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 * (1000000000 / 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) / (1000000000 / HZ))
103 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
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 */
177 /* Default task group's sched entity on each cpu */
178 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
179 /* Default task group's cfs_rq on each cpu */
180 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
182 static struct sched_entity *init_sched_entity_p[NR_CPUS];
183 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
185 /* Default task group.
186 * Every task in system belong to this group at bootup.
188 struct task_group init_task_group = {
189 .se = init_sched_entity_p,
190 .cfs_rq = init_cfs_rq_p,
193 #ifdef CONFIG_FAIR_USER_SCHED
194 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
196 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
199 static int init_task_group_load = INIT_TASK_GRP_LOAD;
201 /* return group to which a task belongs */
202 static inline struct task_group *task_group(struct task_struct *p)
204 struct task_group *tg;
206 #ifdef CONFIG_FAIR_USER_SCHED
208 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
209 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
210 struct task_group, css);
212 tg = &init_task_group;
218 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
219 static inline void set_task_cfs_rq(struct task_struct *p)
221 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
222 p->se.parent = task_group(p)->se[task_cpu(p)];
227 static inline void set_task_cfs_rq(struct task_struct *p) { }
229 #endif /* CONFIG_FAIR_GROUP_SCHED */
231 /* CFS-related fields in a runqueue */
233 struct load_weight load;
234 unsigned long nr_running;
239 struct rb_root tasks_timeline;
240 struct rb_node *rb_leftmost;
241 struct rb_node *rb_load_balance_curr;
242 /* 'curr' points to currently running entity on this cfs_rq.
243 * It is set to NULL otherwise (i.e when none are currently running).
245 struct sched_entity *curr;
247 unsigned long nr_spread_over;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
252 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
253 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
254 * (like users, containers etc.)
256 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
257 * list is used during load balance.
259 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
260 struct task_group *tg; /* group that "owns" this runqueue */
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,
463 SCHED_FEAT_PREEMPT_RESTRICT = 32,
466 const_debug unsigned int sysctl_sched_features =
467 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
468 SCHED_FEAT_START_DEBIT * 1 |
469 SCHED_FEAT_TREE_AVG * 0 |
470 SCHED_FEAT_APPROX_AVG * 0 |
471 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
472 SCHED_FEAT_PREEMPT_RESTRICT * 1;
474 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
477 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
478 * clock constructed from sched_clock():
480 unsigned long long cpu_clock(int cpu)
482 unsigned long long now;
486 local_irq_save(flags);
490 local_irq_restore(flags);
494 EXPORT_SYMBOL_GPL(cpu_clock);
496 #ifndef prepare_arch_switch
497 # define prepare_arch_switch(next) do { } while (0)
499 #ifndef finish_arch_switch
500 # define finish_arch_switch(prev) do { } while (0)
503 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
504 static inline int task_running(struct rq *rq, struct task_struct *p)
506 return rq->curr == p;
509 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
513 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
515 #ifdef CONFIG_DEBUG_SPINLOCK
516 /* this is a valid case when another task releases the spinlock */
517 rq->lock.owner = current;
520 * If we are tracking spinlock dependencies then we have to
521 * fix up the runqueue lock - which gets 'carried over' from
524 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
526 spin_unlock_irq(&rq->lock);
529 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
530 static inline int task_running(struct rq *rq, struct task_struct *p)
535 return rq->curr == p;
539 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
543 * We can optimise this out completely for !SMP, because the
544 * SMP rebalancing from interrupt is the only thing that cares
549 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
550 spin_unlock_irq(&rq->lock);
552 spin_unlock(&rq->lock);
556 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
560 * After ->oncpu is cleared, the task can be moved to a different CPU.
561 * We must ensure this doesn't happen until the switch is completely
567 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
571 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
574 * __task_rq_lock - lock the runqueue a given task resides on.
575 * Must be called interrupts disabled.
577 static inline struct rq *__task_rq_lock(struct task_struct *p)
581 struct rq *rq = task_rq(p);
582 spin_lock(&rq->lock);
583 if (likely(rq == task_rq(p)))
585 spin_unlock(&rq->lock);
590 * task_rq_lock - lock the runqueue a given task resides on and disable
591 * interrupts. Note the ordering: we can safely lookup the task_rq without
592 * explicitly disabling preemption.
594 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
600 local_irq_save(*flags);
602 spin_lock(&rq->lock);
603 if (likely(rq == task_rq(p)))
605 spin_unlock_irqrestore(&rq->lock, *flags);
609 static void __task_rq_unlock(struct rq *rq)
612 spin_unlock(&rq->lock);
615 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
618 spin_unlock_irqrestore(&rq->lock, *flags);
622 * this_rq_lock - lock this runqueue and disable interrupts.
624 static struct rq *this_rq_lock(void)
631 spin_lock(&rq->lock);
637 * We are going deep-idle (irqs are disabled):
639 void sched_clock_idle_sleep_event(void)
641 struct rq *rq = cpu_rq(smp_processor_id());
643 spin_lock(&rq->lock);
644 __update_rq_clock(rq);
645 spin_unlock(&rq->lock);
646 rq->clock_deep_idle_events++;
648 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
651 * We just idled delta nanoseconds (called with irqs disabled):
653 void sched_clock_idle_wakeup_event(u64 delta_ns)
655 struct rq *rq = cpu_rq(smp_processor_id());
656 u64 now = sched_clock();
658 rq->idle_clock += delta_ns;
660 * Override the previous timestamp and ignore all
661 * sched_clock() deltas that occured while we idled,
662 * and use the PM-provided delta_ns to advance the
665 spin_lock(&rq->lock);
666 rq->prev_clock_raw = now;
667 rq->clock += delta_ns;
668 spin_unlock(&rq->lock);
670 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
673 * resched_task - mark a task 'to be rescheduled now'.
675 * On UP this means the setting of the need_resched flag, on SMP it
676 * might also involve a cross-CPU call to trigger the scheduler on
681 #ifndef tsk_is_polling
682 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
685 static void resched_task(struct task_struct *p)
689 assert_spin_locked(&task_rq(p)->lock);
691 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
694 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
697 if (cpu == smp_processor_id())
700 /* NEED_RESCHED must be visible before we test polling */
702 if (!tsk_is_polling(p))
703 smp_send_reschedule(cpu);
706 static void resched_cpu(int cpu)
708 struct rq *rq = cpu_rq(cpu);
711 if (!spin_trylock_irqsave(&rq->lock, flags))
713 resched_task(cpu_curr(cpu));
714 spin_unlock_irqrestore(&rq->lock, flags);
717 static inline void resched_task(struct task_struct *p)
719 assert_spin_locked(&task_rq(p)->lock);
720 set_tsk_need_resched(p);
724 #if BITS_PER_LONG == 32
725 # define WMULT_CONST (~0UL)
727 # define WMULT_CONST (1UL << 32)
730 #define WMULT_SHIFT 32
733 * Shift right and round:
735 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
738 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
739 struct load_weight *lw)
743 if (unlikely(!lw->inv_weight))
744 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
746 tmp = (u64)delta_exec * weight;
748 * Check whether we'd overflow the 64-bit multiplication:
750 if (unlikely(tmp > WMULT_CONST))
751 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
754 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
756 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
759 static inline unsigned long
760 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
762 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
765 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
770 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
776 * To aid in avoiding the subversion of "niceness" due to uneven distribution
777 * of tasks with abnormal "nice" values across CPUs the contribution that
778 * each task makes to its run queue's load is weighted according to its
779 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
780 * scaled version of the new time slice allocation that they receive on time
784 #define WEIGHT_IDLEPRIO 2
785 #define WMULT_IDLEPRIO (1 << 31)
788 * Nice levels are multiplicative, with a gentle 10% change for every
789 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
790 * nice 1, it will get ~10% less CPU time than another CPU-bound task
791 * that remained on nice 0.
793 * The "10% effect" is relative and cumulative: from _any_ nice level,
794 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
795 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
796 * If a task goes up by ~10% and another task goes down by ~10% then
797 * the relative distance between them is ~25%.)
799 static const int prio_to_weight[40] = {
800 /* -20 */ 88761, 71755, 56483, 46273, 36291,
801 /* -15 */ 29154, 23254, 18705, 14949, 11916,
802 /* -10 */ 9548, 7620, 6100, 4904, 3906,
803 /* -5 */ 3121, 2501, 1991, 1586, 1277,
804 /* 0 */ 1024, 820, 655, 526, 423,
805 /* 5 */ 335, 272, 215, 172, 137,
806 /* 10 */ 110, 87, 70, 56, 45,
807 /* 15 */ 36, 29, 23, 18, 15,
811 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
813 * In cases where the weight does not change often, we can use the
814 * precalculated inverse to speed up arithmetics by turning divisions
815 * into multiplications:
817 static const u32 prio_to_wmult[40] = {
818 /* -20 */ 48388, 59856, 76040, 92818, 118348,
819 /* -15 */ 147320, 184698, 229616, 287308, 360437,
820 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
821 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
822 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
823 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
824 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
825 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
828 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
831 * runqueue iterator, to support SMP load-balancing between different
832 * scheduling classes, without having to expose their internal data
833 * structures to the load-balancing proper:
837 struct task_struct *(*start)(void *);
838 struct task_struct *(*next)(void *);
843 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
844 unsigned long max_load_move, struct sched_domain *sd,
845 enum cpu_idle_type idle, int *all_pinned,
846 int *this_best_prio, struct rq_iterator *iterator);
849 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
850 struct sched_domain *sd, enum cpu_idle_type idle,
851 struct rq_iterator *iterator);
853 static inline unsigned long
854 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
855 unsigned long max_load_move, struct sched_domain *sd,
856 enum cpu_idle_type idle, int *all_pinned,
857 int *this_best_prio, struct rq_iterator *iterator)
863 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
864 struct sched_domain *sd, enum cpu_idle_type idle,
865 struct rq_iterator *iterator)
871 #include "sched_stats.h"
872 #include "sched_idletask.c"
873 #include "sched_fair.c"
874 #include "sched_rt.c"
875 #ifdef CONFIG_SCHED_DEBUG
876 # include "sched_debug.c"
879 #define sched_class_highest (&rt_sched_class)
882 * Update delta_exec, delta_fair fields for rq.
884 * delta_fair clock advances at a rate inversely proportional to
885 * total load (rq->load.weight) on the runqueue, while
886 * delta_exec advances at the same rate as wall-clock (provided
889 * delta_exec / delta_fair is a measure of the (smoothened) load on this
890 * runqueue over any given interval. This (smoothened) load is used
891 * during load balance.
893 * This function is called /before/ updating rq->load
894 * and when switching tasks.
896 static inline void inc_load(struct rq *rq, const struct task_struct *p)
898 update_load_add(&rq->load, p->se.load.weight);
901 static inline void dec_load(struct rq *rq, const struct task_struct *p)
903 update_load_sub(&rq->load, p->se.load.weight);
906 static void inc_nr_running(struct task_struct *p, struct rq *rq)
912 static void dec_nr_running(struct task_struct *p, struct rq *rq)
918 static void set_load_weight(struct task_struct *p)
920 if (task_has_rt_policy(p)) {
921 p->se.load.weight = prio_to_weight[0] * 2;
922 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
927 * SCHED_IDLE tasks get minimal weight:
929 if (p->policy == SCHED_IDLE) {
930 p->se.load.weight = WEIGHT_IDLEPRIO;
931 p->se.load.inv_weight = WMULT_IDLEPRIO;
935 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
936 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
939 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
941 sched_info_queued(p);
942 p->sched_class->enqueue_task(rq, p, wakeup);
946 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
948 p->sched_class->dequeue_task(rq, p, sleep);
953 * __normal_prio - return the priority that is based on the static prio
955 static inline int __normal_prio(struct task_struct *p)
957 return p->static_prio;
961 * Calculate the expected normal priority: i.e. priority
962 * without taking RT-inheritance into account. Might be
963 * boosted by interactivity modifiers. Changes upon fork,
964 * setprio syscalls, and whenever the interactivity
965 * estimator recalculates.
967 static inline int normal_prio(struct task_struct *p)
971 if (task_has_rt_policy(p))
972 prio = MAX_RT_PRIO-1 - p->rt_priority;
974 prio = __normal_prio(p);
979 * Calculate the current priority, i.e. the priority
980 * taken into account by the scheduler. This value might
981 * be boosted by RT tasks, or might be boosted by
982 * interactivity modifiers. Will be RT if the task got
983 * RT-boosted. If not then it returns p->normal_prio.
985 static int effective_prio(struct task_struct *p)
987 p->normal_prio = normal_prio(p);
989 * If we are RT tasks or we were boosted to RT priority,
990 * keep the priority unchanged. Otherwise, update priority
991 * to the normal priority:
993 if (!rt_prio(p->prio))
994 return p->normal_prio;
999 * activate_task - move a task to the runqueue.
1001 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1003 if (p->state == TASK_UNINTERRUPTIBLE)
1004 rq->nr_uninterruptible--;
1006 enqueue_task(rq, p, wakeup);
1007 inc_nr_running(p, rq);
1011 * deactivate_task - remove a task from the runqueue.
1013 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1015 if (p->state == TASK_UNINTERRUPTIBLE)
1016 rq->nr_uninterruptible++;
1018 dequeue_task(rq, p, sleep);
1019 dec_nr_running(p, rq);
1023 * task_curr - is this task currently executing on a CPU?
1024 * @p: the task in question.
1026 inline int task_curr(const struct task_struct *p)
1028 return cpu_curr(task_cpu(p)) == p;
1031 /* Used instead of source_load when we know the type == 0 */
1032 unsigned long weighted_cpuload(const int cpu)
1034 return cpu_rq(cpu)->load.weight;
1037 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1040 task_thread_info(p)->cpu = cpu;
1048 * Is this task likely cache-hot:
1051 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1055 if (p->sched_class != &fair_sched_class)
1058 if (sysctl_sched_migration_cost == -1)
1060 if (sysctl_sched_migration_cost == 0)
1063 delta = now - p->se.exec_start;
1065 return delta < (s64)sysctl_sched_migration_cost;
1069 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1071 int old_cpu = task_cpu(p);
1072 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1073 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1074 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1077 clock_offset = old_rq->clock - new_rq->clock;
1079 #ifdef CONFIG_SCHEDSTATS
1080 if (p->se.wait_start)
1081 p->se.wait_start -= clock_offset;
1082 if (p->se.sleep_start)
1083 p->se.sleep_start -= clock_offset;
1084 if (p->se.block_start)
1085 p->se.block_start -= clock_offset;
1086 if (old_cpu != new_cpu) {
1087 schedstat_inc(p, se.nr_migrations);
1088 if (task_hot(p, old_rq->clock, NULL))
1089 schedstat_inc(p, se.nr_forced2_migrations);
1092 p->se.vruntime -= old_cfsrq->min_vruntime -
1093 new_cfsrq->min_vruntime;
1095 __set_task_cpu(p, new_cpu);
1098 struct migration_req {
1099 struct list_head list;
1101 struct task_struct *task;
1104 struct completion done;
1108 * The task's runqueue lock must be held.
1109 * Returns true if you have to wait for migration thread.
1112 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1114 struct rq *rq = task_rq(p);
1117 * If the task is not on a runqueue (and not running), then
1118 * it is sufficient to simply update the task's cpu field.
1120 if (!p->se.on_rq && !task_running(rq, p)) {
1121 set_task_cpu(p, dest_cpu);
1125 init_completion(&req->done);
1127 req->dest_cpu = dest_cpu;
1128 list_add(&req->list, &rq->migration_queue);
1134 * wait_task_inactive - wait for a thread to unschedule.
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1142 void wait_task_inactive(struct task_struct *p)
1144 unsigned long flags;
1150 * We do the initial early heuristics without holding
1151 * any task-queue locks at all. We'll only try to get
1152 * the runqueue lock when things look like they will
1158 * If the task is actively running on another CPU
1159 * still, just relax and busy-wait without holding
1162 * NOTE! Since we don't hold any locks, it's not
1163 * even sure that "rq" stays as the right runqueue!
1164 * But we don't care, since "task_running()" will
1165 * return false if the runqueue has changed and p
1166 * is actually now running somewhere else!
1168 while (task_running(rq, p))
1172 * Ok, time to look more closely! We need the rq
1173 * lock now, to be *sure*. If we're wrong, we'll
1174 * just go back and repeat.
1176 rq = task_rq_lock(p, &flags);
1177 running = task_running(rq, p);
1178 on_rq = p->se.on_rq;
1179 task_rq_unlock(rq, &flags);
1182 * Was it really running after all now that we
1183 * checked with the proper locks actually held?
1185 * Oops. Go back and try again..
1187 if (unlikely(running)) {
1193 * It's not enough that it's not actively running,
1194 * it must be off the runqueue _entirely_, and not
1197 * So if it wa still runnable (but just not actively
1198 * running right now), it's preempted, and we should
1199 * yield - it could be a while.
1201 if (unlikely(on_rq)) {
1202 schedule_timeout_uninterruptible(1);
1207 * Ahh, all good. It wasn't running, and it wasn't
1208 * runnable, which means that it will never become
1209 * running in the future either. We're all done!
1216 * kick_process - kick a running thread to enter/exit the kernel
1217 * @p: the to-be-kicked thread
1219 * Cause a process which is running on another CPU to enter
1220 * kernel-mode, without any delay. (to get signals handled.)
1222 * NOTE: this function doesnt have to take the runqueue lock,
1223 * because all it wants to ensure is that the remote task enters
1224 * the kernel. If the IPI races and the task has been migrated
1225 * to another CPU then no harm is done and the purpose has been
1228 void kick_process(struct task_struct *p)
1234 if ((cpu != smp_processor_id()) && task_curr(p))
1235 smp_send_reschedule(cpu);
1240 * Return a low guess at the load of a migration-source cpu weighted
1241 * according to the scheduling class and "nice" value.
1243 * We want to under-estimate the load of migration sources, to
1244 * balance conservatively.
1246 static unsigned long source_load(int cpu, int type)
1248 struct rq *rq = cpu_rq(cpu);
1249 unsigned long total = weighted_cpuload(cpu);
1254 return min(rq->cpu_load[type-1], total);
1258 * Return a high guess at the load of a migration-target cpu weighted
1259 * according to the scheduling class and "nice" value.
1261 static unsigned long target_load(int cpu, int type)
1263 struct rq *rq = cpu_rq(cpu);
1264 unsigned long total = weighted_cpuload(cpu);
1269 return max(rq->cpu_load[type-1], total);
1273 * Return the average load per task on the cpu's run queue
1275 static inline unsigned long cpu_avg_load_per_task(int cpu)
1277 struct rq *rq = cpu_rq(cpu);
1278 unsigned long total = weighted_cpuload(cpu);
1279 unsigned long n = rq->nr_running;
1281 return n ? total / n : SCHED_LOAD_SCALE;
1285 * find_idlest_group finds and returns the least busy CPU group within the
1288 static struct sched_group *
1289 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1291 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1292 unsigned long min_load = ULONG_MAX, this_load = 0;
1293 int load_idx = sd->forkexec_idx;
1294 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1297 unsigned long load, avg_load;
1301 /* Skip over this group if it has no CPUs allowed */
1302 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1305 local_group = cpu_isset(this_cpu, group->cpumask);
1307 /* Tally up the load of all CPUs in the group */
1310 for_each_cpu_mask(i, group->cpumask) {
1311 /* Bias balancing toward cpus of our domain */
1313 load = source_load(i, load_idx);
1315 load = target_load(i, load_idx);
1320 /* Adjust by relative CPU power of the group */
1321 avg_load = sg_div_cpu_power(group,
1322 avg_load * SCHED_LOAD_SCALE);
1325 this_load = avg_load;
1327 } else if (avg_load < min_load) {
1328 min_load = avg_load;
1331 } while (group = group->next, group != sd->groups);
1333 if (!idlest || 100*this_load < imbalance*min_load)
1339 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1342 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1345 unsigned long load, min_load = ULONG_MAX;
1349 /* Traverse only the allowed CPUs */
1350 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1352 for_each_cpu_mask(i, tmp) {
1353 load = weighted_cpuload(i);
1355 if (load < min_load || (load == min_load && i == this_cpu)) {
1365 * sched_balance_self: balance the current task (running on cpu) in domains
1366 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1369 * Balance, ie. select the least loaded group.
