911 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5912 struct sched_group **sg)
5914 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5915 int group;
5917 cpus_and(nodemask, nodemask, *cpu_map);
5918 group = first_cpu(nodemask);
5920 if (sg)
5921 *sg = &per_cpu(sched_group_allnodes, group);
5922 return group;
5925 static void init_numa_sched_groups_power(struct sched_group *group_head)
5927 struct sched_group *sg = group_head;
5928 int j;
5930 if (!sg)
5931 return;
5932 next_sg:
5933 for_each_cpu_mask(j, sg->cpumask) {
5934 struct sched_domain *sd;
5936 sd = &per_cpu(phys_domains, j);
5937 if (j != first_cpu(sd->groups->cpumask)) {
5939 * Only add "power" once for each
5940 * physical package.
5942 continue;
5945 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5947 sg = sg->next;
5948 if (sg != group_head)
5949 goto next_sg;
5951 #endif
5953 #ifdef CONFIG_NUMA
5954 /* Free memory allocated for various sched_group structures */
5955 static void free_sched_groups(const cpumask_t *cpu_map)
5957 int cpu, i;
5959 for_each_cpu_mask(cpu, *cpu_map) {
5960 struct sched_group **sched_group_nodes
5961 = sched_group_nodes_bycpu[cpu];
5963 if (!sched_group_nodes)
5964 continue;
5966 for (i = 0; i < MAX_NUMNODES; i++) {
5967 cpumask_t nodemask = node_to_cpumask(i);
5968 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5970 cpus_and(nodemask, nodemask, *cpu_map);
5971 if (cpus_empty(nodemask))
5972 continue;
5974 if (sg == NULL)
5975 continue;
5976 sg = sg->next;
5977 next_sg:
5978 oldsg = sg;
5979 sg = sg->next;
5980 kfree(oldsg);
5981 if (oldsg != sched_group_nodes[i])
5982 goto next_sg;
5984 kfree(sched_group_nodes);
5985 sched_group_nodes_bycpu[cpu] = NULL;
5988 #else
5989 static void free_sched_groups(const cpumask_t *cpu_map)
5992 #endif
5995 * Initialize sched groups cpu_power.
5997 * cpu_power indicates the capacity of sched group, which is used while
5998 * distributing the load between different sched groups in a sched domain.
5999 * Typically cpu_power for all the groups in a sched domain will be same unless
6000 * there are asymmetries in the topology. If there are asymmetries, group
6001 * having more cpu_power will pickup more load compared to the group having
6002 * less cpu_power.
6004 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6005 * the maximum number of tasks a group can handle in the presence of other idle
6006 * or lightly loaded groups in the same sched domain.
6008 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6010 struct sched_domain *child;
6011 struct sched_group *group;
6013 WARN_ON(!sd || !sd->groups);
6015 if (cpu != first_cpu(sd->groups->cpumask))
6016 return;
6018 child = sd->child;
6020 sd->groups->__cpu_power = 0;
6023 * For perf policy, if the groups in child domain share resources
6024 * (for example cores sharing some portions of the cache hierarchy
6025 * or SMT), then set this domain groups cpu_power such that each group
6026 * can handle only one task, when there are other idle groups in the
6027 * same sched domain.
6029 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6030 (child->flags &
6031 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6032 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6033 return;
6037 * add cpu_power of each child group to this groups cpu_power
6039 group = child->groups;
6040 do {
6041 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6042 group = group->next;
6043 } while (group != child->groups);
6047 * Build sched domains for a given set of cpus and attach the sched domains
6048 * to the individual cpus
6050 static int build_sched_domains(const cpumask_t *cpu_map)
6052 int i;
6053 #ifdef CONFIG_NUMA
6054 struct sched_group **sched_group_nodes = NULL;
6055 int sd_allnodes = 0;
6058 * Allocate the per-node list of sched groups
6060 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
6061 GFP_KERNEL);
6062 if (!sched_group_nodes) {
6063 printk(KERN_WARNING "Can not alloc sched group node list\n");
6064 return -ENOMEM;
6066 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6067 #endif
6070 * Set up domains for cpus specified by the cpu_map.
