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
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
71 #include <asm/irq_regs.h>
74 * Scheduler clock - returns current time in nanosec units.
75 * This is default implementation.
76 * Architectures and sub-architectures can override this.
78 unsigned long long __attribute__((weak)) sched_clock(void)
80 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
124 return reciprocal_divide(load, sg->reciprocal_cpu_power);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
133 sg->__cpu_power += val;
134 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
138 static inline int rt_policy(int policy)
140 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
145 static inline int task_has_rt_policy(struct task_struct *p)
147 return rt_policy(p->policy);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array {
154 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
155 struct list_head queue[MAX_RT_PRIO];
158 #ifdef CONFIG_FAIR_GROUP_SCHED
160 #include <linux/cgroup.h>
164 static LIST_HEAD(task_groups);
166 /* task group related information */
168 #ifdef CONFIG_FAIR_CGROUP_SCHED
169 struct cgroup_subsys_state css;
171 /* schedulable entities of this group on each cpu */
172 struct sched_entity **se;
173 /* runqueue "owned" by this group on each cpu */
174 struct cfs_rq **cfs_rq;
176 struct sched_rt_entity **rt_se;
177 struct rt_rq **rt_rq;
179 unsigned int rt_ratio;
182 * shares assigned to a task group governs how much of cpu bandwidth
183 * is allocated to the group. The more shares a group has, the more is
184 * the cpu bandwidth allocated to it.
186 * For ex, lets say that there are three task groups, A, B and C which
187 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
188 * cpu bandwidth allocated by the scheduler to task groups A, B and C
191 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
192 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
193 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
195 * The weight assigned to a task group's schedulable entities on every
196 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
197 * group's shares. For ex: lets say that task group A has been
198 * assigned shares of 1000 and there are two CPUs in a system. Then,
200 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
202 * Note: It's not necessary that each of a task's group schedulable
203 * entity have the same weight on all CPUs. If the group
204 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
205 * better distribution of weight could be:
207 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
208 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
210 * rebalance_shares() is responsible for distributing the shares of a
211 * task groups like this among the group's schedulable entities across
215 unsigned long shares;
218 struct list_head list;
221 /* Default task group's sched entity on each cpu */
222 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
223 /* Default task group's cfs_rq on each cpu */
224 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
226 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
227 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
229 static struct sched_entity *init_sched_entity_p[NR_CPUS];
230 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
232 static struct sched_rt_entity *init_sched_rt_entity_p[NR_CPUS];
233 static struct rt_rq *init_rt_rq_p[NR_CPUS];
235 /* task_group_lock serializes add/remove of task groups and also changes to
236 * a task group's cpu shares.
238 static DEFINE_SPINLOCK(task_group_lock);
240 /* doms_cur_mutex serializes access to doms_cur[] array */
241 static DEFINE_MUTEX(doms_cur_mutex);
244 /* kernel thread that runs rebalance_shares() periodically */
245 static struct task_struct *lb_monitor_task;
246 static int load_balance_monitor(void *unused);
249 static void set_se_shares(struct sched_entity *se, unsigned long shares);
251 /* Default task group.
252 * Every task in system belong to this group at bootup.
254 struct task_group init_task_group = {
255 .se = init_sched_entity_p,
256 .cfs_rq = init_cfs_rq_p,
258 .rt_se = init_sched_rt_entity_p,
259 .rt_rq = init_rt_rq_p,
262 #ifdef CONFIG_FAIR_USER_SCHED
263 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
265 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
268 #define MIN_GROUP_SHARES 2
270 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
272 /* return group to which a task belongs */
273 static inline struct task_group *task_group(struct task_struct *p)
275 struct task_group *tg;
277 #ifdef CONFIG_FAIR_USER_SCHED
279 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
280 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
281 struct task_group, css);
283 tg = &init_task_group;
288 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
289 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
291 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
292 p->se.parent = task_group(p)->se[cpu];
294 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
295 p->rt.parent = task_group(p)->rt_se[cpu];
298 static inline void lock_doms_cur(void)
300 mutex_lock(&doms_cur_mutex);
303 static inline void unlock_doms_cur(void)
305 mutex_unlock(&doms_cur_mutex);
310 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
311 static inline void lock_doms_cur(void) { }
312 static inline void unlock_doms_cur(void) { }
314 #endif /* CONFIG_FAIR_GROUP_SCHED */
316 /* CFS-related fields in a runqueue */
318 struct load_weight load;
319 unsigned long nr_running;
324 struct rb_root tasks_timeline;
325 struct rb_node *rb_leftmost;
326 struct rb_node *rb_load_balance_curr;
327 /* 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr;
332 unsigned long nr_spread_over;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 /* Real-Time classes' related field in a runqueue: */
352 struct rt_prio_array active;
353 unsigned long rt_nr_running;
354 #if defined CONFIG_SMP || defined CONFIG_FAIR_GROUP_SCHED
355 int highest_prio; /* highest queued rt task prio */
358 unsigned long rt_nr_migratory;
364 #ifdef CONFIG_FAIR_GROUP_SCHED
365 unsigned long rt_nr_boosted;
368 struct list_head leaf_rt_rq_list;
369 struct task_group *tg;
370 struct sched_rt_entity *rt_se;
377 * We add the notion of a root-domain which will be used to define per-domain
378 * variables. Each exclusive cpuset essentially defines an island domain by
379 * fully partitioning the member cpus from any other cpuset. Whenever a new
380 * exclusive cpuset is created, we also create and attach a new root-domain
390 * The "RT overload" flag: it gets set if a CPU has more than
391 * one runnable RT task.
398 * By default the system creates a single root-domain with all cpus as
399 * members (mimicking the global state we have today).
401 static struct root_domain def_root_domain;
406 * This is the main, per-CPU runqueue data structure.
408 * Locking rule: those places that want to lock multiple runqueues
409 * (such as the load balancing or the thread migration code), lock
410 * acquire operations must be ordered by ascending &runqueue.
417 * nr_running and cpu_load should be in the same cacheline because
418 * remote CPUs use both these fields when doing load calculation.
420 unsigned long nr_running;
421 #define CPU_LOAD_IDX_MAX 5
422 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
423 unsigned char idle_at_tick;
425 unsigned char in_nohz_recently;
427 /* capture load from *all* tasks on this cpu: */
428 struct load_weight load;
429 unsigned long nr_load_updates;
434 u64 rt_period_expire;
437 #ifdef CONFIG_FAIR_GROUP_SCHED
438 /* list of leaf cfs_rq on this cpu: */
439 struct list_head leaf_cfs_rq_list;
440 struct list_head leaf_rt_rq_list;
444 * This is part of a global counter where only the total sum
445 * over all CPUs matters. A task can increase this counter on
446 * one CPU and if it got migrated afterwards it may decrease
447 * it on another CPU. Always updated under the runqueue lock:
449 unsigned long nr_uninterruptible;
451 struct task_struct *curr, *idle;
452 unsigned long next_balance;
453 struct mm_struct *prev_mm;
455 u64 clock, prev_clock_raw;
458 unsigned int clock_warps, clock_overflows, clock_underflows;
460 unsigned int clock_deep_idle_events;
466 struct root_domain *rd;
467 struct sched_domain *sd;
469 /* For active balancing */
472 /* cpu of this runqueue: */
475 struct task_struct *migration_thread;
476 struct list_head migration_queue;
479 #ifdef CONFIG_SCHED_HRTICK
480 unsigned long hrtick_flags;
481 ktime_t hrtick_expire;
482 struct hrtimer hrtick_timer;
485 #ifdef CONFIG_SCHEDSTATS
487 struct sched_info rq_sched_info;
489 /* sys_sched_yield() stats */
490 unsigned int yld_exp_empty;
491 unsigned int yld_act_empty;
492 unsigned int yld_both_empty;
493 unsigned int yld_count;
495 /* schedule() stats */
496 unsigned int sched_switch;
497 unsigned int sched_count;
498 unsigned int sched_goidle;
500 /* try_to_wake_up() stats */
501 unsigned int ttwu_count;
502 unsigned int ttwu_local;
505 unsigned int bkl_count;
507 struct lock_class_key rq_lock_key;
510 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
512 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
514 rq->curr->sched_class->check_preempt_curr(rq, p);
517 static inline int cpu_of(struct rq *rq)
527 * Update the per-runqueue clock, as finegrained as the platform can give
528 * us, but without assuming monotonicity, etc.:
530 static void __update_rq_clock(struct rq *rq)
532 u64 prev_raw = rq->prev_clock_raw;
533 u64 now = sched_clock();
534 s64 delta = now - prev_raw;
535 u64 clock = rq->clock;
537 #ifdef CONFIG_SCHED_DEBUG
538 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
541 * Protect against sched_clock() occasionally going backwards:
543 if (unlikely(delta < 0)) {
548 * Catch too large forward jumps too:
550 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
551 if (clock < rq->tick_timestamp + TICK_NSEC)
552 clock = rq->tick_timestamp + TICK_NSEC;
555 rq->clock_overflows++;
557 if (unlikely(delta > rq->clock_max_delta))
558 rq->clock_max_delta = delta;
563 rq->prev_clock_raw = now;
567 static void update_rq_clock(struct rq *rq)
569 if (likely(smp_processor_id() == cpu_of(rq)))
570 __update_rq_clock(rq);
574 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
575 * See detach_destroy_domains: synchronize_sched for details.
577 * The domain tree of any CPU may only be accessed from within
578 * preempt-disabled sections.
580 #define for_each_domain(cpu, __sd) \
581 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
583 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
584 #define this_rq() (&__get_cpu_var(runqueues))
585 #define task_rq(p) cpu_rq(task_cpu(p))
586 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
588 unsigned long rt_needs_cpu(int cpu)
590 struct rq *rq = cpu_rq(cpu);
593 if (!rq->rt_throttled)
596 if (rq->clock > rq->rt_period_expire)
599 delta = rq->rt_period_expire - rq->clock;
600 do_div(delta, NSEC_PER_SEC / HZ);
602 return (unsigned long)delta;
606 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
608 #ifdef CONFIG_SCHED_DEBUG
609 # define const_debug __read_mostly
611 # define const_debug static const
615 * Debugging: various feature bits
618 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
619 SCHED_FEAT_WAKEUP_PREEMPT = 2,
620 SCHED_FEAT_START_DEBIT = 4,
621 SCHED_FEAT_TREE_AVG = 8,
622 SCHED_FEAT_APPROX_AVG = 16,
623 SCHED_FEAT_HRTICK = 32,
624 SCHED_FEAT_DOUBLE_TICK = 64,
627 const_debug unsigned int sysctl_sched_features =
628 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
629 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
630 SCHED_FEAT_START_DEBIT * 1 |
631 SCHED_FEAT_TREE_AVG * 0 |
632 SCHED_FEAT_APPROX_AVG * 0 |
633 SCHED_FEAT_HRTICK * 1 |
634 SCHED_FEAT_DOUBLE_TICK * 0;
636 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
639 * Number of tasks to iterate in a single balance run.
640 * Limited because this is done with IRQs disabled.
642 const_debug unsigned int sysctl_sched_nr_migrate = 32;
645 * period over which we measure -rt task cpu usage in ms.
648 const_debug unsigned int sysctl_sched_rt_period = 1000;
650 #define SCHED_RT_FRAC_SHIFT 16
651 #define SCHED_RT_FRAC (1UL << SCHED_RT_FRAC_SHIFT)
654 * ratio of time -rt tasks may consume.
657 const_debug unsigned int sysctl_sched_rt_ratio = 62259;
660 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
661 * clock constructed from sched_clock():
663 unsigned long long cpu_clock(int cpu)
665 unsigned long long now;
669 local_irq_save(flags);
672 * Only call sched_clock() if the scheduler has already been
673 * initialized (some code might call cpu_clock() very early):
678 local_irq_restore(flags);
682 EXPORT_SYMBOL_GPL(cpu_clock);
684 #ifndef prepare_arch_switch
685 # define prepare_arch_switch(next) do { } while (0)
687 #ifndef finish_arch_switch
688 # define finish_arch_switch(prev) do { } while (0)
691 static inline int task_current(struct rq *rq, struct task_struct *p)
693 return rq->curr == p;
696 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
697 static inline int task_running(struct rq *rq, struct task_struct *p)
699 return task_current(rq, p);
702 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
706 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
708 #ifdef CONFIG_DEBUG_SPINLOCK
709 /* this is a valid case when another task releases the spinlock */
710 rq->lock.owner = current;
713 * If we are tracking spinlock dependencies then we have to
714 * fix up the runqueue lock - which gets 'carried over' from
717 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
719 spin_unlock_irq(&rq->lock);
722 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
723 static inline int task_running(struct rq *rq, struct task_struct *p)
728 return task_current(rq, p);
732 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
736 * We can optimise this out completely for !SMP, because the
737 * SMP rebalancing from interrupt is the only thing that cares
742 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
743 spin_unlock_irq(&rq->lock);
745 spin_unlock(&rq->lock);
749 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
753 * After ->oncpu is cleared, the task can be moved to a different CPU.