1371 * Returns the target CPU number, or the same CPU if no balancing is needed.
1373 * preempt must be disabled.
1375 static int sched_balance_self(int cpu, int flag)
1377 struct task_struct *t = current;
1378 struct sched_domain *tmp, *sd = NULL;
1380 for_each_domain(cpu, tmp) {
1382 * If power savings logic is enabled for a domain, stop there.
1384 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1386 if (tmp->flags & flag)
1392 struct sched_group *group;
1393 int new_cpu, weight;
1395 if (!(sd->flags & flag)) {
1401 group = find_idlest_group(sd, t, cpu);
1407 new_cpu = find_idlest_cpu(group, t, cpu);
1408 if (new_cpu == -1 || new_cpu == cpu) {
1409 /* Now try balancing at a lower domain level of cpu */
1414 /* Now try balancing at a lower domain level of new_cpu */
1417 weight = cpus_weight(span);
1418 for_each_domain(cpu, tmp) {
1419 if (weight <= cpus_weight(tmp->span))
1421 if (tmp->flags & flag)
1424 /* while loop will break here if sd == NULL */
1430 #endif /* CONFIG_SMP */
1433 * wake_idle() will wake a task on an idle cpu if task->cpu is
1434 * not idle and an idle cpu is available. The span of cpus to
1435 * search starts with cpus closest then further out as needed,
1436 * so we always favor a closer, idle cpu.
1438 * Returns the CPU we should wake onto.
1440 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1441 static int wake_idle(int cpu, struct task_struct *p)
1444 struct sched_domain *sd;
1448 * If it is idle, then it is the best cpu to run this task.
1450 * This cpu is also the best, if it has more than one task already.
1451 * Siblings must be also busy(in most cases) as they didn't already
1452 * pickup the extra load from this cpu and hence we need not check
1453 * sibling runqueue info. This will avoid the checks and cache miss
1454 * penalities associated with that.
1456 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1459 for_each_domain(cpu, sd) {
1460 if (sd->flags & SD_WAKE_IDLE) {
1461 cpus_and(tmp, sd->span, p->cpus_allowed);
1462 for_each_cpu_mask(i, tmp) {
1464 if (i != task_cpu(p)) {
1466 se.nr_wakeups_idle);
1478 static inline int wake_idle(int cpu, struct task_struct *p)
1485 * try_to_wake_up - wake up a thread
1486 * @p: the to-be-woken-up thread
1487 * @state: the mask of task states that can be woken
1488 * @sync: do a synchronous wakeup?
1490 * Put it on the run-queue if it's not already there. The "current"
1491 * thread is always on the run-queue (except when the actual
1492 * re-schedule is in progress), and as such you're allowed to do
1493 * the simpler "current->state = TASK_RUNNING" to mark yourself
1494 * runnable without the overhead of this.
1496 * returns failure only if the task is already active.
1498 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1500 int cpu, orig_cpu, this_cpu, success = 0;
1501 unsigned long flags;
1505 struct sched_domain *sd, *this_sd = NULL;
1506 unsigned long load, this_load;
1510 rq = task_rq_lock(p, &flags);
1511 old_state = p->state;
1512 if (!(old_state & state))
1520 this_cpu = smp_processor_id();
1523 if (unlikely(task_running(rq, p)))
1528 schedstat_inc(rq, ttwu_count);
1529 if (cpu == this_cpu) {
1530 schedstat_inc(rq, ttwu_local);
1534 for_each_domain(this_cpu, sd) {
1535 if (cpu_isset(cpu, sd->span)) {
1536 schedstat_inc(sd, ttwu_wake_remote);
1542 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1546 * Check for affine wakeup and passive balancing possibilities.
1549 int idx = this_sd->wake_idx;
1550 unsigned int imbalance;
1552 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1554 load = source_load(cpu, idx);
1555 this_load = target_load(this_cpu, idx);
1557 new_cpu = this_cpu; /* Wake to this CPU if we can */
1559 if (this_sd->flags & SD_WAKE_AFFINE) {
1560 unsigned long tl = this_load;
1561 unsigned long tl_per_task;
1564 * Attract cache-cold tasks on sync wakeups:
1566 if (sync && !task_hot(p, rq->clock, this_sd))
1569 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1570 tl_per_task = cpu_avg_load_per_task(this_cpu);
1573 * If sync wakeup then subtract the (maximum possible)
1574 * effect of the currently running task from the load
1575 * of the current CPU:
1578 tl -= current->se.load.weight;
1581 tl + target_load(cpu, idx) <= tl_per_task) ||
1582 100*(tl + p->se.load.weight) <= imbalance*load) {
1584 * This domain has SD_WAKE_AFFINE and
1585 * p is cache cold in this domain, and
1586 * there is no bad imbalance.
1588 schedstat_inc(this_sd, ttwu_move_affine);
1589 schedstat_inc(p, se.nr_wakeups_affine);
1595 * Start passive balancing when half the imbalance_pct
1598 if (this_sd->flags & SD_WAKE_BALANCE) {
1599 if (imbalance*this_load <= 100*load) {
1600 schedstat_inc(this_sd, ttwu_move_balance);
1601 schedstat_inc(p, se.nr_wakeups_passive);
1607 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1609 new_cpu = wake_idle(new_cpu, p);
1610 if (new_cpu != cpu) {
1611 set_task_cpu(p, new_cpu);
1612 task_rq_unlock(rq, &flags);
1613 /* might preempt at this point */
1614 rq = task_rq_lock(p, &flags);
1615 old_state = p->state;
1616 if (!(old_state & state))
1621 this_cpu = smp_processor_id();
1626 #endif /* CONFIG_SMP */
1627 schedstat_inc(p, se.nr_wakeups);
1629 schedstat_inc(p, se.nr_wakeups_sync);
1630 if (orig_cpu != cpu)
1631 schedstat_inc(p, se.nr_wakeups_migrate);
1632 if (cpu == this_cpu)
1633 schedstat_inc(p, se.nr_wakeups_local);
1635 schedstat_inc(p, se.nr_wakeups_remote);
1636 update_rq_clock(rq);
1637 activate_task(rq, p, 1);
1638 check_preempt_curr(rq, p);
1642 p->state = TASK_RUNNING;
1644 task_rq_unlock(rq, &flags);
1649 int fastcall wake_up_process(struct task_struct *p)
1651 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1652 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1654 EXPORT_SYMBOL(wake_up_process);
1656 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1658 return try_to_wake_up(p, state, 0);
1662 * Perform scheduler related setup for a newly forked process p.
1663 * p is forked by current.
1665 * __sched_fork() is basic setup used by init_idle() too:
1667 static void __sched_fork(struct task_struct *p)
1669 p->se.exec_start = 0;
1670 p->se.sum_exec_runtime = 0;
1671 p->se.prev_sum_exec_runtime = 0;
1673 #ifdef CONFIG_SCHEDSTATS
1674 p->se.wait_start = 0;
1675 p->se.sum_sleep_runtime = 0;
1676 p->se.sleep_start = 0;
1677 p->se.block_start = 0;
1678 p->se.sleep_max = 0;
1679 p->se.block_max = 0;
1681 p->se.slice_max = 0;
1685 INIT_LIST_HEAD(&p->run_list);
1688 #ifdef CONFIG_PREEMPT_NOTIFIERS
1689 INIT_HLIST_HEAD(&p->preempt_notifiers);
1693 * We mark the process as running here, but have not actually
1694 * inserted it onto the runqueue yet. This guarantees that
1695 * nobody will actually run it, and a signal or other external
1696 * event cannot wake it up and insert it on the runqueue either.
1698 p->state = TASK_RUNNING;
1702 * fork()/clone()-time setup:
1704 void sched_fork(struct task_struct *p, int clone_flags)
1706 int cpu = get_cpu();
1711 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1713 set_task_cpu(p, cpu);
1716 * Make sure we do not leak PI boosting priority to the child:
1718 p->prio = current->normal_prio;
1719 if (!rt_prio(p->prio))
1720 p->sched_class = &fair_sched_class;
1722 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1723 if (likely(sched_info_on()))
1724 memset(&p->sched_info, 0, sizeof(p->sched_info));
1726 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1729 #ifdef CONFIG_PREEMPT
1730 /* Want to start with kernel preemption disabled. */
1731 task_thread_info(p)->preempt_count = 1;
1737 * wake_up_new_task - wake up a newly created task for the first time.
1739 * This function will do some initial scheduler statistics housekeeping
1740 * that must be done for every newly created context, then puts the task
1741 * on the runqueue and wakes it.
1743 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1745 unsigned long flags;
1748 rq = task_rq_lock(p, &flags);
1749 BUG_ON(p->state != TASK_RUNNING);
1750 update_rq_clock(rq);
1752 p->prio = effective_prio(p);
1754 if (!p->sched_class->task_new || !current->se.on_rq) {
1755 activate_task(rq, p, 0);
1758 * Let the scheduling class do new task startup
1759 * management (if any):
1761 p->sched_class->task_new(rq, p);
1762 inc_nr_running(p, rq);
1764 check_preempt_curr(rq, p);
1765 task_rq_unlock(rq, &flags);
1768 #ifdef CONFIG_PREEMPT_NOTIFIERS
1771 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1772 * @notifier: notifier struct to register
1774 void preempt_notifier_register(struct preempt_notifier *notifier)
1776 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1778 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1781 * preempt_notifier_unregister - no longer interested in preemption notifications
1782 * @notifier: notifier struct to unregister
1784 * This is safe to call from within a preemption notifier.
1786 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1788 hlist_del(¬ifier->link);
1790 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1792 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1794 struct preempt_notifier *notifier;
1795 struct hlist_node *node;
1797 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1798 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1802 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1803 struct task_struct *next)
1805 struct preempt_notifier *notifier;
1806 struct hlist_node *node;
1808 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1809 notifier->ops->sched_out(notifier, next);
1814 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1819 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1820 struct task_struct *next)
1827 * prepare_task_switch - prepare to switch tasks
1828 * @rq: the runqueue preparing to switch
1829 * @prev: the current task that is being switched out
1830 * @next: the task we are going to switch to.
1832 * This is called with the rq lock held and interrupts off. It must
1833 * be paired with a subsequent finish_task_switch after the context
1836 * prepare_task_switch sets up locking and calls architecture specific
1840 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1841 struct task_struct *next)
1843 fire_sched_out_preempt_notifiers(prev, next);
1844 prepare_lock_switch(rq, next);
1845 prepare_arch_switch(next);
1849 * finish_task_switch - clean up after a task-switch
1850 * @rq: runqueue associated with task-switch
1851 * @prev: the thread we just switched away from.
1853 * finish_task_switch must be called after the context switch, paired
1854 * with a prepare_task_switch call before the context switch.
1855 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1856 * and do any other architecture-specific cleanup actions.
1858 * Note that we may have delayed dropping an mm in context_switch(). If
1859 * so, we finish that here outside of the runqueue lock. (Doing it
1860 * with the lock held can cause deadlocks; see schedule() for
1863 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1864 __releases(rq->lock)
1866 struct mm_struct *mm = rq->prev_mm;
1872 * A task struct has one reference for the use as "current".
1873 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1874 * schedule one last time. The schedule call will never return, and
1875 * the scheduled task must drop that reference.
1876 * The test for TASK_DEAD must occur while the runqueue locks are
1877 * still held, otherwise prev could be scheduled on another cpu, die
1878 * there before we look at prev->state, and then the reference would
1880 * Manfred Spraul <manfred@colorfullife.com>
1882 prev_state = prev->state;
1883 finish_arch_switch(prev);
1884 finish_lock_switch(rq, prev);
1885 fire_sched_in_preempt_notifiers(current);
1888 if (unlikely(prev_state == TASK_DEAD)) {
1890 * Remove function-return probe instances associated with this
1891 * task and put them back on the free list.
1893 kprobe_flush_task(prev);
1894 put_task_struct(prev);
1899 * schedule_tail - first thing a freshly forked thread must call.
1900 * @prev: the thread we just switched away from.
1902 asmlinkage void schedule_tail(struct task_struct *prev)
1903 __releases(rq->lock)
1905 struct rq *rq = this_rq();
1907 finish_task_switch(rq, prev);
1908 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1909 /* In this case, finish_task_switch does not reenable preemption */
1912 if (current->set_child_tid)
1913 put_user(task_pid_vnr(current), current->set_child_tid);
1917 * context_switch - switch to the new MM and the new
1918 * thread's register state.
1921 context_switch(struct rq *rq, struct task_struct *prev,
1922 struct task_struct *next)
1924 struct mm_struct *mm, *oldmm;
1926 prepare_task_switch(rq, prev, next);
1928 oldmm = prev->active_mm;
1930 * For paravirt, this is coupled with an exit in switch_to to
1931 * combine the page table reload and the switch backend into
1934 arch_enter_lazy_cpu_mode();
1936 if (unlikely(!mm)) {
1937 next->active_mm = oldmm;
1938 atomic_inc(&oldmm->mm_count);
1939 enter_lazy_tlb(oldmm, next);
1941 switch_mm(oldmm, mm, next);
1943 if (unlikely(!prev->mm)) {
1944 prev->active_mm = NULL;
1945 rq->prev_mm = oldmm;
1948 * Since the runqueue lock will be released by the next
1949 * task (which is an invalid locking op but in the case
1950 * of the scheduler it's an obvious special-case), so we
1951 * do an early lockdep release here:
1953 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1954 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1957 /* Here we just switch the register state and the stack. */
1958 switch_to(prev, next, prev);
1962 * this_rq must be evaluated again because prev may have moved
1963 * CPUs since it called schedule(), thus the 'rq' on its stack
1964 * frame will be invalid.
1966 finish_task_switch(this_rq(), prev);
1970 * nr_running, nr_uninterruptible and nr_context_switches:
1972 * externally visible scheduler statistics: current number of runnable
1973 * threads, current number of uninterruptible-sleeping threads, total
1974 * number of context switches performed since bootup.
1976 unsigned long nr_running(void)
1978 unsigned long i, sum = 0;
1980 for_each_online_cpu(i)
1981 sum += cpu_rq(i)->nr_running;
1986 unsigned long nr_uninterruptible(void)
1988 unsigned long i, sum = 0;
1990 for_each_possible_cpu(i)
1991 sum += cpu_rq(i)->nr_uninterruptible;
1994 * Since we read the counters lockless, it might be slightly
1995 * inaccurate. Do not allow it to go below zero though:
1997 if (unlikely((long)sum < 0))
2003 unsigned long long nr_context_switches(void)
2006 unsigned long long sum = 0;
2008 for_each_possible_cpu(i)
2009 sum += cpu_rq(i)->nr_switches;
2014 unsigned long nr_iowait(void)
2016 unsigned long i, sum = 0;
2018 for_each_possible_cpu(i)
2019 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2024 unsigned long nr_active(void)
2026 unsigned long i, running = 0, uninterruptible = 0;
2028 for_each_online_cpu(i) {
2029 running += cpu_rq(i)->nr_running;
2030 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2033 if (unlikely((long)uninterruptible < 0))
2034 uninterruptible = 0;
2036 return running + uninterruptible;
2040 * Update rq->cpu_load[] statistics. This function is usually called every
2041 * scheduler tick (TICK_NSEC).
2043 static void update_cpu_load(struct rq *this_rq)
2045 unsigned long this_load = this_rq->load.weight;
2048 this_rq->nr_load_updates++;
2050 /* Update our load: */
2051 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2052 unsigned long old_load, new_load;
2054 /* scale is effectively 1 << i now, and >> i divides by scale */
2056 old_load = this_rq->cpu_load[i];
2057 new_load = this_load;
2059 * Round up the averaging division if load is increasing. This
2060 * prevents us from getting stuck on 9 if the load is 10, for
2063 if (new_load > old_load)
2064 new_load += scale-1;
2065 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2072 * double_rq_lock - safely lock two runqueues
2074 * Note this does not disable interrupts like task_rq_lock,
2075 * you need to do so manually before calling.
2077 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2078 __acquires(rq1->lock)
2079 __acquires(rq2->lock)
2081 BUG_ON(!irqs_disabled());
2083 spin_lock(&rq1->lock);
2084 __acquire(rq2->lock); /* Fake it out ;) */
2087 spin_lock(&rq1->lock);
2088 spin_lock(&rq2->lock);
2090 spin_lock(&rq2->lock);
2091 spin_lock(&rq1->lock);
2094 update_rq_clock(rq1);
2095 update_rq_clock(rq2);
2099 * double_rq_unlock - safely unlock two runqueues
2101 * Note this does not restore interrupts like task_rq_unlock,
2102 * you need to do so manually after calling.
2104 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2105 __releases(rq1->lock)
2106 __releases(rq2->lock)
2108 spin_unlock(&rq1->lock);
2110 spin_unlock(&rq2->lock);
2112 __release(rq2->lock);
2116 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2118 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2119 __releases(this_rq->lock)
2120 __acquires(busiest->lock)
2121 __acquires(this_rq->lock)
2123 if (unlikely(!irqs_disabled())) {
2124 /* printk() doesn't work good under rq->lock */
2125 spin_unlock(&this_rq->lock);
2128 if (unlikely(!spin_trylock(&busiest->lock))) {
2129 if (busiest < this_rq) {
2130 spin_unlock(&this_rq->lock);
2131 spin_lock(&busiest->lock);
2132 spin_lock(&this_rq->lock);
2134 spin_lock(&busiest->lock);
2139 * If dest_cpu is allowed for this process, migrate the task to it.
2140 * This is accomplished by forcing the cpu_allowed mask to only
2141 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2142 * the cpu_allowed mask is restored.
2144 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2146 struct migration_req req;
2147 unsigned long flags;
2150 rq = task_rq_lock(p, &flags);
2151 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2152 || unlikely(cpu_is_offline(dest_cpu)))
2155 /* force the process onto the specified CPU */
2156 if (migrate_task(p, dest_cpu, &req)) {
2157 /* Need to wait for migration thread (might exit: take ref). */
2158 struct task_struct *mt = rq->migration_thread;
2160 get_task_struct(mt);
2161 task_rq_unlock(rq, &flags);
2162 wake_up_process(mt);
2163 put_task_struct(mt);
2164 wait_for_completion(&req.done);
2169 task_rq_unlock(rq, &flags);
2173 * sched_exec - execve() is a valuable balancing opportunity, because at
2174 * this point the task has the smallest effective memory and cache footprint.