6072 for_each_cpu_mask(i, *cpu_map) {
6073 struct sched_domain *sd = NULL, *p;
6074 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6076 cpus_and(nodemask, nodemask, *cpu_map);
6078 #ifdef CONFIG_NUMA
6079 if (cpus_weight(*cpu_map) >
6080 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6081 sd = &per_cpu(allnodes_domains, i);
6082 *sd = SD_ALLNODES_INIT;
6083 sd->span = *cpu_map;
6084 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6085 p = sd;
6086 sd_allnodes = 1;
6087 } else
6088 p = NULL;
6090 sd = &per_cpu(node_domains, i);
6091 *sd = SD_NODE_INIT;
6092 sd->span = sched_domain_node_span(cpu_to_node(i));
6093 sd->parent = p;
6094 if (p)
6095 p->child = sd;
6096 cpus_and(sd->span, sd->span, *cpu_map);
6097 #endif
6099 p = sd;
6100 sd = &per_cpu(phys_domains, i);
6101 *sd = SD_CPU_INIT;
6102 sd->span = nodemask;
6103 sd->parent = p;
6104 if (p)
6105 p->child = sd;
6106 cpu_to_phys_group(i, cpu_map, &sd->groups);
6108 #ifdef CONFIG_SCHED_MC
6109 p = sd;
6110 sd = &per_cpu(core_domains, i);
6111 *sd = SD_MC_INIT;
6112 sd->span = cpu_coregroup_map(i);
6113 cpus_and(sd->span, sd->span, *cpu_map);
6114 sd->parent = p;
6115 p->child = sd;
6116 cpu_to_core_group(i, cpu_map, &sd->groups);
6117 #endif
6119 #ifdef CONFIG_SCHED_SMT
6120 p = sd;
6121 sd = &per_cpu(cpu_domains, i);
6122 *sd = SD_SIBLING_INIT;
6123 sd->span = cpu_sibling_map[i];
6124 cpus_and(sd->span, sd->span, *cpu_map);
6125 sd->parent = p;
6126 p->child = sd;
6127 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6128 #endif
6131 #ifdef CONFIG_SCHED_SMT
6132 /* Set up CPU (sibling) groups */
6133 for_each_cpu_mask(i, *cpu_map) {
6134 cpumask_t this_sibling_map = cpu_sibling_map[i];
6135 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6136 if (i != first_cpu(this_sibling_map))
6137 continue;
6139 init_sched_build_groups(this_sibling_map, cpu_map,
6140 &cpu_to_cpu_group);
6142 #endif
6144 #ifdef CONFIG_SCHED_MC
6145 /* Set up multi-core groups */
6146 for_each_cpu_mask(i, *cpu_map) {
6147 cpumask_t this_core_map = cpu_coregroup_map(i);
6148 cpus_and(this_core_map, this_core_map, *cpu_map);
6149 if (i != first_cpu(this_core_map))
6150 continue;
6151 init_sched_build_groups(this_core_map, cpu_map,
6152 &cpu_to_core_group);
6154 #endif
6156 /* Set up physical groups */
6157 for (i = 0; i < MAX_NUMNODES; i++) {
6158 cpumask_t nodemask = node_to_cpumask(i);
6160 cpus_and(nodemask, nodemask, *cpu_map);
6161 if (cpus_empty(nodemask))
6162 continue;
6164 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6167 #ifdef CONFIG_NUMA
6168 /* Set up node groups */
6169 if (sd_allnodes)
6170 init_sched_build_groups(*cpu_map, cpu_map,
6171 &cpu_to_allnodes_group);
6173 for (i = 0; i < MAX_NUMNODES; i++) {
6174 /* Set up node groups */
6175 struct sched_group *sg, *prev;
6176 cpumask_t nodemask = node_to_cpumask(i);
6177 cpumask_t domainspan;
6178 cpumask_t covered = CPU_MASK_NONE;
6179 int j;
6181 cpus_and(nodemask, nodemask, *cpu_map);
6182 if (cpus_empty(nodemask)) {
6183 sched_group_nodes[i] = NULL;
6184 continue;
6187 domainspan = sched_domain_node_span(i);
6188 cpus_and(domainspan, domainspan, *cpu_map);
6190 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6191 if (!sg) {
6192 printk(KERN_WARNING "Can not alloc domain group for "
6193 "node %d\n", i);
