4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/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>
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 * (NSEC_PER_SEC / HZ);
82 * Convert user-nice values [ -20 ... 0 ... 19 ]
83 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
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) / (NSEC_PER_SEC / HZ))
103 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / 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)
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);
137 static inline int rt_policy(int policy)
139 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
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>
163 /* task group related information */
165 #ifdef CONFIG_FAIR_CGROUP_SCHED
166 struct cgroup_subsys_state css;
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 */
178 /* Default task group's sched entity on each cpu */
179 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
180 /* Default task group's cfs_rq on each cpu */
181 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
183 static struct sched_entity *init_sched_entity_p[NR_CPUS];
184 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
186 /* Default task group.
187 * Every task in system belong to this group at bootup.
189 struct task_group init_task_group = {
190 .se = init_sched_entity_p,
191 .cfs_rq = init_cfs_rq_p,
194 #ifdef CONFIG_FAIR_USER_SCHED
195 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
197 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
200 static int init_task_group_load = INIT_TASK_GRP_LOAD;
202 /* return group to which a task belongs */
203 static inline struct task_group *task_group(struct task_struct *p)
205 struct task_group *tg;
207 #ifdef CONFIG_FAIR_USER_SCHED
209 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
210 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
211 struct task_group, css);
213 tg = &init_task_group;
219 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
220 static inline void set_task_cfs_rq(struct task_struct *p)
222 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
223 p->se.parent = task_group(p)->se[task_cpu(p)];
228 static inline void set_task_cfs_rq(struct task_struct *p) { }
230 #endif /* CONFIG_FAIR_GROUP_SCHED */
232 /* CFS-related fields in a runqueue */
234 struct load_weight load;
235 unsigned long nr_running;
240 struct rb_root tasks_timeline;
241 struct rb_node *rb_leftmost;
242 struct rb_node *rb_load_balance_curr;
243 /* 'curr' points to currently running entity on this cfs_rq.
244 * It is set to NULL otherwise (i.e when none are currently running).
246 struct sched_entity *curr;
248 unsigned long nr_spread_over;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
253 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
254 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
255 * (like users, containers etc.)
257 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
258 * list is used during load balance.
260 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
261 struct task_group *tg; /* group that "owns" this runqueue */
265 /* Real-Time classes' related field in a runqueue: */
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.
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;
292 unsigned char in_nohz_recently;
294 /* capture load from *all* tasks on this cpu: */
295 struct load_weight load;
296 unsigned long nr_load_updates;
300 #ifdef CONFIG_FAIR_GROUP_SCHED
301 /* list of leaf cfs_rq on this cpu: */
302 struct list_head leaf_cfs_rq_list;
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;
321 unsigned int clock_warps, clock_overflows;
323 unsigned int clock_deep_idle_events;
329 struct sched_domain *sd;
331 /* For active balancing */
334 /* cpu of this runqueue: */
337 struct task_struct *migration_thread;
338 struct list_head migration_queue;
341 #ifdef CONFIG_SCHEDSTATS
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;
361 unsigned int bkl_count;
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)
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());
398 * Protect against sched_clock() occasionally going backwards:
400 if (unlikely(delta < 0)) {
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;
412 rq->clock_overflows++;
414 if (unlikely(delta > rq->clock_max_delta))
415 rq->clock_max_delta = delta;
420 rq->prev_clock_raw = now;
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
451 # define const_debug static const
455 * Debugging: various feature bits
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,
465 const_debug unsigned int sysctl_sched_features =
466 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
467 SCHED_FEAT_START_DEBIT * 1 |
468 SCHED_FEAT_TREE_AVG * 0 |
469 SCHED_FEAT_APPROX_AVG * 0 |
470 SCHED_FEAT_WAKEUP_PREEMPT * 1;
472 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
475 * Number of tasks to iterate in a single balance run.
476 * Limited because this is done with IRQs disabled.
478 const_debug unsigned int sysctl_sched_nr_migrate = 32;
481 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
482 * clock constructed from sched_clock():
484 unsigned long long cpu_clock(int cpu)
486 unsigned long long now;
490 local_irq_save(flags);
494 local_irq_restore(flags);
498 EXPORT_SYMBOL_GPL(cpu_clock);
500 #ifndef prepare_arch_switch
501 # define prepare_arch_switch(next) do { } while (0)
503 #ifndef finish_arch_switch
504 # define finish_arch_switch(prev) do { } while (0)
507 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
508 static inline int task_running(struct rq *rq, struct task_struct *p)
510 return rq->curr == p;
513 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
517 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
519 #ifdef CONFIG_DEBUG_SPINLOCK
520 /* this is a valid case when another task releases the spinlock */
521 rq->lock.owner = current;
524 * If we are tracking spinlock dependencies then we have to
525 * fix up the runqueue lock - which gets 'carried over' from
528 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
530 spin_unlock_irq(&rq->lock);
533 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
534 static inline int task_running(struct rq *rq, struct task_struct *p)
539 return rq->curr == p;
543 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
547 * We can optimise this out completely for !SMP, because the
548 * SMP rebalancing from interrupt is the only thing that cares
553 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
554 spin_unlock_irq(&rq->lock);
556 spin_unlock(&rq->lock);
560 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
564 * After ->oncpu is cleared, the task can be moved to a different CPU.
565 * We must ensure this doesn't happen until the switch is completely
571 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
575 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
578 * __task_rq_lock - lock the runqueue a given task resides on.
579 * Must be called interrupts disabled.
581 static inline struct rq *__task_rq_lock(struct task_struct *p)
585 struct rq *rq = task_rq(p);
586 spin_lock(&rq->lock);
587 if (likely(rq == task_rq(p)))
589 spin_unlock(&rq->lock);
594 * task_rq_lock - lock the runqueue a given task resides on and disable
595 * interrupts. Note the ordering: we can safely lookup the task_rq without
596 * explicitly disabling preemption.
598 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
604 local_irq_save(*flags);
606 spin_lock(&rq->lock);
607 if (likely(rq == task_rq(p)))
609 spin_unlock_irqrestore(&rq->lock, *flags);
613 static void __task_rq_unlock(struct rq *rq)
616 spin_unlock(&rq->lock);
619 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
622 spin_unlock_irqrestore(&rq->lock, *flags);
626 * this_rq_lock - lock this runqueue and disable interrupts.
628 static struct rq *this_rq_lock(void)
635 spin_lock(&rq->lock);
641 * We are going deep-idle (irqs are disabled):
643 void sched_clock_idle_sleep_event(void)
645 struct rq *rq = cpu_rq(smp_processor_id());
647 spin_lock(&rq->lock);
648 __update_rq_clock(rq);
649 spin_unlock(&rq->lock);
650 rq->clock_deep_idle_events++;
652 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
655 * We just idled delta nanoseconds (called with irqs disabled):
657 void sched_clock_idle_wakeup_event(u64 delta_ns)
659 struct rq *rq = cpu_rq(smp_processor_id());
660 u64 now = sched_clock();
662 rq->idle_clock += delta_ns;
664 * Override the previous timestamp and ignore all
665 * sched_clock() deltas that occured while we idled,
666 * and use the PM-provided delta_ns to advance the
669 spin_lock(&rq->lock);
670 rq->prev_clock_raw = now;
671 rq->clock += delta_ns;
672 spin_unlock(&rq->lock);
674 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
677 * resched_task - mark a task 'to be rescheduled now'.
679 * On UP this means the setting of the need_resched flag, on SMP it
680 * might also involve a cross-CPU call to trigger the scheduler on
685 #ifndef tsk_is_polling
686 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
689 static void resched_task(struct task_struct *p)
693 assert_spin_locked(&task_rq(p)->lock);
695 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
698 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
701 if (cpu == smp_processor_id())
704 /* NEED_RESCHED must be visible before we test polling */
706 if (!tsk_is_polling(p))
707 smp_send_reschedule(cpu);
710 static void resched_cpu(int cpu)
712 struct rq *rq = cpu_rq(cpu);
715 if (!spin_trylock_irqsave(&rq->lock, flags))
717 resched_task(cpu_curr(cpu));
718 spin_unlock_irqrestore(&rq->lock, flags);
721 static inline void resched_task(struct task_struct *p)
723 assert_spin_locked(&task_rq(p)->lock);
724 set_tsk_need_resched(p);
728 #if BITS_PER_LONG == 32
729 # define WMULT_CONST (~0UL)
731 # define WMULT_CONST (1UL << 32)
734 #define WMULT_SHIFT 32
737 * Shift right and round:
739 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
742 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
743 struct load_weight *lw)
747 if (unlikely(!lw->inv_weight))
748 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
750 tmp = (u64)delta_exec * weight;
752 * Check whether we'd overflow the 64-bit multiplication:
754 if (unlikely(tmp > WMULT_CONST))
755 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
758 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
760 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
763 static inline unsigned long
764 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
766 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
769 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
774 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
780 * To aid in avoiding the subversion of "niceness" due to uneven distribution
781 * of tasks with abnormal "nice" values across CPUs the contribution that
782 * each task makes to its run queue's load is weighted according to its
783 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
784 * scaled version of the new time slice allocation that they receive on time
788 #define WEIGHT_IDLEPRIO 2
789 #define WMULT_IDLEPRIO (1 << 31)
792 * Nice levels are multiplicative, with a gentle 10% change for every
793 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
794 * nice 1, it will get ~10% less CPU time than another CPU-bound task
795 * that remained on nice 0.
797 * The "10% effect" is relative and cumulative: from _any_ nice level,
798 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
799 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
800 * If a task goes up by ~10% and another task goes down by ~10% then
801 * the relative distance between them is ~25%.)
803 static const int prio_to_weight[40] = {
804 /* -20 */ 88761, 71755, 56483, 46273, 36291,
805 /* -15 */ 29154, 23254, 18705, 14949, 11916,
806 /* -10 */ 9548, 7620, 6100, 4904, 3906,
807 /* -5 */ 3121, 2501, 1991, 1586, 1277,
808 /* 0 */ 1024, 820, 655, 526, 423,
809 /* 5 */ 335, 272, 215, 172, 137,
810 /* 10 */ 110, 87, 70, 56, 45,
811 /* 15 */ 36, 29, 23, 18, 15,
815 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
817 * In cases where the weight does not change often, we can use the
818 * precalculated inverse to speed up arithmetics by turning divisions
819 * into multiplications:
821 static const u32 prio_to_wmult[40] = {
822 /* -20 */ 48388, 59856, 76040, 92818, 118348,
823 /* -15 */ 147320, 184698, 229616, 287308, 360437,
824 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
825 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
826 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
827 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
828 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
829 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
832 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
835 * runqueue iterator, to support SMP load-balancing between different
836 * scheduling classes, without having to expose their internal data
837 * structures to the load-balancing proper:
841 struct task_struct *(*start)(void *);
842 struct task_struct *(*next)(void *);
847 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
848 unsigned long max_load_move, struct sched_domain *sd,
849 enum cpu_idle_type idle, int *all_pinned,
850 int *this_best_prio, struct rq_iterator *iterator);
853 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
854 struct sched_domain *sd, enum cpu_idle_type idle,
855 struct rq_iterator *iterator);
858 #include "sched_stats.h"
859 #include "sched_idletask.c"
860 #include "sched_fair.c"
861 #include "sched_rt.c"
862 #ifdef CONFIG_SCHED_DEBUG
863 # include "sched_debug.c"
866 #define sched_class_highest (&rt_sched_class)
869 * Update delta_exec, delta_fair fields for rq.
871 * delta_fair clock advances at a rate inversely proportional to
872 * total load (rq->load.weight) on the runqueue, while
873 * delta_exec advances at the same rate as wall-clock (provided
876 * delta_exec / delta_fair is a measure of the (smoothened) load on this
877 * runqueue over any given interval. This (smoothened) load is used
878 * during load balance.
880 * This function is called /before/ updating rq->load
881 * and when switching tasks.
883 static inline void inc_load(struct rq *rq, const struct task_struct *p)
885 update_load_add(&rq->load, p->se.load.weight);
888 static inline void dec_load(struct rq *rq, const struct task_struct *p)
890 update_load_sub(&rq->load, p->se.load.weight);
893 static void inc_nr_running(struct task_struct *p, struct rq *rq)
899 static void dec_nr_running(struct task_struct *p, struct rq *rq)
905 static void set_load_weight(struct task_struct *p)
907 if (task_has_rt_policy(p)) {
908 p->se.load.weight = prio_to_weight[0] * 2;
909 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
914 * SCHED_IDLE tasks get minimal weight:
916 if (p->policy == SCHED_IDLE) {
917 p->se.load.weight = WEIGHT_IDLEPRIO;
918 p->se.load.inv_weight = WMULT_IDLEPRIO;
922 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
923 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
926 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
928 sched_info_queued(p);
929 p->sched_class->enqueue_task(rq, p, wakeup);
933 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
935 p->sched_class->dequeue_task(rq, p, sleep);
940 * __normal_prio - return the priority that is based on the static prio
942 static inline int __normal_prio(struct task_struct *p)
944 return p->static_prio;
948 * Calculate the expected normal priority: i.e. priority
949 * without taking RT-inheritance into account. Might be
950 * boosted by interactivity modifiers. Changes upon fork,
951 * setprio syscalls, and whenever the interactivity
952 * estimator recalculates.
954 static inline int normal_prio(struct task_struct *p)
958 if (task_has_rt_policy(p))
959 prio = MAX_RT_PRIO-1 - p->rt_priority;
961 prio = __normal_prio(p);
966 * Calculate the current priority, i.e. the priority
967 * taken into account by the scheduler. This value might
968 * be boosted by RT tasks, or might be boosted by
969 * interactivity modifiers. Will be RT if the task got
970 * RT-boosted. If not then it returns p->normal_prio.
972 static int effective_prio(struct task_struct *p)
974 p->normal_prio = normal_prio(p);
976 * If we are RT tasks or we were boosted to RT priority,
977 * keep the priority unchanged. Otherwise, update priority
978 * to the normal priority:
980 if (!rt_prio(p->prio))
981 return p->normal_prio;
986 * activate_task - move a task to the runqueue.
988 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
990 if (p->state == TASK_UNINTERRUPTIBLE)
991 rq->nr_uninterruptible--;
993 enqueue_task(rq, p, wakeup);
994 inc_nr_running(p, rq);
998 * deactivate_task - remove a task from the runqueue.
1000 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1002 if (p->state == TASK_UNINTERRUPTIBLE)
1003 rq->nr_uninterruptible++;
1005 dequeue_task(rq, p, sleep);
1006 dec_nr_running(p, rq);
1010 * task_curr - is this task currently executing on a CPU?
1011 * @p: the task in question.
1013 inline int task_curr(const struct task_struct *p)
1015 return cpu_curr(task_cpu(p)) == p;
1018 /* Used instead of source_load when we know the type == 0 */
1019 unsigned long weighted_cpuload(const int cpu)
1021 return cpu_rq(cpu)->load.weight;
1024 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1027 task_thread_info(p)->cpu = cpu;
1035 * Is this task likely cache-hot:
1038 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1042 if (p->sched_class != &fair_sched_class)
1045 if (sysctl_sched_migration_cost == -1)
1047 if (sysctl_sched_migration_cost == 0)
1050 delta = now - p->se.exec_start;
1052 return delta < (s64)sysctl_sched_migration_cost;
1056 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1058 int old_cpu = task_cpu(p);
1059 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1060 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1061 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1064 clock_offset = old_rq->clock - new_rq->clock;
1066 #ifdef CONFIG_SCHEDSTATS
1067 if (p->se.wait_start)
1068 p->se.wait_start -= clock_offset;
1069 if (p->se.sleep_start)
1070 p->se.sleep_start -= clock_offset;
1071 if (p->se.block_start)
1072 p->se.block_start -= clock_offset;
1073 if (old_cpu != new_cpu) {
1074 schedstat_inc(p, se.nr_migrations);
1075 if (task_hot(p, old_rq->clock, NULL))
1076 schedstat_inc(p, se.nr_forced2_migrations);
1079 p->se.vruntime -= old_cfsrq->min_vruntime -
1080 new_cfsrq->min_vruntime;
1082 __set_task_cpu(p, new_cpu);
1085 struct migration_req {
1086 struct list_head list;
1088 struct task_struct *task;
1091 struct completion done;
1095 * The task's runqueue lock must be held.
1096 * Returns true if you have to wait for migration thread.
1099 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1101 struct rq *rq = task_rq(p);
1104 * If the task is not on a runqueue (and not running), then
1105 * it is sufficient to simply update the task's cpu field.
1107 if (!p->se.on_rq && !task_running(rq, p)) {
1108 set_task_cpu(p, dest_cpu);
1112 init_completion(&req->done);
1114 req->dest_cpu = dest_cpu;
1115 list_add(&req->list, &rq->migration_queue);
1121 * wait_task_inactive - wait for a thread to unschedule.
1123 * The caller must ensure that the task *will* unschedule sometime soon,
1124 * else this function might spin for a *long* time. This function can't
1125 * be called with interrupts off, or it may introduce deadlock with
1126 * smp_call_function() if an IPI is sent by the same process we are
1127 * waiting to become inactive.
1129 void wait_task_inactive(struct task_struct *p)
1131 unsigned long flags;
1137 * We do the initial early heuristics without holding
1138 * any task-queue locks at all. We'll only try to get
1139 * the runqueue lock when things look like they will
1145 * If the task is actively running on another CPU
1146 * still, just relax and busy-wait without holding
1149 * NOTE! Since we don't hold any locks, it's not
1150 * even sure that "rq" stays as the right runqueue!
1151 * But we don't care, since "task_running()" will
1152 * return false if the runqueue has changed and p
1153 * is actually now running somewhere else!
1155 while (task_running(rq, p))
1159 * Ok, time to look more closely! We need the rq
1160 * lock now, to be *sure*. If we're wrong, we'll
1161 * just go back and repeat.
1163 rq = task_rq_lock(p, &flags);
1164 running = task_running(rq, p);
1165 on_rq = p->se.on_rq;
1166 task_rq_unlock(rq, &flags);
1169 * Was it really running after all now that we
1170 * checked with the proper locks actually held?
1172 * Oops. Go back and try again..
1174 if (unlikely(running)) {
1180 * It's not enough that it's not actively running,
1181 * it must be off the runqueue _entirely_, and not
1184 * So if it wa still runnable (but just not actively
1185 * running right now), it's preempted, and we should
1186 * yield - it could be a while.
1188 if (unlikely(on_rq)) {
1189 schedule_timeout_uninterruptible(1);
1194 * Ahh, all good. It wasn't running, and it wasn't
1195 * runnable, which means that it will never become
1196 * running in the future either. We're all done!
1203 * kick_process - kick a running thread to enter/exit the kernel
1204 * @p: the to-be-kicked thread
1206 * Cause a process which is running on another CPU to enter
1207 * kernel-mode, without any delay. (to get signals handled.)
1209 * NOTE: this function doesnt have to take the runqueue lock,
1210 * because all it wants to ensure is that the remote task enters
1211 * the kernel. If the IPI races and the task has been migrated
1212 * to another CPU then no harm is done and the purpose has been
1215 void kick_process(struct task_struct *p)
1221 if ((cpu != smp_processor_id()) && task_curr(p))
1222 smp_send_reschedule(cpu);
1227 * Return a low guess at the load of a migration-source cpu weighted
1228 * according to the scheduling class and "nice" value.
1230 * We want to under-estimate the load of migration sources, to
1231 * balance conservatively.
1233 static unsigned long source_load(int cpu, int type)
1235 struct rq *rq = cpu_rq(cpu);
1236 unsigned long total = weighted_cpuload(cpu);
1241 return min(rq->cpu_load[type-1], total);
1245 * Return a high guess at the load of a migration-target cpu weighted
1246 * according to the scheduling class and "nice" value.
