Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jmorris...
[linux-2.6/verdex.git] / kernel / sched.c
blobdcd553cc4ee89b52ec511f70361c4f7fd586d976
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
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
70 #include <asm/tlb.h>
71 #include <asm/irq_regs.h>
74 * Scheduler clock - returns current time in nanosec units.
75 * This is default implementation.
76 * Architectures and sub-architectures can override this.
78 unsigned long long __attribute__((weak)) sched_clock(void)
80 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
86 * and back.
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
117 #ifdef CONFIG_SMP
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
124 return reciprocal_divide(load, sg->reciprocal_cpu_power);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
133 sg->__cpu_power += val;
134 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 #endif
138 static inline int rt_policy(int policy)
140 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
141 return 1;
142 return 0;
145 static inline int task_has_rt_policy(struct task_struct *p)
147 return rt_policy(p->policy);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array {
154 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
155 struct list_head queue[MAX_RT_PRIO];
158 #ifdef CONFIG_GROUP_SCHED
160 #include <linux/cgroup.h>
162 struct cfs_rq;
164 static LIST_HEAD(task_groups);
166 /* task group related information */
167 struct task_group {
168 #ifdef CONFIG_CGROUP_SCHED
169 struct cgroup_subsys_state css;
170 #endif
172 #ifdef CONFIG_FAIR_GROUP_SCHED
173 /* schedulable entities of this group on each cpu */
174 struct sched_entity **se;
175 /* runqueue "owned" by this group on each cpu */
176 struct cfs_rq **cfs_rq;
177 unsigned long shares;
178 #endif
180 #ifdef CONFIG_RT_GROUP_SCHED
181 struct sched_rt_entity **rt_se;
182 struct rt_rq **rt_rq;
184 u64 rt_runtime;
185 #endif
187 struct rcu_head rcu;
188 struct list_head list;
191 #ifdef CONFIG_FAIR_GROUP_SCHED
192 /* Default task group's sched entity on each cpu */
193 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
194 /* Default task group's cfs_rq on each cpu */
195 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
197 static struct sched_entity *init_sched_entity_p[NR_CPUS];
198 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
199 #endif
201 #ifdef CONFIG_RT_GROUP_SCHED
202 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
203 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
205 static struct sched_rt_entity *init_sched_rt_entity_p[NR_CPUS];
206 static struct rt_rq *init_rt_rq_p[NR_CPUS];
207 #endif
209 /* task_group_lock serializes add/remove of task groups and also changes to
210 * a task group's cpu shares.
212 static DEFINE_SPINLOCK(task_group_lock);
214 /* doms_cur_mutex serializes access to doms_cur[] array */
215 static DEFINE_MUTEX(doms_cur_mutex);
217 #ifdef CONFIG_FAIR_GROUP_SCHED
218 #ifdef CONFIG_USER_SCHED
219 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
220 #else
221 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
222 #endif
224 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
225 #endif
227 /* Default task group.
228 * Every task in system belong to this group at bootup.
230 struct task_group init_task_group = {
231 #ifdef CONFIG_FAIR_GROUP_SCHED
232 .se = init_sched_entity_p,
233 .cfs_rq = init_cfs_rq_p,
234 #endif
236 #ifdef CONFIG_RT_GROUP_SCHED
237 .rt_se = init_sched_rt_entity_p,
238 .rt_rq = init_rt_rq_p,
239 #endif
242 /* return group to which a task belongs */
243 static inline struct task_group *task_group(struct task_struct *p)
245 struct task_group *tg;
247 #ifdef CONFIG_USER_SCHED
248 tg = p->user->tg;
249 #elif defined(CONFIG_CGROUP_SCHED)
250 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
251 struct task_group, css);
252 #else
253 tg = &init_task_group;
254 #endif
255 return tg;
258 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
259 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
261 #ifdef CONFIG_FAIR_GROUP_SCHED
262 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
263 p->se.parent = task_group(p)->se[cpu];
264 #endif
266 #ifdef CONFIG_RT_GROUP_SCHED
267 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
268 p->rt.parent = task_group(p)->rt_se[cpu];
269 #endif
272 static inline void lock_doms_cur(void)
274 mutex_lock(&doms_cur_mutex);
277 static inline void unlock_doms_cur(void)
279 mutex_unlock(&doms_cur_mutex);
282 #else
284 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
285 static inline void lock_doms_cur(void) { }
286 static inline void unlock_doms_cur(void) { }
288 #endif /* CONFIG_GROUP_SCHED */
290 /* CFS-related fields in a runqueue */
291 struct cfs_rq {
292 struct load_weight load;
293 unsigned long nr_running;
295 u64 exec_clock;
296 u64 min_vruntime;
298 struct rb_root tasks_timeline;
299 struct rb_node *rb_leftmost;
300 struct rb_node *rb_load_balance_curr;
301 /* 'curr' points to currently running entity on this cfs_rq.
302 * It is set to NULL otherwise (i.e when none are currently running).
304 struct sched_entity *curr;
306 unsigned long nr_spread_over;
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
312 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
313 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
314 * (like users, containers etc.)
316 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
317 * list is used during load balance.
319 struct list_head leaf_cfs_rq_list;
320 struct task_group *tg; /* group that "owns" this runqueue */
321 #endif
324 /* Real-Time classes' related field in a runqueue: */
325 struct rt_rq {
326 struct rt_prio_array active;
327 unsigned long rt_nr_running;
328 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
329 int highest_prio; /* highest queued rt task prio */
330 #endif
331 #ifdef CONFIG_SMP
332 unsigned long rt_nr_migratory;
333 int overloaded;
334 #endif
335 int rt_throttled;
336 u64 rt_time;
338 #ifdef CONFIG_RT_GROUP_SCHED
339 unsigned long rt_nr_boosted;
341 struct rq *rq;
342 struct list_head leaf_rt_rq_list;
343 struct task_group *tg;
344 struct sched_rt_entity *rt_se;
345 #endif
348 #ifdef CONFIG_SMP
351 * We add the notion of a root-domain which will be used to define per-domain
352 * variables. Each exclusive cpuset essentially defines an island domain by
353 * fully partitioning the member cpus from any other cpuset. Whenever a new
354 * exclusive cpuset is created, we also create and attach a new root-domain
355 * object.
358 struct root_domain {
359 atomic_t refcount;
360 cpumask_t span;
361 cpumask_t online;
364 * The "RT overload" flag: it gets set if a CPU has more than
365 * one runnable RT task.
367 cpumask_t rto_mask;
368 atomic_t rto_count;
372 * By default the system creates a single root-domain with all cpus as
373 * members (mimicking the global state we have today).
375 static struct root_domain def_root_domain;
377 #endif
380 * This is the main, per-CPU runqueue data structure.
382 * Locking rule: those places that want to lock multiple runqueues
383 * (such as the load balancing or the thread migration code), lock
384 * acquire operations must be ordered by ascending &runqueue.
386 struct rq {
387 /* runqueue lock: */
388 spinlock_t lock;
391 * nr_running and cpu_load should be in the same cacheline because
392 * remote CPUs use both these fields when doing load calculation.
394 unsigned long nr_running;
395 #define CPU_LOAD_IDX_MAX 5
396 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
397 unsigned char idle_at_tick;
398 #ifdef CONFIG_NO_HZ
399 unsigned char in_nohz_recently;
400 #endif
401 /* capture load from *all* tasks on this cpu: */
402 struct load_weight load;
403 unsigned long nr_load_updates;
404 u64 nr_switches;
406 struct cfs_rq cfs;
407 struct rt_rq rt;
408 u64 rt_period_expire;
409 int rt_throttled;
411 #ifdef CONFIG_FAIR_GROUP_SCHED
412 /* list of leaf cfs_rq on this cpu: */
413 struct list_head leaf_cfs_rq_list;
414 #endif
415 #ifdef CONFIG_RT_GROUP_SCHED
416 struct list_head leaf_rt_rq_list;
417 #endif
420 * This is part of a global counter where only the total sum
421 * over all CPUs matters. A task can increase this counter on
422 * one CPU and if it got migrated afterwards it may decrease
423 * it on another CPU. Always updated under the runqueue lock:
425 unsigned long nr_uninterruptible;
427 struct task_struct *curr, *idle;
428 unsigned long next_balance;
429 struct mm_struct *prev_mm;
431 u64 clock, prev_clock_raw;
432 s64 clock_max_delta;
434 unsigned int clock_warps, clock_overflows, clock_underflows;
435 u64 idle_clock;
436 unsigned int clock_deep_idle_events;
437 u64 tick_timestamp;
439 atomic_t nr_iowait;
441 #ifdef CONFIG_SMP
442 struct root_domain *rd;
443 struct sched_domain *sd;
445 /* For active balancing */
446 int active_balance;
447 int push_cpu;
448 /* cpu of this runqueue: */
449 int cpu;
451 struct task_struct *migration_thread;
452 struct list_head migration_queue;
453 #endif
455 #ifdef CONFIG_SCHED_HRTICK
456 unsigned long hrtick_flags;
457 ktime_t hrtick_expire;
458 struct hrtimer hrtick_timer;
459 #endif
461 #ifdef CONFIG_SCHEDSTATS
462 /* latency stats */
463 struct sched_info rq_sched_info;
465 /* sys_sched_yield() stats */
466 unsigned int yld_exp_empty;
467 unsigned int yld_act_empty;
468 unsigned int yld_both_empty;
469 unsigned int yld_count;
471 /* schedule() stats */
472 unsigned int sched_switch;
473 unsigned int sched_count;
474 unsigned int sched_goidle;
476 /* try_to_wake_up() stats */
477 unsigned int ttwu_count;
478 unsigned int ttwu_local;
480 /* BKL stats */
481 unsigned int bkl_count;
482 #endif
483 struct lock_class_key rq_lock_key;
486 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
488 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
490 rq->curr->sched_class->check_preempt_curr(rq, p);
493 static inline int cpu_of(struct rq *rq)
495 #ifdef CONFIG_SMP
496 return rq->cpu;
497 #else
498 return 0;
499 #endif
503 * Update the per-runqueue clock, as finegrained as the platform can give
504 * us, but without assuming monotonicity, etc.:
506 static void __update_rq_clock(struct rq *rq)
508 u64 prev_raw = rq->prev_clock_raw;
509 u64 now = sched_clock();
510 s64 delta = now - prev_raw;
511 u64 clock = rq->clock;
513 #ifdef CONFIG_SCHED_DEBUG
514 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
515 #endif
517 * Protect against sched_clock() occasionally going backwards:
519 if (unlikely(delta < 0)) {
520 clock++;
521 rq->clock_warps++;
522 } else {
524 * Catch too large forward jumps too:
526 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
527 if (clock < rq->tick_timestamp + TICK_NSEC)
528 clock = rq->tick_timestamp + TICK_NSEC;
529 else
530 clock++;
531 rq->clock_overflows++;
532 } else {
533 if (unlikely(delta > rq->clock_max_delta))
534 rq->clock_max_delta = delta;
535 clock += delta;
539 rq->prev_clock_raw = now;
540 rq->clock = clock;
543 static void update_rq_clock(struct rq *rq)
545 if (likely(smp_processor_id() == cpu_of(rq)))
546 __update_rq_clock(rq);
550 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
551 * See detach_destroy_domains: synchronize_sched for details.
553 * The domain tree of any CPU may only be accessed from within
554 * preempt-disabled sections.
556 #define for_each_domain(cpu, __sd) \
557 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
559 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
560 #define this_rq() (&__get_cpu_var(runqueues))
561 #define task_rq(p) cpu_rq(task_cpu(p))
562 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
564 unsigned long rt_needs_cpu(int cpu)
566 struct rq *rq = cpu_rq(cpu);
567 u64 delta;
569 if (!rq->rt_throttled)
570 return 0;
572 if (rq->clock > rq->rt_period_expire)
573 return 1;
575 delta = rq->rt_period_expire - rq->clock;
576 do_div(delta, NSEC_PER_SEC / HZ);
578 return (unsigned long)delta;
582 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
584 #ifdef CONFIG_SCHED_DEBUG
585 # define const_debug __read_mostly
586 #else
587 # define const_debug static const
588 #endif
591 * Debugging: various feature bits
593 enum {
594 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
595 SCHED_FEAT_WAKEUP_PREEMPT = 2,
596 SCHED_FEAT_START_DEBIT = 4,
597 SCHED_FEAT_TREE_AVG = 8,
598 SCHED_FEAT_APPROX_AVG = 16,
599 SCHED_FEAT_HRTICK = 32,
600 SCHED_FEAT_DOUBLE_TICK = 64,
603 const_debug unsigned int sysctl_sched_features =
604 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
605 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
606 SCHED_FEAT_START_DEBIT * 1 |
607 SCHED_FEAT_TREE_AVG * 0 |
608 SCHED_FEAT_APPROX_AVG * 0 |
609 SCHED_FEAT_HRTICK * 1 |
610 SCHED_FEAT_DOUBLE_TICK * 0;
612 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
615 * Number of tasks to iterate in a single balance run.
616 * Limited because this is done with IRQs disabled.
618 const_debug unsigned int sysctl_sched_nr_migrate = 32;
621 * period over which we measure -rt task cpu usage in us.
622 * default: 1s
624 unsigned int sysctl_sched_rt_period = 1000000;
626 static __read_mostly int scheduler_running;
629 * part of the period that we allow rt tasks to run in us.
630 * default: 0.95s
632 int sysctl_sched_rt_runtime = 950000;
635 * single value that denotes runtime == period, ie unlimited time.
637 #define RUNTIME_INF ((u64)~0ULL)
640 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
641 * clock constructed from sched_clock():
643 unsigned long long cpu_clock(int cpu)
645 unsigned long long now;
646 unsigned long flags;
647 struct rq *rq;
650 * Only call sched_clock() if the scheduler has already been
651 * initialized (some code might call cpu_clock() very early):
653 if (unlikely(!scheduler_running))
654 return 0;
656 local_irq_save(flags);
657 rq = cpu_rq(cpu);
658 update_rq_clock(rq);
659 now = rq->clock;
660 local_irq_restore(flags);
662 return now;
664 EXPORT_SYMBOL_GPL(cpu_clock);
666 #ifndef prepare_arch_switch
667 # define prepare_arch_switch(next) do { } while (0)
668 #endif
669 #ifndef finish_arch_switch
670 # define finish_arch_switch(prev) do { } while (0)
671 #endif
673 static inline int task_current(struct rq *rq, struct task_struct *p)
675 return rq->curr == p;
678 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
679 static inline int task_running(struct rq *rq, struct task_struct *p)
681 return task_current(rq, p);
684 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
688 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
690 #ifdef CONFIG_DEBUG_SPINLOCK
691 /* this is a valid case when another task releases the spinlock */
692 rq->lock.owner = current;
693 #endif
695 * If we are tracking spinlock dependencies then we have to
696 * fix up the runqueue lock - which gets 'carried over' from
697 * prev into current:
699 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
701 spin_unlock_irq(&rq->lock);
704 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
705 static inline int task_running(struct rq *rq, struct task_struct *p)
707 #ifdef CONFIG_SMP
708 return p->oncpu;
709 #else
710 return task_current(rq, p);
711 #endif
714 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
716 #ifdef CONFIG_SMP
718 * We can optimise this out completely for !SMP, because the
719 * SMP rebalancing from interrupt is the only thing that cares
720 * here.
722 next->oncpu = 1;
723 #endif
724 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
725 spin_unlock_irq(&rq->lock);
726 #else
727 spin_unlock(&rq->lock);
728 #endif
731 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
733 #ifdef CONFIG_SMP
735 * After ->oncpu is cleared, the task can be moved to a different CPU.
736 * We must ensure this doesn't happen until the switch is completely
737 * finished.
739 smp_wmb();
740 prev->oncpu = 0;
741 #endif
742 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
743 local_irq_enable();
744 #endif
746 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
749 * __task_rq_lock - lock the runqueue a given task resides on.
750 * Must be called interrupts disabled.
752 static inline struct rq *__task_rq_lock(struct task_struct *p)
753 __acquires(rq->lock)
755 for (;;) {
756 struct rq *rq = task_rq(p);
757 spin_lock(&rq->lock);
758 if (likely(rq == task_rq(p)))
759 return rq;
760 spin_unlock(&rq->lock);
765 * task_rq_lock - lock the runqueue a given task resides on and disable
766 * interrupts. Note the ordering: we can safely lookup the task_rq without
767 * explicitly disabling preemption.
769 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
770 __acquires(rq->lock)
772 struct rq *rq;
774 for (;;) {
775 local_irq_save(*flags);
776 rq = task_rq(p);
777 spin_lock(&rq->lock);
778 if (likely(rq == task_rq(p)))
779 return rq;
780 spin_unlock_irqrestore(&rq->lock, *flags);
784 static void __task_rq_unlock(struct rq *rq)
785 __releases(rq->lock)
787 spin_unlock(&rq->lock);
790 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
791 __releases(rq->lock)
793 spin_unlock_irqrestore(&rq->lock, *flags);
797 * this_rq_lock - lock this runqueue and disable interrupts.
799 static struct rq *this_rq_lock(void)
800 __acquires(rq->lock)
802 struct rq *rq;
804 local_irq_disable();
805 rq = this_rq();
806 spin_lock(&rq->lock);
808 return rq;
812 * We are going deep-idle (irqs are disabled):
814 void sched_clock_idle_sleep_event(void)
816 struct rq *rq = cpu_rq(smp_processor_id());
818 spin_lock(&rq->lock);
819 __update_rq_clock(rq);
820 spin_unlock(&rq->lock);
821 rq->clock_deep_idle_events++;
823 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
826 * We just idled delta nanoseconds (called with irqs disabled):
828 void sched_clock_idle_wakeup_event(u64 delta_ns)
830 struct rq *rq = cpu_rq(smp_processor_id());
831 u64 now = sched_clock();
833 rq->idle_clock += delta_ns;
835 * Override the previous timestamp and ignore all
836 * sched_clock() deltas that occured while we idled,
837 * and use the PM-provided delta_ns to advance the
838 * rq clock:
840 spin_lock(&rq->lock);
841 rq->prev_clock_raw = now;
842 rq->clock += delta_ns;
843 spin_unlock(&rq->lock);
844 touch_softlockup_watchdog();
846 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
848 static void __resched_task(struct task_struct *p, int tif_bit);
850 static inline void resched_task(struct task_struct *p)
852 __resched_task(p, TIF_NEED_RESCHED);
855 #ifdef CONFIG_SCHED_HRTICK
857 * Use HR-timers to deliver accurate preemption points.
859 * Its all a bit involved since we cannot program an hrt while holding the
860 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
861 * reschedule event.
863 * When we get rescheduled we reprogram the hrtick_timer outside of the
864 * rq->lock.
866 static inline void resched_hrt(struct task_struct *p)
868 __resched_task(p, TIF_HRTICK_RESCHED);
871 static inline void resched_rq(struct rq *rq)
873 unsigned long flags;
875 spin_lock_irqsave(&rq->lock, flags);
876 resched_task(rq->curr);
877 spin_unlock_irqrestore(&rq->lock, flags);
880 enum {
881 HRTICK_SET, /* re-programm hrtick_timer */
882 HRTICK_RESET, /* not a new slice */
886 * Use hrtick when:
887 * - enabled by features
888 * - hrtimer is actually high res
890 static inline int hrtick_enabled(struct rq *rq)
892 if (!sched_feat(HRTICK))
893 return 0;
894 return hrtimer_is_hres_active(&rq->hrtick_timer);
898 * Called to set the hrtick timer state.
900 * called with rq->lock held and irqs disabled
902 static void hrtick_start(struct rq *rq, u64 delay, int reset)
904 assert_spin_locked(&rq->lock);
907 * preempt at: now + delay
909 rq->hrtick_expire =
910 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
912 * indicate we need to program the timer
914 __set_bit(HRTICK_SET, &rq->hrtick_flags);
915 if (reset)
916 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
919 * New slices are called from the schedule path and don't need a
920 * forced reschedule.
922 if (reset)
923 resched_hrt(rq->curr);
926 static void hrtick_clear(struct rq *rq)
928 if (hrtimer_active(&rq->hrtick_timer))
929 hrtimer_cancel(&rq->hrtick_timer);
933 * Update the timer from the possible pending state.
935 static void hrtick_set(struct rq *rq)
937 ktime_t time;
938 int set, reset;
939 unsigned long flags;
941 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
943 spin_lock_irqsave(&rq->lock, flags);
944 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
945 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
946 time = rq->hrtick_expire;
947 clear_thread_flag(TIF_HRTICK_RESCHED);
948 spin_unlock_irqrestore(&rq->lock, flags);
950 if (set) {
951 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
952 if (reset && !hrtimer_active(&rq->hrtick_timer))
953 resched_rq(rq);
954 } else
955 hrtick_clear(rq);
959 * High-resolution timer tick.
960 * Runs from hardirq context with interrupts disabled.
962 static enum hrtimer_restart hrtick(struct hrtimer *timer)
964 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
966 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
968 spin_lock(&rq->lock);
969 __update_rq_clock(rq);
970 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
971 spin_unlock(&rq->lock);
973 return HRTIMER_NORESTART;
976 static inline void init_rq_hrtick(struct rq *rq)
978 rq->hrtick_flags = 0;
979 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
980 rq->hrtick_timer.function = hrtick;
981 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
984 void hrtick_resched(void)
986 struct rq *rq;
987 unsigned long flags;
989 if (!test_thread_flag(TIF_HRTICK_RESCHED))
990 return;
992 local_irq_save(flags);
993 rq = cpu_rq(smp_processor_id());
994 hrtick_set(rq);
995 local_irq_restore(flags);
997 #else
998 static inline void hrtick_clear(struct rq *rq)
1002 static inline void hrtick_set(struct rq *rq)
1006 static inline void init_rq_hrtick(struct rq *rq)
1010 void hrtick_resched(void)
1013 #endif
1016 * resched_task - mark a task 'to be rescheduled now'.
1018 * On UP this means the setting of the need_resched flag, on SMP it
1019 * might also involve a cross-CPU call to trigger the scheduler on
1020 * the target CPU.
1022 #ifdef CONFIG_SMP
1024 #ifndef tsk_is_polling
1025 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1026 #endif
1028 static void __resched_task(struct task_struct *p, int tif_bit)
1030 int cpu;
1032 assert_spin_locked(&task_rq(p)->lock);
1034 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1035 return;
1037 set_tsk_thread_flag(p, tif_bit);
1039 cpu = task_cpu(p);
1040 if (cpu == smp_processor_id())
1041 return;
1043 /* NEED_RESCHED must be visible before we test polling */
1044 smp_mb();
1045 if (!tsk_is_polling(p))
1046 smp_send_reschedule(cpu);
1049 static void resched_cpu(int cpu)
1051 struct rq *rq = cpu_rq(cpu);
1052 unsigned long flags;
1054 if (!spin_trylock_irqsave(&rq->lock, flags))
1055 return;
1056 resched_task(cpu_curr(cpu));
1057 spin_unlock_irqrestore(&rq->lock, flags);
1059 #else
1060 static void __resched_task(struct task_struct *p, int tif_bit)
1062 assert_spin_locked(&task_rq(p)->lock);
1063 set_tsk_thread_flag(p, tif_bit);
1065 #endif
1067 #if BITS_PER_LONG == 32
1068 # define WMULT_CONST (~0UL)
1069 #else
1070 # define WMULT_CONST (1UL << 32)
1071 #endif
1073 #define WMULT_SHIFT 32
1076 * Shift right and round:
1078 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1080 static unsigned long
1081 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1082 struct load_weight *lw)
1084 u64 tmp;
1086 if (unlikely(!lw->inv_weight))
1087 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
1089 tmp = (u64)delta_exec * weight;
1091 * Check whether we'd overflow the 64-bit multiplication:
1093 if (unlikely(tmp > WMULT_CONST))
1094 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1095 WMULT_SHIFT/2);
1096 else
1097 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1099 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1102 static inline unsigned long
1103 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1105 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1108 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1110 lw->weight += inc;
1113 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1115 lw->weight -= dec;
1119 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1120 * of tasks with abnormal "nice" values across CPUs the contribution that
1121 * each task makes to its run queue's load is weighted according to its
1122 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1123 * scaled version of the new time slice allocation that they receive on time
1124 * slice expiry etc.
1127 #define WEIGHT_IDLEPRIO 2
1128 #define WMULT_IDLEPRIO (1 << 31)
1131 * Nice levels are multiplicative, with a gentle 10% change for every
1132 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1133 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1134 * that remained on nice 0.
1136 * The "10% effect" is relative and cumulative: from _any_ nice level,
1137 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1138 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1139 * If a task goes up by ~10% and another task goes down by ~10% then
1140 * the relative distance between them is ~25%.)
