[MTD] [OneNAND] Use pre-alloced oob buffer instead of local buffer
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / sched.c
blobb4fbbc440453c8d76c686128fad818939d70740e
1 /*
2 * kernel/sched.c
4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
27 #include <linux/mm.h>
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/cpu_acct.h>
56 #include <linux/kthread.h>
57 #include <linux/seq_file.h>
58 #include <linux/sysctl.h>
59 #include <linux/syscalls.h>
60 #include <linux/times.h>
61 #include <linux/tsacct_kern.h>
62 #include <linux/kprobes.h>
63 #include <linux/delayacct.h>
64 #include <linux/reciprocal_div.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
68 #include <asm/tlb.h>
69 #include <asm/irq_regs.h>
72 * Scheduler clock - returns current time in nanosec units.
73 * This is default implementation.
74 * Architectures and sub-architectures can override this.
76 unsigned long long __attribute__((weak)) sched_clock(void)
78 return (unsigned long long)jiffies * (1000000000 / HZ);
82 * Convert user-nice values [ -20 ... 0 ... 19 ]
83 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 * and back.
86 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
87 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
88 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
91 * 'User priority' is the nice value converted to something we
92 * can work with better when scaling various scheduler parameters,
93 * it's a [ 0 ... 39 ] range.
95 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
96 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
97 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
100 * Some helpers for converting nanosecond timing to jiffy resolution
102 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (1000000000 / HZ))
103 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
116 #ifdef CONFIG_SMP
118 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
119 * Since cpu_power is a 'constant', we can use a reciprocal divide.
121 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
123 return reciprocal_divide(load, sg->reciprocal_cpu_power);
127 * Each time a sched group cpu_power is changed,
128 * we must compute its reciprocal value
130 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
132 sg->__cpu_power += val;
133 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
135 #endif
137 static inline int rt_policy(int policy)
139 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
140 return 1;
141 return 0;
144 static inline int task_has_rt_policy(struct task_struct *p)
146 return rt_policy(p->policy);
150 * This is the priority-queue data structure of the RT scheduling class:
152 struct rt_prio_array {
153 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
154 struct list_head queue[MAX_RT_PRIO];
157 #ifdef CONFIG_FAIR_GROUP_SCHED
159 #include <linux/cgroup.h>
161 struct cfs_rq;
163 /* task group related information */
164 struct task_group {
165 #ifdef CONFIG_FAIR_CGROUP_SCHED
166 struct cgroup_subsys_state css;
167 #endif
168 /* schedulable entities of this group on each cpu */
169 struct sched_entity **se;
170 /* runqueue "owned" by this group on each cpu */
171 struct cfs_rq **cfs_rq;
172 unsigned long shares;
173 /* spinlock to serialize modification to shares */
174 spinlock_t lock;
177 /* Default task group's sched entity on each cpu */
178 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
179 /* Default task group's cfs_rq on each cpu */
180 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
182 static struct sched_entity *init_sched_entity_p[NR_CPUS];
183 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
185 /* Default task group.
186 * Every task in system belong to this group at bootup.
188 struct task_group init_task_group = {
189 .se = init_sched_entity_p,
190 .cfs_rq = init_cfs_rq_p,
193 #ifdef CONFIG_FAIR_USER_SCHED
194 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
195 #else
196 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
197 #endif
199 static int init_task_group_load = INIT_TASK_GRP_LOAD;
201 /* return group to which a task belongs */
202 static inline struct task_group *task_group(struct task_struct *p)
204 struct task_group *tg;
206 #ifdef CONFIG_FAIR_USER_SCHED
207 tg = p->user->tg;
208 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
209 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
210 struct task_group, css);
211 #else
212 tg = &init_task_group;
213 #endif
215 return tg;
218 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
219 static inline void set_task_cfs_rq(struct task_struct *p)
221 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
222 p->se.parent = task_group(p)->se[task_cpu(p)];
225 #else
227 static inline void set_task_cfs_rq(struct task_struct *p) { }
229 #endif /* CONFIG_FAIR_GROUP_SCHED */
231 /* CFS-related fields in a runqueue */
232 struct cfs_rq {
233 struct load_weight load;
234 unsigned long nr_running;
236 u64 exec_clock;
237 u64 min_vruntime;
239 struct rb_root tasks_timeline;
240 struct rb_node *rb_leftmost;
241 struct rb_node *rb_load_balance_curr;
242 /* 'curr' points to currently running entity on this cfs_rq.
243 * It is set to NULL otherwise (i.e when none are currently running).
245 struct sched_entity *curr;
247 unsigned long nr_spread_over;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
252 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
253 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
254 * (like users, containers etc.)
256 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
257 * list is used during load balance.
259 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
260 struct task_group *tg; /* group that "owns" this runqueue */
261 struct rcu_head rcu;
262 #endif
265 /* Real-Time classes' related field in a runqueue: */
266 struct rt_rq {
267 struct rt_prio_array active;
268 int rt_load_balance_idx;
269 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
273 * This is the main, per-CPU runqueue data structure.
275 * Locking rule: those places that want to lock multiple runqueues
276 * (such as the load balancing or the thread migration code), lock
277 * acquire operations must be ordered by ascending &runqueue.
279 struct rq {
280 /* runqueue lock: */
281 spinlock_t lock;
284 * nr_running and cpu_load should be in the same cacheline because
285 * remote CPUs use both these fields when doing load calculation.
287 unsigned long nr_running;
288 #define CPU_LOAD_IDX_MAX 5
289 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
290 unsigned char idle_at_tick;
291 #ifdef CONFIG_NO_HZ
292 unsigned char in_nohz_recently;
293 #endif
294 /* capture load from *all* tasks on this cpu: */
295 struct load_weight load;
296 unsigned long nr_load_updates;
297 u64 nr_switches;
299 struct cfs_rq cfs;
300 #ifdef CONFIG_FAIR_GROUP_SCHED
301 /* list of leaf cfs_rq on this cpu: */
302 struct list_head leaf_cfs_rq_list;
303 #endif
304 struct rt_rq rt;
307 * This is part of a global counter where only the total sum
308 * over all CPUs matters. A task can increase this counter on
309 * one CPU and if it got migrated afterwards it may decrease
310 * it on another CPU. Always updated under the runqueue lock:
312 unsigned long nr_uninterruptible;
314 struct task_struct *curr, *idle;
315 unsigned long next_balance;
316 struct mm_struct *prev_mm;
318 u64 clock, prev_clock_raw;
319 s64 clock_max_delta;
321 unsigned int clock_warps, clock_overflows;
322 u64 idle_clock;
323 unsigned int clock_deep_idle_events;
324 u64 tick_timestamp;
326 atomic_t nr_iowait;
328 #ifdef CONFIG_SMP
329 struct sched_domain *sd;
331 /* For active balancing */
332 int active_balance;
333 int push_cpu;
334 /* cpu of this runqueue: */
335 int cpu;
337 struct task_struct *migration_thread;
338 struct list_head migration_queue;
339 #endif
341 #ifdef CONFIG_SCHEDSTATS
342 /* latency stats */
343 struct sched_info rq_sched_info;
345 /* sys_sched_yield() stats */
346 unsigned int yld_exp_empty;
347 unsigned int yld_act_empty;
348 unsigned int yld_both_empty;
349 unsigned int yld_count;
351 /* schedule() stats */
352 unsigned int sched_switch;
353 unsigned int sched_count;
354 unsigned int sched_goidle;
356 /* try_to_wake_up() stats */
357 unsigned int ttwu_count;
358 unsigned int ttwu_local;
360 /* BKL stats */
361 unsigned int bkl_count;
362 #endif
363 struct lock_class_key rq_lock_key;
366 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
367 static DEFINE_MUTEX(sched_hotcpu_mutex);
369 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
371 rq->curr->sched_class->check_preempt_curr(rq, p);
374 static inline int cpu_of(struct rq *rq)
376 #ifdef CONFIG_SMP
377 return rq->cpu;
378 #else
379 return 0;
380 #endif
384 * Update the per-runqueue clock, as finegrained as the platform can give
385 * us, but without assuming monotonicity, etc.:
387 static void __update_rq_clock(struct rq *rq)
389 u64 prev_raw = rq->prev_clock_raw;
390 u64 now = sched_clock();
391 s64 delta = now - prev_raw;
392 u64 clock = rq->clock;
394 #ifdef CONFIG_SCHED_DEBUG
395 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
396 #endif
398 * Protect against sched_clock() occasionally going backwards:
400 if (unlikely(delta < 0)) {
401 clock++;
402 rq->clock_warps++;
403 } else {
405 * Catch too large forward jumps too:
407 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
408 if (clock < rq->tick_timestamp + TICK_NSEC)
409 clock = rq->tick_timestamp + TICK_NSEC;
410 else
411 clock++;
412 rq->clock_overflows++;
413 } else {
414 if (unlikely(delta > rq->clock_max_delta))
415 rq->clock_max_delta = delta;
416 clock += delta;
420 rq->prev_clock_raw = now;
421 rq->clock = clock;
424 static void update_rq_clock(struct rq *rq)
426 if (likely(smp_processor_id() == cpu_of(rq)))
427 __update_rq_clock(rq);
431 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
432 * See detach_destroy_domains: synchronize_sched for details.
434 * The domain tree of any CPU may only be accessed from within
435 * preempt-disabled sections.
437 #define for_each_domain(cpu, __sd) \
438 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
440 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
441 #define this_rq() (&__get_cpu_var(runqueues))
442 #define task_rq(p) cpu_rq(task_cpu(p))
443 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
446 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
448 #ifdef CONFIG_SCHED_DEBUG
449 # define const_debug __read_mostly
450 #else
451 # define const_debug static const
452 #endif
455 * Debugging: various feature bits
457 enum {
458 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
459 SCHED_FEAT_START_DEBIT = 2,
460 SCHED_FEAT_TREE_AVG = 4,
461 SCHED_FEAT_APPROX_AVG = 8,
462 SCHED_FEAT_WAKEUP_PREEMPT = 16,
463 SCHED_FEAT_PREEMPT_RESTRICT = 32,
466 const_debug unsigned int sysctl_sched_features =
467 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
468 SCHED_FEAT_START_DEBIT * 1 |
469 SCHED_FEAT_TREE_AVG * 0 |
470 SCHED_FEAT_APPROX_AVG * 0 |
471 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
472 SCHED_FEAT_PREEMPT_RESTRICT * 1;
474 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
477 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
478 * clock constructed from sched_clock():
480 unsigned long long cpu_clock(int cpu)
482 unsigned long long now;
483 unsigned long flags;
484 struct rq *rq;
486 local_irq_save(flags);
487 rq = cpu_rq(cpu);
488 update_rq_clock(rq);
489 now = rq->clock;
490 local_irq_restore(flags);
492 return now;
494 EXPORT_SYMBOL_GPL(cpu_clock);
496 #ifndef prepare_arch_switch
497 # define prepare_arch_switch(next) do { } while (0)
498 #endif
499 #ifndef finish_arch_switch
500 # define finish_arch_switch(prev) do { } while (0)
501 #endif
503 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
504 static inline int task_running(struct rq *rq, struct task_struct *p)
506 return rq->curr == p;
509 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
513 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
515 #ifdef CONFIG_DEBUG_SPINLOCK
516 /* this is a valid case when another task releases the spinlock */
517 rq->lock.owner = current;
518 #endif
520 * If we are tracking spinlock dependencies then we have to
521 * fix up the runqueue lock - which gets 'carried over' from
522 * prev into current:
524 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
526 spin_unlock_irq(&rq->lock);
529 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
530 static inline int task_running(struct rq *rq, struct task_struct *p)
532 #ifdef CONFIG_SMP
533 return p->oncpu;
534 #else
535 return rq->curr == p;
536 #endif
539 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
541 #ifdef CONFIG_SMP
543 * We can optimise this out completely for !SMP, because the
544 * SMP rebalancing from interrupt is the only thing that cares
545 * here.
547 next->oncpu = 1;
548 #endif
549 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
550 spin_unlock_irq(&rq->lock);
551 #else
552 spin_unlock(&rq->lock);
553 #endif
556 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
558 #ifdef CONFIG_SMP
560 * After ->oncpu is cleared, the task can be moved to a different CPU.
561 * We must ensure this doesn't happen until the switch is completely
562 * finished.
564 smp_wmb();
565 prev->oncpu = 0;
566 #endif
567 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
568 local_irq_enable();
569 #endif
571 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
574 * __task_rq_lock - lock the runqueue a given task resides on.
575 * Must be called interrupts disabled.
577 static inline struct rq *__task_rq_lock(struct task_struct *p)
578 __acquires(rq->lock)
580 for (;;) {
581 struct rq *rq = task_rq(p);
582 spin_lock(&rq->lock);
583 if (likely(rq == task_rq(p)))
584 return rq;
585 spin_unlock(&rq->lock);
590 * task_rq_lock - lock the runqueue a given task resides on and disable
591 * interrupts. Note the ordering: we can safely lookup the task_rq without
592 * explicitly disabling preemption.
594 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
595 __acquires(rq->lock)
597 struct rq *rq;
599 for (;;) {
600 local_irq_save(*flags);
601 rq = task_rq(p);
602 spin_lock(&rq->lock);
603 if (likely(rq == task_rq(p)))
604 return rq;
605 spin_unlock_irqrestore(&rq->lock, *flags);
609 static void __task_rq_unlock(struct rq *rq)
610 __releases(rq->lock)
612 spin_unlock(&rq->lock);
615 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
616 __releases(rq->lock)
618 spin_unlock_irqrestore(&rq->lock, *flags);
622 * this_rq_lock - lock this runqueue and disable interrupts.
624 static struct rq *this_rq_lock(void)
625 __acquires(rq->lock)
627 struct rq *rq;
629 local_irq_disable();
630 rq = this_rq();
631 spin_lock(&rq->lock);
633 return rq;
637 * We are going deep-idle (irqs are disabled):
639 void sched_clock_idle_sleep_event(void)
641 struct rq *rq = cpu_rq(smp_processor_id());
643 spin_lock(&rq->lock);
644 __update_rq_clock(rq);
645 spin_unlock(&rq->lock);
646 rq->clock_deep_idle_events++;
648 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
651 * We just idled delta nanoseconds (called with irqs disabled):
653 void sched_clock_idle_wakeup_event(u64 delta_ns)
655 struct rq *rq = cpu_rq(smp_processor_id());
656 u64 now = sched_clock();
658 rq->idle_clock += delta_ns;
660 * Override the previous timestamp and ignore all
661 * sched_clock() deltas that occured while we idled,
662 * and use the PM-provided delta_ns to advance the
663 * rq clock:
665 spin_lock(&rq->lock);
666 rq->prev_clock_raw = now;
667 rq->clock += delta_ns;
668 spin_unlock(&rq->lock);
670 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
673 * resched_task - mark a task 'to be rescheduled now'.
675 * On UP this means the setting of the need_resched flag, on SMP it
676 * might also involve a cross-CPU call to trigger the scheduler on
677 * the target CPU.
679 #ifdef CONFIG_SMP
681 #ifndef tsk_is_polling
682 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
683 #endif
685 static void resched_task(struct task_struct *p)
687 int cpu;
689 assert_spin_locked(&task_rq(p)->lock);
691 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
692 return;
694 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
696 cpu = task_cpu(p);
697 if (cpu == smp_processor_id())
698 return;
700 /* NEED_RESCHED must be visible before we test polling */
701 smp_mb();
702 if (!tsk_is_polling(p))
703 smp_send_reschedule(cpu);
706 static void resched_cpu(int cpu)
708 struct rq *rq = cpu_rq(cpu);
709 unsigned long flags;
711 if (!spin_trylock_irqsave(&rq->lock, flags))
712 return;
713 resched_task(cpu_curr(cpu));
714 spin_unlock_irqrestore(&rq->lock, flags);
716 #else
717 static inline void resched_task(struct task_struct *p)
719 assert_spin_locked(&task_rq(p)->lock);
720 set_tsk_need_resched(p);
722 #endif
724 #if BITS_PER_LONG == 32
725 # define WMULT_CONST (~0UL)
726 #else
727 # define WMULT_CONST (1UL << 32)
728 #endif
730 #define WMULT_SHIFT 32
733 * Shift right and round:
735 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
737 static unsigned long
738 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
739 struct load_weight *lw)
741 u64 tmp;
743 if (unlikely(!lw->inv_weight))
744 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
746 tmp = (u64)delta_exec * weight;
748 * Check whether we'd overflow the 64-bit multiplication:
750 if (unlikely(tmp > WMULT_CONST))
751 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
752 WMULT_SHIFT/2);
753 else
754 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
756 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
759 static inline unsigned long
760 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
762 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
765 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
767 lw->weight += inc;
770 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
772 lw->weight -= dec;
776 * To aid in avoiding the subversion of "niceness" due to uneven distribution
777 * of tasks with abnormal "nice" values across CPUs the contribution that
778 * each task makes to its run queue's load is weighted according to its
779 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
780 * scaled version of the new time slice allocation that they receive on time
781 * slice expiry etc.
784 #define WEIGHT_IDLEPRIO 2
785 #define WMULT_IDLEPRIO (1 << 31)
788 * Nice levels are multiplicative, with a gentle 10% change for every
789 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
790 * nice 1, it will get ~10% less CPU time than another CPU-bound task
791 * that remained on nice 0.
793 * The "10% effect" is relative and cumulative: from _any_ nice level,
794 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
795 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
796 * If a task goes up by ~10% and another task goes down by ~10% then
797 * the relative distance between them is ~25%.)
799 static const int prio_to_weight[40] = {
800 /* -20 */ 88761, 71755, 56483, 46273, 36291,
801 /* -15 */ 29154, 23254, 18705, 14949, 11916,
802 /* -10 */ 9548, 7620, 6100, 4904, 3906,
803 /* -5 */ 3121, 2501, 1991, 1586, 1277,
804 /* 0 */ 1024, 820, 655, 526, 423,
805 /* 5 */ 335, 272, 215, 172, 137,
806 /* 10 */ 110, 87, 70, 56, 45,
807 /* 15 */ 36, 29, 23, 18, 15,
811 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
813 * In cases where the weight does not change often, we can use the
814 * precalculated inverse to speed up arithmetics by turning divisions
815 * into multiplications:
817 static const u32 prio_to_wmult[40] = {
818 /* -20 */ 48388, 59856, 76040, 92818, 118348,
819 /* -15 */ 147320, 184698, 229616, 287308, 360437,
820 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
821 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
822 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
823 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
824 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
825 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
828 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
831 * runqueue iterator, to support SMP load-balancing between different
832 * scheduling classes, without having to expose their internal data
833 * structures to the load-balancing proper:
835 struct rq_iterator {
836 void *arg;
837 struct task_struct *(*start)(void *);
838 struct task_struct *(*next)(void *);
841 #ifdef CONFIG_SMP
842 static unsigned long
843 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
844 unsigned long max_load_move, struct sched_domain *sd,
845 enum cpu_idle_type idle, int *all_pinned,
846 int *this_best_prio, struct rq_iterator *iterator);
848 static int
849 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
850 struct sched_domain *sd, enum cpu_idle_type idle,
851 struct rq_iterator *iterator);
852 #endif
854 #include "sched_stats.h"
855 #include "sched_idletask.c"
856 #include "sched_fair.c"
857 #include "sched_rt.c"
858 #ifdef CONFIG_SCHED_DEBUG
859 # include "sched_debug.c"
860 #endif
862 #define sched_class_highest (&rt_sched_class)
865 * Update delta_exec, delta_fair fields for rq.
867 * delta_fair clock advances at a rate inversely proportional to
868 * total load (rq->load.weight) on the runqueue, while
869 * delta_exec advances at the same rate as wall-clock (provided
870 * cpu is not idle).
872 * delta_exec / delta_fair is a measure of the (smoothened) load on this
873 * runqueue over any given interval. This (smoothened) load is used
874 * during load balance.
876 * This function is called /before/ updating rq->load
877 * and when switching tasks.
879 static inline void inc_load(struct rq *rq, const struct task_struct *p)
881 update_load_add(&rq->load, p->se.load.weight);
884 static inline void dec_load(struct rq *rq, const struct task_struct *p)
886 update_load_sub(&rq->load, p->se.load.weight);
889 static void inc_nr_running(struct task_struct *p, struct rq *rq)
891 rq->nr_running++;
892 inc_load(rq, p);
895 static void dec_nr_running(struct task_struct *p, struct rq *rq)
897 rq->nr_running--;
898 dec_load(rq, p);
901 static void set_load_weight(struct task_struct *p)
903 if (task_has_rt_policy(p)) {
904 p->se.load.weight = prio_to_weight[0] * 2;
905 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
906 return;
910 * SCHED_IDLE tasks get minimal weight:
912 if (p->policy == SCHED_IDLE) {
913 p->se.load.weight = WEIGHT_IDLEPRIO;
914 p->se.load.inv_weight = WMULT_IDLEPRIO;
915 return;
918 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
919 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
922 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
924 sched_info_queued(p);
925 p->sched_class->enqueue_task(rq, p, wakeup);
926 p->se.on_rq = 1;
929 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
931 p->sched_class->dequeue_task(rq, p, sleep);
932 p->se.on_rq = 0;
936 * __normal_prio - return the priority that is based on the static prio
938 static inline int __normal_prio(struct task_struct *p)
940 return p->static_prio;
944 * Calculate the expected normal priority: i.e. priority
945 * without taking RT-inheritance into account. Might be
946 * boosted by interactivity modifiers. Changes upon fork,
947 * setprio syscalls, and whenever the interactivity
948 * estimator recalculates.
950 static inline int normal_prio(struct task_struct *p)
952 int prio;
954 if (task_has_rt_policy(p))
955 prio = MAX_RT_PRIO-1 - p->rt_priority;
956 else
957 prio = __normal_prio(p);
958 return prio;
962 * Calculate the current priority, i.e. the priority
963 * taken into account by the scheduler. This value might
964 * be boosted by RT tasks, or might be boosted by
965 * interactivity modifiers. Will be RT if the task got
966 * RT-boosted. If not then it returns p->normal_prio.
968 static int effective_prio(struct task_struct *p)
970 p->normal_prio = normal_prio(p);
972 * If we are RT tasks or we were boosted to RT priority,
973 * keep the priority unchanged. Otherwise, update priority
974 * to the normal priority:
976 if (!rt_prio(p->prio))
977 return p->normal_prio;
978 return p->prio;
982 * activate_task - move a task to the runqueue.
984 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
986 if (p->state == TASK_UNINTERRUPTIBLE)
987 rq->nr_uninterruptible--;
989 enqueue_task(rq, p, wakeup);
990 inc_nr_running(p, rq);
994 * deactivate_task - remove a task from the runqueue.
996 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
998 if (p->state == TASK_UNINTERRUPTIBLE)
999 rq->nr_uninterruptible++;
1001 dequeue_task(rq, p, sleep);
1002 dec_nr_running(p, rq);
1006 * task_curr - is this task currently executing on a CPU?
1007 * @p: the task in question.