6194 goto error;
6196 sched_group_nodes[i] = sg;
6197 for_each_cpu_mask(j, nodemask) {
6198 struct sched_domain *sd;
6200 sd = &per_cpu(node_domains, j);
6201 sd->groups = sg;
6203 sg->__cpu_power = 0;
6204 sg->cpumask = nodemask;
6205 sg->next = sg;
6206 cpus_or(covered, covered, nodemask);
6207 prev = sg;
6209 for (j = 0; j < MAX_NUMNODES; j++) {
6210 cpumask_t tmp, notcovered;
6211 int n = (i + j) % MAX_NUMNODES;
6213 cpus_complement(notcovered, covered);
6214 cpus_and(tmp, notcovered, *cpu_map);
6215 cpus_and(tmp, tmp, domainspan);
6216 if (cpus_empty(tmp))
6217 break;
6219 nodemask = node_to_cpumask(n);
6220 cpus_and(tmp, tmp, nodemask);
6221 if (cpus_empty(tmp))
6222 continue;
6224 sg = kmalloc_node(sizeof(struct sched_group),
6225 GFP_KERNEL, i);
6226 if (!sg) {
6227 printk(KERN_WARNING
6228 "Can not alloc domain group for node %d\n", j);
6229 goto error;
6231 sg->__cpu_power = 0;
6232 sg->cpumask = tmp;
6233 sg->next = prev->next;
6234 cpus_or(covered, covered, tmp);
6235 prev->next = sg;
6236 prev = sg;
6239 #endif
6241 /* Calculate CPU power for physical packages and nodes */
6242 #ifdef CONFIG_SCHED_SMT
6243 for_each_cpu_mask(i, *cpu_map) {
6244 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6246 init_sched_groups_power(i, sd);
6248 #endif
6249 #ifdef CONFIG_SCHED_MC
6250 for_each_cpu_mask(i, *cpu_map) {
6251 struct sched_domain *sd = &per_cpu(core_domains, i);
6253 init_sched_groups_power(i, sd);
6255 #endif
6257 for_each_cpu_mask(i, *cpu_map) {
6258 struct sched_domain *sd = &per_cpu(phys_domains, i);
6260 init_sched_groups_power(i, sd);
6263 #ifdef CONFIG_NUMA
6264 for (i = 0; i < MAX_NUMNODES; i++)
6265 init_numa_sched_groups_power(sched_group_nodes[i]);
6267 if (sd_allnodes) {
6268 struct sched_group *sg;
6270 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6271 init_numa_sched_groups_power(sg);
6273 #endif
6275 /* Attach the domains */
6276 for_each_cpu_mask(i, *cpu_map) {
6277 struct sched_domain *sd;
6278 #ifdef CONFIG_SCHED_SMT
6279 sd = &per_cpu(cpu_domains, i);
6280 #elif defined(CONFIG_SCHED_MC)
6281 sd = &per_cpu(core_domains, i);
6282 #else
6283 sd = &per_cpu(phys_domains, i);
6284 #endif
6285 cpu_attach_domain(sd, i);
6288 return 0;
6290 #ifdef CONFIG_NUMA
6291 error:
6292 free_sched_groups(cpu_map);
6293 return -ENOMEM;
6294 #endif
6297 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6299 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6301 cpumask_t cpu_default_map;
6302 int err;
6305 * Setup mask for cpus without special case scheduling requirements.
6306 * For now this just excludes isolated cpus, but could be used to
6307 * exclude other special cases in the future.
6309 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6311 err = build_sched_domains(&cpu_default_map);
6313 return err;
6316 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6318 free_sched_groups(cpu_map);
6322 * Detach sched domains from a group of cpus specified in cpu_map
6323 * These cpus will now be attached to the NULL domain
6325 static void detach_destroy_domains(const cpumask_t *cpu_map)
6327 int i;
6329 for_each_cpu_mask(i, *cpu_map)
6330 cpu_attach_domain(NULL, i);
6331 synchronize_sched();
6332 arch_destroy_sched_domains(cpu_map);
6336 * Partition sched domains as specified by the cpumasks below.