2176 void sched_exec(void)
2178 int new_cpu, this_cpu = get_cpu();
2179 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2181 if (new_cpu != this_cpu)
2182 sched_migrate_task(current, new_cpu);
2186 * pull_task - move a task from a remote runqueue to the local runqueue.
2187 * Both runqueues must be locked.
2189 static void pull_task(struct rq *src_rq, struct task_struct *p,
2190 struct rq *this_rq, int this_cpu)
2192 deactivate_task(src_rq, p, 0);
2193 set_task_cpu(p, this_cpu);
2194 activate_task(this_rq, p, 0);
2196 * Note that idle threads have a prio of MAX_PRIO, for this test
2197 * to be always true for them.
2199 check_preempt_curr(this_rq, p);
2203 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2206 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2207 struct sched_domain *sd, enum cpu_idle_type idle,
2211 * We do not migrate tasks that are:
2212 * 1) running (obviously), or
2213 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2214 * 3) are cache-hot on their current CPU.
2216 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2217 schedstat_inc(p, se.nr_failed_migrations_affine);
2222 if (task_running(rq, p)) {
2223 schedstat_inc(p, se.nr_failed_migrations_running);
2228 * Aggressive migration if:
2229 * 1) task is cache cold, or
2230 * 2) too many balance attempts have failed.
2233 if (!task_hot(p, rq->clock, sd) ||
2234 sd->nr_balance_failed > sd->cache_nice_tries) {
2235 #ifdef CONFIG_SCHEDSTATS
2236 if (task_hot(p, rq->clock, sd)) {
2237 schedstat_inc(sd, lb_hot_gained[idle]);
2238 schedstat_inc(p, se.nr_forced_migrations);
2244 if (task_hot(p, rq->clock, sd)) {
2245 schedstat_inc(p, se.nr_failed_migrations_hot);
2251 static unsigned long
2252 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2253 unsigned long max_load_move, struct sched_domain *sd,
2254 enum cpu_idle_type idle, int *all_pinned,
2255 int *this_best_prio, struct rq_iterator *iterator)
2257 int pulled = 0, pinned = 0, skip_for_load;
2258 struct task_struct *p;
2259 long rem_load_move = max_load_move;
2261 if (max_load_move == 0)
2267 * Start the load-balancing iterator:
2269 p = iterator->start(iterator->arg);
2274 * To help distribute high priority tasks accross CPUs we don't
2275 * skip a task if it will be the highest priority task (i.e. smallest
2276 * prio value) on its new queue regardless of its load weight
2278 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2279 SCHED_LOAD_SCALE_FUZZ;
2280 if ((skip_for_load && p->prio >= *this_best_prio) ||
2281 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2282 p = iterator->next(iterator->arg);
2286 pull_task(busiest, p, this_rq, this_cpu);
2288 rem_load_move -= p->se.load.weight;
2291 * We only want to steal up to the prescribed number of tasks
2292 * and the prescribed amount of weighted load.
2294 if (rem_load_move > 0) {
2295 if (p->prio < *this_best_prio)
2296 *this_best_prio = p->prio;
2297 p = iterator->next(iterator->arg);
2302 * Right now, this is one of only two places pull_task() is called,
2303 * so we can safely collect pull_task() stats here rather than
2304 * inside pull_task().
2306 schedstat_add(sd, lb_gained[idle], pulled);
2309 *all_pinned = pinned;
2311 return max_load_move - rem_load_move;
2315 * move_tasks tries to move up to max_load_move weighted load from busiest to
2316 * this_rq, as part of a balancing operation within domain "sd".
2317 * Returns 1 if successful and 0 otherwise.
2319 * Called with both runqueues locked.
2321 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2322 unsigned long max_load_move,
2323 struct sched_domain *sd, enum cpu_idle_type idle,
2326 const struct sched_class *class = sched_class_highest;
2327 unsigned long total_load_moved = 0;
2328 int this_best_prio = this_rq->curr->prio;
2332 class->load_balance(this_rq, this_cpu, busiest,
2333 max_load_move - total_load_moved,
2334 sd, idle, all_pinned, &this_best_prio);
2335 class = class->next;
2336 } while (class && max_load_move > total_load_moved);
2338 return total_load_moved > 0;
2342 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2343 struct sched_domain *sd, enum cpu_idle_type idle,
2344 struct rq_iterator *iterator)
2346 struct task_struct *p = iterator->start(iterator->arg);
2350 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2351 pull_task(busiest, p, this_rq, this_cpu);
2353 * Right now, this is only the second place pull_task()
2354 * is called, so we can safely collect pull_task()
2355 * stats here rather than inside pull_task().
2357 schedstat_inc(sd, lb_gained[idle]);
2361 p = iterator->next(iterator->arg);
2368 * move_one_task tries to move exactly one task from busiest to this_rq, as
2369 * part of active balancing operations within "domain".
2370 * Returns 1 if successful and 0 otherwise.
2372 * Called with both runqueues locked.
2374 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2375 struct sched_domain *sd, enum cpu_idle_type idle)
2377 const struct sched_class *class;
2379 for (class = sched_class_highest; class; class = class->next)
2380 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2387 * find_busiest_group finds and returns the busiest CPU group within the
2388 * domain. It calculates and returns the amount of weighted load which
2389 * should be moved to restore balance via the imbalance parameter.
2391 static struct sched_group *
2392 find_busiest_group(struct sched_domain *sd, int this_cpu,
2393 unsigned long *imbalance, enum cpu_idle_type idle,
2394 int *sd_idle, cpumask_t *cpus, int *balance)
2396 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2397 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2398 unsigned long max_pull;
2399 unsigned long busiest_load_per_task, busiest_nr_running;
2400 unsigned long this_load_per_task, this_nr_running;
2401 int load_idx, group_imb = 0;
2402 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2403 int power_savings_balance = 1;
2404 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2405 unsigned long min_nr_running = ULONG_MAX;
2406 struct sched_group *group_min = NULL, *group_leader = NULL;
2409 max_load = this_load = total_load = total_pwr = 0;
2410 busiest_load_per_task = busiest_nr_running = 0;
2411 this_load_per_task = this_nr_running = 0;
2412 if (idle == CPU_NOT_IDLE)
2413 load_idx = sd->busy_idx;
2414 else if (idle == CPU_NEWLY_IDLE)
2415 load_idx = sd->newidle_idx;
2417 load_idx = sd->idle_idx;
2420 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2423 int __group_imb = 0;
2424 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2425 unsigned long sum_nr_running, sum_weighted_load;
2427 local_group = cpu_isset(this_cpu, group->cpumask);
2430 balance_cpu = first_cpu(group->cpumask);
2432 /* Tally up the load of all CPUs in the group */
2433 sum_weighted_load = sum_nr_running = avg_load = 0;
2435 min_cpu_load = ~0UL;
2437 for_each_cpu_mask(i, group->cpumask) {
2440 if (!cpu_isset(i, *cpus))
2445 if (*sd_idle && rq->nr_running)
2448 /* Bias balancing toward cpus of our domain */
2450 if (idle_cpu(i) && !first_idle_cpu) {
2455 load = target_load(i, load_idx);
2457 load = source_load(i, load_idx);
2458 if (load > max_cpu_load)
2459 max_cpu_load = load;
2460 if (min_cpu_load > load)
2461 min_cpu_load = load;
2465 sum_nr_running += rq->nr_running;
2466 sum_weighted_load += weighted_cpuload(i);
2470 * First idle cpu or the first cpu(busiest) in this sched group
2471 * is eligible for doing load balancing at this and above
2472 * domains. In the newly idle case, we will allow all the cpu's
2473 * to do the newly idle load balance.
2475 if (idle != CPU_NEWLY_IDLE && local_group &&
2476 balance_cpu != this_cpu && balance) {
2481 total_load += avg_load;
2482 total_pwr += group->__cpu_power;
2484 /* Adjust by relative CPU power of the group */
2485 avg_load = sg_div_cpu_power(group,
2486 avg_load * SCHED_LOAD_SCALE);
2488 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2491 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2494 this_load = avg_load;
2496 this_nr_running = sum_nr_running;
2497 this_load_per_task = sum_weighted_load;
2498 } else if (avg_load > max_load &&
2499 (sum_nr_running > group_capacity || __group_imb)) {
2500 max_load = avg_load;
2502 busiest_nr_running = sum_nr_running;
2503 busiest_load_per_task = sum_weighted_load;
2504 group_imb = __group_imb;
2507 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2509 * Busy processors will not participate in power savings
2512 if (idle == CPU_NOT_IDLE ||
2513 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2517 * If the local group is idle or completely loaded
2518 * no need to do power savings balance at this domain
2520 if (local_group && (this_nr_running >= group_capacity ||
2522 power_savings_balance = 0;
2525 * If a group is already running at full capacity or idle,
2526 * don't include that group in power savings calculations
2528 if (!power_savings_balance || sum_nr_running >= group_capacity
2533 * Calculate the group which has the least non-idle load.
2534 * This is the group from where we need to pick up the load
2537 if ((sum_nr_running < min_nr_running) ||
2538 (sum_nr_running == min_nr_running &&
2539 first_cpu(group->cpumask) <
2540 first_cpu(group_min->cpumask))) {
2542 min_nr_running = sum_nr_running;
2543 min_load_per_task = sum_weighted_load /
2548 * Calculate the group which is almost near its
2549 * capacity but still has some space to pick up some load
2550 * from other group and save more power
2552 if (sum_nr_running <= group_capacity - 1) {
2553 if (sum_nr_running > leader_nr_running ||
2554 (sum_nr_running == leader_nr_running &&
2555 first_cpu(group->cpumask) >
2556 first_cpu(group_leader->cpumask))) {
2557 group_leader = group;
2558 leader_nr_running = sum_nr_running;
2563 group = group->next;
2564 } while (group != sd->groups);
2566 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2569 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2571 if (this_load >= avg_load ||
2572 100*max_load <= sd->imbalance_pct*this_load)
2575 busiest_load_per_task /= busiest_nr_running;
2577 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2580 * We're trying to get all the cpus to the average_load, so we don't
2581 * want to push ourselves above the average load, nor do we wish to
2582 * reduce the max loaded cpu below the average load, as either of these
2583 * actions would just result in more rebalancing later, and ping-pong
2584 * tasks around. Thus we look for the minimum possible imbalance.
2585 * Negative imbalances (*we* are more loaded than anyone else) will
2586 * be counted as no imbalance for these purposes -- we can't fix that
2587 * by pulling tasks to us. Be careful of negative numbers as they'll
2588 * appear as very large values with unsigned longs.
2590 if (max_load <= busiest_load_per_task)
2594 * In the presence of smp nice balancing, certain scenarios can have
2595 * max load less than avg load(as we skip the groups at or below
2596 * its cpu_power, while calculating max_load..)
2598 if (max_load < avg_load) {
2600 goto small_imbalance;
2603 /* Don't want to pull so many tasks that a group would go idle */
2604 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2606 /* How much load to actually move to equalise the imbalance */
2607 *imbalance = min(max_pull * busiest->__cpu_power,
2608 (avg_load - this_load) * this->__cpu_power)
2612 * if *imbalance is less than the average load per runnable task
2613 * there is no gaurantee that any tasks will be moved so we'll have
2614 * a think about bumping its value to force at least one task to be
2617 if (*imbalance < busiest_load_per_task) {
2618 unsigned long tmp, pwr_now, pwr_move;
2622 pwr_move = pwr_now = 0;
2624 if (this_nr_running) {
2625 this_load_per_task /= this_nr_running;
2626 if (busiest_load_per_task > this_load_per_task)
2629 this_load_per_task = SCHED_LOAD_SCALE;
2631 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2632 busiest_load_per_task * imbn) {
2633 *imbalance = busiest_load_per_task;
2638 * OK, we don't have enough imbalance to justify moving tasks,
2639 * however we may be able to increase total CPU power used by
2643 pwr_now += busiest->__cpu_power *
2644 min(busiest_load_per_task, max_load);
2645 pwr_now += this->__cpu_power *
2646 min(this_load_per_task, this_load);
2647 pwr_now /= SCHED_LOAD_SCALE;
2649 /* Amount of load we'd subtract */
2650 tmp = sg_div_cpu_power(busiest,
2651 busiest_load_per_task * SCHED_LOAD_SCALE);
2653 pwr_move += busiest->__cpu_power *
2654 min(busiest_load_per_task, max_load - tmp);
2656 /* Amount of load we'd add */
2657 if (max_load * busiest->__cpu_power <
2658 busiest_load_per_task * SCHED_LOAD_SCALE)
2659 tmp = sg_div_cpu_power(this,
2660 max_load * busiest->__cpu_power);
2662 tmp = sg_div_cpu_power(this,
2663 busiest_load_per_task * SCHED_LOAD_SCALE);
2664 pwr_move += this->__cpu_power *
2665 min(this_load_per_task, this_load + tmp);
2666 pwr_move /= SCHED_LOAD_SCALE;
2668 /* Move if we gain throughput */
2669 if (pwr_move > pwr_now)
2670 *imbalance = busiest_load_per_task;
2676 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2677 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2680 if (this == group_leader && group_leader != group_min) {
2681 *imbalance = min_load_per_task;
2691 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2694 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2695 unsigned long imbalance, cpumask_t *cpus)
2697 struct rq *busiest = NULL, *rq;
2698 unsigned long max_load = 0;
2701 for_each_cpu_mask(i, group->cpumask) {
2704 if (!cpu_isset(i, *cpus))
2708 wl = weighted_cpuload(i);
2710 if (rq->nr_running == 1 && wl > imbalance)
2713 if (wl > max_load) {
2723 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2724 * so long as it is large enough.
2726 #define MAX_PINNED_INTERVAL 512
2729 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2730 * tasks if there is an imbalance.
2732 static int load_balance(int this_cpu, struct rq *this_rq,
2733 struct sched_domain *sd, enum cpu_idle_type idle,
2736 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2737 struct sched_group *group;
2738 unsigned long imbalance;
2740 cpumask_t cpus = CPU_MASK_ALL;
2741 unsigned long flags;
2744 * When power savings policy is enabled for the parent domain, idle
2745 * sibling can pick up load irrespective of busy siblings. In this case,
2746 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2747 * portraying it as CPU_NOT_IDLE.
2749 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2750 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2753 schedstat_inc(sd, lb_count[idle]);
2756 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2763 schedstat_inc(sd, lb_nobusyg[idle]);
2767 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2769 schedstat_inc(sd, lb_nobusyq[idle]);
2773 BUG_ON(busiest == this_rq);
2775 schedstat_add(sd, lb_imbalance[idle], imbalance);
2778 if (busiest->nr_running > 1) {
2780 * Attempt to move tasks. If find_busiest_group has found
2781 * an imbalance but busiest->nr_running <= 1, the group is
2782 * still unbalanced. ld_moved simply stays zero, so it is
2783 * correctly treated as an imbalance.
2785 local_irq_save(flags);
2786 double_rq_lock(this_rq, busiest);
2787 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2788 imbalance, sd, idle, &all_pinned);
2789 double_rq_unlock(this_rq, busiest);
2790 local_irq_restore(flags);
2793 * some other cpu did the load balance for us.
2795 if (ld_moved && this_cpu != smp_processor_id())
2796 resched_cpu(this_cpu);
2798 /* All tasks on this runqueue were pinned by CPU affinity */
2799 if (unlikely(all_pinned)) {
2800 cpu_clear(cpu_of(busiest), cpus);
2801 if (!cpus_empty(cpus))
2808 schedstat_inc(sd, lb_failed[idle]);
2809 sd->nr_balance_failed++;
2811 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2813 spin_lock_irqsave(&busiest->lock, flags);
2815 /* don't kick the migration_thread, if the curr
2816 * task on busiest cpu can't be moved to this_cpu
2818 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2819 spin_unlock_irqrestore(&busiest->lock, flags);
2821 goto out_one_pinned;
2824 if (!busiest->active_balance) {
2825 busiest->active_balance = 1;
2826 busiest->push_cpu = this_cpu;
2829 spin_unlock_irqrestore(&busiest->lock, flags);
2831 wake_up_process(busiest->migration_thread);
2834 * We've kicked active balancing, reset the failure
2837 sd->nr_balance_failed = sd->cache_nice_tries+1;
2840 sd->nr_balance_failed = 0;
2842 if (likely(!active_balance)) {
2843 /* We were unbalanced, so reset the balancing interval */
2844 sd->balance_interval = sd->min_interval;
2847 * If we've begun active balancing, start to back off. This
2848 * case may not be covered by the all_pinned logic if there
2849 * is only 1 task on the busy runqueue (because we don't call
2852 if (sd->balance_interval < sd->max_interval)
2853 sd->balance_interval *= 2;
2856 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2857 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2862 schedstat_inc(sd, lb_balanced[idle]);
2864 sd->nr_balance_failed = 0;
2867 /* tune up the balancing interval */
2868 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2869 (sd->balance_interval < sd->max_interval))
2870 sd->balance_interval *= 2;
2872 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2873 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2879 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2880 * tasks if there is an imbalance.
2882 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2883 * this_rq is locked.
2886 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2888 struct sched_group *group;
2889 struct rq *busiest = NULL;
2890 unsigned long imbalance;
2894 cpumask_t cpus = CPU_MASK_ALL;
2897 * When power savings policy is enabled for the parent domain, idle
2898 * sibling can pick up load irrespective of busy siblings. In this case,
2899 * let the state of idle sibling percolate up as IDLE, instead of
2900 * portraying it as CPU_NOT_IDLE.
2902 if (sd->flags & SD_SHARE_CPUPOWER &&
2903 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2906 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2908 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2909 &sd_idle, &cpus, NULL);
2911 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2915 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2918 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2922 BUG_ON(busiest == this_rq);
2924 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2927 if (busiest->nr_running > 1) {
2928 /* Attempt to move tasks */
2929 double_lock_balance(this_rq, busiest);
2930 /* this_rq->clock is already updated */
2931 update_rq_clock(busiest);
2932 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2933 imbalance, sd, CPU_NEWLY_IDLE,
2935 spin_unlock(&busiest->lock);
2937 if (unlikely(all_pinned)) {
2938 cpu_clear(cpu_of(busiest), cpus);
2939 if (!cpus_empty(cpus))
2945 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2946 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2947 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2950 sd->nr_balance_failed = 0;
2955 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2956 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2957 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2959 sd->nr_balance_failed = 0;
2965 * idle_balance is called by schedule() if this_cpu is about to become
2966 * idle. Attempts to pull tasks from other CPUs.
2968 static void idle_balance(int this_cpu, struct rq *this_rq)
2970 struct sched_domain *sd;
2971 int pulled_task = -1;
2972 unsigned long next_balance = jiffies + HZ;
2974 for_each_domain(this_cpu, sd) {
2975 unsigned long interval;
2977 if (!(sd->flags & SD_LOAD_BALANCE))
2980 if (sd->flags & SD_BALANCE_NEWIDLE)
2981 /* If we've pulled tasks over stop searching: */
2982 pulled_task = load_balance_newidle(this_cpu,
2985 interval = msecs_to_jiffies(sd->balance_interval);
2986 if (time_after(next_balance, sd->last_balance + interval))
2987 next_balance = sd->last_balance + interval;
2991 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2993 * We are going idle. next_balance may be set based on
2994 * a busy processor. So reset next_balance.