1248 static unsigned long target_load(int cpu, int type)
1250 struct rq *rq = cpu_rq(cpu);
1251 unsigned long total = weighted_cpuload(cpu);
1256 return max(rq->cpu_load[type-1], total);
1260 * Return the average load per task on the cpu's run queue
1262 static inline unsigned long cpu_avg_load_per_task(int cpu)
1264 struct rq *rq = cpu_rq(cpu);
1265 unsigned long total = weighted_cpuload(cpu);
1266 unsigned long n = rq->nr_running;
1268 return n ? total / n : SCHED_LOAD_SCALE;
1272 * find_idlest_group finds and returns the least busy CPU group within the
1275 static struct sched_group *
1276 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1278 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1279 unsigned long min_load = ULONG_MAX, this_load = 0;
1280 int load_idx = sd->forkexec_idx;
1281 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1284 unsigned long load, avg_load;
1288 /* Skip over this group if it has no CPUs allowed */
1289 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1292 local_group = cpu_isset(this_cpu, group->cpumask);
1294 /* Tally up the load of all CPUs in the group */
1297 for_each_cpu_mask(i, group->cpumask) {
1298 /* Bias balancing toward cpus of our domain */
1300 load = source_load(i, load_idx);
1302 load = target_load(i, load_idx);
1307 /* Adjust by relative CPU power of the group */
1308 avg_load = sg_div_cpu_power(group,
1309 avg_load * SCHED_LOAD_SCALE);
1312 this_load = avg_load;
1314 } else if (avg_load < min_load) {
1315 min_load = avg_load;
1318 } while (group = group->next, group != sd->groups);
1320 if (!idlest || 100*this_load < imbalance*min_load)
1326 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1329 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1332 unsigned long load, min_load = ULONG_MAX;
1336 /* Traverse only the allowed CPUs */
1337 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1339 for_each_cpu_mask(i, tmp) {
1340 load = weighted_cpuload(i);
1342 if (load < min_load || (load == min_load && i == this_cpu)) {
1352 * sched_balance_self: balance the current task (running on cpu) in domains
1353 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1356 * Balance, ie. select the least loaded group.
1358 * Returns the target CPU number, or the same CPU if no balancing is needed.
1360 * preempt must be disabled.
1362 static int sched_balance_self(int cpu, int flag)
1364 struct task_struct *t = current;
1365 struct sched_domain *tmp, *sd = NULL;
1367 for_each_domain(cpu, tmp) {
1369 * If power savings logic is enabled for a domain, stop there.
1371 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1373 if (tmp->flags & flag)
1379 struct sched_group *group;
1380 int new_cpu, weight;
1382 if (!(sd->flags & flag)) {
1388 group = find_idlest_group(sd, t, cpu);
1394 new_cpu = find_idlest_cpu(group, t, cpu);
1395 if (new_cpu == -1 || new_cpu == cpu) {
1396 /* Now try balancing at a lower domain level of cpu */
1401 /* Now try balancing at a lower domain level of new_cpu */
1404 weight = cpus_weight(span);
1405 for_each_domain(cpu, tmp) {
1406 if (weight <= cpus_weight(tmp->span))
1408 if (tmp->flags & flag)
1411 /* while loop will break here if sd == NULL */
1417 #endif /* CONFIG_SMP */
1420 * wake_idle() will wake a task on an idle cpu if task->cpu is
1421 * not idle and an idle cpu is available. The span of cpus to
1422 * search starts with cpus closest then further out as needed,
1423 * so we always favor a closer, idle cpu.
1425 * Returns the CPU we should wake onto.
1427 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1428 static int wake_idle(int cpu, struct task_struct *p)
1431 struct sched_domain *sd;
1435 * If it is idle, then it is the best cpu to run this task.
1437 * This cpu is also the best, if it has more than one task already.
1438 * Siblings must be also busy(in most cases) as they didn't already
1439 * pickup the extra load from this cpu and hence we need not check
1440 * sibling runqueue info. This will avoid the checks and cache miss
1441 * penalities associated with that.
1443 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1446 for_each_domain(cpu, sd) {
1447 if (sd->flags & SD_WAKE_IDLE) {
1448 cpus_and(tmp, sd->span, p->cpus_allowed);
1449 for_each_cpu_mask(i, tmp) {
1451 if (i != task_cpu(p)) {
1453 se.nr_wakeups_idle);
1465 static inline int wake_idle(int cpu, struct task_struct *p)
1472 * try_to_wake_up - wake up a thread
1473 * @p: the to-be-woken-up thread
1474 * @state: the mask of task states that can be woken
1475 * @sync: do a synchronous wakeup?
1477 * Put it on the run-queue if it's not already there. The "current"
1478 * thread is always on the run-queue (except when the actual
1479 * re-schedule is in progress), and as such you're allowed to do
1480 * the simpler "current->state = TASK_RUNNING" to mark yourself
1481 * runnable without the overhead of this.
1483 * returns failure only if the task is already active.
1485 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1487 int cpu, orig_cpu, this_cpu, success = 0;
1488 unsigned long flags;
1492 struct sched_domain *sd, *this_sd = NULL;
1493 unsigned long load, this_load;
1497 rq = task_rq_lock(p, &flags);
1498 old_state = p->state;
1499 if (!(old_state & state))
1507 this_cpu = smp_processor_id();
1510 if (unlikely(task_running(rq, p)))
1515 schedstat_inc(rq, ttwu_count);
1516 if (cpu == this_cpu) {
1517 schedstat_inc(rq, ttwu_local);
1521 for_each_domain(this_cpu, sd) {
1522 if (cpu_isset(cpu, sd->span)) {
1523 schedstat_inc(sd, ttwu_wake_remote);
1529 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1533 * Check for affine wakeup and passive balancing possibilities.
1536 int idx = this_sd->wake_idx;
1537 unsigned int imbalance;
1539 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1541 load = source_load(cpu, idx);
1542 this_load = target_load(this_cpu, idx);
1544 new_cpu = this_cpu; /* Wake to this CPU if we can */
1546 if (this_sd->flags & SD_WAKE_AFFINE) {
1547 unsigned long tl = this_load;
1548 unsigned long tl_per_task;
1551 * Attract cache-cold tasks on sync wakeups:
1553 if (sync && !task_hot(p, rq->clock, this_sd))
1556 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1557 tl_per_task = cpu_avg_load_per_task(this_cpu);
1560 * If sync wakeup then subtract the (maximum possible)
1561 * effect of the currently running task from the load
1562 * of the current CPU:
1565 tl -= current->se.load.weight;
1568 tl + target_load(cpu, idx) <= tl_per_task) ||
1569 100*(tl + p->se.load.weight) <= imbalance*load) {
1571 * This domain has SD_WAKE_AFFINE and
1572 * p is cache cold in this domain, and
1573 * there is no bad imbalance.
1575 schedstat_inc(this_sd, ttwu_move_affine);
1576 schedstat_inc(p, se.nr_wakeups_affine);
1582 * Start passive balancing when half the imbalance_pct
1585 if (this_sd->flags & SD_WAKE_BALANCE) {
1586 if (imbalance*this_load <= 100*load) {
1587 schedstat_inc(this_sd, ttwu_move_balance);
1588 schedstat_inc(p, se.nr_wakeups_passive);
1594 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1596 new_cpu = wake_idle(new_cpu, p);
1597 if (new_cpu != cpu) {
1598 set_task_cpu(p, new_cpu);
1599 task_rq_unlock(rq, &flags);
1600 /* might preempt at this point */
1601 rq = task_rq_lock(p, &flags);
1602 old_state = p->state;
1603 if (!(old_state & state))
1608 this_cpu = smp_processor_id();
1613 #endif /* CONFIG_SMP */
1614 schedstat_inc(p, se.nr_wakeups);
1616 schedstat_inc(p, se.nr_wakeups_sync);
1617 if (orig_cpu != cpu)
1618 schedstat_inc(p, se.nr_wakeups_migrate);
1619 if (cpu == this_cpu)
1620 schedstat_inc(p, se.nr_wakeups_local);
1622 schedstat_inc(p, se.nr_wakeups_remote);
1623 update_rq_clock(rq);
1624 activate_task(rq, p, 1);
1625 check_preempt_curr(rq, p);
1629 p->state = TASK_RUNNING;
1631 task_rq_unlock(rq, &flags);
1636 int fastcall wake_up_process(struct task_struct *p)
1638 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1639 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1641 EXPORT_SYMBOL(wake_up_process);
1643 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1645 return try_to_wake_up(p, state, 0);
1649 * Perform scheduler related setup for a newly forked process p.
1650 * p is forked by current.
1652 * __sched_fork() is basic setup used by init_idle() too:
1654 static void __sched_fork(struct task_struct *p)
1656 p->se.exec_start = 0;
1657 p->se.sum_exec_runtime = 0;
1658 p->se.prev_sum_exec_runtime = 0;
1660 #ifdef CONFIG_SCHEDSTATS
1661 p->se.wait_start = 0;
1662 p->se.sum_sleep_runtime = 0;
1663 p->se.sleep_start = 0;
1664 p->se.block_start = 0;
1665 p->se.sleep_max = 0;
1666 p->se.block_max = 0;
1668 p->se.slice_max = 0;
1672 INIT_LIST_HEAD(&p->run_list);
1675 #ifdef CONFIG_PREEMPT_NOTIFIERS
1676 INIT_HLIST_HEAD(&p->preempt_notifiers);
1680 * We mark the process as running here, but have not actually
1681 * inserted it onto the runqueue yet. This guarantees that
1682 * nobody will actually run it, and a signal or other external
1683 * event cannot wake it up and insert it on the runqueue either.
1685 p->state = TASK_RUNNING;
1689 * fork()/clone()-time setup:
1691 void sched_fork(struct task_struct *p, int clone_flags)
1693 int cpu = get_cpu();
1698 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1700 set_task_cpu(p, cpu);
1703 * Make sure we do not leak PI boosting priority to the child:
1705 p->prio = current->normal_prio;
1706 if (!rt_prio(p->prio))
1707 p->sched_class = &fair_sched_class;
1709 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1710 if (likely(sched_info_on()))
1711 memset(&p->sched_info, 0, sizeof(p->sched_info));
1713 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1716 #ifdef CONFIG_PREEMPT
1717 /* Want to start with kernel preemption disabled. */
1718 task_thread_info(p)->preempt_count = 1;
1724 * wake_up_new_task - wake up a newly created task for the first time.
1726 * This function will do some initial scheduler statistics housekeeping
1727 * that must be done for every newly created context, then puts the task
1728 * on the runqueue and wakes it.
1730 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1732 unsigned long flags;
1735 rq = task_rq_lock(p, &flags);
1736 BUG_ON(p->state != TASK_RUNNING);
1737 update_rq_clock(rq);
1739 p->prio = effective_prio(p);
1741 if (!p->sched_class->task_new || !current->se.on_rq) {
1742 activate_task(rq, p, 0);
1745 * Let the scheduling class do new task startup
1746 * management (if any):
1748 p->sched_class->task_new(rq, p);
1749 inc_nr_running(p, rq);
1751 check_preempt_curr(rq, p);
1752 task_rq_unlock(rq, &flags);
1755 #ifdef CONFIG_PREEMPT_NOTIFIERS
1758 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1759 * @notifier: notifier struct to register
1761 void preempt_notifier_register(struct preempt_notifier *notifier)
1763 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1765 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1768 * preempt_notifier_unregister - no longer interested in preemption notifications
1769 * @notifier: notifier struct to unregister
1771 * This is safe to call from within a preemption notifier.
1773 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1775 hlist_del(¬ifier->link);
1777 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1779 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1781 struct preempt_notifier *notifier;
1782 struct hlist_node *node;
1784 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1785 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1789 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1790 struct task_struct *next)
1792 struct preempt_notifier *notifier;
1793 struct hlist_node *node;
1795 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1796 notifier->ops->sched_out(notifier, next);
1801 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1806 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1807 struct task_struct *next)
1814 * prepare_task_switch - prepare to switch tasks
1815 * @rq: the runqueue preparing to switch
1816 * @prev: the current task that is being switched out
1817 * @next: the task we are going to switch to.
1819 * This is called with the rq lock held and interrupts off. It must
1820 * be paired with a subsequent finish_task_switch after the context
1823 * prepare_task_switch sets up locking and calls architecture specific
1827 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1828 struct task_struct *next)
1830 fire_sched_out_preempt_notifiers(prev, next);
1831 prepare_lock_switch(rq, next);
1832 prepare_arch_switch(next);
1836 * finish_task_switch - clean up after a task-switch
1837 * @rq: runqueue associated with task-switch
1838 * @prev: the thread we just switched away from.
1840 * finish_task_switch must be called after the context switch, paired
1841 * with a prepare_task_switch call before the context switch.
1842 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1843 * and do any other architecture-specific cleanup actions.
1845 * Note that we may have delayed dropping an mm in context_switch(). If
1846 * so, we finish that here outside of the runqueue lock. (Doing it
1847 * with the lock held can cause deadlocks; see schedule() for
1850 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1851 __releases(rq->lock)
1853 struct mm_struct *mm = rq->prev_mm;
1859 * A task struct has one reference for the use as "current".
1860 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1861 * schedule one last time. The schedule call will never return, and
1862 * the scheduled task must drop that reference.
1863 * The test for TASK_DEAD must occur while the runqueue locks are
1864 * still held, otherwise prev could be scheduled on another cpu, die
1865 * there before we look at prev->state, and then the reference would
1867 * Manfred Spraul <manfred@colorfullife.com>
1869 prev_state = prev->state;
1870 finish_arch_switch(prev);
1871 finish_lock_switch(rq, prev);
1872 fire_sched_in_preempt_notifiers(current);
1875 if (unlikely(prev_state == TASK_DEAD)) {
1877 * Remove function-return probe instances associated with this
1878 * task and put them back on the free list.
1880 kprobe_flush_task(prev);
1881 put_task_struct(prev);
1886 * schedule_tail - first thing a freshly forked thread must call.
1887 * @prev: the thread we just switched away from.
1889 asmlinkage void schedule_tail(struct task_struct *prev)
1890 __releases(rq->lock)
1892 struct rq *rq = this_rq();
1894 finish_task_switch(rq, prev);
1895 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1896 /* In this case, finish_task_switch does not reenable preemption */
1899 if (current->set_child_tid)
1900 put_user(task_pid_vnr(current), current->set_child_tid);
1904 * context_switch - switch to the new MM and the new
1905 * thread's register state.
1908 context_switch(struct rq *rq, struct task_struct *prev,
1909 struct task_struct *next)
1911 struct mm_struct *mm, *oldmm;
1913 prepare_task_switch(rq, prev, next);
1915 oldmm = prev->active_mm;
1917 * For paravirt, this is coupled with an exit in switch_to to
1918 * combine the page table reload and the switch backend into
1921 arch_enter_lazy_cpu_mode();
1923 if (unlikely(!mm)) {
1924 next->active_mm = oldmm;
1925 atomic_inc(&oldmm->mm_count);
1926 enter_lazy_tlb(oldmm, next);
1928 switch_mm(oldmm, mm, next);
1930 if (unlikely(!prev->mm)) {
1931 prev->active_mm = NULL;
1932 rq->prev_mm = oldmm;
1935 * Since the runqueue lock will be released by the next
1936 * task (which is an invalid locking op but in the case
1937 * of the scheduler it's an obvious special-case), so we
1938 * do an early lockdep release here:
1940 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1941 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1944 /* Here we just switch the register state and the stack. */
1945 switch_to(prev, next, prev);
1949 * this_rq must be evaluated again because prev may have moved
1950 * CPUs since it called schedule(), thus the 'rq' on its stack
1951 * frame will be invalid.
1953 finish_task_switch(this_rq(), prev);
1957 * nr_running, nr_uninterruptible and nr_context_switches:
1959 * externally visible scheduler statistics: current number of runnable
1960 * threads, current number of uninterruptible-sleeping threads, total
1961 * number of context switches performed since bootup.
1963 unsigned long nr_running(void)
1965 unsigned long i, sum = 0;
1967 for_each_online_cpu(i)
1968 sum += cpu_rq(i)->nr_running;
1973 unsigned long nr_uninterruptible(void)
1975 unsigned long i, sum = 0;
1977 for_each_possible_cpu(i)
1978 sum += cpu_rq(i)->nr_uninterruptible;
1981 * Since we read the counters lockless, it might be slightly
1982 * inaccurate. Do not allow it to go below zero though:
1984 if (unlikely((long)sum < 0))
1990 unsigned long long nr_context_switches(void)
1993 unsigned long long sum = 0;
1995 for_each_possible_cpu(i)
1996 sum += cpu_rq(i)->nr_switches;
2001 unsigned long nr_iowait(void)
2003 unsigned long i, sum = 0;
2005 for_each_possible_cpu(i)
2006 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2011 unsigned long nr_active(void)
2013 unsigned long i, running = 0, uninterruptible = 0;
2015 for_each_online_cpu(i) {
2016 running += cpu_rq(i)->nr_running;
2017 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2020 if (unlikely((long)uninterruptible < 0))
2021 uninterruptible = 0;
2023 return running + uninterruptible;
2027 * Update rq->cpu_load[] statistics. This function is usually called every
2028 * scheduler tick (TICK_NSEC).
2030 static void update_cpu_load(struct rq *this_rq)
2032 unsigned long this_load = this_rq->load.weight;
2035 this_rq->nr_load_updates++;
2037 /* Update our load: */
2038 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2039 unsigned long old_load, new_load;
2041 /* scale is effectively 1 << i now, and >> i divides by scale */
2043 old_load = this_rq->cpu_load[i];
2044 new_load = this_load;
2046 * Round up the averaging division if load is increasing. This
2047 * prevents us from getting stuck on 9 if the load is 10, for
2050 if (new_load > old_load)
2051 new_load += scale-1;
2052 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2059 * double_rq_lock - safely lock two runqueues
2061 * Note this does not disable interrupts like task_rq_lock,
2062 * you need to do so manually before calling.
2064 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2065 __acquires(rq1->lock)
2066 __acquires(rq2->lock)
2068 BUG_ON(!irqs_disabled());
2070 spin_lock(&rq1->lock);
2071 __acquire(rq2->lock); /* Fake it out ;) */
2074 spin_lock(&rq1->lock);
2075 spin_lock(&rq2->lock);
2077 spin_lock(&rq2->lock);
2078 spin_lock(&rq1->lock);
2081 update_rq_clock(rq1);
2082 update_rq_clock(rq2);
2086 * double_rq_unlock - safely unlock two runqueues
2088 * Note this does not restore interrupts like task_rq_unlock,
2089 * you need to do so manually after calling.
2091 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2092 __releases(rq1->lock)
2093 __releases(rq2->lock)
2095 spin_unlock(&rq1->lock);
2097 spin_unlock(&rq2->lock);
2099 __release(rq2->lock);
2103 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2105 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2106 __releases(this_rq->lock)
2107 __acquires(busiest->lock)
2108 __acquires(this_rq->lock)
2110 if (unlikely(!irqs_disabled())) {
2111 /* printk() doesn't work good under rq->lock */
2112 spin_unlock(&this_rq->lock);
2115 if (unlikely(!spin_trylock(&busiest->lock))) {
2116 if (busiest < this_rq) {
2117 spin_unlock(&this_rq->lock);
2118 spin_lock(&busiest->lock);
2119 spin_lock(&this_rq->lock);
2121 spin_lock(&busiest->lock);
2126 * If dest_cpu is allowed for this process, migrate the task to it.
2127 * This is accomplished by forcing the cpu_allowed mask to only
2128 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2129 * the cpu_allowed mask is restored.