1009 inline int task_curr(const struct task_struct *p)
1011 return cpu_curr(task_cpu(p)) == p;
1014 /* Used instead of source_load when we know the type == 0 */
1015 unsigned long weighted_cpuload(const int cpu)
1017 return cpu_rq(cpu)->load.weight;
1020 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1022 #ifdef CONFIG_SMP
1023 task_thread_info(p)->cpu = cpu;
1024 #endif
1025 set_task_cfs_rq(p);
1028 #ifdef CONFIG_SMP
1031 * Is this task likely cache-hot:
1033 static inline int
1034 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1036 s64 delta;
1038 if (p->sched_class != &fair_sched_class)
1039 return 0;
1041 if (sysctl_sched_migration_cost == -1)
1042 return 1;
1043 if (sysctl_sched_migration_cost == 0)
1044 return 0;
1046 delta = now - p->se.exec_start;
1048 return delta < (s64)sysctl_sched_migration_cost;
1052 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1054 int old_cpu = task_cpu(p);
1055 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1056 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1057 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1058 u64 clock_offset;
1060 clock_offset = old_rq->clock - new_rq->clock;
1062 #ifdef CONFIG_SCHEDSTATS
1063 if (p->se.wait_start)
1064 p->se.wait_start -= clock_offset;
1065 if (p->se.sleep_start)
1066 p->se.sleep_start -= clock_offset;
1067 if (p->se.block_start)
1068 p->se.block_start -= clock_offset;
1069 if (old_cpu != new_cpu) {
1070 schedstat_inc(p, se.nr_migrations);
1071 if (task_hot(p, old_rq->clock, NULL))
1072 schedstat_inc(p, se.nr_forced2_migrations);
1074 #endif
1075 p->se.vruntime -= old_cfsrq->min_vruntime -
1076 new_cfsrq->min_vruntime;
1078 __set_task_cpu(p, new_cpu);
1081 struct migration_req {
1082 struct list_head list;
1084 struct task_struct *task;
1085 int dest_cpu;
1087 struct completion done;
1091 * The task's runqueue lock must be held.
1092 * Returns true if you have to wait for migration thread.
1094 static int
1095 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1097 struct rq *rq = task_rq(p);
1100 * If the task is not on a runqueue (and not running), then
1101 * it is sufficient to simply update the task's cpu field.
1103 if (!p->se.on_rq && !task_running(rq, p)) {
1104 set_task_cpu(p, dest_cpu);
1105 return 0;
1108 init_completion(&req->done);
1109 req->task = p;
1110 req->dest_cpu = dest_cpu;
1111 list_add(&req->list, &rq->migration_queue);
1113 return 1;
1117 * wait_task_inactive - wait for a thread to unschedule.
1119 * The caller must ensure that the task *will* unschedule sometime soon,
1120 * else this function might spin for a *long* time. This function can't
1121 * be called with interrupts off, or it may introduce deadlock with
1122 * smp_call_function() if an IPI is sent by the same process we are
1123 * waiting to become inactive.
1125 void wait_task_inactive(struct task_struct *p)
1127 unsigned long flags;
1128 int running, on_rq;
1129 struct rq *rq;
1131 for (;;) {
1133 * We do the initial early heuristics without holding
1134 * any task-queue locks at all. We'll only try to get
1135 * the runqueue lock when things look like they will
1136 * work out!
1138 rq = task_rq(p);
1141 * If the task is actively running on another CPU
1142 * still, just relax and busy-wait without holding
1143 * any locks.
1145 * NOTE! Since we don't hold any locks, it's not
1146 * even sure that "rq" stays as the right runqueue!
1147 * But we don't care, since "task_running()" will
1148 * return false if the runqueue has changed and p
1149 * is actually now running somewhere else!
1151 while (task_running(rq, p))
1152 cpu_relax();
1155 * Ok, time to look more closely! We need the rq
1156 * lock now, to be *sure*. If we're wrong, we'll
1157 * just go back and repeat.
1159 rq = task_rq_lock(p, &flags);
1160 running = task_running(rq, p);
1161 on_rq = p->se.on_rq;
1162 task_rq_unlock(rq, &flags);
1165 * Was it really running after all now that we
1166 * checked with the proper locks actually held?
1168 * Oops. Go back and try again..
1170 if (unlikely(running)) {
1171 cpu_relax();
1172 continue;
1176 * It's not enough that it's not actively running,
1177 * it must be off the runqueue _entirely_, and not
1178 * preempted!
1180 * So if it wa still runnable (but just not actively
1181 * running right now), it's preempted, and we should
1182 * yield - it could be a while.
1184 if (unlikely(on_rq)) {
1185 schedule_timeout_uninterruptible(1);
1186 continue;
1190 * Ahh, all good. It wasn't running, and it wasn't
1191 * runnable, which means that it will never become
1192 * running in the future either. We're all done!
1194 break;
1198 /***
1199 * kick_process - kick a running thread to enter/exit the kernel
1200 * @p: the to-be-kicked thread
1202 * Cause a process which is running on another CPU to enter
1203 * kernel-mode, without any delay. (to get signals handled.)
1205 * NOTE: this function doesnt have to take the runqueue lock,
1206 * because all it wants to ensure is that the remote task enters
1207 * the kernel. If the IPI races and the task has been migrated
1208 * to another CPU then no harm is done and the purpose has been
1209 * achieved as well.
1211 void kick_process(struct task_struct *p)
1213 int cpu;
1215 preempt_disable();
1216 cpu = task_cpu(p);
1217 if ((cpu != smp_processor_id()) && task_curr(p))
1218 smp_send_reschedule(cpu);
1219 preempt_enable();
1223 * Return a low guess at the load of a migration-source cpu weighted
1224 * according to the scheduling class and "nice" value.
1226 * We want to under-estimate the load of migration sources, to
1227 * balance conservatively.
1229 static unsigned long source_load(int cpu, int type)
1231 struct rq *rq = cpu_rq(cpu);
1232 unsigned long total = weighted_cpuload(cpu);
1234 if (type == 0)
1235 return total;
1237 return min(rq->cpu_load[type-1], total);
1241 * Return a high guess at the load of a migration-target cpu weighted
1242 * according to the scheduling class and "nice" value.
1244 static unsigned long target_load(int cpu, int type)
1246 struct rq *rq = cpu_rq(cpu);
1247 unsigned long total = weighted_cpuload(cpu);
1249 if (type == 0)
1250 return total;
1252 return max(rq->cpu_load[type-1], total);
1256 * Return the average load per task on the cpu's run queue
1258 static inline unsigned long cpu_avg_load_per_task(int cpu)
1260 struct rq *rq = cpu_rq(cpu);
1261 unsigned long total = weighted_cpuload(cpu);
1262 unsigned long n = rq->nr_running;
1264 return n ? total / n : SCHED_LOAD_SCALE;
1268 * find_idlest_group finds and returns the least busy CPU group within the
1269 * domain.
1271 static struct sched_group *
1272 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1274 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1275 unsigned long min_load = ULONG_MAX, this_load = 0;
1276 int load_idx = sd->forkexec_idx;
1277 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1279 do {
1280 unsigned long load, avg_load;
1281 int local_group;
1282 int i;
1284 /* Skip over this group if it has no CPUs allowed */
1285 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1286 continue;
1288 local_group = cpu_isset(this_cpu, group->cpumask);
1290 /* Tally up the load of all CPUs in the group */
1291 avg_load = 0;
1293 for_each_cpu_mask(i, group->cpumask) {
1294 /* Bias balancing toward cpus of our domain */
1295 if (local_group)
1296 load = source_load(i, load_idx);
1297 else
1298 load = target_load(i, load_idx);
1300 avg_load += load;
1303 /* Adjust by relative CPU power of the group */
1304 avg_load = sg_div_cpu_power(group,
1305 avg_load * SCHED_LOAD_SCALE);
1307 if (local_group) {
1308 this_load = avg_load;
1309 this = group;
1310 } else if (avg_load < min_load) {
1311 min_load = avg_load;
1312 idlest = group;
1314 } while (group = group->next, group != sd->groups);
1316 if (!idlest || 100*this_load < imbalance*min_load)
1317 return NULL;
1318 return idlest;
1322 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1324 static int
1325 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1327 cpumask_t tmp;
1328 unsigned long load, min_load = ULONG_MAX;
1329 int idlest = -1;
1330 int i;
1332 /* Traverse only the allowed CPUs */
1333 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1335 for_each_cpu_mask(i, tmp) {
1336 load = weighted_cpuload(i);
1338 if (load < min_load || (load == min_load && i == this_cpu)) {
1339 min_load = load;
1340 idlest = i;
1344 return idlest;
1348 * sched_balance_self: balance the current task (running on cpu) in domains
1349 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1350 * SD_BALANCE_EXEC.
1352 * Balance, ie. select the least loaded group.
1354 * Returns the target CPU number, or the same CPU if no balancing is needed.
1356 * preempt must be disabled.
1358 static int sched_balance_self(int cpu, int flag)
1360 struct task_struct *t = current;
1361 struct sched_domain *tmp, *sd = NULL;
1363 for_each_domain(cpu, tmp) {
1365 * If power savings logic is enabled for a domain, stop there.
1367 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1368 break;
1369 if (tmp->flags & flag)
1370 sd = tmp;
1373 while (sd) {
1374 cpumask_t span;
1375 struct sched_group *group;
1376 int new_cpu, weight;
1378 if (!(sd->flags & flag)) {
1379 sd = sd->child;
1380 continue;
1383 span = sd->span;
1384 group = find_idlest_group(sd, t, cpu);
1385 if (!group) {
1386 sd = sd->child;
1387 continue;
1390 new_cpu = find_idlest_cpu(group, t, cpu);
1391 if (new_cpu == -1 || new_cpu == cpu) {
1392 /* Now try balancing at a lower domain level of cpu */
1393 sd = sd->child;
1394 continue;
1397 /* Now try balancing at a lower domain level of new_cpu */
1398 cpu = new_cpu;
1399 sd = NULL;
1400 weight = cpus_weight(span);
1401 for_each_domain(cpu, tmp) {
1402 if (weight <= cpus_weight(tmp->span))
1403 break;
1404 if (tmp->flags & flag)
1405 sd = tmp;
1407 /* while loop will break here if sd == NULL */
1410 return cpu;
1413 #endif /* CONFIG_SMP */
1416 * wake_idle() will wake a task on an idle cpu if task->cpu is
1417 * not idle and an idle cpu is available. The span of cpus to
1418 * search starts with cpus closest then further out as needed,
1419 * so we always favor a closer, idle cpu.
1421 * Returns the CPU we should wake onto.
1423 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1424 static int wake_idle(int cpu, struct task_struct *p)
1426 cpumask_t tmp;
1427 struct sched_domain *sd;
1428 int i;
1431 * If it is idle, then it is the best cpu to run this task.
1433 * This cpu is also the best, if it has more than one task already.
1434 * Siblings must be also busy(in most cases) as they didn't already
1435 * pickup the extra load from this cpu and hence we need not check
1436 * sibling runqueue info. This will avoid the checks and cache miss
1437 * penalities associated with that.
1439 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1440 return cpu;
1442 for_each_domain(cpu, sd) {
1443 if (sd->flags & SD_WAKE_IDLE) {
1444 cpus_and(tmp, sd->span, p->cpus_allowed);
1445 for_each_cpu_mask(i, tmp) {
1446 if (idle_cpu(i)) {
1447 if (i != task_cpu(p)) {
1448 schedstat_inc(p,
1449 se.nr_wakeups_idle);
1451 return i;
1454 } else {
1455 break;
1458 return cpu;
1460 #else
1461 static inline int wake_idle(int cpu, struct task_struct *p)
1463 return cpu;
1465 #endif
1467 /***
1468 * try_to_wake_up - wake up a thread
1469 * @p: the to-be-woken-up thread
1470 * @state: the mask of task states that can be woken
1471 * @sync: do a synchronous wakeup?
1473 * Put it on the run-queue if it's not already there. The "current"
1474 * thread is always on the run-queue (except when the actual
1475 * re-schedule is in progress), and as such you're allowed to do
1476 * the simpler "current->state = TASK_RUNNING" to mark yourself
1477 * runnable without the overhead of this.
1479 * returns failure only if the task is already active.
1481 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1483 int cpu, orig_cpu, this_cpu, success = 0;
1484 unsigned long flags;
1485 long old_state;
1486 struct rq *rq;
1487 #ifdef CONFIG_SMP
1488 struct sched_domain *sd, *this_sd = NULL;
1489 unsigned long load, this_load;
1490 int new_cpu;
1491 #endif
1493 rq = task_rq_lock(p, &flags);
1494 old_state = p->state;
1495 if (!(old_state & state))
1496 goto out;
1498 if (p->se.on_rq)
1499 goto out_running;
1501 cpu = task_cpu(p);
1502 orig_cpu = cpu;
1503 this_cpu = smp_processor_id();
1505 #ifdef CONFIG_SMP
1506 if (unlikely(task_running(rq, p)))
1507 goto out_activate;
1509 new_cpu = cpu;
1511 schedstat_inc(rq, ttwu_count);
1512 if (cpu == this_cpu) {
1513 schedstat_inc(rq, ttwu_local);
1514 goto out_set_cpu;
1517 for_each_domain(this_cpu, sd) {
1518 if (cpu_isset(cpu, sd->span)) {
1519 schedstat_inc(sd, ttwu_wake_remote);
1520 this_sd = sd;
1521 break;
1525 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1526 goto out_set_cpu;
1529 * Check for affine wakeup and passive balancing possibilities.
1531 if (this_sd) {
1532 int idx = this_sd->wake_idx;
1533 unsigned int imbalance;
1535 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1537 load = source_load(cpu, idx);
1538 this_load = target_load(this_cpu, idx);
1540 new_cpu = this_cpu; /* Wake to this CPU if we can */
1542 if (this_sd->flags & SD_WAKE_AFFINE) {
1543 unsigned long tl = this_load;
1544 unsigned long tl_per_task;
1547 * Attract cache-cold tasks on sync wakeups:
1549 if (sync && !task_hot(p, rq->clock, this_sd))
1550 goto out_set_cpu;
1552 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1553 tl_per_task = cpu_avg_load_per_task(this_cpu);
1556 * If sync wakeup then subtract the (maximum possible)
1557 * effect of the currently running task from the load
1558 * of the current CPU:
1560 if (sync)
1561 tl -= current->se.load.weight;
1563 if ((tl <= load &&
1564 tl + target_load(cpu, idx) <= tl_per_task) ||
1565 100*(tl + p->se.load.weight) <= imbalance*load) {
1567 * This domain has SD_WAKE_AFFINE and
1568 * p is cache cold in this domain, and
1569 * there is no bad imbalance.
1571 schedstat_inc(this_sd, ttwu_move_affine);
1572 schedstat_inc(p, se.nr_wakeups_affine);
1573 goto out_set_cpu;
1578 * Start passive balancing when half the imbalance_pct
1579 * limit is reached.
1581 if (this_sd->flags & SD_WAKE_BALANCE) {
1582 if (imbalance*this_load <= 100*load) {
1583 schedstat_inc(this_sd, ttwu_move_balance);
1584 schedstat_inc(p, se.nr_wakeups_passive);
1585 goto out_set_cpu;
1590 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1591 out_set_cpu:
1592 new_cpu = wake_idle(new_cpu, p);
1593 if (new_cpu != cpu) {
1594 set_task_cpu(p, new_cpu);
1595 task_rq_unlock(rq, &flags);
1596 /* might preempt at this point */
1597 rq = task_rq_lock(p, &flags);
1598 old_state = p->state;
1599 if (!(old_state & state))
1600 goto out;
1601 if (p->se.on_rq)
1602 goto out_running;
1604 this_cpu = smp_processor_id();
1605 cpu = task_cpu(p);
1608 out_activate:
1609 #endif /* CONFIG_SMP */
1610 schedstat_inc(p, se.nr_wakeups);
1611 if (sync)
1612 schedstat_inc(p, se.nr_wakeups_sync);
1613 if (orig_cpu != cpu)
1614 schedstat_inc(p, se.nr_wakeups_migrate);
1615 if (cpu == this_cpu)
1616 schedstat_inc(p, se.nr_wakeups_local);
1617 else
1618 schedstat_inc(p, se.nr_wakeups_remote);
1619 update_rq_clock(rq);
1620 activate_task(rq, p, 1);
1621 check_preempt_curr(rq, p);
1622 success = 1;
1624 out_running:
1625 p->state = TASK_RUNNING;
1626 out:
1627 task_rq_unlock(rq, &flags);
1629 return success;
1632 int fastcall wake_up_process(struct task_struct *p)
1634 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1635 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1637 EXPORT_SYMBOL(wake_up_process);
1639 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1641 return try_to_wake_up(p, state, 0);
1645 * Perform scheduler related setup for a newly forked process p.
1646 * p is forked by current.
1648 * __sched_fork() is basic setup used by init_idle() too:
1650 static void __sched_fork(struct task_struct *p)
1652 p->se.exec_start = 0;
1653 p->se.sum_exec_runtime = 0;
1654 p->se.prev_sum_exec_runtime = 0;
1656 #ifdef CONFIG_SCHEDSTATS
1657 p->se.wait_start = 0;
1658 p->se.sum_sleep_runtime = 0;
1659 p->se.sleep_start = 0;
1660 p->se.block_start = 0;
1661 p->se.sleep_max = 0;
1662 p->se.block_max = 0;
1663 p->se.exec_max = 0;
1664 p->se.slice_max = 0;
1665 p->se.wait_max = 0;
1666 #endif
1668 INIT_LIST_HEAD(&p->run_list);
1669 p->se.on_rq = 0;
1671 #ifdef CONFIG_PREEMPT_NOTIFIERS
1672 INIT_HLIST_HEAD(&p->preempt_notifiers);
1673 #endif
1676 * We mark the process as running here, but have not actually
1677 * inserted it onto the runqueue yet. This guarantees that
1678 * nobody will actually run it, and a signal or other external
1679 * event cannot wake it up and insert it on the runqueue either.
1681 p->state = TASK_RUNNING;
1685 * fork()/clone()-time setup:
1687 void sched_fork(struct task_struct *p, int clone_flags)
1689 int cpu = get_cpu();
1691 __sched_fork(p);
1693 #ifdef CONFIG_SMP
1694 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1695 #endif
1696 set_task_cpu(p, cpu);
1699 * Make sure we do not leak PI boosting priority to the child:
1701 p->prio = current->normal_prio;
1702 if (!rt_prio(p->prio))
1703 p->sched_class = &fair_sched_class;
1705 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1706 if (likely(sched_info_on()))
1707 memset(&p->sched_info, 0, sizeof(p->sched_info));
1708 #endif
1709 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1710 p->oncpu = 0;
1711 #endif
1712 #ifdef CONFIG_PREEMPT
1713 /* Want to start with kernel preemption disabled. */
1714 task_thread_info(p)->preempt_count = 1;
1715 #endif
1716 put_cpu();
1720 * wake_up_new_task - wake up a newly created task for the first time.
1722 * This function will do some initial scheduler statistics housekeeping
1723 * that must be done for every newly created context, then puts the task
1724 * on the runqueue and wakes it.
1726 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1728 unsigned long flags;
1729 struct rq *rq;
1731 rq = task_rq_lock(p, &flags);
1732 BUG_ON(p->state != TASK_RUNNING);
1733 update_rq_clock(rq);
1735 p->prio = effective_prio(p);
1737 if (!p->sched_class->task_new || !current->se.on_rq) {
1738 activate_task(rq, p, 0);
1739 } else {
1741 * Let the scheduling class do new task startup
1742 * management (if any):
1744 p->sched_class->task_new(rq, p);
1745 inc_nr_running(p, rq);
1747 check_preempt_curr(rq, p);
1748 task_rq_unlock(rq, &flags);
1751 #ifdef CONFIG_PREEMPT_NOTIFIERS
1754 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1755 * @notifier: notifier struct to register
1757 void preempt_notifier_register(struct preempt_notifier *notifier)
1759 hlist_add_head(&notifier->link, &current->preempt_notifiers);
1761 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1764 * preempt_notifier_unregister - no longer interested in preemption notifications
1765 * @notifier: notifier struct to unregister
1767 * This is safe to call from within a preemption notifier.
1769 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1771 hlist_del(&notifier->link);
1773 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1775 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1777 struct preempt_notifier *notifier;
1778 struct hlist_node *node;
1780 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1781 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1784 static void
1785 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1786 struct task_struct *next)
1788 struct preempt_notifier *notifier;
1789 struct hlist_node *node;
1791 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1792 notifier->ops->sched_out(notifier, next);
1795 #else
1797 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1801 static void
1802 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1803 struct task_struct *next)
1807 #endif
1810 * prepare_task_switch - prepare to switch tasks
1811 * @rq: the runqueue preparing to switch
1812 * @prev: the current task that is being switched out
1813 * @next: the task we are going to switch to.
1815 * This is called with the rq lock held and interrupts off. It must
1816 * be paired with a subsequent finish_task_switch after the context
1817 * switch.
1819 * prepare_task_switch sets up locking and calls architecture specific
1820 * hooks.
1822 static inline void
1823 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1824 struct task_struct *next)
1826 fire_sched_out_preempt_notifiers(prev, next);
1827 prepare_lock_switch(rq, next);
1828 prepare_arch_switch(next);
1832 * finish_task_switch - clean up after a task-switch
1833 * @rq: runqueue associated with task-switch
1834 * @prev: the thread we just switched away from.
1836 * finish_task_switch must be called after the context switch, paired
1837 * with a prepare_task_switch call before the context switch.
1838 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1839 * and do any other architecture-specific cleanup actions.
1841 * Note that we may have delayed dropping an mm in context_switch(). If
1842 * so, we finish that here outside of the runqueue lock. (Doing it
1843 * with the lock held can cause deadlocks; see schedule() for
1844 * details.)
1846 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1847 __releases(rq->lock)
1849 struct mm_struct *mm = rq->prev_mm;
1850 long prev_state;
1852 rq->prev_mm = NULL;
1855 * A task struct has one reference for the use as "current".
1856 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1857 * schedule one last time. The schedule call will never return, and
1858 * the scheduled task must drop that reference.
1859 * The test for TASK_DEAD must occur while the runqueue locks are
1860 * still held, otherwise prev could be scheduled on another cpu, die
1861 * there before we look at prev->state, and then the reference would
1862 * be dropped twice.
1863 * Manfred Spraul <manfred@colorfullife.com>
1865 prev_state = prev->state;
1866 finish_arch_switch(prev);
1867 finish_lock_switch(rq, prev);
1868 fire_sched_in_preempt_notifiers(current);
1869 if (mm)
1870 mmdrop(mm);
1871 if (unlikely(prev_state == TASK_DEAD)) {
1873 * Remove function-return probe instances associated with this
1874 * task and put them back on the free list.
1876 kprobe_flush_task(prev);
1877 put_task_struct(prev);
1882 * schedule_tail - first thing a freshly forked thread must call.
1883 * @prev: the thread we just switched away from.
1885 asmlinkage void schedule_tail(struct task_struct *prev)
1886 __releases(rq->lock)
1888 struct rq *rq = this_rq();
1890 finish_task_switch(rq, prev);
1891 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1892 /* In this case, finish_task_switch does not reenable preemption */
1893 preempt_enable();
1894 #endif
1895 if (current->set_child_tid)
1896 put_user(task_pid_vnr(current), current->set_child_tid);
1900 * context_switch - switch to the new MM and the new
1901 * thread's register state.
1903 static inline void
1904 context_switch(struct rq *rq, struct task_struct *prev,
1905 struct task_struct *next)
1907 struct mm_struct *mm, *oldmm;
1909 prepare_task_switch(rq, prev, next);
1910 mm = next->mm;
1911 oldmm = prev->active_mm;
1913 * For paravirt, this is coupled with an exit in switch_to to
1914 * combine the page table reload and the switch backend into
1915 * one hypercall.
1917 arch_enter_lazy_cpu_mode();
1919 if (unlikely(!mm)) {
1920 next->active_mm = oldmm;
1921 atomic_inc(&oldmm->mm_count);
1922 enter_lazy_tlb(oldmm, next);
1923 } else
1924 switch_mm(oldmm, mm, next);
1926 if (unlikely(!prev->mm)) {
1927 prev->active_mm = NULL;
1928 rq->prev_mm = oldmm;
1931 * Since the runqueue lock will be released by the next
1932 * task (which is an invalid locking op but in the case
1933 * of the scheduler it's an obvious special-case), so we
1934 * do an early lockdep release here:
1936 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1937 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1938 #endif
1940 /* Here we just switch the register state and the stack. */
1941 switch_to(prev, next, prev);
1943 barrier();
1945 * this_rq must be evaluated again because prev may have moved
1946 * CPUs since it called schedule(), thus the 'rq' on its stack
1947 * frame will be invalid.