2996 this_rq->next_balance = next_balance;
3001 * active_load_balance is run by migration threads. It pushes running tasks
3002 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3003 * running on each physical CPU where possible, and avoids physical /
3004 * logical imbalances.
3006 * Called with busiest_rq locked.
3008 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3010 int target_cpu = busiest_rq->push_cpu;
3011 struct sched_domain *sd;
3012 struct rq *target_rq;
3014 /* Is there any task to move? */
3015 if (busiest_rq->nr_running <= 1)
3018 target_rq = cpu_rq(target_cpu);
3021 * This condition is "impossible", if it occurs
3022 * we need to fix it. Originally reported by
3023 * Bjorn Helgaas on a 128-cpu setup.
3025 BUG_ON(busiest_rq == target_rq);
3027 /* move a task from busiest_rq to target_rq */
3028 double_lock_balance(busiest_rq, target_rq);
3029 update_rq_clock(busiest_rq);
3030 update_rq_clock(target_rq);
3032 /* Search for an sd spanning us and the target CPU. */
3033 for_each_domain(target_cpu, sd) {
3034 if ((sd->flags & SD_LOAD_BALANCE) &&
3035 cpu_isset(busiest_cpu, sd->span))
3040 schedstat_inc(sd, alb_count);
3042 if (move_one_task(target_rq, target_cpu, busiest_rq,
3044 schedstat_inc(sd, alb_pushed);
3046 schedstat_inc(sd, alb_failed);
3048 spin_unlock(&target_rq->lock);
3053 atomic_t load_balancer;
3055 } nohz ____cacheline_aligned = {
3056 .load_balancer = ATOMIC_INIT(-1),
3057 .cpu_mask = CPU_MASK_NONE,
3061 * This routine will try to nominate the ilb (idle load balancing)
3062 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3063 * load balancing on behalf of all those cpus. If all the cpus in the system
3064 * go into this tickless mode, then there will be no ilb owner (as there is
3065 * no need for one) and all the cpus will sleep till the next wakeup event
3068 * For the ilb owner, tick is not stopped. And this tick will be used
3069 * for idle load balancing. ilb owner will still be part of
3072 * While stopping the tick, this cpu will become the ilb owner if there
3073 * is no other owner. And will be the owner till that cpu becomes busy
3074 * or if all cpus in the system stop their ticks at which point
3075 * there is no need for ilb owner.
3077 * When the ilb owner becomes busy, it nominates another owner, during the
3078 * next busy scheduler_tick()
3080 int select_nohz_load_balancer(int stop_tick)
3082 int cpu = smp_processor_id();
3085 cpu_set(cpu, nohz.cpu_mask);
3086 cpu_rq(cpu)->in_nohz_recently = 1;
3089 * If we are going offline and still the leader, give up!
3091 if (cpu_is_offline(cpu) &&
3092 atomic_read(&nohz.load_balancer) == cpu) {
3093 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3098 /* time for ilb owner also to sleep */
3099 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3100 if (atomic_read(&nohz.load_balancer) == cpu)
3101 atomic_set(&nohz.load_balancer, -1);
3105 if (atomic_read(&nohz.load_balancer) == -1) {
3106 /* make me the ilb owner */
3107 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3109 } else if (atomic_read(&nohz.load_balancer) == cpu)
3112 if (!cpu_isset(cpu, nohz.cpu_mask))
3115 cpu_clear(cpu, nohz.cpu_mask);
3117 if (atomic_read(&nohz.load_balancer) == cpu)
3118 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3125 static DEFINE_SPINLOCK(balancing);
3128 * It checks each scheduling domain to see if it is due to be balanced,
3129 * and initiates a balancing operation if so.
3131 * Balancing parameters are set up in arch_init_sched_domains.
3133 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3136 struct rq *rq = cpu_rq(cpu);
3137 unsigned long interval;
3138 struct sched_domain *sd;
3139 /* Earliest time when we have to do rebalance again */
3140 unsigned long next_balance = jiffies + 60*HZ;
3141 int update_next_balance = 0;
3143 for_each_domain(cpu, sd) {
3144 if (!(sd->flags & SD_LOAD_BALANCE))
3147 interval = sd->balance_interval;
3148 if (idle != CPU_IDLE)
3149 interval *= sd->busy_factor;
3151 /* scale ms to jiffies */
3152 interval = msecs_to_jiffies(interval);
3153 if (unlikely(!interval))
3155 if (interval > HZ*NR_CPUS/10)
3156 interval = HZ*NR_CPUS/10;
3159 if (sd->flags & SD_SERIALIZE) {
3160 if (!spin_trylock(&balancing))
3164 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3165 if (load_balance(cpu, rq, sd, idle, &balance)) {
3167 * We've pulled tasks over so either we're no
3168 * longer idle, or one of our SMT siblings is
3171 idle = CPU_NOT_IDLE;
3173 sd->last_balance = jiffies;
3175 if (sd->flags & SD_SERIALIZE)
3176 spin_unlock(&balancing);
3178 if (time_after(next_balance, sd->last_balance + interval)) {
3179 next_balance = sd->last_balance + interval;
3180 update_next_balance = 1;
3184 * Stop the load balance at this level. There is another
3185 * CPU in our sched group which is doing load balancing more
3193 * next_balance will be updated only when there is a need.
3194 * When the cpu is attached to null domain for ex, it will not be
3197 if (likely(update_next_balance))
3198 rq->next_balance = next_balance;
3202 * run_rebalance_domains is triggered when needed from the scheduler tick.
3203 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3204 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3206 static void run_rebalance_domains(struct softirq_action *h)
3208 int this_cpu = smp_processor_id();
3209 struct rq *this_rq = cpu_rq(this_cpu);
3210 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3211 CPU_IDLE : CPU_NOT_IDLE;
3213 rebalance_domains(this_cpu, idle);
3217 * If this cpu is the owner for idle load balancing, then do the
3218 * balancing on behalf of the other idle cpus whose ticks are
3221 if (this_rq->idle_at_tick &&
3222 atomic_read(&nohz.load_balancer) == this_cpu) {
3223 cpumask_t cpus = nohz.cpu_mask;
3227 cpu_clear(this_cpu, cpus);
3228 for_each_cpu_mask(balance_cpu, cpus) {
3230 * If this cpu gets work to do, stop the load balancing
3231 * work being done for other cpus. Next load
3232 * balancing owner will pick it up.
3237 rebalance_domains(balance_cpu, CPU_IDLE);
3239 rq = cpu_rq(balance_cpu);
3240 if (time_after(this_rq->next_balance, rq->next_balance))
3241 this_rq->next_balance = rq->next_balance;
3248 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3250 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3251 * idle load balancing owner or decide to stop the periodic load balancing,
3252 * if the whole system is idle.
3254 static inline void trigger_load_balance(struct rq *rq, int cpu)
3258 * If we were in the nohz mode recently and busy at the current
3259 * scheduler tick, then check if we need to nominate new idle
3262 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3263 rq->in_nohz_recently = 0;
3265 if (atomic_read(&nohz.load_balancer) == cpu) {
3266 cpu_clear(cpu, nohz.cpu_mask);
3267 atomic_set(&nohz.load_balancer, -1);
3270 if (atomic_read(&nohz.load_balancer) == -1) {
3272 * simple selection for now: Nominate the
3273 * first cpu in the nohz list to be the next
3276 * TBD: Traverse the sched domains and nominate
3277 * the nearest cpu in the nohz.cpu_mask.
3279 int ilb = first_cpu(nohz.cpu_mask);
3287 * If this cpu is idle and doing idle load balancing for all the
3288 * cpus with ticks stopped, is it time for that to stop?
3290 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3291 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3297 * If this cpu is idle and the idle load balancing is done by
3298 * someone else, then no need raise the SCHED_SOFTIRQ
3300 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3301 cpu_isset(cpu, nohz.cpu_mask))
3304 if (time_after_eq(jiffies, rq->next_balance))
3305 raise_softirq(SCHED_SOFTIRQ);
3308 #else /* CONFIG_SMP */
3311 * on UP we do not need to balance between CPUs:
3313 static inline void idle_balance(int cpu, struct rq *rq)
3319 DEFINE_PER_CPU(struct kernel_stat, kstat);
3321 EXPORT_PER_CPU_SYMBOL(kstat);
3324 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3325 * that have not yet been banked in case the task is currently running.
3327 unsigned long long task_sched_runtime(struct task_struct *p)
3329 unsigned long flags;
3333 rq = task_rq_lock(p, &flags);
3334 ns = p->se.sum_exec_runtime;
3335 if (rq->curr == p) {
3336 update_rq_clock(rq);
3337 delta_exec = rq->clock - p->se.exec_start;
3338 if ((s64)delta_exec > 0)
3341 task_rq_unlock(rq, &flags);
3347 * Account user cpu time to a process.
3348 * @p: the process that the cpu time gets accounted to
3349 * @cputime: the cpu time spent in user space since the last update
3351 void account_user_time(struct task_struct *p, cputime_t cputime)
3353 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3355 struct rq *rq = this_rq();
3357 p->utime = cputime_add(p->utime, cputime);
3360 cpuacct_charge(p, cputime);
3362 /* Add user time to cpustat. */
3363 tmp = cputime_to_cputime64(cputime);
3364 if (TASK_NICE(p) > 0)
3365 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3367 cpustat->user = cputime64_add(cpustat->user, tmp);
3371 * Account guest cpu time to a process.
3372 * @p: the process that the cpu time gets accounted to
3373 * @cputime: the cpu time spent in virtual machine since the last update
3375 void account_guest_time(struct task_struct *p, cputime_t cputime)
3378 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3380 tmp = cputime_to_cputime64(cputime);
3382 p->utime = cputime_add(p->utime, cputime);
3383 p->gtime = cputime_add(p->gtime, cputime);
3385 cpustat->user = cputime64_add(cpustat->user, tmp);
3386 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3390 * Account scaled user cpu time to a process.
3391 * @p: the process that the cpu time gets accounted to
3392 * @cputime: the cpu time spent in user space since the last update
3394 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3396 p->utimescaled = cputime_add(p->utimescaled, cputime);
3400 * Account system cpu time to a process.
3401 * @p: the process that the cpu time gets accounted to
3402 * @hardirq_offset: the offset to subtract from hardirq_count()
3403 * @cputime: the cpu time spent in kernel space since the last update
3405 void account_system_time(struct task_struct *p, int hardirq_offset,
3408 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3409 struct rq *rq = this_rq();
3412 if (p->flags & PF_VCPU) {
3413 account_guest_time(p, cputime);
3417 p->stime = cputime_add(p->stime, cputime);
3419 /* Add system time to cpustat. */
3420 tmp = cputime_to_cputime64(cputime);
3421 if (hardirq_count() - hardirq_offset)
3422 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3423 else if (softirq_count())
3424 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3425 else if (p != rq->idle) {
3426 cpustat->system = cputime64_add(cpustat->system, tmp);
3427 cpuacct_charge(p, cputime);
3428 } else if (atomic_read(&rq->nr_iowait) > 0)
3429 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3431 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3432 /* Account for system time used */
3433 acct_update_integrals(p);
3437 * Account scaled system cpu time to a process.
3438 * @p: the process that the cpu time gets accounted to
3439 * @hardirq_offset: the offset to subtract from hardirq_count()
3440 * @cputime: the cpu time spent in kernel space since the last update
3442 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3444 p->stimescaled = cputime_add(p->stimescaled, cputime);
3448 * Account for involuntary wait time.
3449 * @p: the process from which the cpu time has been stolen
3450 * @steal: the cpu time spent in involuntary wait
3452 void account_steal_time(struct task_struct *p, cputime_t steal)
3454 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3455 cputime64_t tmp = cputime_to_cputime64(steal);
3456 struct rq *rq = this_rq();
3458 if (p == rq->idle) {
3459 p->stime = cputime_add(p->stime, steal);
3460 if (atomic_read(&rq->nr_iowait) > 0)
3461 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3463 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3465 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3466 cpuacct_charge(p, -tmp);
3471 * This function gets called by the timer code, with HZ frequency.
3472 * We call it with interrupts disabled.
3474 * It also gets called by the fork code, when changing the parent's
3477 void scheduler_tick(void)
3479 int cpu = smp_processor_id();
3480 struct rq *rq = cpu_rq(cpu);
3481 struct task_struct *curr = rq->curr;
3482 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3484 spin_lock(&rq->lock);
3485 __update_rq_clock(rq);
3487 * Let rq->clock advance by at least TICK_NSEC:
3489 if (unlikely(rq->clock < next_tick))
3490 rq->clock = next_tick;
3491 rq->tick_timestamp = rq->clock;
3492 update_cpu_load(rq);
3493 if (curr != rq->idle) /* FIXME: needed? */
3494 curr->sched_class->task_tick(rq, curr);
3495 spin_unlock(&rq->lock);
3498 rq->idle_at_tick = idle_cpu(cpu);
3499 trigger_load_balance(rq, cpu);
3503 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3505 void fastcall add_preempt_count(int val)
3510 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3512 preempt_count() += val;
3514 * Spinlock count overflowing soon?
3516 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3519 EXPORT_SYMBOL(add_preempt_count);
3521 void fastcall sub_preempt_count(int val)
3526 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3529 * Is the spinlock portion underflowing?
3531 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3532 !(preempt_count() & PREEMPT_MASK)))
3535 preempt_count() -= val;
3537 EXPORT_SYMBOL(sub_preempt_count);
3542 * Print scheduling while atomic bug:
3544 static noinline void __schedule_bug(struct task_struct *prev)
3546 struct pt_regs *regs = get_irq_regs();
3548 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3549 prev->comm, prev->pid, preempt_count());
3551 debug_show_held_locks(prev);
3552 if (irqs_disabled())
3553 print_irqtrace_events(prev);
3562 * Various schedule()-time debugging checks and statistics:
3564 static inline void schedule_debug(struct task_struct *prev)
3567 * Test if we are atomic. Since do_exit() needs to call into
3568 * schedule() atomically, we ignore that path for now.
3569 * Otherwise, whine if we are scheduling when we should not be.
3571 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3572 __schedule_bug(prev);
3574 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3576 schedstat_inc(this_rq(), sched_count);
3577 #ifdef CONFIG_SCHEDSTATS
3578 if (unlikely(prev->lock_depth >= 0)) {
3579 schedstat_inc(this_rq(), bkl_count);
3580 schedstat_inc(prev, sched_info.bkl_count);
3586 * Pick up the highest-prio task:
3588 static inline struct task_struct *
3589 pick_next_task(struct rq *rq, struct task_struct *prev)
3591 const struct sched_class *class;
3592 struct task_struct *p;
3595 * Optimization: we know that if all tasks are in
3596 * the fair class we can call that function directly:
3598 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3599 p = fair_sched_class.pick_next_task(rq);
3604 class = sched_class_highest;
3606 p = class->pick_next_task(rq);
3610 * Will never be NULL as the idle class always
3611 * returns a non-NULL p:
3613 class = class->next;
3618 * schedule() is the main scheduler function.
3620 asmlinkage void __sched schedule(void)
3622 struct task_struct *prev, *next;
3629 cpu = smp_processor_id();
3633 switch_count = &prev->nivcsw;
3635 release_kernel_lock(prev);
3636 need_resched_nonpreemptible:
3638 schedule_debug(prev);
3641 * Do the rq-clock update outside the rq lock:
3643 local_irq_disable();
3644 __update_rq_clock(rq);
3645 spin_lock(&rq->lock);
3646 clear_tsk_need_resched(prev);
3648 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3649 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3650 unlikely(signal_pending(prev)))) {
3651 prev->state = TASK_RUNNING;
3653 deactivate_task(rq, prev, 1);
3655 switch_count = &prev->nvcsw;
3658 if (unlikely(!rq->nr_running))
3659 idle_balance(cpu, rq);
3661 prev->sched_class->put_prev_task(rq, prev);
3662 next = pick_next_task(rq, prev);
3664 sched_info_switch(prev, next);
3666 if (likely(prev != next)) {
3671 context_switch(rq, prev, next); /* unlocks the rq */
3673 spin_unlock_irq(&rq->lock);
3675 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3676 cpu = smp_processor_id();
3678 goto need_resched_nonpreemptible;
3680 preempt_enable_no_resched();
3681 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3684 EXPORT_SYMBOL(schedule);
3686 #ifdef CONFIG_PREEMPT
3688 * this is the entry point to schedule() from in-kernel preemption
3689 * off of preempt_enable. Kernel preemptions off return from interrupt
3690 * occur there and call schedule directly.
3692 asmlinkage void __sched preempt_schedule(void)
3694 struct thread_info *ti = current_thread_info();
3695 #ifdef CONFIG_PREEMPT_BKL
3696 struct task_struct *task = current;
3697 int saved_lock_depth;
3700 * If there is a non-zero preempt_count or interrupts are disabled,
3701 * we do not want to preempt the current task. Just return..
3703 if (likely(ti->preempt_count || irqs_disabled()))
3707 add_preempt_count(PREEMPT_ACTIVE);
3710 * We keep the big kernel semaphore locked, but we
3711 * clear ->lock_depth so that schedule() doesnt
3712 * auto-release the semaphore:
3714 #ifdef CONFIG_PREEMPT_BKL
3715 saved_lock_depth = task->lock_depth;
3716 task->lock_depth = -1;
3719 #ifdef CONFIG_PREEMPT_BKL
3720 task->lock_depth = saved_lock_depth;
3722 sub_preempt_count(PREEMPT_ACTIVE);
3725 * Check again in case we missed a preemption opportunity
3726 * between schedule and now.
3729 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3731 EXPORT_SYMBOL(preempt_schedule);
3734 * this is the entry point to schedule() from kernel preemption
3735 * off of irq context.
3736 * Note, that this is called and return with irqs disabled. This will
3737 * protect us against recursive calling from irq.
3739 asmlinkage void __sched preempt_schedule_irq(void)
3741 struct thread_info *ti = current_thread_info();
3742 #ifdef CONFIG_PREEMPT_BKL
3743 struct task_struct *task = current;
3744 int saved_lock_depth;
3746 /* Catch callers which need to be fixed */
3747 BUG_ON(ti->preempt_count || !irqs_disabled());
3750 add_preempt_count(PREEMPT_ACTIVE);
3753 * We keep the big kernel semaphore locked, but we
3754 * clear ->lock_depth so that schedule() doesnt
3755 * auto-release the semaphore:
3757 #ifdef CONFIG_PREEMPT_BKL
3758 saved_lock_depth = task->lock_depth;
3759 task->lock_depth = -1;
3763 local_irq_disable();
3764 #ifdef CONFIG_PREEMPT_BKL
3765 task->lock_depth = saved_lock_depth;
3767 sub_preempt_count(PREEMPT_ACTIVE);
3770 * Check again in case we missed a preemption opportunity
3771 * between schedule and now.
3774 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3777 #endif /* CONFIG_PREEMPT */
3779 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3782 return try_to_wake_up(curr->private, mode, sync);
3784 EXPORT_SYMBOL(default_wake_function);
3787 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3788 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3789 * number) then we wake all the non-exclusive tasks and one exclusive task.