2131 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2133 struct migration_req req;
2134 unsigned long flags;
2137 rq = task_rq_lock(p, &flags);
2138 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2139 || unlikely(cpu_is_offline(dest_cpu)))
2142 /* force the process onto the specified CPU */
2143 if (migrate_task(p, dest_cpu, &req)) {
2144 /* Need to wait for migration thread (might exit: take ref). */
2145 struct task_struct *mt = rq->migration_thread;
2147 get_task_struct(mt);
2148 task_rq_unlock(rq, &flags);
2149 wake_up_process(mt);
2150 put_task_struct(mt);
2151 wait_for_completion(&req.done);
2156 task_rq_unlock(rq, &flags);
2160 * sched_exec - execve() is a valuable balancing opportunity, because at
2161 * this point the task has the smallest effective memory and cache footprint.
2163 void sched_exec(void)
2165 int new_cpu, this_cpu = get_cpu();
2166 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2168 if (new_cpu != this_cpu)
2169 sched_migrate_task(current, new_cpu);
2173 * pull_task - move a task from a remote runqueue to the local runqueue.
2174 * Both runqueues must be locked.
2176 static void pull_task(struct rq *src_rq, struct task_struct *p,
2177 struct rq *this_rq, int this_cpu)
2179 deactivate_task(src_rq, p, 0);
2180 set_task_cpu(p, this_cpu);
2181 activate_task(this_rq, p, 0);
2183 * Note that idle threads have a prio of MAX_PRIO, for this test
2184 * to be always true for them.
2186 check_preempt_curr(this_rq, p);
2190 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2193 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2194 struct sched_domain *sd, enum cpu_idle_type idle,
2198 * We do not migrate tasks that are:
2199 * 1) running (obviously), or
2200 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2201 * 3) are cache-hot on their current CPU.
2203 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2204 schedstat_inc(p, se.nr_failed_migrations_affine);
2209 if (task_running(rq, p)) {
2210 schedstat_inc(p, se.nr_failed_migrations_running);
2215 * Aggressive migration if:
2216 * 1) task is cache cold, or
2217 * 2) too many balance attempts have failed.
2220 if (!task_hot(p, rq->clock, sd) ||
2221 sd->nr_balance_failed > sd->cache_nice_tries) {
2222 #ifdef CONFIG_SCHEDSTATS
2223 if (task_hot(p, rq->clock, sd)) {
2224 schedstat_inc(sd, lb_hot_gained[idle]);
2225 schedstat_inc(p, se.nr_forced_migrations);
2231 if (task_hot(p, rq->clock, sd)) {
2232 schedstat_inc(p, se.nr_failed_migrations_hot);
2238 static unsigned long
2239 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2240 unsigned long max_load_move, struct sched_domain *sd,
2241 enum cpu_idle_type idle, int *all_pinned,
2242 int *this_best_prio, struct rq_iterator *iterator)
2244 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2245 struct task_struct *p;
2246 long rem_load_move = max_load_move;
2248 if (max_load_move == 0)
2254 * Start the load-balancing iterator:
2256 p = iterator->start(iterator->arg);
2258 if (!p || loops++ > sysctl_sched_nr_migrate)
2261 * To help distribute high priority tasks across CPUs we don't
2262 * skip a task if it will be the highest priority task (i.e. smallest
2263 * prio value) on its new queue regardless of its load weight
2265 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2266 SCHED_LOAD_SCALE_FUZZ;
2267 if ((skip_for_load && p->prio >= *this_best_prio) ||
2268 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2269 p = iterator->next(iterator->arg);
2273 pull_task(busiest, p, this_rq, this_cpu);
2275 rem_load_move -= p->se.load.weight;
2278 * We only want to steal up to the prescribed amount of weighted load.
2280 if (rem_load_move > 0) {
2281 if (p->prio < *this_best_prio)
2282 *this_best_prio = p->prio;
2283 p = iterator->next(iterator->arg);
2288 * Right now, this is one of only two places pull_task() is called,
2289 * so we can safely collect pull_task() stats here rather than
2290 * inside pull_task().
2292 schedstat_add(sd, lb_gained[idle], pulled);
2295 *all_pinned = pinned;
2297 return max_load_move - rem_load_move;
2301 * move_tasks tries to move up to max_load_move weighted load from busiest to
2302 * this_rq, as part of a balancing operation within domain "sd".
2303 * Returns 1 if successful and 0 otherwise.
2305 * Called with both runqueues locked.
2307 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2308 unsigned long max_load_move,
2309 struct sched_domain *sd, enum cpu_idle_type idle,
2312 const struct sched_class *class = sched_class_highest;
2313 unsigned long total_load_moved = 0;
2314 int this_best_prio = this_rq->curr->prio;
2318 class->load_balance(this_rq, this_cpu, busiest,
2319 max_load_move - total_load_moved,
2320 sd, idle, all_pinned, &this_best_prio);
2321 class = class->next;
2322 } while (class && max_load_move > total_load_moved);
2324 return total_load_moved > 0;
2328 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2329 struct sched_domain *sd, enum cpu_idle_type idle,
2330 struct rq_iterator *iterator)
2332 struct task_struct *p = iterator->start(iterator->arg);
2336 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2337 pull_task(busiest, p, this_rq, this_cpu);
2339 * Right now, this is only the second place pull_task()
2340 * is called, so we can safely collect pull_task()
2341 * stats here rather than inside pull_task().
2343 schedstat_inc(sd, lb_gained[idle]);
2347 p = iterator->next(iterator->arg);
2354 * move_one_task tries to move exactly one task from busiest to this_rq, as
2355 * part of active balancing operations within "domain".
2356 * Returns 1 if successful and 0 otherwise.
2358 * Called with both runqueues locked.
2360 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2361 struct sched_domain *sd, enum cpu_idle_type idle)
2363 const struct sched_class *class;
2365 for (class = sched_class_highest; class; class = class->next)
2366 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2373 * find_busiest_group finds and returns the busiest CPU group within the
2374 * domain. It calculates and returns the amount of weighted load which
2375 * should be moved to restore balance via the imbalance parameter.
2377 static struct sched_group *
2378 find_busiest_group(struct sched_domain *sd, int this_cpu,
2379 unsigned long *imbalance, enum cpu_idle_type idle,
2380 int *sd_idle, cpumask_t *cpus, int *balance)
2382 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2383 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2384 unsigned long max_pull;
2385 unsigned long busiest_load_per_task, busiest_nr_running;
2386 unsigned long this_load_per_task, this_nr_running;
2387 int load_idx, group_imb = 0;
2388 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2389 int power_savings_balance = 1;
2390 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2391 unsigned long min_nr_running = ULONG_MAX;
2392 struct sched_group *group_min = NULL, *group_leader = NULL;
2395 max_load = this_load = total_load = total_pwr = 0;
2396 busiest_load_per_task = busiest_nr_running = 0;
2397 this_load_per_task = this_nr_running = 0;
2398 if (idle == CPU_NOT_IDLE)
2399 load_idx = sd->busy_idx;
2400 else if (idle == CPU_NEWLY_IDLE)
2401 load_idx = sd->newidle_idx;
2403 load_idx = sd->idle_idx;
2406 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2409 int __group_imb = 0;
2410 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2411 unsigned long sum_nr_running, sum_weighted_load;
2413 local_group = cpu_isset(this_cpu, group->cpumask);
2416 balance_cpu = first_cpu(group->cpumask);
2418 /* Tally up the load of all CPUs in the group */
2419 sum_weighted_load = sum_nr_running = avg_load = 0;
2421 min_cpu_load = ~0UL;
2423 for_each_cpu_mask(i, group->cpumask) {
2426 if (!cpu_isset(i, *cpus))
2431 if (*sd_idle && rq->nr_running)
2434 /* Bias balancing toward cpus of our domain */
2436 if (idle_cpu(i) && !first_idle_cpu) {
2441 load = target_load(i, load_idx);
2443 load = source_load(i, load_idx);
2444 if (load > max_cpu_load)
2445 max_cpu_load = load;
2446 if (min_cpu_load > load)
2447 min_cpu_load = load;
2451 sum_nr_running += rq->nr_running;
2452 sum_weighted_load += weighted_cpuload(i);
2456 * First idle cpu or the first cpu(busiest) in this sched group
2457 * is eligible for doing load balancing at this and above
2458 * domains. In the newly idle case, we will allow all the cpu's
2459 * to do the newly idle load balance.
2461 if (idle != CPU_NEWLY_IDLE && local_group &&
2462 balance_cpu != this_cpu && balance) {
2467 total_load += avg_load;
2468 total_pwr += group->__cpu_power;
2470 /* Adjust by relative CPU power of the group */
2471 avg_load = sg_div_cpu_power(group,
2472 avg_load * SCHED_LOAD_SCALE);
2474 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2477 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2480 this_load = avg_load;
2482 this_nr_running = sum_nr_running;
2483 this_load_per_task = sum_weighted_load;
2484 } else if (avg_load > max_load &&
2485 (sum_nr_running > group_capacity || __group_imb)) {
2486 max_load = avg_load;
2488 busiest_nr_running = sum_nr_running;
2489 busiest_load_per_task = sum_weighted_load;
2490 group_imb = __group_imb;
2493 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2495 * Busy processors will not participate in power savings
2498 if (idle == CPU_NOT_IDLE ||
2499 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2503 * If the local group is idle or completely loaded
2504 * no need to do power savings balance at this domain
2506 if (local_group && (this_nr_running >= group_capacity ||
2508 power_savings_balance = 0;
2511 * If a group is already running at full capacity or idle,
2512 * don't include that group in power savings calculations
2514 if (!power_savings_balance || sum_nr_running >= group_capacity
2519 * Calculate the group which has the least non-idle load.
2520 * This is the group from where we need to pick up the load
2523 if ((sum_nr_running < min_nr_running) ||
2524 (sum_nr_running == min_nr_running &&
2525 first_cpu(group->cpumask) <
2526 first_cpu(group_min->cpumask))) {
2528 min_nr_running = sum_nr_running;
2529 min_load_per_task = sum_weighted_load /
2534 * Calculate the group which is almost near its
2535 * capacity but still has some space to pick up some load
2536 * from other group and save more power
2538 if (sum_nr_running <= group_capacity - 1) {
2539 if (sum_nr_running > leader_nr_running ||
2540 (sum_nr_running == leader_nr_running &&
2541 first_cpu(group->cpumask) >
2542 first_cpu(group_leader->cpumask))) {
2543 group_leader = group;
2544 leader_nr_running = sum_nr_running;
2549 group = group->next;
2550 } while (group != sd->groups);
2552 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2555 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2557 if (this_load >= avg_load ||
2558 100*max_load <= sd->imbalance_pct*this_load)
2561 busiest_load_per_task /= busiest_nr_running;
2563 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2566 * We're trying to get all the cpus to the average_load, so we don't
2567 * want to push ourselves above the average load, nor do we wish to
2568 * reduce the max loaded cpu below the average load, as either of these
2569 * actions would just result in more rebalancing later, and ping-pong
2570 * tasks around. Thus we look for the minimum possible imbalance.
2571 * Negative imbalances (*we* are more loaded than anyone else) will
2572 * be counted as no imbalance for these purposes -- we can't fix that
2573 * by pulling tasks to us. Be careful of negative numbers as they'll
2574 * appear as very large values with unsigned longs.
2576 if (max_load <= busiest_load_per_task)
2580 * In the presence of smp nice balancing, certain scenarios can have
2581 * max load less than avg load(as we skip the groups at or below
2582 * its cpu_power, while calculating max_load..)
2584 if (max_load < avg_load) {
2586 goto small_imbalance;
2589 /* Don't want to pull so many tasks that a group would go idle */
2590 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2592 /* How much load to actually move to equalise the imbalance */
2593 *imbalance = min(max_pull * busiest->__cpu_power,
2594 (avg_load - this_load) * this->__cpu_power)
2598 * if *imbalance is less than the average load per runnable task
2599 * there is no gaurantee that any tasks will be moved so we'll have
2600 * a think about bumping its value to force at least one task to be
2603 if (*imbalance < busiest_load_per_task) {
2604 unsigned long tmp, pwr_now, pwr_move;
2608 pwr_move = pwr_now = 0;
2610 if (this_nr_running) {
2611 this_load_per_task /= this_nr_running;
2612 if (busiest_load_per_task > this_load_per_task)
2615 this_load_per_task = SCHED_LOAD_SCALE;
2617 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2618 busiest_load_per_task * imbn) {
2619 *imbalance = busiest_load_per_task;
2624 * OK, we don't have enough imbalance to justify moving tasks,
2625 * however we may be able to increase total CPU power used by
2629 pwr_now += busiest->__cpu_power *
2630 min(busiest_load_per_task, max_load);
2631 pwr_now += this->__cpu_power *
2632 min(this_load_per_task, this_load);
2633 pwr_now /= SCHED_LOAD_SCALE;
2635 /* Amount of load we'd subtract */
2636 tmp = sg_div_cpu_power(busiest,
2637 busiest_load_per_task * SCHED_LOAD_SCALE);
2639 pwr_move += busiest->__cpu_power *
2640 min(busiest_load_per_task, max_load - tmp);
2642 /* Amount of load we'd add */
2643 if (max_load * busiest->__cpu_power <
2644 busiest_load_per_task * SCHED_LOAD_SCALE)
2645 tmp = sg_div_cpu_power(this,
2646 max_load * busiest->__cpu_power);
2648 tmp = sg_div_cpu_power(this,
2649 busiest_load_per_task * SCHED_LOAD_SCALE);
2650 pwr_move += this->__cpu_power *
2651 min(this_load_per_task, this_load + tmp);
2652 pwr_move /= SCHED_LOAD_SCALE;
2654 /* Move if we gain throughput */
2655 if (pwr_move > pwr_now)
2656 *imbalance = busiest_load_per_task;
2662 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2663 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2666 if (this == group_leader && group_leader != group_min) {
2667 *imbalance = min_load_per_task;
2677 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2680 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2681 unsigned long imbalance, cpumask_t *cpus)
2683 struct rq *busiest = NULL, *rq;
2684 unsigned long max_load = 0;
2687 for_each_cpu_mask(i, group->cpumask) {
2690 if (!cpu_isset(i, *cpus))
2694 wl = weighted_cpuload(i);
2696 if (rq->nr_running == 1 && wl > imbalance)
2699 if (wl > max_load) {
2709 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2710 * so long as it is large enough.
2712 #define MAX_PINNED_INTERVAL 512
2715 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2716 * tasks if there is an imbalance.
2718 static int load_balance(int this_cpu, struct rq *this_rq,
2719 struct sched_domain *sd, enum cpu_idle_type idle,
2722 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2723 struct sched_group *group;
2724 unsigned long imbalance;
2726 cpumask_t cpus = CPU_MASK_ALL;
2727 unsigned long flags;
2730 * When power savings policy is enabled for the parent domain, idle
2731 * sibling can pick up load irrespective of busy siblings. In this case,
2732 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2733 * portraying it as CPU_NOT_IDLE.
2735 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2736 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2739 schedstat_inc(sd, lb_count[idle]);
2742 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2749 schedstat_inc(sd, lb_nobusyg[idle]);
2753 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2755 schedstat_inc(sd, lb_nobusyq[idle]);
2759 BUG_ON(busiest == this_rq);
2761 schedstat_add(sd, lb_imbalance[idle], imbalance);
2764 if (busiest->nr_running > 1) {
2766 * Attempt to move tasks. If find_busiest_group has found
2767 * an imbalance but busiest->nr_running <= 1, the group is
2768 * still unbalanced. ld_moved simply stays zero, so it is
2769 * correctly treated as an imbalance.
2771 local_irq_save(flags);
2772 double_rq_lock(this_rq, busiest);
2773 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2774 imbalance, sd, idle, &all_pinned);
2775 double_rq_unlock(this_rq, busiest);
2776 local_irq_restore(flags);
2779 * some other cpu did the load balance for us.
2781 if (ld_moved && this_cpu != smp_processor_id())
2782 resched_cpu(this_cpu);
2784 /* All tasks on this runqueue were pinned by CPU affinity */
2785 if (unlikely(all_pinned)) {
2786 cpu_clear(cpu_of(busiest), cpus);
2787 if (!cpus_empty(cpus))
2794 schedstat_inc(sd, lb_failed[idle]);
2795 sd->nr_balance_failed++;
2797 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2799 spin_lock_irqsave(&busiest->lock, flags);
2801 /* don't kick the migration_thread, if the curr
2802 * task on busiest cpu can't be moved to this_cpu
2804 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2805 spin_unlock_irqrestore(&busiest->lock, flags);
2807 goto out_one_pinned;
2810 if (!busiest->active_balance) {
2811 busiest->active_balance = 1;
2812 busiest->push_cpu = this_cpu;
2815 spin_unlock_irqrestore(&busiest->lock, flags);
2817 wake_up_process(busiest->migration_thread);
2820 * We've kicked active balancing, reset the failure
2823 sd->nr_balance_failed = sd->cache_nice_tries+1;
2826 sd->nr_balance_failed = 0;
2828 if (likely(!active_balance)) {
2829 /* We were unbalanced, so reset the balancing interval */
2830 sd->balance_interval = sd->min_interval;
2833 * If we've begun active balancing, start to back off. This
2834 * case may not be covered by the all_pinned logic if there
2835 * is only 1 task on the busy runqueue (because we don't call
2838 if (sd->balance_interval < sd->max_interval)
2839 sd->balance_interval *= 2;
2842 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2843 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2848 schedstat_inc(sd, lb_balanced[idle]);
2850 sd->nr_balance_failed = 0;
2853 /* tune up the balancing interval */
2854 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2855 (sd->balance_interval < sd->max_interval))
2856 sd->balance_interval *= 2;
2858 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2859 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2865 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2866 * tasks if there is an imbalance.
2868 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2869 * this_rq is locked.
2872 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2874 struct sched_group *group;
2875 struct rq *busiest = NULL;
2876 unsigned long imbalance;
2880 cpumask_t cpus = CPU_MASK_ALL;
2883 * When power savings policy is enabled for the parent domain, idle
2884 * sibling can pick up load irrespective of busy siblings. In this case,
2885 * let the state of idle sibling percolate up as IDLE, instead of
2886 * portraying it as CPU_NOT_IDLE.
2888 if (sd->flags & SD_SHARE_CPUPOWER &&
2889 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2892 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2894 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2895 &sd_idle, &cpus, NULL);
2897 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2901 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2904 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2908 BUG_ON(busiest == this_rq);
2910 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2913 if (busiest->nr_running > 1) {
2914 /* Attempt to move tasks */
2915 double_lock_balance(this_rq, busiest);
2916 /* this_rq->clock is already updated */
2917 update_rq_clock(busiest);
2918 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2919 imbalance, sd, CPU_NEWLY_IDLE,
2921 spin_unlock(&busiest->lock);
2923 if (unlikely(all_pinned)) {
2924 cpu_clear(cpu_of(busiest), cpus);
2925 if (!cpus_empty(cpus))
2931 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2932 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2933 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2936 sd->nr_balance_failed = 0;
2941 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2942 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2943 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2945 sd->nr_balance_failed = 0;
2951 * idle_balance is called by schedule() if this_cpu is about to become
2952 * idle. Attempts to pull tasks from other CPUs.