754 * We must ensure this doesn't happen until the switch is completely
760 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
764 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
767 * __task_rq_lock - lock the runqueue a given task resides on.
768 * Must be called interrupts disabled.
770 static inline struct rq *__task_rq_lock(struct task_struct *p)
774 struct rq *rq = task_rq(p);
775 spin_lock(&rq->lock);
776 if (likely(rq == task_rq(p)))
778 spin_unlock(&rq->lock);
783 * task_rq_lock - lock the runqueue a given task resides on and disable
784 * interrupts. Note the ordering: we can safely lookup the task_rq without
785 * explicitly disabling preemption.
787 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
793 local_irq_save(*flags);
795 spin_lock(&rq->lock);
796 if (likely(rq == task_rq(p)))
798 spin_unlock_irqrestore(&rq->lock, *flags);
802 static void __task_rq_unlock(struct rq *rq)
805 spin_unlock(&rq->lock);
808 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
811 spin_unlock_irqrestore(&rq->lock, *flags);
815 * this_rq_lock - lock this runqueue and disable interrupts.
817 static struct rq *this_rq_lock(void)
824 spin_lock(&rq->lock);
830 * We are going deep-idle (irqs are disabled):
832 void sched_clock_idle_sleep_event(void)
834 struct rq *rq = cpu_rq(smp_processor_id());
836 spin_lock(&rq->lock);
837 __update_rq_clock(rq);
838 spin_unlock(&rq->lock);
839 rq->clock_deep_idle_events++;
841 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
844 * We just idled delta nanoseconds (called with irqs disabled):
846 void sched_clock_idle_wakeup_event(u64 delta_ns)
848 struct rq *rq = cpu_rq(smp_processor_id());
849 u64 now = sched_clock();
851 rq->idle_clock += delta_ns;
853 * Override the previous timestamp and ignore all
854 * sched_clock() deltas that occured while we idled,
855 * and use the PM-provided delta_ns to advance the
858 spin_lock(&rq->lock);
859 rq->prev_clock_raw = now;
860 rq->clock += delta_ns;
861 spin_unlock(&rq->lock);
862 touch_softlockup_watchdog();
864 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
866 static void __resched_task(struct task_struct *p, int tif_bit);
868 static inline void resched_task(struct task_struct *p)
870 __resched_task(p, TIF_NEED_RESCHED);
873 #ifdef CONFIG_SCHED_HRTICK
875 * Use HR-timers to deliver accurate preemption points.
877 * Its all a bit involved since we cannot program an hrt while holding the
878 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
881 * When we get rescheduled we reprogram the hrtick_timer outside of the
884 static inline void resched_hrt(struct task_struct *p)
886 __resched_task(p, TIF_HRTICK_RESCHED);
889 static inline void resched_rq(struct rq *rq)
893 spin_lock_irqsave(&rq->lock, flags);
894 resched_task(rq->curr);
895 spin_unlock_irqrestore(&rq->lock, flags);
899 HRTICK_SET, /* re-programm hrtick_timer */
900 HRTICK_RESET, /* not a new slice */
905 * - enabled by features
906 * - hrtimer is actually high res
908 static inline int hrtick_enabled(struct rq *rq)
910 if (!sched_feat(HRTICK))
912 return hrtimer_is_hres_active(&rq->hrtick_timer);
916 * Called to set the hrtick timer state.
918 * called with rq->lock held and irqs disabled
920 static void hrtick_start(struct rq *rq, u64 delay, int reset)
922 assert_spin_locked(&rq->lock);
925 * preempt at: now + delay
928 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
930 * indicate we need to program the timer
932 __set_bit(HRTICK_SET, &rq->hrtick_flags);
934 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
937 * New slices are called from the schedule path and don't need a
941 resched_hrt(rq->curr);
944 static void hrtick_clear(struct rq *rq)
946 if (hrtimer_active(&rq->hrtick_timer))
947 hrtimer_cancel(&rq->hrtick_timer);
951 * Update the timer from the possible pending state.
953 static void hrtick_set(struct rq *rq)
959 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
961 spin_lock_irqsave(&rq->lock, flags);
962 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
963 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
964 time = rq->hrtick_expire;
965 clear_thread_flag(TIF_HRTICK_RESCHED);
966 spin_unlock_irqrestore(&rq->lock, flags);
969 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
970 if (reset && !hrtimer_active(&rq->hrtick_timer))
977 * High-resolution timer tick.
978 * Runs from hardirq context with interrupts disabled.
980 static enum hrtimer_restart hrtick(struct hrtimer *timer)
982 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
984 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
986 spin_lock(&rq->lock);
987 __update_rq_clock(rq);
988 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
989 spin_unlock(&rq->lock);
991 return HRTIMER_NORESTART;
994 static inline void init_rq_hrtick(struct rq *rq)
996 rq->hrtick_flags = 0;
997 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
998 rq->hrtick_timer.function = hrtick;
999 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1002 void hrtick_resched(void)
1005 unsigned long flags;
1007 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1010 local_irq_save(flags);
1011 rq = cpu_rq(smp_processor_id());
1013 local_irq_restore(flags);
1016 static inline void hrtick_clear(struct rq *rq)
1020 static inline void hrtick_set(struct rq *rq)
1024 static inline void init_rq_hrtick(struct rq *rq)
1028 void hrtick_resched(void)
1034 * resched_task - mark a task 'to be rescheduled now'.
1036 * On UP this means the setting of the need_resched flag, on SMP it
1037 * might also involve a cross-CPU call to trigger the scheduler on
1042 #ifndef tsk_is_polling
1043 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1046 static void __resched_task(struct task_struct *p, int tif_bit)
1050 assert_spin_locked(&task_rq(p)->lock);
1052 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1055 set_tsk_thread_flag(p, tif_bit);
1058 if (cpu == smp_processor_id())
1061 /* NEED_RESCHED must be visible before we test polling */
1063 if (!tsk_is_polling(p))
1064 smp_send_reschedule(cpu);
1067 static void resched_cpu(int cpu)
1069 struct rq *rq = cpu_rq(cpu);
1070 unsigned long flags;
1072 if (!spin_trylock_irqsave(&rq->lock, flags))
1074 resched_task(cpu_curr(cpu));
1075 spin_unlock_irqrestore(&rq->lock, flags);
1078 static void __resched_task(struct task_struct *p, int tif_bit)
1080 assert_spin_locked(&task_rq(p)->lock);
1081 set_tsk_thread_flag(p, tif_bit);
1085 #if BITS_PER_LONG == 32
1086 # define WMULT_CONST (~0UL)
1088 # define WMULT_CONST (1UL << 32)
1091 #define WMULT_SHIFT 32
1094 * Shift right and round:
1096 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1098 static unsigned long
1099 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1100 struct load_weight *lw)
1104 if (unlikely(!lw->inv_weight))
1105 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
1107 tmp = (u64)delta_exec * weight;
1109 * Check whether we'd overflow the 64-bit multiplication:
1111 if (unlikely(tmp > WMULT_CONST))
1112 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1115 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1117 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1120 static inline unsigned long
1121 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1123 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1126 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1131 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1137 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1138 * of tasks with abnormal "nice" values across CPUs the contribution that
1139 * each task makes to its run queue's load is weighted according to its
1140 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1141 * scaled version of the new time slice allocation that they receive on time
1145 #define WEIGHT_IDLEPRIO 2
1146 #define WMULT_IDLEPRIO (1 << 31)
1149 * Nice levels are multiplicative, with a gentle 10% change for every
1150 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1151 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1152 * that remained on nice 0.
1154 * The "10% effect" is relative and cumulative: from _any_ nice level,
1155 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1156 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1157 * If a task goes up by ~10% and another task goes down by ~10% then
1158 * the relative distance between them is ~25%.)
1160 static const int prio_to_weight[40] = {
1161 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1162 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1163 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1164 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1165 /* 0 */ 1024, 820, 655, 526, 423,
1166 /* 5 */ 335, 272, 215, 172, 137,
1167 /* 10 */ 110, 87, 70, 56, 45,
1168 /* 15 */ 36, 29, 23, 18, 15,
1172 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1174 * In cases where the weight does not change often, we can use the
1175 * precalculated inverse to speed up arithmetics by turning divisions
1176 * into multiplications:
1178 static const u32 prio_to_wmult[40] = {
1179 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1180 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1181 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1182 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1183 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1184 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1185 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1186 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1189 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1192 * runqueue iterator, to support SMP load-balancing between different
1193 * scheduling classes, without having to expose their internal data
1194 * structures to the load-balancing proper:
1196 struct rq_iterator {
1198 struct task_struct *(*start)(void *);
1199 struct task_struct *(*next)(void *);
1203 static unsigned long
1204 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1205 unsigned long max_load_move, struct sched_domain *sd,
1206 enum cpu_idle_type idle, int *all_pinned,
1207 int *this_best_prio, struct rq_iterator *iterator);
1210 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1211 struct sched_domain *sd, enum cpu_idle_type idle,
1212 struct rq_iterator *iterator);
1215 #ifdef CONFIG_CGROUP_CPUACCT
1216 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1218 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1221 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1223 update_load_add(&rq->load, load);
1226 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1228 update_load_sub(&rq->load, load);
1232 static unsigned long source_load(int cpu, int type);
1233 static unsigned long target_load(int cpu, int type);
1234 static unsigned long cpu_avg_load_per_task(int cpu);
1235 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1236 #endif /* CONFIG_SMP */
1238 #include "sched_stats.h"
1239 #include "sched_idletask.c"
1240 #include "sched_fair.c"
1241 #include "sched_rt.c"
1242 #ifdef CONFIG_SCHED_DEBUG
1243 # include "sched_debug.c"
1246 #define sched_class_highest (&rt_sched_class)
1248 static void inc_nr_running(struct rq *rq)
1253 static void dec_nr_running(struct rq *rq)
1258 static void set_load_weight(struct task_struct *p)
1260 if (task_has_rt_policy(p)) {
1261 p->se.load.weight = prio_to_weight[0] * 2;
1262 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1267 * SCHED_IDLE tasks get minimal weight:
1269 if (p->policy == SCHED_IDLE) {
1270 p->se.load.weight = WEIGHT_IDLEPRIO;
1271 p->se.load.inv_weight = WMULT_IDLEPRIO;
1275 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1276 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1279 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1281 sched_info_queued(p);
1282 p->sched_class->enqueue_task(rq, p, wakeup);
1286 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1288 p->sched_class->dequeue_task(rq, p, sleep);
1293 * __normal_prio - return the priority that is based on the static prio
1295 static inline int __normal_prio(struct task_struct *p)
1297 return p->static_prio;
1301 * Calculate the expected normal priority: i.e. priority
1302 * without taking RT-inheritance into account. Might be
1303 * boosted by interactivity modifiers. Changes upon fork,
1304 * setprio syscalls, and whenever the interactivity
1305 * estimator recalculates.
1307 static inline int normal_prio(struct task_struct *p)
1311 if (task_has_rt_policy(p))
1312 prio = MAX_RT_PRIO-1 - p->rt_priority;
1314 prio = __normal_prio(p);
1319 * Calculate the current priority, i.e. the priority
1320 * taken into account by the scheduler. This value might
1321 * be boosted by RT tasks, or might be boosted by
1322 * interactivity modifiers. Will be RT if the task got
1323 * RT-boosted. If not then it returns p->normal_prio.
1325 static int effective_prio(struct task_struct *p)
1327 p->normal_prio = normal_prio(p);
1329 * If we are RT tasks or we were boosted to RT priority,
1330 * keep the priority unchanged. Otherwise, update priority
1331 * to the normal priority:
1333 if (!rt_prio(p->prio))
1334 return p->normal_prio;
1339 * activate_task - move a task to the runqueue.
1341 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1343 if (task_contributes_to_load(p))
1344 rq->nr_uninterruptible--;
1346 enqueue_task(rq, p, wakeup);
1351 * deactivate_task - remove a task from the runqueue.
1353 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1355 if (task_contributes_to_load(p))
1356 rq->nr_uninterruptible++;
1358 dequeue_task(rq, p, sleep);
1363 * task_curr - is this task currently executing on a CPU?
1364 * @p: the task in question.
1366 inline int task_curr(const struct task_struct *p)
1368 return cpu_curr(task_cpu(p)) == p;
1371 /* Used instead of source_load when we know the type == 0 */
1372 unsigned long weighted_cpuload(const int cpu)
1374 return cpu_rq(cpu)->load.weight;
1377 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1379 set_task_rq(p, cpu);
1382 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1383 * successfuly executed on another CPU. We must ensure that updates of
1384 * per-task data have been completed by this moment.