3791 * There are circumstances in which we can try to wake a task which has already
3792 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3793 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3795 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3796 int nr_exclusive, int sync, void *key)
3798 wait_queue_t *curr, *next;
3800 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3801 unsigned flags = curr->flags;
3803 if (curr->func(curr, mode, sync, key) &&
3804 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3810 * __wake_up - wake up threads blocked on a waitqueue.
3812 * @mode: which threads
3813 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3814 * @key: is directly passed to the wakeup function
3816 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3817 int nr_exclusive, void *key)
3819 unsigned long flags;
3821 spin_lock_irqsave(&q->lock, flags);
3822 __wake_up_common(q, mode, nr_exclusive, 0, key);
3823 spin_unlock_irqrestore(&q->lock, flags);
3825 EXPORT_SYMBOL(__wake_up);
3828 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3830 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3832 __wake_up_common(q, mode, 1, 0, NULL);
3836 * __wake_up_sync - wake up threads blocked on a waitqueue.
3838 * @mode: which threads
3839 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3841 * The sync wakeup differs that the waker knows that it will schedule
3842 * away soon, so while the target thread will be woken up, it will not
3843 * be migrated to another CPU - ie. the two threads are 'synchronized'
3844 * with each other. This can prevent needless bouncing between CPUs.
3846 * On UP it can prevent extra preemption.
3849 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3851 unsigned long flags;
3857 if (unlikely(!nr_exclusive))
3860 spin_lock_irqsave(&q->lock, flags);
3861 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3862 spin_unlock_irqrestore(&q->lock, flags);
3864 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3866 void complete(struct completion *x)
3868 unsigned long flags;
3870 spin_lock_irqsave(&x->wait.lock, flags);
3872 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3874 spin_unlock_irqrestore(&x->wait.lock, flags);
3876 EXPORT_SYMBOL(complete);
3878 void complete_all(struct completion *x)
3880 unsigned long flags;
3882 spin_lock_irqsave(&x->wait.lock, flags);
3883 x->done += UINT_MAX/2;
3884 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3886 spin_unlock_irqrestore(&x->wait.lock, flags);
3888 EXPORT_SYMBOL(complete_all);
3890 static inline long __sched
3891 do_wait_for_common(struct completion *x, long timeout, int state)
3894 DECLARE_WAITQUEUE(wait, current);
3896 wait.flags |= WQ_FLAG_EXCLUSIVE;
3897 __add_wait_queue_tail(&x->wait, &wait);
3899 if (state == TASK_INTERRUPTIBLE &&
3900 signal_pending(current)) {
3901 __remove_wait_queue(&x->wait, &wait);
3902 return -ERESTARTSYS;
3904 __set_current_state(state);
3905 spin_unlock_irq(&x->wait.lock);
3906 timeout = schedule_timeout(timeout);
3907 spin_lock_irq(&x->wait.lock);
3909 __remove_wait_queue(&x->wait, &wait);
3913 __remove_wait_queue(&x->wait, &wait);
3920 wait_for_common(struct completion *x, long timeout, int state)
3924 spin_lock_irq(&x->wait.lock);
3925 timeout = do_wait_for_common(x, timeout, state);
3926 spin_unlock_irq(&x->wait.lock);
3930 void __sched wait_for_completion(struct completion *x)
3932 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3934 EXPORT_SYMBOL(wait_for_completion);
3936 unsigned long __sched
3937 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3939 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3941 EXPORT_SYMBOL(wait_for_completion_timeout);
3943 int __sched wait_for_completion_interruptible(struct completion *x)
3945 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3946 if (t == -ERESTARTSYS)
3950 EXPORT_SYMBOL(wait_for_completion_interruptible);
3952 unsigned long __sched
3953 wait_for_completion_interruptible_timeout(struct completion *x,
3954 unsigned long timeout)
3956 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3958 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3961 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3963 unsigned long flags;
3966 init_waitqueue_entry(&wait, current);
3968 __set_current_state(state);
3970 spin_lock_irqsave(&q->lock, flags);
3971 __add_wait_queue(q, &wait);
3972 spin_unlock(&q->lock);
3973 timeout = schedule_timeout(timeout);
3974 spin_lock_irq(&q->lock);
3975 __remove_wait_queue(q, &wait);
3976 spin_unlock_irqrestore(&q->lock, flags);
3981 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3983 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3985 EXPORT_SYMBOL(interruptible_sleep_on);
3988 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3990 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3992 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3994 void __sched sleep_on(wait_queue_head_t *q)
3996 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3998 EXPORT_SYMBOL(sleep_on);
4000 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4002 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4004 EXPORT_SYMBOL(sleep_on_timeout);
4006 #ifdef CONFIG_RT_MUTEXES
4009 * rt_mutex_setprio - set the current priority of a task
4011 * @prio: prio value (kernel-internal form)
4013 * This function changes the 'effective' priority of a task. It does
4014 * not touch ->normal_prio like __setscheduler().
4016 * Used by the rt_mutex code to implement priority inheritance logic.
4018 void rt_mutex_setprio(struct task_struct *p, int prio)
4020 unsigned long flags;
4021 int oldprio, on_rq, running;
4024 BUG_ON(prio < 0 || prio > MAX_PRIO);
4026 rq = task_rq_lock(p, &flags);
4027 update_rq_clock(rq);
4030 on_rq = p->se.on_rq;
4031 running = task_running(rq, p);
4033 dequeue_task(rq, p, 0);
4035 p->sched_class->put_prev_task(rq, p);
4039 p->sched_class = &rt_sched_class;
4041 p->sched_class = &fair_sched_class;
4047 p->sched_class->set_curr_task(rq);
4048 enqueue_task(rq, p, 0);
4050 * Reschedule if we are currently running on this runqueue and
4051 * our priority decreased, or if we are not currently running on
4052 * this runqueue and our priority is higher than the current's
4055 if (p->prio > oldprio)
4056 resched_task(rq->curr);
4058 check_preempt_curr(rq, p);
4061 task_rq_unlock(rq, &flags);
4066 void set_user_nice(struct task_struct *p, long nice)
4068 int old_prio, delta, on_rq;
4069 unsigned long flags;
4072 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4075 * We have to be careful, if called from sys_setpriority(),
4076 * the task might be in the middle of scheduling on another CPU.
4078 rq = task_rq_lock(p, &flags);
4079 update_rq_clock(rq);
4081 * The RT priorities are set via sched_setscheduler(), but we still
4082 * allow the 'normal' nice value to be set - but as expected
4083 * it wont have any effect on scheduling until the task is
4084 * SCHED_FIFO/SCHED_RR:
4086 if (task_has_rt_policy(p)) {
4087 p->static_prio = NICE_TO_PRIO(nice);
4090 on_rq = p->se.on_rq;
4092 dequeue_task(rq, p, 0);
4096 p->static_prio = NICE_TO_PRIO(nice);
4099 p->prio = effective_prio(p);
4100 delta = p->prio - old_prio;
4103 enqueue_task(rq, p, 0);
4106 * If the task increased its priority or is running and
4107 * lowered its priority, then reschedule its CPU:
4109 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4110 resched_task(rq->curr);
4113 task_rq_unlock(rq, &flags);
4115 EXPORT_SYMBOL(set_user_nice);
4118 * can_nice - check if a task can reduce its nice value
4122 int can_nice(const struct task_struct *p, const int nice)
4124 /* convert nice value [19,-20] to rlimit style value [1,40] */
4125 int nice_rlim = 20 - nice;
4127 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4128 capable(CAP_SYS_NICE));
4131 #ifdef __ARCH_WANT_SYS_NICE
4134 * sys_nice - change the priority of the current process.
4135 * @increment: priority increment
4137 * sys_setpriority is a more generic, but much slower function that
4138 * does similar things.
4140 asmlinkage long sys_nice(int increment)
4145 * Setpriority might change our priority at the same moment.
4146 * We don't have to worry. Conceptually one call occurs first
4147 * and we have a single winner.
4149 if (increment < -40)
4154 nice = PRIO_TO_NICE(current->static_prio) + increment;
4160 if (increment < 0 && !can_nice(current, nice))
4163 retval = security_task_setnice(current, nice);
4167 set_user_nice(current, nice);
4174 * task_prio - return the priority value of a given task.
4175 * @p: the task in question.
4177 * This is the priority value as seen by users in /proc.
4178 * RT tasks are offset by -200. Normal tasks are centered
4179 * around 0, value goes from -16 to +15.
4181 int task_prio(const struct task_struct *p)
4183 return p->prio - MAX_RT_PRIO;
4187 * task_nice - return the nice value of a given task.
4188 * @p: the task in question.
4190 int task_nice(const struct task_struct *p)
4192 return TASK_NICE(p);
4194 EXPORT_SYMBOL_GPL(task_nice);
4197 * idle_cpu - is a given cpu idle currently?
4198 * @cpu: the processor in question.
4200 int idle_cpu(int cpu)
4202 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4206 * idle_task - return the idle task for a given cpu.
4207 * @cpu: the processor in question.
4209 struct task_struct *idle_task(int cpu)
4211 return cpu_rq(cpu)->idle;
4215 * find_process_by_pid - find a process with a matching PID value.
4216 * @pid: the pid in question.
4218 static struct task_struct *find_process_by_pid(pid_t pid)
4220 return pid ? find_task_by_vpid(pid) : current;
4223 /* Actually do priority change: must hold rq lock. */
4225 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4227 BUG_ON(p->se.on_rq);
4230 switch (p->policy) {
4234 p->sched_class = &fair_sched_class;
4238 p->sched_class = &rt_sched_class;
4242 p->rt_priority = prio;
4243 p->normal_prio = normal_prio(p);
4244 /* we are holding p->pi_lock already */
4245 p->prio = rt_mutex_getprio(p);
4250 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4251 * @p: the task in question.
4252 * @policy: new policy.
4253 * @param: structure containing the new RT priority.
4255 * NOTE that the task may be already dead.
4257 int sched_setscheduler(struct task_struct *p, int policy,
4258 struct sched_param *param)
4260 int retval, oldprio, oldpolicy = -1, on_rq, running;
4261 unsigned long flags;
4264 /* may grab non-irq protected spin_locks */
4265 BUG_ON(in_interrupt());
4267 /* double check policy once rq lock held */
4269 policy = oldpolicy = p->policy;
4270 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4271 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4272 policy != SCHED_IDLE)
4275 * Valid priorities for SCHED_FIFO and SCHED_RR are
4276 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4277 * SCHED_BATCH and SCHED_IDLE is 0.
4279 if (param->sched_priority < 0 ||
4280 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4281 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4283 if (rt_policy(policy) != (param->sched_priority != 0))
4287 * Allow unprivileged RT tasks to decrease priority:
4289 if (!capable(CAP_SYS_NICE)) {
4290 if (rt_policy(policy)) {
4291 unsigned long rlim_rtprio;
4293 if (!lock_task_sighand(p, &flags))
4295 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4296 unlock_task_sighand(p, &flags);
4298 /* can't set/change the rt policy */
4299 if (policy != p->policy && !rlim_rtprio)
4302 /* can't increase priority */
4303 if (param->sched_priority > p->rt_priority &&
4304 param->sched_priority > rlim_rtprio)
4308 * Like positive nice levels, dont allow tasks to
4309 * move out of SCHED_IDLE either:
4311 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4314 /* can't change other user's priorities */
4315 if ((current->euid != p->euid) &&
4316 (current->euid != p->uid))
4320 retval = security_task_setscheduler(p, policy, param);
4324 * make sure no PI-waiters arrive (or leave) while we are
4325 * changing the priority of the task:
4327 spin_lock_irqsave(&p->pi_lock, flags);
4329 * To be able to change p->policy safely, the apropriate
4330 * runqueue lock must be held.
4332 rq = __task_rq_lock(p);
4333 /* recheck policy now with rq lock held */
4334 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4335 policy = oldpolicy = -1;
4336 __task_rq_unlock(rq);
4337 spin_unlock_irqrestore(&p->pi_lock, flags);
4340 update_rq_clock(rq);
4341 on_rq = p->se.on_rq;
4342 running = task_running(rq, p);
4344 deactivate_task(rq, p, 0);
4346 p->sched_class->put_prev_task(rq, p);
4350 __setscheduler(rq, p, policy, param->sched_priority);
4354 p->sched_class->set_curr_task(rq);
4355 activate_task(rq, p, 0);
4357 * Reschedule if we are currently running on this runqueue and
4358 * our priority decreased, or if we are not currently running on
4359 * this runqueue and our priority is higher than the current's
4362 if (p->prio > oldprio)
4363 resched_task(rq->curr);
4365 check_preempt_curr(rq, p);
4368 __task_rq_unlock(rq);
4369 spin_unlock_irqrestore(&p->pi_lock, flags);
4371 rt_mutex_adjust_pi(p);
4375 EXPORT_SYMBOL_GPL(sched_setscheduler);
4378 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4380 struct sched_param lparam;
4381 struct task_struct *p;
4384 if (!param || pid < 0)
4386 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4391 p = find_process_by_pid(pid);
4393 retval = sched_setscheduler(p, policy, &lparam);
4400 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4401 * @pid: the pid in question.
4402 * @policy: new policy.
4403 * @param: structure containing the new RT priority.
4405 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4406 struct sched_param __user *param)
4408 /* negative values for policy are not valid */
4412 return do_sched_setscheduler(pid, policy, param);
4416 * sys_sched_setparam - set/change the RT priority of a thread
4417 * @pid: the pid in question.
4418 * @param: structure containing the new RT priority.
4420 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4422 return do_sched_setscheduler(pid, -1, param);
4426 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4427 * @pid: the pid in question.
4429 asmlinkage long sys_sched_getscheduler(pid_t pid)
4431 struct task_struct *p;
4438 read_lock(&tasklist_lock);
4439 p = find_process_by_pid(pid);
4441 retval = security_task_getscheduler(p);
4445 read_unlock(&tasklist_lock);
4450 * sys_sched_getscheduler - get the RT priority of a thread
4451 * @pid: the pid in question.
4452 * @param: structure containing the RT priority.
4454 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4456 struct sched_param lp;
4457 struct task_struct *p;
4460 if (!param || pid < 0)
4463 read_lock(&tasklist_lock);
4464 p = find_process_by_pid(pid);
4469 retval = security_task_getscheduler(p);
4473 lp.sched_priority = p->rt_priority;
4474 read_unlock(&tasklist_lock);
4477 * This one might sleep, we cannot do it with a spinlock held ...
4479 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4484 read_unlock(&tasklist_lock);
4488 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4490 cpumask_t cpus_allowed;
4491 struct task_struct *p;
4494 mutex_lock(&sched_hotcpu_mutex);
4495 read_lock(&tasklist_lock);
4497 p = find_process_by_pid(pid);
4499 read_unlock(&tasklist_lock);
4500 mutex_unlock(&sched_hotcpu_mutex);
4505 * It is not safe to call set_cpus_allowed with the
4506 * tasklist_lock held. We will bump the task_struct's
4507 * usage count and then drop tasklist_lock.
4510 read_unlock(&tasklist_lock);
4513 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4514 !capable(CAP_SYS_NICE))
4517 retval = security_task_setscheduler(p, 0, NULL);
4521 cpus_allowed = cpuset_cpus_allowed(p);
4522 cpus_and(new_mask, new_mask, cpus_allowed);
4524 retval = set_cpus_allowed(p, new_mask);
4527 cpus_allowed = cpuset_cpus_allowed(p);
4528 if (!cpus_subset(new_mask, cpus_allowed)) {
4530 * We must have raced with a concurrent cpuset
4531 * update. Just reset the cpus_allowed to the
4532 * cpuset's cpus_allowed
4534 new_mask = cpus_allowed;
4540 mutex_unlock(&sched_hotcpu_mutex);
4544 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4545 cpumask_t *new_mask)
4547 if (len < sizeof(cpumask_t)) {
4548 memset(new_mask, 0, sizeof(cpumask_t));
4549 } else if (len > sizeof(cpumask_t)) {
4550 len = sizeof(cpumask_t);
4552 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4556 * sys_sched_setaffinity - set the cpu affinity of a process
4557 * @pid: pid of the process
4558 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4559 * @user_mask_ptr: user-space pointer to the new cpu mask
4561 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4562 unsigned long __user *user_mask_ptr)
4567 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4571 return sched_setaffinity(pid, new_mask);
4575 * Represents all cpu's present in the system
4576 * In systems capable of hotplug, this map could dynamically grow
4577 * as new cpu's are detected in the system via any platform specific
4578 * method, such as ACPI for e.g.
4581 cpumask_t cpu_present_map __read_mostly;
4582 EXPORT_SYMBOL(cpu_present_map);
4585 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4586 EXPORT_SYMBOL(cpu_online_map);
4588 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4589 EXPORT_SYMBOL(cpu_possible_map);
4592 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4594 struct task_struct *p;
4597 mutex_lock(&sched_hotcpu_mutex);
4598 read_lock(&tasklist_lock);
4601 p = find_process_by_pid(pid);
4605 retval = security_task_getscheduler(p);
4609 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4612 read_unlock(&tasklist_lock);
4613 mutex_unlock(&sched_hotcpu_mutex);
4619 * sys_sched_getaffinity - get the cpu affinity of a process
4620 * @pid: pid of the process
4621 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4622 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4624 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4625 unsigned long __user *user_mask_ptr)
4630 if (len < sizeof(cpumask_t))
4633 ret = sched_getaffinity(pid, &mask);
4637 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4640 return sizeof(cpumask_t);
4644 * sys_sched_yield - yield the current processor to other threads.
4646 * This function yields the current CPU to other tasks. If there are no
4647 * other threads running on this CPU then this function will return.
4649 asmlinkage long sys_sched_yield(void)
4651 struct rq *rq = this_rq_lock();
4653 schedstat_inc(rq, yld_count);
4654 current->sched_class->yield_task(rq);
4657 * Since we are going to call schedule() anyway, there's
4658 * no need to preempt or enable interrupts:
4660 __release(rq->lock);
4661 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4662 _raw_spin_unlock(&rq->lock);
4663 preempt_enable_no_resched();
4670 static void __cond_resched(void)
4672 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4673 __might_sleep(__FILE__, __LINE__);
4676 * The BKS might be reacquired before we have dropped
4677 * PREEMPT_ACTIVE, which could trigger a second
4678 * cond_resched() call.
4681 add_preempt_count(PREEMPT_ACTIVE);
4683 sub_preempt_count(PREEMPT_ACTIVE);
4684 } while (need_resched());
4687 int __sched cond_resched(void)
4689 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4690 system_state == SYSTEM_RUNNING) {
4696 EXPORT_SYMBOL(cond_resched);
4699 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4700 * call schedule, and on return reacquire the lock.
4702 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4703 * operations here to prevent schedule() from being called twice (once via
4704 * spin_unlock(), once by hand).
4706 int cond_resched_lock(spinlock_t *lock)
4710 if (need_lockbreak(lock)) {
4716 if (need_resched() && system_state == SYSTEM_RUNNING) {
4717 spin_release(&lock->dep_map, 1, _THIS_IP_);
4718 _raw_spin_unlock(lock);
4719 preempt_enable_no_resched();
4726 EXPORT_SYMBOL(cond_resched_lock);
4728 int __sched cond_resched_softirq(void)
4730 BUG_ON(!in_softirq());
4732 if (need_resched() && system_state == SYSTEM_RUNNING) {
4740 EXPORT_SYMBOL(cond_resched_softirq);
4743 * yield - yield the current processor to other threads.