2954 static void idle_balance(int this_cpu, struct rq *this_rq)
2956 struct sched_domain *sd;
2957 int pulled_task = -1;
2958 unsigned long next_balance = jiffies + HZ;
2960 for_each_domain(this_cpu, sd) {
2961 unsigned long interval;
2963 if (!(sd->flags & SD_LOAD_BALANCE))
2966 if (sd->flags & SD_BALANCE_NEWIDLE)
2967 /* If we've pulled tasks over stop searching: */
2968 pulled_task = load_balance_newidle(this_cpu,
2971 interval = msecs_to_jiffies(sd->balance_interval);
2972 if (time_after(next_balance, sd->last_balance + interval))
2973 next_balance = sd->last_balance + interval;
2977 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2979 * We are going idle. next_balance may be set based on
2980 * a busy processor. So reset next_balance.
2982 this_rq->next_balance = next_balance;
2987 * active_load_balance is run by migration threads. It pushes running tasks
2988 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2989 * running on each physical CPU where possible, and avoids physical /
2990 * logical imbalances.
2992 * Called with busiest_rq locked.
2994 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2996 int target_cpu = busiest_rq->push_cpu;
2997 struct sched_domain *sd;
2998 struct rq *target_rq;
3000 /* Is there any task to move? */
3001 if (busiest_rq->nr_running <= 1)
3004 target_rq = cpu_rq(target_cpu);
3007 * This condition is "impossible", if it occurs
3008 * we need to fix it. Originally reported by
3009 * Bjorn Helgaas on a 128-cpu setup.
3011 BUG_ON(busiest_rq == target_rq);
3013 /* move a task from busiest_rq to target_rq */
3014 double_lock_balance(busiest_rq, target_rq);
3015 update_rq_clock(busiest_rq);
3016 update_rq_clock(target_rq);
3018 /* Search for an sd spanning us and the target CPU. */
3019 for_each_domain(target_cpu, sd) {
3020 if ((sd->flags & SD_LOAD_BALANCE) &&
3021 cpu_isset(busiest_cpu, sd->span))
3026 schedstat_inc(sd, alb_count);
3028 if (move_one_task(target_rq, target_cpu, busiest_rq,
3030 schedstat_inc(sd, alb_pushed);
3032 schedstat_inc(sd, alb_failed);
3034 spin_unlock(&target_rq->lock);
3039 atomic_t load_balancer;
3041 } nohz ____cacheline_aligned = {
3042 .load_balancer = ATOMIC_INIT(-1),
3043 .cpu_mask = CPU_MASK_NONE,
3047 * This routine will try to nominate the ilb (idle load balancing)
3048 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3049 * load balancing on behalf of all those cpus. If all the cpus in the system
3050 * go into this tickless mode, then there will be no ilb owner (as there is
3051 * no need for one) and all the cpus will sleep till the next wakeup event
3054 * For the ilb owner, tick is not stopped. And this tick will be used
3055 * for idle load balancing. ilb owner will still be part of
3058 * While stopping the tick, this cpu will become the ilb owner if there
3059 * is no other owner. And will be the owner till that cpu becomes busy
3060 * or if all cpus in the system stop their ticks at which point
3061 * there is no need for ilb owner.
3063 * When the ilb owner becomes busy, it nominates another owner, during the
3064 * next busy scheduler_tick()
3066 int select_nohz_load_balancer(int stop_tick)
3068 int cpu = smp_processor_id();
3071 cpu_set(cpu, nohz.cpu_mask);
3072 cpu_rq(cpu)->in_nohz_recently = 1;
3075 * If we are going offline and still the leader, give up!
3077 if (cpu_is_offline(cpu) &&
3078 atomic_read(&nohz.load_balancer) == cpu) {
3079 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3084 /* time for ilb owner also to sleep */
3085 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3086 if (atomic_read(&nohz.load_balancer) == cpu)
3087 atomic_set(&nohz.load_balancer, -1);
3091 if (atomic_read(&nohz.load_balancer) == -1) {
3092 /* make me the ilb owner */
3093 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3095 } else if (atomic_read(&nohz.load_balancer) == cpu)
3098 if (!cpu_isset(cpu, nohz.cpu_mask))
3101 cpu_clear(cpu, nohz.cpu_mask);
3103 if (atomic_read(&nohz.load_balancer) == cpu)
3104 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3111 static DEFINE_SPINLOCK(balancing);
3114 * It checks each scheduling domain to see if it is due to be balanced,
3115 * and initiates a balancing operation if so.
3117 * Balancing parameters are set up in arch_init_sched_domains.
3119 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3122 struct rq *rq = cpu_rq(cpu);
3123 unsigned long interval;
3124 struct sched_domain *sd;
3125 /* Earliest time when we have to do rebalance again */
3126 unsigned long next_balance = jiffies + 60*HZ;
3127 int update_next_balance = 0;
3129 for_each_domain(cpu, sd) {
3130 if (!(sd->flags & SD_LOAD_BALANCE))
3133 interval = sd->balance_interval;
3134 if (idle != CPU_IDLE)
3135 interval *= sd->busy_factor;
3137 /* scale ms to jiffies */
3138 interval = msecs_to_jiffies(interval);
3139 if (unlikely(!interval))
3141 if (interval > HZ*NR_CPUS/10)
3142 interval = HZ*NR_CPUS/10;
3145 if (sd->flags & SD_SERIALIZE) {
3146 if (!spin_trylock(&balancing))
3150 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3151 if (load_balance(cpu, rq, sd, idle, &balance)) {
3153 * We've pulled tasks over so either we're no
3154 * longer idle, or one of our SMT siblings is
3157 idle = CPU_NOT_IDLE;
3159 sd->last_balance = jiffies;
3161 if (sd->flags & SD_SERIALIZE)
3162 spin_unlock(&balancing);
3164 if (time_after(next_balance, sd->last_balance + interval)) {
3165 next_balance = sd->last_balance + interval;
3166 update_next_balance = 1;
3170 * Stop the load balance at this level. There is another
3171 * CPU in our sched group which is doing load balancing more
3179 * next_balance will be updated only when there is a need.
3180 * When the cpu is attached to null domain for ex, it will not be
3183 if (likely(update_next_balance))
3184 rq->next_balance = next_balance;
3188 * run_rebalance_domains is triggered when needed from the scheduler tick.
3189 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3190 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3192 static void run_rebalance_domains(struct softirq_action *h)
3194 int this_cpu = smp_processor_id();
3195 struct rq *this_rq = cpu_rq(this_cpu);
3196 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3197 CPU_IDLE : CPU_NOT_IDLE;
3199 rebalance_domains(this_cpu, idle);
3203 * If this cpu is the owner for idle load balancing, then do the
3204 * balancing on behalf of the other idle cpus whose ticks are
3207 if (this_rq->idle_at_tick &&
3208 atomic_read(&nohz.load_balancer) == this_cpu) {
3209 cpumask_t cpus = nohz.cpu_mask;
3213 cpu_clear(this_cpu, cpus);
3214 for_each_cpu_mask(balance_cpu, cpus) {
3216 * If this cpu gets work to do, stop the load balancing
3217 * work being done for other cpus. Next load
3218 * balancing owner will pick it up.
3223 rebalance_domains(balance_cpu, CPU_IDLE);
3225 rq = cpu_rq(balance_cpu);
3226 if (time_after(this_rq->next_balance, rq->next_balance))
3227 this_rq->next_balance = rq->next_balance;
3234 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3236 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3237 * idle load balancing owner or decide to stop the periodic load balancing,
3238 * if the whole system is idle.
3240 static inline void trigger_load_balance(struct rq *rq, int cpu)
3244 * If we were in the nohz mode recently and busy at the current
3245 * scheduler tick, then check if we need to nominate new idle
3248 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3249 rq->in_nohz_recently = 0;
3251 if (atomic_read(&nohz.load_balancer) == cpu) {
3252 cpu_clear(cpu, nohz.cpu_mask);
3253 atomic_set(&nohz.load_balancer, -1);
3256 if (atomic_read(&nohz.load_balancer) == -1) {
3258 * simple selection for now: Nominate the
3259 * first cpu in the nohz list to be the next
3262 * TBD: Traverse the sched domains and nominate
3263 * the nearest cpu in the nohz.cpu_mask.
3265 int ilb = first_cpu(nohz.cpu_mask);
3273 * If this cpu is idle and doing idle load balancing for all the
3274 * cpus with ticks stopped, is it time for that to stop?
3276 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3277 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3283 * If this cpu is idle and the idle load balancing is done by
3284 * someone else, then no need raise the SCHED_SOFTIRQ
3286 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3287 cpu_isset(cpu, nohz.cpu_mask))
3290 if (time_after_eq(jiffies, rq->next_balance))
3291 raise_softirq(SCHED_SOFTIRQ);
3294 #else /* CONFIG_SMP */
3297 * on UP we do not need to balance between CPUs:
3299 static inline void idle_balance(int cpu, struct rq *rq)
3305 DEFINE_PER_CPU(struct kernel_stat, kstat);
3307 EXPORT_PER_CPU_SYMBOL(kstat);
3310 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3311 * that have not yet been banked in case the task is currently running.
3313 unsigned long long task_sched_runtime(struct task_struct *p)
3315 unsigned long flags;
3319 rq = task_rq_lock(p, &flags);
3320 ns = p->se.sum_exec_runtime;
3321 if (rq->curr == p) {
3322 update_rq_clock(rq);
3323 delta_exec = rq->clock - p->se.exec_start;
3324 if ((s64)delta_exec > 0)
3327 task_rq_unlock(rq, &flags);
3333 * Account user cpu time to a process.
3334 * @p: the process that the cpu time gets accounted to
3335 * @cputime: the cpu time spent in user space since the last update
3337 void account_user_time(struct task_struct *p, cputime_t cputime)
3339 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3341 struct rq *rq = this_rq();
3343 p->utime = cputime_add(p->utime, cputime);
3346 cpuacct_charge(p, cputime);
3348 /* Add user time to cpustat. */
3349 tmp = cputime_to_cputime64(cputime);
3350 if (TASK_NICE(p) > 0)
3351 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3353 cpustat->user = cputime64_add(cpustat->user, tmp);
3357 * Account guest cpu time to a process.
3358 * @p: the process that the cpu time gets accounted to
3359 * @cputime: the cpu time spent in virtual machine since the last update
3361 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3364 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3366 tmp = cputime_to_cputime64(cputime);
3368 p->utime = cputime_add(p->utime, cputime);
3369 p->gtime = cputime_add(p->gtime, cputime);
3371 cpustat->user = cputime64_add(cpustat->user, tmp);
3372 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3376 * Account scaled user cpu time to a process.
3377 * @p: the process that the cpu time gets accounted to
3378 * @cputime: the cpu time spent in user space since the last update
3380 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3382 p->utimescaled = cputime_add(p->utimescaled, cputime);
3386 * Account system cpu time to a process.
3387 * @p: the process that the cpu time gets accounted to
3388 * @hardirq_offset: the offset to subtract from hardirq_count()
3389 * @cputime: the cpu time spent in kernel space since the last update
3391 void account_system_time(struct task_struct *p, int hardirq_offset,
3394 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3395 struct rq *rq = this_rq();
3398 if (p->flags & PF_VCPU) {
3399 account_guest_time(p, cputime);
3403 p->stime = cputime_add(p->stime, cputime);
3405 /* Add system time to cpustat. */
3406 tmp = cputime_to_cputime64(cputime);
3407 if (hardirq_count() - hardirq_offset)
3408 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3409 else if (softirq_count())
3410 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3411 else if (p != rq->idle) {
3412 cpustat->system = cputime64_add(cpustat->system, tmp);
3413 cpuacct_charge(p, cputime);
3414 } else if (atomic_read(&rq->nr_iowait) > 0)
3415 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3417 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3418 /* Account for system time used */
3419 acct_update_integrals(p);
3423 * Account scaled system cpu time to a process.
3424 * @p: the process that the cpu time gets accounted to
3425 * @hardirq_offset: the offset to subtract from hardirq_count()
3426 * @cputime: the cpu time spent in kernel space since the last update
3428 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3430 p->stimescaled = cputime_add(p->stimescaled, cputime);
3434 * Account for involuntary wait time.
3435 * @p: the process from which the cpu time has been stolen
3436 * @steal: the cpu time spent in involuntary wait
3438 void account_steal_time(struct task_struct *p, cputime_t steal)
3440 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3441 cputime64_t tmp = cputime_to_cputime64(steal);
3442 struct rq *rq = this_rq();
3444 if (p == rq->idle) {
3445 p->stime = cputime_add(p->stime, steal);
3446 if (atomic_read(&rq->nr_iowait) > 0)
3447 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3449 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3451 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3452 cpuacct_charge(p, -tmp);
3457 * This function gets called by the timer code, with HZ frequency.
3458 * We call it with interrupts disabled.
3460 * It also gets called by the fork code, when changing the parent's
3463 void scheduler_tick(void)
3465 int cpu = smp_processor_id();
3466 struct rq *rq = cpu_rq(cpu);
3467 struct task_struct *curr = rq->curr;
3468 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3470 spin_lock(&rq->lock);
3471 __update_rq_clock(rq);
3473 * Let rq->clock advance by at least TICK_NSEC:
3475 if (unlikely(rq->clock < next_tick))
3476 rq->clock = next_tick;
3477 rq->tick_timestamp = rq->clock;
3478 update_cpu_load(rq);
3479 if (curr != rq->idle) /* FIXME: needed? */
3480 curr->sched_class->task_tick(rq, curr);
3481 spin_unlock(&rq->lock);
3484 rq->idle_at_tick = idle_cpu(cpu);
3485 trigger_load_balance(rq, cpu);
3489 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3491 void fastcall add_preempt_count(int val)
3496 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3498 preempt_count() += val;
3500 * Spinlock count overflowing soon?
3502 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3505 EXPORT_SYMBOL(add_preempt_count);
3507 void fastcall sub_preempt_count(int val)
3512 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3515 * Is the spinlock portion underflowing?
3517 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3518 !(preempt_count() & PREEMPT_MASK)))
3521 preempt_count() -= val;
3523 EXPORT_SYMBOL(sub_preempt_count);
3528 * Print scheduling while atomic bug:
3530 static noinline void __schedule_bug(struct task_struct *prev)
3532 struct pt_regs *regs = get_irq_regs();
3534 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3535 prev->comm, prev->pid, preempt_count());
3537 debug_show_held_locks(prev);
3538 if (irqs_disabled())
3539 print_irqtrace_events(prev);
3548 * Various schedule()-time debugging checks and statistics:
3550 static inline void schedule_debug(struct task_struct *prev)
3553 * Test if we are atomic. Since do_exit() needs to call into
3554 * schedule() atomically, we ignore that path for now.
3555 * Otherwise, whine if we are scheduling when we should not be.
3557 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3558 __schedule_bug(prev);
3560 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3562 schedstat_inc(this_rq(), sched_count);
3563 #ifdef CONFIG_SCHEDSTATS
3564 if (unlikely(prev->lock_depth >= 0)) {
3565 schedstat_inc(this_rq(), bkl_count);
3566 schedstat_inc(prev, sched_info.bkl_count);
3572 * Pick up the highest-prio task:
3574 static inline struct task_struct *
3575 pick_next_task(struct rq *rq, struct task_struct *prev)
3577 const struct sched_class *class;
3578 struct task_struct *p;
3581 * Optimization: we know that if all tasks are in
3582 * the fair class we can call that function directly:
3584 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3585 p = fair_sched_class.pick_next_task(rq);
3590 class = sched_class_highest;
3592 p = class->pick_next_task(rq);
3596 * Will never be NULL as the idle class always
3597 * returns a non-NULL p:
3599 class = class->next;
3604 * schedule() is the main scheduler function.
3606 asmlinkage void __sched schedule(void)
3608 struct task_struct *prev, *next;
3615 cpu = smp_processor_id();
3619 switch_count = &prev->nivcsw;
3621 release_kernel_lock(prev);
3622 need_resched_nonpreemptible:
3624 schedule_debug(prev);
3627 * Do the rq-clock update outside the rq lock:
3629 local_irq_disable();
3630 __update_rq_clock(rq);
3631 spin_lock(&rq->lock);
3632 clear_tsk_need_resched(prev);
3634 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3635 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3636 unlikely(signal_pending(prev)))) {
3637 prev->state = TASK_RUNNING;
3639 deactivate_task(rq, prev, 1);
3641 switch_count = &prev->nvcsw;
3644 if (unlikely(!rq->nr_running))
3645 idle_balance(cpu, rq);
3647 prev->sched_class->put_prev_task(rq, prev);
3648 next = pick_next_task(rq, prev);
3650 sched_info_switch(prev, next);
3652 if (likely(prev != next)) {
3657 context_switch(rq, prev, next); /* unlocks the rq */
3659 spin_unlock_irq(&rq->lock);
3661 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3662 cpu = smp_processor_id();
3664 goto need_resched_nonpreemptible;
3666 preempt_enable_no_resched();
3667 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3670 EXPORT_SYMBOL(schedule);
3672 #ifdef CONFIG_PREEMPT
3674 * this is the entry point to schedule() from in-kernel preemption
3675 * off of preempt_enable. Kernel preemptions off return from interrupt
3676 * occur there and call schedule directly.
3678 asmlinkage void __sched preempt_schedule(void)
3680 struct thread_info *ti = current_thread_info();
3681 #ifdef CONFIG_PREEMPT_BKL
3682 struct task_struct *task = current;
3683 int saved_lock_depth;
3686 * If there is a non-zero preempt_count or interrupts are disabled,
3687 * we do not want to preempt the current task. Just return..
3689 if (likely(ti->preempt_count || irqs_disabled()))
3693 add_preempt_count(PREEMPT_ACTIVE);
3696 * We keep the big kernel semaphore locked, but we
3697 * clear ->lock_depth so that schedule() doesnt
3698 * auto-release the semaphore:
3700 #ifdef CONFIG_PREEMPT_BKL
3701 saved_lock_depth = task->lock_depth;
3702 task->lock_depth = -1;
3705 #ifdef CONFIG_PREEMPT_BKL
3706 task->lock_depth = saved_lock_depth;
3708 sub_preempt_count(PREEMPT_ACTIVE);
3711 * Check again in case we missed a preemption opportunity
3712 * between schedule and now.
3715 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3717 EXPORT_SYMBOL(preempt_schedule);
3720 * this is the entry point to schedule() from kernel preemption
3721 * off of irq context.
3722 * Note, that this is called and return with irqs disabled. This will
3723 * protect us against recursive calling from irq.
3725 asmlinkage void __sched preempt_schedule_irq(void)
3727 struct thread_info *ti = current_thread_info();
3728 #ifdef CONFIG_PREEMPT_BKL
3729 struct task_struct *task = current;
3730 int saved_lock_depth;
3732 /* Catch callers which need to be fixed */
3733 BUG_ON(ti->preempt_count || !irqs_disabled());
3736 add_preempt_count(PREEMPT_ACTIVE);
3739 * We keep the big kernel semaphore locked, but we
3740 * clear ->lock_depth so that schedule() doesnt
3741 * auto-release the semaphore:
3743 #ifdef CONFIG_PREEMPT_BKL
3744 saved_lock_depth = task->lock_depth;
3745 task->lock_depth = -1;
3749 local_irq_disable();
3750 #ifdef CONFIG_PREEMPT_BKL
3751 task->lock_depth = saved_lock_depth;
3753 sub_preempt_count(PREEMPT_ACTIVE);
3756 * Check again in case we missed a preemption opportunity
3757 * between schedule and now.
3760 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3763 #endif /* CONFIG_PREEMPT */
3765 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3768 return try_to_wake_up(curr->private, mode, sync);
3770 EXPORT_SYMBOL(default_wake_function);
3773 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3774 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3775 * number) then we wake all the non-exclusive tasks and one exclusive task.
3777 * There are circumstances in which we can try to wake a task which has already
3778 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3779 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3781 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3782 int nr_exclusive, int sync, void *key)
3784 wait_queue_t *curr, *next;
3786 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3787 unsigned flags = curr->flags;
3789 if (curr->func(curr, mode, sync, key) &&
3790 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3796 * __wake_up - wake up threads blocked on a waitqueue.