1387 task_thread_info(p)->cpu = cpu;
1391 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1392 const struct sched_class *prev_class,
1393 int oldprio, int running)
1395 if (prev_class != p->sched_class) {
1396 if (prev_class->switched_from)
1397 prev_class->switched_from(rq, p, running);
1398 p->sched_class->switched_to(rq, p, running);
1400 p->sched_class->prio_changed(rq, p, oldprio, running);
1406 * Is this task likely cache-hot:
1409 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1413 if (p->sched_class != &fair_sched_class)
1416 if (sysctl_sched_migration_cost == -1)
1418 if (sysctl_sched_migration_cost == 0)
1421 delta = now - p->se.exec_start;
1423 return delta < (s64)sysctl_sched_migration_cost;
1427 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1429 int old_cpu = task_cpu(p);
1430 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1431 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1432 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1435 clock_offset = old_rq->clock - new_rq->clock;
1437 #ifdef CONFIG_SCHEDSTATS
1438 if (p->se.wait_start)
1439 p->se.wait_start -= clock_offset;
1440 if (p->se.sleep_start)
1441 p->se.sleep_start -= clock_offset;
1442 if (p->se.block_start)
1443 p->se.block_start -= clock_offset;
1444 if (old_cpu != new_cpu) {
1445 schedstat_inc(p, se.nr_migrations);
1446 if (task_hot(p, old_rq->clock, NULL))
1447 schedstat_inc(p, se.nr_forced2_migrations);
1450 p->se.vruntime -= old_cfsrq->min_vruntime -
1451 new_cfsrq->min_vruntime;
1453 __set_task_cpu(p, new_cpu);
1456 struct migration_req {
1457 struct list_head list;
1459 struct task_struct *task;
1462 struct completion done;
1466 * The task's runqueue lock must be held.
1467 * Returns true if you have to wait for migration thread.
1470 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1472 struct rq *rq = task_rq(p);
1475 * If the task is not on a runqueue (and not running), then
1476 * it is sufficient to simply update the task's cpu field.
1478 if (!p->se.on_rq && !task_running(rq, p)) {
1479 set_task_cpu(p, dest_cpu);
1483 init_completion(&req->done);
1485 req->dest_cpu = dest_cpu;
1486 list_add(&req->list, &rq->migration_queue);
1492 * wait_task_inactive - wait for a thread to unschedule.
1494 * The caller must ensure that the task *will* unschedule sometime soon,
1495 * else this function might spin for a *long* time. This function can't
1496 * be called with interrupts off, or it may introduce deadlock with
1497 * smp_call_function() if an IPI is sent by the same process we are
1498 * waiting to become inactive.
1500 void wait_task_inactive(struct task_struct *p)
1502 unsigned long flags;
1508 * We do the initial early heuristics without holding
1509 * any task-queue locks at all. We'll only try to get
1510 * the runqueue lock when things look like they will
1516 * If the task is actively running on another CPU
1517 * still, just relax and busy-wait without holding
1520 * NOTE! Since we don't hold any locks, it's not
1521 * even sure that "rq" stays as the right runqueue!
1522 * But we don't care, since "task_running()" will
1523 * return false if the runqueue has changed and p
1524 * is actually now running somewhere else!
1526 while (task_running(rq, p))
1530 * Ok, time to look more closely! We need the rq
1531 * lock now, to be *sure*. If we're wrong, we'll
1532 * just go back and repeat.
1534 rq = task_rq_lock(p, &flags);
1535 running = task_running(rq, p);
1536 on_rq = p->se.on_rq;
1537 task_rq_unlock(rq, &flags);
1540 * Was it really running after all now that we
1541 * checked with the proper locks actually held?
1543 * Oops. Go back and try again..
1545 if (unlikely(running)) {
1551 * It's not enough that it's not actively running,
1552 * it must be off the runqueue _entirely_, and not
1555 * So if it wa still runnable (but just not actively
1556 * running right now), it's preempted, and we should
1557 * yield - it could be a while.
1559 if (unlikely(on_rq)) {
1560 schedule_timeout_uninterruptible(1);
1565 * Ahh, all good. It wasn't running, and it wasn't
1566 * runnable, which means that it will never become
1567 * running in the future either. We're all done!
1574 * kick_process - kick a running thread to enter/exit the kernel
1575 * @p: the to-be-kicked thread
1577 * Cause a process which is running on another CPU to enter
1578 * kernel-mode, without any delay. (to get signals handled.)
1580 * NOTE: this function doesnt have to take the runqueue lock,
1581 * because all it wants to ensure is that the remote task enters
1582 * the kernel. If the IPI races and the task has been migrated
1583 * to another CPU then no harm is done and the purpose has been
1586 void kick_process(struct task_struct *p)
1592 if ((cpu != smp_processor_id()) && task_curr(p))
1593 smp_send_reschedule(cpu);
1598 * Return a low guess at the load of a migration-source cpu weighted
1599 * according to the scheduling class and "nice" value.
1601 * We want to under-estimate the load of migration sources, to
1602 * balance conservatively.
1604 static unsigned long source_load(int cpu, int type)
1606 struct rq *rq = cpu_rq(cpu);
1607 unsigned long total = weighted_cpuload(cpu);
1612 return min(rq->cpu_load[type-1], total);
1616 * Return a high guess at the load of a migration-target cpu weighted
1617 * according to the scheduling class and "nice" value.
1619 static unsigned long target_load(int cpu, int type)
1621 struct rq *rq = cpu_rq(cpu);
1622 unsigned long total = weighted_cpuload(cpu);
1627 return max(rq->cpu_load[type-1], total);
1631 * Return the average load per task on the cpu's run queue
1633 static unsigned long cpu_avg_load_per_task(int cpu)
1635 struct rq *rq = cpu_rq(cpu);
1636 unsigned long total = weighted_cpuload(cpu);
1637 unsigned long n = rq->nr_running;
1639 return n ? total / n : SCHED_LOAD_SCALE;
1643 * find_idlest_group finds and returns the least busy CPU group within the
1646 static struct sched_group *
1647 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1649 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1650 unsigned long min_load = ULONG_MAX, this_load = 0;
1651 int load_idx = sd->forkexec_idx;
1652 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1655 unsigned long load, avg_load;
1659 /* Skip over this group if it has no CPUs allowed */
1660 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1663 local_group = cpu_isset(this_cpu, group->cpumask);
1665 /* Tally up the load of all CPUs in the group */
1668 for_each_cpu_mask(i, group->cpumask) {
1669 /* Bias balancing toward cpus of our domain */
1671 load = source_load(i, load_idx);
1673 load = target_load(i, load_idx);
1678 /* Adjust by relative CPU power of the group */
1679 avg_load = sg_div_cpu_power(group,
1680 avg_load * SCHED_LOAD_SCALE);
1683 this_load = avg_load;
1685 } else if (avg_load < min_load) {
1686 min_load = avg_load;
1689 } while (group = group->next, group != sd->groups);
1691 if (!idlest || 100*this_load < imbalance*min_load)
1697 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1700 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1703 unsigned long load, min_load = ULONG_MAX;
1707 /* Traverse only the allowed CPUs */
1708 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1710 for_each_cpu_mask(i, tmp) {
1711 load = weighted_cpuload(i);
1713 if (load < min_load || (load == min_load && i == this_cpu)) {
1723 * sched_balance_self: balance the current task (running on cpu) in domains
1724 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1727 * Balance, ie. select the least loaded group.
1729 * Returns the target CPU number, or the same CPU if no balancing is needed.
1731 * preempt must be disabled.
1733 static int sched_balance_self(int cpu, int flag)
1735 struct task_struct *t = current;
1736 struct sched_domain *tmp, *sd = NULL;
1738 for_each_domain(cpu, tmp) {
1740 * If power savings logic is enabled for a domain, stop there.
1742 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1744 if (tmp->flags & flag)
1750 struct sched_group *group;
1751 int new_cpu, weight;
1753 if (!(sd->flags & flag)) {
1759 group = find_idlest_group(sd, t, cpu);
1765 new_cpu = find_idlest_cpu(group, t, cpu);
1766 if (new_cpu == -1 || new_cpu == cpu) {
1767 /* Now try balancing at a lower domain level of cpu */
1772 /* Now try balancing at a lower domain level of new_cpu */
1775 weight = cpus_weight(span);
1776 for_each_domain(cpu, tmp) {
1777 if (weight <= cpus_weight(tmp->span))
1779 if (tmp->flags & flag)
1782 /* while loop will break here if sd == NULL */
1788 #endif /* CONFIG_SMP */
1791 * try_to_wake_up - wake up a thread
1792 * @p: the to-be-woken-up thread
1793 * @state: the mask of task states that can be woken
1794 * @sync: do a synchronous wakeup?
1796 * Put it on the run-queue if it's not already there. The "current"
1797 * thread is always on the run-queue (except when the actual
1798 * re-schedule is in progress), and as such you're allowed to do
1799 * the simpler "current->state = TASK_RUNNING" to mark yourself
1800 * runnable without the overhead of this.
1802 * returns failure only if the task is already active.
1804 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1806 int cpu, orig_cpu, this_cpu, success = 0;
1807 unsigned long flags;
1811 rq = task_rq_lock(p, &flags);
1812 old_state = p->state;
1813 if (!(old_state & state))
1821 this_cpu = smp_processor_id();
1824 if (unlikely(task_running(rq, p)))
1827 cpu = p->sched_class->select_task_rq(p, sync);
1828 if (cpu != orig_cpu) {
1829 set_task_cpu(p, cpu);
1830 task_rq_unlock(rq, &flags);
1831 /* might preempt at this point */
1832 rq = task_rq_lock(p, &flags);
1833 old_state = p->state;
1834 if (!(old_state & state))
1839 this_cpu = smp_processor_id();
1843 #ifdef CONFIG_SCHEDSTATS
1844 schedstat_inc(rq, ttwu_count);
1845 if (cpu == this_cpu)
1846 schedstat_inc(rq, ttwu_local);
1848 struct sched_domain *sd;
1849 for_each_domain(this_cpu, sd) {
1850 if (cpu_isset(cpu, sd->span)) {
1851 schedstat_inc(sd, ttwu_wake_remote);
1859 #endif /* CONFIG_SMP */
1860 schedstat_inc(p, se.nr_wakeups);
1862 schedstat_inc(p, se.nr_wakeups_sync);
1863 if (orig_cpu != cpu)
1864 schedstat_inc(p, se.nr_wakeups_migrate);
1865 if (cpu == this_cpu)
1866 schedstat_inc(p, se.nr_wakeups_local);
1868 schedstat_inc(p, se.nr_wakeups_remote);
1869 update_rq_clock(rq);
1870 activate_task(rq, p, 1);
1871 check_preempt_curr(rq, p);
1875 p->state = TASK_RUNNING;
1877 if (p->sched_class->task_wake_up)
1878 p->sched_class->task_wake_up(rq, p);
1881 task_rq_unlock(rq, &flags);
1886 int wake_up_process(struct task_struct *p)
1888 return try_to_wake_up(p, TASK_ALL, 0);
1890 EXPORT_SYMBOL(wake_up_process);
1892 int wake_up_state(struct task_struct *p, unsigned int state)
1894 return try_to_wake_up(p, state, 0);
1898 * Perform scheduler related setup for a newly forked process p.
1899 * p is forked by current.
1901 * __sched_fork() is basic setup used by init_idle() too:
1903 static void __sched_fork(struct task_struct *p)
1905 p->se.exec_start = 0;
1906 p->se.sum_exec_runtime = 0;
1907 p->se.prev_sum_exec_runtime = 0;
1909 #ifdef CONFIG_SCHEDSTATS
1910 p->se.wait_start = 0;
1911 p->se.sum_sleep_runtime = 0;
1912 p->se.sleep_start = 0;
1913 p->se.block_start = 0;
1914 p->se.sleep_max = 0;
1915 p->se.block_max = 0;
1917 p->se.slice_max = 0;
1921 INIT_LIST_HEAD(&p->rt.run_list);
1924 #ifdef CONFIG_PREEMPT_NOTIFIERS
1925 INIT_HLIST_HEAD(&p->preempt_notifiers);
1929 * We mark the process as running here, but have not actually
1930 * inserted it onto the runqueue yet. This guarantees that
1931 * nobody will actually run it, and a signal or other external
1932 * event cannot wake it up and insert it on the runqueue either.
1934 p->state = TASK_RUNNING;
1938 * fork()/clone()-time setup:
1940 void sched_fork(struct task_struct *p, int clone_flags)
1942 int cpu = get_cpu();
1947 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1949 set_task_cpu(p, cpu);
1952 * Make sure we do not leak PI boosting priority to the child:
1954 p->prio = current->normal_prio;
1955 if (!rt_prio(p->prio))
1956 p->sched_class = &fair_sched_class;
1958 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1959 if (likely(sched_info_on()))
1960 memset(&p->sched_info, 0, sizeof(p->sched_info));
1962 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1965 #ifdef CONFIG_PREEMPT
1966 /* Want to start with kernel preemption disabled. */
1967 task_thread_info(p)->preempt_count = 1;
1973 * wake_up_new_task - wake up a newly created task for the first time.
1975 * This function will do some initial scheduler statistics housekeeping
1976 * that must be done for every newly created context, then puts the task
1977 * on the runqueue and wakes it.