4745 * This is a shortcut for kernel-space yielding - it marks the
4746 * thread runnable and calls sys_sched_yield().
4748 void __sched yield(void)
4750 set_current_state(TASK_RUNNING);
4753 EXPORT_SYMBOL(yield);
4756 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4757 * that process accounting knows that this is a task in IO wait state.
4759 * But don't do that if it is a deliberate, throttling IO wait (this task
4760 * has set its backing_dev_info: the queue against which it should throttle)
4762 void __sched io_schedule(void)
4764 struct rq *rq = &__raw_get_cpu_var(runqueues);
4766 delayacct_blkio_start();
4767 atomic_inc(&rq->nr_iowait);
4769 atomic_dec(&rq->nr_iowait);
4770 delayacct_blkio_end();
4772 EXPORT_SYMBOL(io_schedule);
4774 long __sched io_schedule_timeout(long timeout)
4776 struct rq *rq = &__raw_get_cpu_var(runqueues);
4779 delayacct_blkio_start();
4780 atomic_inc(&rq->nr_iowait);
4781 ret = schedule_timeout(timeout);
4782 atomic_dec(&rq->nr_iowait);
4783 delayacct_blkio_end();
4788 * sys_sched_get_priority_max - return maximum RT priority.
4789 * @policy: scheduling class.
4791 * this syscall returns the maximum rt_priority that can be used
4792 * by a given scheduling class.
4794 asmlinkage long sys_sched_get_priority_max(int policy)
4801 ret = MAX_USER_RT_PRIO-1;
4813 * sys_sched_get_priority_min - return minimum RT priority.
4814 * @policy: scheduling class.
4816 * this syscall returns the minimum rt_priority that can be used
4817 * by a given scheduling class.
4819 asmlinkage long sys_sched_get_priority_min(int policy)
4837 * sys_sched_rr_get_interval - return the default timeslice of a process.
4838 * @pid: pid of the process.
4839 * @interval: userspace pointer to the timeslice value.
4841 * this syscall writes the default timeslice value of a given process
4842 * into the user-space timespec buffer. A value of '0' means infinity.
4845 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4847 struct task_struct *p;
4848 unsigned int time_slice;
4856 read_lock(&tasklist_lock);
4857 p = find_process_by_pid(pid);
4861 retval = security_task_getscheduler(p);
4865 if (p->policy == SCHED_FIFO)
4867 else if (p->policy == SCHED_RR)
4868 time_slice = DEF_TIMESLICE;
4870 struct sched_entity *se = &p->se;
4871 unsigned long flags;
4874 rq = task_rq_lock(p, &flags);
4875 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4876 task_rq_unlock(rq, &flags);
4878 read_unlock(&tasklist_lock);
4879 jiffies_to_timespec(time_slice, &t);
4880 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4884 read_unlock(&tasklist_lock);
4888 static const char stat_nam[] = "RSDTtZX";
4890 static void show_task(struct task_struct *p)
4892 unsigned long free = 0;
4895 state = p->state ? __ffs(p->state) + 1 : 0;
4896 printk(KERN_INFO "%-13.13s %c", p->comm,
4897 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4898 #if BITS_PER_LONG == 32
4899 if (state == TASK_RUNNING)
4900 printk(KERN_CONT " running ");
4902 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4904 if (state == TASK_RUNNING)
4905 printk(KERN_CONT " running task ");
4907 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4909 #ifdef CONFIG_DEBUG_STACK_USAGE
4911 unsigned long *n = end_of_stack(p);
4914 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4917 printk(KERN_CONT "%5lu %5d %6d\n", free,
4918 task_pid_nr(p), task_pid_nr(p->parent));
4920 if (state != TASK_RUNNING)
4921 show_stack(p, NULL);
4924 void show_state_filter(unsigned long state_filter)
4926 struct task_struct *g, *p;
4928 #if BITS_PER_LONG == 32
4930 " task PC stack pid father\n");
4933 " task PC stack pid father\n");
4935 read_lock(&tasklist_lock);
4936 do_each_thread(g, p) {
4938 * reset the NMI-timeout, listing all files on a slow
4939 * console might take alot of time:
4941 touch_nmi_watchdog();
4942 if (!state_filter || (p->state & state_filter))
4944 } while_each_thread(g, p);
4946 touch_all_softlockup_watchdogs();
4948 #ifdef CONFIG_SCHED_DEBUG
4949 sysrq_sched_debug_show();
4951 read_unlock(&tasklist_lock);
4953 * Only show locks if all tasks are dumped:
4955 if (state_filter == -1)
4956 debug_show_all_locks();
4959 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4961 idle->sched_class = &idle_sched_class;
4965 * init_idle - set up an idle thread for a given CPU
4966 * @idle: task in question
4967 * @cpu: cpu the idle task belongs to
4969 * NOTE: this function does not set the idle thread's NEED_RESCHED
4970 * flag, to make booting more robust.
4972 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4974 struct rq *rq = cpu_rq(cpu);
4975 unsigned long flags;
4978 idle->se.exec_start = sched_clock();
4980 idle->prio = idle->normal_prio = MAX_PRIO;
4981 idle->cpus_allowed = cpumask_of_cpu(cpu);
4982 __set_task_cpu(idle, cpu);
4984 spin_lock_irqsave(&rq->lock, flags);
4985 rq->curr = rq->idle = idle;
4986 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4989 spin_unlock_irqrestore(&rq->lock, flags);
4991 /* Set the preempt count _outside_ the spinlocks! */
4992 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4993 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4995 task_thread_info(idle)->preempt_count = 0;
4998 * The idle tasks have their own, simple scheduling class:
5000 idle->sched_class = &idle_sched_class;
5004 * In a system that switches off the HZ timer nohz_cpu_mask
5005 * indicates which cpus entered this state. This is used
5006 * in the rcu update to wait only for active cpus. For system
5007 * which do not switch off the HZ timer nohz_cpu_mask should
5008 * always be CPU_MASK_NONE.
5010 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5014 * This is how migration works:
5016 * 1) we queue a struct migration_req structure in the source CPU's
5017 * runqueue and wake up that CPU's migration thread.
5018 * 2) we down() the locked semaphore => thread blocks.
5019 * 3) migration thread wakes up (implicitly it forces the migrated
5020 * thread off the CPU)
5021 * 4) it gets the migration request and checks whether the migrated
5022 * task is still in the wrong runqueue.
5023 * 5) if it's in the wrong runqueue then the migration thread removes
5024 * it and puts it into the right queue.
5025 * 6) migration thread up()s the semaphore.
5026 * 7) we wake up and the migration is done.
5030 * Change a given task's CPU affinity. Migrate the thread to a
5031 * proper CPU and schedule it away if the CPU it's executing on
5032 * is removed from the allowed bitmask.
5034 * NOTE: the caller must have a valid reference to the task, the
5035 * task must not exit() & deallocate itself prematurely. The
5036 * call is not atomic; no spinlocks may be held.
5038 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5040 struct migration_req req;
5041 unsigned long flags;
5045 rq = task_rq_lock(p, &flags);
5046 if (!cpus_intersects(new_mask, cpu_online_map)) {
5051 p->cpus_allowed = new_mask;
5052 /* Can the task run on the task's current CPU? If so, we're done */
5053 if (cpu_isset(task_cpu(p), new_mask))
5056 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5057 /* Need help from migration thread: drop lock and wait. */
5058 task_rq_unlock(rq, &flags);
5059 wake_up_process(rq->migration_thread);
5060 wait_for_completion(&req.done);
5061 tlb_migrate_finish(p->mm);
5065 task_rq_unlock(rq, &flags);
5069 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5072 * Move (not current) task off this cpu, onto dest cpu. We're doing
5073 * this because either it can't run here any more (set_cpus_allowed()
5074 * away from this CPU, or CPU going down), or because we're
5075 * attempting to rebalance this task on exec (sched_exec).
5077 * So we race with normal scheduler movements, but that's OK, as long
5078 * as the task is no longer on this CPU.
5080 * Returns non-zero if task was successfully migrated.
5082 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5084 struct rq *rq_dest, *rq_src;
5087 if (unlikely(cpu_is_offline(dest_cpu)))
5090 rq_src = cpu_rq(src_cpu);
5091 rq_dest = cpu_rq(dest_cpu);
5093 double_rq_lock(rq_src, rq_dest);
5094 /* Already moved. */
5095 if (task_cpu(p) != src_cpu)
5097 /* Affinity changed (again). */
5098 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5101 on_rq = p->se.on_rq;
5103 deactivate_task(rq_src, p, 0);
5105 set_task_cpu(p, dest_cpu);
5107 activate_task(rq_dest, p, 0);
5108 check_preempt_curr(rq_dest, p);
5112 double_rq_unlock(rq_src, rq_dest);
5117 * migration_thread - this is a highprio system thread that performs
5118 * thread migration by bumping thread off CPU then 'pushing' onto
5121 static int migration_thread(void *data)
5123 int cpu = (long)data;
5127 BUG_ON(rq->migration_thread != current);
5129 set_current_state(TASK_INTERRUPTIBLE);
5130 while (!kthread_should_stop()) {
5131 struct migration_req *req;
5132 struct list_head *head;
5134 spin_lock_irq(&rq->lock);
5136 if (cpu_is_offline(cpu)) {
5137 spin_unlock_irq(&rq->lock);
5141 if (rq->active_balance) {
5142 active_load_balance(rq, cpu);
5143 rq->active_balance = 0;
5146 head = &rq->migration_queue;
5148 if (list_empty(head)) {
5149 spin_unlock_irq(&rq->lock);
5151 set_current_state(TASK_INTERRUPTIBLE);
5154 req = list_entry(head->next, struct migration_req, list);
5155 list_del_init(head->next);
5157 spin_unlock(&rq->lock);
5158 __migrate_task(req->task, cpu, req->dest_cpu);
5161 complete(&req->done);
5163 __set_current_state(TASK_RUNNING);
5167 /* Wait for kthread_stop */
5168 set_current_state(TASK_INTERRUPTIBLE);
5169 while (!kthread_should_stop()) {
5171 set_current_state(TASK_INTERRUPTIBLE);
5173 __set_current_state(TASK_RUNNING);
5177 #ifdef CONFIG_HOTPLUG_CPU
5179 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5183 local_irq_disable();
5184 ret = __migrate_task(p, src_cpu, dest_cpu);
5190 * Figure out where task on dead CPU should go, use force if necessary.
5191 * NOTE: interrupts should be disabled by the caller
5193 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5195 unsigned long flags;
5202 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5203 cpus_and(mask, mask, p->cpus_allowed);
5204 dest_cpu = any_online_cpu(mask);
5206 /* On any allowed CPU? */
5207 if (dest_cpu == NR_CPUS)
5208 dest_cpu = any_online_cpu(p->cpus_allowed);
5210 /* No more Mr. Nice Guy. */
5211 if (dest_cpu == NR_CPUS) {
5212 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5214 * Try to stay on the same cpuset, where the
5215 * current cpuset may be a subset of all cpus.
5216 * The cpuset_cpus_allowed_locked() variant of
5217 * cpuset_cpus_allowed() will not block. It must be
5218 * called within calls to cpuset_lock/cpuset_unlock.
5220 rq = task_rq_lock(p, &flags);
5221 p->cpus_allowed = cpus_allowed;
5222 dest_cpu = any_online_cpu(p->cpus_allowed);
5223 task_rq_unlock(rq, &flags);
5226 * Don't tell them about moving exiting tasks or
5227 * kernel threads (both mm NULL), since they never
5230 if (p->mm && printk_ratelimit())
5231 printk(KERN_INFO "process %d (%s) no "
5232 "longer affine to cpu%d\n",
5233 task_pid_nr(p), p->comm, dead_cpu);
5235 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5239 * While a dead CPU has no uninterruptible tasks queued at this point,
5240 * it might still have a nonzero ->nr_uninterruptible counter, because
5241 * for performance reasons the counter is not stricly tracking tasks to
5242 * their home CPUs. So we just add the counter to another CPU's counter,
5243 * to keep the global sum constant after CPU-down:
5245 static void migrate_nr_uninterruptible(struct rq *rq_src)
5247 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5248 unsigned long flags;
5250 local_irq_save(flags);
5251 double_rq_lock(rq_src, rq_dest);
5252 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5253 rq_src->nr_uninterruptible = 0;
5254 double_rq_unlock(rq_src, rq_dest);
5255 local_irq_restore(flags);
5258 /* Run through task list and migrate tasks from the dead cpu. */
5259 static void migrate_live_tasks(int src_cpu)
5261 struct task_struct *p, *t;
5263 read_lock(&tasklist_lock);
5265 do_each_thread(t, p) {
5269 if (task_cpu(p) == src_cpu)
5270 move_task_off_dead_cpu(src_cpu, p);
5271 } while_each_thread(t, p);
5273 read_unlock(&tasklist_lock);
5277 * activate_idle_task - move idle task to the _front_ of runqueue.
5279 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5281 update_rq_clock(rq);
5283 if (p->state == TASK_UNINTERRUPTIBLE)
5284 rq->nr_uninterruptible--;
5286 enqueue_task(rq, p, 0);
5287 inc_nr_running(p, rq);
5291 * Schedules idle task to be the next runnable task on current CPU.
5292 * It does so by boosting its priority to highest possible and adding it to
5293 * the _front_ of the runqueue. Used by CPU offline code.
5295 void sched_idle_next(void)
5297 int this_cpu = smp_processor_id();
5298 struct rq *rq = cpu_rq(this_cpu);
5299 struct task_struct *p = rq->idle;
5300 unsigned long flags;
5302 /* cpu has to be offline */
5303 BUG_ON(cpu_online(this_cpu));
5306 * Strictly not necessary since rest of the CPUs are stopped by now
5307 * and interrupts disabled on the current cpu.
5309 spin_lock_irqsave(&rq->lock, flags);
5311 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5313 /* Add idle task to the _front_ of its priority queue: */
5314 activate_idle_task(p, rq);
5316 spin_unlock_irqrestore(&rq->lock, flags);
5320 * Ensures that the idle task is using init_mm right before its cpu goes
5323 void idle_task_exit(void)
5325 struct mm_struct *mm = current->active_mm;
5327 BUG_ON(cpu_online(smp_processor_id()));
5330 switch_mm(mm, &init_mm, current);
5334 /* called under rq->lock with disabled interrupts */
5335 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5337 struct rq *rq = cpu_rq(dead_cpu);
5339 /* Must be exiting, otherwise would be on tasklist. */
5340 BUG_ON(!p->exit_state);
5342 /* Cannot have done final schedule yet: would have vanished. */
5343 BUG_ON(p->state == TASK_DEAD);
5348 * Drop lock around migration; if someone else moves it,
5349 * that's OK. No task can be added to this CPU, so iteration is
5352 spin_unlock_irq(&rq->lock);
5353 move_task_off_dead_cpu(dead_cpu, p);
5354 spin_lock_irq(&rq->lock);
5359 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5360 static void migrate_dead_tasks(unsigned int dead_cpu)
5362 struct rq *rq = cpu_rq(dead_cpu);
5363 struct task_struct *next;
5366 if (!rq->nr_running)
5368 update_rq_clock(rq);
5369 next = pick_next_task(rq, rq->curr);
5372 migrate_dead(dead_cpu, next);
5376 #endif /* CONFIG_HOTPLUG_CPU */
5378 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5380 static struct ctl_table sd_ctl_dir[] = {
5382 .procname = "sched_domain",
5388 static struct ctl_table sd_ctl_root[] = {
5390 .ctl_name = CTL_KERN,
5391 .procname = "kernel",
5393 .child = sd_ctl_dir,
5398 static struct ctl_table *sd_alloc_ctl_entry(int n)
5400 struct ctl_table *entry =
5401 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5406 static void sd_free_ctl_entry(struct ctl_table **tablep)
5408 struct ctl_table *entry;
5411 * In the intermediate directories, both the child directory and
5412 * procname are dynamically allocated and could fail but the mode
5413 * will always be set. In the lowest directory the names are
5414 * static strings and all have proc handlers.
5416 for (entry = *tablep; entry->mode; entry++) {
5418 sd_free_ctl_entry(&entry->child);
5419 if (entry->proc_handler == NULL)
5420 kfree(entry->procname);
5428 set_table_entry(struct ctl_table *entry,
5429 const char *procname, void *data, int maxlen,
5430 mode_t mode, proc_handler *proc_handler)
5432 entry->procname = procname;
5434 entry->maxlen = maxlen;
5436 entry->proc_handler = proc_handler;
5439 static struct ctl_table *
5440 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5442 struct ctl_table *table = sd_alloc_ctl_entry(12);
5447 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5448 sizeof(long), 0644, proc_doulongvec_minmax);
5449 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5450 sizeof(long), 0644, proc_doulongvec_minmax);
5451 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5452 sizeof(int), 0644, proc_dointvec_minmax);
5453 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5454 sizeof(int), 0644, proc_dointvec_minmax);
5455 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5456 sizeof(int), 0644, proc_dointvec_minmax);
5457 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5458 sizeof(int), 0644, proc_dointvec_minmax);
5459 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5460 sizeof(int), 0644, proc_dointvec_minmax);
5461 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5462 sizeof(int), 0644, proc_dointvec_minmax);
5463 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5464 sizeof(int), 0644, proc_dointvec_minmax);
5465 set_table_entry(&table[9], "cache_nice_tries",
5466 &sd->cache_nice_tries,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[10], "flags", &sd->flags,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 /* &table[11] is terminator */
5475 static ctl_table * sd_alloc_ctl_cpu_table(int cpu)
5477 struct ctl_table *entry, *table;
5478 struct sched_domain *sd;
5479 int domain_num = 0, i;
5482 for_each_domain(cpu, sd)
5484 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5489 for_each_domain(cpu, sd) {
5490 snprintf(buf, 32, "domain%d", i);
5491 entry->procname = kstrdup(buf, GFP_KERNEL);
5493 entry->child = sd_alloc_ctl_domain_table(sd);
5500 static struct ctl_table_header *sd_sysctl_header;
5501 static void register_sched_domain_sysctl(void)
5503 int i, cpu_num = num_online_cpus();
5504 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5507 WARN_ON(sd_ctl_dir[0].child);
5508 sd_ctl_dir[0].child = entry;
5513 for_each_online_cpu(i) {
5514 snprintf(buf, 32, "cpu%d", i);
5515 entry->procname = kstrdup(buf, GFP_KERNEL);
5517 entry->child = sd_alloc_ctl_cpu_table(i);
5521 WARN_ON(sd_sysctl_header);
5522 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5525 /* may be called multiple times per register */
5526 static void unregister_sched_domain_sysctl(void)
5528 if (sd_sysctl_header)
5529 unregister_sysctl_table(sd_sysctl_header);
5530 sd_sysctl_header = NULL;
5531 if (sd_ctl_dir[0].child)
5532 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5535 static void register_sched_domain_sysctl(void)
5538 static void unregister_sched_domain_sysctl(void)
5544 * migration_call - callback that gets triggered when a CPU is added.
5545 * Here we can start up the necessary migration thread for the new CPU.