1142 static const int prio_to_weight[40] = {
1143 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1144 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1145 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1146 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1147 /* 0 */ 1024, 820, 655, 526, 423,
1148 /* 5 */ 335, 272, 215, 172, 137,
1149 /* 10 */ 110, 87, 70, 56, 45,
1150 /* 15 */ 36, 29, 23, 18, 15,
1154 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1156 * In cases where the weight does not change often, we can use the
1157 * precalculated inverse to speed up arithmetics by turning divisions
1158 * into multiplications:
1160 static const u32 prio_to_wmult[40] = {
1161 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1162 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1163 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1164 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1165 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1166 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1167 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1168 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1171 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1174 * runqueue iterator, to support SMP load-balancing between different
1175 * scheduling classes, without having to expose their internal data
1176 * structures to the load-balancing proper:
1178 struct rq_iterator {
1179 void *arg;
1180 struct task_struct *(*start)(void *);
1181 struct task_struct *(*next)(void *);
1184 #ifdef CONFIG_SMP
1185 static unsigned long
1186 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1187 unsigned long max_load_move, struct sched_domain *sd,
1188 enum cpu_idle_type idle, int *all_pinned,
1189 int *this_best_prio, struct rq_iterator *iterator);
1191 static int
1192 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1193 struct sched_domain *sd, enum cpu_idle_type idle,
1194 struct rq_iterator *iterator);
1195 #endif
1197 #ifdef CONFIG_CGROUP_CPUACCT
1198 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1199 #else
1200 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1201 #endif
1203 #ifdef CONFIG_SMP
1204 static unsigned long source_load(int cpu, int type);
1205 static unsigned long target_load(int cpu, int type);
1206 static unsigned long cpu_avg_load_per_task(int cpu);
1207 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1208 #endif /* CONFIG_SMP */
1210 #include "sched_stats.h"
1211 #include "sched_idletask.c"
1212 #include "sched_fair.c"
1213 #include "sched_rt.c"
1214 #ifdef CONFIG_SCHED_DEBUG
1215 # include "sched_debug.c"
1216 #endif
1218 #define sched_class_highest (&rt_sched_class)
1220 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1222 update_load_add(&rq->load, p->se.load.weight);
1225 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1227 update_load_sub(&rq->load, p->se.load.weight);
1230 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1232 rq->nr_running++;
1233 inc_load(rq, p);
1236 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1238 rq->nr_running--;
1239 dec_load(rq, p);
1242 static void set_load_weight(struct task_struct *p)
1244 if (task_has_rt_policy(p)) {
1245 p->se.load.weight = prio_to_weight[0] * 2;
1246 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1247 return;
1251 * SCHED_IDLE tasks get minimal weight:
1253 if (p->policy == SCHED_IDLE) {
1254 p->se.load.weight = WEIGHT_IDLEPRIO;
1255 p->se.load.inv_weight = WMULT_IDLEPRIO;
1256 return;
1259 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1260 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1263 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1265 sched_info_queued(p);
1266 p->sched_class->enqueue_task(rq, p, wakeup);
1267 p->se.on_rq = 1;
1270 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1272 p->sched_class->dequeue_task(rq, p, sleep);
1273 p->se.on_rq = 0;
1277 * __normal_prio - return the priority that is based on the static prio
1279 static inline int __normal_prio(struct task_struct *p)
1281 return p->static_prio;
1285 * Calculate the expected normal priority: i.e. priority
1286 * without taking RT-inheritance into account. Might be
1287 * boosted by interactivity modifiers. Changes upon fork,
1288 * setprio syscalls, and whenever the interactivity
1289 * estimator recalculates.
1291 static inline int normal_prio(struct task_struct *p)
1293 int prio;
1295 if (task_has_rt_policy(p))
1296 prio = MAX_RT_PRIO-1 - p->rt_priority;
1297 else
1298 prio = __normal_prio(p);
1299 return prio;
1303 * Calculate the current priority, i.e. the priority
1304 * taken into account by the scheduler. This value might
1305 * be boosted by RT tasks, or might be boosted by
1306 * interactivity modifiers. Will be RT if the task got
1307 * RT-boosted. If not then it returns p->normal_prio.
1309 static int effective_prio(struct task_struct *p)
1311 p->normal_prio = normal_prio(p);
1313 * If we are RT tasks or we were boosted to RT priority,
1314 * keep the priority unchanged. Otherwise, update priority
1315 * to the normal priority:
1317 if (!rt_prio(p->prio))
1318 return p->normal_prio;
1319 return p->prio;
1323 * activate_task - move a task to the runqueue.
1325 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1327 if (task_contributes_to_load(p))
1328 rq->nr_uninterruptible--;
1330 enqueue_task(rq, p, wakeup);
1331 inc_nr_running(p, rq);
1335 * deactivate_task - remove a task from the runqueue.
1337 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1339 if (task_contributes_to_load(p))
1340 rq->nr_uninterruptible++;
1342 dequeue_task(rq, p, sleep);
1343 dec_nr_running(p, rq);
1347 * task_curr - is this task currently executing on a CPU?
1348 * @p: the task in question.
1350 inline int task_curr(const struct task_struct *p)
1352 return cpu_curr(task_cpu(p)) == p;
1355 /* Used instead of source_load when we know the type == 0 */
1356 unsigned long weighted_cpuload(const int cpu)
1358 return cpu_rq(cpu)->load.weight;
1361 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1363 set_task_rq(p, cpu);
1364 #ifdef CONFIG_SMP
1366 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1367 * successfuly executed on another CPU. We must ensure that updates of
1368 * per-task data have been completed by this moment.
1370 smp_wmb();
1371 task_thread_info(p)->cpu = cpu;
1372 #endif
1375 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1376 const struct sched_class *prev_class,
1377 int oldprio, int running)
1379 if (prev_class != p->sched_class) {
1380 if (prev_class->switched_from)
1381 prev_class->switched_from(rq, p, running);
1382 p->sched_class->switched_to(rq, p, running);
1383 } else
1384 p->sched_class->prio_changed(rq, p, oldprio, running);
1387 #ifdef CONFIG_SMP
1390 * Is this task likely cache-hot:
1392 static int
1393 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1395 s64 delta;
1397 if (p->sched_class != &fair_sched_class)
1398 return 0;
1400 if (sysctl_sched_migration_cost == -1)
1401 return 1;
1402 if (sysctl_sched_migration_cost == 0)
1403 return 0;
1405 delta = now - p->se.exec_start;
1407 return delta < (s64)sysctl_sched_migration_cost;
1411 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1413 int old_cpu = task_cpu(p);
1414 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1415 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1416 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1417 u64 clock_offset;
1419 clock_offset = old_rq->clock - new_rq->clock;
1421 #ifdef CONFIG_SCHEDSTATS
1422 if (p->se.wait_start)
1423 p->se.wait_start -= clock_offset;
1424 if (p->se.sleep_start)
1425 p->se.sleep_start -= clock_offset;
1426 if (p->se.block_start)
1427 p->se.block_start -= clock_offset;
1428 if (old_cpu != new_cpu) {
1429 schedstat_inc(p, se.nr_migrations);
1430 if (task_hot(p, old_rq->clock, NULL))
1431 schedstat_inc(p, se.nr_forced2_migrations);
1433 #endif
1434 p->se.vruntime -= old_cfsrq->min_vruntime -
1435 new_cfsrq->min_vruntime;
1437 __set_task_cpu(p, new_cpu);
1440 struct migration_req {
1441 struct list_head list;
1443 struct task_struct *task;
1444 int dest_cpu;
1446 struct completion done;
1450 * The task's runqueue lock must be held.
1451 * Returns true if you have to wait for migration thread.
1453 static int
1454 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1456 struct rq *rq = task_rq(p);
1459 * If the task is not on a runqueue (and not running), then
1460 * it is sufficient to simply update the task's cpu field.
1462 if (!p->se.on_rq && !task_running(rq, p)) {
1463 set_task_cpu(p, dest_cpu);
1464 return 0;
1467 init_completion(&req->done);
1468 req->task = p;
1469 req->dest_cpu = dest_cpu;
1470 list_add(&req->list, &rq->migration_queue);
1472 return 1;
1476 * wait_task_inactive - wait for a thread to unschedule.
1478 * The caller must ensure that the task *will* unschedule sometime soon,
1479 * else this function might spin for a *long* time. This function can't
1480 * be called with interrupts off, or it may introduce deadlock with
1481 * smp_call_function() if an IPI is sent by the same process we are
1482 * waiting to become inactive.
1484 void wait_task_inactive(struct task_struct *p)
1486 unsigned long flags;
1487 int running, on_rq;
1488 struct rq *rq;
1490 for (;;) {
1492 * We do the initial early heuristics without holding
1493 * any task-queue locks at all. We'll only try to get
1494 * the runqueue lock when things look like they will
1495 * work out!
1497 rq = task_rq(p);
1500 * If the task is actively running on another CPU
1501 * still, just relax and busy-wait without holding
1502 * any locks.
1504 * NOTE! Since we don't hold any locks, it's not
1505 * even sure that "rq" stays as the right runqueue!
1506 * But we don't care, since "task_running()" will
1507 * return false if the runqueue has changed and p
1508 * is actually now running somewhere else!
1510 while (task_running(rq, p))
1511 cpu_relax();
1514 * Ok, time to look more closely! We need the rq
1515 * lock now, to be *sure*. If we're wrong, we'll
1516 * just go back and repeat.
1518 rq = task_rq_lock(p, &flags);
1519 running = task_running(rq, p);
1520 on_rq = p->se.on_rq;
1521 task_rq_unlock(rq, &flags);
1524 * Was it really running after all now that we
1525 * checked with the proper locks actually held?
1527 * Oops. Go back and try again..
1529 if (unlikely(running)) {
1530 cpu_relax();
1531 continue;
1535 * It's not enough that it's not actively running,
1536 * it must be off the runqueue _entirely_, and not
1537 * preempted!
1539 * So if it wa still runnable (but just not actively
1540 * running right now), it's preempted, and we should
1541 * yield - it could be a while.
1543 if (unlikely(on_rq)) {
1544 schedule_timeout_uninterruptible(1);
1545 continue;
1549 * Ahh, all good. It wasn't running, and it wasn't
1550 * runnable, which means that it will never become
1551 * running in the future either. We're all done!
1553 break;
1557 /***
1558 * kick_process - kick a running thread to enter/exit the kernel
1559 * @p: the to-be-kicked thread
1561 * Cause a process which is running on another CPU to enter
1562 * kernel-mode, without any delay. (to get signals handled.)
1564 * NOTE: this function doesnt have to take the runqueue lock,
1565 * because all it wants to ensure is that the remote task enters
1566 * the kernel. If the IPI races and the task has been migrated
1567 * to another CPU then no harm is done and the purpose has been
1568 * achieved as well.
1570 void kick_process(struct task_struct *p)
1572 int cpu;
1574 preempt_disable();
1575 cpu = task_cpu(p);
1576 if ((cpu != smp_processor_id()) && task_curr(p))
1577 smp_send_reschedule(cpu);
1578 preempt_enable();
1582 * Return a low guess at the load of a migration-source cpu weighted
1583 * according to the scheduling class and "nice" value.
1585 * We want to under-estimate the load of migration sources, to
1586 * balance conservatively.
1588 static unsigned long source_load(int cpu, int type)
1590 struct rq *rq = cpu_rq(cpu);
1591 unsigned long total = weighted_cpuload(cpu);
1593 if (type == 0)
1594 return total;
1596 return min(rq->cpu_load[type-1], total);
1600 * Return a high guess at the load of a migration-target cpu weighted
1601 * according to the scheduling class and "nice" value.
1603 static unsigned long target_load(int cpu, int type)
1605 struct rq *rq = cpu_rq(cpu);
1606 unsigned long total = weighted_cpuload(cpu);
1608 if (type == 0)
1609 return total;
1611 return max(rq->cpu_load[type-1], total);
1615 * Return the average load per task on the cpu's run queue
1617 static unsigned long cpu_avg_load_per_task(int cpu)
1619 struct rq *rq = cpu_rq(cpu);
1620 unsigned long total = weighted_cpuload(cpu);
1621 unsigned long n = rq->nr_running;
1623 return n ? total / n : SCHED_LOAD_SCALE;
1627 * find_idlest_group finds and returns the least busy CPU group within the
1628 * domain.
1630 static struct sched_group *
1631 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1633 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1634 unsigned long min_load = ULONG_MAX, this_load = 0;
1635 int load_idx = sd->forkexec_idx;
1636 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1638 do {
1639 unsigned long load, avg_load;
1640 int local_group;
1641 int i;
1643 /* Skip over this group if it has no CPUs allowed */
1644 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1645 continue;
1647 local_group = cpu_isset(this_cpu, group->cpumask);
1649 /* Tally up the load of all CPUs in the group */
1650 avg_load = 0;
1652 for_each_cpu_mask(i, group->cpumask) {
1653 /* Bias balancing toward cpus of our domain */
1654 if (local_group)
1655 load = source_load(i, load_idx);
1656 else
1657 load = target_load(i, load_idx);
1659 avg_load += load;
1662 /* Adjust by relative CPU power of the group */
1663 avg_load = sg_div_cpu_power(group,
1664 avg_load * SCHED_LOAD_SCALE);
1666 if (local_group) {
1667 this_load = avg_load;
1668 this = group;
1669 } else if (avg_load < min_load) {
1670 min_load = avg_load;
1671 idlest = group;
1673 } while (group = group->next, group != sd->groups);
1675 if (!idlest || 100*this_load < imbalance*min_load)
1676 return NULL;
1677 return idlest;
1681 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1683 static int
1684 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1686 cpumask_t tmp;
1687 unsigned long load, min_load = ULONG_MAX;
1688 int idlest = -1;
1689 int i;
1691 /* Traverse only the allowed CPUs */
1692 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1694 for_each_cpu_mask(i, tmp) {
1695 load = weighted_cpuload(i);
1697 if (load < min_load || (load == min_load && i == this_cpu)) {
1698 min_load = load;
1699 idlest = i;
1703 return idlest;
1707 * sched_balance_self: balance the current task (running on cpu) in domains
1708 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1709 * SD_BALANCE_EXEC.
1711 * Balance, ie. select the least loaded group.
1713 * Returns the target CPU number, or the same CPU if no balancing is needed.
1715 * preempt must be disabled.
1717 static int sched_balance_self(int cpu, int flag)
1719 struct task_struct *t = current;
1720 struct sched_domain *tmp, *sd = NULL;
1722 for_each_domain(cpu, tmp) {
1724 * If power savings logic is enabled for a domain, stop there.
1726 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1727 break;
1728 if (tmp->flags & flag)
1729 sd = tmp;
1732 while (sd) {
1733 cpumask_t span;
1734 struct sched_group *group;
1735 int new_cpu, weight;
1737 if (!(sd->flags & flag)) {
1738 sd = sd->child;
1739 continue;
1742 span = sd->span;
1743 group = find_idlest_group(sd, t, cpu);
1744 if (!group) {
1745 sd = sd->child;
1746 continue;
1749 new_cpu = find_idlest_cpu(group, t, cpu);
1750 if (new_cpu == -1 || new_cpu == cpu) {
1751 /* Now try balancing at a lower domain level of cpu */
1752 sd = sd->child;
1753 continue;
1756 /* Now try balancing at a lower domain level of new_cpu */
1757 cpu = new_cpu;
1758 sd = NULL;
1759 weight = cpus_weight(span);
1760 for_each_domain(cpu, tmp) {
1761 if (weight <= cpus_weight(tmp->span))
1762 break;
1763 if (tmp->flags & flag)
1764 sd = tmp;
1766 /* while loop will break here if sd == NULL */
1769 return cpu;
1772 #endif /* CONFIG_SMP */
1774 /***
1775 * try_to_wake_up - wake up a thread
1776 * @p: the to-be-woken-up thread
1777 * @state: the mask of task states that can be woken
1778 * @sync: do a synchronous wakeup?
1780 * Put it on the run-queue if it's not already there. The "current"
1781 * thread is always on the run-queue (except when the actual
1782 * re-schedule is in progress), and as such you're allowed to do
1783 * the simpler "current->state = TASK_RUNNING" to mark yourself
1784 * runnable without the overhead of this.
1786 * returns failure only if the task is already active.
1788 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1790 int cpu, orig_cpu, this_cpu, success = 0;
1791 unsigned long flags;
1792 long old_state;
1793 struct rq *rq;
1795 smp_wmb();
1796 rq = task_rq_lock(p, &flags);
1797 old_state = p->state;
1798 if (!(old_state & state))
1799 goto out;
1801 if (p->se.on_rq)
1802 goto out_running;
1804 cpu = task_cpu(p);
1805 orig_cpu = cpu;
1806 this_cpu = smp_processor_id();
1808 #ifdef CONFIG_SMP
1809 if (unlikely(task_running(rq, p)))
1810 goto out_activate;
1812 cpu = p->sched_class->select_task_rq(p, sync);
1813 if (cpu != orig_cpu) {
1814 set_task_cpu(p, cpu);
1815 task_rq_unlock(rq, &flags);
1816 /* might preempt at this point */
1817 rq = task_rq_lock(p, &flags);
1818 old_state = p->state;
1819 if (!(old_state & state))
1820 goto out;
1821 if (p->se.on_rq)
1822 goto out_running;
1824 this_cpu = smp_processor_id();
1825 cpu = task_cpu(p);
1828 #ifdef CONFIG_SCHEDSTATS
1829 schedstat_inc(rq, ttwu_count);
1830 if (cpu == this_cpu)
1831 schedstat_inc(rq, ttwu_local);
1832 else {
1833 struct sched_domain *sd;
1834 for_each_domain(this_cpu, sd) {
1835 if (cpu_isset(cpu, sd->span)) {
1836 schedstat_inc(sd, ttwu_wake_remote);
1837 break;
1841 #endif
1843 out_activate:
1844 #endif /* CONFIG_SMP */
1845 schedstat_inc(p, se.nr_wakeups);
1846 if (sync)
1847 schedstat_inc(p, se.nr_wakeups_sync);
1848 if (orig_cpu != cpu)
1849 schedstat_inc(p, se.nr_wakeups_migrate);
1850 if (cpu == this_cpu)
1851 schedstat_inc(p, se.nr_wakeups_local);
1852 else
1853 schedstat_inc(p, se.nr_wakeups_remote);
1854 update_rq_clock(rq);
1855 activate_task(rq, p, 1);
1856 check_preempt_curr(rq, p);
1857 success = 1;
1859 out_running:
1860 p->state = TASK_RUNNING;
1861 #ifdef CONFIG_SMP
1862 if (p->sched_class->task_wake_up)
1863 p->sched_class->task_wake_up(rq, p);
1864 #endif
1865 out:
1866 task_rq_unlock(rq, &flags);
1868 return success;
1871 int wake_up_process(struct task_struct *p)
1873 return try_to_wake_up(p, TASK_ALL, 0);
1875 EXPORT_SYMBOL(wake_up_process);
1877 int wake_up_state(struct task_struct *p, unsigned int state)
1879 return try_to_wake_up(p, state, 0);
1883 * Perform scheduler related setup for a newly forked process p.
1884 * p is forked by current.
1886 * __sched_fork() is basic setup used by init_idle() too:
1888 static void __sched_fork(struct task_struct *p)
1890 p->se.exec_start = 0;
1891 p->se.sum_exec_runtime = 0;
1892 p->se.prev_sum_exec_runtime = 0;
1894 #ifdef CONFIG_SCHEDSTATS
1895 p->se.wait_start = 0;
1896 p->se.sum_sleep_runtime = 0;
1897 p->se.sleep_start = 0;
1898 p->se.block_start = 0;
1899 p->se.sleep_max = 0;
1900 p->se.block_max = 0;
1901 p->se.exec_max = 0;
1902 p->se.slice_max = 0;
1903 p->se.wait_max = 0;
1904 #endif
1906 INIT_LIST_HEAD(&p->rt.run_list);
1907 p->se.on_rq = 0;
1909 #ifdef CONFIG_PREEMPT_NOTIFIERS
1910 INIT_HLIST_HEAD(&p->preempt_notifiers);
1911 #endif
1914 * We mark the process as running here, but have not actually
1915 * inserted it onto the runqueue yet. This guarantees that
1916 * nobody will actually run it, and a signal or other external
1917 * event cannot wake it up and insert it on the runqueue either.
1919 p->state = TASK_RUNNING;
1923 * fork()/clone()-time setup:
1925 void sched_fork(struct task_struct *p, int clone_flags)
1927 int cpu = get_cpu();
1929 __sched_fork(p);
1931 #ifdef CONFIG_SMP
1932 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1933 #endif
1934 set_task_cpu(p, cpu);
1937 * Make sure we do not leak PI boosting priority to the child:
1939 p->prio = current->normal_prio;
1940 if (!rt_prio(p->prio))
1941 p->sched_class = &fair_sched_class;
1943 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1944 if (likely(sched_info_on()))
1945 memset(&p->sched_info, 0, sizeof(p->sched_info));
1946 #endif
1947 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1948 p->oncpu = 0;
1949 #endif
1950 #ifdef CONFIG_PREEMPT
1951 /* Want to start with kernel preemption disabled. */
1952 task_thread_info(p)->preempt_count = 1;
1953 #endif
1954 put_cpu();
1958 * wake_up_new_task - wake up a newly created task for the first time.
1960 * This function will do some initial scheduler statistics housekeeping
1961 * that must be done for every newly created context, then puts the task
1962 * on the runqueue and wakes it.
1964 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1966 unsigned long flags;
1967 struct rq *rq;
1969 rq = task_rq_lock(p, &flags);
1970 BUG_ON(p->state != TASK_RUNNING);
1971 update_rq_clock(rq);
1973 p->prio = effective_prio(p);
1975 if (!p->sched_class->task_new || !current->se.on_rq) {
1976 activate_task(rq, p, 0);
1977 } else {
1979 * Let the scheduling class do new task startup
1980 * management (if any):
1982 p->sched_class->task_new(rq, p);
1983 inc_nr_running(p, rq);
1985 check_preempt_curr(rq, p);
1986 #ifdef CONFIG_SMP
1987 if (p->sched_class->task_wake_up)
1988 p->sched_class->task_wake_up(rq, p);
1989 #endif
1990 task_rq_unlock(rq, &flags);
1993 #ifdef CONFIG_PREEMPT_NOTIFIERS
1996 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1997 * @notifier: notifier struct to register
1999 void preempt_notifier_register(struct preempt_notifier *notifier)
2001 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2003 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2006 * preempt_notifier_unregister - no longer interested in preemption notifications
2007 * @notifier: notifier struct to unregister
2009 * This is safe to call from within a preemption notifier.
2011 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2013 hlist_del(&notifier->link);
2015 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2017 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2019 struct preempt_notifier *notifier;
2020 struct hlist_node *node;
2022 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2023 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2026 static void
2027 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2028 struct task_struct *next)
2030 struct preempt_notifier *notifier;
2031 struct hlist_node *node;
2033 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2034 notifier->ops->sched_out(notifier, next);
2037 #else
2039 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2043 static void
2044 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2045 struct task_struct *next)
2049 #endif
2052 * prepare_task_switch - prepare to switch tasks
2053 * @rq: the runqueue preparing to switch
2054 * @prev: the current task that is being switched out
2055 * @next: the task we are going to switch to.
2057 * This is called with the rq lock held and interrupts off. It must
2058 * be paired with a subsequent finish_task_switch after the context
2059 * switch.
2061 * prepare_task_switch sets up locking and calls architecture specific
2062 * hooks.
2064 static inline void
2065 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2066 struct task_struct *next)
2068 fire_sched_out_preempt_notifiers(prev, next);
2069 prepare_lock_switch(rq, next);
2070 prepare_arch_switch(next);
2074 * finish_task_switch - clean up after a task-switch
2075 * @rq: runqueue associated with task-switch
2076 * @prev: the thread we just switched away from.
2078 * finish_task_switch must be called after the context switch, paired
2079 * with a prepare_task_switch call before the context switch.
2080 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2081 * and do any other architecture-specific cleanup actions.
2083 * Note that we may have delayed dropping an mm in context_switch(). If
2084 * so, we finish that here outside of the runqueue lock. (Doing it
2085 * with the lock held can cause deadlocks; see schedule() for
2086 * details.)
2088 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2089 __releases(rq->lock)
2091 struct mm_struct *mm = rq->prev_mm;
2092 long prev_state;
2094 rq->prev_mm = NULL;
2097 * A task struct has one reference for the use as "current".
2098 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2099 * schedule one last time. The schedule call will never return, and
2100 * the scheduled task must drop that reference.
2101 * The test for TASK_DEAD must occur while the runqueue locks are
2102 * still held, otherwise prev could be scheduled on another cpu, die
2103 * there before we look at prev->state, and then the reference would
2104 * be dropped twice.
2105 * Manfred Spraul <manfred@colorfullife.com>
2107 prev_state = prev->state;
2108 finish_arch_switch(prev);
2109 finish_lock_switch(rq, prev);
2110 #ifdef CONFIG_SMP
2111 if (current->sched_class->post_schedule)
2112 current->sched_class->post_schedule(rq);
2113 #endif
2115 fire_sched_in_preempt_notifiers(current);
2116 if (mm)
2117 mmdrop(mm);
2118 if (unlikely(prev_state == TASK_DEAD)) {
2120 * Remove function-return probe instances associated with this
2121 * task and put them back on the free list.
2123 kprobe_flush_task(prev);
2124 put_task_struct(prev);
2129 * schedule_tail - first thing a freshly forked thread must call.
2130 * @prev: the thread we just switched away from.
2132 asmlinkage void schedule_tail(struct task_struct *prev)
2133 __releases(rq->lock)
2135 struct rq *rq = this_rq();
2137 finish_task_switch(rq, prev);
2138 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2139 /* In this case, finish_task_switch does not reenable preemption */
2140 preempt_enable();
2141 #endif
2142 if (current->set_child_tid)
2143 put_user(task_pid_vnr(current), current->set_child_tid);
2147 * context_switch - switch to the new MM and the new
2148 * thread's register state.
2150 static inline void
2151 context_switch(struct rq *rq, struct task_struct *prev,
2152 struct task_struct *next)
2154 struct mm_struct *mm, *oldmm;
2156 prepare_task_switch(rq, prev, next);
2157 mm = next->mm;
2158 oldmm = prev->active_mm;
2160 * For paravirt, this is coupled with an exit in switch_to to
2161 * combine the page table reload and the switch backend into
2162 * one hypercall.
2164 arch_enter_lazy_cpu_mode();
2166 if (unlikely(!mm)) {
2167 next->active_mm = oldmm;
2168 atomic_inc(&oldmm->mm_count);
2169 enter_lazy_tlb(oldmm, next);
2170 } else
2171 switch_mm(oldmm, mm, next);
2173 if (unlikely(!prev->mm)) {
2174 prev->active_mm = NULL;
2175 rq->prev_mm = oldmm;
2178 * Since the runqueue lock will be released by the next
2179 * task (which is an invalid locking op but in the case
2180 * of the scheduler it's an obvious special-case), so we
2181 * do an early lockdep release here:
2183 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2184 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2185 #endif
2187 /* Here we just switch the register state and the stack. */
2188 switch_to(prev, next, prev);
2190 barrier();
2192 * this_rq must be evaluated again because prev may have moved
2193 * CPUs since it called schedule(), thus the 'rq' on its stack
2194 * frame will be invalid.
2196 finish_task_switch(this_rq(), prev);
2200 * nr_running, nr_uninterruptible and nr_context_switches:
2202 * externally visible scheduler statistics: current number of runnable
2203 * threads, current number of uninterruptible-sleeping threads, total
2204 * number of context switches performed since bootup.