1979 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1981 unsigned long flags;
1984 rq = task_rq_lock(p, &flags);
1985 BUG_ON(p->state != TASK_RUNNING);
1986 update_rq_clock(rq);
1988 p->prio = effective_prio(p);
1990 if (!p->sched_class->task_new || !current->se.on_rq) {
1991 activate_task(rq, p, 0);
1994 * Let the scheduling class do new task startup
1995 * management (if any):
1997 p->sched_class->task_new(rq, p);
2000 check_preempt_curr(rq, p);
2002 if (p->sched_class->task_wake_up)
2003 p->sched_class->task_wake_up(rq, p);
2005 task_rq_unlock(rq, &flags);
2008 #ifdef CONFIG_PREEMPT_NOTIFIERS
2011 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2012 * @notifier: notifier struct to register
2014 void preempt_notifier_register(struct preempt_notifier *notifier)
2016 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2018 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2021 * preempt_notifier_unregister - no longer interested in preemption notifications
2022 * @notifier: notifier struct to unregister
2024 * This is safe to call from within a preemption notifier.
2026 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2028 hlist_del(¬ifier->link);
2030 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2032 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2034 struct preempt_notifier *notifier;
2035 struct hlist_node *node;
2037 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2038 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2042 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2043 struct task_struct *next)
2045 struct preempt_notifier *notifier;
2046 struct hlist_node *node;
2048 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2049 notifier->ops->sched_out(notifier, next);
2054 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2059 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2060 struct task_struct *next)
2067 * prepare_task_switch - prepare to switch tasks
2068 * @rq: the runqueue preparing to switch
2069 * @prev: the current task that is being switched out
2070 * @next: the task we are going to switch to.
2072 * This is called with the rq lock held and interrupts off. It must
2073 * be paired with a subsequent finish_task_switch after the context
2076 * prepare_task_switch sets up locking and calls architecture specific
2080 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2081 struct task_struct *next)
2083 fire_sched_out_preempt_notifiers(prev, next);
2084 prepare_lock_switch(rq, next);
2085 prepare_arch_switch(next);
2089 * finish_task_switch - clean up after a task-switch
2090 * @rq: runqueue associated with task-switch
2091 * @prev: the thread we just switched away from.
2093 * finish_task_switch must be called after the context switch, paired
2094 * with a prepare_task_switch call before the context switch.
2095 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2096 * and do any other architecture-specific cleanup actions.
2098 * Note that we may have delayed dropping an mm in context_switch(). If
2099 * so, we finish that here outside of the runqueue lock. (Doing it
2100 * with the lock held can cause deadlocks; see schedule() for
2103 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2104 __releases(rq->lock)
2106 struct mm_struct *mm = rq->prev_mm;
2112 * A task struct has one reference for the use as "current".
2113 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2114 * schedule one last time. The schedule call will never return, and
2115 * the scheduled task must drop that reference.
2116 * The test for TASK_DEAD must occur while the runqueue locks are
2117 * still held, otherwise prev could be scheduled on another cpu, die
2118 * there before we look at prev->state, and then the reference would
2120 * Manfred Spraul <manfred@colorfullife.com>
2122 prev_state = prev->state;
2123 finish_arch_switch(prev);
2124 finish_lock_switch(rq, prev);
2126 if (current->sched_class->post_schedule)
2127 current->sched_class->post_schedule(rq);
2130 fire_sched_in_preempt_notifiers(current);
2133 if (unlikely(prev_state == TASK_DEAD)) {
2135 * Remove function-return probe instances associated with this
2136 * task and put them back on the free list.
2138 kprobe_flush_task(prev);
2139 put_task_struct(prev);
2144 * schedule_tail - first thing a freshly forked thread must call.
2145 * @prev: the thread we just switched away from.
2147 asmlinkage void schedule_tail(struct task_struct *prev)
2148 __releases(rq->lock)
2150 struct rq *rq = this_rq();
2152 finish_task_switch(rq, prev);
2153 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2154 /* In this case, finish_task_switch does not reenable preemption */
2157 if (current->set_child_tid)
2158 put_user(task_pid_vnr(current), current->set_child_tid);
2162 * context_switch - switch to the new MM and the new
2163 * thread's register state.
2166 context_switch(struct rq *rq, struct task_struct *prev,
2167 struct task_struct *next)
2169 struct mm_struct *mm, *oldmm;
2171 prepare_task_switch(rq, prev, next);
2173 oldmm = prev->active_mm;
2175 * For paravirt, this is coupled with an exit in switch_to to
2176 * combine the page table reload and the switch backend into
2179 arch_enter_lazy_cpu_mode();
2181 if (unlikely(!mm)) {
2182 next->active_mm = oldmm;
2183 atomic_inc(&oldmm->mm_count);
2184 enter_lazy_tlb(oldmm, next);
2186 switch_mm(oldmm, mm, next);
2188 if (unlikely(!prev->mm)) {
2189 prev->active_mm = NULL;
2190 rq->prev_mm = oldmm;
2193 * Since the runqueue lock will be released by the next
2194 * task (which is an invalid locking op but in the case
2195 * of the scheduler it's an obvious special-case), so we
2196 * do an early lockdep release here:
2198 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2199 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2202 /* Here we just switch the register state and the stack. */
2203 switch_to(prev, next, prev);
2207 * this_rq must be evaluated again because prev may have moved
2208 * CPUs since it called schedule(), thus the 'rq' on its stack
2209 * frame will be invalid.
2211 finish_task_switch(this_rq(), prev);
2215 * nr_running, nr_uninterruptible and nr_context_switches:
2217 * externally visible scheduler statistics: current number of runnable
2218 * threads, current number of uninterruptible-sleeping threads, total
2219 * number of context switches performed since bootup.
2221 unsigned long nr_running(void)
2223 unsigned long i, sum = 0;
2225 for_each_online_cpu(i)
2226 sum += cpu_rq(i)->nr_running;
2231 unsigned long nr_uninterruptible(void)
2233 unsigned long i, sum = 0;
2235 for_each_possible_cpu(i)
2236 sum += cpu_rq(i)->nr_uninterruptible;
2239 * Since we read the counters lockless, it might be slightly
2240 * inaccurate. Do not allow it to go below zero though:
2242 if (unlikely((long)sum < 0))
2248 unsigned long long nr_context_switches(void)
2251 unsigned long long sum = 0;
2253 for_each_possible_cpu(i)
2254 sum += cpu_rq(i)->nr_switches;
2259 unsigned long nr_iowait(void)
2261 unsigned long i, sum = 0;
2263 for_each_possible_cpu(i)
2264 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2269 unsigned long nr_active(void)
2271 unsigned long i, running = 0, uninterruptible = 0;
2273 for_each_online_cpu(i) {
2274 running += cpu_rq(i)->nr_running;
2275 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2278 if (unlikely((long)uninterruptible < 0))
2279 uninterruptible = 0;
2281 return running + uninterruptible;
2285 * Update rq->cpu_load[] statistics. This function is usually called every
2286 * scheduler tick (TICK_NSEC).
2288 static void update_cpu_load(struct rq *this_rq)
2290 unsigned long this_load = this_rq->load.weight;
2293 this_rq->nr_load_updates++;
2295 /* Update our load: */
2296 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2297 unsigned long old_load, new_load;
2299 /* scale is effectively 1 << i now, and >> i divides by scale */
2301 old_load = this_rq->cpu_load[i];
2302 new_load = this_load;
2304 * Round up the averaging division if load is increasing. This
2305 * prevents us from getting stuck on 9 if the load is 10, for
2308 if (new_load > old_load)
2309 new_load += scale-1;
2310 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2317 * double_rq_lock - safely lock two runqueues
2319 * Note this does not disable interrupts like task_rq_lock,
2320 * you need to do so manually before calling.
2322 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2323 __acquires(rq1->lock)
2324 __acquires(rq2->lock)
2326 BUG_ON(!irqs_disabled());
2328 spin_lock(&rq1->lock);
2329 __acquire(rq2->lock); /* Fake it out ;) */
2332 spin_lock(&rq1->lock);
2333 spin_lock(&rq2->lock);
2335 spin_lock(&rq2->lock);
2336 spin_lock(&rq1->lock);
2339 update_rq_clock(rq1);
2340 update_rq_clock(rq2);
2344 * double_rq_unlock - safely unlock two runqueues
2346 * Note this does not restore interrupts like task_rq_unlock,
2347 * you need to do so manually after calling.
2349 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2350 __releases(rq1->lock)
2351 __releases(rq2->lock)
2353 spin_unlock(&rq1->lock);
2355 spin_unlock(&rq2->lock);
2357 __release(rq2->lock);
2361 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2363 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2364 __releases(this_rq->lock)
2365 __acquires(busiest->lock)
2366 __acquires(this_rq->lock)
2370 if (unlikely(!irqs_disabled())) {
2371 /* printk() doesn't work good under rq->lock */
2372 spin_unlock(&this_rq->lock);
2375 if (unlikely(!spin_trylock(&busiest->lock))) {
2376 if (busiest < this_rq) {
2377 spin_unlock(&this_rq->lock);
2378 spin_lock(&busiest->lock);
2379 spin_lock(&this_rq->lock);
2382 spin_lock(&busiest->lock);
2388 * If dest_cpu is allowed for this process, migrate the task to it.
2389 * This is accomplished by forcing the cpu_allowed mask to only
2390 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2391 * the cpu_allowed mask is restored.
2393 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2395 struct migration_req req;
2396 unsigned long flags;
2399 rq = task_rq_lock(p, &flags);
2400 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2401 || unlikely(cpu_is_offline(dest_cpu)))
2404 /* force the process onto the specified CPU */
2405 if (migrate_task(p, dest_cpu, &req)) {
2406 /* Need to wait for migration thread (might exit: take ref). */
2407 struct task_struct *mt = rq->migration_thread;
2409 get_task_struct(mt);
2410 task_rq_unlock(rq, &flags);
2411 wake_up_process(mt);
2412 put_task_struct(mt);
2413 wait_for_completion(&req.done);
2418 task_rq_unlock(rq, &flags);
2422 * sched_exec - execve() is a valuable balancing opportunity, because at
2423 * this point the task has the smallest effective memory and cache footprint.
2425 void sched_exec(void)
2427 int new_cpu, this_cpu = get_cpu();
2428 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2430 if (new_cpu != this_cpu)
2431 sched_migrate_task(current, new_cpu);
2435 * pull_task - move a task from a remote runqueue to the local runqueue.
2436 * Both runqueues must be locked.
2438 static void pull_task(struct rq *src_rq, struct task_struct *p,
2439 struct rq *this_rq, int this_cpu)
2441 deactivate_task(src_rq, p, 0);
2442 set_task_cpu(p, this_cpu);
2443 activate_task(this_rq, p, 0);
2445 * Note that idle threads have a prio of MAX_PRIO, for this test
2446 * to be always true for them.
2448 check_preempt_curr(this_rq, p);
2452 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2455 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2456 struct sched_domain *sd, enum cpu_idle_type idle,
2460 * We do not migrate tasks that are:
2461 * 1) running (obviously), or
2462 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2463 * 3) are cache-hot on their current CPU.
2465 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2466 schedstat_inc(p, se.nr_failed_migrations_affine);
2471 if (task_running(rq, p)) {
2472 schedstat_inc(p, se.nr_failed_migrations_running);
2477 * Aggressive migration if:
2478 * 1) task is cache cold, or
2479 * 2) too many balance attempts have failed.
2482 if (!task_hot(p, rq->clock, sd) ||
2483 sd->nr_balance_failed > sd->cache_nice_tries) {
2484 #ifdef CONFIG_SCHEDSTATS
2485 if (task_hot(p, rq->clock, sd)) {
2486 schedstat_inc(sd, lb_hot_gained[idle]);
2487 schedstat_inc(p, se.nr_forced_migrations);
2493 if (task_hot(p, rq->clock, sd)) {
2494 schedstat_inc(p, se.nr_failed_migrations_hot);
2500 static unsigned long
2501 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2502 unsigned long max_load_move, struct sched_domain *sd,
2503 enum cpu_idle_type idle, int *all_pinned,
2504 int *this_best_prio, struct rq_iterator *iterator)
2506 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2507 struct task_struct *p;
2508 long rem_load_move = max_load_move;
2510 if (max_load_move == 0)
2516 * Start the load-balancing iterator:
2518 p = iterator->start(iterator->arg);
2520 if (!p || loops++ > sysctl_sched_nr_migrate)
2523 * To help distribute high priority tasks across CPUs we don't
2524 * skip a task if it will be the highest priority task (i.e. smallest
2525 * prio value) on its new queue regardless of its load weight
2527 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2528 SCHED_LOAD_SCALE_FUZZ;
2529 if ((skip_for_load && p->prio >= *this_best_prio) ||
2530 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2531 p = iterator->next(iterator->arg);
2535 pull_task(busiest, p, this_rq, this_cpu);
2537 rem_load_move -= p->se.load.weight;
2540 * We only want to steal up to the prescribed amount of weighted load.
2542 if (rem_load_move > 0) {
2543 if (p->prio < *this_best_prio)
2544 *this_best_prio = p->prio;
2545 p = iterator->next(iterator->arg);
2550 * Right now, this is one of only two places pull_task() is called,
2551 * so we can safely collect pull_task() stats here rather than
2552 * inside pull_task().