6337 * This attaches all cpus from the cpumasks to the NULL domain,
6338 * waits for a RCU quiescent period, recalculates sched
6339 * domain information and then attaches them back to the
6340 * correct sched domains
6341 * Call with hotplug lock held
6343 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6345 cpumask_t change_map;
6346 int err = 0;
6348 cpus_and(*partition1, *partition1, cpu_online_map);
6349 cpus_and(*partition2, *partition2, cpu_online_map);
6350 cpus_or(change_map, *partition1, *partition2);
6352 /* Detach sched domains from all of the affected cpus */
6353 detach_destroy_domains(&change_map);
6354 if (!cpus_empty(*partition1))
6355 err = build_sched_domains(partition1);
6356 if (!err && !cpus_empty(*partition2))
6357 err = build_sched_domains(partition2);
6359 return err;
6362 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6363 static int arch_reinit_sched_domains(void)
6365 int err;
6367 mutex_lock(&sched_hotcpu_mutex);
6368 detach_destroy_domains(&cpu_online_map);
6369 err = arch_init_sched_domains(&cpu_online_map);
6370 mutex_unlock(&sched_hotcpu_mutex);
6372 return err;
6375 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6377 int ret;
6379 if (buf[0] != '0' && buf[0] != '1')
6380 return -EINVAL;
6382 if (smt)
6383 sched_smt_power_savings = (buf[0] == '1');
6384 else
6385 sched_mc_power_savings = (buf[0] == '1');
6387 ret = arch_reinit_sched_domains();
6389 return ret ? ret : count;
6392 #ifdef CONFIG_SCHED_MC
6393 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6395 return sprintf(page, "%u\n", sched_mc_power_savings);
6397 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6398 const char *buf, size_t count)
6400 return sched_power_savings_store(buf, count, 0);
6402 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6403 sched_mc_power_savings_store);
6404 #endif
6406 #ifdef CONFIG_SCHED_SMT
6407 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6409 return sprintf(page, "%u\n", sched_smt_power_savings);
6411 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6412 const char *buf, size_t count)
6414 return sched_power_savings_store(buf, count, 1);
6416 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6417 sched_smt_power_savings_store);
6418 #endif
6420 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6422 int err = 0;
6424 #ifdef CONFIG_SCHED_SMT
6425 if (smt_capable())
6426 err = sysfs_create_file(&cls->kset.kobj,
6427 &attr_sched_smt_power_savings.attr);
6428 #endif
6429 #ifdef CONFIG_SCHED_MC
6430 if (!err && mc_capable())
6431 err = sysfs_create_file(&cls->kset.kobj,
6432 &attr_sched_mc_power_savings.attr);
6433 #endif
6434 return err;
6436 #endif
6439 * Force a reinitialization of the sched domains hierarchy. The domains
6440 * and groups cannot be updated in place without racing with the balancing
6441 * code, so we temporarily attach all running cpus to the NULL domain
6442 * which will prevent rebalancing while the sched domains are recalculated.
6444 static int update_sched_domains(struct notifier_block *nfb,
6445 unsigned long action, void *hcpu)
6447 switch (action) {
6448 case CPU_UP_PREPARE:
6449 case CPU_UP_PREPARE_FROZEN:
6450 case CPU_DOWN_PREPARE:
6451 case CPU_DOWN_PREPARE_FROZEN:
6452 detach_destroy_domains(&cpu_online_map);
6453 return NOTIFY_OK;
6455 case CPU_UP_CANCELED:
6456 case CPU_UP_CANCELED_FROZEN:
6457 case CPU_DOWN_FAILED:
6458 case CPU_DOWN_FAILED_FROZEN:
6459 case CPU_ONLINE:
6460 case CPU_ONLINE_FROZEN:
6461 case CPU_DEAD:
6462 case CPU_DEAD_FROZEN:
6464 * Fall through and re-initialise the domains.