3798 * @mode: which threads
3799 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3800 * @key: is directly passed to the wakeup function
3802 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3803 int nr_exclusive, void *key)
3805 unsigned long flags;
3807 spin_lock_irqsave(&q->lock, flags);
3808 __wake_up_common(q, mode, nr_exclusive, 0, key);
3809 spin_unlock_irqrestore(&q->lock, flags);
3811 EXPORT_SYMBOL(__wake_up);
3814 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3816 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3818 __wake_up_common(q, mode, 1, 0, NULL);
3822 * __wake_up_sync - wake up threads blocked on a waitqueue.
3824 * @mode: which threads
3825 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3827 * The sync wakeup differs that the waker knows that it will schedule
3828 * away soon, so while the target thread will be woken up, it will not
3829 * be migrated to another CPU - ie. the two threads are 'synchronized'
3830 * with each other. This can prevent needless bouncing between CPUs.
3832 * On UP it can prevent extra preemption.
3835 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3837 unsigned long flags;
3843 if (unlikely(!nr_exclusive))
3846 spin_lock_irqsave(&q->lock, flags);
3847 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3848 spin_unlock_irqrestore(&q->lock, flags);
3850 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3852 void complete(struct completion *x)
3854 unsigned long flags;
3856 spin_lock_irqsave(&x->wait.lock, flags);
3858 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3860 spin_unlock_irqrestore(&x->wait.lock, flags);
3862 EXPORT_SYMBOL(complete);
3864 void complete_all(struct completion *x)
3866 unsigned long flags;
3868 spin_lock_irqsave(&x->wait.lock, flags);
3869 x->done += UINT_MAX/2;
3870 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3872 spin_unlock_irqrestore(&x->wait.lock, flags);
3874 EXPORT_SYMBOL(complete_all);
3876 static inline long __sched
3877 do_wait_for_common(struct completion *x, long timeout, int state)
3880 DECLARE_WAITQUEUE(wait, current);
3882 wait.flags |= WQ_FLAG_EXCLUSIVE;
3883 __add_wait_queue_tail(&x->wait, &wait);
3885 if (state == TASK_INTERRUPTIBLE &&
3886 signal_pending(current)) {
3887 __remove_wait_queue(&x->wait, &wait);
3888 return -ERESTARTSYS;
3890 __set_current_state(state);
3891 spin_unlock_irq(&x->wait.lock);
3892 timeout = schedule_timeout(timeout);
3893 spin_lock_irq(&x->wait.lock);
3895 __remove_wait_queue(&x->wait, &wait);
3899 __remove_wait_queue(&x->wait, &wait);
3906 wait_for_common(struct completion *x, long timeout, int state)
3910 spin_lock_irq(&x->wait.lock);
3911 timeout = do_wait_for_common(x, timeout, state);
3912 spin_unlock_irq(&x->wait.lock);
3916 void __sched wait_for_completion(struct completion *x)
3918 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3920 EXPORT_SYMBOL(wait_for_completion);
3922 unsigned long __sched
3923 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3925 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3927 EXPORT_SYMBOL(wait_for_completion_timeout);
3929 int __sched wait_for_completion_interruptible(struct completion *x)
3931 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3932 if (t == -ERESTARTSYS)
3936 EXPORT_SYMBOL(wait_for_completion_interruptible);
3938 unsigned long __sched
3939 wait_for_completion_interruptible_timeout(struct completion *x,
3940 unsigned long timeout)
3942 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3944 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3947 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3949 unsigned long flags;
3952 init_waitqueue_entry(&wait, current);
3954 __set_current_state(state);
3956 spin_lock_irqsave(&q->lock, flags);
3957 __add_wait_queue(q, &wait);
3958 spin_unlock(&q->lock);
3959 timeout = schedule_timeout(timeout);
3960 spin_lock_irq(&q->lock);
3961 __remove_wait_queue(q, &wait);
3962 spin_unlock_irqrestore(&q->lock, flags);
3967 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3969 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3971 EXPORT_SYMBOL(interruptible_sleep_on);
3974 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3976 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3978 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3980 void __sched sleep_on(wait_queue_head_t *q)
3982 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3984 EXPORT_SYMBOL(sleep_on);
3986 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3988 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3990 EXPORT_SYMBOL(sleep_on_timeout);
3992 #ifdef CONFIG_RT_MUTEXES
3995 * rt_mutex_setprio - set the current priority of a task
3997 * @prio: prio value (kernel-internal form)
3999 * This function changes the 'effective' priority of a task. It does
4000 * not touch ->normal_prio like __setscheduler().
4002 * Used by the rt_mutex code to implement priority inheritance logic.
4004 void rt_mutex_setprio(struct task_struct *p, int prio)
4006 unsigned long flags;
4007 int oldprio, on_rq, running;
4010 BUG_ON(prio < 0 || prio > MAX_PRIO);
4012 rq = task_rq_lock(p, &flags);
4013 update_rq_clock(rq);
4016 on_rq = p->se.on_rq;
4017 running = task_running(rq, p);
4019 dequeue_task(rq, p, 0);
4021 p->sched_class->put_prev_task(rq, p);
4025 p->sched_class = &rt_sched_class;
4027 p->sched_class = &fair_sched_class;
4033 p->sched_class->set_curr_task(rq);
4034 enqueue_task(rq, p, 0);
4036 * Reschedule if we are currently running on this runqueue and
4037 * our priority decreased, or if we are not currently running on
4038 * this runqueue and our priority is higher than the current's
4041 if (p->prio > oldprio)
4042 resched_task(rq->curr);
4044 check_preempt_curr(rq, p);
4047 task_rq_unlock(rq, &flags);
4052 void set_user_nice(struct task_struct *p, long nice)
4054 int old_prio, delta, on_rq;
4055 unsigned long flags;
4058 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4061 * We have to be careful, if called from sys_setpriority(),
4062 * the task might be in the middle of scheduling on another CPU.
4064 rq = task_rq_lock(p, &flags);
4065 update_rq_clock(rq);
4067 * The RT priorities are set via sched_setscheduler(), but we still
4068 * allow the 'normal' nice value to be set - but as expected
4069 * it wont have any effect on scheduling until the task is
4070 * SCHED_FIFO/SCHED_RR:
4072 if (task_has_rt_policy(p)) {
4073 p->static_prio = NICE_TO_PRIO(nice);
4076 on_rq = p->se.on_rq;
4078 dequeue_task(rq, p, 0);
4082 p->static_prio = NICE_TO_PRIO(nice);
4085 p->prio = effective_prio(p);
4086 delta = p->prio - old_prio;
4089 enqueue_task(rq, p, 0);
4092 * If the task increased its priority or is running and
4093 * lowered its priority, then reschedule its CPU:
4095 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4096 resched_task(rq->curr);
4099 task_rq_unlock(rq, &flags);
4101 EXPORT_SYMBOL(set_user_nice);
4104 * can_nice - check if a task can reduce its nice value
4108 int can_nice(const struct task_struct *p, const int nice)
4110 /* convert nice value [19,-20] to rlimit style value [1,40] */
4111 int nice_rlim = 20 - nice;
4113 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4114 capable(CAP_SYS_NICE));
4117 #ifdef __ARCH_WANT_SYS_NICE
4120 * sys_nice - change the priority of the current process.
4121 * @increment: priority increment
4123 * sys_setpriority is a more generic, but much slower function that
4124 * does similar things.
4126 asmlinkage long sys_nice(int increment)
4131 * Setpriority might change our priority at the same moment.
4132 * We don't have to worry. Conceptually one call occurs first
4133 * and we have a single winner.
4135 if (increment < -40)
4140 nice = PRIO_TO_NICE(current->static_prio) + increment;
4146 if (increment < 0 && !can_nice(current, nice))
4149 retval = security_task_setnice(current, nice);
4153 set_user_nice(current, nice);
4160 * task_prio - return the priority value of a given task.
4161 * @p: the task in question.
4163 * This is the priority value as seen by users in /proc.
4164 * RT tasks are offset by -200. Normal tasks are centered
4165 * around 0, value goes from -16 to +15.
4167 int task_prio(const struct task_struct *p)
4169 return p->prio - MAX_RT_PRIO;
4173 * task_nice - return the nice value of a given task.
4174 * @p: the task in question.
4176 int task_nice(const struct task_struct *p)
4178 return TASK_NICE(p);
4180 EXPORT_SYMBOL_GPL(task_nice);
4183 * idle_cpu - is a given cpu idle currently?
4184 * @cpu: the processor in question.
4186 int idle_cpu(int cpu)
4188 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4192 * idle_task - return the idle task for a given cpu.
4193 * @cpu: the processor in question.
4195 struct task_struct *idle_task(int cpu)
4197 return cpu_rq(cpu)->idle;
4201 * find_process_by_pid - find a process with a matching PID value.
4202 * @pid: the pid in question.
4204 static struct task_struct *find_process_by_pid(pid_t pid)
4206 return pid ? find_task_by_vpid(pid) : current;
4209 /* Actually do priority change: must hold rq lock. */
4211 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4213 BUG_ON(p->se.on_rq);
4216 switch (p->policy) {
4220 p->sched_class = &fair_sched_class;
4224 p->sched_class = &rt_sched_class;
4228 p->rt_priority = prio;
4229 p->normal_prio = normal_prio(p);
4230 /* we are holding p->pi_lock already */
4231 p->prio = rt_mutex_getprio(p);
4236 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4237 * @p: the task in question.
4238 * @policy: new policy.
4239 * @param: structure containing the new RT priority.
4241 * NOTE that the task may be already dead.
4243 int sched_setscheduler(struct task_struct *p, int policy,
4244 struct sched_param *param)
4246 int retval, oldprio, oldpolicy = -1, on_rq, running;
4247 unsigned long flags;
4250 /* may grab non-irq protected spin_locks */
4251 BUG_ON(in_interrupt());
4253 /* double check policy once rq lock held */
4255 policy = oldpolicy = p->policy;
4256 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4257 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4258 policy != SCHED_IDLE)
4261 * Valid priorities for SCHED_FIFO and SCHED_RR are
4262 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4263 * SCHED_BATCH and SCHED_IDLE is 0.
4265 if (param->sched_priority < 0 ||
4266 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4267 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4269 if (rt_policy(policy) != (param->sched_priority != 0))
4273 * Allow unprivileged RT tasks to decrease priority:
4275 if (!capable(CAP_SYS_NICE)) {
4276 if (rt_policy(policy)) {
4277 unsigned long rlim_rtprio;
4279 if (!lock_task_sighand(p, &flags))
4281 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4282 unlock_task_sighand(p, &flags);
4284 /* can't set/change the rt policy */
4285 if (policy != p->policy && !rlim_rtprio)
4288 /* can't increase priority */
4289 if (param->sched_priority > p->rt_priority &&
4290 param->sched_priority > rlim_rtprio)
4294 * Like positive nice levels, dont allow tasks to
4295 * move out of SCHED_IDLE either:
4297 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4300 /* can't change other user's priorities */
4301 if ((current->euid != p->euid) &&
4302 (current->euid != p->uid))
4306 retval = security_task_setscheduler(p, policy, param);
4310 * make sure no PI-waiters arrive (or leave) while we are
4311 * changing the priority of the task:
4313 spin_lock_irqsave(&p->pi_lock, flags);
4315 * To be able to change p->policy safely, the apropriate
4316 * runqueue lock must be held.
4318 rq = __task_rq_lock(p);
4319 /* recheck policy now with rq lock held */
4320 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4321 policy = oldpolicy = -1;
4322 __task_rq_unlock(rq);
4323 spin_unlock_irqrestore(&p->pi_lock, flags);
4326 update_rq_clock(rq);
4327 on_rq = p->se.on_rq;
4328 running = task_running(rq, p);
4330 deactivate_task(rq, p, 0);
4332 p->sched_class->put_prev_task(rq, p);
4336 __setscheduler(rq, p, policy, param->sched_priority);
4340 p->sched_class->set_curr_task(rq);
4341 activate_task(rq, p, 0);
4343 * Reschedule if we are currently running on this runqueue and
4344 * our priority decreased, or if we are not currently running on
4345 * this runqueue and our priority is higher than the current's
4348 if (p->prio > oldprio)
4349 resched_task(rq->curr);
4351 check_preempt_curr(rq, p);
4354 __task_rq_unlock(rq);
4355 spin_unlock_irqrestore(&p->pi_lock, flags);
4357 rt_mutex_adjust_pi(p);
4361 EXPORT_SYMBOL_GPL(sched_setscheduler);
4364 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4366 struct sched_param lparam;
4367 struct task_struct *p;
4370 if (!param || pid < 0)
4372 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4377 p = find_process_by_pid(pid);
4379 retval = sched_setscheduler(p, policy, &lparam);
4386 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4387 * @pid: the pid in question.
4388 * @policy: new policy.
4389 * @param: structure containing the new RT priority.
4391 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4392 struct sched_param __user *param)
4394 /* negative values for policy are not valid */
4398 return do_sched_setscheduler(pid, policy, param);
4402 * sys_sched_setparam - set/change the RT priority of a thread
4403 * @pid: the pid in question.
4404 * @param: structure containing the new RT priority.
4406 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4408 return do_sched_setscheduler(pid, -1, param);
4412 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4413 * @pid: the pid in question.
4415 asmlinkage long sys_sched_getscheduler(pid_t pid)
4417 struct task_struct *p;
4424 read_lock(&tasklist_lock);
4425 p = find_process_by_pid(pid);
4427 retval = security_task_getscheduler(p);
4431 read_unlock(&tasklist_lock);
4436 * sys_sched_getscheduler - get the RT priority of a thread
4437 * @pid: the pid in question.
4438 * @param: structure containing the RT priority.
4440 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4442 struct sched_param lp;
4443 struct task_struct *p;
4446 if (!param || pid < 0)
4449 read_lock(&tasklist_lock);
4450 p = find_process_by_pid(pid);
4455 retval = security_task_getscheduler(p);
4459 lp.sched_priority = p->rt_priority;
4460 read_unlock(&tasklist_lock);
4463 * This one might sleep, we cannot do it with a spinlock held ...
4465 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4470 read_unlock(&tasklist_lock);
4474 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4476 cpumask_t cpus_allowed;
4477 struct task_struct *p;
4480 mutex_lock(&sched_hotcpu_mutex);
4481 read_lock(&tasklist_lock);
4483 p = find_process_by_pid(pid);
4485 read_unlock(&tasklist_lock);
4486 mutex_unlock(&sched_hotcpu_mutex);
4491 * It is not safe to call set_cpus_allowed with the
4492 * tasklist_lock held. We will bump the task_struct's
4493 * usage count and then drop tasklist_lock.
4496 read_unlock(&tasklist_lock);
4499 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4500 !capable(CAP_SYS_NICE))
4503 retval = security_task_setscheduler(p, 0, NULL);
4507 cpus_allowed = cpuset_cpus_allowed(p);
4508 cpus_and(new_mask, new_mask, cpus_allowed);
4510 retval = set_cpus_allowed(p, new_mask);
4513 cpus_allowed = cpuset_cpus_allowed(p);
4514 if (!cpus_subset(new_mask, cpus_allowed)) {
4516 * We must have raced with a concurrent cpuset
4517 * update. Just reset the cpus_allowed to the
4518 * cpuset's cpus_allowed
4520 new_mask = cpus_allowed;
4526 mutex_unlock(&sched_hotcpu_mutex);
4530 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4531 cpumask_t *new_mask)
4533 if (len < sizeof(cpumask_t)) {
4534 memset(new_mask, 0, sizeof(cpumask_t));
4535 } else if (len > sizeof(cpumask_t)) {
4536 len = sizeof(cpumask_t);
4538 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4542 * sys_sched_setaffinity - set the cpu affinity of a process
4543 * @pid: pid of the process
4544 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4545 * @user_mask_ptr: user-space pointer to the new cpu mask
4547 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4548 unsigned long __user *user_mask_ptr)
4553 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4557 return sched_setaffinity(pid, new_mask);
4561 * Represents all cpu's present in the system
4562 * In systems capable of hotplug, this map could dynamically grow
4563 * as new cpu's are detected in the system via any platform specific
4564 * method, such as ACPI for e.g.
4567 cpumask_t cpu_present_map __read_mostly;
4568 EXPORT_SYMBOL(cpu_present_map);
4571 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4572 EXPORT_SYMBOL(cpu_online_map);
4574 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4575 EXPORT_SYMBOL(cpu_possible_map);
4578 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4580 struct task_struct *p;
4583 mutex_lock(&sched_hotcpu_mutex);
4584 read_lock(&tasklist_lock);
4587 p = find_process_by_pid(pid);
4591 retval = security_task_getscheduler(p);
4595 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4598 read_unlock(&tasklist_lock);
4599 mutex_unlock(&sched_hotcpu_mutex);
4605 * sys_sched_getaffinity - get the cpu affinity of a process
4606 * @pid: pid of the process
4607 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4608 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4610 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4611 unsigned long __user *user_mask_ptr)
4616 if (len < sizeof(cpumask_t))
4619 ret = sched_getaffinity(pid, &mask);
4623 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4626 return sizeof(cpumask_t);
4630 * sys_sched_yield - yield the current processor to other threads.
4632 * This function yields the current CPU to other tasks. If there are no
4633 * other threads running on this CPU then this function will return.
4635 asmlinkage long sys_sched_yield(void)
4637 struct rq *rq = this_rq_lock();
4639 schedstat_inc(rq, yld_count);
4640 current->sched_class->yield_task(rq);
4643 * Since we are going to call schedule() anyway, there's
4644 * no need to preempt or enable interrupts:
4646 __release(rq->lock);
4647 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4648 _raw_spin_unlock(&rq->lock);
4649 preempt_enable_no_resched();
4656 static void __cond_resched(void)
4658 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4659 __might_sleep(__FILE__, __LINE__);
4662 * The BKS might be reacquired before we have dropped
4663 * PREEMPT_ACTIVE, which could trigger a second
4664 * cond_resched() call.
4667 add_preempt_count(PREEMPT_ACTIVE);
4669 sub_preempt_count(PREEMPT_ACTIVE);
4670 } while (need_resched());
4673 int __sched cond_resched(void)
4675 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4676 system_state == SYSTEM_RUNNING) {
4682 EXPORT_SYMBOL(cond_resched);
4685 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4686 * call schedule, and on return reacquire the lock.
4688 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4689 * operations here to prevent schedule() from being called twice (once via
4690 * spin_unlock(), once by hand).
4692 int cond_resched_lock(spinlock_t *lock)
4696 if (need_lockbreak(lock)) {
4702 if (need_resched() && system_state == SYSTEM_RUNNING) {
4703 spin_release(&lock->dep_map, 1, _THIS_IP_);
4704 _raw_spin_unlock(lock);
4705 preempt_enable_no_resched();
4712 EXPORT_SYMBOL(cond_resched_lock);
4714 int __sched cond_resched_softirq(void)
4716 BUG_ON(!in_softirq());
4718 if (need_resched() && system_state == SYSTEM_RUNNING) {
4726 EXPORT_SYMBOL(cond_resched_softirq);
4729 * yield - yield the current processor to other threads.
4731 * This is a shortcut for kernel-space yielding - it marks the
4732 * thread runnable and calls sys_sched_yield().
4734 void __sched yield(void)
4736 set_current_state(TASK_RUNNING);
4739 EXPORT_SYMBOL(yield);
4742 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4743 * that process accounting knows that this is a task in IO wait state.