2206 unsigned long nr_running(void)
2208 unsigned long i, sum = 0;
2210 for_each_online_cpu(i)
2211 sum += cpu_rq(i)->nr_running;
2213 return sum;
2216 unsigned long nr_uninterruptible(void)
2218 unsigned long i, sum = 0;
2220 for_each_possible_cpu(i)
2221 sum += cpu_rq(i)->nr_uninterruptible;
2224 * Since we read the counters lockless, it might be slightly
2225 * inaccurate. Do not allow it to go below zero though:
2227 if (unlikely((long)sum < 0))
2228 sum = 0;
2230 return sum;
2233 unsigned long long nr_context_switches(void)
2235 int i;
2236 unsigned long long sum = 0;
2238 for_each_possible_cpu(i)
2239 sum += cpu_rq(i)->nr_switches;
2241 return sum;
2244 unsigned long nr_iowait(void)
2246 unsigned long i, sum = 0;
2248 for_each_possible_cpu(i)
2249 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2251 return sum;
2254 unsigned long nr_active(void)
2256 unsigned long i, running = 0, uninterruptible = 0;
2258 for_each_online_cpu(i) {
2259 running += cpu_rq(i)->nr_running;
2260 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2263 if (unlikely((long)uninterruptible < 0))
2264 uninterruptible = 0;
2266 return running + uninterruptible;
2270 * Update rq->cpu_load[] statistics. This function is usually called every
2271 * scheduler tick (TICK_NSEC).
2273 static void update_cpu_load(struct rq *this_rq)
2275 unsigned long this_load = this_rq->load.weight;
2276 int i, scale;
2278 this_rq->nr_load_updates++;
2280 /* Update our load: */
2281 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2282 unsigned long old_load, new_load;
2284 /* scale is effectively 1 << i now, and >> i divides by scale */
2286 old_load = this_rq->cpu_load[i];
2287 new_load = this_load;
2289 * Round up the averaging division if load is increasing. This
2290 * prevents us from getting stuck on 9 if the load is 10, for
2291 * example.
2293 if (new_load > old_load)
2294 new_load += scale-1;
2295 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2299 #ifdef CONFIG_SMP
2302 * double_rq_lock - safely lock two runqueues
2304 * Note this does not disable interrupts like task_rq_lock,
2305 * you need to do so manually before calling.
2307 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2308 __acquires(rq1->lock)
2309 __acquires(rq2->lock)
2311 BUG_ON(!irqs_disabled());
2312 if (rq1 == rq2) {
2313 spin_lock(&rq1->lock);
2314 __acquire(rq2->lock); /* Fake it out ;) */
2315 } else {
2316 if (rq1 < rq2) {
2317 spin_lock(&rq1->lock);
2318 spin_lock(&rq2->lock);
2319 } else {
2320 spin_lock(&rq2->lock);
2321 spin_lock(&rq1->lock);
2324 update_rq_clock(rq1);
2325 update_rq_clock(rq2);
2329 * double_rq_unlock - safely unlock two runqueues
2331 * Note this does not restore interrupts like task_rq_unlock,
2332 * you need to do so manually after calling.
2334 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2335 __releases(rq1->lock)
2336 __releases(rq2->lock)
2338 spin_unlock(&rq1->lock);
2339 if (rq1 != rq2)
2340 spin_unlock(&rq2->lock);
2341 else
2342 __release(rq2->lock);
2346 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2348 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2349 __releases(this_rq->lock)
2350 __acquires(busiest->lock)
2351 __acquires(this_rq->lock)
2353 int ret = 0;
2355 if (unlikely(!irqs_disabled())) {
2356 /* printk() doesn't work good under rq->lock */
2357 spin_unlock(&this_rq->lock);
2358 BUG_ON(1);
2360 if (unlikely(!spin_trylock(&busiest->lock))) {
2361 if (busiest < this_rq) {
2362 spin_unlock(&this_rq->lock);
2363 spin_lock(&busiest->lock);
2364 spin_lock(&this_rq->lock);
2365 ret = 1;
2366 } else
2367 spin_lock(&busiest->lock);
2369 return ret;
2373 * If dest_cpu is allowed for this process, migrate the task to it.
2374 * This is accomplished by forcing the cpu_allowed mask to only
2375 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2376 * the cpu_allowed mask is restored.
2378 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2380 struct migration_req req;
2381 unsigned long flags;
2382 struct rq *rq;
2384 rq = task_rq_lock(p, &flags);
2385 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2386 || unlikely(cpu_is_offline(dest_cpu)))
2387 goto out;
2389 /* force the process onto the specified CPU */
2390 if (migrate_task(p, dest_cpu, &req)) {
2391 /* Need to wait for migration thread (might exit: take ref). */
2392 struct task_struct *mt = rq->migration_thread;
2394 get_task_struct(mt);
2395 task_rq_unlock(rq, &flags);
2396 wake_up_process(mt);
2397 put_task_struct(mt);
2398 wait_for_completion(&req.done);
2400 return;
2402 out:
2403 task_rq_unlock(rq, &flags);
2407 * sched_exec - execve() is a valuable balancing opportunity, because at
2408 * this point the task has the smallest effective memory and cache footprint.
2410 void sched_exec(void)
2412 int new_cpu, this_cpu = get_cpu();
2413 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2414 put_cpu();
2415 if (new_cpu != this_cpu)
2416 sched_migrate_task(current, new_cpu);
2420 * pull_task - move a task from a remote runqueue to the local runqueue.
2421 * Both runqueues must be locked.
2423 static void pull_task(struct rq *src_rq, struct task_struct *p,
2424 struct rq *this_rq, int this_cpu)
2426 deactivate_task(src_rq, p, 0);
2427 set_task_cpu(p, this_cpu);
2428 activate_task(this_rq, p, 0);
2430 * Note that idle threads have a prio of MAX_PRIO, for this test
2431 * to be always true for them.
2433 check_preempt_curr(this_rq, p);
2437 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2439 static
2440 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2441 struct sched_domain *sd, enum cpu_idle_type idle,
2442 int *all_pinned)
2445 * We do not migrate tasks that are:
2446 * 1) running (obviously), or
2447 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2448 * 3) are cache-hot on their current CPU.
2450 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2451 schedstat_inc(p, se.nr_failed_migrations_affine);
2452 return 0;
2454 *all_pinned = 0;
2456 if (task_running(rq, p)) {
2457 schedstat_inc(p, se.nr_failed_migrations_running);
2458 return 0;
2462 * Aggressive migration if:
2463 * 1) task is cache cold, or
2464 * 2) too many balance attempts have failed.
2467 if (!task_hot(p, rq->clock, sd) ||
2468 sd->nr_balance_failed > sd->cache_nice_tries) {
2469 #ifdef CONFIG_SCHEDSTATS
2470 if (task_hot(p, rq->clock, sd)) {
2471 schedstat_inc(sd, lb_hot_gained[idle]);
2472 schedstat_inc(p, se.nr_forced_migrations);
2474 #endif
2475 return 1;
2478 if (task_hot(p, rq->clock, sd)) {
2479 schedstat_inc(p, se.nr_failed_migrations_hot);
2480 return 0;
2482 return 1;
2485 static unsigned long
2486 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2487 unsigned long max_load_move, struct sched_domain *sd,
2488 enum cpu_idle_type idle, int *all_pinned,
2489 int *this_best_prio, struct rq_iterator *iterator)
2491 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2492 struct task_struct *p;
2493 long rem_load_move = max_load_move;
2495 if (max_load_move == 0)
2496 goto out;
2498 pinned = 1;
2501 * Start the load-balancing iterator:
2503 p = iterator->start(iterator->arg);
2504 next:
2505 if (!p || loops++ > sysctl_sched_nr_migrate)
2506 goto out;
2508 * To help distribute high priority tasks across CPUs we don't
2509 * skip a task if it will be the highest priority task (i.e. smallest
2510 * prio value) on its new queue regardless of its load weight
2512 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2513 SCHED_LOAD_SCALE_FUZZ;
2514 if ((skip_for_load && p->prio >= *this_best_prio) ||
2515 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2516 p = iterator->next(iterator->arg);
2517 goto next;
2520 pull_task(busiest, p, this_rq, this_cpu);
2521 pulled++;
2522 rem_load_move -= p->se.load.weight;
2525 * We only want to steal up to the prescribed amount of weighted load.
2527 if (rem_load_move > 0) {
2528 if (p->prio < *this_best_prio)
2529 *this_best_prio = p->prio;
2530 p = iterator->next(iterator->arg);
2531 goto next;
2533 out:
2535 * Right now, this is one of only two places pull_task() is called,
2536 * so we can safely collect pull_task() stats here rather than
2537 * inside pull_task().
2539 schedstat_add(sd, lb_gained[idle], pulled);
2541 if (all_pinned)
2542 *all_pinned = pinned;
2544 return max_load_move - rem_load_move;
2548 * move_tasks tries to move up to max_load_move weighted load from busiest to
2549 * this_rq, as part of a balancing operation within domain "sd".
2550 * Returns 1 if successful and 0 otherwise.
2552 * Called with both runqueues locked.
2554 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2555 unsigned long max_load_move,
2556 struct sched_domain *sd, enum cpu_idle_type idle,
2557 int *all_pinned)
2559 const struct sched_class *class = sched_class_highest;
2560 unsigned long total_load_moved = 0;
2561 int this_best_prio = this_rq->curr->prio;
2563 do {
2564 total_load_moved +=
2565 class->load_balance(this_rq, this_cpu, busiest,
2566 max_load_move - total_load_moved,
2567 sd, idle, all_pinned, &this_best_prio);
2568 class = class->next;
2569 } while (class && max_load_move > total_load_moved);
2571 return total_load_moved > 0;
2574 static int
2575 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2576 struct sched_domain *sd, enum cpu_idle_type idle,
2577 struct rq_iterator *iterator)
2579 struct task_struct *p = iterator->start(iterator->arg);
2580 int pinned = 0;
2582 while (p) {
2583 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2584 pull_task(busiest, p, this_rq, this_cpu);
2586 * Right now, this is only the second place pull_task()
2587 * is called, so we can safely collect pull_task()
2588 * stats here rather than inside pull_task().
2590 schedstat_inc(sd, lb_gained[idle]);
2592 return 1;
2594 p = iterator->next(iterator->arg);
2597 return 0;
2601 * move_one_task tries to move exactly one task from busiest to this_rq, as
2602 * part of active balancing operations within "domain".
2603 * Returns 1 if successful and 0 otherwise.
2605 * Called with both runqueues locked.
2607 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2608 struct sched_domain *sd, enum cpu_idle_type idle)
2610 const struct sched_class *class;
2612 for (class = sched_class_highest; class; class = class->next)
2613 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2614 return 1;
2616 return 0;
2620 * find_busiest_group finds and returns the busiest CPU group within the
2621 * domain. It calculates and returns the amount of weighted load which
2622 * should be moved to restore balance via the imbalance parameter.
2624 static struct sched_group *
2625 find_busiest_group(struct sched_domain *sd, int this_cpu,
2626 unsigned long *imbalance, enum cpu_idle_type idle,
2627 int *sd_idle, cpumask_t *cpus, int *balance)
2629 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2630 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2631 unsigned long max_pull;
2632 unsigned long busiest_load_per_task, busiest_nr_running;
2633 unsigned long this_load_per_task, this_nr_running;
2634 int load_idx, group_imb = 0;
2635 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2636 int power_savings_balance = 1;
2637 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2638 unsigned long min_nr_running = ULONG_MAX;
2639 struct sched_group *group_min = NULL, *group_leader = NULL;
2640 #endif
2642 max_load = this_load = total_load = total_pwr = 0;
2643 busiest_load_per_task = busiest_nr_running = 0;
2644 this_load_per_task = this_nr_running = 0;
2645 if (idle == CPU_NOT_IDLE)
2646 load_idx = sd->busy_idx;
2647 else if (idle == CPU_NEWLY_IDLE)
2648 load_idx = sd->newidle_idx;
2649 else
2650 load_idx = sd->idle_idx;
2652 do {
2653 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2654 int local_group;
2655 int i;
2656 int __group_imb = 0;
2657 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2658 unsigned long sum_nr_running, sum_weighted_load;
2660 local_group = cpu_isset(this_cpu, group->cpumask);
2662 if (local_group)
2663 balance_cpu = first_cpu(group->cpumask);
2665 /* Tally up the load of all CPUs in the group */
2666 sum_weighted_load = sum_nr_running = avg_load = 0;
2667 max_cpu_load = 0;
2668 min_cpu_load = ~0UL;
2670 for_each_cpu_mask(i, group->cpumask) {
2671 struct rq *rq;
2673 if (!cpu_isset(i, *cpus))
2674 continue;
2676 rq = cpu_rq(i);
2678 if (*sd_idle && rq->nr_running)
2679 *sd_idle = 0;
2681 /* Bias balancing toward cpus of our domain */
2682 if (local_group) {
2683 if (idle_cpu(i) && !first_idle_cpu) {
2684 first_idle_cpu = 1;
2685 balance_cpu = i;
2688 load = target_load(i, load_idx);
2689 } else {
2690 load = source_load(i, load_idx);
2691 if (load > max_cpu_load)
2692 max_cpu_load = load;
2693 if (min_cpu_load > load)
2694 min_cpu_load = load;
2697 avg_load += load;
2698 sum_nr_running += rq->nr_running;
2699 sum_weighted_load += weighted_cpuload(i);
2703 * First idle cpu or the first cpu(busiest) in this sched group
2704 * is eligible for doing load balancing at this and above
2705 * domains. In the newly idle case, we will allow all the cpu's
2706 * to do the newly idle load balance.
2708 if (idle != CPU_NEWLY_IDLE && local_group &&
2709 balance_cpu != this_cpu && balance) {
2710 *balance = 0;
2711 goto ret;
2714 total_load += avg_load;
2715 total_pwr += group->__cpu_power;
2717 /* Adjust by relative CPU power of the group */
2718 avg_load = sg_div_cpu_power(group,
2719 avg_load * SCHED_LOAD_SCALE);
2721 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2722 __group_imb = 1;
2724 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2726 if (local_group) {
2727 this_load = avg_load;
2728 this = group;
2729 this_nr_running = sum_nr_running;
2730 this_load_per_task = sum_weighted_load;
2731 } else if (avg_load > max_load &&
2732 (sum_nr_running > group_capacity || __group_imb)) {
2733 max_load = avg_load;
2734 busiest = group;
2735 busiest_nr_running = sum_nr_running;
2736 busiest_load_per_task = sum_weighted_load;
2737 group_imb = __group_imb;
2740 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2742 * Busy processors will not participate in power savings
2743 * balance.
2745 if (idle == CPU_NOT_IDLE ||
2746 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2747 goto group_next;
2750 * If the local group is idle or completely loaded
2751 * no need to do power savings balance at this domain
2753 if (local_group && (this_nr_running >= group_capacity ||
2754 !this_nr_running))
2755 power_savings_balance = 0;
2758 * If a group is already running at full capacity or idle,
2759 * don't include that group in power savings calculations
2761 if (!power_savings_balance || sum_nr_running >= group_capacity
2762 || !sum_nr_running)
2763 goto group_next;
2766 * Calculate the group which has the least non-idle load.
2767 * This is the group from where we need to pick up the load
2768 * for saving power
2770 if ((sum_nr_running < min_nr_running) ||
2771 (sum_nr_running == min_nr_running &&
2772 first_cpu(group->cpumask) <
2773 first_cpu(group_min->cpumask))) {
2774 group_min = group;
2775 min_nr_running = sum_nr_running;
2776 min_load_per_task = sum_weighted_load /
2777 sum_nr_running;
2781 * Calculate the group which is almost near its
2782 * capacity but still has some space to pick up some load
2783 * from other group and save more power
2785 if (sum_nr_running <= group_capacity - 1) {
2786 if (sum_nr_running > leader_nr_running ||
2787 (sum_nr_running == leader_nr_running &&
2788 first_cpu(group->cpumask) >
2789 first_cpu(group_leader->cpumask))) {
2790 group_leader = group;
2791 leader_nr_running = sum_nr_running;
2794 group_next:
2795 #endif
2796 group = group->next;
2797 } while (group != sd->groups);
2799 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2800 goto out_balanced;
2802 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2804 if (this_load >= avg_load ||
2805 100*max_load <= sd->imbalance_pct*this_load)
2806 goto out_balanced;
2808 busiest_load_per_task /= busiest_nr_running;
2809 if (group_imb)
2810 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2813 * We're trying to get all the cpus to the average_load, so we don't
2814 * want to push ourselves above the average load, nor do we wish to
2815 * reduce the max loaded cpu below the average load, as either of these
2816 * actions would just result in more rebalancing later, and ping-pong
2817 * tasks around. Thus we look for the minimum possible imbalance.
2818 * Negative imbalances (*we* are more loaded than anyone else) will
2819 * be counted as no imbalance for these purposes -- we can't fix that
2820 * by pulling tasks to us. Be careful of negative numbers as they'll
2821 * appear as very large values with unsigned longs.
2823 if (max_load <= busiest_load_per_task)
2824 goto out_balanced;
2827 * In the presence of smp nice balancing, certain scenarios can have
2828 * max load less than avg load(as we skip the groups at or below
2829 * its cpu_power, while calculating max_load..)
2831 if (max_load < avg_load) {
2832 *imbalance = 0;
2833 goto small_imbalance;
2836 /* Don't want to pull so many tasks that a group would go idle */
2837 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2839 /* How much load to actually move to equalise the imbalance */
2840 *imbalance = min(max_pull * busiest->__cpu_power,
2841 (avg_load - this_load) * this->__cpu_power)
2842 / SCHED_LOAD_SCALE;
2845 * if *imbalance is less than the average load per runnable task
2846 * there is no gaurantee that any tasks will be moved so we'll have
2847 * a think about bumping its value to force at least one task to be
2848 * moved
2850 if (*imbalance < busiest_load_per_task) {
2851 unsigned long tmp, pwr_now, pwr_move;
2852 unsigned int imbn;
2854 small_imbalance:
2855 pwr_move = pwr_now = 0;
2856 imbn = 2;
2857 if (this_nr_running) {
2858 this_load_per_task /= this_nr_running;
2859 if (busiest_load_per_task > this_load_per_task)
2860 imbn = 1;
2861 } else
2862 this_load_per_task = SCHED_LOAD_SCALE;
2864 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2865 busiest_load_per_task * imbn) {
2866 *imbalance = busiest_load_per_task;
2867 return busiest;
2871 * OK, we don't have enough imbalance to justify moving tasks,
2872 * however we may be able to increase total CPU power used by
2873 * moving them.
2876 pwr_now += busiest->__cpu_power *
2877 min(busiest_load_per_task, max_load);
2878 pwr_now += this->__cpu_power *
2879 min(this_load_per_task, this_load);
2880 pwr_now /= SCHED_LOAD_SCALE;
2882 /* Amount of load we'd subtract */
2883 tmp = sg_div_cpu_power(busiest,
2884 busiest_load_per_task * SCHED_LOAD_SCALE);
2885 if (max_load > tmp)
2886 pwr_move += busiest->__cpu_power *
2887 min(busiest_load_per_task, max_load - tmp);
2889 /* Amount of load we'd add */
2890 if (max_load * busiest->__cpu_power <
2891 busiest_load_per_task * SCHED_LOAD_SCALE)
2892 tmp = sg_div_cpu_power(this,
2893 max_load * busiest->__cpu_power);
2894 else
2895 tmp = sg_div_cpu_power(this,
2896 busiest_load_per_task * SCHED_LOAD_SCALE);
2897 pwr_move += this->__cpu_power *
2898 min(this_load_per_task, this_load + tmp);
2899 pwr_move /= SCHED_LOAD_SCALE;
2901 /* Move if we gain throughput */
2902 if (pwr_move > pwr_now)
2903 *imbalance = busiest_load_per_task;
2906 return busiest;
2908 out_balanced:
2909 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2910 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2911 goto ret;
2913 if (this == group_leader && group_leader != group_min) {
2914 *imbalance = min_load_per_task;
2915 return group_min;
2917 #endif
2918 ret:
2919 *imbalance = 0;
2920 return NULL;
2924 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2926 static struct rq *
2927 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2928 unsigned long imbalance, cpumask_t *cpus)
2930 struct rq *busiest = NULL, *rq;
2931 unsigned long max_load = 0;
2932 int i;
2934 for_each_cpu_mask(i, group->cpumask) {
2935 unsigned long wl;
2937 if (!cpu_isset(i, *cpus))
2938 continue;
2940 rq = cpu_rq(i);
2941 wl = weighted_cpuload(i);
2943 if (rq->nr_running == 1 && wl > imbalance)
2944 continue;
2946 if (wl > max_load) {
2947 max_load = wl;
2948 busiest = rq;
2952 return busiest;
2956 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2957 * so long as it is large enough.
2959 #define MAX_PINNED_INTERVAL 512
2962 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2963 * tasks if there is an imbalance.
2965 static int load_balance(int this_cpu, struct rq *this_rq,
2966 struct sched_domain *sd, enum cpu_idle_type idle,
2967 int *balance)
2969 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2970 struct sched_group *group;
2971 unsigned long imbalance;
2972 struct rq *busiest;
2973 cpumask_t cpus = CPU_MASK_ALL;
2974 unsigned long flags;
2977 * When power savings policy is enabled for the parent domain, idle
2978 * sibling can pick up load irrespective of busy siblings. In this case,
2979 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2980 * portraying it as CPU_NOT_IDLE.
2982 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2983 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2984 sd_idle = 1;
2986 schedstat_inc(sd, lb_count[idle]);
2988 redo:
2989 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2990 &cpus, balance);
2992 if (*balance == 0)
2993 goto out_balanced;
2995 if (!group) {
2996 schedstat_inc(sd, lb_nobusyg[idle]);
2997 goto out_balanced;
3000 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
3001 if (!busiest) {
3002 schedstat_inc(sd, lb_nobusyq[idle]);
3003 goto out_balanced;
3006 BUG_ON(busiest == this_rq);
3008 schedstat_add(sd, lb_imbalance[idle], imbalance);
3010 ld_moved = 0;
3011 if (busiest->nr_running > 1) {
3013 * Attempt to move tasks. If find_busiest_group has found
3014 * an imbalance but busiest->nr_running <= 1, the group is
3015 * still unbalanced. ld_moved simply stays zero, so it is
3016 * correctly treated as an imbalance.
3018 local_irq_save(flags);
3019 double_rq_lock(this_rq, busiest);
3020 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3021 imbalance, sd, idle, &all_pinned);
3022 double_rq_unlock(this_rq, busiest);
3023 local_irq_restore(flags);
3026 * some other cpu did the load balance for us.
3028 if (ld_moved && this_cpu != smp_processor_id())
3029 resched_cpu(this_cpu);
3031 /* All tasks on this runqueue were pinned by CPU affinity */
3032 if (unlikely(all_pinned)) {
3033 cpu_clear(cpu_of(busiest), cpus);
3034 if (!cpus_empty(cpus))
3035 goto redo;
3036 goto out_balanced;
3040 if (!ld_moved) {
3041 schedstat_inc(sd, lb_failed[idle]);
3042 sd->nr_balance_failed++;
3044 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3046 spin_lock_irqsave(&busiest->lock, flags);
3048 /* don't kick the migration_thread, if the curr
3049 * task on busiest cpu can't be moved to this_cpu
3051 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3052 spin_unlock_irqrestore(&busiest->lock, flags);
3053 all_pinned = 1;
3054 goto out_one_pinned;
3057 if (!busiest->active_balance) {
3058 busiest->active_balance = 1;
3059 busiest->push_cpu = this_cpu;
3060 active_balance = 1;
3062 spin_unlock_irqrestore(&busiest->lock, flags);
3063 if (active_balance)
3064 wake_up_process(busiest->migration_thread);
3067 * We've kicked active balancing, reset the failure
3068 * counter.
3070 sd->nr_balance_failed = sd->cache_nice_tries+1;
3072 } else
3073 sd->nr_balance_failed = 0;
3075 if (likely(!active_balance)) {
3076 /* We were unbalanced, so reset the balancing interval */
3077 sd->balance_interval = sd->min_interval;
3078 } else {
3080 * If we've begun active balancing, start to back off. This
3081 * case may not be covered by the all_pinned logic if there
3082 * is only 1 task on the busy runqueue (because we don't call
3083 * move_tasks).
3085 if (sd->balance_interval < sd->max_interval)
3086 sd->balance_interval *= 2;
3089 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3090 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3091 return -1;
3092 return ld_moved;
3094 out_balanced:
3095 schedstat_inc(sd, lb_balanced[idle]);
3097 sd->nr_balance_failed = 0;
3099 out_one_pinned:
3100 /* tune up the balancing interval */
3101 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3102 (sd->balance_interval < sd->max_interval))
3103 sd->balance_interval *= 2;
3105 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3106 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3107 return -1;
3108 return 0;
3112 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3113 * tasks if there is an imbalance.
3115 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3116 * this_rq is locked.
3118 static int
3119 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3121 struct sched_group *group;
3122 struct rq *busiest = NULL;
3123 unsigned long imbalance;
3124 int ld_moved = 0;
3125 int sd_idle = 0;
3126 int all_pinned = 0;
3127 cpumask_t cpus = CPU_MASK_ALL;
3130 * When power savings policy is enabled for the parent domain, idle
3131 * sibling can pick up load irrespective of busy siblings. In this case,
3132 * let the state of idle sibling percolate up as IDLE, instead of
3133 * portraying it as CPU_NOT_IDLE.
3135 if (sd->flags & SD_SHARE_CPUPOWER &&
3136 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3137 sd_idle = 1;
3139 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3140 redo:
3141 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3142 &sd_idle, &cpus, NULL);
3143 if (!group) {
3144 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3145 goto out_balanced;
3148 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
3149 &cpus);
3150 if (!busiest) {
3151 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3152 goto out_balanced;
3155 BUG_ON(busiest == this_rq);
3157 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3159 ld_moved = 0;
3160 if (busiest->nr_running > 1) {
3161 /* Attempt to move tasks */
3162 double_lock_balance(this_rq, busiest);
3163 /* this_rq->clock is already updated */
3164 update_rq_clock(busiest);
3165 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3166 imbalance, sd, CPU_NEWLY_IDLE,
3167 &all_pinned);
3168 spin_unlock(&busiest->lock);
3170 if (unlikely(all_pinned)) {
3171 cpu_clear(cpu_of(busiest), cpus);
3172 if (!cpus_empty(cpus))
3173 goto redo;
3177 if (!ld_moved) {
3178 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3179 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3180 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3181 return -1;
3182 } else
3183 sd->nr_balance_failed = 0;
3185 return ld_moved;
3187 out_balanced:
3188 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3189 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3190 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3191 return -1;
3192 sd->nr_balance_failed = 0;
3194 return 0;
3198 * idle_balance is called by schedule() if this_cpu is about to become
3199 * idle. Attempts to pull tasks from other CPUs.