1949 finish_task_switch(this_rq(), prev);
1953 * nr_running, nr_uninterruptible and nr_context_switches:
1955 * externally visible scheduler statistics: current number of runnable
1956 * threads, current number of uninterruptible-sleeping threads, total
1957 * number of context switches performed since bootup.
1959 unsigned long nr_running(void)
1961 unsigned long i, sum = 0;
1963 for_each_online_cpu(i)
1964 sum += cpu_rq(i)->nr_running;
1966 return sum;
1969 unsigned long nr_uninterruptible(void)
1971 unsigned long i, sum = 0;
1973 for_each_possible_cpu(i)
1974 sum += cpu_rq(i)->nr_uninterruptible;
1977 * Since we read the counters lockless, it might be slightly
1978 * inaccurate. Do not allow it to go below zero though:
1980 if (unlikely((long)sum < 0))
1981 sum = 0;
1983 return sum;
1986 unsigned long long nr_context_switches(void)
1988 int i;
1989 unsigned long long sum = 0;
1991 for_each_possible_cpu(i)
1992 sum += cpu_rq(i)->nr_switches;
1994 return sum;
1997 unsigned long nr_iowait(void)
1999 unsigned long i, sum = 0;
2001 for_each_possible_cpu(i)
2002 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2004 return sum;
2007 unsigned long nr_active(void)
2009 unsigned long i, running = 0, uninterruptible = 0;
2011 for_each_online_cpu(i) {
2012 running += cpu_rq(i)->nr_running;
2013 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2016 if (unlikely((long)uninterruptible < 0))
2017 uninterruptible = 0;
2019 return running + uninterruptible;
2023 * Update rq->cpu_load[] statistics. This function is usually called every
2024 * scheduler tick (TICK_NSEC).
2026 static void update_cpu_load(struct rq *this_rq)
2028 unsigned long this_load = this_rq->load.weight;
2029 int i, scale;
2031 this_rq->nr_load_updates++;
2033 /* Update our load: */
2034 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2035 unsigned long old_load, new_load;
2037 /* scale is effectively 1 << i now, and >> i divides by scale */
2039 old_load = this_rq->cpu_load[i];
2040 new_load = this_load;
2042 * Round up the averaging division if load is increasing. This
2043 * prevents us from getting stuck on 9 if the load is 10, for
2044 * example.
2046 if (new_load > old_load)
2047 new_load += scale-1;
2048 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2052 #ifdef CONFIG_SMP
2055 * double_rq_lock - safely lock two runqueues
2057 * Note this does not disable interrupts like task_rq_lock,
2058 * you need to do so manually before calling.
2060 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2061 __acquires(rq1->lock)
2062 __acquires(rq2->lock)
2064 BUG_ON(!irqs_disabled());
2065 if (rq1 == rq2) {
2066 spin_lock(&rq1->lock);
2067 __acquire(rq2->lock); /* Fake it out ;) */
2068 } else {
2069 if (rq1 < rq2) {
2070 spin_lock(&rq1->lock);
2071 spin_lock(&rq2->lock);
2072 } else {
2073 spin_lock(&rq2->lock);
2074 spin_lock(&rq1->lock);
2077 update_rq_clock(rq1);
2078 update_rq_clock(rq2);
2082 * double_rq_unlock - safely unlock two runqueues
2084 * Note this does not restore interrupts like task_rq_unlock,
2085 * you need to do so manually after calling.
2087 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2088 __releases(rq1->lock)
2089 __releases(rq2->lock)
2091 spin_unlock(&rq1->lock);
2092 if (rq1 != rq2)
2093 spin_unlock(&rq2->lock);
2094 else
2095 __release(rq2->lock);
2099 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2101 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2102 __releases(this_rq->lock)
2103 __acquires(busiest->lock)
2104 __acquires(this_rq->lock)
2106 if (unlikely(!irqs_disabled())) {
2107 /* printk() doesn't work good under rq->lock */
2108 spin_unlock(&this_rq->lock);
2109 BUG_ON(1);
2111 if (unlikely(!spin_trylock(&busiest->lock))) {
2112 if (busiest < this_rq) {
2113 spin_unlock(&this_rq->lock);
2114 spin_lock(&busiest->lock);
2115 spin_lock(&this_rq->lock);
2116 } else
2117 spin_lock(&busiest->lock);
2122 * If dest_cpu is allowed for this process, migrate the task to it.
2123 * This is accomplished by forcing the cpu_allowed mask to only
2124 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2125 * the cpu_allowed mask is restored.
2127 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2129 struct migration_req req;
2130 unsigned long flags;
2131 struct rq *rq;
2133 rq = task_rq_lock(p, &flags);
2134 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2135 || unlikely(cpu_is_offline(dest_cpu)))
2136 goto out;
2138 /* force the process onto the specified CPU */
2139 if (migrate_task(p, dest_cpu, &req)) {
2140 /* Need to wait for migration thread (might exit: take ref). */
2141 struct task_struct *mt = rq->migration_thread;
2143 get_task_struct(mt);
2144 task_rq_unlock(rq, &flags);
2145 wake_up_process(mt);
2146 put_task_struct(mt);
2147 wait_for_completion(&req.done);
2149 return;
2151 out:
2152 task_rq_unlock(rq, &flags);
2156 * sched_exec - execve() is a valuable balancing opportunity, because at
2157 * this point the task has the smallest effective memory and cache footprint.
2159 void sched_exec(void)
2161 int new_cpu, this_cpu = get_cpu();
2162 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2163 put_cpu();
2164 if (new_cpu != this_cpu)
2165 sched_migrate_task(current, new_cpu);
2169 * pull_task - move a task from a remote runqueue to the local runqueue.
2170 * Both runqueues must be locked.
2172 static void pull_task(struct rq *src_rq, struct task_struct *p,
2173 struct rq *this_rq, int this_cpu)
2175 deactivate_task(src_rq, p, 0);
2176 set_task_cpu(p, this_cpu);
2177 activate_task(this_rq, p, 0);
2179 * Note that idle threads have a prio of MAX_PRIO, for this test
2180 * to be always true for them.
2182 check_preempt_curr(this_rq, p);
2186 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2188 static
2189 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2190 struct sched_domain *sd, enum cpu_idle_type idle,
2191 int *all_pinned)
2194 * We do not migrate tasks that are:
2195 * 1) running (obviously), or
2196 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2197 * 3) are cache-hot on their current CPU.
2199 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2200 schedstat_inc(p, se.nr_failed_migrations_affine);
2201 return 0;
2203 *all_pinned = 0;
2205 if (task_running(rq, p)) {
2206 schedstat_inc(p, se.nr_failed_migrations_running);
2207 return 0;
2211 * Aggressive migration if:
2212 * 1) task is cache cold, or
2213 * 2) too many balance attempts have failed.
2216 if (!task_hot(p, rq->clock, sd) ||
2217 sd->nr_balance_failed > sd->cache_nice_tries) {
2218 #ifdef CONFIG_SCHEDSTATS
2219 if (task_hot(p, rq->clock, sd)) {
2220 schedstat_inc(sd, lb_hot_gained[idle]);
2221 schedstat_inc(p, se.nr_forced_migrations);
2223 #endif
2224 return 1;
2227 if (task_hot(p, rq->clock, sd)) {
2228 schedstat_inc(p, se.nr_failed_migrations_hot);
2229 return 0;
2231 return 1;
2234 static unsigned long
2235 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2236 unsigned long max_load_move, struct sched_domain *sd,
2237 enum cpu_idle_type idle, int *all_pinned,
2238 int *this_best_prio, struct rq_iterator *iterator)
2240 int pulled = 0, pinned = 0, skip_for_load;
2241 struct task_struct *p;
2242 long rem_load_move = max_load_move;
2244 if (max_load_move == 0)
2245 goto out;
2247 pinned = 1;
2250 * Start the load-balancing iterator:
2252 p = iterator->start(iterator->arg);
2253 next:
2254 if (!p)
2255 goto out;
2257 * To help distribute high priority tasks accross CPUs we don't
2258 * skip a task if it will be the highest priority task (i.e. smallest
2259 * prio value) on its new queue regardless of its load weight
2261 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2262 SCHED_LOAD_SCALE_FUZZ;
2263 if ((skip_for_load && p->prio >= *this_best_prio) ||
2264 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2265 p = iterator->next(iterator->arg);
2266 goto next;
2269 pull_task(busiest, p, this_rq, this_cpu);
2270 pulled++;
2271 rem_load_move -= p->se.load.weight;
2274 * We only want to steal up to the prescribed number of tasks
2275 * and the prescribed amount of weighted load.
2277 if (rem_load_move > 0) {
2278 if (p->prio < *this_best_prio)
2279 *this_best_prio = p->prio;
2280 p = iterator->next(iterator->arg);
2281 goto next;
2283 out:
2285 * Right now, this is one of only two places pull_task() is called,
2286 * so we can safely collect pull_task() stats here rather than
2287 * inside pull_task().
2289 schedstat_add(sd, lb_gained[idle], pulled);
2291 if (all_pinned)
2292 *all_pinned = pinned;
2294 return max_load_move - rem_load_move;
2298 * move_tasks tries to move up to max_load_move weighted load from busiest to
2299 * this_rq, as part of a balancing operation within domain "sd".
2300 * Returns 1 if successful and 0 otherwise.
2302 * Called with both runqueues locked.
2304 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2305 unsigned long max_load_move,
2306 struct sched_domain *sd, enum cpu_idle_type idle,
2307 int *all_pinned)
2309 const struct sched_class *class = sched_class_highest;
2310 unsigned long total_load_moved = 0;
2311 int this_best_prio = this_rq->curr->prio;
2313 do {
2314 total_load_moved +=
2315 class->load_balance(this_rq, this_cpu, busiest,
2316 max_load_move - total_load_moved,
2317 sd, idle, all_pinned, &this_best_prio);
2318 class = class->next;
2319 } while (class && max_load_move > total_load_moved);
2321 return total_load_moved > 0;
2324 static int
2325 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2326 struct sched_domain *sd, enum cpu_idle_type idle,
2327 struct rq_iterator *iterator)
2329 struct task_struct *p = iterator->start(iterator->arg);
2330 int pinned = 0;
2332 while (p) {
2333 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2334 pull_task(busiest, p, this_rq, this_cpu);
2336 * Right now, this is only the second place pull_task()
2337 * is called, so we can safely collect pull_task()
2338 * stats here rather than inside pull_task().
2340 schedstat_inc(sd, lb_gained[idle]);
2342 return 1;
2344 p = iterator->next(iterator->arg);
2347 return 0;
2351 * move_one_task tries to move exactly one task from busiest to this_rq, as
2352 * part of active balancing operations within "domain".
2353 * Returns 1 if successful and 0 otherwise.
2355 * Called with both runqueues locked.
2357 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2358 struct sched_domain *sd, enum cpu_idle_type idle)
2360 const struct sched_class *class;
2362 for (class = sched_class_highest; class; class = class->next)
2363 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2364 return 1;
2366 return 0;
2370 * find_busiest_group finds and returns the busiest CPU group within the
2371 * domain. It calculates and returns the amount of weighted load which
2372 * should be moved to restore balance via the imbalance parameter.
2374 static struct sched_group *
2375 find_busiest_group(struct sched_domain *sd, int this_cpu,
2376 unsigned long *imbalance, enum cpu_idle_type idle,
2377 int *sd_idle, cpumask_t *cpus, int *balance)
2379 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2380 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2381 unsigned long max_pull;
2382 unsigned long busiest_load_per_task, busiest_nr_running;
2383 unsigned long this_load_per_task, this_nr_running;
2384 int load_idx, group_imb = 0;
2385 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2386 int power_savings_balance = 1;
2387 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2388 unsigned long min_nr_running = ULONG_MAX;
2389 struct sched_group *group_min = NULL, *group_leader = NULL;
2390 #endif
2392 max_load = this_load = total_load = total_pwr = 0;
2393 busiest_load_per_task = busiest_nr_running = 0;
2394 this_load_per_task = this_nr_running = 0;
2395 if (idle == CPU_NOT_IDLE)
2396 load_idx = sd->busy_idx;
2397 else if (idle == CPU_NEWLY_IDLE)
2398 load_idx = sd->newidle_idx;
2399 else
2400 load_idx = sd->idle_idx;
2402 do {
2403 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2404 int local_group;
2405 int i;
2406 int __group_imb = 0;
2407 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2408 unsigned long sum_nr_running, sum_weighted_load;
2410 local_group = cpu_isset(this_cpu, group->cpumask);
2412 if (local_group)
2413 balance_cpu = first_cpu(group->cpumask);
2415 /* Tally up the load of all CPUs in the group */
2416 sum_weighted_load = sum_nr_running = avg_load = 0;
2417 max_cpu_load = 0;
2418 min_cpu_load = ~0UL;
2420 for_each_cpu_mask(i, group->cpumask) {
2421 struct rq *rq;
2423 if (!cpu_isset(i, *cpus))
2424 continue;
2426 rq = cpu_rq(i);
2428 if (*sd_idle && rq->nr_running)
2429 *sd_idle = 0;
2431 /* Bias balancing toward cpus of our domain */
2432 if (local_group) {
2433 if (idle_cpu(i) && !first_idle_cpu) {
2434 first_idle_cpu = 1;
2435 balance_cpu = i;
2438 load = target_load(i, load_idx);
2439 } else {
2440 load = source_load(i, load_idx);
2441 if (load > max_cpu_load)
2442 max_cpu_load = load;
2443 if (min_cpu_load > load)
2444 min_cpu_load = load;
2447 avg_load += load;
2448 sum_nr_running += rq->nr_running;
2449 sum_weighted_load += weighted_cpuload(i);
2453 * First idle cpu or the first cpu(busiest) in this sched group
2454 * is eligible for doing load balancing at this and above
2455 * domains. In the newly idle case, we will allow all the cpu's
2456 * to do the newly idle load balance.
2458 if (idle != CPU_NEWLY_IDLE && local_group &&
2459 balance_cpu != this_cpu && balance) {
2460 *balance = 0;
2461 goto ret;
2464 total_load += avg_load;
2465 total_pwr += group->__cpu_power;
2467 /* Adjust by relative CPU power of the group */
2468 avg_load = sg_div_cpu_power(group,
2469 avg_load * SCHED_LOAD_SCALE);
2471 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2472 __group_imb = 1;
2474 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2476 if (local_group) {
2477 this_load = avg_load;
2478 this = group;
2479 this_nr_running = sum_nr_running;
2480 this_load_per_task = sum_weighted_load;
2481 } else if (avg_load > max_load &&
2482 (sum_nr_running > group_capacity || __group_imb)) {
2483 max_load = avg_load;
2484 busiest = group;
2485 busiest_nr_running = sum_nr_running;
2486 busiest_load_per_task = sum_weighted_load;
2487 group_imb = __group_imb;
2490 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2492 * Busy processors will not participate in power savings
2493 * balance.
2495 if (idle == CPU_NOT_IDLE ||
2496 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2497 goto group_next;
2500 * If the local group is idle or completely loaded
2501 * no need to do power savings balance at this domain
2503 if (local_group && (this_nr_running >= group_capacity ||
2504 !this_nr_running))
2505 power_savings_balance = 0;
2508 * If a group is already running at full capacity or idle,
2509 * don't include that group in power savings calculations
2511 if (!power_savings_balance || sum_nr_running >= group_capacity
2512 || !sum_nr_running)
2513 goto group_next;
2516 * Calculate the group which has the least non-idle load.
2517 * This is the group from where we need to pick up the load
2518 * for saving power
2520 if ((sum_nr_running < min_nr_running) ||
2521 (sum_nr_running == min_nr_running &&
2522 first_cpu(group->cpumask) <
2523 first_cpu(group_min->cpumask))) {
2524 group_min = group;
2525 min_nr_running = sum_nr_running;
2526 min_load_per_task = sum_weighted_load /
2527 sum_nr_running;
2531 * Calculate the group which is almost near its
2532 * capacity but still has some space to pick up some load
2533 * from other group and save more power
2535 if (sum_nr_running <= group_capacity - 1) {
2536 if (sum_nr_running > leader_nr_running ||
2537 (sum_nr_running == leader_nr_running &&
2538 first_cpu(group->cpumask) >
2539 first_cpu(group_leader->cpumask))) {
2540 group_leader = group;
2541 leader_nr_running = sum_nr_running;
2544 group_next:
2545 #endif
2546 group = group->next;
2547 } while (group != sd->groups);
2549 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2550 goto out_balanced;
2552 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2554 if (this_load >= avg_load ||
2555 100*max_load <= sd->imbalance_pct*this_load)
2556 goto out_balanced;
2558 busiest_load_per_task /= busiest_nr_running;
2559 if (group_imb)
2560 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2563 * We're trying to get all the cpus to the average_load, so we don't
2564 * want to push ourselves above the average load, nor do we wish to
2565 * reduce the max loaded cpu below the average load, as either of these
2566 * actions would just result in more rebalancing later, and ping-pong
2567 * tasks around. Thus we look for the minimum possible imbalance.
2568 * Negative imbalances (*we* are more loaded than anyone else) will
2569 * be counted as no imbalance for these purposes -- we can't fix that
2570 * by pulling tasks to us. Be careful of negative numbers as they'll
2571 * appear as very large values with unsigned longs.
2573 if (max_load <= busiest_load_per_task)
2574 goto out_balanced;
2577 * In the presence of smp nice balancing, certain scenarios can have
2578 * max load less than avg load(as we skip the groups at or below
2579 * its cpu_power, while calculating max_load..)
2581 if (max_load < avg_load) {
2582 *imbalance = 0;
2583 goto small_imbalance;
2586 /* Don't want to pull so many tasks that a group would go idle */
2587 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2589 /* How much load to actually move to equalise the imbalance */
2590 *imbalance = min(max_pull * busiest->__cpu_power,
2591 (avg_load - this_load) * this->__cpu_power)
2592 / SCHED_LOAD_SCALE;
2595 * if *imbalance is less than the average load per runnable task
2596 * there is no gaurantee that any tasks will be moved so we'll have
2597 * a think about bumping its value to force at least one task to be
2598 * moved
2600 if (*imbalance < busiest_load_per_task) {
2601 unsigned long tmp, pwr_now, pwr_move;
2602 unsigned int imbn;
2604 small_imbalance:
2605 pwr_move = pwr_now = 0;
2606 imbn = 2;
2607 if (this_nr_running) {
2608 this_load_per_task /= this_nr_running;
2609 if (busiest_load_per_task > this_load_per_task)
2610 imbn = 1;
2611 } else
2612 this_load_per_task = SCHED_LOAD_SCALE;
2614 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2615 busiest_load_per_task * imbn) {
2616 *imbalance = busiest_load_per_task;
2617 return busiest;
2621 * OK, we don't have enough imbalance to justify moving tasks,
2622 * however we may be able to increase total CPU power used by
2623 * moving them.
2626 pwr_now += busiest->__cpu_power *
2627 min(busiest_load_per_task, max_load);
2628 pwr_now += this->__cpu_power *
2629 min(this_load_per_task, this_load);
2630 pwr_now /= SCHED_LOAD_SCALE;
2632 /* Amount of load we'd subtract */
2633 tmp = sg_div_cpu_power(busiest,
2634 busiest_load_per_task * SCHED_LOAD_SCALE);
2635 if (max_load > tmp)
2636 pwr_move += busiest->__cpu_power *
2637 min(busiest_load_per_task, max_load - tmp);
2639 /* Amount of load we'd add */
2640 if (max_load * busiest->__cpu_power <
2641 busiest_load_per_task * SCHED_LOAD_SCALE)
2642 tmp = sg_div_cpu_power(this,
2643 max_load * busiest->__cpu_power);
2644 else
2645 tmp = sg_div_cpu_power(this,
2646 busiest_load_per_task * SCHED_LOAD_SCALE);
2647 pwr_move += this->__cpu_power *
2648 min(this_load_per_task, this_load + tmp);
2649 pwr_move /= SCHED_LOAD_SCALE;
2651 /* Move if we gain throughput */
2652 if (pwr_move > pwr_now)
2653 *imbalance = busiest_load_per_task;
2656 return busiest;
2658 out_balanced:
2659 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2660 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2661 goto ret;
2663 if (this == group_leader && group_leader != group_min) {
2664 *imbalance = min_load_per_task;
2665 return group_min;
2667 #endif
2668 ret:
2669 *imbalance = 0;
2670 return NULL;
2674 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2676 static struct rq *
2677 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2678 unsigned long imbalance, cpumask_t *cpus)
2680 struct rq *busiest = NULL, *rq;
2681 unsigned long max_load = 0;
2682 int i;
2684 for_each_cpu_mask(i, group->cpumask) {
2685 unsigned long wl;
2687 if (!cpu_isset(i, *cpus))
2688 continue;
2690 rq = cpu_rq(i);
2691 wl = weighted_cpuload(i);
2693 if (rq->nr_running == 1 && wl > imbalance)
2694 continue;
2696 if (wl > max_load) {
2697 max_load = wl;
2698 busiest = rq;
2702 return busiest;
2706 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2707 * so long as it is large enough.
2709 #define MAX_PINNED_INTERVAL 512
2712 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2713 * tasks if there is an imbalance.
2715 static int load_balance(int this_cpu, struct rq *this_rq,
2716 struct sched_domain *sd, enum cpu_idle_type idle,
2717 int *balance)
2719 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2720 struct sched_group *group;
2721 unsigned long imbalance;
2722 struct rq *busiest;
2723 cpumask_t cpus = CPU_MASK_ALL;
2724 unsigned long flags;
2727 * When power savings policy is enabled for the parent domain, idle
2728 * sibling can pick up load irrespective of busy siblings. In this case,
2729 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2730 * portraying it as CPU_NOT_IDLE.
2732 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2733 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2734 sd_idle = 1;
2736 schedstat_inc(sd, lb_count[idle]);
2738 redo:
2739 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2740 &cpus, balance);
2742 if (*balance == 0)
2743 goto out_balanced;
2745 if (!group) {
2746 schedstat_inc(sd, lb_nobusyg[idle]);
2747 goto out_balanced;
2750 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2751 if (!busiest) {
2752 schedstat_inc(sd, lb_nobusyq[idle]);
2753 goto out_balanced;
2756 BUG_ON(busiest == this_rq);
2758 schedstat_add(sd, lb_imbalance[idle], imbalance);
2760 ld_moved = 0;
2761 if (busiest->nr_running > 1) {
2763 * Attempt to move tasks. If find_busiest_group has found
2764 * an imbalance but busiest->nr_running <= 1, the group is
2765 * still unbalanced. ld_moved simply stays zero, so it is
2766 * correctly treated as an imbalance.
2768 local_irq_save(flags);
2769 double_rq_lock(this_rq, busiest);
2770 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2771 imbalance, sd, idle, &all_pinned);
2772 double_rq_unlock(this_rq, busiest);
2773 local_irq_restore(flags);
2776 * some other cpu did the load balance for us.
2778 if (ld_moved && this_cpu != smp_processor_id())
2779 resched_cpu(this_cpu);
2781 /* All tasks on this runqueue were pinned by CPU affinity */
2782 if (unlikely(all_pinned)) {
2783 cpu_clear(cpu_of(busiest), cpus);
2784 if (!cpus_empty(cpus))
2785 goto redo;
2786 goto out_balanced;
2790 if (!ld_moved) {
2791 schedstat_inc(sd, lb_failed[idle]);
2792 sd->nr_balance_failed++;
2794 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2796 spin_lock_irqsave(&busiest->lock, flags);
2798 /* don't kick the migration_thread, if the curr
2799 * task on busiest cpu can't be moved to this_cpu
2801 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2802 spin_unlock_irqrestore(&busiest->lock, flags);
2803 all_pinned = 1;
2804 goto out_one_pinned;
2807 if (!busiest->active_balance) {
2808 busiest->active_balance = 1;
2809 busiest->push_cpu = this_cpu;
2810 active_balance = 1;
2812 spin_unlock_irqrestore(&busiest->lock, flags);
2813 if (active_balance)
2814 wake_up_process(busiest->migration_thread);
2817 * We've kicked active balancing, reset the failure
2818 * counter.