4745 * But don't do that if it is a deliberate, throttling IO wait (this task
4746 * has set its backing_dev_info: the queue against which it should throttle)
4748 void __sched io_schedule(void)
4750 struct rq *rq = &__raw_get_cpu_var(runqueues);
4752 delayacct_blkio_start();
4753 atomic_inc(&rq->nr_iowait);
4755 atomic_dec(&rq->nr_iowait);
4756 delayacct_blkio_end();
4758 EXPORT_SYMBOL(io_schedule);
4760 long __sched io_schedule_timeout(long timeout)
4762 struct rq *rq = &__raw_get_cpu_var(runqueues);
4765 delayacct_blkio_start();
4766 atomic_inc(&rq->nr_iowait);
4767 ret = schedule_timeout(timeout);
4768 atomic_dec(&rq->nr_iowait);
4769 delayacct_blkio_end();
4774 * sys_sched_get_priority_max - return maximum RT priority.
4775 * @policy: scheduling class.
4777 * this syscall returns the maximum rt_priority that can be used
4778 * by a given scheduling class.
4780 asmlinkage long sys_sched_get_priority_max(int policy)
4787 ret = MAX_USER_RT_PRIO-1;
4799 * sys_sched_get_priority_min - return minimum RT priority.
4800 * @policy: scheduling class.
4802 * this syscall returns the minimum rt_priority that can be used
4803 * by a given scheduling class.
4805 asmlinkage long sys_sched_get_priority_min(int policy)
4823 * sys_sched_rr_get_interval - return the default timeslice of a process.
4824 * @pid: pid of the process.
4825 * @interval: userspace pointer to the timeslice value.
4827 * this syscall writes the default timeslice value of a given process
4828 * into the user-space timespec buffer. A value of '0' means infinity.
4831 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4833 struct task_struct *p;
4834 unsigned int time_slice;
4842 read_lock(&tasklist_lock);
4843 p = find_process_by_pid(pid);
4847 retval = security_task_getscheduler(p);
4851 if (p->policy == SCHED_FIFO)
4853 else if (p->policy == SCHED_RR)
4854 time_slice = DEF_TIMESLICE;
4856 struct sched_entity *se = &p->se;
4857 unsigned long flags;
4860 rq = task_rq_lock(p, &flags);
4861 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4862 task_rq_unlock(rq, &flags);
4864 read_unlock(&tasklist_lock);
4865 jiffies_to_timespec(time_slice, &t);
4866 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4870 read_unlock(&tasklist_lock);
4874 static const char stat_nam[] = "RSDTtZX";
4876 static void show_task(struct task_struct *p)
4878 unsigned long free = 0;
4881 state = p->state ? __ffs(p->state) + 1 : 0;
4882 printk(KERN_INFO "%-13.13s %c", p->comm,
4883 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4884 #if BITS_PER_LONG == 32
4885 if (state == TASK_RUNNING)
4886 printk(KERN_CONT " running ");
4888 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4890 if (state == TASK_RUNNING)
4891 printk(KERN_CONT " running task ");
4893 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4895 #ifdef CONFIG_DEBUG_STACK_USAGE
4897 unsigned long *n = end_of_stack(p);
4900 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4903 printk(KERN_CONT "%5lu %5d %6d\n", free,
4904 task_pid_nr(p), task_pid_nr(p->parent));
4906 if (state != TASK_RUNNING)
4907 show_stack(p, NULL);
4910 void show_state_filter(unsigned long state_filter)
4912 struct task_struct *g, *p;
4914 #if BITS_PER_LONG == 32
4916 " task PC stack pid father\n");
4919 " task PC stack pid father\n");
4921 read_lock(&tasklist_lock);
4922 do_each_thread(g, p) {
4924 * reset the NMI-timeout, listing all files on a slow
4925 * console might take alot of time:
4927 touch_nmi_watchdog();
4928 if (!state_filter || (p->state & state_filter))
4930 } while_each_thread(g, p);
4932 touch_all_softlockup_watchdogs();
4934 #ifdef CONFIG_SCHED_DEBUG
4935 sysrq_sched_debug_show();
4937 read_unlock(&tasklist_lock);
4939 * Only show locks if all tasks are dumped:
4941 if (state_filter == -1)
4942 debug_show_all_locks();
4945 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4947 idle->sched_class = &idle_sched_class;
4951 * init_idle - set up an idle thread for a given CPU
4952 * @idle: task in question
4953 * @cpu: cpu the idle task belongs to
4955 * NOTE: this function does not set the idle thread's NEED_RESCHED
4956 * flag, to make booting more robust.
4958 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4960 struct rq *rq = cpu_rq(cpu);
4961 unsigned long flags;
4964 idle->se.exec_start = sched_clock();
4966 idle->prio = idle->normal_prio = MAX_PRIO;
4967 idle->cpus_allowed = cpumask_of_cpu(cpu);
4968 __set_task_cpu(idle, cpu);
4970 spin_lock_irqsave(&rq->lock, flags);
4971 rq->curr = rq->idle = idle;
4972 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4975 spin_unlock_irqrestore(&rq->lock, flags);
4977 /* Set the preempt count _outside_ the spinlocks! */
4978 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4979 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4981 task_thread_info(idle)->preempt_count = 0;
4984 * The idle tasks have their own, simple scheduling class:
4986 idle->sched_class = &idle_sched_class;
4990 * In a system that switches off the HZ timer nohz_cpu_mask
4991 * indicates which cpus entered this state. This is used
4992 * in the rcu update to wait only for active cpus. For system
4993 * which do not switch off the HZ timer nohz_cpu_mask should
4994 * always be CPU_MASK_NONE.
4996 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4999 * Increase the granularity value when there are more CPUs,
5000 * because with more CPUs the 'effective latency' as visible
5001 * to users decreases. But the relationship is not linear,
5002 * so pick a second-best guess by going with the log2 of the
5005 * This idea comes from the SD scheduler of Con Kolivas:
5007 static inline void sched_init_granularity(void)
5009 unsigned int factor = 1 + ilog2(num_online_cpus());
5010 const unsigned long limit = 200000000;
5012 sysctl_sched_min_granularity *= factor;
5013 if (sysctl_sched_min_granularity > limit)
5014 sysctl_sched_min_granularity = limit;
5016 sysctl_sched_latency *= factor;
5017 if (sysctl_sched_latency > limit)
5018 sysctl_sched_latency = limit;
5020 sysctl_sched_wakeup_granularity *= factor;
5021 sysctl_sched_batch_wakeup_granularity *= factor;
5026 * This is how migration works:
5028 * 1) we queue a struct migration_req structure in the source CPU's
5029 * runqueue and wake up that CPU's migration thread.
5030 * 2) we down() the locked semaphore => thread blocks.
5031 * 3) migration thread wakes up (implicitly it forces the migrated
5032 * thread off the CPU)
5033 * 4) it gets the migration request and checks whether the migrated
5034 * task is still in the wrong runqueue.
5035 * 5) if it's in the wrong runqueue then the migration thread removes
5036 * it and puts it into the right queue.
5037 * 6) migration thread up()s the semaphore.
5038 * 7) we wake up and the migration is done.
5042 * Change a given task's CPU affinity. Migrate the thread to a
5043 * proper CPU and schedule it away if the CPU it's executing on
5044 * is removed from the allowed bitmask.
5046 * NOTE: the caller must have a valid reference to the task, the
5047 * task must not exit() & deallocate itself prematurely. The
5048 * call is not atomic; no spinlocks may be held.
5050 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5052 struct migration_req req;
5053 unsigned long flags;
5057 rq = task_rq_lock(p, &flags);
5058 if (!cpus_intersects(new_mask, cpu_online_map)) {
5063 p->cpus_allowed = new_mask;
5064 /* Can the task run on the task's current CPU? If so, we're done */
5065 if (cpu_isset(task_cpu(p), new_mask))
5068 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5069 /* Need help from migration thread: drop lock and wait. */
5070 task_rq_unlock(rq, &flags);
5071 wake_up_process(rq->migration_thread);
5072 wait_for_completion(&req.done);
5073 tlb_migrate_finish(p->mm);
5077 task_rq_unlock(rq, &flags);
5081 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5084 * Move (not current) task off this cpu, onto dest cpu. We're doing
5085 * this because either it can't run here any more (set_cpus_allowed()
5086 * away from this CPU, or CPU going down), or because we're
5087 * attempting to rebalance this task on exec (sched_exec).
5089 * So we race with normal scheduler movements, but that's OK, as long
5090 * as the task is no longer on this CPU.
5092 * Returns non-zero if task was successfully migrated.
5094 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5096 struct rq *rq_dest, *rq_src;
5099 if (unlikely(cpu_is_offline(dest_cpu)))
5102 rq_src = cpu_rq(src_cpu);
5103 rq_dest = cpu_rq(dest_cpu);
5105 double_rq_lock(rq_src, rq_dest);
5106 /* Already moved. */
5107 if (task_cpu(p) != src_cpu)
5109 /* Affinity changed (again). */
5110 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5113 on_rq = p->se.on_rq;
5115 deactivate_task(rq_src, p, 0);
5117 set_task_cpu(p, dest_cpu);
5119 activate_task(rq_dest, p, 0);
5120 check_preempt_curr(rq_dest, p);
5124 double_rq_unlock(rq_src, rq_dest);
5129 * migration_thread - this is a highprio system thread that performs
5130 * thread migration by bumping thread off CPU then 'pushing' onto
5133 static int migration_thread(void *data)
5135 int cpu = (long)data;
5139 BUG_ON(rq->migration_thread != current);
5141 set_current_state(TASK_INTERRUPTIBLE);
5142 while (!kthread_should_stop()) {
5143 struct migration_req *req;
5144 struct list_head *head;
5146 spin_lock_irq(&rq->lock);
5148 if (cpu_is_offline(cpu)) {
5149 spin_unlock_irq(&rq->lock);
5153 if (rq->active_balance) {
5154 active_load_balance(rq, cpu);
5155 rq->active_balance = 0;
5158 head = &rq->migration_queue;
5160 if (list_empty(head)) {
5161 spin_unlock_irq(&rq->lock);
5163 set_current_state(TASK_INTERRUPTIBLE);
5166 req = list_entry(head->next, struct migration_req, list);
5167 list_del_init(head->next);
5169 spin_unlock(&rq->lock);
5170 __migrate_task(req->task, cpu, req->dest_cpu);
5173 complete(&req->done);
5175 __set_current_state(TASK_RUNNING);
5179 /* Wait for kthread_stop */
5180 set_current_state(TASK_INTERRUPTIBLE);
5181 while (!kthread_should_stop()) {
5183 set_current_state(TASK_INTERRUPTIBLE);
5185 __set_current_state(TASK_RUNNING);
5189 #ifdef CONFIG_HOTPLUG_CPU
5191 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5195 local_irq_disable();
5196 ret = __migrate_task(p, src_cpu, dest_cpu);
5202 * Figure out where task on dead CPU should go, use force if necessary.
5203 * NOTE: interrupts should be disabled by the caller
5205 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5207 unsigned long flags;
5214 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5215 cpus_and(mask, mask, p->cpus_allowed);
5216 dest_cpu = any_online_cpu(mask);
5218 /* On any allowed CPU? */
5219 if (dest_cpu == NR_CPUS)
5220 dest_cpu = any_online_cpu(p->cpus_allowed);
5222 /* No more Mr. Nice Guy. */
5223 if (dest_cpu == NR_CPUS) {
5224 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5226 * Try to stay on the same cpuset, where the
5227 * current cpuset may be a subset of all cpus.
5228 * The cpuset_cpus_allowed_locked() variant of
5229 * cpuset_cpus_allowed() will not block. It must be
5230 * called within calls to cpuset_lock/cpuset_unlock.
5232 rq = task_rq_lock(p, &flags);
5233 p->cpus_allowed = cpus_allowed;
5234 dest_cpu = any_online_cpu(p->cpus_allowed);
5235 task_rq_unlock(rq, &flags);
5238 * Don't tell them about moving exiting tasks or
5239 * kernel threads (both mm NULL), since they never
5242 if (p->mm && printk_ratelimit())
5243 printk(KERN_INFO "process %d (%s) no "
5244 "longer affine to cpu%d\n",
5245 task_pid_nr(p), p->comm, dead_cpu);
5247 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5251 * While a dead CPU has no uninterruptible tasks queued at this point,
5252 * it might still have a nonzero ->nr_uninterruptible counter, because
5253 * for performance reasons the counter is not stricly tracking tasks to
5254 * their home CPUs. So we just add the counter to another CPU's counter,
5255 * to keep the global sum constant after CPU-down:
5257 static void migrate_nr_uninterruptible(struct rq *rq_src)
5259 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5260 unsigned long flags;
5262 local_irq_save(flags);
5263 double_rq_lock(rq_src, rq_dest);
5264 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5265 rq_src->nr_uninterruptible = 0;
5266 double_rq_unlock(rq_src, rq_dest);
5267 local_irq_restore(flags);
5270 /* Run through task list and migrate tasks from the dead cpu. */
5271 static void migrate_live_tasks(int src_cpu)
5273 struct task_struct *p, *t;
5275 read_lock(&tasklist_lock);
5277 do_each_thread(t, p) {
5281 if (task_cpu(p) == src_cpu)
5282 move_task_off_dead_cpu(src_cpu, p);
5283 } while_each_thread(t, p);
5285 read_unlock(&tasklist_lock);
5289 * activate_idle_task - move idle task to the _front_ of runqueue.
5291 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5293 update_rq_clock(rq);
5295 if (p->state == TASK_UNINTERRUPTIBLE)
5296 rq->nr_uninterruptible--;
5298 enqueue_task(rq, p, 0);
5299 inc_nr_running(p, rq);
5303 * Schedules idle task to be the next runnable task on current CPU.
5304 * It does so by boosting its priority to highest possible and adding it to
5305 * the _front_ of the runqueue. Used by CPU offline code.
5307 void sched_idle_next(void)
5309 int this_cpu = smp_processor_id();
5310 struct rq *rq = cpu_rq(this_cpu);
5311 struct task_struct *p = rq->idle;
5312 unsigned long flags;
5314 /* cpu has to be offline */
5315 BUG_ON(cpu_online(this_cpu));
5318 * Strictly not necessary since rest of the CPUs are stopped by now
5319 * and interrupts disabled on the current cpu.
5321 spin_lock_irqsave(&rq->lock, flags);
5323 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5325 /* Add idle task to the _front_ of its priority queue: */
5326 activate_idle_task(p, rq);
5328 spin_unlock_irqrestore(&rq->lock, flags);
5332 * Ensures that the idle task is using init_mm right before its cpu goes
5335 void idle_task_exit(void)
5337 struct mm_struct *mm = current->active_mm;
5339 BUG_ON(cpu_online(smp_processor_id()));
5342 switch_mm(mm, &init_mm, current);
5346 /* called under rq->lock with disabled interrupts */
5347 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5349 struct rq *rq = cpu_rq(dead_cpu);
5351 /* Must be exiting, otherwise would be on tasklist. */
5352 BUG_ON(!p->exit_state);
5354 /* Cannot have done final schedule yet: would have vanished. */
5355 BUG_ON(p->state == TASK_DEAD);
5360 * Drop lock around migration; if someone else moves it,
5361 * that's OK. No task can be added to this CPU, so iteration is
5364 spin_unlock_irq(&rq->lock);
5365 move_task_off_dead_cpu(dead_cpu, p);
5366 spin_lock_irq(&rq->lock);
5371 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5372 static void migrate_dead_tasks(unsigned int dead_cpu)
5374 struct rq *rq = cpu_rq(dead_cpu);
5375 struct task_struct *next;
5378 if (!rq->nr_running)
5380 update_rq_clock(rq);
5381 next = pick_next_task(rq, rq->curr);
5384 migrate_dead(dead_cpu, next);
5388 #endif /* CONFIG_HOTPLUG_CPU */
5390 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5392 static struct ctl_table sd_ctl_dir[] = {
5394 .procname = "sched_domain",
5400 static struct ctl_table sd_ctl_root[] = {
5402 .ctl_name = CTL_KERN,
5403 .procname = "kernel",
5405 .child = sd_ctl_dir,
5410 static struct ctl_table *sd_alloc_ctl_entry(int n)
5412 struct ctl_table *entry =
5413 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5418 static void sd_free_ctl_entry(struct ctl_table **tablep)
5420 struct ctl_table *entry;
5423 * In the intermediate directories, both the child directory and
5424 * procname are dynamically allocated and could fail but the mode
5425 * will always be set. In the lowest directory the names are
5426 * static strings and all have proc handlers.
5428 for (entry = *tablep; entry->mode; entry++) {
5430 sd_free_ctl_entry(&entry->child);
5431 if (entry->proc_handler == NULL)
5432 kfree(entry->procname);
5440 set_table_entry(struct ctl_table *entry,
5441 const char *procname, void *data, int maxlen,
5442 mode_t mode, proc_handler *proc_handler)
5444 entry->procname = procname;
5446 entry->maxlen = maxlen;
5448 entry->proc_handler = proc_handler;
5451 static struct ctl_table *
5452 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5454 struct ctl_table *table = sd_alloc_ctl_entry(12);
5459 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5460 sizeof(long), 0644, proc_doulongvec_minmax);
5461 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5462 sizeof(long), 0644, proc_doulongvec_minmax);
5463 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5464 sizeof(int), 0644, proc_dointvec_minmax);
5465 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5466 sizeof(int), 0644, proc_dointvec_minmax);
5467 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5468 sizeof(int), 0644, proc_dointvec_minmax);
5469 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5470 sizeof(int), 0644, proc_dointvec_minmax);
5471 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5472 sizeof(int), 0644, proc_dointvec_minmax);
5473 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5474 sizeof(int), 0644, proc_dointvec_minmax);
5475 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5476 sizeof(int), 0644, proc_dointvec_minmax);
5477 set_table_entry(&table[9], "cache_nice_tries",
5478 &sd->cache_nice_tries,
5479 sizeof(int), 0644, proc_dointvec_minmax);
5480 set_table_entry(&table[10], "flags", &sd->flags,
5481 sizeof(int), 0644, proc_dointvec_minmax);
5482 /* &table[11] is terminator */
5487 static ctl_table * sd_alloc_ctl_cpu_table(int cpu)
5489 struct ctl_table *entry, *table;
5490 struct sched_domain *sd;
5491 int domain_num = 0, i;
5494 for_each_domain(cpu, sd)
5496 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5501 for_each_domain(cpu, sd) {
5502 snprintf(buf, 32, "domain%d", i);
5503 entry->procname = kstrdup(buf, GFP_KERNEL);
5505 entry->child = sd_alloc_ctl_domain_table(sd);
5512 static struct ctl_table_header *sd_sysctl_header;
5513 static void register_sched_domain_sysctl(void)
5515 int i, cpu_num = num_online_cpus();
5516 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5519 WARN_ON(sd_ctl_dir[0].child);
5520 sd_ctl_dir[0].child = entry;
5525 for_each_online_cpu(i) {
5526 snprintf(buf, 32, "cpu%d", i);
5527 entry->procname = kstrdup(buf, GFP_KERNEL);
5529 entry->child = sd_alloc_ctl_cpu_table(i);
5533 WARN_ON(sd_sysctl_header);
5534 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5537 /* may be called multiple times per register */
5538 static void unregister_sched_domain_sysctl(void)
5540 if (sd_sysctl_header)
5541 unregister_sysctl_table(sd_sysctl_header);
5542 sd_sysctl_header = NULL;
5543 if (sd_ctl_dir[0].child)
5544 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5547 static void register_sched_domain_sysctl(void)
5550 static void unregister_sched_domain_sysctl(void)
5556 * migration_call - callback that gets triggered when a CPU is added.
5557 * Here we can start up the necessary migration thread for the new CPU.