3201 static void idle_balance(int this_cpu, struct rq *this_rq)
3203 struct sched_domain *sd;
3204 int pulled_task = -1;
3205 unsigned long next_balance = jiffies + HZ;
3207 for_each_domain(this_cpu, sd) {
3208 unsigned long interval;
3210 if (!(sd->flags & SD_LOAD_BALANCE))
3211 continue;
3213 if (sd->flags & SD_BALANCE_NEWIDLE)
3214 /* If we've pulled tasks over stop searching: */
3215 pulled_task = load_balance_newidle(this_cpu,
3216 this_rq, sd);
3218 interval = msecs_to_jiffies(sd->balance_interval);
3219 if (time_after(next_balance, sd->last_balance + interval))
3220 next_balance = sd->last_balance + interval;
3221 if (pulled_task)
3222 break;
3224 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3226 * We are going idle. next_balance may be set based on
3227 * a busy processor. So reset next_balance.
3229 this_rq->next_balance = next_balance;
3234 * active_load_balance is run by migration threads. It pushes running tasks
3235 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3236 * running on each physical CPU where possible, and avoids physical /
3237 * logical imbalances.
3239 * Called with busiest_rq locked.
3241 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3243 int target_cpu = busiest_rq->push_cpu;
3244 struct sched_domain *sd;
3245 struct rq *target_rq;
3247 /* Is there any task to move? */
3248 if (busiest_rq->nr_running <= 1)
3249 return;
3251 target_rq = cpu_rq(target_cpu);
3254 * This condition is "impossible", if it occurs
3255 * we need to fix it. Originally reported by
3256 * Bjorn Helgaas on a 128-cpu setup.
3258 BUG_ON(busiest_rq == target_rq);
3260 /* move a task from busiest_rq to target_rq */
3261 double_lock_balance(busiest_rq, target_rq);
3262 update_rq_clock(busiest_rq);
3263 update_rq_clock(target_rq);
3265 /* Search for an sd spanning us and the target CPU. */
3266 for_each_domain(target_cpu, sd) {
3267 if ((sd->flags & SD_LOAD_BALANCE) &&
3268 cpu_isset(busiest_cpu, sd->span))
3269 break;
3272 if (likely(sd)) {
3273 schedstat_inc(sd, alb_count);
3275 if (move_one_task(target_rq, target_cpu, busiest_rq,
3276 sd, CPU_IDLE))
3277 schedstat_inc(sd, alb_pushed);
3278 else
3279 schedstat_inc(sd, alb_failed);
3281 spin_unlock(&target_rq->lock);
3284 #ifdef CONFIG_NO_HZ
3285 static struct {
3286 atomic_t load_balancer;
3287 cpumask_t cpu_mask;
3288 } nohz ____cacheline_aligned = {
3289 .load_balancer = ATOMIC_INIT(-1),
3290 .cpu_mask = CPU_MASK_NONE,
3294 * This routine will try to nominate the ilb (idle load balancing)
3295 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3296 * load balancing on behalf of all those cpus. If all the cpus in the system
3297 * go into this tickless mode, then there will be no ilb owner (as there is
3298 * no need for one) and all the cpus will sleep till the next wakeup event
3299 * arrives...
3301 * For the ilb owner, tick is not stopped. And this tick will be used
3302 * for idle load balancing. ilb owner will still be part of
3303 * nohz.cpu_mask..
3305 * While stopping the tick, this cpu will become the ilb owner if there
3306 * is no other owner. And will be the owner till that cpu becomes busy
3307 * or if all cpus in the system stop their ticks at which point
3308 * there is no need for ilb owner.
3310 * When the ilb owner becomes busy, it nominates another owner, during the
3311 * next busy scheduler_tick()
3313 int select_nohz_load_balancer(int stop_tick)
3315 int cpu = smp_processor_id();
3317 if (stop_tick) {
3318 cpu_set(cpu, nohz.cpu_mask);
3319 cpu_rq(cpu)->in_nohz_recently = 1;
3322 * If we are going offline and still the leader, give up!
3324 if (cpu_is_offline(cpu) &&
3325 atomic_read(&nohz.load_balancer) == cpu) {
3326 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3327 BUG();
3328 return 0;
3331 /* time for ilb owner also to sleep */
3332 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3333 if (atomic_read(&nohz.load_balancer) == cpu)
3334 atomic_set(&nohz.load_balancer, -1);
3335 return 0;
3338 if (atomic_read(&nohz.load_balancer) == -1) {
3339 /* make me the ilb owner */
3340 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3341 return 1;
3342 } else if (atomic_read(&nohz.load_balancer) == cpu)
3343 return 1;
3344 } else {
3345 if (!cpu_isset(cpu, nohz.cpu_mask))
3346 return 0;
3348 cpu_clear(cpu, nohz.cpu_mask);
3350 if (atomic_read(&nohz.load_balancer) == cpu)
3351 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3352 BUG();
3354 return 0;
3356 #endif
3358 static DEFINE_SPINLOCK(balancing);
3361 * It checks each scheduling domain to see if it is due to be balanced,
3362 * and initiates a balancing operation if so.
3364 * Balancing parameters are set up in arch_init_sched_domains.
3366 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3368 int balance = 1;
3369 struct rq *rq = cpu_rq(cpu);
3370 unsigned long interval;
3371 struct sched_domain *sd;
3372 /* Earliest time when we have to do rebalance again */
3373 unsigned long next_balance = jiffies + 60*HZ;
3374 int update_next_balance = 0;
3376 for_each_domain(cpu, sd) {
3377 if (!(sd->flags & SD_LOAD_BALANCE))
3378 continue;
3380 interval = sd->balance_interval;
3381 if (idle != CPU_IDLE)
3382 interval *= sd->busy_factor;
3384 /* scale ms to jiffies */
3385 interval = msecs_to_jiffies(interval);
3386 if (unlikely(!interval))
3387 interval = 1;
3388 if (interval > HZ*NR_CPUS/10)
3389 interval = HZ*NR_CPUS/10;
3392 if (sd->flags & SD_SERIALIZE) {
3393 if (!spin_trylock(&balancing))
3394 goto out;
3397 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3398 if (load_balance(cpu, rq, sd, idle, &balance)) {
3400 * We've pulled tasks over so either we're no
3401 * longer idle, or one of our SMT siblings is
3402 * not idle.
3404 idle = CPU_NOT_IDLE;
3406 sd->last_balance = jiffies;
3408 if (sd->flags & SD_SERIALIZE)
3409 spin_unlock(&balancing);
3410 out:
3411 if (time_after(next_balance, sd->last_balance + interval)) {
3412 next_balance = sd->last_balance + interval;
3413 update_next_balance = 1;
3417 * Stop the load balance at this level. There is another
3418 * CPU in our sched group which is doing load balancing more
3419 * actively.
3421 if (!balance)
3422 break;
3426 * next_balance will be updated only when there is a need.
3427 * When the cpu is attached to null domain for ex, it will not be
3428 * updated.
3430 if (likely(update_next_balance))
3431 rq->next_balance = next_balance;
3435 * run_rebalance_domains is triggered when needed from the scheduler tick.
3436 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3437 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3439 static void run_rebalance_domains(struct softirq_action *h)
3441 int this_cpu = smp_processor_id();
3442 struct rq *this_rq = cpu_rq(this_cpu);
3443 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3444 CPU_IDLE : CPU_NOT_IDLE;
3446 rebalance_domains(this_cpu, idle);
3448 #ifdef CONFIG_NO_HZ
3450 * If this cpu is the owner for idle load balancing, then do the
3451 * balancing on behalf of the other idle cpus whose ticks are
3452 * stopped.
3454 if (this_rq->idle_at_tick &&
3455 atomic_read(&nohz.load_balancer) == this_cpu) {
3456 cpumask_t cpus = nohz.cpu_mask;
3457 struct rq *rq;
3458 int balance_cpu;
3460 cpu_clear(this_cpu, cpus);
3461 for_each_cpu_mask(balance_cpu, cpus) {
3463 * If this cpu gets work to do, stop the load balancing
3464 * work being done for other cpus. Next load
3465 * balancing owner will pick it up.
3467 if (need_resched())
3468 break;
3470 rebalance_domains(balance_cpu, CPU_IDLE);
3472 rq = cpu_rq(balance_cpu);
3473 if (time_after(this_rq->next_balance, rq->next_balance))
3474 this_rq->next_balance = rq->next_balance;
3477 #endif
3481 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3483 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3484 * idle load balancing owner or decide to stop the periodic load balancing,
3485 * if the whole system is idle.
3487 static inline void trigger_load_balance(struct rq *rq, int cpu)
3489 #ifdef CONFIG_NO_HZ
3491 * If we were in the nohz mode recently and busy at the current
3492 * scheduler tick, then check if we need to nominate new idle
3493 * load balancer.
3495 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3496 rq->in_nohz_recently = 0;
3498 if (atomic_read(&nohz.load_balancer) == cpu) {
3499 cpu_clear(cpu, nohz.cpu_mask);
3500 atomic_set(&nohz.load_balancer, -1);
3503 if (atomic_read(&nohz.load_balancer) == -1) {
3505 * simple selection for now: Nominate the
3506 * first cpu in the nohz list to be the next
3507 * ilb owner.
3509 * TBD: Traverse the sched domains and nominate
3510 * the nearest cpu in the nohz.cpu_mask.
3512 int ilb = first_cpu(nohz.cpu_mask);
3514 if (ilb != NR_CPUS)
3515 resched_cpu(ilb);
3520 * If this cpu is idle and doing idle load balancing for all the
3521 * cpus with ticks stopped, is it time for that to stop?
3523 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3524 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3525 resched_cpu(cpu);
3526 return;
3530 * If this cpu is idle and the idle load balancing is done by
3531 * someone else, then no need raise the SCHED_SOFTIRQ
3533 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3534 cpu_isset(cpu, nohz.cpu_mask))
3535 return;
3536 #endif
3537 if (time_after_eq(jiffies, rq->next_balance))
3538 raise_softirq(SCHED_SOFTIRQ);
3541 #else /* CONFIG_SMP */
3544 * on UP we do not need to balance between CPUs:
3546 static inline void idle_balance(int cpu, struct rq *rq)
3550 #endif
3552 DEFINE_PER_CPU(struct kernel_stat, kstat);
3554 EXPORT_PER_CPU_SYMBOL(kstat);
3557 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3558 * that have not yet been banked in case the task is currently running.
3560 unsigned long long task_sched_runtime(struct task_struct *p)
3562 unsigned long flags;
3563 u64 ns, delta_exec;
3564 struct rq *rq;
3566 rq = task_rq_lock(p, &flags);
3567 ns = p->se.sum_exec_runtime;
3568 if (task_current(rq, p)) {
3569 update_rq_clock(rq);
3570 delta_exec = rq->clock - p->se.exec_start;
3571 if ((s64)delta_exec > 0)
3572 ns += delta_exec;
3574 task_rq_unlock(rq, &flags);
3576 return ns;
3580 * Account user cpu time to a process.
3581 * @p: the process that the cpu time gets accounted to
3582 * @cputime: the cpu time spent in user space since the last update
3584 void account_user_time(struct task_struct *p, cputime_t cputime)
3586 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3587 cputime64_t tmp;
3589 p->utime = cputime_add(p->utime, cputime);
3591 /* Add user time to cpustat. */
3592 tmp = cputime_to_cputime64(cputime);
3593 if (TASK_NICE(p) > 0)
3594 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3595 else
3596 cpustat->user = cputime64_add(cpustat->user, tmp);
3600 * Account guest cpu time to a process.
3601 * @p: the process that the cpu time gets accounted to
3602 * @cputime: the cpu time spent in virtual machine since the last update
3604 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3606 cputime64_t tmp;
3607 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3609 tmp = cputime_to_cputime64(cputime);
3611 p->utime = cputime_add(p->utime, cputime);
3612 p->gtime = cputime_add(p->gtime, cputime);
3614 cpustat->user = cputime64_add(cpustat->user, tmp);
3615 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3619 * Account scaled user cpu time to a process.
3620 * @p: the process that the cpu time gets accounted to
3621 * @cputime: the cpu time spent in user space since the last update
3623 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3625 p->utimescaled = cputime_add(p->utimescaled, cputime);
3629 * Account system cpu time to a process.
3630 * @p: the process that the cpu time gets accounted to
3631 * @hardirq_offset: the offset to subtract from hardirq_count()
3632 * @cputime: the cpu time spent in kernel space since the last update
3634 void account_system_time(struct task_struct *p, int hardirq_offset,
3635 cputime_t cputime)
3637 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3638 struct rq *rq = this_rq();
3639 cputime64_t tmp;
3641 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3642 return account_guest_time(p, cputime);
3644 p->stime = cputime_add(p->stime, cputime);
3646 /* Add system time to cpustat. */
3647 tmp = cputime_to_cputime64(cputime);
3648 if (hardirq_count() - hardirq_offset)
3649 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3650 else if (softirq_count())
3651 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3652 else if (p != rq->idle)
3653 cpustat->system = cputime64_add(cpustat->system, tmp);
3654 else if (atomic_read(&rq->nr_iowait) > 0)
3655 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3656 else
3657 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3658 /* Account for system time used */
3659 acct_update_integrals(p);
3663 * Account scaled system cpu time to a process.
3664 * @p: the process that the cpu time gets accounted to
3665 * @hardirq_offset: the offset to subtract from hardirq_count()
3666 * @cputime: the cpu time spent in kernel space since the last update
3668 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3670 p->stimescaled = cputime_add(p->stimescaled, cputime);
3674 * Account for involuntary wait time.
3675 * @p: the process from which the cpu time has been stolen
3676 * @steal: the cpu time spent in involuntary wait
3678 void account_steal_time(struct task_struct *p, cputime_t steal)
3680 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3681 cputime64_t tmp = cputime_to_cputime64(steal);
3682 struct rq *rq = this_rq();
3684 if (p == rq->idle) {
3685 p->stime = cputime_add(p->stime, steal);
3686 if (atomic_read(&rq->nr_iowait) > 0)
3687 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3688 else
3689 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3690 } else
3691 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3695 * This function gets called by the timer code, with HZ frequency.
3696 * We call it with interrupts disabled.
3698 * It also gets called by the fork code, when changing the parent's
3699 * timeslices.
3701 void scheduler_tick(void)
3703 int cpu = smp_processor_id();
3704 struct rq *rq = cpu_rq(cpu);
3705 struct task_struct *curr = rq->curr;
3706 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3708 spin_lock(&rq->lock);
3709 __update_rq_clock(rq);
3711 * Let rq->clock advance by at least TICK_NSEC:
3713 if (unlikely(rq->clock < next_tick)) {
3714 rq->clock = next_tick;
3715 rq->clock_underflows++;
3717 rq->tick_timestamp = rq->clock;
3718 update_cpu_load(rq);
3719 curr->sched_class->task_tick(rq, curr, 0);
3720 update_sched_rt_period(rq);
3721 spin_unlock(&rq->lock);
3723 #ifdef CONFIG_SMP
3724 rq->idle_at_tick = idle_cpu(cpu);
3725 trigger_load_balance(rq, cpu);
3726 #endif
3729 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3731 void __kprobes add_preempt_count(int val)
3734 * Underflow?
3736 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3737 return;
3738 preempt_count() += val;
3740 * Spinlock count overflowing soon?
3742 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3743 PREEMPT_MASK - 10);
3745 EXPORT_SYMBOL(add_preempt_count);
3747 void __kprobes sub_preempt_count(int val)
3750 * Underflow?
3752 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3753 return;
3755 * Is the spinlock portion underflowing?
3757 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3758 !(preempt_count() & PREEMPT_MASK)))
3759 return;
3761 preempt_count() -= val;
3763 EXPORT_SYMBOL(sub_preempt_count);
3765 #endif
3768 * Print scheduling while atomic bug:
3770 static noinline void __schedule_bug(struct task_struct *prev)
3772 struct pt_regs *regs = get_irq_regs();
3774 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3775 prev->comm, prev->pid, preempt_count());
3777 debug_show_held_locks(prev);
3778 if (irqs_disabled())
3779 print_irqtrace_events(prev);
3781 if (regs)
3782 show_regs(regs);
3783 else
3784 dump_stack();
3788 * Various schedule()-time debugging checks and statistics:
3790 static inline void schedule_debug(struct task_struct *prev)
3793 * Test if we are atomic. Since do_exit() needs to call into
3794 * schedule() atomically, we ignore that path for now.
3795 * Otherwise, whine if we are scheduling when we should not be.
3797 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3798 __schedule_bug(prev);
3800 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3802 schedstat_inc(this_rq(), sched_count);
3803 #ifdef CONFIG_SCHEDSTATS
3804 if (unlikely(prev->lock_depth >= 0)) {
3805 schedstat_inc(this_rq(), bkl_count);
3806 schedstat_inc(prev, sched_info.bkl_count);
3808 #endif
3812 * Pick up the highest-prio task:
3814 static inline struct task_struct *
3815 pick_next_task(struct rq *rq, struct task_struct *prev)
3817 const struct sched_class *class;
3818 struct task_struct *p;
3821 * Optimization: we know that if all tasks are in
3822 * the fair class we can call that function directly:
3824 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3825 p = fair_sched_class.pick_next_task(rq);
3826 if (likely(p))
3827 return p;
3830 class = sched_class_highest;
3831 for ( ; ; ) {
3832 p = class->pick_next_task(rq);
3833 if (p)
3834 return p;
3836 * Will never be NULL as the idle class always
3837 * returns a non-NULL p:
3839 class = class->next;
3844 * schedule() is the main scheduler function.
3846 asmlinkage void __sched schedule(void)
3848 struct task_struct *prev, *next;
3849 unsigned long *switch_count;
3850 struct rq *rq;
3851 int cpu;
3853 need_resched:
3854 preempt_disable();
3855 cpu = smp_processor_id();
3856 rq = cpu_rq(cpu);
3857 rcu_qsctr_inc(cpu);
3858 prev = rq->curr;
3859 switch_count = &prev->nivcsw;
3861 release_kernel_lock(prev);
3862 need_resched_nonpreemptible:
3864 schedule_debug(prev);
3866 hrtick_clear(rq);
3869 * Do the rq-clock update outside the rq lock:
3871 local_irq_disable();
3872 __update_rq_clock(rq);
3873 spin_lock(&rq->lock);
3874 clear_tsk_need_resched(prev);
3876 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3877 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3878 unlikely(signal_pending(prev)))) {
3879 prev->state = TASK_RUNNING;
3880 } else {
3881 deactivate_task(rq, prev, 1);
3883 switch_count = &prev->nvcsw;
3886 #ifdef CONFIG_SMP
3887 if (prev->sched_class->pre_schedule)
3888 prev->sched_class->pre_schedule(rq, prev);
3889 #endif
3891 if (unlikely(!rq->nr_running))
3892 idle_balance(cpu, rq);
3894 prev->sched_class->put_prev_task(rq, prev);
3895 next = pick_next_task(rq, prev);
3897 sched_info_switch(prev, next);
3899 if (likely(prev != next)) {
3900 rq->nr_switches++;
3901 rq->curr = next;
3902 ++*switch_count;
3904 context_switch(rq, prev, next); /* unlocks the rq */
3906 * the context switch might have flipped the stack from under
3907 * us, hence refresh the local variables.
3909 cpu = smp_processor_id();
3910 rq = cpu_rq(cpu);
3911 } else
3912 spin_unlock_irq(&rq->lock);
3914 hrtick_set(rq);
3916 if (unlikely(reacquire_kernel_lock(current) < 0))
3917 goto need_resched_nonpreemptible;
3919 preempt_enable_no_resched();
3920 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3921 goto need_resched;
3923 EXPORT_SYMBOL(schedule);
3925 #ifdef CONFIG_PREEMPT
3927 * this is the entry point to schedule() from in-kernel preemption
3928 * off of preempt_enable. Kernel preemptions off return from interrupt
3929 * occur there and call schedule directly.
3931 asmlinkage void __sched preempt_schedule(void)
3933 struct thread_info *ti = current_thread_info();
3934 struct task_struct *task = current;
3935 int saved_lock_depth;
3938 * If there is a non-zero preempt_count or interrupts are disabled,
3939 * we do not want to preempt the current task. Just return..
3941 if (likely(ti->preempt_count || irqs_disabled()))
3942 return;
3944 do {
3945 add_preempt_count(PREEMPT_ACTIVE);
3948 * We keep the big kernel semaphore locked, but we
3949 * clear ->lock_depth so that schedule() doesnt
3950 * auto-release the semaphore:
3952 saved_lock_depth = task->lock_depth;
3953 task->lock_depth = -1;
3954 schedule();
3955 task->lock_depth = saved_lock_depth;
3956 sub_preempt_count(PREEMPT_ACTIVE);
3959 * Check again in case we missed a preemption opportunity
3960 * between schedule and now.
3962 barrier();
3963 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3965 EXPORT_SYMBOL(preempt_schedule);
3968 * this is the entry point to schedule() from kernel preemption
3969 * off of irq context.
3970 * Note, that this is called and return with irqs disabled. This will
3971 * protect us against recursive calling from irq.
3973 asmlinkage void __sched preempt_schedule_irq(void)
3975 struct thread_info *ti = current_thread_info();
3976 struct task_struct *task = current;
3977 int saved_lock_depth;
3979 /* Catch callers which need to be fixed */
3980 BUG_ON(ti->preempt_count || !irqs_disabled());
3982 do {
3983 add_preempt_count(PREEMPT_ACTIVE);
3986 * We keep the big kernel semaphore locked, but we
3987 * clear ->lock_depth so that schedule() doesnt
3988 * auto-release the semaphore:
3990 saved_lock_depth = task->lock_depth;
3991 task->lock_depth = -1;
3992 local_irq_enable();
3993 schedule();
3994 local_irq_disable();
3995 task->lock_depth = saved_lock_depth;
3996 sub_preempt_count(PREEMPT_ACTIVE);
3999 * Check again in case we missed a preemption opportunity
4000 * between schedule and now.
4002 barrier();
4003 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4006 #endif /* CONFIG_PREEMPT */
4008 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4009 void *key)
4011 return try_to_wake_up(curr->private, mode, sync);
4013 EXPORT_SYMBOL(default_wake_function);
4016 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4017 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4018 * number) then we wake all the non-exclusive tasks and one exclusive task.
4020 * There are circumstances in which we can try to wake a task which has already
4021 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4022 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4024 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4025 int nr_exclusive, int sync, void *key)
4027 wait_queue_t *curr, *next;
4029 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4030 unsigned flags = curr->flags;
4032 if (curr->func(curr, mode, sync, key) &&
4033 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4034 break;
4039 * __wake_up - wake up threads blocked on a waitqueue.
4040 * @q: the waitqueue
4041 * @mode: which threads
4042 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4043 * @key: is directly passed to the wakeup function
4045 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4046 int nr_exclusive, void *key)
4048 unsigned long flags;
4050 spin_lock_irqsave(&q->lock, flags);
4051 __wake_up_common(q, mode, nr_exclusive, 0, key);
4052 spin_unlock_irqrestore(&q->lock, flags);
4054 EXPORT_SYMBOL(__wake_up);
4057 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4059 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4061 __wake_up_common(q, mode, 1, 0, NULL);
4065 * __wake_up_sync - wake up threads blocked on a waitqueue.
4066 * @q: the waitqueue
4067 * @mode: which threads
4068 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4070 * The sync wakeup differs that the waker knows that it will schedule
4071 * away soon, so while the target thread will be woken up, it will not
4072 * be migrated to another CPU - ie. the two threads are 'synchronized'
4073 * with each other. This can prevent needless bouncing between CPUs.
4075 * On UP it can prevent extra preemption.