6466 break;
6467 default:
6468 return NOTIFY_DONE;
6471 /* The hotplug lock is already held by cpu_up/cpu_down */
6472 arch_init_sched_domains(&cpu_online_map);
6474 return NOTIFY_OK;
6477 void __init sched_init_smp(void)
6479 cpumask_t non_isolated_cpus;
6481 mutex_lock(&sched_hotcpu_mutex);
6482 arch_init_sched_domains(&cpu_online_map);
6483 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6484 if (cpus_empty(non_isolated_cpus))
6485 cpu_set(smp_processor_id(), non_isolated_cpus);
6486 mutex_unlock(&sched_hotcpu_mutex);
6487 /* XXX: Theoretical race here - CPU may be hotplugged now */
6488 hotcpu_notifier(update_sched_domains, 0);
6490 init_sched_domain_sysctl();
6492 /* Move init over to a non-isolated CPU */
6493 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6494 BUG();
6495 sched_init_granularity();
6497 #else
6498 void __init sched_init_smp(void)
6500 sched_init_granularity();
6502 #endif /* CONFIG_SMP */
6504 int in_sched_functions(unsigned long addr)
6506 /* Linker adds these: start and end of __sched functions */
6507 extern char __sched_text_start[], __sched_text_end[];
6509 return in_lock_functions(addr) ||
6510 (addr >= (unsigned long)__sched_text_start
6511 && addr < (unsigned long)__sched_text_end);
6514 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6516 cfs_rq->tasks_timeline = RB_ROOT;
6517 cfs_rq->fair_clock = 1;
6518 #ifdef CONFIG_FAIR_GROUP_SCHED
6519 cfs_rq->rq = rq;
6520 #endif
6523 void __init sched_init(void)
6525 u64 now = sched_clock();
6526 int highest_cpu = 0;
6527 int i, j;
6530 * Link up the scheduling class hierarchy:
6532 rt_sched_class.next = &fair_sched_class;
6533 fair_sched_class.next = &idle_sched_class;
6534 idle_sched_class.next = NULL;
6536 for_each_possible_cpu(i) {
6537 struct rt_prio_array *array;
6538 struct rq *rq;
6540 rq = cpu_rq(i);
6541 spin_lock_init(&rq->lock);
6542 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6543 rq->nr_running = 0;
6544 rq->clock = 1;
6545 init_cfs_rq(&rq->cfs, rq);
6546 #ifdef CONFIG_FAIR_GROUP_SCHED
6547 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6548 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6549 #endif
6550 rq->ls.load_update_last = now;
6551 rq->ls.load_update_start = now;
6553 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6554 rq->cpu_load[j] = 0;
6555 #ifdef CONFIG_SMP
6556 rq->sd = NULL;
6557 rq->active_balance = 0;
6558 rq->next_balance = jiffies;
6559 rq->push_cpu = 0;
6560 rq->cpu = i;
6561 rq->migration_thread = NULL;
6562 INIT_LIST_HEAD(&rq->migration_queue);
6563 #endif
6564 atomic_set(&rq->nr_iowait, 0);
6566 array = &rq->rt.active;
6567 for (j = 0; j < MAX_RT_PRIO; j++) {
6568 INIT_LIST_HEAD(array->queue + j);
6569 __clear_bit(j, array->bitmap);
6571 highest_cpu = i;
6572 /* delimiter for bitsearch: */
6573 __set_bit(MAX_RT_PRIO, array->bitmap);
6576 set_load_weight(&init_task);
6578 #ifdef CONFIG_PREEMPT_NOTIFIERS
6579 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6580 #endif
6582 #ifdef CONFIG_SMP
6583 nr_cpu_ids = highest_cpu + 1;
6584 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6585 #endif
6587 #ifdef CONFIG_RT_MUTEXES
6588 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6589 #endif
6592 * The boot idle thread does lazy MMU switching as well:
6594 atomic_inc(&init_mm.mm_count);
6595 enter_lazy_tlb(&init_mm, current);
6598 * Make us the idle thread. Technically, schedule() should not be
6599 * called from this thread, however somewhere below it might be,
6600 * but because we are the idle thread, we just pick up running again
6601 * when this runqueue becomes "idle".