2554 schedstat_add(sd, lb_gained[idle], pulled);
2557 *all_pinned = pinned;
2559 return max_load_move - rem_load_move;
2563 * move_tasks tries to move up to max_load_move weighted load from busiest to
2564 * this_rq, as part of a balancing operation within domain "sd".
2565 * Returns 1 if successful and 0 otherwise.
2567 * Called with both runqueues locked.
2569 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2570 unsigned long max_load_move,
2571 struct sched_domain *sd, enum cpu_idle_type idle,
2574 const struct sched_class *class = sched_class_highest;
2575 unsigned long total_load_moved = 0;
2576 int this_best_prio = this_rq->curr->prio;
2580 class->load_balance(this_rq, this_cpu, busiest,
2581 max_load_move - total_load_moved,
2582 sd, idle, all_pinned, &this_best_prio);
2583 class = class->next;
2584 } while (class && max_load_move > total_load_moved);
2586 return total_load_moved > 0;
2590 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2591 struct sched_domain *sd, enum cpu_idle_type idle,
2592 struct rq_iterator *iterator)
2594 struct task_struct *p = iterator->start(iterator->arg);
2598 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2599 pull_task(busiest, p, this_rq, this_cpu);
2601 * Right now, this is only the second place pull_task()
2602 * is called, so we can safely collect pull_task()
2603 * stats here rather than inside pull_task().
2605 schedstat_inc(sd, lb_gained[idle]);
2609 p = iterator->next(iterator->arg);
2616 * move_one_task tries to move exactly one task from busiest to this_rq, as
2617 * part of active balancing operations within "domain".
2618 * Returns 1 if successful and 0 otherwise.
2620 * Called with both runqueues locked.
2622 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2623 struct sched_domain *sd, enum cpu_idle_type idle)
2625 const struct sched_class *class;
2627 for (class = sched_class_highest; class; class = class->next)
2628 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2635 * find_busiest_group finds and returns the busiest CPU group within the
2636 * domain. It calculates and returns the amount of weighted load which
2637 * should be moved to restore balance via the imbalance parameter.
2639 static struct sched_group *
2640 find_busiest_group(struct sched_domain *sd, int this_cpu,
2641 unsigned long *imbalance, enum cpu_idle_type idle,
2642 int *sd_idle, cpumask_t *cpus, int *balance)
2644 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2645 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2646 unsigned long max_pull;
2647 unsigned long busiest_load_per_task, busiest_nr_running;
2648 unsigned long this_load_per_task, this_nr_running;
2649 int load_idx, group_imb = 0;
2650 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2651 int power_savings_balance = 1;
2652 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2653 unsigned long min_nr_running = ULONG_MAX;
2654 struct sched_group *group_min = NULL, *group_leader = NULL;
2657 max_load = this_load = total_load = total_pwr = 0;
2658 busiest_load_per_task = busiest_nr_running = 0;
2659 this_load_per_task = this_nr_running = 0;
2660 if (idle == CPU_NOT_IDLE)
2661 load_idx = sd->busy_idx;
2662 else if (idle == CPU_NEWLY_IDLE)
2663 load_idx = sd->newidle_idx;
2665 load_idx = sd->idle_idx;
2668 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2671 int __group_imb = 0;
2672 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2673 unsigned long sum_nr_running, sum_weighted_load;
2675 local_group = cpu_isset(this_cpu, group->cpumask);
2678 balance_cpu = first_cpu(group->cpumask);
2680 /* Tally up the load of all CPUs in the group */
2681 sum_weighted_load = sum_nr_running = avg_load = 0;
2683 min_cpu_load = ~0UL;
2685 for_each_cpu_mask(i, group->cpumask) {
2688 if (!cpu_isset(i, *cpus))
2693 if (*sd_idle && rq->nr_running)
2696 /* Bias balancing toward cpus of our domain */
2698 if (idle_cpu(i) && !first_idle_cpu) {
2703 load = target_load(i, load_idx);
2705 load = source_load(i, load_idx);
2706 if (load > max_cpu_load)
2707 max_cpu_load = load;
2708 if (min_cpu_load > load)
2709 min_cpu_load = load;
2713 sum_nr_running += rq->nr_running;
2714 sum_weighted_load += weighted_cpuload(i);
2718 * First idle cpu or the first cpu(busiest) in this sched group
2719 * is eligible for doing load balancing at this and above
2720 * domains. In the newly idle case, we will allow all the cpu's
2721 * to do the newly idle load balance.
2723 if (idle != CPU_NEWLY_IDLE && local_group &&
2724 balance_cpu != this_cpu && balance) {
2729 total_load += avg_load;
2730 total_pwr += group->__cpu_power;
2732 /* Adjust by relative CPU power of the group */
2733 avg_load = sg_div_cpu_power(group,
2734 avg_load * SCHED_LOAD_SCALE);
2736 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2739 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2742 this_load = avg_load;
2744 this_nr_running = sum_nr_running;
2745 this_load_per_task = sum_weighted_load;
2746 } else if (avg_load > max_load &&
2747 (sum_nr_running > group_capacity || __group_imb)) {
2748 max_load = avg_load;
2750 busiest_nr_running = sum_nr_running;
2751 busiest_load_per_task = sum_weighted_load;
2752 group_imb = __group_imb;
2755 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2757 * Busy processors will not participate in power savings
2760 if (idle == CPU_NOT_IDLE ||
2761 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2765 * If the local group is idle or completely loaded
2766 * no need to do power savings balance at this domain
2768 if (local_group && (this_nr_running >= group_capacity ||
2770 power_savings_balance = 0;
2773 * If a group is already running at full capacity or idle,
2774 * don't include that group in power savings calculations
2776 if (!power_savings_balance || sum_nr_running >= group_capacity
2781 * Calculate the group which has the least non-idle load.
2782 * This is the group from where we need to pick up the load
2785 if ((sum_nr_running < min_nr_running) ||
2786 (sum_nr_running == min_nr_running &&
2787 first_cpu(group->cpumask) <
2788 first_cpu(group_min->cpumask))) {
2790 min_nr_running = sum_nr_running;
2791 min_load_per_task = sum_weighted_load /
2796 * Calculate the group which is almost near its
2797 * capacity but still has some space to pick up some load
2798 * from other group and save more power
2800 if (sum_nr_running <= group_capacity - 1) {
2801 if (sum_nr_running > leader_nr_running ||
2802 (sum_nr_running == leader_nr_running &&
2803 first_cpu(group->cpumask) >
2804 first_cpu(group_leader->cpumask))) {
2805 group_leader = group;
2806 leader_nr_running = sum_nr_running;
2811 group = group->next;
2812 } while (group != sd->groups);
2814 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2817 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2819 if (this_load >= avg_load ||
2820 100*max_load <= sd->imbalance_pct*this_load)
2823 busiest_load_per_task /= busiest_nr_running;
2825 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2828 * We're trying to get all the cpus to the average_load, so we don't
2829 * want to push ourselves above the average load, nor do we wish to
2830 * reduce the max loaded cpu below the average load, as either of these
2831 * actions would just result in more rebalancing later, and ping-pong
2832 * tasks around. Thus we look for the minimum possible imbalance.
2833 * Negative imbalances (*we* are more loaded than anyone else) will
2834 * be counted as no imbalance for these purposes -- we can't fix that
2835 * by pulling tasks to us. Be careful of negative numbers as they'll
2836 * appear as very large values with unsigned longs.
2838 if (max_load <= busiest_load_per_task)
2842 * In the presence of smp nice balancing, certain scenarios can have
2843 * max load less than avg load(as we skip the groups at or below
2844 * its cpu_power, while calculating max_load..)
2846 if (max_load < avg_load) {
2848 goto small_imbalance;
2851 /* Don't want to pull so many tasks that a group would go idle */
2852 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2854 /* How much load to actually move to equalise the imbalance */
2855 *imbalance = min(max_pull * busiest->__cpu_power,
2856 (avg_load - this_load) * this->__cpu_power)
2860 * if *imbalance is less than the average load per runnable task
2861 * there is no gaurantee that any tasks will be moved so we'll have
2862 * a think about bumping its value to force at least one task to be
2865 if (*imbalance < busiest_load_per_task) {
2866 unsigned long tmp, pwr_now, pwr_move;
2870 pwr_move = pwr_now = 0;
2872 if (this_nr_running) {
2873 this_load_per_task /= this_nr_running;
2874 if (busiest_load_per_task > this_load_per_task)
2877 this_load_per_task = SCHED_LOAD_SCALE;
2879 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2880 busiest_load_per_task * imbn) {
2881 *imbalance = busiest_load_per_task;
2886 * OK, we don't have enough imbalance to justify moving tasks,
2887 * however we may be able to increase total CPU power used by
2891 pwr_now += busiest->__cpu_power *
2892 min(busiest_load_per_task, max_load);
2893 pwr_now += this->__cpu_power *
2894 min(this_load_per_task, this_load);
2895 pwr_now /= SCHED_LOAD_SCALE;
2897 /* Amount of load we'd subtract */
2898 tmp = sg_div_cpu_power(busiest,
2899 busiest_load_per_task * SCHED_LOAD_SCALE);
2901 pwr_move += busiest->__cpu_power *
2902 min(busiest_load_per_task, max_load - tmp);
2904 /* Amount of load we'd add */
2905 if (max_load * busiest->__cpu_power <
2906 busiest_load_per_task * SCHED_LOAD_SCALE)
2907 tmp = sg_div_cpu_power(this,
2908 max_load * busiest->__cpu_power);
2910 tmp = sg_div_cpu_power(this,
2911 busiest_load_per_task * SCHED_LOAD_SCALE);
2912 pwr_move += this->__cpu_power *
2913 min(this_load_per_task, this_load + tmp);
2914 pwr_move /= SCHED_LOAD_SCALE;
2916 /* Move if we gain throughput */
2917 if (pwr_move > pwr_now)
2918 *imbalance = busiest_load_per_task;
2924 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2925 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2928 if (this == group_leader && group_leader != group_min) {
2929 *imbalance = min_load_per_task;
2939 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2942 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2943 unsigned long imbalance, cpumask_t *cpus)
2945 struct rq *busiest = NULL, *rq;
2946 unsigned long max_load = 0;
2949 for_each_cpu_mask(i, group->cpumask) {
2952 if (!cpu_isset(i, *cpus))
2956 wl = weighted_cpuload(i);
2958 if (rq->nr_running == 1 && wl > imbalance)
2961 if (wl > max_load) {
2971 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2972 * so long as it is large enough.
2974 #define MAX_PINNED_INTERVAL 512
2977 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2978 * tasks if there is an imbalance.
2980 static int load_balance(int this_cpu, struct rq *this_rq,
2981 struct sched_domain *sd, enum cpu_idle_type idle,
2984 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2985 struct sched_group *group;
2986 unsigned long imbalance;
2988 cpumask_t cpus = CPU_MASK_ALL;
2989 unsigned long flags;
2992 * When power savings policy is enabled for the parent domain, idle
2993 * sibling can pick up load irrespective of busy siblings. In this case,
2994 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2995 * portraying it as CPU_NOT_IDLE.
2997 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2998 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3001 schedstat_inc(sd, lb_count[idle]);
3004 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3011 schedstat_inc(sd, lb_nobusyg[idle]);
3015 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
3017 schedstat_inc(sd, lb_nobusyq[idle]);
3021 BUG_ON(busiest == this_rq);
3023 schedstat_add(sd, lb_imbalance[idle], imbalance);
3026 if (busiest->nr_running > 1) {
3028 * Attempt to move tasks. If find_busiest_group has found
3029 * an imbalance but busiest->nr_running <= 1, the group is
3030 * still unbalanced. ld_moved simply stays zero, so it is
3031 * correctly treated as an imbalance.
3033 local_irq_save(flags);
3034 double_rq_lock(this_rq, busiest);
3035 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3036 imbalance, sd, idle, &all_pinned);
3037 double_rq_unlock(this_rq, busiest);
3038 local_irq_restore(flags);
3041 * some other cpu did the load balance for us.
3043 if (ld_moved && this_cpu != smp_processor_id())
3044 resched_cpu(this_cpu);
3046 /* All tasks on this runqueue were pinned by CPU affinity */
3047 if (unlikely(all_pinned)) {
3048 cpu_clear(cpu_of(busiest), cpus);
3049 if (!cpus_empty(cpus))
3056 schedstat_inc(sd, lb_failed[idle]);
3057 sd->nr_balance_failed++;
3059 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3061 spin_lock_irqsave(&busiest->lock, flags);
3063 /* don't kick the migration_thread, if the curr
3064 * task on busiest cpu can't be moved to this_cpu
3066 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3067 spin_unlock_irqrestore(&busiest->lock, flags);
3069 goto out_one_pinned;
3072 if (!busiest->active_balance) {
3073 busiest->active_balance = 1;
3074 busiest->push_cpu = this_cpu;
3077 spin_unlock_irqrestore(&busiest->lock, flags);
3079 wake_up_process(busiest->migration_thread);
3082 * We've kicked active balancing, reset the failure
3085 sd->nr_balance_failed = sd->cache_nice_tries+1;
3088 sd->nr_balance_failed = 0;
3090 if (likely(!active_balance)) {
3091 /* We were unbalanced, so reset the balancing interval */
3092 sd->balance_interval = sd->min_interval;
3095 * If we've begun active balancing, start to back off. This
3096 * case may not be covered by the all_pinned logic if there
3097 * is only 1 task on the busy runqueue (because we don't call
3100 if (sd->balance_interval < sd->max_interval)
3101 sd->balance_interval *= 2;
3104 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3105 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3110 schedstat_inc(sd, lb_balanced[idle]);
3112 sd->nr_balance_failed = 0;
3115 /* tune up the balancing interval */
3116 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3117 (sd->balance_interval < sd->max_interval))
3118 sd->balance_interval *= 2;
3120 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3121 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3127 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3128 * tasks if there is an imbalance.