5547 static int __cpuinit
5548 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5550 struct task_struct *p;
5551 int cpu = (long)hcpu;
5552 unsigned long flags;
5556 case CPU_LOCK_ACQUIRE:
5557 mutex_lock(&sched_hotcpu_mutex);
5560 case CPU_UP_PREPARE:
5561 case CPU_UP_PREPARE_FROZEN:
5562 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5565 kthread_bind(p, cpu);
5566 /* Must be high prio: stop_machine expects to yield to it. */
5567 rq = task_rq_lock(p, &flags);
5568 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5569 task_rq_unlock(rq, &flags);
5570 cpu_rq(cpu)->migration_thread = p;
5574 case CPU_ONLINE_FROZEN:
5575 /* Strictly unnecessary, as first user will wake it. */
5576 wake_up_process(cpu_rq(cpu)->migration_thread);
5579 #ifdef CONFIG_HOTPLUG_CPU
5580 case CPU_UP_CANCELED:
5581 case CPU_UP_CANCELED_FROZEN:
5582 if (!cpu_rq(cpu)->migration_thread)
5584 /* Unbind it from offline cpu so it can run. Fall thru. */
5585 kthread_bind(cpu_rq(cpu)->migration_thread,
5586 any_online_cpu(cpu_online_map));
5587 kthread_stop(cpu_rq(cpu)->migration_thread);
5588 cpu_rq(cpu)->migration_thread = NULL;
5592 case CPU_DEAD_FROZEN:
5593 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5594 migrate_live_tasks(cpu);
5596 kthread_stop(rq->migration_thread);
5597 rq->migration_thread = NULL;
5598 /* Idle task back to normal (off runqueue, low prio) */
5599 spin_lock_irq(&rq->lock);
5600 update_rq_clock(rq);
5601 deactivate_task(rq, rq->idle, 0);
5602 rq->idle->static_prio = MAX_PRIO;
5603 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5604 rq->idle->sched_class = &idle_sched_class;
5605 migrate_dead_tasks(cpu);
5606 spin_unlock_irq(&rq->lock);
5608 migrate_nr_uninterruptible(rq);
5609 BUG_ON(rq->nr_running != 0);
5611 /* No need to migrate the tasks: it was best-effort if
5612 * they didn't take sched_hotcpu_mutex. Just wake up
5613 * the requestors. */
5614 spin_lock_irq(&rq->lock);
5615 while (!list_empty(&rq->migration_queue)) {
5616 struct migration_req *req;
5618 req = list_entry(rq->migration_queue.next,
5619 struct migration_req, list);
5620 list_del_init(&req->list);
5621 complete(&req->done);
5623 spin_unlock_irq(&rq->lock);
5626 case CPU_LOCK_RELEASE:
5627 mutex_unlock(&sched_hotcpu_mutex);
5633 /* Register at highest priority so that task migration (migrate_all_tasks)
5634 * happens before everything else.
5636 static struct notifier_block __cpuinitdata migration_notifier = {
5637 .notifier_call = migration_call,
5641 int __init migration_init(void)
5643 void *cpu = (void *)(long)smp_processor_id();
5646 /* Start one for the boot CPU: */
5647 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5648 BUG_ON(err == NOTIFY_BAD);
5649 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5650 register_cpu_notifier(&migration_notifier);
5658 /* Number of possible processor ids */
5659 int nr_cpu_ids __read_mostly = NR_CPUS;
5660 EXPORT_SYMBOL(nr_cpu_ids);
5662 #ifdef CONFIG_SCHED_DEBUG
5664 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5666 struct sched_group *group = sd->groups;
5667 cpumask_t groupmask;
5670 cpumask_scnprintf(str, NR_CPUS, sd->span);
5671 cpus_clear(groupmask);
5673 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5675 if (!(sd->flags & SD_LOAD_BALANCE)) {
5676 printk("does not load-balance\n");
5678 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5683 printk(KERN_CONT "span %s\n", str);
5685 if (!cpu_isset(cpu, sd->span)) {
5686 printk(KERN_ERR "ERROR: domain->span does not contain "
5689 if (!cpu_isset(cpu, group->cpumask)) {
5690 printk(KERN_ERR "ERROR: domain->groups does not contain"
5694 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5698 printk(KERN_ERR "ERROR: group is NULL\n");
5702 if (!group->__cpu_power) {
5703 printk(KERN_CONT "\n");
5704 printk(KERN_ERR "ERROR: domain->cpu_power not "
5709 if (!cpus_weight(group->cpumask)) {
5710 printk(KERN_CONT "\n");
5711 printk(KERN_ERR "ERROR: empty group\n");
5715 if (cpus_intersects(groupmask, group->cpumask)) {
5716 printk(KERN_CONT "\n");
5717 printk(KERN_ERR "ERROR: repeated CPUs\n");
5721 cpus_or(groupmask, groupmask, group->cpumask);
5723 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5724 printk(KERN_CONT " %s", str);
5726 group = group->next;
5727 } while (group != sd->groups);
5728 printk(KERN_CONT "\n");
5730 if (!cpus_equal(sd->span, groupmask))
5731 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5733 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5734 printk(KERN_ERR "ERROR: parent span is not a superset "
5735 "of domain->span\n");
5739 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5744 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5748 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5751 if (sched_domain_debug_one(sd, cpu, level))
5760 # define sched_domain_debug(sd, cpu) do { } while (0)
5763 static int sd_degenerate(struct sched_domain *sd)
5765 if (cpus_weight(sd->span) == 1)
5768 /* Following flags need at least 2 groups */
5769 if (sd->flags & (SD_LOAD_BALANCE |
5770 SD_BALANCE_NEWIDLE |
5774 SD_SHARE_PKG_RESOURCES)) {
5775 if (sd->groups != sd->groups->next)
5779 /* Following flags don't use groups */
5780 if (sd->flags & (SD_WAKE_IDLE |
5789 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5791 unsigned long cflags = sd->flags, pflags = parent->flags;
5793 if (sd_degenerate(parent))
5796 if (!cpus_equal(sd->span, parent->span))
5799 /* Does parent contain flags not in child? */
5800 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5801 if (cflags & SD_WAKE_AFFINE)
5802 pflags &= ~SD_WAKE_BALANCE;
5803 /* Flags needing groups don't count if only 1 group in parent */
5804 if (parent->groups == parent->groups->next) {
5805 pflags &= ~(SD_LOAD_BALANCE |
5806 SD_BALANCE_NEWIDLE |
5810 SD_SHARE_PKG_RESOURCES);
5812 if (~cflags & pflags)
5819 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5820 * hold the hotplug lock.
5822 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5824 struct rq *rq = cpu_rq(cpu);
5825 struct sched_domain *tmp;
5827 /* Remove the sched domains which do not contribute to scheduling. */
5828 for (tmp = sd; tmp; tmp = tmp->parent) {
5829 struct sched_domain *parent = tmp->parent;
5832 if (sd_parent_degenerate(tmp, parent)) {
5833 tmp->parent = parent->parent;
5835 parent->parent->child = tmp;
5839 if (sd && sd_degenerate(sd)) {
5845 sched_domain_debug(sd, cpu);
5847 rcu_assign_pointer(rq->sd, sd);
5850 /* cpus with isolated domains */
5851 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5853 /* Setup the mask of cpus configured for isolated domains */
5854 static int __init isolated_cpu_setup(char *str)
5856 int ints[NR_CPUS], i;
5858 str = get_options(str, ARRAY_SIZE(ints), ints);
5859 cpus_clear(cpu_isolated_map);
5860 for (i = 1; i <= ints[0]; i++)
5861 if (ints[i] < NR_CPUS)
5862 cpu_set(ints[i], cpu_isolated_map);
5866 __setup("isolcpus=", isolated_cpu_setup);
5869 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5870 * to a function which identifies what group(along with sched group) a CPU
5871 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5872 * (due to the fact that we keep track of groups covered with a cpumask_t).
5874 * init_sched_build_groups will build a circular linked list of the groups
5875 * covered by the given span, and will set each group's ->cpumask correctly,
5876 * and ->cpu_power to 0.
5879 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5880 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5881 struct sched_group **sg))
5883 struct sched_group *first = NULL, *last = NULL;
5884 cpumask_t covered = CPU_MASK_NONE;
5887 for_each_cpu_mask(i, span) {
5888 struct sched_group *sg;
5889 int group = group_fn(i, cpu_map, &sg);
5892 if (cpu_isset(i, covered))
5895 sg->cpumask = CPU_MASK_NONE;
5896 sg->__cpu_power = 0;
5898 for_each_cpu_mask(j, span) {
5899 if (group_fn(j, cpu_map, NULL) != group)
5902 cpu_set(j, covered);
5903 cpu_set(j, sg->cpumask);
5914 #define SD_NODES_PER_DOMAIN 16
5919 * find_next_best_node - find the next node to include in a sched_domain
5920 * @node: node whose sched_domain we're building
5921 * @used_nodes: nodes already in the sched_domain
5923 * Find the next node to include in a given scheduling domain. Simply
5924 * finds the closest node not already in the @used_nodes map.
5926 * Should use nodemask_t.
5928 static int find_next_best_node(int node, unsigned long *used_nodes)
5930 int i, n, val, min_val, best_node = 0;
5934 for (i = 0; i < MAX_NUMNODES; i++) {
5935 /* Start at @node */
5936 n = (node + i) % MAX_NUMNODES;
5938 if (!nr_cpus_node(n))
5941 /* Skip already used nodes */
5942 if (test_bit(n, used_nodes))
5945 /* Simple min distance search */
5946 val = node_distance(node, n);
5948 if (val < min_val) {
5954 set_bit(best_node, used_nodes);
5959 * sched_domain_node_span - get a cpumask for a node's sched_domain
5960 * @node: node whose cpumask we're constructing
5961 * @size: number of nodes to include in this span
5963 * Given a node, construct a good cpumask for its sched_domain to span. It
5964 * should be one that prevents unnecessary balancing, but also spreads tasks
5967 static cpumask_t sched_domain_node_span(int node)
5969 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5970 cpumask_t span, nodemask;
5974 bitmap_zero(used_nodes, MAX_NUMNODES);
5976 nodemask = node_to_cpumask(node);
5977 cpus_or(span, span, nodemask);
5978 set_bit(node, used_nodes);
5980 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5981 int next_node = find_next_best_node(node, used_nodes);
5983 nodemask = node_to_cpumask(next_node);
5984 cpus_or(span, span, nodemask);
5991 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5994 * SMT sched-domains:
5996 #ifdef CONFIG_SCHED_SMT
5997 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5998 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6000 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6001 struct sched_group **sg)
6004 *sg = &per_cpu(sched_group_cpus, cpu);
6010 * multi-core sched-domains:
6012 #ifdef CONFIG_SCHED_MC
6013 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6014 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6017 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6018 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6019 struct sched_group **sg)
6022 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6023 cpus_and(mask, mask, *cpu_map);
6024 group = first_cpu(mask);
6026 *sg = &per_cpu(sched_group_core, group);
6029 #elif defined(CONFIG_SCHED_MC)
6030 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6031 struct sched_group **sg)
6034 *sg = &per_cpu(sched_group_core, cpu);
6039 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6040 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6042 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6043 struct sched_group **sg)
6046 #ifdef CONFIG_SCHED_MC
6047 cpumask_t mask = cpu_coregroup_map(cpu);
6048 cpus_and(mask, mask, *cpu_map);
6049 group = first_cpu(mask);
6050 #elif defined(CONFIG_SCHED_SMT)
6051 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6052 cpus_and(mask, mask, *cpu_map);
6053 group = first_cpu(mask);
6058 *sg = &per_cpu(sched_group_phys, group);
6064 * The init_sched_build_groups can't handle what we want to do with node
6065 * groups, so roll our own. Now each node has its own list of groups which
6066 * gets dynamically allocated.
6068 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6069 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6071 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6072 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6074 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6075 struct sched_group **sg)
6077 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6080 cpus_and(nodemask, nodemask, *cpu_map);
6081 group = first_cpu(nodemask);
6084 *sg = &per_cpu(sched_group_allnodes, group);
6088 static void init_numa_sched_groups_power(struct sched_group *group_head)
6090 struct sched_group *sg = group_head;
6096 for_each_cpu_mask(j, sg->cpumask) {
6097 struct sched_domain *sd;
6099 sd = &per_cpu(phys_domains, j);
6100 if (j != first_cpu(sd->groups->cpumask)) {
6102 * Only add "power" once for each
6108 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6111 } while (sg != group_head);
6116 /* Free memory allocated for various sched_group structures */
6117 static void free_sched_groups(const cpumask_t *cpu_map)
6121 for_each_cpu_mask(cpu, *cpu_map) {
6122 struct sched_group **sched_group_nodes
6123 = sched_group_nodes_bycpu[cpu];
6125 if (!sched_group_nodes)
6128 for (i = 0; i < MAX_NUMNODES; i++) {
6129 cpumask_t nodemask = node_to_cpumask(i);
6130 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6132 cpus_and(nodemask, nodemask, *cpu_map);
6133 if (cpus_empty(nodemask))
6143 if (oldsg != sched_group_nodes[i])
6146 kfree(sched_group_nodes);
6147 sched_group_nodes_bycpu[cpu] = NULL;
6151 static void free_sched_groups(const cpumask_t *cpu_map)
6157 * Initialize sched groups cpu_power.
6159 * cpu_power indicates the capacity of sched group, which is used while
6160 * distributing the load between different sched groups in a sched domain.
6161 * Typically cpu_power for all the groups in a sched domain will be same unless
6162 * there are asymmetries in the topology. If there are asymmetries, group
6163 * having more cpu_power will pickup more load compared to the group having
6166 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6167 * the maximum number of tasks a group can handle in the presence of other idle
6168 * or lightly loaded groups in the same sched domain.
6170 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6172 struct sched_domain *child;
6173 struct sched_group *group;
6175 WARN_ON(!sd || !sd->groups);
6177 if (cpu != first_cpu(sd->groups->cpumask))
6182 sd->groups->__cpu_power = 0;
6185 * For perf policy, if the groups in child domain share resources
6186 * (for example cores sharing some portions of the cache hierarchy
6187 * or SMT), then set this domain groups cpu_power such that each group
6188 * can handle only one task, when there are other idle groups in the
6189 * same sched domain.
6191 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6193 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6194 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6199 * add cpu_power of each child group to this groups cpu_power
6201 group = child->groups;
6203 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6204 group = group->next;
6205 } while (group != child->groups);
6209 * Build sched domains for a given set of cpus and attach the sched domains
6210 * to the individual cpus
6212 static int build_sched_domains(const cpumask_t *cpu_map)
6216 struct sched_group **sched_group_nodes = NULL;
6217 int sd_allnodes = 0;
6220 * Allocate the per-node list of sched groups
6222 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6224 if (!sched_group_nodes) {
6225 printk(KERN_WARNING "Can not alloc sched group node list\n");
6228 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6232 * Set up domains for cpus specified by the cpu_map.
6234 for_each_cpu_mask(i, *cpu_map) {
6235 struct sched_domain *sd = NULL, *p;
6236 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6238 cpus_and(nodemask, nodemask, *cpu_map);
6241 if (cpus_weight(*cpu_map) >
6242 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6243 sd = &per_cpu(allnodes_domains, i);
6244 *sd = SD_ALLNODES_INIT;
6245 sd->span = *cpu_map;
6246 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6252 sd = &per_cpu(node_domains, i);
6254 sd->span = sched_domain_node_span(cpu_to_node(i));
6258 cpus_and(sd->span, sd->span, *cpu_map);
6262 sd = &per_cpu(phys_domains, i);
6264 sd->span = nodemask;
6268 cpu_to_phys_group(i, cpu_map, &sd->groups);
6270 #ifdef CONFIG_SCHED_MC
6272 sd = &per_cpu(core_domains, i);
6274 sd->span = cpu_coregroup_map(i);
6275 cpus_and(sd->span, sd->span, *cpu_map);
6278 cpu_to_core_group(i, cpu_map, &sd->groups);
6281 #ifdef CONFIG_SCHED_SMT
6283 sd = &per_cpu(cpu_domains, i);
6284 *sd = SD_SIBLING_INIT;
6285 sd->span = per_cpu(cpu_sibling_map, i);
6286 cpus_and(sd->span, sd->span, *cpu_map);
6289 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6293 #ifdef CONFIG_SCHED_SMT
6294 /* Set up CPU (sibling) groups */
6295 for_each_cpu_mask(i, *cpu_map) {
6296 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6297 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6298 if (i != first_cpu(this_sibling_map))
6301 init_sched_build_groups(this_sibling_map, cpu_map,
6306 #ifdef CONFIG_SCHED_MC
6307 /* Set up multi-core groups */
6308 for_each_cpu_mask(i, *cpu_map) {
6309 cpumask_t this_core_map = cpu_coregroup_map(i);
6310 cpus_and(this_core_map, this_core_map, *cpu_map);
6311 if (i != first_cpu(this_core_map))
6313 init_sched_build_groups(this_core_map, cpu_map,
6314 &cpu_to_core_group);
6318 /* Set up physical groups */
6319 for (i = 0; i < MAX_NUMNODES; i++) {
6320 cpumask_t nodemask = node_to_cpumask(i);
6322 cpus_and(nodemask, nodemask, *cpu_map);
6323 if (cpus_empty(nodemask))
6326 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6330 /* Set up node groups */
6332 init_sched_build_groups(*cpu_map, cpu_map,
6333 &cpu_to_allnodes_group);
6335 for (i = 0; i < MAX_NUMNODES; i++) {
6336 /* Set up node groups */
6337 struct sched_group *sg, *prev;
6338 cpumask_t nodemask = node_to_cpumask(i);
6339 cpumask_t domainspan;
6340 cpumask_t covered = CPU_MASK_NONE;
6343 cpus_and(nodemask, nodemask, *cpu_map);
6344 if (cpus_empty(nodemask)) {
6345 sched_group_nodes[i] = NULL;
6349 domainspan = sched_domain_node_span(i);
6350 cpus_and(domainspan, domainspan, *cpu_map);
6352 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6354 printk(KERN_WARNING "Can not alloc domain group for "
6358 sched_group_nodes[i] = sg;
6359 for_each_cpu_mask(j, nodemask) {
6360 struct sched_domain *sd;
6362 sd = &per_cpu(node_domains, j);
6365 sg->__cpu_power = 0;
6366 sg->cpumask = nodemask;
6368 cpus_or(covered, covered, nodemask);
6371 for (j = 0; j < MAX_NUMNODES; j++) {
6372 cpumask_t tmp, notcovered;
6373 int n = (i + j) % MAX_NUMNODES;
6375 cpus_complement(notcovered, covered);
6376 cpus_and(tmp, notcovered, *cpu_map);
6377 cpus_and(tmp, tmp, domainspan);
6378 if (cpus_empty(tmp))
6381 nodemask = node_to_cpumask(n);
6382 cpus_and(tmp, tmp, nodemask);
6383 if (cpus_empty(tmp))
6386 sg = kmalloc_node(sizeof(struct sched_group),
6390 "Can not alloc domain group for node %d\n", j);
6393 sg->__cpu_power = 0;
6395 sg->next = prev->next;
6396 cpus_or(covered, covered, tmp);
6403 /* Calculate CPU power for physical packages and nodes */
6404 #ifdef CONFIG_SCHED_SMT
6405 for_each_cpu_mask(i, *cpu_map) {
6406 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6408 init_sched_groups_power(i, sd);
6411 #ifdef CONFIG_SCHED_MC
6412 for_each_cpu_mask(i, *cpu_map) {
6413 struct sched_domain *sd = &per_cpu(core_domains, i);
6415 init_sched_groups_power(i, sd);
6419 for_each_cpu_mask(i, *cpu_map) {
6420 struct sched_domain *sd = &per_cpu(phys_domains, i);
6422 init_sched_groups_power(i, sd);
6426 for (i = 0; i < MAX_NUMNODES; i++)
6427 init_numa_sched_groups_power(sched_group_nodes[i]);
6430 struct sched_group *sg;
6432 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6433 init_numa_sched_groups_power(sg);
6437 /* Attach the domains */
6438 for_each_cpu_mask(i, *cpu_map) {
6439 struct sched_domain *sd;
6440 #ifdef CONFIG_SCHED_SMT
6441 sd = &per_cpu(cpu_domains, i);
6442 #elif defined(CONFIG_SCHED_MC)
6443 sd = &per_cpu(core_domains, i);
6445 sd = &per_cpu(phys_domains, i);
6447 cpu_attach_domain(sd, i);
6454 free_sched_groups(cpu_map);
6459 static cpumask_t *doms_cur; /* current sched domains */
6460 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6463 * Special case: If a kmalloc of a doms_cur partition (array of
6464 * cpumask_t) fails, then fallback to a single sched domain,
6465 * as determined by the single cpumask_t fallback_doms.