4077 void
4078 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4080 unsigned long flags;
4081 int sync = 1;
4083 if (unlikely(!q))
4084 return;
4086 if (unlikely(!nr_exclusive))
4087 sync = 0;
4089 spin_lock_irqsave(&q->lock, flags);
4090 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4091 spin_unlock_irqrestore(&q->lock, flags);
4093 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4095 void complete(struct completion *x)
4097 unsigned long flags;
4099 spin_lock_irqsave(&x->wait.lock, flags);
4100 x->done++;
4101 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4102 spin_unlock_irqrestore(&x->wait.lock, flags);
4104 EXPORT_SYMBOL(complete);
4106 void complete_all(struct completion *x)
4108 unsigned long flags;
4110 spin_lock_irqsave(&x->wait.lock, flags);
4111 x->done += UINT_MAX/2;
4112 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4113 spin_unlock_irqrestore(&x->wait.lock, flags);
4115 EXPORT_SYMBOL(complete_all);
4117 static inline long __sched
4118 do_wait_for_common(struct completion *x, long timeout, int state)
4120 if (!x->done) {
4121 DECLARE_WAITQUEUE(wait, current);
4123 wait.flags |= WQ_FLAG_EXCLUSIVE;
4124 __add_wait_queue_tail(&x->wait, &wait);
4125 do {
4126 if ((state == TASK_INTERRUPTIBLE &&
4127 signal_pending(current)) ||
4128 (state == TASK_KILLABLE &&
4129 fatal_signal_pending(current))) {
4130 __remove_wait_queue(&x->wait, &wait);
4131 return -ERESTARTSYS;
4133 __set_current_state(state);
4134 spin_unlock_irq(&x->wait.lock);
4135 timeout = schedule_timeout(timeout);
4136 spin_lock_irq(&x->wait.lock);
4137 if (!timeout) {
4138 __remove_wait_queue(&x->wait, &wait);
4139 return timeout;
4141 } while (!x->done);
4142 __remove_wait_queue(&x->wait, &wait);
4144 x->done--;
4145 return timeout;
4148 static long __sched
4149 wait_for_common(struct completion *x, long timeout, int state)
4151 might_sleep();
4153 spin_lock_irq(&x->wait.lock);
4154 timeout = do_wait_for_common(x, timeout, state);
4155 spin_unlock_irq(&x->wait.lock);
4156 return timeout;
4159 void __sched wait_for_completion(struct completion *x)
4161 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4163 EXPORT_SYMBOL(wait_for_completion);
4165 unsigned long __sched
4166 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4168 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4170 EXPORT_SYMBOL(wait_for_completion_timeout);
4172 int __sched wait_for_completion_interruptible(struct completion *x)
4174 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4175 if (t == -ERESTARTSYS)
4176 return t;
4177 return 0;
4179 EXPORT_SYMBOL(wait_for_completion_interruptible);
4181 unsigned long __sched
4182 wait_for_completion_interruptible_timeout(struct completion *x,
4183 unsigned long timeout)
4185 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4187 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4189 int __sched wait_for_completion_killable(struct completion *x)
4191 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4192 if (t == -ERESTARTSYS)
4193 return t;
4194 return 0;
4196 EXPORT_SYMBOL(wait_for_completion_killable);
4198 static long __sched
4199 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4201 unsigned long flags;
4202 wait_queue_t wait;
4204 init_waitqueue_entry(&wait, current);
4206 __set_current_state(state);
4208 spin_lock_irqsave(&q->lock, flags);
4209 __add_wait_queue(q, &wait);
4210 spin_unlock(&q->lock);
4211 timeout = schedule_timeout(timeout);
4212 spin_lock_irq(&q->lock);
4213 __remove_wait_queue(q, &wait);
4214 spin_unlock_irqrestore(&q->lock, flags);
4216 return timeout;
4219 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4221 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4223 EXPORT_SYMBOL(interruptible_sleep_on);
4225 long __sched
4226 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4228 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4230 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4232 void __sched sleep_on(wait_queue_head_t *q)
4234 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4236 EXPORT_SYMBOL(sleep_on);
4238 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4240 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4242 EXPORT_SYMBOL(sleep_on_timeout);
4244 #ifdef CONFIG_RT_MUTEXES
4247 * rt_mutex_setprio - set the current priority of a task
4248 * @p: task
4249 * @prio: prio value (kernel-internal form)
4251 * This function changes the 'effective' priority of a task. It does
4252 * not touch ->normal_prio like __setscheduler().
4254 * Used by the rt_mutex code to implement priority inheritance logic.
4256 void rt_mutex_setprio(struct task_struct *p, int prio)
4258 unsigned long flags;
4259 int oldprio, on_rq, running;
4260 struct rq *rq;
4261 const struct sched_class *prev_class = p->sched_class;
4263 BUG_ON(prio < 0 || prio > MAX_PRIO);
4265 rq = task_rq_lock(p, &flags);
4266 update_rq_clock(rq);
4268 oldprio = p->prio;
4269 on_rq = p->se.on_rq;
4270 running = task_current(rq, p);
4271 if (on_rq) {
4272 dequeue_task(rq, p, 0);
4273 if (running)
4274 p->sched_class->put_prev_task(rq, p);
4277 if (rt_prio(prio))
4278 p->sched_class = &rt_sched_class;
4279 else
4280 p->sched_class = &fair_sched_class;
4282 p->prio = prio;
4284 if (on_rq) {
4285 if (running)
4286 p->sched_class->set_curr_task(rq);
4288 enqueue_task(rq, p, 0);
4290 check_class_changed(rq, p, prev_class, oldprio, running);
4292 task_rq_unlock(rq, &flags);
4295 #endif
4297 void set_user_nice(struct task_struct *p, long nice)
4299 int old_prio, delta, on_rq;
4300 unsigned long flags;
4301 struct rq *rq;
4303 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4304 return;
4306 * We have to be careful, if called from sys_setpriority(),
4307 * the task might be in the middle of scheduling on another CPU.
4309 rq = task_rq_lock(p, &flags);
4310 update_rq_clock(rq);
4312 * The RT priorities are set via sched_setscheduler(), but we still
4313 * allow the 'normal' nice value to be set - but as expected
4314 * it wont have any effect on scheduling until the task is
4315 * SCHED_FIFO/SCHED_RR:
4317 if (task_has_rt_policy(p)) {
4318 p->static_prio = NICE_TO_PRIO(nice);
4319 goto out_unlock;
4321 on_rq = p->se.on_rq;
4322 if (on_rq) {
4323 dequeue_task(rq, p, 0);
4324 dec_load(rq, p);
4327 p->static_prio = NICE_TO_PRIO(nice);
4328 set_load_weight(p);
4329 old_prio = p->prio;
4330 p->prio = effective_prio(p);
4331 delta = p->prio - old_prio;
4333 if (on_rq) {
4334 enqueue_task(rq, p, 0);
4335 inc_load(rq, p);
4337 * If the task increased its priority or is running and
4338 * lowered its priority, then reschedule its CPU:
4340 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4341 resched_task(rq->curr);
4343 out_unlock:
4344 task_rq_unlock(rq, &flags);
4346 EXPORT_SYMBOL(set_user_nice);
4349 * can_nice - check if a task can reduce its nice value
4350 * @p: task
4351 * @nice: nice value
4353 int can_nice(const struct task_struct *p, const int nice)
4355 /* convert nice value [19,-20] to rlimit style value [1,40] */
4356 int nice_rlim = 20 - nice;
4358 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4359 capable(CAP_SYS_NICE));
4362 #ifdef __ARCH_WANT_SYS_NICE
4365 * sys_nice - change the priority of the current process.
4366 * @increment: priority increment
4368 * sys_setpriority is a more generic, but much slower function that
4369 * does similar things.
4371 asmlinkage long sys_nice(int increment)
4373 long nice, retval;
4376 * Setpriority might change our priority at the same moment.
4377 * We don't have to worry. Conceptually one call occurs first
4378 * and we have a single winner.
4380 if (increment < -40)
4381 increment = -40;
4382 if (increment > 40)
4383 increment = 40;
4385 nice = PRIO_TO_NICE(current->static_prio) + increment;
4386 if (nice < -20)
4387 nice = -20;
4388 if (nice > 19)
4389 nice = 19;
4391 if (increment < 0 && !can_nice(current, nice))
4392 return -EPERM;
4394 retval = security_task_setnice(current, nice);
4395 if (retval)
4396 return retval;
4398 set_user_nice(current, nice);
4399 return 0;
4402 #endif
4405 * task_prio - return the priority value of a given task.
4406 * @p: the task in question.
4408 * This is the priority value as seen by users in /proc.
4409 * RT tasks are offset by -200. Normal tasks are centered
4410 * around 0, value goes from -16 to +15.
4412 int task_prio(const struct task_struct *p)
4414 return p->prio - MAX_RT_PRIO;
4418 * task_nice - return the nice value of a given task.
4419 * @p: the task in question.
4421 int task_nice(const struct task_struct *p)
4423 return TASK_NICE(p);
4425 EXPORT_SYMBOL_GPL(task_nice);
4428 * idle_cpu - is a given cpu idle currently?
4429 * @cpu: the processor in question.
4431 int idle_cpu(int cpu)
4433 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4437 * idle_task - return the idle task for a given cpu.
4438 * @cpu: the processor in question.
4440 struct task_struct *idle_task(int cpu)
4442 return cpu_rq(cpu)->idle;
4446 * find_process_by_pid - find a process with a matching PID value.
4447 * @pid: the pid in question.
4449 static struct task_struct *find_process_by_pid(pid_t pid)
4451 return pid ? find_task_by_vpid(pid) : current;
4454 /* Actually do priority change: must hold rq lock. */
4455 static void
4456 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4458 BUG_ON(p->se.on_rq);
4460 p->policy = policy;
4461 switch (p->policy) {
4462 case SCHED_NORMAL:
4463 case SCHED_BATCH:
4464 case SCHED_IDLE:
4465 p->sched_class = &fair_sched_class;
4466 break;
4467 case SCHED_FIFO:
4468 case SCHED_RR:
4469 p->sched_class = &rt_sched_class;
4470 break;
4473 p->rt_priority = prio;
4474 p->normal_prio = normal_prio(p);
4475 /* we are holding p->pi_lock already */
4476 p->prio = rt_mutex_getprio(p);
4477 set_load_weight(p);
4481 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4482 * @p: the task in question.
4483 * @policy: new policy.
4484 * @param: structure containing the new RT priority.
4486 * NOTE that the task may be already dead.
4488 int sched_setscheduler(struct task_struct *p, int policy,
4489 struct sched_param *param)
4491 int retval, oldprio, oldpolicy = -1, on_rq, running;
4492 unsigned long flags;
4493 const struct sched_class *prev_class = p->sched_class;
4494 struct rq *rq;
4496 /* may grab non-irq protected spin_locks */
4497 BUG_ON(in_interrupt());
4498 recheck:
4499 /* double check policy once rq lock held */
4500 if (policy < 0)
4501 policy = oldpolicy = p->policy;
4502 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4503 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4504 policy != SCHED_IDLE)
4505 return -EINVAL;
4507 * Valid priorities for SCHED_FIFO and SCHED_RR are
4508 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4509 * SCHED_BATCH and SCHED_IDLE is 0.
4511 if (param->sched_priority < 0 ||
4512 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4513 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4514 return -EINVAL;
4515 if (rt_policy(policy) != (param->sched_priority != 0))
4516 return -EINVAL;
4519 * Allow unprivileged RT tasks to decrease priority:
4521 if (!capable(CAP_SYS_NICE)) {
4522 if (rt_policy(policy)) {
4523 unsigned long rlim_rtprio;
4525 if (!lock_task_sighand(p, &flags))
4526 return -ESRCH;
4527 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4528 unlock_task_sighand(p, &flags);
4530 /* can't set/change the rt policy */
4531 if (policy != p->policy && !rlim_rtprio)
4532 return -EPERM;
4534 /* can't increase priority */
4535 if (param->sched_priority > p->rt_priority &&
4536 param->sched_priority > rlim_rtprio)
4537 return -EPERM;
4540 * Like positive nice levels, dont allow tasks to
4541 * move out of SCHED_IDLE either:
4543 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4544 return -EPERM;
4546 /* can't change other user's priorities */
4547 if ((current->euid != p->euid) &&
4548 (current->euid != p->uid))
4549 return -EPERM;
4552 #ifdef CONFIG_RT_GROUP_SCHED
4554 * Do not allow realtime tasks into groups that have no runtime
4555 * assigned.
4557 if (rt_policy(policy) && task_group(p)->rt_runtime == 0)
4558 return -EPERM;
4559 #endif
4561 retval = security_task_setscheduler(p, policy, param);
4562 if (retval)
4563 return retval;
4565 * make sure no PI-waiters arrive (or leave) while we are
4566 * changing the priority of the task:
4568 spin_lock_irqsave(&p->pi_lock, flags);
4570 * To be able to change p->policy safely, the apropriate
4571 * runqueue lock must be held.
4573 rq = __task_rq_lock(p);
4574 /* recheck policy now with rq lock held */
4575 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4576 policy = oldpolicy = -1;
4577 __task_rq_unlock(rq);
4578 spin_unlock_irqrestore(&p->pi_lock, flags);
4579 goto recheck;
4581 update_rq_clock(rq);
4582 on_rq = p->se.on_rq;
4583 running = task_current(rq, p);
4584 if (on_rq) {
4585 deactivate_task(rq, p, 0);
4586 if (running)
4587 p->sched_class->put_prev_task(rq, p);
4590 oldprio = p->prio;
4591 __setscheduler(rq, p, policy, param->sched_priority);
4593 if (on_rq) {
4594 if (running)
4595 p->sched_class->set_curr_task(rq);
4597 activate_task(rq, p, 0);
4599 check_class_changed(rq, p, prev_class, oldprio, running);
4601 __task_rq_unlock(rq);
4602 spin_unlock_irqrestore(&p->pi_lock, flags);
4604 rt_mutex_adjust_pi(p);
4606 return 0;
4608 EXPORT_SYMBOL_GPL(sched_setscheduler);
4610 static int
4611 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4613 struct sched_param lparam;
4614 struct task_struct *p;
4615 int retval;
4617 if (!param || pid < 0)
4618 return -EINVAL;
4619 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4620 return -EFAULT;
4622 rcu_read_lock();
4623 retval = -ESRCH;
4624 p = find_process_by_pid(pid);
4625 if (p != NULL)
4626 retval = sched_setscheduler(p, policy, &lparam);
4627 rcu_read_unlock();
4629 return retval;
4633 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4634 * @pid: the pid in question.
4635 * @policy: new policy.
4636 * @param: structure containing the new RT priority.
4638 asmlinkage long
4639 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4641 /* negative values for policy are not valid */
4642 if (policy < 0)
4643 return -EINVAL;
4645 return do_sched_setscheduler(pid, policy, param);
4649 * sys_sched_setparam - set/change the RT priority of a thread
4650 * @pid: the pid in question.
4651 * @param: structure containing the new RT priority.
4653 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4655 return do_sched_setscheduler(pid, -1, param);
4659 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4660 * @pid: the pid in question.
4662 asmlinkage long sys_sched_getscheduler(pid_t pid)
4664 struct task_struct *p;
4665 int retval;
4667 if (pid < 0)
4668 return -EINVAL;
4670 retval = -ESRCH;
4671 read_lock(&tasklist_lock);
4672 p = find_process_by_pid(pid);
4673 if (p) {
4674 retval = security_task_getscheduler(p);
4675 if (!retval)
4676 retval = p->policy;
4678 read_unlock(&tasklist_lock);
4679 return retval;
4683 * sys_sched_getscheduler - get the RT priority of a thread
4684 * @pid: the pid in question.
4685 * @param: structure containing the RT priority.
4687 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4689 struct sched_param lp;
4690 struct task_struct *p;
4691 int retval;
4693 if (!param || pid < 0)
4694 return -EINVAL;
4696 read_lock(&tasklist_lock);
4697 p = find_process_by_pid(pid);
4698 retval = -ESRCH;
4699 if (!p)
4700 goto out_unlock;
4702 retval = security_task_getscheduler(p);
4703 if (retval)
4704 goto out_unlock;
4706 lp.sched_priority = p->rt_priority;
4707 read_unlock(&tasklist_lock);
4710 * This one might sleep, we cannot do it with a spinlock held ...
4712 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4714 return retval;
4716 out_unlock:
4717 read_unlock(&tasklist_lock);
4718 return retval;
4721 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4723 cpumask_t cpus_allowed;
4724 struct task_struct *p;
4725 int retval;
4727 get_online_cpus();
4728 read_lock(&tasklist_lock);
4730 p = find_process_by_pid(pid);
4731 if (!p) {
4732 read_unlock(&tasklist_lock);
4733 put_online_cpus();
4734 return -ESRCH;
4738 * It is not safe to call set_cpus_allowed with the
4739 * tasklist_lock held. We will bump the task_struct's
4740 * usage count and then drop tasklist_lock.
4742 get_task_struct(p);
4743 read_unlock(&tasklist_lock);
4745 retval = -EPERM;
4746 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4747 !capable(CAP_SYS_NICE))
4748 goto out_unlock;
4750 retval = security_task_setscheduler(p, 0, NULL);
4751 if (retval)
4752 goto out_unlock;
4754 cpus_allowed = cpuset_cpus_allowed(p);
4755 cpus_and(new_mask, new_mask, cpus_allowed);
4756 again:
4757 retval = set_cpus_allowed(p, new_mask);
4759 if (!retval) {
4760 cpus_allowed = cpuset_cpus_allowed(p);
4761 if (!cpus_subset(new_mask, cpus_allowed)) {
4763 * We must have raced with a concurrent cpuset
4764 * update. Just reset the cpus_allowed to the
4765 * cpuset's cpus_allowed
4767 new_mask = cpus_allowed;
4768 goto again;
4771 out_unlock:
4772 put_task_struct(p);
4773 put_online_cpus();
4774 return retval;
4777 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4778 cpumask_t *new_mask)
4780 if (len < sizeof(cpumask_t)) {
4781 memset(new_mask, 0, sizeof(cpumask_t));
4782 } else if (len > sizeof(cpumask_t)) {
4783 len = sizeof(cpumask_t);
4785 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4789 * sys_sched_setaffinity - set the cpu affinity of a process
4790 * @pid: pid of the process
4791 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4792 * @user_mask_ptr: user-space pointer to the new cpu mask
4794 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4795 unsigned long __user *user_mask_ptr)
4797 cpumask_t new_mask;
4798 int retval;
4800 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4801 if (retval)
4802 return retval;
4804 return sched_setaffinity(pid, new_mask);
4808 * Represents all cpu's present in the system
4809 * In systems capable of hotplug, this map could dynamically grow
4810 * as new cpu's are detected in the system via any platform specific
4811 * method, such as ACPI for e.g.
4814 cpumask_t cpu_present_map __read_mostly;
4815 EXPORT_SYMBOL(cpu_present_map);
4817 #ifndef CONFIG_SMP
4818 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4819 EXPORT_SYMBOL(cpu_online_map);
4821 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4822 EXPORT_SYMBOL(cpu_possible_map);
4823 #endif
4825 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4827 struct task_struct *p;
4828 int retval;
4830 get_online_cpus();
4831 read_lock(&tasklist_lock);
4833 retval = -ESRCH;
4834 p = find_process_by_pid(pid);
4835 if (!p)
4836 goto out_unlock;
4838 retval = security_task_getscheduler(p);
4839 if (retval)
4840 goto out_unlock;
4842 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4844 out_unlock:
4845 read_unlock(&tasklist_lock);
4846 put_online_cpus();
4848 return retval;
4852 * sys_sched_getaffinity - get the cpu affinity of a process
4853 * @pid: pid of the process
4854 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4855 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4857 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4858 unsigned long __user *user_mask_ptr)
4860 int ret;
4861 cpumask_t mask;
4863 if (len < sizeof(cpumask_t))
4864 return -EINVAL;
4866 ret = sched_getaffinity(pid, &mask);
4867 if (ret < 0)
4868 return ret;
4870 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4871 return -EFAULT;
4873 return sizeof(cpumask_t);
4877 * sys_sched_yield - yield the current processor to other threads.
4879 * This function yields the current CPU to other tasks. If there are no
4880 * other threads running on this CPU then this function will return.
4882 asmlinkage long sys_sched_yield(void)
4884 struct rq *rq = this_rq_lock();
4886 schedstat_inc(rq, yld_count);
4887 current->sched_class->yield_task(rq);
4890 * Since we are going to call schedule() anyway, there's
4891 * no need to preempt or enable interrupts:
4893 __release(rq->lock);
4894 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4895 _raw_spin_unlock(&rq->lock);
4896 preempt_enable_no_resched();
4898 schedule();
4900 return 0;
4903 static void __cond_resched(void)
4905 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4906 __might_sleep(__FILE__, __LINE__);
4907 #endif
4909 * The BKS might be reacquired before we have dropped
4910 * PREEMPT_ACTIVE, which could trigger a second
4911 * cond_resched() call.
4913 do {
4914 add_preempt_count(PREEMPT_ACTIVE);
4915 schedule();
4916 sub_preempt_count(PREEMPT_ACTIVE);
4917 } while (need_resched());
4920 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4921 int __sched _cond_resched(void)
4923 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4924 system_state == SYSTEM_RUNNING) {
4925 __cond_resched();
4926 return 1;
4928 return 0;
4930 EXPORT_SYMBOL(_cond_resched);
4931 #endif
4934 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4935 * call schedule, and on return reacquire the lock.
4937 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4938 * operations here to prevent schedule() from being called twice (once via
4939 * spin_unlock(), once by hand).
4941 int cond_resched_lock(spinlock_t *lock)
4943 int resched = need_resched() && system_state == SYSTEM_RUNNING;
4944 int ret = 0;
4946 if (spin_needbreak(lock) || resched) {
4947 spin_unlock(lock);
4948 if (resched && need_resched())
4949 __cond_resched();
4950 else
4951 cpu_relax();
4952 ret = 1;
4953 spin_lock(lock);
4955 return ret;
4957 EXPORT_SYMBOL(cond_resched_lock);
4959 int __sched cond_resched_softirq(void)
4961 BUG_ON(!in_softirq());
4963 if (need_resched() && system_state == SYSTEM_RUNNING) {
4964 local_bh_enable();
4965 __cond_resched();
4966 local_bh_disable();
4967 return 1;
4969 return 0;
4971 EXPORT_SYMBOL(cond_resched_softirq);
4974 * yield - yield the current processor to other threads.
4976 * This is a shortcut for kernel-space yielding - it marks the
4977 * thread runnable and calls sys_sched_yield().
4979 void __sched yield(void)
4981 set_current_state(TASK_RUNNING);
4982 sys_sched_yield();
4984 EXPORT_SYMBOL(yield);
4987 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4988 * that process accounting knows that this is a task in IO wait state.
4990 * But don't do that if it is a deliberate, throttling IO wait (this task
4991 * has set its backing_dev_info: the queue against which it should throttle)
4993 void __sched io_schedule(void)
4995 struct rq *rq = &__raw_get_cpu_var(runqueues);
4997 delayacct_blkio_start();
4998 atomic_inc(&rq->nr_iowait);
4999 schedule();
5000 atomic_dec(&rq->nr_iowait);
5001 delayacct_blkio_end();
5003 EXPORT_SYMBOL(io_schedule);
5005 long __sched io_schedule_timeout(long timeout)
5007 struct rq *rq = &__raw_get_cpu_var(runqueues);
5008 long ret;
5010 delayacct_blkio_start();
5011 atomic_inc(&rq->nr_iowait);
5012 ret = schedule_timeout(timeout);
5013 atomic_dec(&rq->nr_iowait);
5014 delayacct_blkio_end();
5015 return ret;
5019 * sys_sched_get_priority_max - return maximum RT priority.
5020 * @policy: scheduling class.
5022 * this syscall returns the maximum rt_priority that can be used
5023 * by a given scheduling class.
5025 asmlinkage long sys_sched_get_priority_max(int policy)
5027 int ret = -EINVAL;
5029 switch (policy) {
5030 case SCHED_FIFO:
5031 case SCHED_RR:
5032 ret = MAX_USER_RT_PRIO-1;
5033 break;
5034 case SCHED_NORMAL:
5035 case SCHED_BATCH:
5036 case SCHED_IDLE:
5037 ret = 0;
5038 break;
5040 return ret;
5044 * sys_sched_get_priority_min - return minimum RT priority.
5045 * @policy: scheduling class.
5047 * this syscall returns the minimum rt_priority that can be used
5048 * by a given scheduling class.
5050 asmlinkage long sys_sched_get_priority_min(int policy)
5052 int ret = -EINVAL;
5054 switch (policy) {
5055 case SCHED_FIFO:
5056 case SCHED_RR:
5057 ret = 1;
5058 break;
5059 case SCHED_NORMAL:
5060 case SCHED_BATCH:
5061 case SCHED_IDLE:
5062 ret = 0;
5064 return ret;
5068 * sys_sched_rr_get_interval - return the default timeslice of a process.
5069 * @pid: pid of the process.
5070 * @interval: userspace pointer to the timeslice value.
5072 * this syscall writes the default timeslice value of a given process
5073 * into the user-space timespec buffer. A value of '0' means infinity.
5075 asmlinkage
5076 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5078 struct task_struct *p;
5079 unsigned int time_slice;
5080 int retval;
5081 struct timespec t;
5083 if (pid < 0)
5084 return -EINVAL;
5086 retval = -ESRCH;
5087 read_lock(&tasklist_lock);
5088 p = find_process_by_pid(pid);
5089 if (!p)
5090 goto out_unlock;
5092 retval = security_task_getscheduler(p);
5093 if (retval)
5094 goto out_unlock;
5097 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5098 * tasks that are on an otherwise idle runqueue:
5100 time_slice = 0;
5101 if (p->policy == SCHED_RR) {
5102 time_slice = DEF_TIMESLICE;
5103 } else {
5104 struct sched_entity *se = &p->se;
5105 unsigned long flags;
5106 struct rq *rq;
5108 rq = task_rq_lock(p, &flags);
5109 if (rq->cfs.load.weight)
5110 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5111 task_rq_unlock(rq, &flags);
5113 read_unlock(&tasklist_lock);
5114 jiffies_to_timespec(time_slice, &t);
5115 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5116 return retval;
5118 out_unlock:
5119 read_unlock(&tasklist_lock);
5120 return retval;
5123 static const char stat_nam[] = "RSDTtZX";
5125 void sched_show_task(struct task_struct *p)
5127 unsigned long free = 0;
5128 unsigned state;
5130 state = p->state ? __ffs(p->state) + 1 : 0;
5131 printk(KERN_INFO "%-13.13s %c", p->comm,
5132 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5133 #if BITS_PER_LONG == 32
5134 if (state == TASK_RUNNING)
5135 printk(KERN_CONT " running ");
5136 else
5137 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5138 #else
5139 if (state == TASK_RUNNING)
5140 printk(KERN_CONT " running task ");
5141 else
5142 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5143 #endif
5144 #ifdef CONFIG_DEBUG_STACK_USAGE
5146 unsigned long *n = end_of_stack(p);
5147 while (!*n)
5148 n++;
5149 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5151 #endif
5152 printk(KERN_CONT "%5lu %5d %6d\n", free,
5153 task_pid_nr(p), task_pid_nr(p->real_parent));
5155 show_stack(p, NULL);
5158 void show_state_filter(unsigned long state_filter)
5160 struct task_struct *g, *p;
5162 #if BITS_PER_LONG == 32
5163 printk(KERN_INFO
5164 " task PC stack pid father\n");
5165 #else
5166 printk(KERN_INFO
5167 " task PC stack pid father\n");
5168 #endif
5169 read_lock(&tasklist_lock);
5170 do_each_thread(g, p) {
5172 * reset the NMI-timeout, listing all files on a slow
5173 * console might take alot of time:
5175 touch_nmi_watchdog();
5176 if (!state_filter || (p->state & state_filter))
5177 sched_show_task(p);
5178 } while_each_thread(g, p);
5180 touch_all_softlockup_watchdogs();
5182 #ifdef CONFIG_SCHED_DEBUG
5183 sysrq_sched_debug_show();
5184 #endif
5185 read_unlock(&tasklist_lock);
5187 * Only show locks if all tasks are dumped:
5189 if (state_filter == -1)
5190 debug_show_all_locks();
5193 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5195 idle->sched_class = &idle_sched_class;
5199 * init_idle - set up an idle thread for a given CPU
5200 * @idle: task in question
5201 * @cpu: cpu the idle task belongs to
5203 * NOTE: this function does not set the idle thread's NEED_RESCHED
5204 * flag, to make booting more robust.