3130 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3131 * this_rq is locked.
3134 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3136 struct sched_group *group;
3137 struct rq *busiest = NULL;
3138 unsigned long imbalance;
3142 cpumask_t cpus = CPU_MASK_ALL;
3145 * When power savings policy is enabled for the parent domain, idle
3146 * sibling can pick up load irrespective of busy siblings. In this case,
3147 * let the state of idle sibling percolate up as IDLE, instead of
3148 * portraying it as CPU_NOT_IDLE.
3150 if (sd->flags & SD_SHARE_CPUPOWER &&
3151 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3154 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3156 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3157 &sd_idle, &cpus, NULL);
3159 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3163 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
3166 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3170 BUG_ON(busiest == this_rq);
3172 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3175 if (busiest->nr_running > 1) {
3176 /* Attempt to move tasks */
3177 double_lock_balance(this_rq, busiest);
3178 /* this_rq->clock is already updated */
3179 update_rq_clock(busiest);
3180 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3181 imbalance, sd, CPU_NEWLY_IDLE,
3183 spin_unlock(&busiest->lock);
3185 if (unlikely(all_pinned)) {
3186 cpu_clear(cpu_of(busiest), cpus);
3187 if (!cpus_empty(cpus))
3193 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3194 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3195 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3198 sd->nr_balance_failed = 0;
3203 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3204 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3205 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3207 sd->nr_balance_failed = 0;
3213 * idle_balance is called by schedule() if this_cpu is about to become
3214 * idle. Attempts to pull tasks from other CPUs.
3216 static void idle_balance(int this_cpu, struct rq *this_rq)
3218 struct sched_domain *sd;
3219 int pulled_task = -1;
3220 unsigned long next_balance = jiffies + HZ;
3222 for_each_domain(this_cpu, sd) {
3223 unsigned long interval;
3225 if (!(sd->flags & SD_LOAD_BALANCE))
3228 if (sd->flags & SD_BALANCE_NEWIDLE)
3229 /* If we've pulled tasks over stop searching: */
3230 pulled_task = load_balance_newidle(this_cpu,
3233 interval = msecs_to_jiffies(sd->balance_interval);
3234 if (time_after(next_balance, sd->last_balance + interval))
3235 next_balance = sd->last_balance + interval;
3239 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3241 * We are going idle. next_balance may be set based on
3242 * a busy processor. So reset next_balance.
3244 this_rq->next_balance = next_balance;
3249 * active_load_balance is run by migration threads. It pushes running tasks
3250 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3251 * running on each physical CPU where possible, and avoids physical /
3252 * logical imbalances.
3254 * Called with busiest_rq locked.
3256 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3258 int target_cpu = busiest_rq->push_cpu;
3259 struct sched_domain *sd;
3260 struct rq *target_rq;
3262 /* Is there any task to move? */
3263 if (busiest_rq->nr_running <= 1)
3266 target_rq = cpu_rq(target_cpu);
3269 * This condition is "impossible", if it occurs
3270 * we need to fix it. Originally reported by
3271 * Bjorn Helgaas on a 128-cpu setup.
3273 BUG_ON(busiest_rq == target_rq);
3275 /* move a task from busiest_rq to target_rq */
3276 double_lock_balance(busiest_rq, target_rq);
3277 update_rq_clock(busiest_rq);
3278 update_rq_clock(target_rq);
3280 /* Search for an sd spanning us and the target CPU. */
3281 for_each_domain(target_cpu, sd) {
3282 if ((sd->flags & SD_LOAD_BALANCE) &&
3283 cpu_isset(busiest_cpu, sd->span))
3288 schedstat_inc(sd, alb_count);
3290 if (move_one_task(target_rq, target_cpu, busiest_rq,
3292 schedstat_inc(sd, alb_pushed);
3294 schedstat_inc(sd, alb_failed);
3296 spin_unlock(&target_rq->lock);
3301 atomic_t load_balancer;
3303 } nohz ____cacheline_aligned = {
3304 .load_balancer = ATOMIC_INIT(-1),
3305 .cpu_mask = CPU_MASK_NONE,
3309 * This routine will try to nominate the ilb (idle load balancing)
3310 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3311 * load balancing on behalf of all those cpus. If all the cpus in the system
3312 * go into this tickless mode, then there will be no ilb owner (as there is
3313 * no need for one) and all the cpus will sleep till the next wakeup event
3316 * For the ilb owner, tick is not stopped. And this tick will be used
3317 * for idle load balancing. ilb owner will still be part of
3320 * While stopping the tick, this cpu will become the ilb owner if there
3321 * is no other owner. And will be the owner till that cpu becomes busy
3322 * or if all cpus in the system stop their ticks at which point
3323 * there is no need for ilb owner.
3325 * When the ilb owner becomes busy, it nominates another owner, during the
3326 * next busy scheduler_tick()
3328 int select_nohz_load_balancer(int stop_tick)
3330 int cpu = smp_processor_id();
3333 cpu_set(cpu, nohz.cpu_mask);
3334 cpu_rq(cpu)->in_nohz_recently = 1;
3337 * If we are going offline and still the leader, give up!
3339 if (cpu_is_offline(cpu) &&
3340 atomic_read(&nohz.load_balancer) == cpu) {
3341 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3346 /* time for ilb owner also to sleep */
3347 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3348 if (atomic_read(&nohz.load_balancer) == cpu)
3349 atomic_set(&nohz.load_balancer, -1);
3353 if (atomic_read(&nohz.load_balancer) == -1) {
3354 /* make me the ilb owner */
3355 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3357 } else if (atomic_read(&nohz.load_balancer) == cpu)
3360 if (!cpu_isset(cpu, nohz.cpu_mask))
3363 cpu_clear(cpu, nohz.cpu_mask);
3365 if (atomic_read(&nohz.load_balancer) == cpu)
3366 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3373 static DEFINE_SPINLOCK(balancing);
3376 * It checks each scheduling domain to see if it is due to be balanced,
3377 * and initiates a balancing operation if so.
3379 * Balancing parameters are set up in arch_init_sched_domains.
3381 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3384 struct rq *rq = cpu_rq(cpu);
3385 unsigned long interval;
3386 struct sched_domain *sd;
3387 /* Earliest time when we have to do rebalance again */
3388 unsigned long next_balance = jiffies + 60*HZ;
3389 int update_next_balance = 0;
3391 for_each_domain(cpu, sd) {
3392 if (!(sd->flags & SD_LOAD_BALANCE))
3395 interval = sd->balance_interval;
3396 if (idle != CPU_IDLE)
3397 interval *= sd->busy_factor;
3399 /* scale ms to jiffies */
3400 interval = msecs_to_jiffies(interval);
3401 if (unlikely(!interval))
3403 if (interval > HZ*NR_CPUS/10)
3404 interval = HZ*NR_CPUS/10;
3407 if (sd->flags & SD_SERIALIZE) {
3408 if (!spin_trylock(&balancing))
3412 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3413 if (load_balance(cpu, rq, sd, idle, &balance)) {
3415 * We've pulled tasks over so either we're no
3416 * longer idle, or one of our SMT siblings is
3419 idle = CPU_NOT_IDLE;
3421 sd->last_balance = jiffies;
3423 if (sd->flags & SD_SERIALIZE)
3424 spin_unlock(&balancing);
3426 if (time_after(next_balance, sd->last_balance + interval)) {
3427 next_balance = sd->last_balance + interval;
3428 update_next_balance = 1;
3432 * Stop the load balance at this level. There is another
3433 * CPU in our sched group which is doing load balancing more
3441 * next_balance will be updated only when there is a need.
3442 * When the cpu is attached to null domain for ex, it will not be
3445 if (likely(update_next_balance))
3446 rq->next_balance = next_balance;
3450 * run_rebalance_domains is triggered when needed from the scheduler tick.
3451 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3452 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3454 static void run_rebalance_domains(struct softirq_action *h)
3456 int this_cpu = smp_processor_id();
3457 struct rq *this_rq = cpu_rq(this_cpu);
3458 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3459 CPU_IDLE : CPU_NOT_IDLE;
3461 rebalance_domains(this_cpu, idle);
3465 * If this cpu is the owner for idle load balancing, then do the
3466 * balancing on behalf of the other idle cpus whose ticks are
3469 if (this_rq->idle_at_tick &&
3470 atomic_read(&nohz.load_balancer) == this_cpu) {
3471 cpumask_t cpus = nohz.cpu_mask;
3475 cpu_clear(this_cpu, cpus);
3476 for_each_cpu_mask(balance_cpu, cpus) {
3478 * If this cpu gets work to do, stop the load balancing
3479 * work being done for other cpus. Next load
3480 * balancing owner will pick it up.
3485 rebalance_domains(balance_cpu, CPU_IDLE);
3487 rq = cpu_rq(balance_cpu);
3488 if (time_after(this_rq->next_balance, rq->next_balance))
3489 this_rq->next_balance = rq->next_balance;
3496 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3498 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3499 * idle load balancing owner or decide to stop the periodic load balancing,
3500 * if the whole system is idle.
3502 static inline void trigger_load_balance(struct rq *rq, int cpu)
3506 * If we were in the nohz mode recently and busy at the current
3507 * scheduler tick, then check if we need to nominate new idle
3510 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3511 rq->in_nohz_recently = 0;
3513 if (atomic_read(&nohz.load_balancer) == cpu) {
3514 cpu_clear(cpu, nohz.cpu_mask);
3515 atomic_set(&nohz.load_balancer, -1);
3518 if (atomic_read(&nohz.load_balancer) == -1) {
3520 * simple selection for now: Nominate the
3521 * first cpu in the nohz list to be the next
3524 * TBD: Traverse the sched domains and nominate
3525 * the nearest cpu in the nohz.cpu_mask.
3527 int ilb = first_cpu(nohz.cpu_mask);
3535 * If this cpu is idle and doing idle load balancing for all the
3536 * cpus with ticks stopped, is it time for that to stop?
3538 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3539 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3545 * If this cpu is idle and the idle load balancing is done by
3546 * someone else, then no need raise the SCHED_SOFTIRQ
3548 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3549 cpu_isset(cpu, nohz.cpu_mask))
3552 if (time_after_eq(jiffies, rq->next_balance))
3553 raise_softirq(SCHED_SOFTIRQ);
3556 #else /* CONFIG_SMP */
3559 * on UP we do not need to balance between CPUs:
3561 static inline void idle_balance(int cpu, struct rq *rq)
3567 DEFINE_PER_CPU(struct kernel_stat, kstat);
3569 EXPORT_PER_CPU_SYMBOL(kstat);
3572 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3573 * that have not yet been banked in case the task is currently running.
3575 unsigned long long task_sched_runtime(struct task_struct *p)
3577 unsigned long flags;
3581 rq = task_rq_lock(p, &flags);
3582 ns = p->se.sum_exec_runtime;
3583 if (task_current(rq, p)) {
3584 update_rq_clock(rq);
3585 delta_exec = rq->clock - p->se.exec_start;
3586 if ((s64)delta_exec > 0)
3589 task_rq_unlock(rq, &flags);
3595 * Account user cpu time to a process.
3596 * @p: the process that the cpu time gets accounted to
3597 * @cputime: the cpu time spent in user space since the last update
3599 void account_user_time(struct task_struct *p, cputime_t cputime)
3601 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3604 p->utime = cputime_add(p->utime, cputime);
3606 /* Add user time to cpustat. */
3607 tmp = cputime_to_cputime64(cputime);
3608 if (TASK_NICE(p) > 0)
3609 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3611 cpustat->user = cputime64_add(cpustat->user, tmp);
3615 * Account guest cpu time to a process.
3616 * @p: the process that the cpu time gets accounted to
3617 * @cputime: the cpu time spent in virtual machine since the last update
3619 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3622 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3624 tmp = cputime_to_cputime64(cputime);
3626 p->utime = cputime_add(p->utime, cputime);
3627 p->gtime = cputime_add(p->gtime, cputime);
3629 cpustat->user = cputime64_add(cpustat->user, tmp);
3630 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3634 * Account scaled user cpu time to a process.
3635 * @p: the process that the cpu time gets accounted to
3636 * @cputime: the cpu time spent in user space since the last update
3638 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3640 p->utimescaled = cputime_add(p->utimescaled, cputime);
3644 * Account system cpu time to a process.