2820 sd->nr_balance_failed = sd->cache_nice_tries+1;
2822 } else
2823 sd->nr_balance_failed = 0;
2825 if (likely(!active_balance)) {
2826 /* We were unbalanced, so reset the balancing interval */
2827 sd->balance_interval = sd->min_interval;
2828 } else {
2830 * If we've begun active balancing, start to back off. This
2831 * case may not be covered by the all_pinned logic if there
2832 * is only 1 task on the busy runqueue (because we don't call
2833 * move_tasks).
2835 if (sd->balance_interval < sd->max_interval)
2836 sd->balance_interval *= 2;
2839 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2840 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2841 return -1;
2842 return ld_moved;
2844 out_balanced:
2845 schedstat_inc(sd, lb_balanced[idle]);
2847 sd->nr_balance_failed = 0;
2849 out_one_pinned:
2850 /* tune up the balancing interval */
2851 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2852 (sd->balance_interval < sd->max_interval))
2853 sd->balance_interval *= 2;
2855 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2856 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2857 return -1;
2858 return 0;
2862 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2863 * tasks if there is an imbalance.
2865 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2866 * this_rq is locked.
2868 static int
2869 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2871 struct sched_group *group;
2872 struct rq *busiest = NULL;
2873 unsigned long imbalance;
2874 int ld_moved = 0;
2875 int sd_idle = 0;
2876 int all_pinned = 0;
2877 cpumask_t cpus = CPU_MASK_ALL;
2880 * When power savings policy is enabled for the parent domain, idle
2881 * sibling can pick up load irrespective of busy siblings. In this case,
2882 * let the state of idle sibling percolate up as IDLE, instead of
2883 * portraying it as CPU_NOT_IDLE.
2885 if (sd->flags & SD_SHARE_CPUPOWER &&
2886 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2887 sd_idle = 1;
2889 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2890 redo:
2891 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2892 &sd_idle, &cpus, NULL);
2893 if (!group) {
2894 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2895 goto out_balanced;
2898 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2899 &cpus);
2900 if (!busiest) {
2901 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2902 goto out_balanced;
2905 BUG_ON(busiest == this_rq);
2907 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2909 ld_moved = 0;
2910 if (busiest->nr_running > 1) {
2911 /* Attempt to move tasks */
2912 double_lock_balance(this_rq, busiest);
2913 /* this_rq->clock is already updated */
2914 update_rq_clock(busiest);
2915 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2916 imbalance, sd, CPU_NEWLY_IDLE,
2917 &all_pinned);
2918 spin_unlock(&busiest->lock);
2920 if (unlikely(all_pinned)) {
2921 cpu_clear(cpu_of(busiest), cpus);
2922 if (!cpus_empty(cpus))
2923 goto redo;
2927 if (!ld_moved) {
2928 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2929 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2930 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2931 return -1;
2932 } else
2933 sd->nr_balance_failed = 0;
2935 return ld_moved;
2937 out_balanced:
2938 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2939 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2940 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2941 return -1;
2942 sd->nr_balance_failed = 0;
2944 return 0;
2948 * idle_balance is called by schedule() if this_cpu is about to become
2949 * idle. Attempts to pull tasks from other CPUs.
2951 static void idle_balance(int this_cpu, struct rq *this_rq)
2953 struct sched_domain *sd;
2954 int pulled_task = -1;
2955 unsigned long next_balance = jiffies + HZ;
2957 for_each_domain(this_cpu, sd) {
2958 unsigned long interval;
2960 if (!(sd->flags & SD_LOAD_BALANCE))
2961 continue;
2963 if (sd->flags & SD_BALANCE_NEWIDLE)
2964 /* If we've pulled tasks over stop searching: */
2965 pulled_task = load_balance_newidle(this_cpu,
2966 this_rq, sd);
2968 interval = msecs_to_jiffies(sd->balance_interval);
2969 if (time_after(next_balance, sd->last_balance + interval))
2970 next_balance = sd->last_balance + interval;
2971 if (pulled_task)
2972 break;
2974 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2976 * We are going idle. next_balance may be set based on
2977 * a busy processor. So reset next_balance.
2979 this_rq->next_balance = next_balance;
2984 * active_load_balance is run by migration threads. It pushes running tasks
2985 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2986 * running on each physical CPU where possible, and avoids physical /
2987 * logical imbalances.
2989 * Called with busiest_rq locked.
2991 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2993 int target_cpu = busiest_rq->push_cpu;
2994 struct sched_domain *sd;
2995 struct rq *target_rq;
2997 /* Is there any task to move? */
2998 if (busiest_rq->nr_running <= 1)
2999 return;
3001 target_rq = cpu_rq(target_cpu);
3004 * This condition is "impossible", if it occurs
3005 * we need to fix it. Originally reported by
3006 * Bjorn Helgaas on a 128-cpu setup.
3008 BUG_ON(busiest_rq == target_rq);
3010 /* move a task from busiest_rq to target_rq */
3011 double_lock_balance(busiest_rq, target_rq);
3012 update_rq_clock(busiest_rq);
3013 update_rq_clock(target_rq);
3015 /* Search for an sd spanning us and the target CPU. */
3016 for_each_domain(target_cpu, sd) {
3017 if ((sd->flags & SD_LOAD_BALANCE) &&
3018 cpu_isset(busiest_cpu, sd->span))
3019 break;
3022 if (likely(sd)) {
3023 schedstat_inc(sd, alb_count);
3025 if (move_one_task(target_rq, target_cpu, busiest_rq,
3026 sd, CPU_IDLE))
3027 schedstat_inc(sd, alb_pushed);
3028 else
3029 schedstat_inc(sd, alb_failed);
3031 spin_unlock(&target_rq->lock);
3034 #ifdef CONFIG_NO_HZ
3035 static struct {
3036 atomic_t load_balancer;
3037 cpumask_t cpu_mask;
3038 } nohz ____cacheline_aligned = {
3039 .load_balancer = ATOMIC_INIT(-1),
3040 .cpu_mask = CPU_MASK_NONE,
3044 * This routine will try to nominate the ilb (idle load balancing)
3045 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3046 * load balancing on behalf of all those cpus. If all the cpus in the system
3047 * go into this tickless mode, then there will be no ilb owner (as there is
3048 * no need for one) and all the cpus will sleep till the next wakeup event
3049 * arrives...
3051 * For the ilb owner, tick is not stopped. And this tick will be used
3052 * for idle load balancing. ilb owner will still be part of
3053 * nohz.cpu_mask..
3055 * While stopping the tick, this cpu will become the ilb owner if there
3056 * is no other owner. And will be the owner till that cpu becomes busy
3057 * or if all cpus in the system stop their ticks at which point
3058 * there is no need for ilb owner.
3060 * When the ilb owner becomes busy, it nominates another owner, during the
3061 * next busy scheduler_tick()
3063 int select_nohz_load_balancer(int stop_tick)
3065 int cpu = smp_processor_id();
3067 if (stop_tick) {
3068 cpu_set(cpu, nohz.cpu_mask);
3069 cpu_rq(cpu)->in_nohz_recently = 1;
3072 * If we are going offline and still the leader, give up!
3074 if (cpu_is_offline(cpu) &&
3075 atomic_read(&nohz.load_balancer) == cpu) {
3076 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3077 BUG();
3078 return 0;
3081 /* time for ilb owner also to sleep */
3082 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3083 if (atomic_read(&nohz.load_balancer) == cpu)
3084 atomic_set(&nohz.load_balancer, -1);
3085 return 0;
3088 if (atomic_read(&nohz.load_balancer) == -1) {
3089 /* make me the ilb owner */
3090 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3091 return 1;
3092 } else if (atomic_read(&nohz.load_balancer) == cpu)
3093 return 1;
3094 } else {
3095 if (!cpu_isset(cpu, nohz.cpu_mask))
3096 return 0;
3098 cpu_clear(cpu, nohz.cpu_mask);
3100 if (atomic_read(&nohz.load_balancer) == cpu)
3101 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3102 BUG();
3104 return 0;
3106 #endif
3108 static DEFINE_SPINLOCK(balancing);
3111 * It checks each scheduling domain to see if it is due to be balanced,
3112 * and initiates a balancing operation if so.
3114 * Balancing parameters are set up in arch_init_sched_domains.
3116 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3118 int balance = 1;
3119 struct rq *rq = cpu_rq(cpu);
3120 unsigned long interval;
3121 struct sched_domain *sd;
3122 /* Earliest time when we have to do rebalance again */
3123 unsigned long next_balance = jiffies + 60*HZ;
3124 int update_next_balance = 0;
3126 for_each_domain(cpu, sd) {
3127 if (!(sd->flags & SD_LOAD_BALANCE))
3128 continue;
3130 interval = sd->balance_interval;
3131 if (idle != CPU_IDLE)
3132 interval *= sd->busy_factor;
3134 /* scale ms to jiffies */
3135 interval = msecs_to_jiffies(interval);
3136 if (unlikely(!interval))
3137 interval = 1;
3138 if (interval > HZ*NR_CPUS/10)
3139 interval = HZ*NR_CPUS/10;
3142 if (sd->flags & SD_SERIALIZE) {
3143 if (!spin_trylock(&balancing))
3144 goto out;
3147 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3148 if (load_balance(cpu, rq, sd, idle, &balance)) {
3150 * We've pulled tasks over so either we're no
3151 * longer idle, or one of our SMT siblings is
3152 * not idle.
3154 idle = CPU_NOT_IDLE;
3156 sd->last_balance = jiffies;
3158 if (sd->flags & SD_SERIALIZE)
3159 spin_unlock(&balancing);
3160 out:
3161 if (time_after(next_balance, sd->last_balance + interval)) {
3162 next_balance = sd->last_balance + interval;
3163 update_next_balance = 1;
3167 * Stop the load balance at this level. There is another
3168 * CPU in our sched group which is doing load balancing more
3169 * actively.
3171 if (!balance)
3172 break;
3176 * next_balance will be updated only when there is a need.
3177 * When the cpu is attached to null domain for ex, it will not be
3178 * updated.
3180 if (likely(update_next_balance))
3181 rq->next_balance = next_balance;
3185 * run_rebalance_domains is triggered when needed from the scheduler tick.
3186 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3187 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3189 static void run_rebalance_domains(struct softirq_action *h)
3191 int this_cpu = smp_processor_id();
3192 struct rq *this_rq = cpu_rq(this_cpu);
3193 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3194 CPU_IDLE : CPU_NOT_IDLE;
3196 rebalance_domains(this_cpu, idle);
3198 #ifdef CONFIG_NO_HZ
3200 * If this cpu is the owner for idle load balancing, then do the
3201 * balancing on behalf of the other idle cpus whose ticks are
3202 * stopped.
3204 if (this_rq->idle_at_tick &&
3205 atomic_read(&nohz.load_balancer) == this_cpu) {
3206 cpumask_t cpus = nohz.cpu_mask;
3207 struct rq *rq;
3208 int balance_cpu;
3210 cpu_clear(this_cpu, cpus);
3211 for_each_cpu_mask(balance_cpu, cpus) {
3213 * If this cpu gets work to do, stop the load balancing
3214 * work being done for other cpus. Next load
3215 * balancing owner will pick it up.
3217 if (need_resched())
3218 break;
3220 rebalance_domains(balance_cpu, CPU_IDLE);
3222 rq = cpu_rq(balance_cpu);
3223 if (time_after(this_rq->next_balance, rq->next_balance))
3224 this_rq->next_balance = rq->next_balance;
3227 #endif
3231 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3233 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3234 * idle load balancing owner or decide to stop the periodic load balancing,
3235 * if the whole system is idle.
3237 static inline void trigger_load_balance(struct rq *rq, int cpu)
3239 #ifdef CONFIG_NO_HZ
3241 * If we were in the nohz mode recently and busy at the current
3242 * scheduler tick, then check if we need to nominate new idle
3243 * load balancer.
3245 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3246 rq->in_nohz_recently = 0;
3248 if (atomic_read(&nohz.load_balancer) == cpu) {
3249 cpu_clear(cpu, nohz.cpu_mask);
3250 atomic_set(&nohz.load_balancer, -1);
3253 if (atomic_read(&nohz.load_balancer) == -1) {
3255 * simple selection for now: Nominate the
3256 * first cpu in the nohz list to be the next
3257 * ilb owner.
3259 * TBD: Traverse the sched domains and nominate
3260 * the nearest cpu in the nohz.cpu_mask.
3262 int ilb = first_cpu(nohz.cpu_mask);
3264 if (ilb != NR_CPUS)
3265 resched_cpu(ilb);
3270 * If this cpu is idle and doing idle load balancing for all the
3271 * cpus with ticks stopped, is it time for that to stop?
3273 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3274 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3275 resched_cpu(cpu);
3276 return;
3280 * If this cpu is idle and the idle load balancing is done by
3281 * someone else, then no need raise the SCHED_SOFTIRQ
3283 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3284 cpu_isset(cpu, nohz.cpu_mask))
3285 return;
3286 #endif
3287 if (time_after_eq(jiffies, rq->next_balance))
3288 raise_softirq(SCHED_SOFTIRQ);
3291 #else /* CONFIG_SMP */
3294 * on UP we do not need to balance between CPUs:
3296 static inline void idle_balance(int cpu, struct rq *rq)
3300 #endif
3302 DEFINE_PER_CPU(struct kernel_stat, kstat);
3304 EXPORT_PER_CPU_SYMBOL(kstat);
3307 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3308 * that have not yet been banked in case the task is currently running.
3310 unsigned long long task_sched_runtime(struct task_struct *p)
3312 unsigned long flags;
3313 u64 ns, delta_exec;
3314 struct rq *rq;
3316 rq = task_rq_lock(p, &flags);
3317 ns = p->se.sum_exec_runtime;
3318 if (rq->curr == p) {
3319 update_rq_clock(rq);
3320 delta_exec = rq->clock - p->se.exec_start;
3321 if ((s64)delta_exec > 0)
3322 ns += delta_exec;
3324 task_rq_unlock(rq, &flags);
3326 return ns;
3330 * Account user cpu time to a process.
3331 * @p: the process that the cpu time gets accounted to
3332 * @cputime: the cpu time spent in user space since the last update
3334 void account_user_time(struct task_struct *p, cputime_t cputime)
3336 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3337 cputime64_t tmp;
3338 struct rq *rq = this_rq();
3340 p->utime = cputime_add(p->utime, cputime);
3342 if (p != rq->idle)
3343 cpuacct_charge(p, cputime);
3345 /* Add user time to cpustat. */
3346 tmp = cputime_to_cputime64(cputime);
3347 if (TASK_NICE(p) > 0)
3348 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3349 else
3350 cpustat->user = cputime64_add(cpustat->user, tmp);
3354 * Account guest cpu time to a process.
3355 * @p: the process that the cpu time gets accounted to
3356 * @cputime: the cpu time spent in virtual machine since the last update
3358 void account_guest_time(struct task_struct *p, cputime_t cputime)
3360 cputime64_t tmp;
3361 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3363 tmp = cputime_to_cputime64(cputime);
3365 p->utime = cputime_add(p->utime, cputime);
3366 p->gtime = cputime_add(p->gtime, cputime);
3368 cpustat->user = cputime64_add(cpustat->user, tmp);
3369 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3373 * Account scaled user cpu time to a process.
3374 * @p: the process that the cpu time gets accounted to
3375 * @cputime: the cpu time spent in user space since the last update
3377 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3379 p->utimescaled = cputime_add(p->utimescaled, cputime);
3383 * Account system cpu time to a process.
3384 * @p: the process that the cpu time gets accounted to
3385 * @hardirq_offset: the offset to subtract from hardirq_count()
3386 * @cputime: the cpu time spent in kernel space since the last update
3388 void account_system_time(struct task_struct *p, int hardirq_offset,
3389 cputime_t cputime)
3391 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3392 struct rq *rq = this_rq();
3393 cputime64_t tmp;
3395 if (p->flags & PF_VCPU) {
3396 account_guest_time(p, cputime);
3397 return;
3400 p->stime = cputime_add(p->stime, cputime);
3402 /* Add system time to cpustat. */
3403 tmp = cputime_to_cputime64(cputime);
3404 if (hardirq_count() - hardirq_offset)
3405 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3406 else if (softirq_count())
3407 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3408 else if (p != rq->idle) {
3409 cpustat->system = cputime64_add(cpustat->system, tmp);
3410 cpuacct_charge(p, cputime);
3411 } else if (atomic_read(&rq->nr_iowait) > 0)
3412 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3413 else
3414 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3415 /* Account for system time used */
3416 acct_update_integrals(p);
3420 * Account scaled system cpu time to a process.
3421 * @p: the process that the cpu time gets accounted to
3422 * @hardirq_offset: the offset to subtract from hardirq_count()
3423 * @cputime: the cpu time spent in kernel space since the last update
3425 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3427 p->stimescaled = cputime_add(p->stimescaled, cputime);
3431 * Account for involuntary wait time.
3432 * @p: the process from which the cpu time has been stolen
3433 * @steal: the cpu time spent in involuntary wait
3435 void account_steal_time(struct task_struct *p, cputime_t steal)
3437 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3438 cputime64_t tmp = cputime_to_cputime64(steal);
3439 struct rq *rq = this_rq();
3441 if (p == rq->idle) {
3442 p->stime = cputime_add(p->stime, steal);
3443 if (atomic_read(&rq->nr_iowait) > 0)
3444 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3445 else
3446 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3447 } else {
3448 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3449 cpuacct_charge(p, -tmp);
3454 * This function gets called by the timer code, with HZ frequency.
3455 * We call it with interrupts disabled.
3457 * It also gets called by the fork code, when changing the parent's
3458 * timeslices.
3460 void scheduler_tick(void)
3462 int cpu = smp_processor_id();
3463 struct rq *rq = cpu_rq(cpu);
3464 struct task_struct *curr = rq->curr;
3465 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3467 spin_lock(&rq->lock);
3468 __update_rq_clock(rq);
3470 * Let rq->clock advance by at least TICK_NSEC:
3472 if (unlikely(rq->clock < next_tick))
3473 rq->clock = next_tick;
3474 rq->tick_timestamp = rq->clock;
3475 update_cpu_load(rq);
3476 if (curr != rq->idle) /* FIXME: needed? */
3477 curr->sched_class->task_tick(rq, curr);
3478 spin_unlock(&rq->lock);
3480 #ifdef CONFIG_SMP
3481 rq->idle_at_tick = idle_cpu(cpu);
3482 trigger_load_balance(rq, cpu);
3483 #endif
3486 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3488 void fastcall add_preempt_count(int val)
3491 * Underflow?
3493 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3494 return;
3495 preempt_count() += val;
3497 * Spinlock count overflowing soon?
3499 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3500 PREEMPT_MASK - 10);
3502 EXPORT_SYMBOL(add_preempt_count);
3504 void fastcall sub_preempt_count(int val)
3507 * Underflow?
3509 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3510 return;
3512 * Is the spinlock portion underflowing?
3514 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3515 !(preempt_count() & PREEMPT_MASK)))
3516 return;
3518 preempt_count() -= val;
3520 EXPORT_SYMBOL(sub_preempt_count);
3522 #endif
3525 * Print scheduling while atomic bug:
3527 static noinline void __schedule_bug(struct task_struct *prev)
3529 struct pt_regs *regs = get_irq_regs();
3531 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3532 prev->comm, prev->pid, preempt_count());
3534 debug_show_held_locks(prev);
3535 if (irqs_disabled())
3536 print_irqtrace_events(prev);
3538 if (regs)
3539 show_regs(regs);
3540 else
3541 dump_stack();
3545 * Various schedule()-time debugging checks and statistics:
3547 static inline void schedule_debug(struct task_struct *prev)
3550 * Test if we are atomic. Since do_exit() needs to call into
3551 * schedule() atomically, we ignore that path for now.
3552 * Otherwise, whine if we are scheduling when we should not be.
3554 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3555 __schedule_bug(prev);
3557 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3559 schedstat_inc(this_rq(), sched_count);
3560 #ifdef CONFIG_SCHEDSTATS
3561 if (unlikely(prev->lock_depth >= 0)) {
3562 schedstat_inc(this_rq(), bkl_count);
3563 schedstat_inc(prev, sched_info.bkl_count);
3565 #endif
3569 * Pick up the highest-prio task:
3571 static inline struct task_struct *
3572 pick_next_task(struct rq *rq, struct task_struct *prev)
3574 const struct sched_class *class;
3575 struct task_struct *p;
3578 * Optimization: we know that if all tasks are in
3579 * the fair class we can call that function directly:
3581 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3582 p = fair_sched_class.pick_next_task(rq);
3583 if (likely(p))
3584 return p;
3587 class = sched_class_highest;
3588 for ( ; ; ) {
3589 p = class->pick_next_task(rq);
3590 if (p)
3591 return p;
3593 * Will never be NULL as the idle class always
3594 * returns a non-NULL p:
3596 class = class->next;
3601 * schedule() is the main scheduler function.
3603 asmlinkage void __sched schedule(void)
3605 struct task_struct *prev, *next;
3606 long *switch_count;
3607 struct rq *rq;
3608 int cpu;
3610 need_resched:
3611 preempt_disable();
3612 cpu = smp_processor_id();
3613 rq = cpu_rq(cpu);
3614 rcu_qsctr_inc(cpu);
3615 prev = rq->curr;
3616 switch_count = &prev->nivcsw;
3618 release_kernel_lock(prev);
3619 need_resched_nonpreemptible:
3621 schedule_debug(prev);
3624 * Do the rq-clock update outside the rq lock:
3626 local_irq_disable();
3627 __update_rq_clock(rq);
3628 spin_lock(&rq->lock);
3629 clear_tsk_need_resched(prev);
3631 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3632 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3633 unlikely(signal_pending(prev)))) {
3634 prev->state = TASK_RUNNING;
3635 } else {
3636 deactivate_task(rq, prev, 1);
3638 switch_count = &prev->nvcsw;
3641 if (unlikely(!rq->nr_running))
3642 idle_balance(cpu, rq);
3644 prev->sched_class->put_prev_task(rq, prev);
3645 next = pick_next_task(rq, prev);
3647 sched_info_switch(prev, next);
3649 if (likely(prev != next)) {
3650 rq->nr_switches++;
3651 rq->curr = next;
3652 ++*switch_count;
3654 context_switch(rq, prev, next); /* unlocks the rq */
3655 } else
3656 spin_unlock_irq(&rq->lock);
3658 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3659 cpu = smp_processor_id();
3660 rq = cpu_rq(cpu);
3661 goto need_resched_nonpreemptible;
3663 preempt_enable_no_resched();
3664 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3665 goto need_resched;
3667 EXPORT_SYMBOL(schedule);
3669 #ifdef CONFIG_PREEMPT
3671 * this is the entry point to schedule() from in-kernel preemption
3672 * off of preempt_enable. Kernel preemptions off return from interrupt
3673 * occur there and call schedule directly.
3675 asmlinkage void __sched preempt_schedule(void)
3677 struct thread_info *ti = current_thread_info();
3678 #ifdef CONFIG_PREEMPT_BKL
3679 struct task_struct *task = current;
3680 int saved_lock_depth;
3681 #endif
3683 * If there is a non-zero preempt_count or interrupts are disabled,
3684 * we do not want to preempt the current task. Just return..
3686 if (likely(ti->preempt_count || irqs_disabled()))
3687 return;
3689 do {
3690 add_preempt_count(PREEMPT_ACTIVE);
3693 * We keep the big kernel semaphore locked, but we
3694 * clear ->lock_depth so that schedule() doesnt
3695 * auto-release the semaphore:
3697 #ifdef CONFIG_PREEMPT_BKL
3698 saved_lock_depth = task->lock_depth;
3699 task->lock_depth = -1;
3700 #endif
3701 schedule();
3702 #ifdef CONFIG_PREEMPT_BKL
3703 task->lock_depth = saved_lock_depth;
3704 #endif
3705 sub_preempt_count(PREEMPT_ACTIVE);
3708 * Check again in case we missed a preemption opportunity
3709 * between schedule and now.