5206 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5208 struct rq *rq = cpu_rq(cpu);
5209 unsigned long flags;
5211 __sched_fork(idle);
5212 idle->se.exec_start = sched_clock();
5214 idle->prio = idle->normal_prio = MAX_PRIO;
5215 idle->cpus_allowed = cpumask_of_cpu(cpu);
5216 __set_task_cpu(idle, cpu);
5218 spin_lock_irqsave(&rq->lock, flags);
5219 rq->curr = rq->idle = idle;
5220 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5221 idle->oncpu = 1;
5222 #endif
5223 spin_unlock_irqrestore(&rq->lock, flags);
5225 /* Set the preempt count _outside_ the spinlocks! */
5226 task_thread_info(idle)->preempt_count = 0;
5229 * The idle tasks have their own, simple scheduling class:
5231 idle->sched_class = &idle_sched_class;
5235 * In a system that switches off the HZ timer nohz_cpu_mask
5236 * indicates which cpus entered this state. This is used
5237 * in the rcu update to wait only for active cpus. For system
5238 * which do not switch off the HZ timer nohz_cpu_mask should
5239 * always be CPU_MASK_NONE.
5241 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5244 * Increase the granularity value when there are more CPUs,
5245 * because with more CPUs the 'effective latency' as visible
5246 * to users decreases. But the relationship is not linear,
5247 * so pick a second-best guess by going with the log2 of the
5248 * number of CPUs.
5250 * This idea comes from the SD scheduler of Con Kolivas:
5252 static inline void sched_init_granularity(void)
5254 unsigned int factor = 1 + ilog2(num_online_cpus());
5255 const unsigned long limit = 200000000;
5257 sysctl_sched_min_granularity *= factor;
5258 if (sysctl_sched_min_granularity > limit)
5259 sysctl_sched_min_granularity = limit;
5261 sysctl_sched_latency *= factor;
5262 if (sysctl_sched_latency > limit)
5263 sysctl_sched_latency = limit;
5265 sysctl_sched_wakeup_granularity *= factor;
5266 sysctl_sched_batch_wakeup_granularity *= factor;
5269 #ifdef CONFIG_SMP
5271 * This is how migration works:
5273 * 1) we queue a struct migration_req structure in the source CPU's
5274 * runqueue and wake up that CPU's migration thread.
5275 * 2) we down() the locked semaphore => thread blocks.
5276 * 3) migration thread wakes up (implicitly it forces the migrated
5277 * thread off the CPU)
5278 * 4) it gets the migration request and checks whether the migrated
5279 * task is still in the wrong runqueue.
5280 * 5) if it's in the wrong runqueue then the migration thread removes
5281 * it and puts it into the right queue.
5282 * 6) migration thread up()s the semaphore.
5283 * 7) we wake up and the migration is done.
5287 * Change a given task's CPU affinity. Migrate the thread to a
5288 * proper CPU and schedule it away if the CPU it's executing on
5289 * is removed from the allowed bitmask.
5291 * NOTE: the caller must have a valid reference to the task, the
5292 * task must not exit() & deallocate itself prematurely. The
5293 * call is not atomic; no spinlocks may be held.
5295 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5297 struct migration_req req;
5298 unsigned long flags;
5299 struct rq *rq;
5300 int ret = 0;
5302 rq = task_rq_lock(p, &flags);
5303 if (!cpus_intersects(new_mask, cpu_online_map)) {
5304 ret = -EINVAL;
5305 goto out;
5308 if (p->sched_class->set_cpus_allowed)
5309 p->sched_class->set_cpus_allowed(p, &new_mask);
5310 else {
5311 p->cpus_allowed = new_mask;
5312 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5315 /* Can the task run on the task's current CPU? If so, we're done */
5316 if (cpu_isset(task_cpu(p), new_mask))
5317 goto out;
5319 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5320 /* Need help from migration thread: drop lock and wait. */
5321 task_rq_unlock(rq, &flags);
5322 wake_up_process(rq->migration_thread);
5323 wait_for_completion(&req.done);
5324 tlb_migrate_finish(p->mm);
5325 return 0;
5327 out:
5328 task_rq_unlock(rq, &flags);
5330 return ret;
5332 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5335 * Move (not current) task off this cpu, onto dest cpu. We're doing
5336 * this because either it can't run here any more (set_cpus_allowed()
5337 * away from this CPU, or CPU going down), or because we're
5338 * attempting to rebalance this task on exec (sched_exec).
5340 * So we race with normal scheduler movements, but that's OK, as long
5341 * as the task is no longer on this CPU.
5343 * Returns non-zero if task was successfully migrated.
5345 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5347 struct rq *rq_dest, *rq_src;
5348 int ret = 0, on_rq;
5350 if (unlikely(cpu_is_offline(dest_cpu)))
5351 return ret;
5353 rq_src = cpu_rq(src_cpu);
5354 rq_dest = cpu_rq(dest_cpu);
5356 double_rq_lock(rq_src, rq_dest);
5357 /* Already moved. */
5358 if (task_cpu(p) != src_cpu)
5359 goto out;
5360 /* Affinity changed (again). */
5361 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5362 goto out;
5364 on_rq = p->se.on_rq;
5365 if (on_rq)
5366 deactivate_task(rq_src, p, 0);
5368 set_task_cpu(p, dest_cpu);
5369 if (on_rq) {
5370 activate_task(rq_dest, p, 0);
5371 check_preempt_curr(rq_dest, p);
5373 ret = 1;
5374 out:
5375 double_rq_unlock(rq_src, rq_dest);
5376 return ret;
5380 * migration_thread - this is a highprio system thread that performs
5381 * thread migration by bumping thread off CPU then 'pushing' onto
5382 * another runqueue.
5384 static int migration_thread(void *data)
5386 int cpu = (long)data;
5387 struct rq *rq;
5389 rq = cpu_rq(cpu);
5390 BUG_ON(rq->migration_thread != current);
5392 set_current_state(TASK_INTERRUPTIBLE);
5393 while (!kthread_should_stop()) {
5394 struct migration_req *req;
5395 struct list_head *head;
5397 spin_lock_irq(&rq->lock);
5399 if (cpu_is_offline(cpu)) {
5400 spin_unlock_irq(&rq->lock);
5401 goto wait_to_die;
5404 if (rq->active_balance) {
5405 active_load_balance(rq, cpu);
5406 rq->active_balance = 0;
5409 head = &rq->migration_queue;
5411 if (list_empty(head)) {
5412 spin_unlock_irq(&rq->lock);
5413 schedule();
5414 set_current_state(TASK_INTERRUPTIBLE);
5415 continue;
5417 req = list_entry(head->next, struct migration_req, list);
5418 list_del_init(head->next);
5420 spin_unlock(&rq->lock);
5421 __migrate_task(req->task, cpu, req->dest_cpu);
5422 local_irq_enable();
5424 complete(&req->done);
5426 __set_current_state(TASK_RUNNING);
5427 return 0;
5429 wait_to_die:
5430 /* Wait for kthread_stop */
5431 set_current_state(TASK_INTERRUPTIBLE);
5432 while (!kthread_should_stop()) {
5433 schedule();
5434 set_current_state(TASK_INTERRUPTIBLE);
5436 __set_current_state(TASK_RUNNING);
5437 return 0;
5440 #ifdef CONFIG_HOTPLUG_CPU
5442 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5444 int ret;
5446 local_irq_disable();
5447 ret = __migrate_task(p, src_cpu, dest_cpu);
5448 local_irq_enable();
5449 return ret;
5453 * Figure out where task on dead CPU should go, use force if necessary.
5454 * NOTE: interrupts should be disabled by the caller
5456 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5458 unsigned long flags;
5459 cpumask_t mask;
5460 struct rq *rq;
5461 int dest_cpu;
5463 do {
5464 /* On same node? */
5465 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5466 cpus_and(mask, mask, p->cpus_allowed);
5467 dest_cpu = any_online_cpu(mask);
5469 /* On any allowed CPU? */
5470 if (dest_cpu == NR_CPUS)
5471 dest_cpu = any_online_cpu(p->cpus_allowed);
5473 /* No more Mr. Nice Guy. */
5474 if (dest_cpu == NR_CPUS) {
5475 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5477 * Try to stay on the same cpuset, where the
5478 * current cpuset may be a subset of all cpus.
5479 * The cpuset_cpus_allowed_locked() variant of
5480 * cpuset_cpus_allowed() will not block. It must be
5481 * called within calls to cpuset_lock/cpuset_unlock.
5483 rq = task_rq_lock(p, &flags);
5484 p->cpus_allowed = cpus_allowed;
5485 dest_cpu = any_online_cpu(p->cpus_allowed);
5486 task_rq_unlock(rq, &flags);
5489 * Don't tell them about moving exiting tasks or
5490 * kernel threads (both mm NULL), since they never
5491 * leave kernel.
5493 if (p->mm && printk_ratelimit()) {
5494 printk(KERN_INFO "process %d (%s) no "
5495 "longer affine to cpu%d\n",
5496 task_pid_nr(p), p->comm, dead_cpu);
5499 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5503 * While a dead CPU has no uninterruptible tasks queued at this point,
5504 * it might still have a nonzero ->nr_uninterruptible counter, because
5505 * for performance reasons the counter is not stricly tracking tasks to
5506 * their home CPUs. So we just add the counter to another CPU's counter,
5507 * to keep the global sum constant after CPU-down:
5509 static void migrate_nr_uninterruptible(struct rq *rq_src)
5511 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5512 unsigned long flags;
5514 local_irq_save(flags);
5515 double_rq_lock(rq_src, rq_dest);
5516 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5517 rq_src->nr_uninterruptible = 0;
5518 double_rq_unlock(rq_src, rq_dest);
5519 local_irq_restore(flags);
5522 /* Run through task list and migrate tasks from the dead cpu. */
5523 static void migrate_live_tasks(int src_cpu)
5525 struct task_struct *p, *t;
5527 read_lock(&tasklist_lock);
5529 do_each_thread(t, p) {
5530 if (p == current)
5531 continue;
5533 if (task_cpu(p) == src_cpu)
5534 move_task_off_dead_cpu(src_cpu, p);
5535 } while_each_thread(t, p);
5537 read_unlock(&tasklist_lock);
5541 * Schedules idle task to be the next runnable task on current CPU.
5542 * It does so by boosting its priority to highest possible.
5543 * Used by CPU offline code.
5545 void sched_idle_next(void)
5547 int this_cpu = smp_processor_id();
5548 struct rq *rq = cpu_rq(this_cpu);
5549 struct task_struct *p = rq->idle;
5550 unsigned long flags;
5552 /* cpu has to be offline */
5553 BUG_ON(cpu_online(this_cpu));
5556 * Strictly not necessary since rest of the CPUs are stopped by now
5557 * and interrupts disabled on the current cpu.
5559 spin_lock_irqsave(&rq->lock, flags);
5561 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5563 update_rq_clock(rq);
5564 activate_task(rq, p, 0);
5566 spin_unlock_irqrestore(&rq->lock, flags);
5570 * Ensures that the idle task is using init_mm right before its cpu goes
5571 * offline.
5573 void idle_task_exit(void)
5575 struct mm_struct *mm = current->active_mm;
5577 BUG_ON(cpu_online(smp_processor_id()));
5579 if (mm != &init_mm)
5580 switch_mm(mm, &init_mm, current);
5581 mmdrop(mm);
5584 /* called under rq->lock with disabled interrupts */
5585 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5587 struct rq *rq = cpu_rq(dead_cpu);
5589 /* Must be exiting, otherwise would be on tasklist. */
5590 BUG_ON(!p->exit_state);
5592 /* Cannot have done final schedule yet: would have vanished. */
5593 BUG_ON(p->state == TASK_DEAD);
5595 get_task_struct(p);
5598 * Drop lock around migration; if someone else moves it,
5599 * that's OK. No task can be added to this CPU, so iteration is
5600 * fine.
5602 spin_unlock_irq(&rq->lock);
5603 move_task_off_dead_cpu(dead_cpu, p);
5604 spin_lock_irq(&rq->lock);
5606 put_task_struct(p);
5609 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5610 static void migrate_dead_tasks(unsigned int dead_cpu)
5612 struct rq *rq = cpu_rq(dead_cpu);
5613 struct task_struct *next;
5615 for ( ; ; ) {
5616 if (!rq->nr_running)
5617 break;
5618 update_rq_clock(rq);
5619 next = pick_next_task(rq, rq->curr);
5620 if (!next)
5621 break;
5622 migrate_dead(dead_cpu, next);
5626 #endif /* CONFIG_HOTPLUG_CPU */
5628 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5630 static struct ctl_table sd_ctl_dir[] = {
5632 .procname = "sched_domain",
5633 .mode = 0555,
5635 {0, },
5638 static struct ctl_table sd_ctl_root[] = {
5640 .ctl_name = CTL_KERN,
5641 .procname = "kernel",
5642 .mode = 0555,
5643 .child = sd_ctl_dir,
5645 {0, },
5648 static struct ctl_table *sd_alloc_ctl_entry(int n)
5650 struct ctl_table *entry =
5651 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5653 return entry;
5656 static void sd_free_ctl_entry(struct ctl_table **tablep)
5658 struct ctl_table *entry;
5661 * In the intermediate directories, both the child directory and
5662 * procname are dynamically allocated and could fail but the mode
5663 * will always be set. In the lowest directory the names are
5664 * static strings and all have proc handlers.
5666 for (entry = *tablep; entry->mode; entry++) {
5667 if (entry->child)
5668 sd_free_ctl_entry(&entry->child);
5669 if (entry->proc_handler == NULL)
5670 kfree(entry->procname);
5673 kfree(*tablep);
5674 *tablep = NULL;
5677 static void
5678 set_table_entry(struct ctl_table *entry,
5679 const char *procname, void *data, int maxlen,
5680 mode_t mode, proc_handler *proc_handler)
5682 entry->procname = procname;
5683 entry->data = data;
5684 entry->maxlen = maxlen;
5685 entry->mode = mode;
5686 entry->proc_handler = proc_handler;
5689 static struct ctl_table *
5690 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5692 struct ctl_table *table = sd_alloc_ctl_entry(12);
5694 if (table == NULL)
5695 return NULL;
5697 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5698 sizeof(long), 0644, proc_doulongvec_minmax);
5699 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5700 sizeof(long), 0644, proc_doulongvec_minmax);
5701 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5702 sizeof(int), 0644, proc_dointvec_minmax);
5703 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5704 sizeof(int), 0644, proc_dointvec_minmax);
5705 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5706 sizeof(int), 0644, proc_dointvec_minmax);
5707 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5708 sizeof(int), 0644, proc_dointvec_minmax);
5709 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5710 sizeof(int), 0644, proc_dointvec_minmax);
5711 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5712 sizeof(int), 0644, proc_dointvec_minmax);
5713 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5714 sizeof(int), 0644, proc_dointvec_minmax);
5715 set_table_entry(&table[9], "cache_nice_tries",
5716 &sd->cache_nice_tries,
5717 sizeof(int), 0644, proc_dointvec_minmax);
5718 set_table_entry(&table[10], "flags", &sd->flags,
5719 sizeof(int), 0644, proc_dointvec_minmax);
5720 /* &table[11] is terminator */
5722 return table;
5725 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5727 struct ctl_table *entry, *table;
5728 struct sched_domain *sd;
5729 int domain_num = 0, i;
5730 char buf[32];
5732 for_each_domain(cpu, sd)
5733 domain_num++;
5734 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5735 if (table == NULL)
5736 return NULL;
5738 i = 0;
5739 for_each_domain(cpu, sd) {
5740 snprintf(buf, 32, "domain%d", i);
5741 entry->procname = kstrdup(buf, GFP_KERNEL);
5742 entry->mode = 0555;
5743 entry->child = sd_alloc_ctl_domain_table(sd);
5744 entry++;
5745 i++;
5747 return table;
5750 static struct ctl_table_header *sd_sysctl_header;
5751 static void register_sched_domain_sysctl(void)
5753 int i, cpu_num = num_online_cpus();
5754 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5755 char buf[32];
5757 WARN_ON(sd_ctl_dir[0].child);
5758 sd_ctl_dir[0].child = entry;
5760 if (entry == NULL)
5761 return;
5763 for_each_online_cpu(i) {
5764 snprintf(buf, 32, "cpu%d", i);
5765 entry->procname = kstrdup(buf, GFP_KERNEL);
5766 entry->mode = 0555;
5767 entry->child = sd_alloc_ctl_cpu_table(i);
5768 entry++;
5771 WARN_ON(sd_sysctl_header);
5772 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5775 /* may be called multiple times per register */
5776 static void unregister_sched_domain_sysctl(void)
5778 if (sd_sysctl_header)
5779 unregister_sysctl_table(sd_sysctl_header);
5780 sd_sysctl_header = NULL;
5781 if (sd_ctl_dir[0].child)
5782 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5784 #else
5785 static void register_sched_domain_sysctl(void)
5788 static void unregister_sched_domain_sysctl(void)
5791 #endif
5794 * migration_call - callback that gets triggered when a CPU is added.
5795 * Here we can start up the necessary migration thread for the new CPU.
5797 static int __cpuinit
5798 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5800 struct task_struct *p;
5801 int cpu = (long)hcpu;
5802 unsigned long flags;
5803 struct rq *rq;
5805 switch (action) {
5807 case CPU_UP_PREPARE:
5808 case CPU_UP_PREPARE_FROZEN:
5809 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5810 if (IS_ERR(p))
5811 return NOTIFY_BAD;
5812 kthread_bind(p, cpu);
5813 /* Must be high prio: stop_machine expects to yield to it. */
5814 rq = task_rq_lock(p, &flags);
5815 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5816 task_rq_unlock(rq, &flags);
5817 cpu_rq(cpu)->migration_thread = p;
5818 break;
5820 case CPU_ONLINE:
5821 case CPU_ONLINE_FROZEN:
5822 /* Strictly unnecessary, as first user will wake it. */
5823 wake_up_process(cpu_rq(cpu)->migration_thread);
5825 /* Update our root-domain */
5826 rq = cpu_rq(cpu);
5827 spin_lock_irqsave(&rq->lock, flags);
5828 if (rq->rd) {
5829 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5830 cpu_set(cpu, rq->rd->online);
5832 spin_unlock_irqrestore(&rq->lock, flags);
5833 break;
5835 #ifdef CONFIG_HOTPLUG_CPU
5836 case CPU_UP_CANCELED:
5837 case CPU_UP_CANCELED_FROZEN:
5838 if (!cpu_rq(cpu)->migration_thread)
5839 break;
5840 /* Unbind it from offline cpu so it can run. Fall thru. */
5841 kthread_bind(cpu_rq(cpu)->migration_thread,
5842 any_online_cpu(cpu_online_map));
5843 kthread_stop(cpu_rq(cpu)->migration_thread);
5844 cpu_rq(cpu)->migration_thread = NULL;
5845 break;
5847 case CPU_DEAD:
5848 case CPU_DEAD_FROZEN:
5849 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5850 migrate_live_tasks(cpu);
5851 rq = cpu_rq(cpu);
5852 kthread_stop(rq->migration_thread);
5853 rq->migration_thread = NULL;
5854 /* Idle task back to normal (off runqueue, low prio) */
5855 spin_lock_irq(&rq->lock);
5856 update_rq_clock(rq);
5857 deactivate_task(rq, rq->idle, 0);
5858 rq->idle->static_prio = MAX_PRIO;
5859 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5860 rq->idle->sched_class = &idle_sched_class;
5861 migrate_dead_tasks(cpu);
5862 spin_unlock_irq(&rq->lock);
5863 cpuset_unlock();
5864 migrate_nr_uninterruptible(rq);
5865 BUG_ON(rq->nr_running != 0);
5868 * No need to migrate the tasks: it was best-effort if
5869 * they didn't take sched_hotcpu_mutex. Just wake up
5870 * the requestors.
5872 spin_lock_irq(&rq->lock);
5873 while (!list_empty(&rq->migration_queue)) {
5874 struct migration_req *req;
5876 req = list_entry(rq->migration_queue.next,
5877 struct migration_req, list);
5878 list_del_init(&req->list);
5879 complete(&req->done);
5881 spin_unlock_irq(&rq->lock);
5882 break;
5884 case CPU_DOWN_PREPARE:
5885 /* Update our root-domain */
5886 rq = cpu_rq(cpu);
5887 spin_lock_irqsave(&rq->lock, flags);
5888 if (rq->rd) {
5889 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5890 cpu_clear(cpu, rq->rd->online);
5892 spin_unlock_irqrestore(&rq->lock, flags);
5893 break;
5894 #endif
5896 return NOTIFY_OK;
5899 /* Register at highest priority so that task migration (migrate_all_tasks)
5900 * happens before everything else.
5902 static struct notifier_block __cpuinitdata migration_notifier = {
5903 .notifier_call = migration_call,
5904 .priority = 10
5907 void __init migration_init(void)
5909 void *cpu = (void *)(long)smp_processor_id();
5910 int err;
5912 /* Start one for the boot CPU: */
5913 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5914 BUG_ON(err == NOTIFY_BAD);
5915 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5916 register_cpu_notifier(&migration_notifier);
5918 #endif
5920 #ifdef CONFIG_SMP
5922 /* Number of possible processor ids */
5923 int nr_cpu_ids __read_mostly = NR_CPUS;
5924 EXPORT_SYMBOL(nr_cpu_ids);
5926 #ifdef CONFIG_SCHED_DEBUG
5928 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5930 struct sched_group *group = sd->groups;
5931 cpumask_t groupmask;
5932 char str[NR_CPUS];
5934 cpumask_scnprintf(str, NR_CPUS, sd->span);
5935 cpus_clear(groupmask);
5937 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5939 if (!(sd->flags & SD_LOAD_BALANCE)) {
5940 printk("does not load-balance\n");
5941 if (sd->parent)
5942 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5943 " has parent");
5944 return -1;
5947 printk(KERN_CONT "span %s\n", str);
5949 if (!cpu_isset(cpu, sd->span)) {
5950 printk(KERN_ERR "ERROR: domain->span does not contain "
5951 "CPU%d\n", cpu);
5953 if (!cpu_isset(cpu, group->cpumask)) {
5954 printk(KERN_ERR "ERROR: domain->groups does not contain"
5955 " CPU%d\n", cpu);
5958 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5959 do {
5960 if (!group) {
5961 printk("\n");
5962 printk(KERN_ERR "ERROR: group is NULL\n");
5963 break;
5966 if (!group->__cpu_power) {
5967 printk(KERN_CONT "\n");
5968 printk(KERN_ERR "ERROR: domain->cpu_power not "
5969 "set\n");
5970 break;
5973 if (!cpus_weight(group->cpumask)) {
5974 printk(KERN_CONT "\n");
5975 printk(KERN_ERR "ERROR: empty group\n");
5976 break;
5979 if (cpus_intersects(groupmask, group->cpumask)) {
5980 printk(KERN_CONT "\n");
5981 printk(KERN_ERR "ERROR: repeated CPUs\n");
5982 break;
5985 cpus_or(groupmask, groupmask, group->cpumask);
5987 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5988 printk(KERN_CONT " %s", str);
5990 group = group->next;
5991 } while (group != sd->groups);
5992 printk(KERN_CONT "\n");
5994 if (!cpus_equal(sd->span, groupmask))
5995 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5997 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5998 printk(KERN_ERR "ERROR: parent span is not a superset "
5999 "of domain->span\n");
6000 return 0;
6003 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6005 int level = 0;
6007 if (!sd) {
6008 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6009 return;
6012 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6014 for (;;) {
6015 if (sched_domain_debug_one(sd, cpu, level))
6016 break;
6017 level++;
6018 sd = sd->parent;
6019 if (!sd)
6020 break;
6023 #else
6024 # define sched_domain_debug(sd, cpu) do { } while (0)
6025 #endif
6027 static int sd_degenerate(struct sched_domain *sd)
6029 if (cpus_weight(sd->span) == 1)
6030 return 1;
6032 /* Following flags need at least 2 groups */
6033 if (sd->flags & (SD_LOAD_BALANCE |
6034 SD_BALANCE_NEWIDLE |
6035 SD_BALANCE_FORK |
6036 SD_BALANCE_EXEC |
6037 SD_SHARE_CPUPOWER |
6038 SD_SHARE_PKG_RESOURCES)) {
6039 if (sd->groups != sd->groups->next)
6040 return 0;
6043 /* Following flags don't use groups */
6044 if (sd->flags & (SD_WAKE_IDLE |
6045 SD_WAKE_AFFINE |
6046 SD_WAKE_BALANCE))
6047 return 0;
6049 return 1;
6052 static int
6053 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6055 unsigned long cflags = sd->flags, pflags = parent->flags;
6057 if (sd_degenerate(parent))
6058 return 1;
6060 if (!cpus_equal(sd->span, parent->span))
6061 return 0;
6063 /* Does parent contain flags not in child? */
6064 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6065 if (cflags & SD_WAKE_AFFINE)
6066 pflags &= ~SD_WAKE_BALANCE;
6067 /* Flags needing groups don't count if only 1 group in parent */
6068 if (parent->groups == parent->groups->next) {
6069 pflags &= ~(SD_LOAD_BALANCE |
6070 SD_BALANCE_NEWIDLE |
6071 SD_BALANCE_FORK |
6072 SD_BALANCE_EXEC |
6073 SD_SHARE_CPUPOWER |
6074 SD_SHARE_PKG_RESOURCES);
6076 if (~cflags & pflags)
6077 return 0;
6079 return 1;
6082 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6084 unsigned long flags;
6085 const struct sched_class *class;
6087 spin_lock_irqsave(&rq->lock, flags);
6089 if (rq->rd) {
6090 struct root_domain *old_rd = rq->rd;
6092 for (class = sched_class_highest; class; class = class->next) {
6093 if (class->leave_domain)
6094 class->leave_domain(rq);
6097 cpu_clear(rq->cpu, old_rd->span);
6098 cpu_clear(rq->cpu, old_rd->online);
6100 if (atomic_dec_and_test(&old_rd->refcount))
6101 kfree(old_rd);
6104 atomic_inc(&rd->refcount);
6105 rq->rd = rd;
6107 cpu_set(rq->cpu, rd->span);
6108 if (cpu_isset(rq->cpu, cpu_online_map))
6109 cpu_set(rq->cpu, rd->online);
6111 for (class = sched_class_highest; class; class = class->next) {
6112 if (class->join_domain)
6113 class->join_domain(rq);
6116 spin_unlock_irqrestore(&rq->lock, flags);
6119 static void init_rootdomain(struct root_domain *rd)
6121 memset(rd, 0, sizeof(*rd));
6123 cpus_clear(rd->span);
6124 cpus_clear(rd->online);
6127 static void init_defrootdomain(void)
6129 init_rootdomain(&def_root_domain);
6130 atomic_set(&def_root_domain.refcount, 1);
6133 static struct root_domain *alloc_rootdomain(void)
6135 struct root_domain *rd;
6137 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6138 if (!rd)
6139 return NULL;
6141 init_rootdomain(rd);
6143 return rd;
6147 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6148 * hold the hotplug lock.