3645 * @p: the process that the cpu time gets accounted to
3646 * @hardirq_offset: the offset to subtract from hardirq_count()
3647 * @cputime: the cpu time spent in kernel space since the last update
3649 void account_system_time(struct task_struct *p, int hardirq_offset,
3652 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3653 struct rq *rq = this_rq();
3656 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3657 return account_guest_time(p, cputime);
3659 p->stime = cputime_add(p->stime, cputime);
3661 /* Add system time to cpustat. */
3662 tmp = cputime_to_cputime64(cputime);
3663 if (hardirq_count() - hardirq_offset)
3664 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3665 else if (softirq_count())
3666 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3667 else if (p != rq->idle)
3668 cpustat->system = cputime64_add(cpustat->system, tmp);
3669 else if (atomic_read(&rq->nr_iowait) > 0)
3670 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3672 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3673 /* Account for system time used */
3674 acct_update_integrals(p);
3678 * Account scaled system cpu time to a process.
3679 * @p: the process that the cpu time gets accounted to
3680 * @hardirq_offset: the offset to subtract from hardirq_count()
3681 * @cputime: the cpu time spent in kernel space since the last update
3683 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3685 p->stimescaled = cputime_add(p->stimescaled, cputime);
3689 * Account for involuntary wait time.
3690 * @p: the process from which the cpu time has been stolen
3691 * @steal: the cpu time spent in involuntary wait
3693 void account_steal_time(struct task_struct *p, cputime_t steal)
3695 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3696 cputime64_t tmp = cputime_to_cputime64(steal);
3697 struct rq *rq = this_rq();
3699 if (p == rq->idle) {
3700 p->stime = cputime_add(p->stime, steal);
3701 if (atomic_read(&rq->nr_iowait) > 0)
3702 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3704 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3706 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3710 * This function gets called by the timer code, with HZ frequency.
3711 * We call it with interrupts disabled.
3713 * It also gets called by the fork code, when changing the parent's
3716 void scheduler_tick(void)
3718 int cpu = smp_processor_id();
3719 struct rq *rq = cpu_rq(cpu);
3720 struct task_struct *curr = rq->curr;
3721 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3723 spin_lock(&rq->lock);
3724 __update_rq_clock(rq);
3726 * Let rq->clock advance by at least TICK_NSEC:
3728 if (unlikely(rq->clock < next_tick)) {
3729 rq->clock = next_tick;
3730 rq->clock_underflows++;
3732 rq->tick_timestamp = rq->clock;
3733 update_cpu_load(rq);
3734 curr->sched_class->task_tick(rq, curr, 0);
3735 update_sched_rt_period(rq);
3736 spin_unlock(&rq->lock);
3739 rq->idle_at_tick = idle_cpu(cpu);
3740 trigger_load_balance(rq, cpu);
3744 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3746 void add_preempt_count(int val)
3751 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3753 preempt_count() += val;
3755 * Spinlock count overflowing soon?
3757 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3760 EXPORT_SYMBOL(add_preempt_count);
3762 void sub_preempt_count(int val)
3767 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3770 * Is the spinlock portion underflowing?
3772 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3773 !(preempt_count() & PREEMPT_MASK)))
3776 preempt_count() -= val;
3778 EXPORT_SYMBOL(sub_preempt_count);
3783 * Print scheduling while atomic bug:
3785 static noinline void __schedule_bug(struct task_struct *prev)
3787 struct pt_regs *regs = get_irq_regs();
3789 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3790 prev->comm, prev->pid, preempt_count());
3792 debug_show_held_locks(prev);
3793 if (irqs_disabled())
3794 print_irqtrace_events(prev);
3803 * Various schedule()-time debugging checks and statistics:
3805 static inline void schedule_debug(struct task_struct *prev)
3808 * Test if we are atomic. Since do_exit() needs to call into
3809 * schedule() atomically, we ignore that path for now.
3810 * Otherwise, whine if we are scheduling when we should not be.
3812 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3813 __schedule_bug(prev);
3815 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3817 schedstat_inc(this_rq(), sched_count);
3818 #ifdef CONFIG_SCHEDSTATS
3819 if (unlikely(prev->lock_depth >= 0)) {
3820 schedstat_inc(this_rq(), bkl_count);
3821 schedstat_inc(prev, sched_info.bkl_count);
3827 * Pick up the highest-prio task:
3829 static inline struct task_struct *
3830 pick_next_task(struct rq *rq, struct task_struct *prev)
3832 const struct sched_class *class;
3833 struct task_struct *p;
3836 * Optimization: we know that if all tasks are in
3837 * the fair class we can call that function directly:
3839 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3840 p = fair_sched_class.pick_next_task(rq);
3845 class = sched_class_highest;
3847 p = class->pick_next_task(rq);
3851 * Will never be NULL as the idle class always
3852 * returns a non-NULL p:
3854 class = class->next;
3859 * schedule() is the main scheduler function.
3861 asmlinkage void __sched schedule(void)
3863 struct task_struct *prev, *next;
3870 cpu = smp_processor_id();
3874 switch_count = &prev->nivcsw;
3876 release_kernel_lock(prev);
3877 need_resched_nonpreemptible:
3879 schedule_debug(prev);
3884 * Do the rq-clock update outside the rq lock:
3886 local_irq_disable();
3887 __update_rq_clock(rq);
3888 spin_lock(&rq->lock);
3889 clear_tsk_need_resched(prev);
3891 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3892 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3893 unlikely(signal_pending(prev)))) {
3894 prev->state = TASK_RUNNING;
3896 deactivate_task(rq, prev, 1);
3898 switch_count = &prev->nvcsw;
3902 if (prev->sched_class->pre_schedule)
3903 prev->sched_class->pre_schedule(rq, prev);
3906 if (unlikely(!rq->nr_running))
3907 idle_balance(cpu, rq);
3909 prev->sched_class->put_prev_task(rq, prev);
3910 next = pick_next_task(rq, prev);
3912 sched_info_switch(prev, next);
3914 if (likely(prev != next)) {
3919 context_switch(rq, prev, next); /* unlocks the rq */
3921 * the context switch might have flipped the stack from under
3922 * us, hence refresh the local variables.
3924 cpu = smp_processor_id();
3927 spin_unlock_irq(&rq->lock);
3931 if (unlikely(reacquire_kernel_lock(current) < 0))
3932 goto need_resched_nonpreemptible;
3934 preempt_enable_no_resched();
3935 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3938 EXPORT_SYMBOL(schedule);
3940 #ifdef CONFIG_PREEMPT
3942 * this is the entry point to schedule() from in-kernel preemption
3943 * off of preempt_enable. Kernel preemptions off return from interrupt
3944 * occur there and call schedule directly.
3946 asmlinkage void __sched preempt_schedule(void)
3948 struct thread_info *ti = current_thread_info();
3949 struct task_struct *task = current;
3950 int saved_lock_depth;
3953 * If there is a non-zero preempt_count or interrupts are disabled,
3954 * we do not want to preempt the current task. Just return..
3956 if (likely(ti->preempt_count || irqs_disabled()))
3960 add_preempt_count(PREEMPT_ACTIVE);
3963 * We keep the big kernel semaphore locked, but we
3964 * clear ->lock_depth so that schedule() doesnt
3965 * auto-release the semaphore:
3967 saved_lock_depth = task->lock_depth;
3968 task->lock_depth = -1;
3970 task->lock_depth = saved_lock_depth;
3971 sub_preempt_count(PREEMPT_ACTIVE);
3974 * Check again in case we missed a preemption opportunity
3975 * between schedule and now.
3978 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3980 EXPORT_SYMBOL(preempt_schedule);
3983 * this is the entry point to schedule() from kernel preemption
3984 * off of irq context.
3985 * Note, that this is called and return with irqs disabled. This will
3986 * protect us against recursive calling from irq.
3988 asmlinkage void __sched preempt_schedule_irq(void)
3990 struct thread_info *ti = current_thread_info();
3991 struct task_struct *task = current;
3992 int saved_lock_depth;
3994 /* Catch callers which need to be fixed */
3995 BUG_ON(ti->preempt_count || !irqs_disabled());
3998 add_preempt_count(PREEMPT_ACTIVE);
4001 * We keep the big kernel semaphore locked, but we
4002 * clear ->lock_depth so that schedule() doesnt
4003 * auto-release the semaphore:
4005 saved_lock_depth = task->lock_depth;
4006 task->lock_depth = -1;
4009 local_irq_disable();
4010 task->lock_depth = saved_lock_depth;
4011 sub_preempt_count(PREEMPT_ACTIVE);
4014 * Check again in case we missed a preemption opportunity
4015 * between schedule and now.
4018 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4021 #endif /* CONFIG_PREEMPT */
4023 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4026 return try_to_wake_up(curr->private, mode, sync);
4028 EXPORT_SYMBOL(default_wake_function);
4031 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4032 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4033 * number) then we wake all the non-exclusive tasks and one exclusive task.
4035 * There are circumstances in which we can try to wake a task which has already
4036 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4037 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4039 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4040 int nr_exclusive, int sync, void *key)
4042 wait_queue_t *curr, *next;
4044 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4045 unsigned flags = curr->flags;
4047 if (curr->func(curr, mode, sync, key) &&
4048 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4054 * __wake_up - wake up threads blocked on a waitqueue.
4056 * @mode: which threads
4057 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4058 * @key: is directly passed to the wakeup function
4060 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4061 int nr_exclusive, void *key)
4063 unsigned long flags;
4065 spin_lock_irqsave(&q->lock, flags);
4066 __wake_up_common(q, mode, nr_exclusive, 0, key);
4067 spin_unlock_irqrestore(&q->lock, flags);
4069 EXPORT_SYMBOL(__wake_up);
4072 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4074 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4076 __wake_up_common(q, mode, 1, 0, NULL);
4080 * __wake_up_sync - wake up threads blocked on a waitqueue.
4082 * @mode: which threads
4083 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4085 * The sync wakeup differs that the waker knows that it will schedule
4086 * away soon, so while the target thread will be woken up, it will not
4087 * be migrated to another CPU - ie. the two threads are 'synchronized'
4088 * with each other. This can prevent needless bouncing between CPUs.
4090 * On UP it can prevent extra preemption.
4093 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4095 unsigned long flags;
4101 if (unlikely(!nr_exclusive))
4104 spin_lock_irqsave(&q->lock, flags);
4105 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4106 spin_unlock_irqrestore(&q->lock, flags);
4108 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4110 void complete(struct completion *x)
4112 unsigned long flags;
4114 spin_lock_irqsave(&x->wait.lock, flags);
4116 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4117 spin_unlock_irqrestore(&x->wait.lock, flags);
4119 EXPORT_SYMBOL(complete);
4121 void complete_all(struct completion *x)
4123 unsigned long flags;
4125 spin_lock_irqsave(&x->wait.lock, flags);
4126 x->done += UINT_MAX/2;
4127 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4128 spin_unlock_irqrestore(&x->wait.lock, flags);
4130 EXPORT_SYMBOL(complete_all);
4132 static inline long __sched
4133 do_wait_for_common(struct completion *x, long timeout, int state)
4136 DECLARE_WAITQUEUE(wait, current);
4138 wait.flags |= WQ_FLAG_EXCLUSIVE;
4139 __add_wait_queue_tail(&x->wait, &wait);
4141 if ((state == TASK_INTERRUPTIBLE &&
4142 signal_pending(current)) ||
4143 (state == TASK_KILLABLE &&
4144 fatal_signal_pending(current))) {
4145 __remove_wait_queue(&x->wait, &wait);
4146 return -ERESTARTSYS;
4148 __set_current_state(state);
4149 spin_unlock_irq(&x->wait.lock);
4150 timeout = schedule_timeout(timeout);
4151 spin_lock_irq(&x->wait.lock);
4153 __remove_wait_queue(&x->wait, &wait);
4157 __remove_wait_queue(&x->wait, &wait);
4164 wait_for_common(struct completion *x, long timeout, int state)
4168 spin_lock_irq(&x->wait.lock);
4169 timeout = do_wait_for_common(x, timeout, state);
4170 spin_unlock_irq(&x->wait.lock);
4174 void __sched wait_for_completion(struct completion *x)
4176 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4178 EXPORT_SYMBOL(wait_for_completion);
4180 unsigned long __sched
4181 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4183 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4185 EXPORT_SYMBOL(wait_for_completion_timeout);
4187 int __sched wait_for_completion_interruptible(struct completion *x)
4189 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4190 if (t == -ERESTARTSYS)
4194 EXPORT_SYMBOL(wait_for_completion_interruptible);
4196 unsigned long __sched
4197 wait_for_completion_interruptible_timeout(struct completion *x,
4198 unsigned long timeout)
4200 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4202 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4204 int __sched wait_for_completion_killable(struct completion *x)
4206 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4207 if (t == -ERESTARTSYS)
4211 EXPORT_SYMBOL(wait_for_completion_killable);
4214 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4216 unsigned long flags;
4219 init_waitqueue_entry(&wait, current);
4221 __set_current_state(state);
4223 spin_lock_irqsave(&q->lock, flags);
4224 __add_wait_queue(q, &wait);
4225 spin_unlock(&q->lock);
4226 timeout = schedule_timeout(timeout);
4227 spin_lock_irq(&q->lock);
4228 __remove_wait_queue(q, &wait);
4229 spin_unlock_irqrestore(&q->lock, flags);
4234 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4236 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4238 EXPORT_SYMBOL(interruptible_sleep_on);
4241 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4243 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4245 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4247 void __sched sleep_on(wait_queue_head_t *q)
4249 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4251 EXPORT_SYMBOL(sleep_on);
4253 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4255 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4257 EXPORT_SYMBOL(sleep_on_timeout);
4259 #ifdef CONFIG_RT_MUTEXES
4262 * rt_mutex_setprio - set the current priority of a task
4264 * @prio: prio value (kernel-internal form)
4266 * This function changes the 'effective' priority of a task. It does
4267 * not touch ->normal_prio like __setscheduler().