3711 barrier();
3712 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3714 EXPORT_SYMBOL(preempt_schedule);
3717 * this is the entry point to schedule() from kernel preemption
3718 * off of irq context.
3719 * Note, that this is called and return with irqs disabled. This will
3720 * protect us against recursive calling from irq.
3722 asmlinkage void __sched preempt_schedule_irq(void)
3724 struct thread_info *ti = current_thread_info();
3725 #ifdef CONFIG_PREEMPT_BKL
3726 struct task_struct *task = current;
3727 int saved_lock_depth;
3728 #endif
3729 /* Catch callers which need to be fixed */
3730 BUG_ON(ti->preempt_count || !irqs_disabled());
3732 do {
3733 add_preempt_count(PREEMPT_ACTIVE);
3736 * We keep the big kernel semaphore locked, but we
3737 * clear ->lock_depth so that schedule() doesnt
3738 * auto-release the semaphore:
3740 #ifdef CONFIG_PREEMPT_BKL
3741 saved_lock_depth = task->lock_depth;
3742 task->lock_depth = -1;
3743 #endif
3744 local_irq_enable();
3745 schedule();
3746 local_irq_disable();
3747 #ifdef CONFIG_PREEMPT_BKL
3748 task->lock_depth = saved_lock_depth;
3749 #endif
3750 sub_preempt_count(PREEMPT_ACTIVE);
3753 * Check again in case we missed a preemption opportunity
3754 * between schedule and now.
3756 barrier();
3757 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3760 #endif /* CONFIG_PREEMPT */
3762 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3763 void *key)
3765 return try_to_wake_up(curr->private, mode, sync);
3767 EXPORT_SYMBOL(default_wake_function);
3770 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3771 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3772 * number) then we wake all the non-exclusive tasks and one exclusive task.
3774 * There are circumstances in which we can try to wake a task which has already
3775 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3776 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3778 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3779 int nr_exclusive, int sync, void *key)
3781 wait_queue_t *curr, *next;
3783 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3784 unsigned flags = curr->flags;
3786 if (curr->func(curr, mode, sync, key) &&
3787 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3788 break;
3793 * __wake_up - wake up threads blocked on a waitqueue.
3794 * @q: the waitqueue
3795 * @mode: which threads
3796 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3797 * @key: is directly passed to the wakeup function
3799 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3800 int nr_exclusive, void *key)
3802 unsigned long flags;
3804 spin_lock_irqsave(&q->lock, flags);
3805 __wake_up_common(q, mode, nr_exclusive, 0, key);
3806 spin_unlock_irqrestore(&q->lock, flags);
3808 EXPORT_SYMBOL(__wake_up);
3811 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3813 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3815 __wake_up_common(q, mode, 1, 0, NULL);
3819 * __wake_up_sync - wake up threads blocked on a waitqueue.
3820 * @q: the waitqueue
3821 * @mode: which threads
3822 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3824 * The sync wakeup differs that the waker knows that it will schedule
3825 * away soon, so while the target thread will be woken up, it will not
3826 * be migrated to another CPU - ie. the two threads are 'synchronized'
3827 * with each other. This can prevent needless bouncing between CPUs.
3829 * On UP it can prevent extra preemption.
3831 void fastcall
3832 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3834 unsigned long flags;
3835 int sync = 1;
3837 if (unlikely(!q))
3838 return;
3840 if (unlikely(!nr_exclusive))
3841 sync = 0;
3843 spin_lock_irqsave(&q->lock, flags);
3844 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3845 spin_unlock_irqrestore(&q->lock, flags);
3847 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3849 void complete(struct completion *x)
3851 unsigned long flags;
3853 spin_lock_irqsave(&x->wait.lock, flags);
3854 x->done++;
3855 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3856 1, 0, NULL);
3857 spin_unlock_irqrestore(&x->wait.lock, flags);
3859 EXPORT_SYMBOL(complete);
3861 void complete_all(struct completion *x)
3863 unsigned long flags;
3865 spin_lock_irqsave(&x->wait.lock, flags);
3866 x->done += UINT_MAX/2;
3867 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3868 0, 0, NULL);
3869 spin_unlock_irqrestore(&x->wait.lock, flags);
3871 EXPORT_SYMBOL(complete_all);
3873 static inline long __sched
3874 do_wait_for_common(struct completion *x, long timeout, int state)
3876 if (!x->done) {
3877 DECLARE_WAITQUEUE(wait, current);
3879 wait.flags |= WQ_FLAG_EXCLUSIVE;
3880 __add_wait_queue_tail(&x->wait, &wait);
3881 do {
3882 if (state == TASK_INTERRUPTIBLE &&
3883 signal_pending(current)) {
3884 __remove_wait_queue(&x->wait, &wait);
3885 return -ERESTARTSYS;
3887 __set_current_state(state);
3888 spin_unlock_irq(&x->wait.lock);
3889 timeout = schedule_timeout(timeout);
3890 spin_lock_irq(&x->wait.lock);
3891 if (!timeout) {
3892 __remove_wait_queue(&x->wait, &wait);
3893 return timeout;
3895 } while (!x->done);
3896 __remove_wait_queue(&x->wait, &wait);
3898 x->done--;
3899 return timeout;
3902 static long __sched
3903 wait_for_common(struct completion *x, long timeout, int state)
3905 might_sleep();
3907 spin_lock_irq(&x->wait.lock);
3908 timeout = do_wait_for_common(x, timeout, state);
3909 spin_unlock_irq(&x->wait.lock);
3910 return timeout;
3913 void __sched wait_for_completion(struct completion *x)
3915 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3917 EXPORT_SYMBOL(wait_for_completion);
3919 unsigned long __sched
3920 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3922 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3924 EXPORT_SYMBOL(wait_for_completion_timeout);
3926 int __sched wait_for_completion_interruptible(struct completion *x)
3928 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3929 if (t == -ERESTARTSYS)
3930 return t;
3931 return 0;
3933 EXPORT_SYMBOL(wait_for_completion_interruptible);
3935 unsigned long __sched
3936 wait_for_completion_interruptible_timeout(struct completion *x,
3937 unsigned long timeout)
3939 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3941 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3943 static long __sched
3944 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3946 unsigned long flags;
3947 wait_queue_t wait;
3949 init_waitqueue_entry(&wait, current);
3951 __set_current_state(state);
3953 spin_lock_irqsave(&q->lock, flags);
3954 __add_wait_queue(q, &wait);
3955 spin_unlock(&q->lock);
3956 timeout = schedule_timeout(timeout);
3957 spin_lock_irq(&q->lock);
3958 __remove_wait_queue(q, &wait);
3959 spin_unlock_irqrestore(&q->lock, flags);
3961 return timeout;
3964 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3966 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3968 EXPORT_SYMBOL(interruptible_sleep_on);
3970 long __sched
3971 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3973 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3975 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3977 void __sched sleep_on(wait_queue_head_t *q)
3979 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3981 EXPORT_SYMBOL(sleep_on);
3983 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3985 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3987 EXPORT_SYMBOL(sleep_on_timeout);
3989 #ifdef CONFIG_RT_MUTEXES
3992 * rt_mutex_setprio - set the current priority of a task
3993 * @p: task
3994 * @prio: prio value (kernel-internal form)
3996 * This function changes the 'effective' priority of a task. It does
3997 * not touch ->normal_prio like __setscheduler().
3999 * Used by the rt_mutex code to implement priority inheritance logic.
4001 void rt_mutex_setprio(struct task_struct *p, int prio)
4003 unsigned long flags;
4004 int oldprio, on_rq, running;
4005 struct rq *rq;
4007 BUG_ON(prio < 0 || prio > MAX_PRIO);
4009 rq = task_rq_lock(p, &flags);
4010 update_rq_clock(rq);
4012 oldprio = p->prio;
4013 on_rq = p->se.on_rq;
4014 running = task_running(rq, p);
4015 if (on_rq) {
4016 dequeue_task(rq, p, 0);
4017 if (running)
4018 p->sched_class->put_prev_task(rq, p);
4021 if (rt_prio(prio))
4022 p->sched_class = &rt_sched_class;
4023 else
4024 p->sched_class = &fair_sched_class;
4026 p->prio = prio;
4028 if (on_rq) {
4029 if (running)
4030 p->sched_class->set_curr_task(rq);
4031 enqueue_task(rq, p, 0);
4033 * Reschedule if we are currently running on this runqueue and
4034 * our priority decreased, or if we are not currently running on
4035 * this runqueue and our priority is higher than the current's
4037 if (running) {
4038 if (p->prio > oldprio)
4039 resched_task(rq->curr);
4040 } else {
4041 check_preempt_curr(rq, p);
4044 task_rq_unlock(rq, &flags);
4047 #endif
4049 void set_user_nice(struct task_struct *p, long nice)
4051 int old_prio, delta, on_rq;
4052 unsigned long flags;
4053 struct rq *rq;
4055 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4056 return;
4058 * We have to be careful, if called from sys_setpriority(),
4059 * the task might be in the middle of scheduling on another CPU.
4061 rq = task_rq_lock(p, &flags);
4062 update_rq_clock(rq);
4064 * The RT priorities are set via sched_setscheduler(), but we still
4065 * allow the 'normal' nice value to be set - but as expected
4066 * it wont have any effect on scheduling until the task is
4067 * SCHED_FIFO/SCHED_RR:
4069 if (task_has_rt_policy(p)) {
4070 p->static_prio = NICE_TO_PRIO(nice);
4071 goto out_unlock;
4073 on_rq = p->se.on_rq;
4074 if (on_rq) {
4075 dequeue_task(rq, p, 0);
4076 dec_load(rq, p);
4079 p->static_prio = NICE_TO_PRIO(nice);
4080 set_load_weight(p);
4081 old_prio = p->prio;
4082 p->prio = effective_prio(p);
4083 delta = p->prio - old_prio;
4085 if (on_rq) {
4086 enqueue_task(rq, p, 0);
4087 inc_load(rq, p);
4089 * If the task increased its priority or is running and
4090 * lowered its priority, then reschedule its CPU:
4092 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4093 resched_task(rq->curr);
4095 out_unlock:
4096 task_rq_unlock(rq, &flags);
4098 EXPORT_SYMBOL(set_user_nice);
4101 * can_nice - check if a task can reduce its nice value
4102 * @p: task
4103 * @nice: nice value
4105 int can_nice(const struct task_struct *p, const int nice)
4107 /* convert nice value [19,-20] to rlimit style value [1,40] */
4108 int nice_rlim = 20 - nice;
4110 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4111 capable(CAP_SYS_NICE));
4114 #ifdef __ARCH_WANT_SYS_NICE
4117 * sys_nice - change the priority of the current process.
4118 * @increment: priority increment
4120 * sys_setpriority is a more generic, but much slower function that
4121 * does similar things.
4123 asmlinkage long sys_nice(int increment)
4125 long nice, retval;
4128 * Setpriority might change our priority at the same moment.
4129 * We don't have to worry. Conceptually one call occurs first
4130 * and we have a single winner.
4132 if (increment < -40)
4133 increment = -40;
4134 if (increment > 40)
4135 increment = 40;
4137 nice = PRIO_TO_NICE(current->static_prio) + increment;
4138 if (nice < -20)
4139 nice = -20;
4140 if (nice > 19)
4141 nice = 19;
4143 if (increment < 0 && !can_nice(current, nice))
4144 return -EPERM;
4146 retval = security_task_setnice(current, nice);
4147 if (retval)
4148 return retval;
4150 set_user_nice(current, nice);
4151 return 0;
4154 #endif
4157 * task_prio - return the priority value of a given task.
4158 * @p: the task in question.
4160 * This is the priority value as seen by users in /proc.
4161 * RT tasks are offset by -200. Normal tasks are centered
4162 * around 0, value goes from -16 to +15.
4164 int task_prio(const struct task_struct *p)
4166 return p->prio - MAX_RT_PRIO;
4170 * task_nice - return the nice value of a given task.
4171 * @p: the task in question.
4173 int task_nice(const struct task_struct *p)
4175 return TASK_NICE(p);
4177 EXPORT_SYMBOL_GPL(task_nice);
4180 * idle_cpu - is a given cpu idle currently?
4181 * @cpu: the processor in question.
4183 int idle_cpu(int cpu)
4185 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4189 * idle_task - return the idle task for a given cpu.
4190 * @cpu: the processor in question.
4192 struct task_struct *idle_task(int cpu)
4194 return cpu_rq(cpu)->idle;
4198 * find_process_by_pid - find a process with a matching PID value.
4199 * @pid: the pid in question.
4201 static struct task_struct *find_process_by_pid(pid_t pid)
4203 return pid ? find_task_by_vpid(pid) : current;
4206 /* Actually do priority change: must hold rq lock. */
4207 static void
4208 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4210 BUG_ON(p->se.on_rq);
4212 p->policy = policy;
4213 switch (p->policy) {
4214 case SCHED_NORMAL:
4215 case SCHED_BATCH:
4216 case SCHED_IDLE:
4217 p->sched_class = &fair_sched_class;
4218 break;
4219 case SCHED_FIFO:
4220 case SCHED_RR:
4221 p->sched_class = &rt_sched_class;
4222 break;
4225 p->rt_priority = prio;
4226 p->normal_prio = normal_prio(p);
4227 /* we are holding p->pi_lock already */
4228 p->prio = rt_mutex_getprio(p);
4229 set_load_weight(p);
4233 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4234 * @p: the task in question.
4235 * @policy: new policy.
4236 * @param: structure containing the new RT priority.
4238 * NOTE that the task may be already dead.
4240 int sched_setscheduler(struct task_struct *p, int policy,
4241 struct sched_param *param)
4243 int retval, oldprio, oldpolicy = -1, on_rq, running;
4244 unsigned long flags;
4245 struct rq *rq;
4247 /* may grab non-irq protected spin_locks */
4248 BUG_ON(in_interrupt());
4249 recheck:
4250 /* double check policy once rq lock held */
4251 if (policy < 0)
4252 policy = oldpolicy = p->policy;
4253 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4254 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4255 policy != SCHED_IDLE)
4256 return -EINVAL;
4258 * Valid priorities for SCHED_FIFO and SCHED_RR are
4259 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4260 * SCHED_BATCH and SCHED_IDLE is 0.
4262 if (param->sched_priority < 0 ||
4263 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4264 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4265 return -EINVAL;
4266 if (rt_policy(policy) != (param->sched_priority != 0))
4267 return -EINVAL;
4270 * Allow unprivileged RT tasks to decrease priority:
4272 if (!capable(CAP_SYS_NICE)) {
4273 if (rt_policy(policy)) {
4274 unsigned long rlim_rtprio;
4276 if (!lock_task_sighand(p, &flags))
4277 return -ESRCH;
4278 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4279 unlock_task_sighand(p, &flags);
4281 /* can't set/change the rt policy */
4282 if (policy != p->policy && !rlim_rtprio)
4283 return -EPERM;
4285 /* can't increase priority */
4286 if (param->sched_priority > p->rt_priority &&
4287 param->sched_priority > rlim_rtprio)
4288 return -EPERM;
4291 * Like positive nice levels, dont allow tasks to
4292 * move out of SCHED_IDLE either:
4294 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4295 return -EPERM;
4297 /* can't change other user's priorities */
4298 if ((current->euid != p->euid) &&
4299 (current->euid != p->uid))
4300 return -EPERM;
4303 retval = security_task_setscheduler(p, policy, param);
4304 if (retval)
4305 return retval;
4307 * make sure no PI-waiters arrive (or leave) while we are
4308 * changing the priority of the task:
4310 spin_lock_irqsave(&p->pi_lock, flags);
4312 * To be able to change p->policy safely, the apropriate
4313 * runqueue lock must be held.
4315 rq = __task_rq_lock(p);
4316 /* recheck policy now with rq lock held */
4317 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4318 policy = oldpolicy = -1;
4319 __task_rq_unlock(rq);
4320 spin_unlock_irqrestore(&p->pi_lock, flags);
4321 goto recheck;
4323 update_rq_clock(rq);
4324 on_rq = p->se.on_rq;
4325 running = task_running(rq, p);
4326 if (on_rq) {
4327 deactivate_task(rq, p, 0);
4328 if (running)
4329 p->sched_class->put_prev_task(rq, p);
4332 oldprio = p->prio;
4333 __setscheduler(rq, p, policy, param->sched_priority);
4335 if (on_rq) {
4336 if (running)
4337 p->sched_class->set_curr_task(rq);
4338 activate_task(rq, p, 0);
4340 * Reschedule if we are currently running on this runqueue and
4341 * our priority decreased, or if we are not currently running on
4342 * this runqueue and our priority is higher than the current's
4344 if (running) {
4345 if (p->prio > oldprio)
4346 resched_task(rq->curr);
4347 } else {
4348 check_preempt_curr(rq, p);
4351 __task_rq_unlock(rq);
4352 spin_unlock_irqrestore(&p->pi_lock, flags);
4354 rt_mutex_adjust_pi(p);
4356 return 0;
4358 EXPORT_SYMBOL_GPL(sched_setscheduler);
4360 static int
4361 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4363 struct sched_param lparam;
4364 struct task_struct *p;
4365 int retval;
4367 if (!param || pid < 0)
4368 return -EINVAL;
4369 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4370 return -EFAULT;
4372 rcu_read_lock();
4373 retval = -ESRCH;
4374 p = find_process_by_pid(pid);
4375 if (p != NULL)
4376 retval = sched_setscheduler(p, policy, &lparam);
4377 rcu_read_unlock();
4379 return retval;
4383 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4384 * @pid: the pid in question.
4385 * @policy: new policy.
4386 * @param: structure containing the new RT priority.
4388 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4389 struct sched_param __user *param)
4391 /* negative values for policy are not valid */
4392 if (policy < 0)
4393 return -EINVAL;
4395 return do_sched_setscheduler(pid, policy, param);
4399 * sys_sched_setparam - set/change the RT priority of a thread
4400 * @pid: the pid in question.
4401 * @param: structure containing the new RT priority.
4403 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4405 return do_sched_setscheduler(pid, -1, param);
4409 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4410 * @pid: the pid in question.
4412 asmlinkage long sys_sched_getscheduler(pid_t pid)
4414 struct task_struct *p;
4415 int retval;
4417 if (pid < 0)
4418 return -EINVAL;
4420 retval = -ESRCH;
4421 read_lock(&tasklist_lock);
4422 p = find_process_by_pid(pid);
4423 if (p) {
4424 retval = security_task_getscheduler(p);
4425 if (!retval)
4426 retval = p->policy;
4428 read_unlock(&tasklist_lock);
4429 return retval;
4433 * sys_sched_getscheduler - get the RT priority of a thread
4434 * @pid: the pid in question.
4435 * @param: structure containing the RT priority.
4437 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4439 struct sched_param lp;
4440 struct task_struct *p;
4441 int retval;
4443 if (!param || pid < 0)
4444 return -EINVAL;
4446 read_lock(&tasklist_lock);
4447 p = find_process_by_pid(pid);
4448 retval = -ESRCH;
4449 if (!p)
4450 goto out_unlock;
4452 retval = security_task_getscheduler(p);
4453 if (retval)
4454 goto out_unlock;
4456 lp.sched_priority = p->rt_priority;
4457 read_unlock(&tasklist_lock);
4460 * This one might sleep, we cannot do it with a spinlock held ...
4462 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4464 return retval;
4466 out_unlock:
4467 read_unlock(&tasklist_lock);
4468 return retval;
4471 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4473 cpumask_t cpus_allowed;
4474 struct task_struct *p;
4475 int retval;
4477 mutex_lock(&sched_hotcpu_mutex);
4478 read_lock(&tasklist_lock);
4480 p = find_process_by_pid(pid);
4481 if (!p) {
4482 read_unlock(&tasklist_lock);
4483 mutex_unlock(&sched_hotcpu_mutex);
4484 return -ESRCH;
4488 * It is not safe to call set_cpus_allowed with the
4489 * tasklist_lock held. We will bump the task_struct's
4490 * usage count and then drop tasklist_lock.
4492 get_task_struct(p);
4493 read_unlock(&tasklist_lock);
4495 retval = -EPERM;
4496 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4497 !capable(CAP_SYS_NICE))
4498 goto out_unlock;
4500 retval = security_task_setscheduler(p, 0, NULL);
4501 if (retval)
4502 goto out_unlock;
4504 cpus_allowed = cpuset_cpus_allowed(p);
4505 cpus_and(new_mask, new_mask, cpus_allowed);
4506 again:
4507 retval = set_cpus_allowed(p, new_mask);
4509 if (!retval) {
4510 cpus_allowed = cpuset_cpus_allowed(p);
4511 if (!cpus_subset(new_mask, cpus_allowed)) {
4513 * We must have raced with a concurrent cpuset
4514 * update. Just reset the cpus_allowed to the
4515 * cpuset's cpus_allowed
4517 new_mask = cpus_allowed;
4518 goto again;
4521 out_unlock:
4522 put_task_struct(p);
4523 mutex_unlock(&sched_hotcpu_mutex);
4524 return retval;
4527 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4528 cpumask_t *new_mask)
4530 if (len < sizeof(cpumask_t)) {
4531 memset(new_mask, 0, sizeof(cpumask_t));
4532 } else if (len > sizeof(cpumask_t)) {
4533 len = sizeof(cpumask_t);
4535 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4539 * sys_sched_setaffinity - set the cpu affinity of a process
4540 * @pid: pid of the process
4541 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4542 * @user_mask_ptr: user-space pointer to the new cpu mask
4544 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4545 unsigned long __user *user_mask_ptr)
4547 cpumask_t new_mask;
4548 int retval;
4550 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4551 if (retval)
4552 return retval;
4554 return sched_setaffinity(pid, new_mask);
4558 * Represents all cpu's present in the system
4559 * In systems capable of hotplug, this map could dynamically grow
4560 * as new cpu's are detected in the system via any platform specific
4561 * method, such as ACPI for e.g.
4564 cpumask_t cpu_present_map __read_mostly;
4565 EXPORT_SYMBOL(cpu_present_map);
4567 #ifndef CONFIG_SMP
4568 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4569 EXPORT_SYMBOL(cpu_online_map);
4571 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4572 EXPORT_SYMBOL(cpu_possible_map);
4573 #endif
4575 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4577 struct task_struct *p;
4578 int retval;
4580 mutex_lock(&sched_hotcpu_mutex);
4581 read_lock(&tasklist_lock);
4583 retval = -ESRCH;
4584 p = find_process_by_pid(pid);
4585 if (!p)
4586 goto out_unlock;
4588 retval = security_task_getscheduler(p);
4589 if (retval)
4590 goto out_unlock;
4592 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4594 out_unlock:
4595 read_unlock(&tasklist_lock);
4596 mutex_unlock(&sched_hotcpu_mutex);
4598 return retval;
4602 * sys_sched_getaffinity - get the cpu affinity of a process
4603 * @pid: pid of the process
4604 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4605 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4607 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4608 unsigned long __user *user_mask_ptr)
4610 int ret;
4611 cpumask_t mask;
4613 if (len < sizeof(cpumask_t))
4614 return -EINVAL;
4616 ret = sched_getaffinity(pid, &mask);
4617 if (ret < 0)
4618 return ret;
4620 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4621 return -EFAULT;
4623 return sizeof(cpumask_t);
4627 * sys_sched_yield - yield the current processor to other threads.
4629 * This function yields the current CPU to other tasks. If there are no
4630 * other threads running on this CPU then this function will return.
4632 asmlinkage long sys_sched_yield(void)
4634 struct rq *rq = this_rq_lock();
4636 schedstat_inc(rq, yld_count);
4637 current->sched_class->yield_task(rq);
4640 * Since we are going to call schedule() anyway, there's
4641 * no need to preempt or enable interrupts:
4643 __release(rq->lock);
4644 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4645 _raw_spin_unlock(&rq->lock);
4646 preempt_enable_no_resched();
4648 schedule();
4650 return 0;
4653 static void __cond_resched(void)
4655 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4656 __might_sleep(__FILE__, __LINE__);
4657 #endif
4659 * The BKS might be reacquired before we have dropped
4660 * PREEMPT_ACTIVE, which could trigger a second
4661 * cond_resched() call.