6150 static void
6151 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6153 struct rq *rq = cpu_rq(cpu);
6154 struct sched_domain *tmp;
6156 /* Remove the sched domains which do not contribute to scheduling. */
6157 for (tmp = sd; tmp; tmp = tmp->parent) {
6158 struct sched_domain *parent = tmp->parent;
6159 if (!parent)
6160 break;
6161 if (sd_parent_degenerate(tmp, parent)) {
6162 tmp->parent = parent->parent;
6163 if (parent->parent)
6164 parent->parent->child = tmp;
6168 if (sd && sd_degenerate(sd)) {
6169 sd = sd->parent;
6170 if (sd)
6171 sd->child = NULL;
6174 sched_domain_debug(sd, cpu);
6176 rq_attach_root(rq, rd);
6177 rcu_assign_pointer(rq->sd, sd);
6180 /* cpus with isolated domains */
6181 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6183 /* Setup the mask of cpus configured for isolated domains */
6184 static int __init isolated_cpu_setup(char *str)
6186 int ints[NR_CPUS], i;
6188 str = get_options(str, ARRAY_SIZE(ints), ints);
6189 cpus_clear(cpu_isolated_map);
6190 for (i = 1; i <= ints[0]; i++)
6191 if (ints[i] < NR_CPUS)
6192 cpu_set(ints[i], cpu_isolated_map);
6193 return 1;
6196 __setup("isolcpus=", isolated_cpu_setup);
6199 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6200 * to a function which identifies what group(along with sched group) a CPU
6201 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6202 * (due to the fact that we keep track of groups covered with a cpumask_t).
6204 * init_sched_build_groups will build a circular linked list of the groups
6205 * covered by the given span, and will set each group's ->cpumask correctly,
6206 * and ->cpu_power to 0.
6208 static void
6209 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6210 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6211 struct sched_group **sg))
6213 struct sched_group *first = NULL, *last = NULL;
6214 cpumask_t covered = CPU_MASK_NONE;
6215 int i;
6217 for_each_cpu_mask(i, span) {
6218 struct sched_group *sg;
6219 int group = group_fn(i, cpu_map, &sg);
6220 int j;
6222 if (cpu_isset(i, covered))
6223 continue;
6225 sg->cpumask = CPU_MASK_NONE;
6226 sg->__cpu_power = 0;
6228 for_each_cpu_mask(j, span) {
6229 if (group_fn(j, cpu_map, NULL) != group)
6230 continue;
6232 cpu_set(j, covered);
6233 cpu_set(j, sg->cpumask);
6235 if (!first)
6236 first = sg;
6237 if (last)
6238 last->next = sg;
6239 last = sg;
6241 last->next = first;
6244 #define SD_NODES_PER_DOMAIN 16
6246 #ifdef CONFIG_NUMA
6249 * find_next_best_node - find the next node to include in a sched_domain
6250 * @node: node whose sched_domain we're building
6251 * @used_nodes: nodes already in the sched_domain
6253 * Find the next node to include in a given scheduling domain. Simply
6254 * finds the closest node not already in the @used_nodes map.
6256 * Should use nodemask_t.
6258 static int find_next_best_node(int node, unsigned long *used_nodes)
6260 int i, n, val, min_val, best_node = 0;
6262 min_val = INT_MAX;
6264 for (i = 0; i < MAX_NUMNODES; i++) {
6265 /* Start at @node */
6266 n = (node + i) % MAX_NUMNODES;
6268 if (!nr_cpus_node(n))
6269 continue;
6271 /* Skip already used nodes */
6272 if (test_bit(n, used_nodes))
6273 continue;
6275 /* Simple min distance search */
6276 val = node_distance(node, n);
6278 if (val < min_val) {
6279 min_val = val;
6280 best_node = n;
6284 set_bit(best_node, used_nodes);
6285 return best_node;
6289 * sched_domain_node_span - get a cpumask for a node's sched_domain
6290 * @node: node whose cpumask we're constructing
6291 * @size: number of nodes to include in this span
6293 * Given a node, construct a good cpumask for its sched_domain to span. It
6294 * should be one that prevents unnecessary balancing, but also spreads tasks
6295 * out optimally.
6297 static cpumask_t sched_domain_node_span(int node)
6299 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6300 cpumask_t span, nodemask;
6301 int i;
6303 cpus_clear(span);
6304 bitmap_zero(used_nodes, MAX_NUMNODES);
6306 nodemask = node_to_cpumask(node);
6307 cpus_or(span, span, nodemask);
6308 set_bit(node, used_nodes);
6310 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6311 int next_node = find_next_best_node(node, used_nodes);
6313 nodemask = node_to_cpumask(next_node);
6314 cpus_or(span, span, nodemask);
6317 return span;
6319 #endif
6321 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6324 * SMT sched-domains:
6326 #ifdef CONFIG_SCHED_SMT
6327 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6328 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6330 static int
6331 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6333 if (sg)
6334 *sg = &per_cpu(sched_group_cpus, cpu);
6335 return cpu;
6337 #endif
6340 * multi-core sched-domains:
6342 #ifdef CONFIG_SCHED_MC
6343 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6344 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6345 #endif
6347 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6348 static int
6349 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6351 int group;
6352 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6353 cpus_and(mask, mask, *cpu_map);
6354 group = first_cpu(mask);
6355 if (sg)
6356 *sg = &per_cpu(sched_group_core, group);
6357 return group;
6359 #elif defined(CONFIG_SCHED_MC)
6360 static int
6361 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6363 if (sg)
6364 *sg = &per_cpu(sched_group_core, cpu);
6365 return cpu;
6367 #endif
6369 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6370 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6372 static int
6373 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6375 int group;
6376 #ifdef CONFIG_SCHED_MC
6377 cpumask_t mask = cpu_coregroup_map(cpu);
6378 cpus_and(mask, mask, *cpu_map);
6379 group = first_cpu(mask);
6380 #elif defined(CONFIG_SCHED_SMT)
6381 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6382 cpus_and(mask, mask, *cpu_map);
6383 group = first_cpu(mask);
6384 #else
6385 group = cpu;
6386 #endif
6387 if (sg)
6388 *sg = &per_cpu(sched_group_phys, group);
6389 return group;
6392 #ifdef CONFIG_NUMA
6394 * The init_sched_build_groups can't handle what we want to do with node
6395 * groups, so roll our own. Now each node has its own list of groups which
6396 * gets dynamically allocated.
6398 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6399 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6401 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6402 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6404 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6405 struct sched_group **sg)
6407 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6408 int group;
6410 cpus_and(nodemask, nodemask, *cpu_map);
6411 group = first_cpu(nodemask);
6413 if (sg)
6414 *sg = &per_cpu(sched_group_allnodes, group);
6415 return group;
6418 static void init_numa_sched_groups_power(struct sched_group *group_head)
6420 struct sched_group *sg = group_head;
6421 int j;
6423 if (!sg)
6424 return;
6425 do {
6426 for_each_cpu_mask(j, sg->cpumask) {
6427 struct sched_domain *sd;
6429 sd = &per_cpu(phys_domains, j);
6430 if (j != first_cpu(sd->groups->cpumask)) {
6432 * Only add "power" once for each
6433 * physical package.
6435 continue;
6438 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6440 sg = sg->next;
6441 } while (sg != group_head);
6443 #endif
6445 #ifdef CONFIG_NUMA
6446 /* Free memory allocated for various sched_group structures */
6447 static void free_sched_groups(const cpumask_t *cpu_map)
6449 int cpu, i;
6451 for_each_cpu_mask(cpu, *cpu_map) {
6452 struct sched_group **sched_group_nodes
6453 = sched_group_nodes_bycpu[cpu];
6455 if (!sched_group_nodes)
6456 continue;
6458 for (i = 0; i < MAX_NUMNODES; i++) {
6459 cpumask_t nodemask = node_to_cpumask(i);
6460 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6462 cpus_and(nodemask, nodemask, *cpu_map);
6463 if (cpus_empty(nodemask))
6464 continue;
6466 if (sg == NULL)
6467 continue;
6468 sg = sg->next;
6469 next_sg:
6470 oldsg = sg;
6471 sg = sg->next;
6472 kfree(oldsg);
6473 if (oldsg != sched_group_nodes[i])
6474 goto next_sg;
6476 kfree(sched_group_nodes);
6477 sched_group_nodes_bycpu[cpu] = NULL;
6480 #else
6481 static void free_sched_groups(const cpumask_t *cpu_map)
6484 #endif
6487 * Initialize sched groups cpu_power.
6489 * cpu_power indicates the capacity of sched group, which is used while
6490 * distributing the load between different sched groups in a sched domain.
6491 * Typically cpu_power for all the groups in a sched domain will be same unless
6492 * there are asymmetries in the topology. If there are asymmetries, group
6493 * having more cpu_power will pickup more load compared to the group having
6494 * less cpu_power.
6496 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6497 * the maximum number of tasks a group can handle in the presence of other idle
6498 * or lightly loaded groups in the same sched domain.
6500 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6502 struct sched_domain *child;
6503 struct sched_group *group;
6505 WARN_ON(!sd || !sd->groups);
6507 if (cpu != first_cpu(sd->groups->cpumask))
6508 return;
6510 child = sd->child;
6512 sd->groups->__cpu_power = 0;
6515 * For perf policy, if the groups in child domain share resources
6516 * (for example cores sharing some portions of the cache hierarchy
6517 * or SMT), then set this domain groups cpu_power such that each group
6518 * can handle only one task, when there are other idle groups in the
6519 * same sched domain.
6521 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6522 (child->flags &
6523 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6524 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6525 return;
6529 * add cpu_power of each child group to this groups cpu_power
6531 group = child->groups;
6532 do {
6533 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6534 group = group->next;
6535 } while (group != child->groups);
6539 * Build sched domains for a given set of cpus and attach the sched domains
6540 * to the individual cpus
6542 static int build_sched_domains(const cpumask_t *cpu_map)
6544 int i;
6545 struct root_domain *rd;
6546 #ifdef CONFIG_NUMA
6547 struct sched_group **sched_group_nodes = NULL;
6548 int sd_allnodes = 0;
6551 * Allocate the per-node list of sched groups
6553 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6554 GFP_KERNEL);
6555 if (!sched_group_nodes) {
6556 printk(KERN_WARNING "Can not alloc sched group node list\n");
6557 return -ENOMEM;
6559 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6560 #endif
6562 rd = alloc_rootdomain();
6563 if (!rd) {
6564 printk(KERN_WARNING "Cannot alloc root domain\n");
6565 return -ENOMEM;
6569 * Set up domains for cpus specified by the cpu_map.
6571 for_each_cpu_mask(i, *cpu_map) {
6572 struct sched_domain *sd = NULL, *p;
6573 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6575 cpus_and(nodemask, nodemask, *cpu_map);
6577 #ifdef CONFIG_NUMA
6578 if (cpus_weight(*cpu_map) >
6579 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6580 sd = &per_cpu(allnodes_domains, i);
6581 *sd = SD_ALLNODES_INIT;
6582 sd->span = *cpu_map;
6583 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6584 p = sd;
6585 sd_allnodes = 1;
6586 } else
6587 p = NULL;
6589 sd = &per_cpu(node_domains, i);
6590 *sd = SD_NODE_INIT;
6591 sd->span = sched_domain_node_span(cpu_to_node(i));
6592 sd->parent = p;
6593 if (p)
6594 p->child = sd;
6595 cpus_and(sd->span, sd->span, *cpu_map);
6596 #endif
6598 p = sd;
6599 sd = &per_cpu(phys_domains, i);
6600 *sd = SD_CPU_INIT;
6601 sd->span = nodemask;
6602 sd->parent = p;
6603 if (p)
6604 p->child = sd;
6605 cpu_to_phys_group(i, cpu_map, &sd->groups);
6607 #ifdef CONFIG_SCHED_MC
6608 p = sd;
6609 sd = &per_cpu(core_domains, i);
6610 *sd = SD_MC_INIT;
6611 sd->span = cpu_coregroup_map(i);
6612 cpus_and(sd->span, sd->span, *cpu_map);
6613 sd->parent = p;
6614 p->child = sd;
6615 cpu_to_core_group(i, cpu_map, &sd->groups);
6616 #endif
6618 #ifdef CONFIG_SCHED_SMT
6619 p = sd;
6620 sd = &per_cpu(cpu_domains, i);
6621 *sd = SD_SIBLING_INIT;
6622 sd->span = per_cpu(cpu_sibling_map, i);
6623 cpus_and(sd->span, sd->span, *cpu_map);
6624 sd->parent = p;
6625 p->child = sd;
6626 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6627 #endif
6630 #ifdef CONFIG_SCHED_SMT
6631 /* Set up CPU (sibling) groups */
6632 for_each_cpu_mask(i, *cpu_map) {
6633 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6634 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6635 if (i != first_cpu(this_sibling_map))
6636 continue;
6638 init_sched_build_groups(this_sibling_map, cpu_map,
6639 &cpu_to_cpu_group);
6641 #endif
6643 #ifdef CONFIG_SCHED_MC
6644 /* Set up multi-core groups */
6645 for_each_cpu_mask(i, *cpu_map) {
6646 cpumask_t this_core_map = cpu_coregroup_map(i);
6647 cpus_and(this_core_map, this_core_map, *cpu_map);
6648 if (i != first_cpu(this_core_map))
6649 continue;
6650 init_sched_build_groups(this_core_map, cpu_map,
6651 &cpu_to_core_group);
6653 #endif
6655 /* Set up physical groups */
6656 for (i = 0; i < MAX_NUMNODES; i++) {
6657 cpumask_t nodemask = node_to_cpumask(i);
6659 cpus_and(nodemask, nodemask, *cpu_map);
6660 if (cpus_empty(nodemask))
6661 continue;
6663 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6666 #ifdef CONFIG_NUMA
6667 /* Set up node groups */
6668 if (sd_allnodes)
6669 init_sched_build_groups(*cpu_map, cpu_map,
6670 &cpu_to_allnodes_group);
6672 for (i = 0; i < MAX_NUMNODES; i++) {
6673 /* Set up node groups */
6674 struct sched_group *sg, *prev;
6675 cpumask_t nodemask = node_to_cpumask(i);
6676 cpumask_t domainspan;
6677 cpumask_t covered = CPU_MASK_NONE;
6678 int j;
6680 cpus_and(nodemask, nodemask, *cpu_map);
6681 if (cpus_empty(nodemask)) {
6682 sched_group_nodes[i] = NULL;
6683 continue;
6686 domainspan = sched_domain_node_span(i);
6687 cpus_and(domainspan, domainspan, *cpu_map);
6689 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6690 if (!sg) {
6691 printk(KERN_WARNING "Can not alloc domain group for "
6692 "node %d\n", i);
6693 goto error;
6695 sched_group_nodes[i] = sg;
6696 for_each_cpu_mask(j, nodemask) {
6697 struct sched_domain *sd;
6699 sd = &per_cpu(node_domains, j);
6700 sd->groups = sg;
6702 sg->__cpu_power = 0;
6703 sg->cpumask = nodemask;
6704 sg->next = sg;
6705 cpus_or(covered, covered, nodemask);
6706 prev = sg;
6708 for (j = 0; j < MAX_NUMNODES; j++) {
6709 cpumask_t tmp, notcovered;
6710 int n = (i + j) % MAX_NUMNODES;
6712 cpus_complement(notcovered, covered);
6713 cpus_and(tmp, notcovered, *cpu_map);
6714 cpus_and(tmp, tmp, domainspan);
6715 if (cpus_empty(tmp))
6716 break;
6718 nodemask = node_to_cpumask(n);
6719 cpus_and(tmp, tmp, nodemask);
6720 if (cpus_empty(tmp))
6721 continue;
6723 sg = kmalloc_node(sizeof(struct sched_group),
6724 GFP_KERNEL, i);
6725 if (!sg) {
6726 printk(KERN_WARNING
6727 "Can not alloc domain group for node %d\n", j);
6728 goto error;
6730 sg->__cpu_power = 0;
6731 sg->cpumask = tmp;
6732 sg->next = prev->next;
6733 cpus_or(covered, covered, tmp);
6734 prev->next = sg;
6735 prev = sg;
6738 #endif
6740 /* Calculate CPU power for physical packages and nodes */
6741 #ifdef CONFIG_SCHED_SMT
6742 for_each_cpu_mask(i, *cpu_map) {
6743 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6745 init_sched_groups_power(i, sd);
6747 #endif
6748 #ifdef CONFIG_SCHED_MC
6749 for_each_cpu_mask(i, *cpu_map) {
6750 struct sched_domain *sd = &per_cpu(core_domains, i);
6752 init_sched_groups_power(i, sd);
6754 #endif
6756 for_each_cpu_mask(i, *cpu_map) {
6757 struct sched_domain *sd = &per_cpu(phys_domains, i);
6759 init_sched_groups_power(i, sd);
6762 #ifdef CONFIG_NUMA
6763 for (i = 0; i < MAX_NUMNODES; i++)
6764 init_numa_sched_groups_power(sched_group_nodes[i]);
6766 if (sd_allnodes) {
6767 struct sched_group *sg;
6769 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6770 init_numa_sched_groups_power(sg);
6772 #endif
6774 /* Attach the domains */
6775 for_each_cpu_mask(i, *cpu_map) {
6776 struct sched_domain *sd;
6777 #ifdef CONFIG_SCHED_SMT
6778 sd = &per_cpu(cpu_domains, i);
6779 #elif defined(CONFIG_SCHED_MC)
6780 sd = &per_cpu(core_domains, i);
6781 #else
6782 sd = &per_cpu(phys_domains, i);
6783 #endif
6784 cpu_attach_domain(sd, rd, i);
6787 return 0;
6789 #ifdef CONFIG_NUMA
6790 error:
6791 free_sched_groups(cpu_map);
6792 return -ENOMEM;
6793 #endif
6796 static cpumask_t *doms_cur; /* current sched domains */
6797 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6800 * Special case: If a kmalloc of a doms_cur partition (array of
6801 * cpumask_t) fails, then fallback to a single sched domain,
6802 * as determined by the single cpumask_t fallback_doms.
6804 static cpumask_t fallback_doms;
6807 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6808 * For now this just excludes isolated cpus, but could be used to
6809 * exclude other special cases in the future.
6811 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6813 int err;
6815 ndoms_cur = 1;
6816 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6817 if (!doms_cur)
6818 doms_cur = &fallback_doms;
6819 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6820 err = build_sched_domains(doms_cur);
6821 register_sched_domain_sysctl();
6823 return err;
6826 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6828 free_sched_groups(cpu_map);
6832 * Detach sched domains from a group of cpus specified in cpu_map
6833 * These cpus will now be attached to the NULL domain
6835 static void detach_destroy_domains(const cpumask_t *cpu_map)
6837 int i;
6839 unregister_sched_domain_sysctl();
6841 for_each_cpu_mask(i, *cpu_map)
6842 cpu_attach_domain(NULL, &def_root_domain, i);
6843 synchronize_sched();
6844 arch_destroy_sched_domains(cpu_map);
6848 * Partition sched domains as specified by the 'ndoms_new'
6849 * cpumasks in the array doms_new[] of cpumasks. This compares
6850 * doms_new[] to the current sched domain partitioning, doms_cur[].
6851 * It destroys each deleted domain and builds each new domain.
6853 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6854 * The masks don't intersect (don't overlap.) We should setup one
6855 * sched domain for each mask. CPUs not in any of the cpumasks will
6856 * not be load balanced. If the same cpumask appears both in the
6857 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6858 * it as it is.
6860 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6861 * ownership of it and will kfree it when done with it. If the caller
6862 * failed the kmalloc call, then it can pass in doms_new == NULL,
6863 * and partition_sched_domains() will fallback to the single partition
6864 * 'fallback_doms'.
6866 * Call with hotplug lock held
6868 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6870 int i, j;
6872 lock_doms_cur();
6874 /* always unregister in case we don't destroy any domains */
6875 unregister_sched_domain_sysctl();
6877 if (doms_new == NULL) {
6878 ndoms_new = 1;
6879 doms_new = &fallback_doms;
6880 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6883 /* Destroy deleted domains */
6884 for (i = 0; i < ndoms_cur; i++) {
6885 for (j = 0; j < ndoms_new; j++) {
6886 if (cpus_equal(doms_cur[i], doms_new[j]))
6887 goto match1;
6889 /* no match - a current sched domain not in new doms_new[] */
6890 detach_destroy_domains(doms_cur + i);
6891 match1:
6895 /* Build new domains */
6896 for (i = 0; i < ndoms_new; i++) {
6897 for (j = 0; j < ndoms_cur; j++) {
6898 if (cpus_equal(doms_new[i], doms_cur[j]))
6899 goto match2;
6901 /* no match - add a new doms_new */
6902 build_sched_domains(doms_new + i);
6903 match2:
6907 /* Remember the new sched domains */
6908 if (doms_cur != &fallback_doms)
6909 kfree(doms_cur);
6910 doms_cur = doms_new;
6911 ndoms_cur = ndoms_new;
6913 register_sched_domain_sysctl();
6915 unlock_doms_cur();
6918 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6919 static int arch_reinit_sched_domains(void)
6921 int err;
6923 get_online_cpus();
6924 detach_destroy_domains(&cpu_online_map);
6925 err = arch_init_sched_domains(&cpu_online_map);
6926 put_online_cpus();
6928 return err;
6931 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6933 int ret;
6935 if (buf[0] != '0' && buf[0] != '1')
6936 return -EINVAL;
6938 if (smt)
6939 sched_smt_power_savings = (buf[0] == '1');
6940 else
6941 sched_mc_power_savings = (buf[0] == '1');
6943 ret = arch_reinit_sched_domains();
6945 return ret ? ret : count;
6948 #ifdef CONFIG_SCHED_MC
6949 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6951 return sprintf(page, "%u\n", sched_mc_power_savings);
6953 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6954 const char *buf, size_t count)
6956 return sched_power_savings_store(buf, count, 0);
6958 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6959 sched_mc_power_savings_store);
6960 #endif
6962 #ifdef CONFIG_SCHED_SMT
6963 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6965 return sprintf(page, "%u\n", sched_smt_power_savings);
6967 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6968 const char *buf, size_t count)
6970 return sched_power_savings_store(buf, count, 1);
6972 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6973 sched_smt_power_savings_store);
6974 #endif
6976 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6978 int err = 0;
6980 #ifdef CONFIG_SCHED_SMT
6981 if (smt_capable())
6982 err = sysfs_create_file(&cls->kset.kobj,
6983 &attr_sched_smt_power_savings.attr);
6984 #endif
6985 #ifdef CONFIG_SCHED_MC
6986 if (!err && mc_capable())
6987 err = sysfs_create_file(&cls->kset.kobj,
6988 &attr_sched_mc_power_savings.attr);
6989 #endif
6990 return err;
6992 #endif
6995 * Force a reinitialization of the sched domains hierarchy. The domains
6996 * and groups cannot be updated in place without racing with the balancing
6997 * code, so we temporarily attach all running cpus to the NULL domain
6998 * which will prevent rebalancing while the sched domains are recalculated.
7000 static int update_sched_domains(struct notifier_block *nfb,
7001 unsigned long action, void *hcpu)
7003 switch (action) {
7004 case CPU_UP_PREPARE:
7005 case CPU_UP_PREPARE_FROZEN:
7006 case CPU_DOWN_PREPARE:
7007 case CPU_DOWN_PREPARE_FROZEN:
7008 detach_destroy_domains(&cpu_online_map);
7009 return NOTIFY_OK;
7011 case CPU_UP_CANCELED:
7012 case CPU_UP_CANCELED_FROZEN:
7013 case CPU_DOWN_FAILED:
7014 case CPU_DOWN_FAILED_FROZEN:
7015 case CPU_ONLINE:
7016 case CPU_ONLINE_FROZEN:
7017 case CPU_DEAD:
7018 case CPU_DEAD_FROZEN:
7020 * Fall through and re-initialise the domains.