6467 static cpumask_t fallback_doms;
6470 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6471 * For now this just excludes isolated cpus, but could be used to
6472 * exclude other special cases in the future.
6474 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6479 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6481 doms_cur = &fallback_doms;
6482 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6483 err = build_sched_domains(doms_cur);
6484 register_sched_domain_sysctl();
6489 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6491 free_sched_groups(cpu_map);
6495 * Detach sched domains from a group of cpus specified in cpu_map
6496 * These cpus will now be attached to the NULL domain
6498 static void detach_destroy_domains(const cpumask_t *cpu_map)
6502 unregister_sched_domain_sysctl();
6504 for_each_cpu_mask(i, *cpu_map)
6505 cpu_attach_domain(NULL, i);
6506 synchronize_sched();
6507 arch_destroy_sched_domains(cpu_map);
6511 * Partition sched domains as specified by the 'ndoms_new'
6512 * cpumasks in the array doms_new[] of cpumasks. This compares
6513 * doms_new[] to the current sched domain partitioning, doms_cur[].
6514 * It destroys each deleted domain and builds each new domain.
6516 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6517 * The masks don't intersect (don't overlap.) We should setup one
6518 * sched domain for each mask. CPUs not in any of the cpumasks will
6519 * not be load balanced. If the same cpumask appears both in the
6520 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6523 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6524 * ownership of it and will kfree it when done with it. If the caller
6525 * failed the kmalloc call, then it can pass in doms_new == NULL,
6526 * and partition_sched_domains() will fallback to the single partition
6529 * Call with hotplug lock held
6531 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6535 /* always unregister in case we don't destroy any domains */
6536 unregister_sched_domain_sysctl();
6538 if (doms_new == NULL) {
6540 doms_new = &fallback_doms;
6541 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6544 /* Destroy deleted domains */
6545 for (i = 0; i < ndoms_cur; i++) {
6546 for (j = 0; j < ndoms_new; j++) {
6547 if (cpus_equal(doms_cur[i], doms_new[j]))
6550 /* no match - a current sched domain not in new doms_new[] */
6551 detach_destroy_domains(doms_cur + i);
6556 /* Build new domains */
6557 for (i = 0; i < ndoms_new; i++) {
6558 for (j = 0; j < ndoms_cur; j++) {
6559 if (cpus_equal(doms_new[i], doms_cur[j]))
6562 /* no match - add a new doms_new */
6563 build_sched_domains(doms_new + i);
6568 /* Remember the new sched domains */
6569 if (doms_cur != &fallback_doms)
6571 doms_cur = doms_new;
6572 ndoms_cur = ndoms_new;
6574 register_sched_domain_sysctl();
6577 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6578 static int arch_reinit_sched_domains(void)
6582 mutex_lock(&sched_hotcpu_mutex);
6583 detach_destroy_domains(&cpu_online_map);
6584 err = arch_init_sched_domains(&cpu_online_map);
6585 mutex_unlock(&sched_hotcpu_mutex);
6590 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6594 if (buf[0] != '0' && buf[0] != '1')
6598 sched_smt_power_savings = (buf[0] == '1');
6600 sched_mc_power_savings = (buf[0] == '1');
6602 ret = arch_reinit_sched_domains();
6604 return ret ? ret : count;
6607 #ifdef CONFIG_SCHED_MC
6608 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6610 return sprintf(page, "%u\n", sched_mc_power_savings);
6612 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6613 const char *buf, size_t count)
6615 return sched_power_savings_store(buf, count, 0);
6617 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6618 sched_mc_power_savings_store);
6621 #ifdef CONFIG_SCHED_SMT
6622 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6624 return sprintf(page, "%u\n", sched_smt_power_savings);
6626 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6627 const char *buf, size_t count)
6629 return sched_power_savings_store(buf, count, 1);
6631 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6632 sched_smt_power_savings_store);
6635 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6639 #ifdef CONFIG_SCHED_SMT
6641 err = sysfs_create_file(&cls->kset.kobj,
6642 &attr_sched_smt_power_savings.attr);
6644 #ifdef CONFIG_SCHED_MC
6645 if (!err && mc_capable())
6646 err = sysfs_create_file(&cls->kset.kobj,
6647 &attr_sched_mc_power_savings.attr);
6654 * Force a reinitialization of the sched domains hierarchy. The domains
6655 * and groups cannot be updated in place without racing with the balancing
6656 * code, so we temporarily attach all running cpus to the NULL domain
6657 * which will prevent rebalancing while the sched domains are recalculated.
6659 static int update_sched_domains(struct notifier_block *nfb,
6660 unsigned long action, void *hcpu)
6663 case CPU_UP_PREPARE:
6664 case CPU_UP_PREPARE_FROZEN:
6665 case CPU_DOWN_PREPARE:
6666 case CPU_DOWN_PREPARE_FROZEN:
6667 detach_destroy_domains(&cpu_online_map);
6670 case CPU_UP_CANCELED:
6671 case CPU_UP_CANCELED_FROZEN:
6672 case CPU_DOWN_FAILED:
6673 case CPU_DOWN_FAILED_FROZEN:
6675 case CPU_ONLINE_FROZEN:
6677 case CPU_DEAD_FROZEN:
6679 * Fall through and re-initialise the domains.
6686 /* The hotplug lock is already held by cpu_up/cpu_down */
6687 arch_init_sched_domains(&cpu_online_map);
6692 void __init sched_init_smp(void)
6694 cpumask_t non_isolated_cpus;
6696 mutex_lock(&sched_hotcpu_mutex);
6697 arch_init_sched_domains(&cpu_online_map);
6698 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6699 if (cpus_empty(non_isolated_cpus))
6700 cpu_set(smp_processor_id(), non_isolated_cpus);
6701 mutex_unlock(&sched_hotcpu_mutex);
6702 /* XXX: Theoretical race here - CPU may be hotplugged now */
6703 hotcpu_notifier(update_sched_domains, 0);
6705 /* Move init over to a non-isolated CPU */
6706 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6710 void __init sched_init_smp(void)
6713 #endif /* CONFIG_SMP */
6715 int in_sched_functions(unsigned long addr)
6717 /* Linker adds these: start and end of __sched functions */
6718 extern char __sched_text_start[], __sched_text_end[];
6720 return in_lock_functions(addr) ||
6721 (addr >= (unsigned long)__sched_text_start
6722 && addr < (unsigned long)__sched_text_end);
6725 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6727 cfs_rq->tasks_timeline = RB_ROOT;
6728 #ifdef CONFIG_FAIR_GROUP_SCHED
6731 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6734 void __init sched_init(void)
6736 int highest_cpu = 0;
6739 for_each_possible_cpu(i) {
6740 struct rt_prio_array *array;
6744 spin_lock_init(&rq->lock);
6745 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6748 init_cfs_rq(&rq->cfs, rq);
6749 #ifdef CONFIG_FAIR_GROUP_SCHED
6750 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6752 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6753 struct sched_entity *se =
6754 &per_cpu(init_sched_entity, i);
6756 init_cfs_rq_p[i] = cfs_rq;
6757 init_cfs_rq(cfs_rq, rq);
6758 cfs_rq->tg = &init_task_group;
6759 list_add(&cfs_rq->leaf_cfs_rq_list,
6760 &rq->leaf_cfs_rq_list);
6762 init_sched_entity_p[i] = se;
6763 se->cfs_rq = &rq->cfs;
6765 se->load.weight = init_task_group_load;
6766 se->load.inv_weight =
6767 div64_64(1ULL<<32, init_task_group_load);
6770 init_task_group.shares = init_task_group_load;
6771 spin_lock_init(&init_task_group.lock);
6774 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6775 rq->cpu_load[j] = 0;
6778 rq->active_balance = 0;
6779 rq->next_balance = jiffies;
6782 rq->migration_thread = NULL;
6783 INIT_LIST_HEAD(&rq->migration_queue);
6785 atomic_set(&rq->nr_iowait, 0);
6787 array = &rq->rt.active;
6788 for (j = 0; j < MAX_RT_PRIO; j++) {
6789 INIT_LIST_HEAD(array->queue + j);
6790 __clear_bit(j, array->bitmap);
6793 /* delimiter for bitsearch: */
6794 __set_bit(MAX_RT_PRIO, array->bitmap);
6797 set_load_weight(&init_task);
6799 #ifdef CONFIG_PREEMPT_NOTIFIERS
6800 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6804 nr_cpu_ids = highest_cpu + 1;
6805 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6808 #ifdef CONFIG_RT_MUTEXES
6809 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6813 * The boot idle thread does lazy MMU switching as well:
6815 atomic_inc(&init_mm.mm_count);
6816 enter_lazy_tlb(&init_mm, current);
6819 * Make us the idle thread. Technically, schedule() should not be
6820 * called from this thread, however somewhere below it might be,
6821 * but because we are the idle thread, we just pick up running again
6822 * when this runqueue becomes "idle".
6824 init_idle(current, smp_processor_id());
6826 * During early bootup we pretend to be a normal task:
6828 current->sched_class = &fair_sched_class;
6831 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6832 void __might_sleep(char *file, int line)
6835 static unsigned long prev_jiffy; /* ratelimiting */
6837 if ((in_atomic() || irqs_disabled()) &&
6838 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6839 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6841 prev_jiffy = jiffies;
6842 printk(KERN_ERR "BUG: sleeping function called from invalid"
6843 " context at %s:%d\n", file, line);
6844 printk("in_atomic():%d, irqs_disabled():%d\n",
6845 in_atomic(), irqs_disabled());
6846 debug_show_held_locks(current);
6847 if (irqs_disabled())
6848 print_irqtrace_events(current);
6853 EXPORT_SYMBOL(__might_sleep);
6856 #ifdef CONFIG_MAGIC_SYSRQ
6857 static void normalize_task(struct rq *rq, struct task_struct *p)
6860 update_rq_clock(rq);
6861 on_rq = p->se.on_rq;
6863 deactivate_task(rq, p, 0);
6864 __setscheduler(rq, p, SCHED_NORMAL, 0);
6866 activate_task(rq, p, 0);
6867 resched_task(rq->curr);
6871 void normalize_rt_tasks(void)
6873 struct task_struct *g, *p;
6874 unsigned long flags;
6877 read_lock_irq(&tasklist_lock);
6878 do_each_thread(g, p) {
6880 * Only normalize user tasks:
6885 p->se.exec_start = 0;
6886 #ifdef CONFIG_SCHEDSTATS
6887 p->se.wait_start = 0;
6888 p->se.sleep_start = 0;
6889 p->se.block_start = 0;
6891 task_rq(p)->clock = 0;
6895 * Renice negative nice level userspace
6898 if (TASK_NICE(p) < 0 && p->mm)
6899 set_user_nice(p, 0);
6903 spin_lock_irqsave(&p->pi_lock, flags);
6904 rq = __task_rq_lock(p);
6906 normalize_task(rq, p);
6908 __task_rq_unlock(rq);
6909 spin_unlock_irqrestore(&p->pi_lock, flags);
6910 } while_each_thread(g, p);
6912 read_unlock_irq(&tasklist_lock);
6915 #endif /* CONFIG_MAGIC_SYSRQ */
6919 * These functions are only useful for the IA64 MCA handling.
6921 * They can only be called when the whole system has been
6922 * stopped - every CPU needs to be quiescent, and no scheduling
6923 * activity can take place. Using them for anything else would
6924 * be a serious bug, and as a result, they aren't even visible
6925 * under any other configuration.
6929 * curr_task - return the current task for a given cpu.
6930 * @cpu: the processor in question.
6932 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6934 struct task_struct *curr_task(int cpu)
6936 return cpu_curr(cpu);
6940 * set_curr_task - set the current task for a given cpu.
6941 * @cpu: the processor in question.
6942 * @p: the task pointer to set.
6944 * Description: This function must only be used when non-maskable interrupts
6945 * are serviced on a separate stack. It allows the architecture to switch the
6946 * notion of the current task on a cpu in a non-blocking manner. This function
6947 * must be called with all CPU's synchronized, and interrupts disabled, the
6948 * and caller must save the original value of the current task (see
6949 * curr_task() above) and restore that value before reenabling interrupts and
6950 * re-starting the system.
6952 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6954 void set_curr_task(int cpu, struct task_struct *p)
6961 #ifdef CONFIG_FAIR_GROUP_SCHED
6963 /* allocate runqueue etc for a new task group */
6964 struct task_group *sched_create_group(void)
6966 struct task_group *tg;
6967 struct cfs_rq *cfs_rq;
6968 struct sched_entity *se;
6972 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6974 return ERR_PTR(-ENOMEM);
6976 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6979 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6983 for_each_possible_cpu(i) {
6986 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6991 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6996 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6997 memset(se, 0, sizeof(struct sched_entity));
6999 tg->cfs_rq[i] = cfs_rq;
7000 init_cfs_rq(cfs_rq, rq);
7004 se->cfs_rq = &rq->cfs;
7006 se->load.weight = NICE_0_LOAD;
7007 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7011 for_each_possible_cpu(i) {
7013 cfs_rq = tg->cfs_rq[i];
7014 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7017 tg->shares = NICE_0_LOAD;
7018 spin_lock_init(&tg->lock);
7023 for_each_possible_cpu(i) {
7025 kfree(tg->cfs_rq[i]);
7033 return ERR_PTR(-ENOMEM);
7036 /* rcu callback to free various structures associated with a task group */
7037 static void free_sched_group(struct rcu_head *rhp)
7039 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
7040 struct task_group *tg = cfs_rq->tg;
7041 struct sched_entity *se;
7044 /* now it should be safe to free those cfs_rqs */
7045 for_each_possible_cpu(i) {
7046 cfs_rq = tg->cfs_rq[i];
7058 /* Destroy runqueue etc associated with a task group */
7059 void sched_destroy_group(struct task_group *tg)
7061 struct cfs_rq *cfs_rq;
7064 for_each_possible_cpu(i) {
7065 cfs_rq = tg->cfs_rq[i];
7066 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7069 cfs_rq = tg->cfs_rq[0];
7071 /* wait for possible concurrent references to cfs_rqs complete */
7072 call_rcu(&cfs_rq->rcu, free_sched_group);
7075 /* change task's runqueue when it moves between groups.
7076 * The caller of this function should have put the task in its new group
7077 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7078 * reflect its new group.
7080 void sched_move_task(struct task_struct *tsk)
7083 unsigned long flags;
7086 rq = task_rq_lock(tsk, &flags);
7088 if (tsk->sched_class != &fair_sched_class)
7091 update_rq_clock(rq);
7093 running = task_running(rq, tsk);
7094 on_rq = tsk->se.on_rq;
7097 dequeue_task(rq, tsk, 0);
7098 if (unlikely(running))
7099 tsk->sched_class->put_prev_task(rq, tsk);
7102 set_task_cfs_rq(tsk);
7105 if (unlikely(running))
7106 tsk->sched_class->set_curr_task(rq);
7107 enqueue_task(rq, tsk, 0);
7111 task_rq_unlock(rq, &flags);
7114 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7116 struct cfs_rq *cfs_rq = se->cfs_rq;
7117 struct rq *rq = cfs_rq->rq;
7120 spin_lock_irq(&rq->lock);
7124 dequeue_entity(cfs_rq, se, 0);
7126 se->load.weight = shares;
7127 se->load.inv_weight = div64_64((1ULL<<32), shares);
7130 enqueue_entity(cfs_rq, se, 0);
7132 spin_unlock_irq(&rq->lock);
7135 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7139 spin_lock(&tg->lock);
7140 if (tg->shares == shares)
7143 tg->shares = shares;
7144 for_each_possible_cpu(i)
7145 set_se_shares(tg->se[i], shares);
7148 spin_unlock(&tg->lock);
7152 unsigned long sched_group_shares(struct task_group *tg)
7157 #endif /* CONFIG_FAIR_GROUP_SCHED */
7159 #ifdef CONFIG_FAIR_CGROUP_SCHED
7161 /* return corresponding task_group object of a cgroup */
7162 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7164 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7165 struct task_group, css);
7168 static struct cgroup_subsys_state *
7169 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7171 struct task_group *tg;
7173 if (!cgrp->parent) {
7174 /* This is early initialization for the top cgroup */
7175 init_task_group.css.cgroup = cgrp;
7176 return &init_task_group.css;
7179 /* we support only 1-level deep hierarchical scheduler atm */
7180 if (cgrp->parent->parent)
7181 return ERR_PTR(-EINVAL);
7183 tg = sched_create_group();
7185 return ERR_PTR(-ENOMEM);
7187 /* Bind the cgroup to task_group object we just created */
7188 tg->css.cgroup = cgrp;
7193 static void cpu_cgroup_destroy(struct cgroup_subsys *ss,
7194 struct cgroup *cgrp)
7196 struct task_group *tg = cgroup_tg(cgrp);
7198 sched_destroy_group(tg);
7201 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss,
7202 struct cgroup *cgrp, struct task_struct *tsk)
7204 /* We don't support RT-tasks being in separate groups */
7205 if (tsk->sched_class != &fair_sched_class)
7212 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7213 struct cgroup *old_cont, struct task_struct *tsk)
7215 sched_move_task(tsk);
7218 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7221 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7224 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7226 struct task_group *tg = cgroup_tg(cgrp);
7228 return (u64) tg->shares;
7231 static struct cftype cpu_shares = {
7233 .read_uint = cpu_shares_read_uint,
7234 .write_uint = cpu_shares_write_uint,
7237 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7239 return cgroup_add_file(cont, ss, &cpu_shares);
7242 struct cgroup_subsys cpu_cgroup_subsys = {
7244 .create = cpu_cgroup_create,
7245 .destroy = cpu_cgroup_destroy,
7246 .can_attach = cpu_cgroup_can_attach,
7247 .attach = cpu_cgroup_attach,
7248 .populate = cpu_cgroup_populate,
7249 .subsys_id = cpu_cgroup_subsys_id,
7253 #endif /* CONFIG_FAIR_CGROUP_SCHED */