4663 do {
4664 add_preempt_count(PREEMPT_ACTIVE);
4665 schedule();
4666 sub_preempt_count(PREEMPT_ACTIVE);
4667 } while (need_resched());
4670 int __sched cond_resched(void)
4672 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4673 system_state == SYSTEM_RUNNING) {
4674 __cond_resched();
4675 return 1;
4677 return 0;
4679 EXPORT_SYMBOL(cond_resched);
4682 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4683 * call schedule, and on return reacquire the lock.
4685 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4686 * operations here to prevent schedule() from being called twice (once via
4687 * spin_unlock(), once by hand).
4689 int cond_resched_lock(spinlock_t *lock)
4691 int ret = 0;
4693 if (need_lockbreak(lock)) {
4694 spin_unlock(lock);
4695 cpu_relax();
4696 ret = 1;
4697 spin_lock(lock);
4699 if (need_resched() && system_state == SYSTEM_RUNNING) {
4700 spin_release(&lock->dep_map, 1, _THIS_IP_);
4701 _raw_spin_unlock(lock);
4702 preempt_enable_no_resched();
4703 __cond_resched();
4704 ret = 1;
4705 spin_lock(lock);
4707 return ret;
4709 EXPORT_SYMBOL(cond_resched_lock);
4711 int __sched cond_resched_softirq(void)
4713 BUG_ON(!in_softirq());
4715 if (need_resched() && system_state == SYSTEM_RUNNING) {
4716 local_bh_enable();
4717 __cond_resched();
4718 local_bh_disable();
4719 return 1;
4721 return 0;
4723 EXPORT_SYMBOL(cond_resched_softirq);
4726 * yield - yield the current processor to other threads.
4728 * This is a shortcut for kernel-space yielding - it marks the
4729 * thread runnable and calls sys_sched_yield().
4731 void __sched yield(void)
4733 set_current_state(TASK_RUNNING);
4734 sys_sched_yield();
4736 EXPORT_SYMBOL(yield);
4739 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4740 * that process accounting knows that this is a task in IO wait state.
4742 * But don't do that if it is a deliberate, throttling IO wait (this task
4743 * has set its backing_dev_info: the queue against which it should throttle)
4745 void __sched io_schedule(void)
4747 struct rq *rq = &__raw_get_cpu_var(runqueues);
4749 delayacct_blkio_start();
4750 atomic_inc(&rq->nr_iowait);
4751 schedule();
4752 atomic_dec(&rq->nr_iowait);
4753 delayacct_blkio_end();
4755 EXPORT_SYMBOL(io_schedule);
4757 long __sched io_schedule_timeout(long timeout)
4759 struct rq *rq = &__raw_get_cpu_var(runqueues);
4760 long ret;
4762 delayacct_blkio_start();
4763 atomic_inc(&rq->nr_iowait);
4764 ret = schedule_timeout(timeout);
4765 atomic_dec(&rq->nr_iowait);
4766 delayacct_blkio_end();
4767 return ret;
4771 * sys_sched_get_priority_max - return maximum RT priority.
4772 * @policy: scheduling class.
4774 * this syscall returns the maximum rt_priority that can be used
4775 * by a given scheduling class.
4777 asmlinkage long sys_sched_get_priority_max(int policy)
4779 int ret = -EINVAL;
4781 switch (policy) {
4782 case SCHED_FIFO:
4783 case SCHED_RR:
4784 ret = MAX_USER_RT_PRIO-1;
4785 break;
4786 case SCHED_NORMAL:
4787 case SCHED_BATCH:
4788 case SCHED_IDLE:
4789 ret = 0;
4790 break;
4792 return ret;
4796 * sys_sched_get_priority_min - return minimum RT priority.
4797 * @policy: scheduling class.
4799 * this syscall returns the minimum rt_priority that can be used
4800 * by a given scheduling class.
4802 asmlinkage long sys_sched_get_priority_min(int policy)
4804 int ret = -EINVAL;
4806 switch (policy) {
4807 case SCHED_FIFO:
4808 case SCHED_RR:
4809 ret = 1;
4810 break;
4811 case SCHED_NORMAL:
4812 case SCHED_BATCH:
4813 case SCHED_IDLE:
4814 ret = 0;
4816 return ret;
4820 * sys_sched_rr_get_interval - return the default timeslice of a process.
4821 * @pid: pid of the process.
4822 * @interval: userspace pointer to the timeslice value.
4824 * this syscall writes the default timeslice value of a given process
4825 * into the user-space timespec buffer. A value of '0' means infinity.
4827 asmlinkage
4828 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4830 struct task_struct *p;
4831 unsigned int time_slice;
4832 int retval;
4833 struct timespec t;
4835 if (pid < 0)
4836 return -EINVAL;
4838 retval = -ESRCH;
4839 read_lock(&tasklist_lock);
4840 p = find_process_by_pid(pid);
4841 if (!p)
4842 goto out_unlock;
4844 retval = security_task_getscheduler(p);
4845 if (retval)
4846 goto out_unlock;
4848 if (p->policy == SCHED_FIFO)
4849 time_slice = 0;
4850 else if (p->policy == SCHED_RR)
4851 time_slice = DEF_TIMESLICE;
4852 else {
4853 struct sched_entity *se = &p->se;
4854 unsigned long flags;
4855 struct rq *rq;
4857 rq = task_rq_lock(p, &flags);
4858 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4859 task_rq_unlock(rq, &flags);
4861 read_unlock(&tasklist_lock);
4862 jiffies_to_timespec(time_slice, &t);
4863 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4864 return retval;
4866 out_unlock:
4867 read_unlock(&tasklist_lock);
4868 return retval;
4871 static const char stat_nam[] = "RSDTtZX";
4873 static void show_task(struct task_struct *p)
4875 unsigned long free = 0;
4876 unsigned state;
4878 state = p->state ? __ffs(p->state) + 1 : 0;
4879 printk(KERN_INFO "%-13.13s %c", p->comm,
4880 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4881 #if BITS_PER_LONG == 32
4882 if (state == TASK_RUNNING)
4883 printk(KERN_CONT " running ");
4884 else
4885 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4886 #else
4887 if (state == TASK_RUNNING)
4888 printk(KERN_CONT " running task ");
4889 else
4890 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4891 #endif
4892 #ifdef CONFIG_DEBUG_STACK_USAGE
4894 unsigned long *n = end_of_stack(p);
4895 while (!*n)
4896 n++;
4897 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4899 #endif
4900 printk(KERN_CONT "%5lu %5d %6d\n", free,
4901 task_pid_nr(p), task_pid_nr(p->parent));
4903 if (state != TASK_RUNNING)
4904 show_stack(p, NULL);
4907 void show_state_filter(unsigned long state_filter)
4909 struct task_struct *g, *p;
4911 #if BITS_PER_LONG == 32
4912 printk(KERN_INFO
4913 " task PC stack pid father\n");
4914 #else
4915 printk(KERN_INFO
4916 " task PC stack pid father\n");
4917 #endif
4918 read_lock(&tasklist_lock);
4919 do_each_thread(g, p) {
4921 * reset the NMI-timeout, listing all files on a slow
4922 * console might take alot of time:
4924 touch_nmi_watchdog();
4925 if (!state_filter || (p->state & state_filter))
4926 show_task(p);
4927 } while_each_thread(g, p);
4929 touch_all_softlockup_watchdogs();
4931 #ifdef CONFIG_SCHED_DEBUG
4932 sysrq_sched_debug_show();
4933 #endif
4934 read_unlock(&tasklist_lock);
4936 * Only show locks if all tasks are dumped:
4938 if (state_filter == -1)
4939 debug_show_all_locks();
4942 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4944 idle->sched_class = &idle_sched_class;
4948 * init_idle - set up an idle thread for a given CPU
4949 * @idle: task in question
4950 * @cpu: cpu the idle task belongs to
4952 * NOTE: this function does not set the idle thread's NEED_RESCHED
4953 * flag, to make booting more robust.
4955 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4957 struct rq *rq = cpu_rq(cpu);
4958 unsigned long flags;
4960 __sched_fork(idle);
4961 idle->se.exec_start = sched_clock();
4963 idle->prio = idle->normal_prio = MAX_PRIO;
4964 idle->cpus_allowed = cpumask_of_cpu(cpu);
4965 __set_task_cpu(idle, cpu);
4967 spin_lock_irqsave(&rq->lock, flags);
4968 rq->curr = rq->idle = idle;
4969 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4970 idle->oncpu = 1;
4971 #endif
4972 spin_unlock_irqrestore(&rq->lock, flags);
4974 /* Set the preempt count _outside_ the spinlocks! */
4975 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4976 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4977 #else
4978 task_thread_info(idle)->preempt_count = 0;
4979 #endif
4981 * The idle tasks have their own, simple scheduling class:
4983 idle->sched_class = &idle_sched_class;
4987 * In a system that switches off the HZ timer nohz_cpu_mask
4988 * indicates which cpus entered this state. This is used
4989 * in the rcu update to wait only for active cpus. For system
4990 * which do not switch off the HZ timer nohz_cpu_mask should
4991 * always be CPU_MASK_NONE.
4993 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4995 #ifdef CONFIG_SMP
4997 * This is how migration works:
4999 * 1) we queue a struct migration_req structure in the source CPU's
5000 * runqueue and wake up that CPU's migration thread.
5001 * 2) we down() the locked semaphore => thread blocks.
5002 * 3) migration thread wakes up (implicitly it forces the migrated
5003 * thread off the CPU)
5004 * 4) it gets the migration request and checks whether the migrated
5005 * task is still in the wrong runqueue.
5006 * 5) if it's in the wrong runqueue then the migration thread removes
5007 * it and puts it into the right queue.
5008 * 6) migration thread up()s the semaphore.
5009 * 7) we wake up and the migration is done.
5013 * Change a given task's CPU affinity. Migrate the thread to a
5014 * proper CPU and schedule it away if the CPU it's executing on
5015 * is removed from the allowed bitmask.
5017 * NOTE: the caller must have a valid reference to the task, the
5018 * task must not exit() & deallocate itself prematurely. The
5019 * call is not atomic; no spinlocks may be held.
5021 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5023 struct migration_req req;
5024 unsigned long flags;
5025 struct rq *rq;
5026 int ret = 0;
5028 rq = task_rq_lock(p, &flags);
5029 if (!cpus_intersects(new_mask, cpu_online_map)) {
5030 ret = -EINVAL;
5031 goto out;
5034 p->cpus_allowed = new_mask;
5035 /* Can the task run on the task's current CPU? If so, we're done */
5036 if (cpu_isset(task_cpu(p), new_mask))
5037 goto out;
5039 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5040 /* Need help from migration thread: drop lock and wait. */
5041 task_rq_unlock(rq, &flags);
5042 wake_up_process(rq->migration_thread);
5043 wait_for_completion(&req.done);
5044 tlb_migrate_finish(p->mm);
5045 return 0;
5047 out:
5048 task_rq_unlock(rq, &flags);
5050 return ret;
5052 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5055 * Move (not current) task off this cpu, onto dest cpu. We're doing
5056 * this because either it can't run here any more (set_cpus_allowed()
5057 * away from this CPU, or CPU going down), or because we're
5058 * attempting to rebalance this task on exec (sched_exec).
5060 * So we race with normal scheduler movements, but that's OK, as long
5061 * as the task is no longer on this CPU.
5063 * Returns non-zero if task was successfully migrated.
5065 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5067 struct rq *rq_dest, *rq_src;
5068 int ret = 0, on_rq;
5070 if (unlikely(cpu_is_offline(dest_cpu)))
5071 return ret;
5073 rq_src = cpu_rq(src_cpu);
5074 rq_dest = cpu_rq(dest_cpu);
5076 double_rq_lock(rq_src, rq_dest);
5077 /* Already moved. */
5078 if (task_cpu(p) != src_cpu)
5079 goto out;
5080 /* Affinity changed (again). */
5081 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5082 goto out;
5084 on_rq = p->se.on_rq;
5085 if (on_rq)
5086 deactivate_task(rq_src, p, 0);
5088 set_task_cpu(p, dest_cpu);
5089 if (on_rq) {
5090 activate_task(rq_dest, p, 0);
5091 check_preempt_curr(rq_dest, p);
5093 ret = 1;
5094 out:
5095 double_rq_unlock(rq_src, rq_dest);
5096 return ret;
5100 * migration_thread - this is a highprio system thread that performs
5101 * thread migration by bumping thread off CPU then 'pushing' onto
5102 * another runqueue.
5104 static int migration_thread(void *data)
5106 int cpu = (long)data;
5107 struct rq *rq;
5109 rq = cpu_rq(cpu);
5110 BUG_ON(rq->migration_thread != current);
5112 set_current_state(TASK_INTERRUPTIBLE);
5113 while (!kthread_should_stop()) {
5114 struct migration_req *req;
5115 struct list_head *head;
5117 spin_lock_irq(&rq->lock);
5119 if (cpu_is_offline(cpu)) {
5120 spin_unlock_irq(&rq->lock);
5121 goto wait_to_die;
5124 if (rq->active_balance) {
5125 active_load_balance(rq, cpu);
5126 rq->active_balance = 0;
5129 head = &rq->migration_queue;
5131 if (list_empty(head)) {
5132 spin_unlock_irq(&rq->lock);
5133 schedule();
5134 set_current_state(TASK_INTERRUPTIBLE);
5135 continue;
5137 req = list_entry(head->next, struct migration_req, list);
5138 list_del_init(head->next);
5140 spin_unlock(&rq->lock);
5141 __migrate_task(req->task, cpu, req->dest_cpu);
5142 local_irq_enable();
5144 complete(&req->done);
5146 __set_current_state(TASK_RUNNING);
5147 return 0;
5149 wait_to_die:
5150 /* Wait for kthread_stop */
5151 set_current_state(TASK_INTERRUPTIBLE);
5152 while (!kthread_should_stop()) {
5153 schedule();
5154 set_current_state(TASK_INTERRUPTIBLE);
5156 __set_current_state(TASK_RUNNING);
5157 return 0;
5160 #ifdef CONFIG_HOTPLUG_CPU
5162 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5164 int ret;
5166 local_irq_disable();
5167 ret = __migrate_task(p, src_cpu, dest_cpu);
5168 local_irq_enable();
5169 return ret;
5173 * Figure out where task on dead CPU should go, use force if necessary.
5174 * NOTE: interrupts should be disabled by the caller
5176 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5178 unsigned long flags;
5179 cpumask_t mask;
5180 struct rq *rq;
5181 int dest_cpu;
5183 do {
5184 /* On same node? */
5185 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5186 cpus_and(mask, mask, p->cpus_allowed);
5187 dest_cpu = any_online_cpu(mask);
5189 /* On any allowed CPU? */
5190 if (dest_cpu == NR_CPUS)
5191 dest_cpu = any_online_cpu(p->cpus_allowed);
5193 /* No more Mr. Nice Guy. */
5194 if (dest_cpu == NR_CPUS) {
5195 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5197 * Try to stay on the same cpuset, where the
5198 * current cpuset may be a subset of all cpus.
5199 * The cpuset_cpus_allowed_locked() variant of
5200 * cpuset_cpus_allowed() will not block. It must be
5201 * called within calls to cpuset_lock/cpuset_unlock.
5203 rq = task_rq_lock(p, &flags);
5204 p->cpus_allowed = cpus_allowed;
5205 dest_cpu = any_online_cpu(p->cpus_allowed);
5206 task_rq_unlock(rq, &flags);
5209 * Don't tell them about moving exiting tasks or
5210 * kernel threads (both mm NULL), since they never
5211 * leave kernel.
5213 if (p->mm && printk_ratelimit())
5214 printk(KERN_INFO "process %d (%s) no "
5215 "longer affine to cpu%d\n",
5216 task_pid_nr(p), p->comm, dead_cpu);
5218 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5222 * While a dead CPU has no uninterruptible tasks queued at this point,
5223 * it might still have a nonzero ->nr_uninterruptible counter, because
5224 * for performance reasons the counter is not stricly tracking tasks to
5225 * their home CPUs. So we just add the counter to another CPU's counter,
5226 * to keep the global sum constant after CPU-down:
5228 static void migrate_nr_uninterruptible(struct rq *rq_src)
5230 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5231 unsigned long flags;
5233 local_irq_save(flags);
5234 double_rq_lock(rq_src, rq_dest);
5235 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5236 rq_src->nr_uninterruptible = 0;
5237 double_rq_unlock(rq_src, rq_dest);
5238 local_irq_restore(flags);
5241 /* Run through task list and migrate tasks from the dead cpu. */
5242 static void migrate_live_tasks(int src_cpu)
5244 struct task_struct *p, *t;
5246 read_lock(&tasklist_lock);
5248 do_each_thread(t, p) {
5249 if (p == current)
5250 continue;
5252 if (task_cpu(p) == src_cpu)
5253 move_task_off_dead_cpu(src_cpu, p);
5254 } while_each_thread(t, p);
5256 read_unlock(&tasklist_lock);
5260 * activate_idle_task - move idle task to the _front_ of runqueue.
5262 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5264 update_rq_clock(rq);
5266 if (p->state == TASK_UNINTERRUPTIBLE)
5267 rq->nr_uninterruptible--;
5269 enqueue_task(rq, p, 0);
5270 inc_nr_running(p, rq);
5274 * Schedules idle task to be the next runnable task on current CPU.
5275 * It does so by boosting its priority to highest possible and adding it to
5276 * the _front_ of the runqueue. Used by CPU offline code.
5278 void sched_idle_next(void)
5280 int this_cpu = smp_processor_id();
5281 struct rq *rq = cpu_rq(this_cpu);
5282 struct task_struct *p = rq->idle;
5283 unsigned long flags;
5285 /* cpu has to be offline */
5286 BUG_ON(cpu_online(this_cpu));
5289 * Strictly not necessary since rest of the CPUs are stopped by now
5290 * and interrupts disabled on the current cpu.
5292 spin_lock_irqsave(&rq->lock, flags);
5294 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5296 /* Add idle task to the _front_ of its priority queue: */
5297 activate_idle_task(p, rq);
5299 spin_unlock_irqrestore(&rq->lock, flags);
5303 * Ensures that the idle task is using init_mm right before its cpu goes
5304 * offline.
5306 void idle_task_exit(void)
5308 struct mm_struct *mm = current->active_mm;
5310 BUG_ON(cpu_online(smp_processor_id()));
5312 if (mm != &init_mm)
5313 switch_mm(mm, &init_mm, current);
5314 mmdrop(mm);
5317 /* called under rq->lock with disabled interrupts */
5318 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5320 struct rq *rq = cpu_rq(dead_cpu);
5322 /* Must be exiting, otherwise would be on tasklist. */
5323 BUG_ON(!p->exit_state);
5325 /* Cannot have done final schedule yet: would have vanished. */
5326 BUG_ON(p->state == TASK_DEAD);
5328 get_task_struct(p);
5331 * Drop lock around migration; if someone else moves it,
5332 * that's OK. No task can be added to this CPU, so iteration is
5333 * fine.
5335 spin_unlock_irq(&rq->lock);
5336 move_task_off_dead_cpu(dead_cpu, p);
5337 spin_lock_irq(&rq->lock);
5339 put_task_struct(p);
5342 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5343 static void migrate_dead_tasks(unsigned int dead_cpu)
5345 struct rq *rq = cpu_rq(dead_cpu);
5346 struct task_struct *next;
5348 for ( ; ; ) {
5349 if (!rq->nr_running)
5350 break;
5351 update_rq_clock(rq);
5352 next = pick_next_task(rq, rq->curr);
5353 if (!next)
5354 break;
5355 migrate_dead(dead_cpu, next);
5359 #endif /* CONFIG_HOTPLUG_CPU */
5361 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5363 static struct ctl_table sd_ctl_dir[] = {
5365 .procname = "sched_domain",
5366 .mode = 0555,
5368 {0,},
5371 static struct ctl_table sd_ctl_root[] = {
5373 .ctl_name = CTL_KERN,
5374 .procname = "kernel",
5375 .mode = 0555,
5376 .child = sd_ctl_dir,
5378 {0,},
5381 static struct ctl_table *sd_alloc_ctl_entry(int n)
5383 struct ctl_table *entry =
5384 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5386 return entry;
5389 static void sd_free_ctl_entry(struct ctl_table **tablep)
5391 struct ctl_table *entry;
5394 * In the intermediate directories, both the child directory and
5395 * procname are dynamically allocated and could fail but the mode
5396 * will always be set. In the lowest directory the names are
5397 * static strings and all have proc handlers.
5399 for (entry = *tablep; entry->mode; entry++) {
5400 if (entry->child)
5401 sd_free_ctl_entry(&entry->child);
5402 if (entry->proc_handler == NULL)
5403 kfree(entry->procname);
5406 kfree(*tablep);
5407 *tablep = NULL;
5410 static void
5411 set_table_entry(struct ctl_table *entry,
5412 const char *procname, void *data, int maxlen,
5413 mode_t mode, proc_handler *proc_handler)
5415 entry->procname = procname;
5416 entry->data = data;
5417 entry->maxlen = maxlen;
5418 entry->mode = mode;
5419 entry->proc_handler = proc_handler;
5422 static struct ctl_table *
5423 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5425 struct ctl_table *table = sd_alloc_ctl_entry(12);
5427 if (table == NULL)
5428 return NULL;
5430 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5431 sizeof(long), 0644, proc_doulongvec_minmax);
5432 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5433 sizeof(long), 0644, proc_doulongvec_minmax);
5434 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5435 sizeof(int), 0644, proc_dointvec_minmax);
5436 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5437 sizeof(int), 0644, proc_dointvec_minmax);
5438 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5439 sizeof(int), 0644, proc_dointvec_minmax);
5440 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5441 sizeof(int), 0644, proc_dointvec_minmax);
5442 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5443 sizeof(int), 0644, proc_dointvec_minmax);
5444 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5445 sizeof(int), 0644, proc_dointvec_minmax);
5446 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5447 sizeof(int), 0644, proc_dointvec_minmax);
5448 set_table_entry(&table[9], "cache_nice_tries",
5449 &sd->cache_nice_tries,
5450 sizeof(int), 0644, proc_dointvec_minmax);
5451 set_table_entry(&table[10], "flags", &sd->flags,
5452 sizeof(int), 0644, proc_dointvec_minmax);
5453 /* &table[11] is terminator */
5455 return table;
5458 static ctl_table * sd_alloc_ctl_cpu_table(int cpu)
5460 struct ctl_table *entry, *table;
5461 struct sched_domain *sd;
5462 int domain_num = 0, i;
5463 char buf[32];
5465 for_each_domain(cpu, sd)
5466 domain_num++;
5467 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5468 if (table == NULL)
5469 return NULL;
5471 i = 0;
5472 for_each_domain(cpu, sd) {
5473 snprintf(buf, 32, "domain%d", i);
5474 entry->procname = kstrdup(buf, GFP_KERNEL);
5475 entry->mode = 0555;
5476 entry->child = sd_alloc_ctl_domain_table(sd);
5477 entry++;
5478 i++;
5480 return table;
5483 static struct ctl_table_header *sd_sysctl_header;
5484 static void register_sched_domain_sysctl(void)
5486 int i, cpu_num = num_online_cpus();
5487 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5488 char buf[32];
5490 WARN_ON(sd_ctl_dir[0].child);
5491 sd_ctl_dir[0].child = entry;
5493 if (entry == NULL)
5494 return;
5496 for_each_online_cpu(i) {
5497 snprintf(buf, 32, "cpu%d", i);
5498 entry->procname = kstrdup(buf, GFP_KERNEL);
5499 entry->mode = 0555;
5500 entry->child = sd_alloc_ctl_cpu_table(i);
5501 entry++;
5504 WARN_ON(sd_sysctl_header);
5505 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5508 /* may be called multiple times per register */
5509 static void unregister_sched_domain_sysctl(void)
5511 if (sd_sysctl_header)
5512 unregister_sysctl_table(sd_sysctl_header);
5513 sd_sysctl_header = NULL;
5514 if (sd_ctl_dir[0].child)
5515 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5517 #else
5518 static void register_sched_domain_sysctl(void)
5521 static void unregister_sched_domain_sysctl(void)
5524 #endif
5527 * migration_call - callback that gets triggered when a CPU is added.