7022 break;
7023 default:
7024 return NOTIFY_DONE;
7027 /* The hotplug lock is already held by cpu_up/cpu_down */
7028 arch_init_sched_domains(&cpu_online_map);
7030 return NOTIFY_OK;
7033 void __init sched_init_smp(void)
7035 cpumask_t non_isolated_cpus;
7037 get_online_cpus();
7038 arch_init_sched_domains(&cpu_online_map);
7039 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7040 if (cpus_empty(non_isolated_cpus))
7041 cpu_set(smp_processor_id(), non_isolated_cpus);
7042 put_online_cpus();
7043 /* XXX: Theoretical race here - CPU may be hotplugged now */
7044 hotcpu_notifier(update_sched_domains, 0);
7046 /* Move init over to a non-isolated CPU */
7047 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7048 BUG();
7049 sched_init_granularity();
7051 #else
7052 void __init sched_init_smp(void)
7054 sched_init_granularity();
7056 #endif /* CONFIG_SMP */
7058 int in_sched_functions(unsigned long addr)
7060 return in_lock_functions(addr) ||
7061 (addr >= (unsigned long)__sched_text_start
7062 && addr < (unsigned long)__sched_text_end);
7065 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7067 cfs_rq->tasks_timeline = RB_ROOT;
7068 #ifdef CONFIG_FAIR_GROUP_SCHED
7069 cfs_rq->rq = rq;
7070 #endif
7071 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7074 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7076 struct rt_prio_array *array;
7077 int i;
7079 array = &rt_rq->active;
7080 for (i = 0; i < MAX_RT_PRIO; i++) {
7081 INIT_LIST_HEAD(array->queue + i);
7082 __clear_bit(i, array->bitmap);
7084 /* delimiter for bitsearch: */
7085 __set_bit(MAX_RT_PRIO, array->bitmap);
7087 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7088 rt_rq->highest_prio = MAX_RT_PRIO;
7089 #endif
7090 #ifdef CONFIG_SMP
7091 rt_rq->rt_nr_migratory = 0;
7092 rt_rq->overloaded = 0;
7093 #endif
7095 rt_rq->rt_time = 0;
7096 rt_rq->rt_throttled = 0;
7098 #ifdef CONFIG_RT_GROUP_SCHED
7099 rt_rq->rt_nr_boosted = 0;
7100 rt_rq->rq = rq;
7101 #endif
7104 #ifdef CONFIG_FAIR_GROUP_SCHED
7105 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7106 struct cfs_rq *cfs_rq, struct sched_entity *se,
7107 int cpu, int add)
7109 tg->cfs_rq[cpu] = cfs_rq;
7110 init_cfs_rq(cfs_rq, rq);
7111 cfs_rq->tg = tg;
7112 if (add)
7113 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7115 tg->se[cpu] = se;
7116 se->cfs_rq = &rq->cfs;
7117 se->my_q = cfs_rq;
7118 se->load.weight = tg->shares;
7119 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7120 se->parent = NULL;
7122 #endif
7124 #ifdef CONFIG_RT_GROUP_SCHED
7125 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7126 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7127 int cpu, int add)
7129 tg->rt_rq[cpu] = rt_rq;
7130 init_rt_rq(rt_rq, rq);
7131 rt_rq->tg = tg;
7132 rt_rq->rt_se = rt_se;
7133 if (add)
7134 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7136 tg->rt_se[cpu] = rt_se;
7137 rt_se->rt_rq = &rq->rt;
7138 rt_se->my_q = rt_rq;
7139 rt_se->parent = NULL;
7140 INIT_LIST_HEAD(&rt_se->run_list);
7142 #endif
7144 void __init sched_init(void)
7146 int highest_cpu = 0;
7147 int i, j;
7149 #ifdef CONFIG_SMP
7150 init_defrootdomain();
7151 #endif
7153 #ifdef CONFIG_GROUP_SCHED
7154 list_add(&init_task_group.list, &task_groups);
7155 #endif
7157 for_each_possible_cpu(i) {
7158 struct rq *rq;
7160 rq = cpu_rq(i);
7161 spin_lock_init(&rq->lock);
7162 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7163 rq->nr_running = 0;
7164 rq->clock = 1;
7165 init_cfs_rq(&rq->cfs, rq);
7166 init_rt_rq(&rq->rt, rq);
7167 #ifdef CONFIG_FAIR_GROUP_SCHED
7168 init_task_group.shares = init_task_group_load;
7169 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7170 init_tg_cfs_entry(rq, &init_task_group,
7171 &per_cpu(init_cfs_rq, i),
7172 &per_cpu(init_sched_entity, i), i, 1);
7174 #endif
7175 #ifdef CONFIG_RT_GROUP_SCHED
7176 init_task_group.rt_runtime =
7177 sysctl_sched_rt_runtime * NSEC_PER_USEC;
7178 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7179 init_tg_rt_entry(rq, &init_task_group,
7180 &per_cpu(init_rt_rq, i),
7181 &per_cpu(init_sched_rt_entity, i), i, 1);
7182 #endif
7183 rq->rt_period_expire = 0;
7184 rq->rt_throttled = 0;
7186 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7187 rq->cpu_load[j] = 0;
7188 #ifdef CONFIG_SMP
7189 rq->sd = NULL;
7190 rq->rd = NULL;
7191 rq->active_balance = 0;
7192 rq->next_balance = jiffies;
7193 rq->push_cpu = 0;
7194 rq->cpu = i;
7195 rq->migration_thread = NULL;
7196 INIT_LIST_HEAD(&rq->migration_queue);
7197 rq_attach_root(rq, &def_root_domain);
7198 #endif
7199 init_rq_hrtick(rq);
7200 atomic_set(&rq->nr_iowait, 0);
7201 highest_cpu = i;
7204 set_load_weight(&init_task);
7206 #ifdef CONFIG_PREEMPT_NOTIFIERS
7207 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7208 #endif
7210 #ifdef CONFIG_SMP
7211 nr_cpu_ids = highest_cpu + 1;
7212 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7213 #endif
7215 #ifdef CONFIG_RT_MUTEXES
7216 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7217 #endif
7220 * The boot idle thread does lazy MMU switching as well:
7222 atomic_inc(&init_mm.mm_count);
7223 enter_lazy_tlb(&init_mm, current);
7226 * Make us the idle thread. Technically, schedule() should not be
7227 * called from this thread, however somewhere below it might be,
7228 * but because we are the idle thread, we just pick up running again
7229 * when this runqueue becomes "idle".
7231 init_idle(current, smp_processor_id());
7233 * During early bootup we pretend to be a normal task:
7235 current->sched_class = &fair_sched_class;
7237 scheduler_running = 1;
7240 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7241 void __might_sleep(char *file, int line)
7243 #ifdef in_atomic
7244 static unsigned long prev_jiffy; /* ratelimiting */
7246 if ((in_atomic() || irqs_disabled()) &&
7247 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7248 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7249 return;
7250 prev_jiffy = jiffies;
7251 printk(KERN_ERR "BUG: sleeping function called from invalid"
7252 " context at %s:%d\n", file, line);
7253 printk("in_atomic():%d, irqs_disabled():%d\n",
7254 in_atomic(), irqs_disabled());
7255 debug_show_held_locks(current);
7256 if (irqs_disabled())
7257 print_irqtrace_events(current);
7258 dump_stack();
7260 #endif
7262 EXPORT_SYMBOL(__might_sleep);
7263 #endif
7265 #ifdef CONFIG_MAGIC_SYSRQ
7266 static void normalize_task(struct rq *rq, struct task_struct *p)
7268 int on_rq;
7269 update_rq_clock(rq);
7270 on_rq = p->se.on_rq;
7271 if (on_rq)
7272 deactivate_task(rq, p, 0);
7273 __setscheduler(rq, p, SCHED_NORMAL, 0);
7274 if (on_rq) {
7275 activate_task(rq, p, 0);
7276 resched_task(rq->curr);
7280 void normalize_rt_tasks(void)
7282 struct task_struct *g, *p;
7283 unsigned long flags;
7284 struct rq *rq;
7286 read_lock_irqsave(&tasklist_lock, flags);
7287 do_each_thread(g, p) {
7289 * Only normalize user tasks:
7291 if (!p->mm)
7292 continue;
7294 p->se.exec_start = 0;
7295 #ifdef CONFIG_SCHEDSTATS
7296 p->se.wait_start = 0;
7297 p->se.sleep_start = 0;
7298 p->se.block_start = 0;
7299 #endif
7300 task_rq(p)->clock = 0;
7302 if (!rt_task(p)) {
7304 * Renice negative nice level userspace
7305 * tasks back to 0:
7307 if (TASK_NICE(p) < 0 && p->mm)
7308 set_user_nice(p, 0);
7309 continue;
7312 spin_lock(&p->pi_lock);
7313 rq = __task_rq_lock(p);
7315 normalize_task(rq, p);
7317 __task_rq_unlock(rq);
7318 spin_unlock(&p->pi_lock);
7319 } while_each_thread(g, p);
7321 read_unlock_irqrestore(&tasklist_lock, flags);
7324 #endif /* CONFIG_MAGIC_SYSRQ */
7326 #ifdef CONFIG_IA64
7328 * These functions are only useful for the IA64 MCA handling.
7330 * They can only be called when the whole system has been
7331 * stopped - every CPU needs to be quiescent, and no scheduling
7332 * activity can take place. Using them for anything else would
7333 * be a serious bug, and as a result, they aren't even visible
7334 * under any other configuration.
7338 * curr_task - return the current task for a given cpu.
7339 * @cpu: the processor in question.
7341 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7343 struct task_struct *curr_task(int cpu)
7345 return cpu_curr(cpu);
7349 * set_curr_task - set the current task for a given cpu.
7350 * @cpu: the processor in question.
7351 * @p: the task pointer to set.
7353 * Description: This function must only be used when non-maskable interrupts
7354 * are serviced on a separate stack. It allows the architecture to switch the
7355 * notion of the current task on a cpu in a non-blocking manner. This function
7356 * must be called with all CPU's synchronized, and interrupts disabled, the
7357 * and caller must save the original value of the current task (see
7358 * curr_task() above) and restore that value before reenabling interrupts and
7359 * re-starting the system.
7361 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7363 void set_curr_task(int cpu, struct task_struct *p)
7365 cpu_curr(cpu) = p;
7368 #endif
7370 #ifdef CONFIG_GROUP_SCHED
7372 #ifdef CONFIG_FAIR_GROUP_SCHED
7373 static void free_fair_sched_group(struct task_group *tg)
7375 int i;
7377 for_each_possible_cpu(i) {
7378 if (tg->cfs_rq)
7379 kfree(tg->cfs_rq[i]);
7380 if (tg->se)
7381 kfree(tg->se[i]);
7384 kfree(tg->cfs_rq);
7385 kfree(tg->se);
7388 static int alloc_fair_sched_group(struct task_group *tg)
7390 struct cfs_rq *cfs_rq;
7391 struct sched_entity *se;
7392 struct rq *rq;
7393 int i;
7395 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7396 if (!tg->cfs_rq)
7397 goto err;
7398 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7399 if (!tg->se)
7400 goto err;
7402 tg->shares = NICE_0_LOAD;
7404 for_each_possible_cpu(i) {
7405 rq = cpu_rq(i);
7407 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7408 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7409 if (!cfs_rq)
7410 goto err;
7412 se = kmalloc_node(sizeof(struct sched_entity),
7413 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7414 if (!se)
7415 goto err;
7417 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7420 return 1;
7422 err:
7423 return 0;
7426 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7428 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7429 &cpu_rq(cpu)->leaf_cfs_rq_list);
7432 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7434 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7436 #else
7437 static inline void free_fair_sched_group(struct task_group *tg)
7441 static inline int alloc_fair_sched_group(struct task_group *tg)
7443 return 1;
7446 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7450 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7453 #endif
7455 #ifdef CONFIG_RT_GROUP_SCHED
7456 static void free_rt_sched_group(struct task_group *tg)
7458 int i;
7460 for_each_possible_cpu(i) {
7461 if (tg->rt_rq)
7462 kfree(tg->rt_rq[i]);
7463 if (tg->rt_se)
7464 kfree(tg->rt_se[i]);
7467 kfree(tg->rt_rq);
7468 kfree(tg->rt_se);
7471 static int alloc_rt_sched_group(struct task_group *tg)
7473 struct rt_rq *rt_rq;
7474 struct sched_rt_entity *rt_se;
7475 struct rq *rq;
7476 int i;
7478 tg->rt_rq = kzalloc(sizeof(rt_rq) * NR_CPUS, GFP_KERNEL);
7479 if (!tg->rt_rq)
7480 goto err;
7481 tg->rt_se = kzalloc(sizeof(rt_se) * NR_CPUS, GFP_KERNEL);
7482 if (!tg->rt_se)
7483 goto err;
7485 tg->rt_runtime = 0;
7487 for_each_possible_cpu(i) {
7488 rq = cpu_rq(i);
7490 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7491 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7492 if (!rt_rq)
7493 goto err;
7495 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7496 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7497 if (!rt_se)
7498 goto err;
7500 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7503 return 1;
7505 err:
7506 return 0;
7509 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7511 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7512 &cpu_rq(cpu)->leaf_rt_rq_list);
7515 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7517 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7519 #else
7520 static inline void free_rt_sched_group(struct task_group *tg)
7524 static inline int alloc_rt_sched_group(struct task_group *tg)
7526 return 1;
7529 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7533 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7536 #endif
7538 static void free_sched_group(struct task_group *tg)
7540 free_fair_sched_group(tg);
7541 free_rt_sched_group(tg);
7542 kfree(tg);
7545 /* allocate runqueue etc for a new task group */
7546 struct task_group *sched_create_group(void)
7548 struct task_group *tg;
7549 unsigned long flags;
7550 int i;
7552 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7553 if (!tg)
7554 return ERR_PTR(-ENOMEM);
7556 if (!alloc_fair_sched_group(tg))
7557 goto err;
7559 if (!alloc_rt_sched_group(tg))
7560 goto err;
7562 spin_lock_irqsave(&task_group_lock, flags);
7563 for_each_possible_cpu(i) {
7564 register_fair_sched_group(tg, i);
7565 register_rt_sched_group(tg, i);
7567 list_add_rcu(&tg->list, &task_groups);
7568 spin_unlock_irqrestore(&task_group_lock, flags);
7570 return tg;
7572 err:
7573 free_sched_group(tg);
7574 return ERR_PTR(-ENOMEM);
7577 /* rcu callback to free various structures associated with a task group */
7578 static void free_sched_group_rcu(struct rcu_head *rhp)
7580 /* now it should be safe to free those cfs_rqs */
7581 free_sched_group(container_of(rhp, struct task_group, rcu));
7584 /* Destroy runqueue etc associated with a task group */
7585 void sched_destroy_group(struct task_group *tg)
7587 unsigned long flags;
7588 int i;
7590 spin_lock_irqsave(&task_group_lock, flags);
7591 for_each_possible_cpu(i) {
7592 unregister_fair_sched_group(tg, i);
7593 unregister_rt_sched_group(tg, i);
7595 list_del_rcu(&tg->list);
7596 spin_unlock_irqrestore(&task_group_lock, flags);
7598 /* wait for possible concurrent references to cfs_rqs complete */
7599 call_rcu(&tg->rcu, free_sched_group_rcu);
7602 /* change task's runqueue when it moves between groups.
7603 * The caller of this function should have put the task in its new group
7604 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7605 * reflect its new group.
7607 void sched_move_task(struct task_struct *tsk)
7609 int on_rq, running;
7610 unsigned long flags;
7611 struct rq *rq;
7613 rq = task_rq_lock(tsk, &flags);
7615 update_rq_clock(rq);
7617 running = task_current(rq, tsk);
7618 on_rq = tsk->se.on_rq;
7620 if (on_rq) {
7621 dequeue_task(rq, tsk, 0);
7622 if (unlikely(running))
7623 tsk->sched_class->put_prev_task(rq, tsk);
7626 set_task_rq(tsk, task_cpu(tsk));
7628 if (on_rq) {
7629 if (unlikely(running))
7630 tsk->sched_class->set_curr_task(rq);
7631 enqueue_task(rq, tsk, 0);
7634 task_rq_unlock(rq, &flags);
7637 #ifdef CONFIG_FAIR_GROUP_SCHED
7638 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7640 struct cfs_rq *cfs_rq = se->cfs_rq;
7641 struct rq *rq = cfs_rq->rq;
7642 int on_rq;
7644 spin_lock_irq(&rq->lock);
7646 on_rq = se->on_rq;
7647 if (on_rq)
7648 dequeue_entity(cfs_rq, se, 0);
7650 se->load.weight = shares;
7651 se->load.inv_weight = div64_64((1ULL<<32), shares);
7653 if (on_rq)
7654 enqueue_entity(cfs_rq, se, 0);
7656 spin_unlock_irq(&rq->lock);
7659 static DEFINE_MUTEX(shares_mutex);
7661 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7663 int i;
7664 unsigned long flags;
7667 * A weight of 0 or 1 can cause arithmetics problems.
7668 * (The default weight is 1024 - so there's no practical
7669 * limitation from this.)
7671 if (shares < 2)
7672 shares = 2;
7674 mutex_lock(&shares_mutex);
7675 if (tg->shares == shares)
7676 goto done;
7678 spin_lock_irqsave(&task_group_lock, flags);
7679 for_each_possible_cpu(i)
7680 unregister_fair_sched_group(tg, i);
7681 spin_unlock_irqrestore(&task_group_lock, flags);
7683 /* wait for any ongoing reference to this group to finish */
7684 synchronize_sched();
7687 * Now we are free to modify the group's share on each cpu
7688 * w/o tripping rebalance_share or load_balance_fair.
7690 tg->shares = shares;
7691 for_each_possible_cpu(i)
7692 set_se_shares(tg->se[i], shares);
7695 * Enable load balance activity on this group, by inserting it back on
7696 * each cpu's rq->leaf_cfs_rq_list.
7698 spin_lock_irqsave(&task_group_lock, flags);
7699 for_each_possible_cpu(i)
7700 register_fair_sched_group(tg, i);
7701 spin_unlock_irqrestore(&task_group_lock, flags);
7702 done:
7703 mutex_unlock(&shares_mutex);
7704 return 0;
7707 unsigned long sched_group_shares(struct task_group *tg)
7709 return tg->shares;
7711 #endif
7713 #ifdef CONFIG_RT_GROUP_SCHED
7715 * Ensure that the real time constraints are schedulable.
7717 static DEFINE_MUTEX(rt_constraints_mutex);
7719 static unsigned long to_ratio(u64 period, u64 runtime)
7721 if (runtime == RUNTIME_INF)
7722 return 1ULL << 16;
7724 runtime *= (1ULL << 16);
7725 div64_64(runtime, period);
7726 return runtime;
7729 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7731 struct task_group *tgi;
7732 unsigned long total = 0;
7733 unsigned long global_ratio =
7734 to_ratio(sysctl_sched_rt_period,
7735 sysctl_sched_rt_runtime < 0 ?
7736 RUNTIME_INF : sysctl_sched_rt_runtime);
7738 rcu_read_lock();
7739 list_for_each_entry_rcu(tgi, &task_groups, list) {
7740 if (tgi == tg)
7741 continue;
7743 total += to_ratio(period, tgi->rt_runtime);
7745 rcu_read_unlock();
7747 return total + to_ratio(period, runtime) < global_ratio;
7750 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7752 u64 rt_runtime, rt_period;
7753 int err = 0;
7755 rt_period = sysctl_sched_rt_period * NSEC_PER_USEC;
7756 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7757 if (rt_runtime_us == -1)
7758 rt_runtime = rt_period;
7760 mutex_lock(&rt_constraints_mutex);
7761 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
7762 err = -EINVAL;
7763 goto unlock;
7765 if (rt_runtime_us == -1)
7766 rt_runtime = RUNTIME_INF;
7767 tg->rt_runtime = rt_runtime;
7768 unlock:
7769 mutex_unlock(&rt_constraints_mutex);
7771 return err;
7774 long sched_group_rt_runtime(struct task_group *tg)
7776 u64 rt_runtime_us;
7778 if (tg->rt_runtime == RUNTIME_INF)
7779 return -1;
7781 rt_runtime_us = tg->rt_runtime;
7782 do_div(rt_runtime_us, NSEC_PER_USEC);
7783 return rt_runtime_us;
7785 #endif
7786 #endif /* CONFIG_GROUP_SCHED */
7788 #ifdef CONFIG_CGROUP_SCHED
7790 /* return corresponding task_group object of a cgroup */
7791 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7793 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7794 struct task_group, css);
7797 static struct cgroup_subsys_state *
7798 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7800 struct task_group *tg;
7802 if (!cgrp->parent) {
7803 /* This is early initialization for the top cgroup */
7804 init_task_group.css.cgroup = cgrp;
7805 return &init_task_group.css;
7808 /* we support only 1-level deep hierarchical scheduler atm */
7809 if (cgrp->parent->parent)
7810 return ERR_PTR(-EINVAL);
7812 tg = sched_create_group();
7813 if (IS_ERR(tg))
7814 return ERR_PTR(-ENOMEM);
7816 /* Bind the cgroup to task_group object we just created */
7817 tg->css.cgroup = cgrp;
7819 return &tg->css;
7822 static void
7823 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7825 struct task_group *tg = cgroup_tg(cgrp);
7827 sched_destroy_group(tg);
7830 static int
7831 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7832 struct task_struct *tsk)
7834 #ifdef CONFIG_RT_GROUP_SCHED
7835 /* Don't accept realtime tasks when there is no way for them to run */
7836 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_runtime == 0)
7837 return -EINVAL;
7838 #else
7839 /* We don't support RT-tasks being in separate groups */
7840 if (tsk->sched_class != &fair_sched_class)
7841 return -EINVAL;
7842 #endif
7844 return 0;
7847 static void
7848 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7849 struct cgroup *old_cont, struct task_struct *tsk)
7851 sched_move_task(tsk);
7854 #ifdef CONFIG_FAIR_GROUP_SCHED
7855 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7856 u64 shareval)
7858 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7861 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7863 struct task_group *tg = cgroup_tg(cgrp);
7865 return (u64) tg->shares;
7867 #endif
7869 #ifdef CONFIG_RT_GROUP_SCHED
7870 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7871 struct file *file,
7872 const char __user *userbuf,
7873 size_t nbytes, loff_t *unused_ppos)
7875 char buffer[64];
7876 int retval = 0;
7877 s64 val;
7878 char *end;
7880 if (!nbytes)
7881 return -EINVAL;
7882 if (nbytes >= sizeof(buffer))
7883 return -E2BIG;
7884 if (copy_from_user(buffer, userbuf, nbytes))
7885 return -EFAULT;
7887 buffer[nbytes] = 0; /* nul-terminate */
7889 /* strip newline if necessary */
7890 if (nbytes && (buffer[nbytes-1] == '\n'))
7891 buffer[nbytes-1] = 0;
7892 val = simple_strtoll(buffer, &end, 0);
7893 if (*end)
7894 return -EINVAL;
7896 /* Pass to subsystem */
7897 retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7898 if (!retval)
7899 retval = nbytes;
7900 return retval;
7903 static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
7904 struct file *file,
7905 char __user *buf, size_t nbytes,
7906 loff_t *ppos)
7908 char tmp[64];
7909 long val = sched_group_rt_runtime(cgroup_tg(cgrp));
7910 int len = sprintf(tmp, "%ld\n", val);
7912 return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
7914 #endif
7916 static struct cftype cpu_files[] = {
7917 #ifdef CONFIG_FAIR_GROUP_SCHED
7919 .name = "shares",
7920 .read_uint = cpu_shares_read_uint,
7921 .write_uint = cpu_shares_write_uint,
7923 #endif
7924 #ifdef CONFIG_RT_GROUP_SCHED
7926 .name = "rt_runtime_us",
7927 .read = cpu_rt_runtime_read,
7928 .write = cpu_rt_runtime_write,
7930 #endif
7933 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7935 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7938 struct cgroup_subsys cpu_cgroup_subsys = {
7939 .name = "cpu",
7940 .create = cpu_cgroup_create,
7941 .destroy = cpu_cgroup_destroy,
7942 .can_attach = cpu_cgroup_can_attach,
7943 .attach = cpu_cgroup_attach,
7944 .populate = cpu_cgroup_populate,
7945 .subsys_id = cpu_cgroup_subsys_id,
7946 .early_init = 1,
7949 #endif /* CONFIG_CGROUP_SCHED */
7951 #ifdef CONFIG_CGROUP_CPUACCT
7954 * CPU accounting code for task groups.
7956 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7957 * (balbir@in.ibm.com).
7960 /* track cpu usage of a group of tasks */
7961 struct cpuacct {
7962 struct cgroup_subsys_state css;
7963 /* cpuusage holds pointer to a u64-type object on every cpu */
7964 u64 *cpuusage;
7967 struct cgroup_subsys cpuacct_subsys;
7969 /* return cpu accounting group corresponding to this container */
7970 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7972 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7973 struct cpuacct, css);
7976 /* return cpu accounting group to which this task belongs */
7977 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7979 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7980 struct cpuacct, css);
7983 /* create a new cpu accounting group */
7984 static struct cgroup_subsys_state *cpuacct_create(
7985 struct cgroup_subsys *ss, struct cgroup *cont)
7987 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7989 if (!ca)
7990 return ERR_PTR(-ENOMEM);
7992 ca->cpuusage = alloc_percpu(u64);
7993 if (!ca->cpuusage) {
7994 kfree(ca);
7995 return ERR_PTR(-ENOMEM);
7998 return &ca->css;
8001 /* destroy an existing cpu accounting group */
8002 static void
8003 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
8005 struct cpuacct *ca = cgroup_ca(cont);
8007 free_percpu(ca->cpuusage);
8008 kfree(ca);
8011 /* return total cpu usage (in nanoseconds) of a group */
8012 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
8014 struct cpuacct *ca = cgroup_ca(cont);
8015 u64 totalcpuusage = 0;
8016 int i;
8018 for_each_possible_cpu(i) {
8019 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8022 * Take rq->lock to make 64-bit addition safe on 32-bit
8023 * platforms.
8025 spin_lock_irq(&cpu_rq(i)->lock);
8026 totalcpuusage += *cpuusage;
8027 spin_unlock_irq(&cpu_rq(i)->lock);
8030 return totalcpuusage;
8033 static struct cftype files[] = {
8035 .name = "usage",
8036 .read_uint = cpuusage_read,
8040 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8042 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
8046 * charge this task's execution time to its accounting group.
8048 * called with rq->lock held.
8050 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8052 struct cpuacct *ca;
8054 if (!cpuacct_subsys.active)
8055 return;
8057 ca = task_ca(tsk);
8058 if (ca) {
8059 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8061 *cpuusage += cputime;
8065 struct cgroup_subsys cpuacct_subsys = {
8066 .name = "cpuacct",
8067 .create = cpuacct_create,
8068 .destroy = cpuacct_destroy,
8069 .populate = cpuacct_populate,
8070 .subsys_id = cpuacct_subsys_id,
8072 #endif /* CONFIG_CGROUP_CPUACCT */