5559 static int __cpuinit
5560 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5562 struct task_struct *p;
5563 int cpu = (long)hcpu;
5564 unsigned long flags;
5568 case CPU_LOCK_ACQUIRE:
5569 mutex_lock(&sched_hotcpu_mutex);
5572 case CPU_UP_PREPARE:
5573 case CPU_UP_PREPARE_FROZEN:
5574 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5577 kthread_bind(p, cpu);
5578 /* Must be high prio: stop_machine expects to yield to it. */
5579 rq = task_rq_lock(p, &flags);
5580 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5581 task_rq_unlock(rq, &flags);
5582 cpu_rq(cpu)->migration_thread = p;
5586 case CPU_ONLINE_FROZEN:
5587 /* Strictly unnecessary, as first user will wake it. */
5588 wake_up_process(cpu_rq(cpu)->migration_thread);
5591 #ifdef CONFIG_HOTPLUG_CPU
5592 case CPU_UP_CANCELED:
5593 case CPU_UP_CANCELED_FROZEN:
5594 if (!cpu_rq(cpu)->migration_thread)
5596 /* Unbind it from offline cpu so it can run. Fall thru. */
5597 kthread_bind(cpu_rq(cpu)->migration_thread,
5598 any_online_cpu(cpu_online_map));
5599 kthread_stop(cpu_rq(cpu)->migration_thread);
5600 cpu_rq(cpu)->migration_thread = NULL;
5604 case CPU_DEAD_FROZEN:
5605 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5606 migrate_live_tasks(cpu);
5608 kthread_stop(rq->migration_thread);
5609 rq->migration_thread = NULL;
5610 /* Idle task back to normal (off runqueue, low prio) */
5611 spin_lock_irq(&rq->lock);
5612 update_rq_clock(rq);
5613 deactivate_task(rq, rq->idle, 0);
5614 rq->idle->static_prio = MAX_PRIO;
5615 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5616 rq->idle->sched_class = &idle_sched_class;
5617 migrate_dead_tasks(cpu);
5618 spin_unlock_irq(&rq->lock);
5620 migrate_nr_uninterruptible(rq);
5621 BUG_ON(rq->nr_running != 0);
5623 /* No need to migrate the tasks: it was best-effort if
5624 * they didn't take sched_hotcpu_mutex. Just wake up
5625 * the requestors. */
5626 spin_lock_irq(&rq->lock);
5627 while (!list_empty(&rq->migration_queue)) {
5628 struct migration_req *req;
5630 req = list_entry(rq->migration_queue.next,
5631 struct migration_req, list);
5632 list_del_init(&req->list);
5633 complete(&req->done);
5635 spin_unlock_irq(&rq->lock);
5638 case CPU_LOCK_RELEASE:
5639 mutex_unlock(&sched_hotcpu_mutex);
5645 /* Register at highest priority so that task migration (migrate_all_tasks)
5646 * happens before everything else.
5648 static struct notifier_block __cpuinitdata migration_notifier = {
5649 .notifier_call = migration_call,
5653 void __init migration_init(void)
5655 void *cpu = (void *)(long)smp_processor_id();
5658 /* Start one for the boot CPU: */
5659 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5660 BUG_ON(err == NOTIFY_BAD);
5661 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5662 register_cpu_notifier(&migration_notifier);
5668 /* Number of possible processor ids */
5669 int nr_cpu_ids __read_mostly = NR_CPUS;
5670 EXPORT_SYMBOL(nr_cpu_ids);
5672 #ifdef CONFIG_SCHED_DEBUG
5674 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5676 struct sched_group *group = sd->groups;
5677 cpumask_t groupmask;
5680 cpumask_scnprintf(str, NR_CPUS, sd->span);
5681 cpus_clear(groupmask);
5683 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5685 if (!(sd->flags & SD_LOAD_BALANCE)) {
5686 printk("does not load-balance\n");
5688 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5693 printk(KERN_CONT "span %s\n", str);
5695 if (!cpu_isset(cpu, sd->span)) {
5696 printk(KERN_ERR "ERROR: domain->span does not contain "
5699 if (!cpu_isset(cpu, group->cpumask)) {
5700 printk(KERN_ERR "ERROR: domain->groups does not contain"
5704 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5708 printk(KERN_ERR "ERROR: group is NULL\n");
5712 if (!group->__cpu_power) {
5713 printk(KERN_CONT "\n");
5714 printk(KERN_ERR "ERROR: domain->cpu_power not "
5719 if (!cpus_weight(group->cpumask)) {
5720 printk(KERN_CONT "\n");
5721 printk(KERN_ERR "ERROR: empty group\n");
5725 if (cpus_intersects(groupmask, group->cpumask)) {
5726 printk(KERN_CONT "\n");
5727 printk(KERN_ERR "ERROR: repeated CPUs\n");
5731 cpus_or(groupmask, groupmask, group->cpumask);
5733 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5734 printk(KERN_CONT " %s", str);
5736 group = group->next;
5737 } while (group != sd->groups);
5738 printk(KERN_CONT "\n");
5740 if (!cpus_equal(sd->span, groupmask))
5741 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5743 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5744 printk(KERN_ERR "ERROR: parent span is not a superset "
5745 "of domain->span\n");
5749 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5754 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5758 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5761 if (sched_domain_debug_one(sd, cpu, level))
5770 # define sched_domain_debug(sd, cpu) do { } while (0)
5773 static int sd_degenerate(struct sched_domain *sd)
5775 if (cpus_weight(sd->span) == 1)
5778 /* Following flags need at least 2 groups */
5779 if (sd->flags & (SD_LOAD_BALANCE |
5780 SD_BALANCE_NEWIDLE |
5784 SD_SHARE_PKG_RESOURCES)) {
5785 if (sd->groups != sd->groups->next)
5789 /* Following flags don't use groups */
5790 if (sd->flags & (SD_WAKE_IDLE |
5799 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5801 unsigned long cflags = sd->flags, pflags = parent->flags;
5803 if (sd_degenerate(parent))
5806 if (!cpus_equal(sd->span, parent->span))
5809 /* Does parent contain flags not in child? */
5810 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5811 if (cflags & SD_WAKE_AFFINE)
5812 pflags &= ~SD_WAKE_BALANCE;
5813 /* Flags needing groups don't count if only 1 group in parent */
5814 if (parent->groups == parent->groups->next) {
5815 pflags &= ~(SD_LOAD_BALANCE |
5816 SD_BALANCE_NEWIDLE |
5820 SD_SHARE_PKG_RESOURCES);
5822 if (~cflags & pflags)
5829 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5830 * hold the hotplug lock.
5832 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5834 struct rq *rq = cpu_rq(cpu);
5835 struct sched_domain *tmp;
5837 /* Remove the sched domains which do not contribute to scheduling. */
5838 for (tmp = sd; tmp; tmp = tmp->parent) {
5839 struct sched_domain *parent = tmp->parent;
5842 if (sd_parent_degenerate(tmp, parent)) {
5843 tmp->parent = parent->parent;
5845 parent->parent->child = tmp;
5849 if (sd && sd_degenerate(sd)) {
5855 sched_domain_debug(sd, cpu);
5857 rcu_assign_pointer(rq->sd, sd);
5860 /* cpus with isolated domains */
5861 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5863 /* Setup the mask of cpus configured for isolated domains */
5864 static int __init isolated_cpu_setup(char *str)
5866 int ints[NR_CPUS], i;
5868 str = get_options(str, ARRAY_SIZE(ints), ints);
5869 cpus_clear(cpu_isolated_map);
5870 for (i = 1; i <= ints[0]; i++)
5871 if (ints[i] < NR_CPUS)
5872 cpu_set(ints[i], cpu_isolated_map);
5876 __setup("isolcpus=", isolated_cpu_setup);
5879 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5880 * to a function which identifies what group(along with sched group) a CPU
5881 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5882 * (due to the fact that we keep track of groups covered with a cpumask_t).
5884 * init_sched_build_groups will build a circular linked list of the groups
5885 * covered by the given span, and will set each group's ->cpumask correctly,
5886 * and ->cpu_power to 0.
5889 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5890 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5891 struct sched_group **sg))
5893 struct sched_group *first = NULL, *last = NULL;
5894 cpumask_t covered = CPU_MASK_NONE;
5897 for_each_cpu_mask(i, span) {
5898 struct sched_group *sg;
5899 int group = group_fn(i, cpu_map, &sg);
5902 if (cpu_isset(i, covered))
5905 sg->cpumask = CPU_MASK_NONE;
5906 sg->__cpu_power = 0;
5908 for_each_cpu_mask(j, span) {
5909 if (group_fn(j, cpu_map, NULL) != group)
5912 cpu_set(j, covered);
5913 cpu_set(j, sg->cpumask);
5924 #define SD_NODES_PER_DOMAIN 16
5929 * find_next_best_node - find the next node to include in a sched_domain
5930 * @node: node whose sched_domain we're building
5931 * @used_nodes: nodes already in the sched_domain
5933 * Find the next node to include in a given scheduling domain. Simply
5934 * finds the closest node not already in the @used_nodes map.
5936 * Should use nodemask_t.
5938 static int find_next_best_node(int node, unsigned long *used_nodes)
5940 int i, n, val, min_val, best_node = 0;
5944 for (i = 0; i < MAX_NUMNODES; i++) {
5945 /* Start at @node */
5946 n = (node + i) % MAX_NUMNODES;
5948 if (!nr_cpus_node(n))
5951 /* Skip already used nodes */
5952 if (test_bit(n, used_nodes))
5955 /* Simple min distance search */
5956 val = node_distance(node, n);
5958 if (val < min_val) {
5964 set_bit(best_node, used_nodes);
5969 * sched_domain_node_span - get a cpumask for a node's sched_domain
5970 * @node: node whose cpumask we're constructing
5971 * @size: number of nodes to include in this span
5973 * Given a node, construct a good cpumask for its sched_domain to span. It
5974 * should be one that prevents unnecessary balancing, but also spreads tasks
5977 static cpumask_t sched_domain_node_span(int node)
5979 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5980 cpumask_t span, nodemask;
5984 bitmap_zero(used_nodes, MAX_NUMNODES);
5986 nodemask = node_to_cpumask(node);
5987 cpus_or(span, span, nodemask);
5988 set_bit(node, used_nodes);
5990 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5991 int next_node = find_next_best_node(node, used_nodes);
5993 nodemask = node_to_cpumask(next_node);
5994 cpus_or(span, span, nodemask);
6001 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6004 * SMT sched-domains:
6006 #ifdef CONFIG_SCHED_SMT
6007 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6008 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6010 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6011 struct sched_group **sg)
6014 *sg = &per_cpu(sched_group_cpus, cpu);
6020 * multi-core sched-domains:
6022 #ifdef CONFIG_SCHED_MC
6023 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6024 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6027 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6028 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6029 struct sched_group **sg)
6032 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6033 cpus_and(mask, mask, *cpu_map);
6034 group = first_cpu(mask);
6036 *sg = &per_cpu(sched_group_core, group);
6039 #elif defined(CONFIG_SCHED_MC)
6040 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6041 struct sched_group **sg)
6044 *sg = &per_cpu(sched_group_core, cpu);
6049 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6050 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6052 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6053 struct sched_group **sg)
6056 #ifdef CONFIG_SCHED_MC
6057 cpumask_t mask = cpu_coregroup_map(cpu);
6058 cpus_and(mask, mask, *cpu_map);
6059 group = first_cpu(mask);
6060 #elif defined(CONFIG_SCHED_SMT)
6061 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6062 cpus_and(mask, mask, *cpu_map);
6063 group = first_cpu(mask);
6068 *sg = &per_cpu(sched_group_phys, group);
6074 * The init_sched_build_groups can't handle what we want to do with node
6075 * groups, so roll our own. Now each node has its own list of groups which
6076 * gets dynamically allocated.
6078 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6079 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6081 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6082 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6084 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6085 struct sched_group **sg)
6087 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6090 cpus_and(nodemask, nodemask, *cpu_map);
6091 group = first_cpu(nodemask);
6094 *sg = &per_cpu(sched_group_allnodes, group);
6098 static void init_numa_sched_groups_power(struct sched_group *group_head)
6100 struct sched_group *sg = group_head;
6106 for_each_cpu_mask(j, sg->cpumask) {
6107 struct sched_domain *sd;
6109 sd = &per_cpu(phys_domains, j);
6110 if (j != first_cpu(sd->groups->cpumask)) {
6112 * Only add "power" once for each
6118 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6121 } while (sg != group_head);
6126 /* Free memory allocated for various sched_group structures */
6127 static void free_sched_groups(const cpumask_t *cpu_map)
6131 for_each_cpu_mask(cpu, *cpu_map) {
6132 struct sched_group **sched_group_nodes
6133 = sched_group_nodes_bycpu[cpu];
6135 if (!sched_group_nodes)
6138 for (i = 0; i < MAX_NUMNODES; i++) {
6139 cpumask_t nodemask = node_to_cpumask(i);
6140 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6142 cpus_and(nodemask, nodemask, *cpu_map);
6143 if (cpus_empty(nodemask))
6153 if (oldsg != sched_group_nodes[i])
6156 kfree(sched_group_nodes);
6157 sched_group_nodes_bycpu[cpu] = NULL;
6161 static void free_sched_groups(const cpumask_t *cpu_map)
6167 * Initialize sched groups cpu_power.
6169 * cpu_power indicates the capacity of sched group, which is used while
6170 * distributing the load between different sched groups in a sched domain.
6171 * Typically cpu_power for all the groups in a sched domain will be same unless
6172 * there are asymmetries in the topology. If there are asymmetries, group
6173 * having more cpu_power will pickup more load compared to the group having
6176 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6177 * the maximum number of tasks a group can handle in the presence of other idle
6178 * or lightly loaded groups in the same sched domain.
6180 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6182 struct sched_domain *child;
6183 struct sched_group *group;
6185 WARN_ON(!sd || !sd->groups);
6187 if (cpu != first_cpu(sd->groups->cpumask))
6192 sd->groups->__cpu_power = 0;
6195 * For perf policy, if the groups in child domain share resources
6196 * (for example cores sharing some portions of the cache hierarchy
6197 * or SMT), then set this domain groups cpu_power such that each group
6198 * can handle only one task, when there are other idle groups in the
6199 * same sched domain.
6201 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6203 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6204 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6209 * add cpu_power of each child group to this groups cpu_power
6211 group = child->groups;
6213 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6214 group = group->next;
6215 } while (group != child->groups);
6219 * Build sched domains for a given set of cpus and attach the sched domains
6220 * to the individual cpus
6222 static int build_sched_domains(const cpumask_t *cpu_map)
6226 struct sched_group **sched_group_nodes = NULL;
6227 int sd_allnodes = 0;
6230 * Allocate the per-node list of sched groups
6232 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6234 if (!sched_group_nodes) {
6235 printk(KERN_WARNING "Can not alloc sched group node list\n");
6238 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6242 * Set up domains for cpus specified by the cpu_map.
6244 for_each_cpu_mask(i, *cpu_map) {
6245 struct sched_domain *sd = NULL, *p;
6246 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6248 cpus_and(nodemask, nodemask, *cpu_map);
6251 if (cpus_weight(*cpu_map) >
6252 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6253 sd = &per_cpu(allnodes_domains, i);
6254 *sd = SD_ALLNODES_INIT;
6255 sd->span = *cpu_map;
6256 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6262 sd = &per_cpu(node_domains, i);
6264 sd->span = sched_domain_node_span(cpu_to_node(i));
6268 cpus_and(sd->span, sd->span, *cpu_map);
6272 sd = &per_cpu(phys_domains, i);
6274 sd->span = nodemask;
6278 cpu_to_phys_group(i, cpu_map, &sd->groups);
6280 #ifdef CONFIG_SCHED_MC
6282 sd = &per_cpu(core_domains, i);
6284 sd->span = cpu_coregroup_map(i);
6285 cpus_and(sd->span, sd->span, *cpu_map);
6288 cpu_to_core_group(i, cpu_map, &sd->groups);
6291 #ifdef CONFIG_SCHED_SMT
6293 sd = &per_cpu(cpu_domains, i);
6294 *sd = SD_SIBLING_INIT;
6295 sd->span = per_cpu(cpu_sibling_map, i);
6296 cpus_and(sd->span, sd->span, *cpu_map);
6299 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6303 #ifdef CONFIG_SCHED_SMT
6304 /* Set up CPU (sibling) groups */
6305 for_each_cpu_mask(i, *cpu_map) {
6306 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6307 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6308 if (i != first_cpu(this_sibling_map))
6311 init_sched_build_groups(this_sibling_map, cpu_map,
6316 #ifdef CONFIG_SCHED_MC
6317 /* Set up multi-core groups */
6318 for_each_cpu_mask(i, *cpu_map) {
6319 cpumask_t this_core_map = cpu_coregroup_map(i);
6320 cpus_and(this_core_map, this_core_map, *cpu_map);
6321 if (i != first_cpu(this_core_map))
6323 init_sched_build_groups(this_core_map, cpu_map,
6324 &cpu_to_core_group);
6328 /* Set up physical groups */
6329 for (i = 0; i < MAX_NUMNODES; i++) {
6330 cpumask_t nodemask = node_to_cpumask(i);
6332 cpus_and(nodemask, nodemask, *cpu_map);
6333 if (cpus_empty(nodemask))
6336 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6340 /* Set up node groups */
6342 init_sched_build_groups(*cpu_map, cpu_map,
6343 &cpu_to_allnodes_group);
6345 for (i = 0; i < MAX_NUMNODES; i++) {
6346 /* Set up node groups */
6347 struct sched_group *sg, *prev;
6348 cpumask_t nodemask = node_to_cpumask(i);
6349 cpumask_t domainspan;
6350 cpumask_t covered = CPU_MASK_NONE;
6353 cpus_and(nodemask, nodemask, *cpu_map);
6354 if (cpus_empty(nodemask)) {
6355 sched_group_nodes[i] = NULL;
6359 domainspan = sched_domain_node_span(i);
6360 cpus_and(domainspan, domainspan, *cpu_map);
6362 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6364 printk(KERN_WARNING "Can not alloc domain group for "
6368 sched_group_nodes[i] = sg;
6369 for_each_cpu_mask(j, nodemask) {
6370 struct sched_domain *sd;
6372 sd = &per_cpu(node_domains, j);
6375 sg->__cpu_power = 0;
6376 sg->cpumask = nodemask;
6378 cpus_or(covered, covered, nodemask);
6381 for (j = 0; j < MAX_NUMNODES; j++) {
6382 cpumask_t tmp, notcovered;
6383 int n = (i + j) % MAX_NUMNODES;
6385 cpus_complement(notcovered, covered);
6386 cpus_and(tmp, notcovered, *cpu_map);
6387 cpus_and(tmp, tmp, domainspan);
6388 if (cpus_empty(tmp))
6391 nodemask = node_to_cpumask(n);
6392 cpus_and(tmp, tmp, nodemask);
6393 if (cpus_empty(tmp))
6396 sg = kmalloc_node(sizeof(struct sched_group),
6400 "Can not alloc domain group for node %d\n", j);
6403 sg->__cpu_power = 0;
6405 sg->next = prev->next;
6406 cpus_or(covered, covered, tmp);
6413 /* Calculate CPU power for physical packages and nodes */
6414 #ifdef CONFIG_SCHED_SMT
6415 for_each_cpu_mask(i, *cpu_map) {
6416 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6418 init_sched_groups_power(i, sd);
6421 #ifdef CONFIG_SCHED_MC
6422 for_each_cpu_mask(i, *cpu_map) {
6423 struct sched_domain *sd = &per_cpu(core_domains, i);
6425 init_sched_groups_power(i, sd);
6429 for_each_cpu_mask(i, *cpu_map) {
6430 struct sched_domain *sd = &per_cpu(phys_domains, i);
6432 init_sched_groups_power(i, sd);
6436 for (i = 0; i < MAX_NUMNODES; i++)
6437 init_numa_sched_groups_power(sched_group_nodes[i]);
6440 struct sched_group *sg;
6442 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6443 init_numa_sched_groups_power(sg);
6447 /* Attach the domains */
6448 for_each_cpu_mask(i, *cpu_map) {
6449 struct sched_domain *sd;
6450 #ifdef CONFIG_SCHED_SMT
6451 sd = &per_cpu(cpu_domains, i);
6452 #elif defined(CONFIG_SCHED_MC)
6453 sd = &per_cpu(core_domains, i);
6455 sd = &per_cpu(phys_domains, i);
6457 cpu_attach_domain(sd, i);
6464 free_sched_groups(cpu_map);
6469 static cpumask_t *doms_cur; /* current sched domains */
6470 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6473 * Special case: If a kmalloc of a doms_cur partition (array of
6474 * cpumask_t) fails, then fallback to a single sched domain,
6475 * as determined by the single cpumask_t fallback_doms.