5528 * Here we can start up the necessary migration thread for the new CPU.
5530 static int __cpuinit
5531 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5533 struct task_struct *p;
5534 int cpu = (long)hcpu;
5535 unsigned long flags;
5536 struct rq *rq;
5538 switch (action) {
5539 case CPU_LOCK_ACQUIRE:
5540 mutex_lock(&sched_hotcpu_mutex);
5541 break;
5543 case CPU_UP_PREPARE:
5544 case CPU_UP_PREPARE_FROZEN:
5545 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5546 if (IS_ERR(p))
5547 return NOTIFY_BAD;
5548 kthread_bind(p, cpu);
5549 /* Must be high prio: stop_machine expects to yield to it. */
5550 rq = task_rq_lock(p, &flags);
5551 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5552 task_rq_unlock(rq, &flags);
5553 cpu_rq(cpu)->migration_thread = p;
5554 break;
5556 case CPU_ONLINE:
5557 case CPU_ONLINE_FROZEN:
5558 /* Strictly unnecessary, as first user will wake it. */
5559 wake_up_process(cpu_rq(cpu)->migration_thread);
5560 break;
5562 #ifdef CONFIG_HOTPLUG_CPU
5563 case CPU_UP_CANCELED:
5564 case CPU_UP_CANCELED_FROZEN:
5565 if (!cpu_rq(cpu)->migration_thread)
5566 break;
5567 /* Unbind it from offline cpu so it can run. Fall thru. */
5568 kthread_bind(cpu_rq(cpu)->migration_thread,
5569 any_online_cpu(cpu_online_map));
5570 kthread_stop(cpu_rq(cpu)->migration_thread);
5571 cpu_rq(cpu)->migration_thread = NULL;
5572 break;
5574 case CPU_DEAD:
5575 case CPU_DEAD_FROZEN:
5576 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5577 migrate_live_tasks(cpu);
5578 rq = cpu_rq(cpu);
5579 kthread_stop(rq->migration_thread);
5580 rq->migration_thread = NULL;
5581 /* Idle task back to normal (off runqueue, low prio) */
5582 spin_lock_irq(&rq->lock);
5583 update_rq_clock(rq);
5584 deactivate_task(rq, rq->idle, 0);
5585 rq->idle->static_prio = MAX_PRIO;
5586 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5587 rq->idle->sched_class = &idle_sched_class;
5588 migrate_dead_tasks(cpu);
5589 spin_unlock_irq(&rq->lock);
5590 cpuset_unlock();
5591 migrate_nr_uninterruptible(rq);
5592 BUG_ON(rq->nr_running != 0);
5594 /* No need to migrate the tasks: it was best-effort if
5595 * they didn't take sched_hotcpu_mutex. Just wake up
5596 * the requestors. */
5597 spin_lock_irq(&rq->lock);
5598 while (!list_empty(&rq->migration_queue)) {
5599 struct migration_req *req;
5601 req = list_entry(rq->migration_queue.next,
5602 struct migration_req, list);
5603 list_del_init(&req->list);
5604 complete(&req->done);
5606 spin_unlock_irq(&rq->lock);
5607 break;
5608 #endif
5609 case CPU_LOCK_RELEASE:
5610 mutex_unlock(&sched_hotcpu_mutex);
5611 break;
5613 return NOTIFY_OK;
5616 /* Register at highest priority so that task migration (migrate_all_tasks)
5617 * happens before everything else.
5619 static struct notifier_block __cpuinitdata migration_notifier = {
5620 .notifier_call = migration_call,
5621 .priority = 10
5624 int __init migration_init(void)
5626 void *cpu = (void *)(long)smp_processor_id();
5627 int err;
5629 /* Start one for the boot CPU: */
5630 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5631 BUG_ON(err == NOTIFY_BAD);
5632 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5633 register_cpu_notifier(&migration_notifier);
5635 return 0;
5637 #endif
5639 #ifdef CONFIG_SMP
5641 /* Number of possible processor ids */
5642 int nr_cpu_ids __read_mostly = NR_CPUS;
5643 EXPORT_SYMBOL(nr_cpu_ids);
5645 #ifdef CONFIG_SCHED_DEBUG
5647 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5649 struct sched_group *group = sd->groups;
5650 cpumask_t groupmask;
5651 char str[NR_CPUS];
5653 cpumask_scnprintf(str, NR_CPUS, sd->span);
5654 cpus_clear(groupmask);
5656 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5658 if (!(sd->flags & SD_LOAD_BALANCE)) {
5659 printk("does not load-balance\n");
5660 if (sd->parent)
5661 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5662 " has parent");
5663 return -1;
5666 printk(KERN_CONT "span %s\n", str);
5668 if (!cpu_isset(cpu, sd->span)) {
5669 printk(KERN_ERR "ERROR: domain->span does not contain "
5670 "CPU%d\n", cpu);
5672 if (!cpu_isset(cpu, group->cpumask)) {
5673 printk(KERN_ERR "ERROR: domain->groups does not contain"
5674 " CPU%d\n", cpu);
5677 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5678 do {
5679 if (!group) {
5680 printk("\n");
5681 printk(KERN_ERR "ERROR: group is NULL\n");
5682 break;
5685 if (!group->__cpu_power) {
5686 printk(KERN_CONT "\n");
5687 printk(KERN_ERR "ERROR: domain->cpu_power not "
5688 "set\n");
5689 break;
5692 if (!cpus_weight(group->cpumask)) {
5693 printk(KERN_CONT "\n");
5694 printk(KERN_ERR "ERROR: empty group\n");
5695 break;
5698 if (cpus_intersects(groupmask, group->cpumask)) {
5699 printk(KERN_CONT "\n");
5700 printk(KERN_ERR "ERROR: repeated CPUs\n");
5701 break;
5704 cpus_or(groupmask, groupmask, group->cpumask);
5706 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5707 printk(KERN_CONT " %s", str);
5709 group = group->next;
5710 } while (group != sd->groups);
5711 printk(KERN_CONT "\n");
5713 if (!cpus_equal(sd->span, groupmask))
5714 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5716 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5717 printk(KERN_ERR "ERROR: parent span is not a superset "
5718 "of domain->span\n");
5719 return 0;
5722 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5724 int level = 0;
5726 if (!sd) {
5727 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5728 return;
5731 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5733 for (;;) {
5734 if (sched_domain_debug_one(sd, cpu, level))
5735 break;
5736 level++;
5737 sd = sd->parent;
5738 if (!sd)
5739 break;
5742 #else
5743 # define sched_domain_debug(sd, cpu) do { } while (0)
5744 #endif
5746 static int sd_degenerate(struct sched_domain *sd)
5748 if (cpus_weight(sd->span) == 1)
5749 return 1;
5751 /* Following flags need at least 2 groups */
5752 if (sd->flags & (SD_LOAD_BALANCE |
5753 SD_BALANCE_NEWIDLE |
5754 SD_BALANCE_FORK |
5755 SD_BALANCE_EXEC |
5756 SD_SHARE_CPUPOWER |
5757 SD_SHARE_PKG_RESOURCES)) {
5758 if (sd->groups != sd->groups->next)
5759 return 0;
5762 /* Following flags don't use groups */
5763 if (sd->flags & (SD_WAKE_IDLE |
5764 SD_WAKE_AFFINE |
5765 SD_WAKE_BALANCE))
5766 return 0;
5768 return 1;
5771 static int
5772 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5774 unsigned long cflags = sd->flags, pflags = parent->flags;
5776 if (sd_degenerate(parent))
5777 return 1;
5779 if (!cpus_equal(sd->span, parent->span))
5780 return 0;
5782 /* Does parent contain flags not in child? */
5783 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5784 if (cflags & SD_WAKE_AFFINE)
5785 pflags &= ~SD_WAKE_BALANCE;
5786 /* Flags needing groups don't count if only 1 group in parent */
5787 if (parent->groups == parent->groups->next) {
5788 pflags &= ~(SD_LOAD_BALANCE |
5789 SD_BALANCE_NEWIDLE |
5790 SD_BALANCE_FORK |
5791 SD_BALANCE_EXEC |
5792 SD_SHARE_CPUPOWER |
5793 SD_SHARE_PKG_RESOURCES);
5795 if (~cflags & pflags)
5796 return 0;
5798 return 1;
5802 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5803 * hold the hotplug lock.
5805 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5807 struct rq *rq = cpu_rq(cpu);
5808 struct sched_domain *tmp;
5810 /* Remove the sched domains which do not contribute to scheduling. */
5811 for (tmp = sd; tmp; tmp = tmp->parent) {
5812 struct sched_domain *parent = tmp->parent;
5813 if (!parent)
5814 break;
5815 if (sd_parent_degenerate(tmp, parent)) {
5816 tmp->parent = parent->parent;
5817 if (parent->parent)
5818 parent->parent->child = tmp;
5822 if (sd && sd_degenerate(sd)) {
5823 sd = sd->parent;
5824 if (sd)
5825 sd->child = NULL;
5828 sched_domain_debug(sd, cpu);
5830 rcu_assign_pointer(rq->sd, sd);
5833 /* cpus with isolated domains */
5834 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5836 /* Setup the mask of cpus configured for isolated domains */
5837 static int __init isolated_cpu_setup(char *str)
5839 int ints[NR_CPUS], i;
5841 str = get_options(str, ARRAY_SIZE(ints), ints);
5842 cpus_clear(cpu_isolated_map);
5843 for (i = 1; i <= ints[0]; i++)
5844 if (ints[i] < NR_CPUS)
5845 cpu_set(ints[i], cpu_isolated_map);
5846 return 1;
5849 __setup("isolcpus=", isolated_cpu_setup);
5852 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5853 * to a function which identifies what group(along with sched group) a CPU
5854 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5855 * (due to the fact that we keep track of groups covered with a cpumask_t).
5857 * init_sched_build_groups will build a circular linked list of the groups
5858 * covered by the given span, and will set each group's ->cpumask correctly,
5859 * and ->cpu_power to 0.
5861 static void
5862 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5863 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5864 struct sched_group **sg))
5866 struct sched_group *first = NULL, *last = NULL;
5867 cpumask_t covered = CPU_MASK_NONE;
5868 int i;
5870 for_each_cpu_mask(i, span) {
5871 struct sched_group *sg;
5872 int group = group_fn(i, cpu_map, &sg);
5873 int j;
5875 if (cpu_isset(i, covered))
5876 continue;
5878 sg->cpumask = CPU_MASK_NONE;
5879 sg->__cpu_power = 0;
5881 for_each_cpu_mask(j, span) {
5882 if (group_fn(j, cpu_map, NULL) != group)
5883 continue;
5885 cpu_set(j, covered);
5886 cpu_set(j, sg->cpumask);
5888 if (!first)
5889 first = sg;
5890 if (last)
5891 last->next = sg;
5892 last = sg;
5894 last->next = first;
5897 #define SD_NODES_PER_DOMAIN 16
5899 #ifdef CONFIG_NUMA
5902 * find_next_best_node - find the next node to include in a sched_domain
5903 * @node: node whose sched_domain we're building
5904 * @used_nodes: nodes already in the sched_domain
5906 * Find the next node to include in a given scheduling domain. Simply
5907 * finds the closest node not already in the @used_nodes map.
5909 * Should use nodemask_t.
5911 static int find_next_best_node(int node, unsigned long *used_nodes)
5913 int i, n, val, min_val, best_node = 0;
5915 min_val = INT_MAX;
5917 for (i = 0; i < MAX_NUMNODES; i++) {
5918 /* Start at @node */
5919 n = (node + i) % MAX_NUMNODES;
5921 if (!nr_cpus_node(n))
5922 continue;
5924 /* Skip already used nodes */
5925 if (test_bit(n, used_nodes))
5926 continue;
5928 /* Simple min distance search */
5929 val = node_distance(node, n);
5931 if (val < min_val) {
5932 min_val = val;
5933 best_node = n;
5937 set_bit(best_node, used_nodes);
5938 return best_node;
5942 * sched_domain_node_span - get a cpumask for a node's sched_domain
5943 * @node: node whose cpumask we're constructing
5944 * @size: number of nodes to include in this span
5946 * Given a node, construct a good cpumask for its sched_domain to span. It
5947 * should be one that prevents unnecessary balancing, but also spreads tasks
5948 * out optimally.
5950 static cpumask_t sched_domain_node_span(int node)
5952 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5953 cpumask_t span, nodemask;
5954 int i;
5956 cpus_clear(span);
5957 bitmap_zero(used_nodes, MAX_NUMNODES);
5959 nodemask = node_to_cpumask(node);
5960 cpus_or(span, span, nodemask);
5961 set_bit(node, used_nodes);
5963 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5964 int next_node = find_next_best_node(node, used_nodes);
5966 nodemask = node_to_cpumask(next_node);
5967 cpus_or(span, span, nodemask);
5970 return span;
5972 #endif
5974 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5977 * SMT sched-domains:
5979 #ifdef CONFIG_SCHED_SMT
5980 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5981 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5983 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5984 struct sched_group **sg)
5986 if (sg)
5987 *sg = &per_cpu(sched_group_cpus, cpu);
5988 return cpu;
5990 #endif
5993 * multi-core sched-domains:
5995 #ifdef CONFIG_SCHED_MC
5996 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5997 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5998 #endif
6000 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6001 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6002 struct sched_group **sg)
6004 int group;
6005 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6006 cpus_and(mask, mask, *cpu_map);
6007 group = first_cpu(mask);
6008 if (sg)
6009 *sg = &per_cpu(sched_group_core, group);
6010 return group;
6012 #elif defined(CONFIG_SCHED_MC)
6013 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6014 struct sched_group **sg)
6016 if (sg)
6017 *sg = &per_cpu(sched_group_core, cpu);
6018 return cpu;
6020 #endif
6022 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6023 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6025 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6026 struct sched_group **sg)
6028 int group;
6029 #ifdef CONFIG_SCHED_MC
6030 cpumask_t mask = cpu_coregroup_map(cpu);
6031 cpus_and(mask, mask, *cpu_map);
6032 group = first_cpu(mask);
6033 #elif defined(CONFIG_SCHED_SMT)
6034 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6035 cpus_and(mask, mask, *cpu_map);
6036 group = first_cpu(mask);
6037 #else
6038 group = cpu;
6039 #endif
6040 if (sg)
6041 *sg = &per_cpu(sched_group_phys, group);
6042 return group;
6045 #ifdef CONFIG_NUMA
6047 * The init_sched_build_groups can't handle what we want to do with node
6048 * groups, so roll our own. Now each node has its own list of groups which
6049 * gets dynamically allocated.
6051 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6052 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6054 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6055 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6057 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6058 struct sched_group **sg)
6060 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6061 int group;
6063 cpus_and(nodemask, nodemask, *cpu_map);
6064 group = first_cpu(nodemask);
6066 if (sg)
6067 *sg = &per_cpu(sched_group_allnodes, group);
6068 return group;
6071 static void init_numa_sched_groups_power(struct sched_group *group_head)
6073 struct sched_group *sg = group_head;
6074 int j;
6076 if (!sg)
6077 return;
6078 do {
6079 for_each_cpu_mask(j, sg->cpumask) {
6080 struct sched_domain *sd;
6082 sd = &per_cpu(phys_domains, j);
6083 if (j != first_cpu(sd->groups->cpumask)) {
6085 * Only add "power" once for each
6086 * physical package.
6088 continue;
6091 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6093 sg = sg->next;
6094 } while (sg != group_head);
6096 #endif
6098 #ifdef CONFIG_NUMA
6099 /* Free memory allocated for various sched_group structures */
6100 static void free_sched_groups(const cpumask_t *cpu_map)
6102 int cpu, i;
6104 for_each_cpu_mask(cpu, *cpu_map) {
6105 struct sched_group **sched_group_nodes
6106 = sched_group_nodes_bycpu[cpu];
6108 if (!sched_group_nodes)
6109 continue;
6111 for (i = 0; i < MAX_NUMNODES; i++) {
6112 cpumask_t nodemask = node_to_cpumask(i);
6113 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6115 cpus_and(nodemask, nodemask, *cpu_map);
6116 if (cpus_empty(nodemask))
6117 continue;
6119 if (sg == NULL)
6120 continue;
6121 sg = sg->next;
6122 next_sg:
6123 oldsg = sg;
6124 sg = sg->next;
6125 kfree(oldsg);
6126 if (oldsg != sched_group_nodes[i])
6127 goto next_sg;
6129 kfree(sched_group_nodes);
6130 sched_group_nodes_bycpu[cpu] = NULL;
6133 #else
6134 static void free_sched_groups(const cpumask_t *cpu_map)
6137 #endif
6140 * Initialize sched groups cpu_power.
6142 * cpu_power indicates the capacity of sched group, which is used while
6143 * distributing the load between different sched groups in a sched domain.
6144 * Typically cpu_power for all the groups in a sched domain will be same unless
6145 * there are asymmetries in the topology. If there are asymmetries, group
6146 * having more cpu_power will pickup more load compared to the group having
6147 * less cpu_power.
6149 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6150 * the maximum number of tasks a group can handle in the presence of other idle
6151 * or lightly loaded groups in the same sched domain.
6153 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6155 struct sched_domain *child;
6156 struct sched_group *group;
6158 WARN_ON(!sd || !sd->groups);
6160 if (cpu != first_cpu(sd->groups->cpumask))
6161 return;
6163 child = sd->child;
6165 sd->groups->__cpu_power = 0;
6168 * For perf policy, if the groups in child domain share resources
6169 * (for example cores sharing some portions of the cache hierarchy
6170 * or SMT), then set this domain groups cpu_power such that each group
6171 * can handle only one task, when there are other idle groups in the
6172 * same sched domain.
6174 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6175 (child->flags &
6176 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6177 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6178 return;
6182 * add cpu_power of each child group to this groups cpu_power
6184 group = child->groups;
6185 do {
6186 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6187 group = group->next;
6188 } while (group != child->groups);
6192 * Build sched domains for a given set of cpus and attach the sched domains
6193 * to the individual cpus
6195 static int build_sched_domains(const cpumask_t *cpu_map)
6197 int i;
6198 #ifdef CONFIG_NUMA
6199 struct sched_group **sched_group_nodes = NULL;
6200 int sd_allnodes = 0;
6203 * Allocate the per-node list of sched groups
6205 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6206 GFP_KERNEL);
6207 if (!sched_group_nodes) {
6208 printk(KERN_WARNING "Can not alloc sched group node list\n");
6209 return -ENOMEM;
6211 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6212 #endif
6215 * Set up domains for cpus specified by the cpu_map.
6217 for_each_cpu_mask(i, *cpu_map) {
6218 struct sched_domain *sd = NULL, *p;
6219 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6221 cpus_and(nodemask, nodemask, *cpu_map);
6223 #ifdef CONFIG_NUMA
6224 if (cpus_weight(*cpu_map) >
6225 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6226 sd = &per_cpu(allnodes_domains, i);
6227 *sd = SD_ALLNODES_INIT;
6228 sd->span = *cpu_map;
6229 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6230 p = sd;
6231 sd_allnodes = 1;
6232 } else
6233 p = NULL;
6235 sd = &per_cpu(node_domains, i);
6236 *sd = SD_NODE_INIT;
6237 sd->span = sched_domain_node_span(cpu_to_node(i));
6238 sd->parent = p;
6239 if (p)
6240 p->child = sd;
6241 cpus_and(sd->span, sd->span, *cpu_map);
6242 #endif
6244 p = sd;
6245 sd = &per_cpu(phys_domains, i);
6246 *sd = SD_CPU_INIT;
6247 sd->span = nodemask;
6248 sd->parent = p;
6249 if (p)
6250 p->child = sd;
6251 cpu_to_phys_group(i, cpu_map, &sd->groups);
6253 #ifdef CONFIG_SCHED_MC
6254 p = sd;
6255 sd = &per_cpu(core_domains, i);
6256 *sd = SD_MC_INIT;
6257 sd->span = cpu_coregroup_map(i);
6258 cpus_and(sd->span, sd->span, *cpu_map);
6259 sd->parent = p;
6260 p->child = sd;
6261 cpu_to_core_group(i, cpu_map, &sd->groups);
6262 #endif
6264 #ifdef CONFIG_SCHED_SMT
6265 p = sd;
6266 sd = &per_cpu(cpu_domains, i);
6267 *sd = SD_SIBLING_INIT;
6268 sd->span = per_cpu(cpu_sibling_map, i);
6269 cpus_and(sd->span, sd->span, *cpu_map);
6270 sd->parent = p;
6271 p->child = sd;
6272 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6273 #endif
6276 #ifdef CONFIG_SCHED_SMT
6277 /* Set up CPU (sibling) groups */
6278 for_each_cpu_mask(i, *cpu_map) {
6279 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6280 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6281 if (i != first_cpu(this_sibling_map))
6282 continue;
6284 init_sched_build_groups(this_sibling_map, cpu_map,
6285 &cpu_to_cpu_group);
6287 #endif
6289 #ifdef CONFIG_SCHED_MC
6290 /* Set up multi-core groups */
6291 for_each_cpu_mask(i, *cpu_map) {
6292 cpumask_t this_core_map = cpu_coregroup_map(i);
6293 cpus_and(this_core_map, this_core_map, *cpu_map);
6294 if (i != first_cpu(this_core_map))
6295 continue;
6296 init_sched_build_groups(this_core_map, cpu_map,
6297 &cpu_to_core_group);
6299 #endif
6301 /* Set up physical groups */
6302 for (i = 0; i < MAX_NUMNODES; i++) {
6303 cpumask_t nodemask = node_to_cpumask(i);
6305 cpus_and(nodemask, nodemask, *cpu_map);
6306 if (cpus_empty(nodemask))
6307 continue;
6309 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6312 #ifdef CONFIG_NUMA
6313 /* Set up node groups */
6314 if (sd_allnodes)
6315 init_sched_build_groups(*cpu_map, cpu_map,
6316 &cpu_to_allnodes_group);
6318 for (i = 0; i < MAX_NUMNODES; i++) {
6319 /* Set up node groups */
6320 struct sched_group *sg, *prev;
6321 cpumask_t nodemask = node_to_cpumask(i);
6322 cpumask_t domainspan;
6323 cpumask_t covered = CPU_MASK_NONE;
6324 int j;
6326 cpus_and(nodemask, nodemask, *cpu_map);
6327 if (cpus_empty(nodemask)) {
6328 sched_group_nodes[i] = NULL;
6329 continue;
6332 domainspan = sched_domain_node_span(i);
6333 cpus_and(domainspan, domainspan, *cpu_map);
6335 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6336 if (!sg) {
6337 printk(KERN_WARNING "Can not alloc domain group for "
6338 "node %d\n", i);
6339 goto error;
6341 sched_group_nodes[i] = sg;
6342 for_each_cpu_mask(j, nodemask) {
6343 struct sched_domain *sd;
6345 sd = &per_cpu(node_domains, j);
6346 sd->groups = sg;
6348 sg->__cpu_power = 0;
6349 sg->cpumask = nodemask;
6350 sg->next = sg;
6351 cpus_or(covered, covered, nodemask);
6352 prev = sg;
6354 for (j = 0; j < MAX_NUMNODES; j++) {
6355 cpumask_t tmp, notcovered;
6356 int n = (i + j) % MAX_NUMNODES;
6358 cpus_complement(notcovered, covered);
6359 cpus_and(tmp, notcovered, *cpu_map);
6360 cpus_and(tmp, tmp, domainspan);
6361 if (cpus_empty(tmp))
6362 break;
6364 nodemask = node_to_cpumask(n);
6365 cpus_and(tmp, tmp, nodemask);
6366 if (cpus_empty(tmp))
6367 continue;
6369 sg = kmalloc_node(sizeof(struct sched_group),
6370 GFP_KERNEL, i);
6371 if (!sg) {
6372 printk(KERN_WARNING
6373 "Can not alloc domain group for node %d\n", j);
6374 goto error;
6376 sg->__cpu_power = 0;
6377 sg->cpumask = tmp;
6378 sg->next = prev->next;
6379 cpus_or(covered, covered, tmp);
6380 prev->next = sg;
6381 prev = sg;
6384 #endif
6386 /* Calculate CPU power for physical packages and nodes */
6387 #ifdef CONFIG_SCHED_SMT
6388 for_each_cpu_mask(i, *cpu_map) {
6389 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6391 init_sched_groups_power(i, sd);
6393 #endif
6394 #ifdef CONFIG_SCHED_MC
6395 for_each_cpu_mask(i, *cpu_map) {
6396 struct sched_domain *sd = &per_cpu(core_domains, i);
6398 init_sched_groups_power(i, sd);
6400 #endif
6402 for_each_cpu_mask(i, *cpu_map) {
6403 struct sched_domain *sd = &per_cpu(phys_domains, i);
6405 init_sched_groups_power(i, sd);
6408 #ifdef CONFIG_NUMA
6409 for (i = 0; i < MAX_NUMNODES; i++)
6410 init_numa_sched_groups_power(sched_group_nodes[i]);
6412 if (sd_allnodes) {
6413 struct sched_group *sg;
6415 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6416 init_numa_sched_groups_power(sg);
6418 #endif
6420 /* Attach the domains */
6421 for_each_cpu_mask(i, *cpu_map) {
6422 struct sched_domain *sd;
6423 #ifdef CONFIG_SCHED_SMT
6424 sd = &per_cpu(cpu_domains, i);
6425 #elif defined(CONFIG_SCHED_MC)
6426 sd = &per_cpu(core_domains, i);
6427 #else
6428 sd = &per_cpu(phys_domains, i);
6429 #endif
6430 cpu_attach_domain(sd, i);
6433 return 0;
6435 #ifdef CONFIG_NUMA
6436 error:
6437 free_sched_groups(cpu_map);
6438 return -ENOMEM;
6439 #endif
6442 static cpumask_t *doms_cur; /* current sched domains */
6443 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6446 * Special case: If a kmalloc of a doms_cur partition (array of
6447 * cpumask_t) fails, then fallback to a single sched domain,
6448 * as determined by the single cpumask_t fallback_doms.