6603 init_idle(current, smp_processor_id());
6605 * During early bootup we pretend to be a normal task:
6607 current->sched_class = &fair_sched_class;
6610 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6611 void __might_sleep(char *file, int line)
6613 #ifdef in_atomic
6614 static unsigned long prev_jiffy; /* ratelimiting */
6616 if ((in_atomic() || irqs_disabled()) &&
6617 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6618 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6619 return;
6620 prev_jiffy = jiffies;
6621 printk(KERN_ERR "BUG: sleeping function called from invalid"
6622 " context at %s:%d\n", file, line);
6623 printk("in_atomic():%d, irqs_disabled():%d\n",
6624 in_atomic(), irqs_disabled());
6625 debug_show_held_locks(current);
6626 if (irqs_disabled())
6627 print_irqtrace_events(current);
6628 dump_stack();
6630 #endif
6632 EXPORT_SYMBOL(__might_sleep);
6633 #endif
6635 #ifdef CONFIG_MAGIC_SYSRQ
6636 void normalize_rt_tasks(void)
6638 struct task_struct *g, *p;
6639 unsigned long flags;
6640 struct rq *rq;
6641 int on_rq;
6643 read_lock_irq(&tasklist_lock);
6644 do_each_thread(g, p) {
6645 p->se.fair_key = 0;
6646 p->se.wait_runtime = 0;
6647 p->se.exec_start = 0;
6648 p->se.wait_start_fair = 0;
6649 p->se.sleep_start_fair = 0;
6650 #ifdef CONFIG_SCHEDSTATS
6651 p->se.wait_start = 0;
6652 p->se.sleep_start = 0;
6653 p->se.block_start = 0;
6654 #endif
6655 task_rq(p)->cfs.fair_clock = 0;
6656 task_rq(p)->clock = 0;
6658 if (!rt_task(p)) {
6660 * Renice negative nice level userspace
6661 * tasks back to 0:
6663 if (TASK_NICE(p) < 0 && p->mm)
6664 set_user_nice(p, 0);
6665 continue;
6668 spin_lock_irqsave(&p->pi_lock, flags);
6669 rq = __task_rq_lock(p);
6670 #ifdef CONFIG_SMP
6672 * Do not touch the migration thread:
6674 if (p == rq->migration_thread)
6675 goto out_unlock;
6676 #endif
6678 update_rq_clock(rq);
6679 on_rq = p->se.on_rq;
6680 if (on_rq)
6681 deactivate_task(rq, p, 0);
6682 __setscheduler(rq, p, SCHED_NORMAL, 0);
6683 if (on_rq) {
6684 activate_task(rq, p, 0);
6685 resched_task(rq->curr);
6687 #ifdef CONFIG_SMP
6688 out_unlock:
6689 #endif
6690 __task_rq_unlock(rq);
6691 spin_unlock_irqrestore(&p->pi_lock, flags);
6692 } while_each_thread(g, p);
6694 read_unlock_irq(&tasklist_lock);
6697 #endif /* CONFIG_MAGIC_SYSRQ */
6699 #ifdef CONFIG_IA64
6701 * These functions are only useful for the IA64 MCA handling.
6703 * They can only be called when the whole system has been
6704 * stopped - every CPU needs to be quiescent, and no scheduling
6705 * activity can take place. Using them for anything else would
6706 * be a serious bug, and as a result, they aren't even visible
6707 * under any other configuration.
6711 * curr_task - return the current task for a given cpu.
6712 * @cpu: the processor in question.
6714 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6716 struct task_struct *curr_task(int cpu)
6718 return cpu_curr(cpu);
6722 * set_curr_task - set the current task for a given cpu.
6723 * @cpu: the processor in question.
6724 * @p: the task pointer to set.
6726 * Description: This function must only be used when non-maskable interrupts
6727 * are serviced on a separate stack. It allows the architecture to switch the
6728 * notion of the current task on a cpu in a non-blocking manner. This function
6729 * must be called with all CPU's synchronized, and interrupts disabled, the
6730 * and caller must save the original value of the current task (see
6731 * curr_task() above) and restore that value before reenabling interrupts and
6732 * re-starting the system.
6734 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6736 void set_curr_task(int cpu, struct task_struct *p)
6738 cpu_curr(cpu) = p;
6741 #endif