6477 static cpumask_t fallback_doms;
6480 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6481 * For now this just excludes isolated cpus, but could be used to
6482 * exclude other special cases in the future.
6484 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6489 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6491 doms_cur = &fallback_doms;
6492 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6493 err = build_sched_domains(doms_cur);
6494 register_sched_domain_sysctl();
6499 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6501 free_sched_groups(cpu_map);
6505 * Detach sched domains from a group of cpus specified in cpu_map
6506 * These cpus will now be attached to the NULL domain
6508 static void detach_destroy_domains(const cpumask_t *cpu_map)
6512 unregister_sched_domain_sysctl();
6514 for_each_cpu_mask(i, *cpu_map)
6515 cpu_attach_domain(NULL, i);
6516 synchronize_sched();
6517 arch_destroy_sched_domains(cpu_map);
6521 * Partition sched domains as specified by the 'ndoms_new'
6522 * cpumasks in the array doms_new[] of cpumasks. This compares
6523 * doms_new[] to the current sched domain partitioning, doms_cur[].
6524 * It destroys each deleted domain and builds each new domain.
6526 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6527 * The masks don't intersect (don't overlap.) We should setup one
6528 * sched domain for each mask. CPUs not in any of the cpumasks will
6529 * not be load balanced. If the same cpumask appears both in the
6530 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6533 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6534 * ownership of it and will kfree it when done with it. If the caller
6535 * failed the kmalloc call, then it can pass in doms_new == NULL,
6536 * and partition_sched_domains() will fallback to the single partition
6539 * Call with hotplug lock held
6541 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6545 /* always unregister in case we don't destroy any domains */
6546 unregister_sched_domain_sysctl();
6548 if (doms_new == NULL) {
6550 doms_new = &fallback_doms;
6551 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6554 /* Destroy deleted domains */
6555 for (i = 0; i < ndoms_cur; i++) {
6556 for (j = 0; j < ndoms_new; j++) {
6557 if (cpus_equal(doms_cur[i], doms_new[j]))
6560 /* no match - a current sched domain not in new doms_new[] */
6561 detach_destroy_domains(doms_cur + i);
6566 /* Build new domains */
6567 for (i = 0; i < ndoms_new; i++) {
6568 for (j = 0; j < ndoms_cur; j++) {
6569 if (cpus_equal(doms_new[i], doms_cur[j]))
6572 /* no match - add a new doms_new */
6573 build_sched_domains(doms_new + i);
6578 /* Remember the new sched domains */
6579 if (doms_cur != &fallback_doms)
6581 doms_cur = doms_new;
6582 ndoms_cur = ndoms_new;
6584 register_sched_domain_sysctl();
6587 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6588 static int arch_reinit_sched_domains(void)
6592 mutex_lock(&sched_hotcpu_mutex);
6593 detach_destroy_domains(&cpu_online_map);
6594 err = arch_init_sched_domains(&cpu_online_map);
6595 mutex_unlock(&sched_hotcpu_mutex);
6600 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6604 if (buf[0] != '0' && buf[0] != '1')
6608 sched_smt_power_savings = (buf[0] == '1');
6610 sched_mc_power_savings = (buf[0] == '1');
6612 ret = arch_reinit_sched_domains();
6614 return ret ? ret : count;
6617 #ifdef CONFIG_SCHED_MC
6618 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6620 return sprintf(page, "%u\n", sched_mc_power_savings);
6622 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6623 const char *buf, size_t count)
6625 return sched_power_savings_store(buf, count, 0);
6627 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6628 sched_mc_power_savings_store);
6631 #ifdef CONFIG_SCHED_SMT
6632 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6634 return sprintf(page, "%u\n", sched_smt_power_savings);
6636 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6637 const char *buf, size_t count)
6639 return sched_power_savings_store(buf, count, 1);
6641 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6642 sched_smt_power_savings_store);
6645 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6649 #ifdef CONFIG_SCHED_SMT
6651 err = sysfs_create_file(&cls->kset.kobj,
6652 &attr_sched_smt_power_savings.attr);
6654 #ifdef CONFIG_SCHED_MC
6655 if (!err && mc_capable())
6656 err = sysfs_create_file(&cls->kset.kobj,
6657 &attr_sched_mc_power_savings.attr);
6664 * Force a reinitialization of the sched domains hierarchy. The domains
6665 * and groups cannot be updated in place without racing with the balancing
6666 * code, so we temporarily attach all running cpus to the NULL domain
6667 * which will prevent rebalancing while the sched domains are recalculated.
6669 static int update_sched_domains(struct notifier_block *nfb,
6670 unsigned long action, void *hcpu)
6673 case CPU_UP_PREPARE:
6674 case CPU_UP_PREPARE_FROZEN:
6675 case CPU_DOWN_PREPARE:
6676 case CPU_DOWN_PREPARE_FROZEN:
6677 detach_destroy_domains(&cpu_online_map);
6680 case CPU_UP_CANCELED:
6681 case CPU_UP_CANCELED_FROZEN:
6682 case CPU_DOWN_FAILED:
6683 case CPU_DOWN_FAILED_FROZEN:
6685 case CPU_ONLINE_FROZEN:
6687 case CPU_DEAD_FROZEN:
6689 * Fall through and re-initialise the domains.
6696 /* The hotplug lock is already held by cpu_up/cpu_down */
6697 arch_init_sched_domains(&cpu_online_map);
6702 void __init sched_init_smp(void)
6704 cpumask_t non_isolated_cpus;
6706 mutex_lock(&sched_hotcpu_mutex);
6707 arch_init_sched_domains(&cpu_online_map);
6708 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6709 if (cpus_empty(non_isolated_cpus))
6710 cpu_set(smp_processor_id(), non_isolated_cpus);
6711 mutex_unlock(&sched_hotcpu_mutex);
6712 /* XXX: Theoretical race here - CPU may be hotplugged now */
6713 hotcpu_notifier(update_sched_domains, 0);
6715 /* Move init over to a non-isolated CPU */
6716 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6718 sched_init_granularity();
6721 void __init sched_init_smp(void)
6723 sched_init_granularity();
6725 #endif /* CONFIG_SMP */
6727 int in_sched_functions(unsigned long addr)
6729 /* Linker adds these: start and end of __sched functions */
6730 extern char __sched_text_start[], __sched_text_end[];
6732 return in_lock_functions(addr) ||
6733 (addr >= (unsigned long)__sched_text_start
6734 && addr < (unsigned long)__sched_text_end);
6737 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6739 cfs_rq->tasks_timeline = RB_ROOT;
6740 #ifdef CONFIG_FAIR_GROUP_SCHED
6743 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6746 void __init sched_init(void)
6748 int highest_cpu = 0;
6751 for_each_possible_cpu(i) {
6752 struct rt_prio_array *array;
6756 spin_lock_init(&rq->lock);
6757 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6760 init_cfs_rq(&rq->cfs, rq);
6761 #ifdef CONFIG_FAIR_GROUP_SCHED
6762 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6764 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6765 struct sched_entity *se =
6766 &per_cpu(init_sched_entity, i);
6768 init_cfs_rq_p[i] = cfs_rq;
6769 init_cfs_rq(cfs_rq, rq);
6770 cfs_rq->tg = &init_task_group;
6771 list_add(&cfs_rq->leaf_cfs_rq_list,
6772 &rq->leaf_cfs_rq_list);
6774 init_sched_entity_p[i] = se;
6775 se->cfs_rq = &rq->cfs;
6777 se->load.weight = init_task_group_load;
6778 se->load.inv_weight =
6779 div64_64(1ULL<<32, init_task_group_load);
6782 init_task_group.shares = init_task_group_load;
6783 spin_lock_init(&init_task_group.lock);
6786 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6787 rq->cpu_load[j] = 0;
6790 rq->active_balance = 0;
6791 rq->next_balance = jiffies;
6794 rq->migration_thread = NULL;
6795 INIT_LIST_HEAD(&rq->migration_queue);
6797 atomic_set(&rq->nr_iowait, 0);
6799 array = &rq->rt.active;
6800 for (j = 0; j < MAX_RT_PRIO; j++) {
6801 INIT_LIST_HEAD(array->queue + j);
6802 __clear_bit(j, array->bitmap);
6805 /* delimiter for bitsearch: */
6806 __set_bit(MAX_RT_PRIO, array->bitmap);
6809 set_load_weight(&init_task);
6811 #ifdef CONFIG_PREEMPT_NOTIFIERS
6812 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6816 nr_cpu_ids = highest_cpu + 1;
6817 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6820 #ifdef CONFIG_RT_MUTEXES
6821 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6825 * The boot idle thread does lazy MMU switching as well:
6827 atomic_inc(&init_mm.mm_count);
6828 enter_lazy_tlb(&init_mm, current);
6831 * Make us the idle thread. Technically, schedule() should not be
6832 * called from this thread, however somewhere below it might be,
6833 * but because we are the idle thread, we just pick up running again
6834 * when this runqueue becomes "idle".
6836 init_idle(current, smp_processor_id());
6838 * During early bootup we pretend to be a normal task:
6840 current->sched_class = &fair_sched_class;
6843 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6844 void __might_sleep(char *file, int line)
6847 static unsigned long prev_jiffy; /* ratelimiting */
6849 if ((in_atomic() || irqs_disabled()) &&
6850 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6851 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6853 prev_jiffy = jiffies;
6854 printk(KERN_ERR "BUG: sleeping function called from invalid"
6855 " context at %s:%d\n", file, line);
6856 printk("in_atomic():%d, irqs_disabled():%d\n",
6857 in_atomic(), irqs_disabled());
6858 debug_show_held_locks(current);
6859 if (irqs_disabled())
6860 print_irqtrace_events(current);
6865 EXPORT_SYMBOL(__might_sleep);
6868 #ifdef CONFIG_MAGIC_SYSRQ
6869 static void normalize_task(struct rq *rq, struct task_struct *p)
6872 update_rq_clock(rq);
6873 on_rq = p->se.on_rq;
6875 deactivate_task(rq, p, 0);
6876 __setscheduler(rq, p, SCHED_NORMAL, 0);
6878 activate_task(rq, p, 0);
6879 resched_task(rq->curr);
6883 void normalize_rt_tasks(void)
6885 struct task_struct *g, *p;
6886 unsigned long flags;
6889 read_lock_irq(&tasklist_lock);
6890 do_each_thread(g, p) {
6892 * Only normalize user tasks:
6897 p->se.exec_start = 0;
6898 #ifdef CONFIG_SCHEDSTATS
6899 p->se.wait_start = 0;
6900 p->se.sleep_start = 0;
6901 p->se.block_start = 0;
6903 task_rq(p)->clock = 0;
6907 * Renice negative nice level userspace
6910 if (TASK_NICE(p) < 0 && p->mm)
6911 set_user_nice(p, 0);
6915 spin_lock_irqsave(&p->pi_lock, flags);
6916 rq = __task_rq_lock(p);
6918 normalize_task(rq, p);
6920 __task_rq_unlock(rq);
6921 spin_unlock_irqrestore(&p->pi_lock, flags);
6922 } while_each_thread(g, p);
6924 read_unlock_irq(&tasklist_lock);
6927 #endif /* CONFIG_MAGIC_SYSRQ */
6931 * These functions are only useful for the IA64 MCA handling.
6933 * They can only be called when the whole system has been
6934 * stopped - every CPU needs to be quiescent, and no scheduling
6935 * activity can take place. Using them for anything else would
6936 * be a serious bug, and as a result, they aren't even visible
6937 * under any other configuration.
6941 * curr_task - return the current task for a given cpu.
6942 * @cpu: the processor in question.
6944 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6946 struct task_struct *curr_task(int cpu)
6948 return cpu_curr(cpu);
6952 * set_curr_task - set the current task for a given cpu.
6953 * @cpu: the processor in question.
6954 * @p: the task pointer to set.
6956 * Description: This function must only be used when non-maskable interrupts
6957 * are serviced on a separate stack. It allows the architecture to switch the
6958 * notion of the current task on a cpu in a non-blocking manner. This function
6959 * must be called with all CPU's synchronized, and interrupts disabled, the
6960 * and caller must save the original value of the current task (see
6961 * curr_task() above) and restore that value before reenabling interrupts and
6962 * re-starting the system.
6964 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6966 void set_curr_task(int cpu, struct task_struct *p)
6973 #ifdef CONFIG_FAIR_GROUP_SCHED
6975 /* allocate runqueue etc for a new task group */
6976 struct task_group *sched_create_group(void)
6978 struct task_group *tg;
6979 struct cfs_rq *cfs_rq;
6980 struct sched_entity *se;
6984 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6986 return ERR_PTR(-ENOMEM);
6988 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6991 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6995 for_each_possible_cpu(i) {
6998 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7003 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7008 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7009 memset(se, 0, sizeof(struct sched_entity));
7011 tg->cfs_rq[i] = cfs_rq;
7012 init_cfs_rq(cfs_rq, rq);
7016 se->cfs_rq = &rq->cfs;
7018 se->load.weight = NICE_0_LOAD;
7019 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7023 for_each_possible_cpu(i) {
7025 cfs_rq = tg->cfs_rq[i];
7026 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7029 tg->shares = NICE_0_LOAD;
7030 spin_lock_init(&tg->lock);
7035 for_each_possible_cpu(i) {
7037 kfree(tg->cfs_rq[i]);
7045 return ERR_PTR(-ENOMEM);
7048 /* rcu callback to free various structures associated with a task group */
7049 static void free_sched_group(struct rcu_head *rhp)
7051 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7052 struct cfs_rq *cfs_rq;
7053 struct sched_entity *se;
7056 /* now it should be safe to free those cfs_rqs */
7057 for_each_possible_cpu(i) {
7058 cfs_rq = tg->cfs_rq[i];
7070 /* Destroy runqueue etc associated with a task group */
7071 void sched_destroy_group(struct task_group *tg)
7073 struct cfs_rq *cfs_rq = NULL;
7076 for_each_possible_cpu(i) {
7077 cfs_rq = tg->cfs_rq[i];
7078 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7083 /* wait for possible concurrent references to cfs_rqs complete */
7084 call_rcu(&tg->rcu, free_sched_group);
7087 /* change task's runqueue when it moves between groups.
7088 * The caller of this function should have put the task in its new group
7089 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7090 * reflect its new group.
7092 void sched_move_task(struct task_struct *tsk)
7095 unsigned long flags;
7098 rq = task_rq_lock(tsk, &flags);
7100 if (tsk->sched_class != &fair_sched_class)
7103 update_rq_clock(rq);
7105 running = task_running(rq, tsk);
7106 on_rq = tsk->se.on_rq;
7109 dequeue_task(rq, tsk, 0);
7110 if (unlikely(running))
7111 tsk->sched_class->put_prev_task(rq, tsk);
7114 set_task_cfs_rq(tsk);
7117 if (unlikely(running))
7118 tsk->sched_class->set_curr_task(rq);
7119 enqueue_task(rq, tsk, 0);
7123 task_rq_unlock(rq, &flags);
7126 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7128 struct cfs_rq *cfs_rq = se->cfs_rq;
7129 struct rq *rq = cfs_rq->rq;
7132 spin_lock_irq(&rq->lock);
7136 dequeue_entity(cfs_rq, se, 0);
7138 se->load.weight = shares;
7139 se->load.inv_weight = div64_64((1ULL<<32), shares);
7142 enqueue_entity(cfs_rq, se, 0);
7144 spin_unlock_irq(&rq->lock);
7147 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7151 spin_lock(&tg->lock);
7152 if (tg->shares == shares)
7155 tg->shares = shares;
7156 for_each_possible_cpu(i)
7157 set_se_shares(tg->se[i], shares);
7160 spin_unlock(&tg->lock);
7164 unsigned long sched_group_shares(struct task_group *tg)
7169 #endif /* CONFIG_FAIR_GROUP_SCHED */
7171 #ifdef CONFIG_FAIR_CGROUP_SCHED
7173 /* return corresponding task_group object of a cgroup */
7174 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7176 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7177 struct task_group, css);
7180 static struct cgroup_subsys_state *
7181 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7183 struct task_group *tg;
7185 if (!cgrp->parent) {
7186 /* This is early initialization for the top cgroup */
7187 init_task_group.css.cgroup = cgrp;
7188 return &init_task_group.css;
7191 /* we support only 1-level deep hierarchical scheduler atm */
7192 if (cgrp->parent->parent)
7193 return ERR_PTR(-EINVAL);
7195 tg = sched_create_group();
7197 return ERR_PTR(-ENOMEM);
7199 /* Bind the cgroup to task_group object we just created */
7200 tg->css.cgroup = cgrp;
7205 static void cpu_cgroup_destroy(struct cgroup_subsys *ss,
7206 struct cgroup *cgrp)
7208 struct task_group *tg = cgroup_tg(cgrp);
7210 sched_destroy_group(tg);
7213 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss,
7214 struct cgroup *cgrp, struct task_struct *tsk)
7216 /* We don't support RT-tasks being in separate groups */
7217 if (tsk->sched_class != &fair_sched_class)
7224 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7225 struct cgroup *old_cont, struct task_struct *tsk)
7227 sched_move_task(tsk);
7230 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7233 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7236 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7238 struct task_group *tg = cgroup_tg(cgrp);
7240 return (u64) tg->shares;
7243 static u64 cpu_usage_read(struct cgroup *cgrp, struct cftype *cft)
7245 struct task_group *tg = cgroup_tg(cgrp);
7246 unsigned long flags;
7250 for_each_possible_cpu(i) {
7252 * Lock to prevent races with updating 64-bit counters
7255 spin_lock_irqsave(&cpu_rq(i)->lock, flags);
7256 res += tg->se[i]->sum_exec_runtime;
7257 spin_unlock_irqrestore(&cpu_rq(i)->lock, flags);
7259 /* Convert from ns to ms */
7260 do_div(res, NSEC_PER_MSEC);
7265 static struct cftype cpu_files[] = {
7268 .read_uint = cpu_shares_read_uint,
7269 .write_uint = cpu_shares_write_uint,
7273 .read_uint = cpu_usage_read,
7277 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7279 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7282 struct cgroup_subsys cpu_cgroup_subsys = {
7284 .create = cpu_cgroup_create,
7285 .destroy = cpu_cgroup_destroy,
7286 .can_attach = cpu_cgroup_can_attach,
7287 .attach = cpu_cgroup_attach,
7288 .populate = cpu_cgroup_populate,
7289 .subsys_id = cpu_cgroup_subsys_id,
7293 #endif /* CONFIG_FAIR_CGROUP_SCHED */