6450 static cpumask_t fallback_doms;
6453 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6454 * For now this just excludes isolated cpus, but could be used to
6455 * exclude other special cases in the future.
6457 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6459 int err;
6461 ndoms_cur = 1;
6462 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6463 if (!doms_cur)
6464 doms_cur = &fallback_doms;
6465 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6466 err = build_sched_domains(doms_cur);
6467 register_sched_domain_sysctl();
6469 return err;
6472 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6474 free_sched_groups(cpu_map);
6478 * Detach sched domains from a group of cpus specified in cpu_map
6479 * These cpus will now be attached to the NULL domain
6481 static void detach_destroy_domains(const cpumask_t *cpu_map)
6483 int i;
6485 unregister_sched_domain_sysctl();
6487 for_each_cpu_mask(i, *cpu_map)
6488 cpu_attach_domain(NULL, i);
6489 synchronize_sched();
6490 arch_destroy_sched_domains(cpu_map);
6494 * Partition sched domains as specified by the 'ndoms_new'
6495 * cpumasks in the array doms_new[] of cpumasks. This compares
6496 * doms_new[] to the current sched domain partitioning, doms_cur[].
6497 * It destroys each deleted domain and builds each new domain.
6499 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6500 * The masks don't intersect (don't overlap.) We should setup one
6501 * sched domain for each mask. CPUs not in any of the cpumasks will
6502 * not be load balanced. If the same cpumask appears both in the
6503 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6504 * it as it is.
6506 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6507 * ownership of it and will kfree it when done with it. If the caller
6508 * failed the kmalloc call, then it can pass in doms_new == NULL,
6509 * and partition_sched_domains() will fallback to the single partition
6510 * 'fallback_doms'.
6512 * Call with hotplug lock held
6514 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6516 int i, j;
6518 /* always unregister in case we don't destroy any domains */
6519 unregister_sched_domain_sysctl();
6521 if (doms_new == NULL) {
6522 ndoms_new = 1;
6523 doms_new = &fallback_doms;
6524 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6527 /* Destroy deleted domains */
6528 for (i = 0; i < ndoms_cur; i++) {
6529 for (j = 0; j < ndoms_new; j++) {
6530 if (cpus_equal(doms_cur[i], doms_new[j]))
6531 goto match1;
6533 /* no match - a current sched domain not in new doms_new[] */
6534 detach_destroy_domains(doms_cur + i);
6535 match1:
6539 /* Build new domains */
6540 for (i = 0; i < ndoms_new; i++) {
6541 for (j = 0; j < ndoms_cur; j++) {
6542 if (cpus_equal(doms_new[i], doms_cur[j]))
6543 goto match2;
6545 /* no match - add a new doms_new */
6546 build_sched_domains(doms_new + i);
6547 match2:
6551 /* Remember the new sched domains */
6552 if (doms_cur != &fallback_doms)
6553 kfree(doms_cur);
6554 doms_cur = doms_new;
6555 ndoms_cur = ndoms_new;
6557 register_sched_domain_sysctl();
6560 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6561 static int arch_reinit_sched_domains(void)
6563 int err;
6565 mutex_lock(&sched_hotcpu_mutex);
6566 detach_destroy_domains(&cpu_online_map);
6567 err = arch_init_sched_domains(&cpu_online_map);
6568 mutex_unlock(&sched_hotcpu_mutex);
6570 return err;
6573 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6575 int ret;
6577 if (buf[0] != '0' && buf[0] != '1')
6578 return -EINVAL;
6580 if (smt)
6581 sched_smt_power_savings = (buf[0] == '1');
6582 else
6583 sched_mc_power_savings = (buf[0] == '1');
6585 ret = arch_reinit_sched_domains();
6587 return ret ? ret : count;
6590 #ifdef CONFIG_SCHED_MC
6591 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6593 return sprintf(page, "%u\n", sched_mc_power_savings);
6595 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6596 const char *buf, size_t count)
6598 return sched_power_savings_store(buf, count, 0);
6600 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6601 sched_mc_power_savings_store);
6602 #endif
6604 #ifdef CONFIG_SCHED_SMT
6605 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6607 return sprintf(page, "%u\n", sched_smt_power_savings);
6609 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6610 const char *buf, size_t count)
6612 return sched_power_savings_store(buf, count, 1);
6614 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6615 sched_smt_power_savings_store);
6616 #endif
6618 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6620 int err = 0;
6622 #ifdef CONFIG_SCHED_SMT
6623 if (smt_capable())
6624 err = sysfs_create_file(&cls->kset.kobj,
6625 &attr_sched_smt_power_savings.attr);
6626 #endif
6627 #ifdef CONFIG_SCHED_MC
6628 if (!err && mc_capable())
6629 err = sysfs_create_file(&cls->kset.kobj,
6630 &attr_sched_mc_power_savings.attr);
6631 #endif
6632 return err;
6634 #endif
6637 * Force a reinitialization of the sched domains hierarchy. The domains
6638 * and groups cannot be updated in place without racing with the balancing
6639 * code, so we temporarily attach all running cpus to the NULL domain
6640 * which will prevent rebalancing while the sched domains are recalculated.
6642 static int update_sched_domains(struct notifier_block *nfb,
6643 unsigned long action, void *hcpu)
6645 switch (action) {
6646 case CPU_UP_PREPARE:
6647 case CPU_UP_PREPARE_FROZEN:
6648 case CPU_DOWN_PREPARE:
6649 case CPU_DOWN_PREPARE_FROZEN:
6650 detach_destroy_domains(&cpu_online_map);
6651 return NOTIFY_OK;
6653 case CPU_UP_CANCELED:
6654 case CPU_UP_CANCELED_FROZEN:
6655 case CPU_DOWN_FAILED:
6656 case CPU_DOWN_FAILED_FROZEN:
6657 case CPU_ONLINE:
6658 case CPU_ONLINE_FROZEN:
6659 case CPU_DEAD:
6660 case CPU_DEAD_FROZEN:
6662 * Fall through and re-initialise the domains.
6664 break;
6665 default:
6666 return NOTIFY_DONE;
6669 /* The hotplug lock is already held by cpu_up/cpu_down */
6670 arch_init_sched_domains(&cpu_online_map);
6672 return NOTIFY_OK;
6675 void __init sched_init_smp(void)
6677 cpumask_t non_isolated_cpus;
6679 mutex_lock(&sched_hotcpu_mutex);
6680 arch_init_sched_domains(&cpu_online_map);
6681 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6682 if (cpus_empty(non_isolated_cpus))
6683 cpu_set(smp_processor_id(), non_isolated_cpus);
6684 mutex_unlock(&sched_hotcpu_mutex);
6685 /* XXX: Theoretical race here - CPU may be hotplugged now */
6686 hotcpu_notifier(update_sched_domains, 0);
6688 /* Move init over to a non-isolated CPU */
6689 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6690 BUG();
6692 #else
6693 void __init sched_init_smp(void)
6696 #endif /* CONFIG_SMP */
6698 int in_sched_functions(unsigned long addr)
6700 /* Linker adds these: start and end of __sched functions */
6701 extern char __sched_text_start[], __sched_text_end[];
6703 return in_lock_functions(addr) ||
6704 (addr >= (unsigned long)__sched_text_start
6705 && addr < (unsigned long)__sched_text_end);
6708 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6710 cfs_rq->tasks_timeline = RB_ROOT;
6711 #ifdef CONFIG_FAIR_GROUP_SCHED
6712 cfs_rq->rq = rq;
6713 #endif
6714 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6717 void __init sched_init(void)
6719 int highest_cpu = 0;
6720 int i, j;
6722 for_each_possible_cpu(i) {
6723 struct rt_prio_array *array;
6724 struct rq *rq;
6726 rq = cpu_rq(i);
6727 spin_lock_init(&rq->lock);
6728 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6729 rq->nr_running = 0;
6730 rq->clock = 1;
6731 init_cfs_rq(&rq->cfs, rq);
6732 #ifdef CONFIG_FAIR_GROUP_SCHED
6733 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6735 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6736 struct sched_entity *se =
6737 &per_cpu(init_sched_entity, i);
6739 init_cfs_rq_p[i] = cfs_rq;
6740 init_cfs_rq(cfs_rq, rq);
6741 cfs_rq->tg = &init_task_group;
6742 list_add(&cfs_rq->leaf_cfs_rq_list,
6743 &rq->leaf_cfs_rq_list);
6745 init_sched_entity_p[i] = se;
6746 se->cfs_rq = &rq->cfs;
6747 se->my_q = cfs_rq;
6748 se->load.weight = init_task_group_load;
6749 se->load.inv_weight =
6750 div64_64(1ULL<<32, init_task_group_load);
6751 se->parent = NULL;
6753 init_task_group.shares = init_task_group_load;
6754 spin_lock_init(&init_task_group.lock);
6755 #endif
6757 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6758 rq->cpu_load[j] = 0;
6759 #ifdef CONFIG_SMP
6760 rq->sd = NULL;
6761 rq->active_balance = 0;
6762 rq->next_balance = jiffies;
6763 rq->push_cpu = 0;
6764 rq->cpu = i;
6765 rq->migration_thread = NULL;
6766 INIT_LIST_HEAD(&rq->migration_queue);
6767 #endif
6768 atomic_set(&rq->nr_iowait, 0);
6770 array = &rq->rt.active;
6771 for (j = 0; j < MAX_RT_PRIO; j++) {
6772 INIT_LIST_HEAD(array->queue + j);
6773 __clear_bit(j, array->bitmap);
6775 highest_cpu = i;
6776 /* delimiter for bitsearch: */
6777 __set_bit(MAX_RT_PRIO, array->bitmap);
6780 set_load_weight(&init_task);
6782 #ifdef CONFIG_PREEMPT_NOTIFIERS
6783 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6784 #endif
6786 #ifdef CONFIG_SMP
6787 nr_cpu_ids = highest_cpu + 1;
6788 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6789 #endif
6791 #ifdef CONFIG_RT_MUTEXES
6792 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6793 #endif
6796 * The boot idle thread does lazy MMU switching as well:
6798 atomic_inc(&init_mm.mm_count);
6799 enter_lazy_tlb(&init_mm, current);
6802 * Make us the idle thread. Technically, schedule() should not be
6803 * called from this thread, however somewhere below it might be,
6804 * but because we are the idle thread, we just pick up running again
6805 * when this runqueue becomes "idle".
6807 init_idle(current, smp_processor_id());
6809 * During early bootup we pretend to be a normal task:
6811 current->sched_class = &fair_sched_class;
6814 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6815 void __might_sleep(char *file, int line)
6817 #ifdef in_atomic
6818 static unsigned long prev_jiffy; /* ratelimiting */
6820 if ((in_atomic() || irqs_disabled()) &&
6821 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6822 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6823 return;
6824 prev_jiffy = jiffies;
6825 printk(KERN_ERR "BUG: sleeping function called from invalid"
6826 " context at %s:%d\n", file, line);
6827 printk("in_atomic():%d, irqs_disabled():%d\n",
6828 in_atomic(), irqs_disabled());
6829 debug_show_held_locks(current);
6830 if (irqs_disabled())
6831 print_irqtrace_events(current);
6832 dump_stack();
6834 #endif
6836 EXPORT_SYMBOL(__might_sleep);
6837 #endif
6839 #ifdef CONFIG_MAGIC_SYSRQ
6840 static void normalize_task(struct rq *rq, struct task_struct *p)
6842 int on_rq;
6843 update_rq_clock(rq);
6844 on_rq = p->se.on_rq;
6845 if (on_rq)
6846 deactivate_task(rq, p, 0);
6847 __setscheduler(rq, p, SCHED_NORMAL, 0);
6848 if (on_rq) {
6849 activate_task(rq, p, 0);
6850 resched_task(rq->curr);
6854 void normalize_rt_tasks(void)
6856 struct task_struct *g, *p;
6857 unsigned long flags;
6858 struct rq *rq;
6860 read_lock_irq(&tasklist_lock);
6861 do_each_thread(g, p) {
6863 * Only normalize user tasks:
6865 if (!p->mm)
6866 continue;
6868 p->se.exec_start = 0;
6869 #ifdef CONFIG_SCHEDSTATS
6870 p->se.wait_start = 0;
6871 p->se.sleep_start = 0;
6872 p->se.block_start = 0;
6873 #endif
6874 task_rq(p)->clock = 0;
6876 if (!rt_task(p)) {
6878 * Renice negative nice level userspace
6879 * tasks back to 0:
6881 if (TASK_NICE(p) < 0 && p->mm)
6882 set_user_nice(p, 0);
6883 continue;
6886 spin_lock_irqsave(&p->pi_lock, flags);
6887 rq = __task_rq_lock(p);
6889 normalize_task(rq, p);
6891 __task_rq_unlock(rq);
6892 spin_unlock_irqrestore(&p->pi_lock, flags);
6893 } while_each_thread(g, p);
6895 read_unlock_irq(&tasklist_lock);
6898 #endif /* CONFIG_MAGIC_SYSRQ */
6900 #ifdef CONFIG_IA64
6902 * These functions are only useful for the IA64 MCA handling.
6904 * They can only be called when the whole system has been
6905 * stopped - every CPU needs to be quiescent, and no scheduling
6906 * activity can take place. Using them for anything else would
6907 * be a serious bug, and as a result, they aren't even visible
6908 * under any other configuration.
6912 * curr_task - return the current task for a given cpu.
6913 * @cpu: the processor in question.
6915 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6917 struct task_struct *curr_task(int cpu)
6919 return cpu_curr(cpu);
6923 * set_curr_task - set the current task for a given cpu.
6924 * @cpu: the processor in question.
6925 * @p: the task pointer to set.
6927 * Description: This function must only be used when non-maskable interrupts
6928 * are serviced on a separate stack. It allows the architecture to switch the
6929 * notion of the current task on a cpu in a non-blocking manner. This function
6930 * must be called with all CPU's synchronized, and interrupts disabled, the
6931 * and caller must save the original value of the current task (see
6932 * curr_task() above) and restore that value before reenabling interrupts and
6933 * re-starting the system.
6935 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6937 void set_curr_task(int cpu, struct task_struct *p)
6939 cpu_curr(cpu) = p;
6942 #endif
6944 #ifdef CONFIG_FAIR_GROUP_SCHED
6946 /* allocate runqueue etc for a new task group */
6947 struct task_group *sched_create_group(void)
6949 struct task_group *tg;
6950 struct cfs_rq *cfs_rq;
6951 struct sched_entity *se;
6952 struct rq *rq;
6953 int i;
6955 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6956 if (!tg)
6957 return ERR_PTR(-ENOMEM);
6959 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6960 if (!tg->cfs_rq)
6961 goto err;
6962 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6963 if (!tg->se)
6964 goto err;
6966 for_each_possible_cpu(i) {
6967 rq = cpu_rq(i);
6969 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6970 cpu_to_node(i));
6971 if (!cfs_rq)
6972 goto err;
6974 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6975 cpu_to_node(i));
6976 if (!se)
6977 goto err;
6979 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6980 memset(se, 0, sizeof(struct sched_entity));
6982 tg->cfs_rq[i] = cfs_rq;
6983 init_cfs_rq(cfs_rq, rq);
6984 cfs_rq->tg = tg;
6986 tg->se[i] = se;
6987 se->cfs_rq = &rq->cfs;
6988 se->my_q = cfs_rq;
6989 se->load.weight = NICE_0_LOAD;
6990 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
6991 se->parent = NULL;
6994 for_each_possible_cpu(i) {
6995 rq = cpu_rq(i);
6996 cfs_rq = tg->cfs_rq[i];
6997 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7000 tg->shares = NICE_0_LOAD;
7001 spin_lock_init(&tg->lock);
7003 return tg;
7005 err:
7006 for_each_possible_cpu(i) {
7007 if (tg->cfs_rq)
7008 kfree(tg->cfs_rq[i]);
7009 if (tg->se)
7010 kfree(tg->se[i]);
7012 kfree(tg->cfs_rq);
7013 kfree(tg->se);
7014 kfree(tg);
7016 return ERR_PTR(-ENOMEM);
7019 /* rcu callback to free various structures associated with a task group */
7020 static void free_sched_group(struct rcu_head *rhp)
7022 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
7023 struct task_group *tg = cfs_rq->tg;
7024 struct sched_entity *se;
7025 int i;
7027 /* now it should be safe to free those cfs_rqs */
7028 for_each_possible_cpu(i) {
7029 cfs_rq = tg->cfs_rq[i];
7030 kfree(cfs_rq);
7032 se = tg->se[i];
7033 kfree(se);
7036 kfree(tg->cfs_rq);
7037 kfree(tg->se);
7038 kfree(tg);
7041 /* Destroy runqueue etc associated with a task group */
7042 void sched_destroy_group(struct task_group *tg)
7044 struct cfs_rq *cfs_rq;
7045 int i;
7047 for_each_possible_cpu(i) {
7048 cfs_rq = tg->cfs_rq[i];
7049 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7052 cfs_rq = tg->cfs_rq[0];
7054 /* wait for possible concurrent references to cfs_rqs complete */
7055 call_rcu(&cfs_rq->rcu, free_sched_group);
7058 /* change task's runqueue when it moves between groups.
7059 * The caller of this function should have put the task in its new group
7060 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7061 * reflect its new group.
7063 void sched_move_task(struct task_struct *tsk)
7065 int on_rq, running;
7066 unsigned long flags;
7067 struct rq *rq;
7069 rq = task_rq_lock(tsk, &flags);
7071 if (tsk->sched_class != &fair_sched_class)
7072 goto done;
7074 update_rq_clock(rq);
7076 running = task_running(rq, tsk);
7077 on_rq = tsk->se.on_rq;
7079 if (on_rq) {
7080 dequeue_task(rq, tsk, 0);
7081 if (unlikely(running))
7082 tsk->sched_class->put_prev_task(rq, tsk);
7085 set_task_cfs_rq(tsk);
7087 if (on_rq) {
7088 if (unlikely(running))
7089 tsk->sched_class->set_curr_task(rq);
7090 enqueue_task(rq, tsk, 0);
7093 done:
7094 task_rq_unlock(rq, &flags);
7097 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7099 struct cfs_rq *cfs_rq = se->cfs_rq;
7100 struct rq *rq = cfs_rq->rq;
7101 int on_rq;
7103 spin_lock_irq(&rq->lock);
7105 on_rq = se->on_rq;
7106 if (on_rq)
7107 dequeue_entity(cfs_rq, se, 0);
7109 se->load.weight = shares;
7110 se->load.inv_weight = div64_64((1ULL<<32), shares);
7112 if (on_rq)
7113 enqueue_entity(cfs_rq, se, 0);
7115 spin_unlock_irq(&rq->lock);
7118 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7120 int i;
7122 spin_lock(&tg->lock);
7123 if (tg->shares == shares)
7124 goto done;
7126 tg->shares = shares;
7127 for_each_possible_cpu(i)
7128 set_se_shares(tg->se[i], shares);
7130 done:
7131 spin_unlock(&tg->lock);
7132 return 0;
7135 unsigned long sched_group_shares(struct task_group *tg)
7137 return tg->shares;
7140 #endif /* CONFIG_FAIR_GROUP_SCHED */
7142 #ifdef CONFIG_FAIR_CGROUP_SCHED
7144 /* return corresponding task_group object of a cgroup */
7145 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7147 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7148 struct task_group, css);
7151 static struct cgroup_subsys_state *
7152 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7154 struct task_group *tg;
7156 if (!cgrp->parent) {
7157 /* This is early initialization for the top cgroup */
7158 init_task_group.css.cgroup = cgrp;
7159 return &init_task_group.css;
7162 /* we support only 1-level deep hierarchical scheduler atm */
7163 if (cgrp->parent->parent)
7164 return ERR_PTR(-EINVAL);
7166 tg = sched_create_group();
7167 if (IS_ERR(tg))
7168 return ERR_PTR(-ENOMEM);
7170 /* Bind the cgroup to task_group object we just created */
7171 tg->css.cgroup = cgrp;
7173 return &tg->css;
7176 static void cpu_cgroup_destroy(struct cgroup_subsys *ss,
7177 struct cgroup *cgrp)
7179 struct task_group *tg = cgroup_tg(cgrp);
7181 sched_destroy_group(tg);
7184 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss,
7185 struct cgroup *cgrp, struct task_struct *tsk)
7187 /* We don't support RT-tasks being in separate groups */
7188 if (tsk->sched_class != &fair_sched_class)
7189 return -EINVAL;
7191 return 0;
7194 static void
7195 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7196 struct cgroup *old_cont, struct task_struct *tsk)
7198 sched_move_task(tsk);
7201 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7202 u64 shareval)
7204 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7207 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7209 struct task_group *tg = cgroup_tg(cgrp);
7211 return (u64) tg->shares;
7214 static struct cftype cpu_shares = {
7215 .name = "shares",
7216 .read_uint = cpu_shares_read_uint,
7217 .write_uint = cpu_shares_write_uint,
7220 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7222 return cgroup_add_file(cont, ss, &cpu_shares);
7225 struct cgroup_subsys cpu_cgroup_subsys = {
7226 .name = "cpu",
7227 .create = cpu_cgroup_create,
7228 .destroy = cpu_cgroup_destroy,
7229 .can_attach = cpu_cgroup_can_attach,
7230 .attach = cpu_cgroup_attach,
7231 .populate = cpu_cgroup_populate,
7232 .subsys_id = cpu_cgroup_subsys_id,
7233 .early_init = 1,
7236 #endif /* CONFIG_FAIR_CGROUP_SCHED */