4269 * Used by the rt_mutex code to implement priority inheritance logic.
4271 void rt_mutex_setprio(struct task_struct *p, int prio)
4273 unsigned long flags;
4274 int oldprio, on_rq, running;
4276 const struct sched_class *prev_class = p->sched_class;
4278 BUG_ON(prio < 0 || prio > MAX_PRIO);
4280 rq = task_rq_lock(p, &flags);
4281 update_rq_clock(rq);
4284 on_rq = p->se.on_rq;
4285 running = task_current(rq, p);
4287 dequeue_task(rq, p, 0);
4289 p->sched_class->put_prev_task(rq, p);
4293 p->sched_class = &rt_sched_class;
4295 p->sched_class = &fair_sched_class;
4301 p->sched_class->set_curr_task(rq);
4303 enqueue_task(rq, p, 0);
4305 check_class_changed(rq, p, prev_class, oldprio, running);
4307 task_rq_unlock(rq, &flags);
4312 void set_user_nice(struct task_struct *p, long nice)
4314 int old_prio, delta, on_rq;
4315 unsigned long flags;
4318 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4321 * We have to be careful, if called from sys_setpriority(),
4322 * the task might be in the middle of scheduling on another CPU.
4324 rq = task_rq_lock(p, &flags);
4325 update_rq_clock(rq);
4327 * The RT priorities are set via sched_setscheduler(), but we still
4328 * allow the 'normal' nice value to be set - but as expected
4329 * it wont have any effect on scheduling until the task is
4330 * SCHED_FIFO/SCHED_RR:
4332 if (task_has_rt_policy(p)) {
4333 p->static_prio = NICE_TO_PRIO(nice);
4336 on_rq = p->se.on_rq;
4338 dequeue_task(rq, p, 0);
4340 p->static_prio = NICE_TO_PRIO(nice);
4343 p->prio = effective_prio(p);
4344 delta = p->prio - old_prio;
4347 enqueue_task(rq, p, 0);
4349 * If the task increased its priority or is running and
4350 * lowered its priority, then reschedule its CPU:
4352 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4353 resched_task(rq->curr);
4356 task_rq_unlock(rq, &flags);
4358 EXPORT_SYMBOL(set_user_nice);
4361 * can_nice - check if a task can reduce its nice value
4365 int can_nice(const struct task_struct *p, const int nice)
4367 /* convert nice value [19,-20] to rlimit style value [1,40] */
4368 int nice_rlim = 20 - nice;
4370 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4371 capable(CAP_SYS_NICE));
4374 #ifdef __ARCH_WANT_SYS_NICE
4377 * sys_nice - change the priority of the current process.
4378 * @increment: priority increment
4380 * sys_setpriority is a more generic, but much slower function that
4381 * does similar things.
4383 asmlinkage long sys_nice(int increment)
4388 * Setpriority might change our priority at the same moment.
4389 * We don't have to worry. Conceptually one call occurs first
4390 * and we have a single winner.
4392 if (increment < -40)
4397 nice = PRIO_TO_NICE(current->static_prio) + increment;
4403 if (increment < 0 && !can_nice(current, nice))
4406 retval = security_task_setnice(current, nice);
4410 set_user_nice(current, nice);
4417 * task_prio - return the priority value of a given task.
4418 * @p: the task in question.
4420 * This is the priority value as seen by users in /proc.
4421 * RT tasks are offset by -200. Normal tasks are centered
4422 * around 0, value goes from -16 to +15.
4424 int task_prio(const struct task_struct *p)
4426 return p->prio - MAX_RT_PRIO;
4430 * task_nice - return the nice value of a given task.
4431 * @p: the task in question.
4433 int task_nice(const struct task_struct *p)
4435 return TASK_NICE(p);
4437 EXPORT_SYMBOL_GPL(task_nice);
4440 * idle_cpu - is a given cpu idle currently?
4441 * @cpu: the processor in question.
4443 int idle_cpu(int cpu)
4445 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4449 * idle_task - return the idle task for a given cpu.
4450 * @cpu: the processor in question.
4452 struct task_struct *idle_task(int cpu)
4454 return cpu_rq(cpu)->idle;
4458 * find_process_by_pid - find a process with a matching PID value.
4459 * @pid: the pid in question.
4461 static struct task_struct *find_process_by_pid(pid_t pid)
4463 return pid ? find_task_by_vpid(pid) : current;
4466 /* Actually do priority change: must hold rq lock. */
4468 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4470 BUG_ON(p->se.on_rq);
4473 switch (p->policy) {
4477 p->sched_class = &fair_sched_class;
4481 p->sched_class = &rt_sched_class;
4485 p->rt_priority = prio;
4486 p->normal_prio = normal_prio(p);
4487 /* we are holding p->pi_lock already */
4488 p->prio = rt_mutex_getprio(p);
4493 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4494 * @p: the task in question.
4495 * @policy: new policy.
4496 * @param: structure containing the new RT priority.
4498 * NOTE that the task may be already dead.
4500 int sched_setscheduler(struct task_struct *p, int policy,
4501 struct sched_param *param)
4503 int retval, oldprio, oldpolicy = -1, on_rq, running;
4504 unsigned long flags;
4505 const struct sched_class *prev_class = p->sched_class;
4508 /* may grab non-irq protected spin_locks */
4509 BUG_ON(in_interrupt());
4511 /* double check policy once rq lock held */
4513 policy = oldpolicy = p->policy;
4514 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4515 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4516 policy != SCHED_IDLE)
4519 * Valid priorities for SCHED_FIFO and SCHED_RR are
4520 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4521 * SCHED_BATCH and SCHED_IDLE is 0.
4523 if (param->sched_priority < 0 ||
4524 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4525 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4527 if (rt_policy(policy) != (param->sched_priority != 0))
4531 * Allow unprivileged RT tasks to decrease priority:
4533 if (!capable(CAP_SYS_NICE)) {
4534 if (rt_policy(policy)) {
4535 unsigned long rlim_rtprio;
4537 if (!lock_task_sighand(p, &flags))
4539 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4540 unlock_task_sighand(p, &flags);
4542 /* can't set/change the rt policy */
4543 if (policy != p->policy && !rlim_rtprio)
4546 /* can't increase priority */
4547 if (param->sched_priority > p->rt_priority &&
4548 param->sched_priority > rlim_rtprio)
4552 * Like positive nice levels, dont allow tasks to
4553 * move out of SCHED_IDLE either:
4555 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4558 /* can't change other user's priorities */
4559 if ((current->euid != p->euid) &&
4560 (current->euid != p->uid))
4564 retval = security_task_setscheduler(p, policy, param);
4568 * make sure no PI-waiters arrive (or leave) while we are
4569 * changing the priority of the task:
4571 spin_lock_irqsave(&p->pi_lock, flags);
4573 * To be able to change p->policy safely, the apropriate
4574 * runqueue lock must be held.
4576 rq = __task_rq_lock(p);
4577 /* recheck policy now with rq lock held */
4578 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4579 policy = oldpolicy = -1;
4580 __task_rq_unlock(rq);
4581 spin_unlock_irqrestore(&p->pi_lock, flags);
4584 update_rq_clock(rq);
4585 on_rq = p->se.on_rq;
4586 running = task_current(rq, p);
4588 deactivate_task(rq, p, 0);
4590 p->sched_class->put_prev_task(rq, p);
4594 __setscheduler(rq, p, policy, param->sched_priority);
4598 p->sched_class->set_curr_task(rq);
4600 activate_task(rq, p, 0);
4602 check_class_changed(rq, p, prev_class, oldprio, running);
4604 __task_rq_unlock(rq);
4605 spin_unlock_irqrestore(&p->pi_lock, flags);
4607 rt_mutex_adjust_pi(p);
4611 EXPORT_SYMBOL_GPL(sched_setscheduler);
4614 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4616 struct sched_param lparam;
4617 struct task_struct *p;
4620 if (!param || pid < 0)
4622 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4627 p = find_process_by_pid(pid);
4629 retval = sched_setscheduler(p, policy, &lparam);
4636 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4637 * @pid: the pid in question.
4638 * @policy: new policy.
4639 * @param: structure containing the new RT priority.
4642 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4644 /* negative values for policy are not valid */
4648 return do_sched_setscheduler(pid, policy, param);
4652 * sys_sched_setparam - set/change the RT priority of a thread
4653 * @pid: the pid in question.
4654 * @param: structure containing the new RT priority.
4656 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4658 return do_sched_setscheduler(pid, -1, param);
4662 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4663 * @pid: the pid in question.
4665 asmlinkage long sys_sched_getscheduler(pid_t pid)
4667 struct task_struct *p;
4674 read_lock(&tasklist_lock);
4675 p = find_process_by_pid(pid);
4677 retval = security_task_getscheduler(p);
4681 read_unlock(&tasklist_lock);
4686 * sys_sched_getscheduler - get the RT priority of a thread
4687 * @pid: the pid in question.
4688 * @param: structure containing the RT priority.
4690 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4692 struct sched_param lp;
4693 struct task_struct *p;
4696 if (!param || pid < 0)
4699 read_lock(&tasklist_lock);
4700 p = find_process_by_pid(pid);
4705 retval = security_task_getscheduler(p);
4709 lp.sched_priority = p->rt_priority;
4710 read_unlock(&tasklist_lock);
4713 * This one might sleep, we cannot do it with a spinlock held ...
4715 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4720 read_unlock(&tasklist_lock);
4724 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4726 cpumask_t cpus_allowed;
4727 struct task_struct *p;
4731 read_lock(&tasklist_lock);
4733 p = find_process_by_pid(pid);
4735 read_unlock(&tasklist_lock);
4741 * It is not safe to call set_cpus_allowed with the
4742 * tasklist_lock held. We will bump the task_struct's
4743 * usage count and then drop tasklist_lock.
4746 read_unlock(&tasklist_lock);
4749 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4750 !capable(CAP_SYS_NICE))
4753 retval = security_task_setscheduler(p, 0, NULL);
4757 cpus_allowed = cpuset_cpus_allowed(p);
4758 cpus_and(new_mask, new_mask, cpus_allowed);
4760 retval = set_cpus_allowed(p, new_mask);
4763 cpus_allowed = cpuset_cpus_allowed(p);
4764 if (!cpus_subset(new_mask, cpus_allowed)) {
4766 * We must have raced with a concurrent cpuset
4767 * update. Just reset the cpus_allowed to the
4768 * cpuset's cpus_allowed
4770 new_mask = cpus_allowed;
4780 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4781 cpumask_t *new_mask)
4783 if (len < sizeof(cpumask_t)) {
4784 memset(new_mask, 0, sizeof(cpumask_t));
4785 } else if (len > sizeof(cpumask_t)) {
4786 len = sizeof(cpumask_t);
4788 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4792 * sys_sched_setaffinity - set the cpu affinity of a process
4793 * @pid: pid of the process
4794 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4795 * @user_mask_ptr: user-space pointer to the new cpu mask
4797 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4798 unsigned long __user *user_mask_ptr)
4803 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4807 return sched_setaffinity(pid, new_mask);
4811 * Represents all cpu's present in the system
4812 * In systems capable of hotplug, this map could dynamically grow
4813 * as new cpu's are detected in the system via any platform specific
4814 * method, such as ACPI for e.g.
4817 cpumask_t cpu_present_map __read_mostly;
4818 EXPORT_SYMBOL(cpu_present_map);
4821 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4822 EXPORT_SYMBOL(cpu_online_map);
4824 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4825 EXPORT_SYMBOL(cpu_possible_map);
4828 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4830 struct task_struct *p;
4834 read_lock(&tasklist_lock);
4837 p = find_process_by_pid(pid);
4841 retval = security_task_getscheduler(p);
4845 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4848 read_unlock(&tasklist_lock);
4855 * sys_sched_getaffinity - get the cpu affinity of a process
4856 * @pid: pid of the process
4857 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4858 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4860 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4861 unsigned long __user *user_mask_ptr)
4866 if (len < sizeof(cpumask_t))
4869 ret = sched_getaffinity(pid, &mask);
4873 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4876 return sizeof(cpumask_t);
4880 * sys_sched_yield - yield the current processor to other threads.
4882 * This function yields the current CPU to other tasks. If there are no
4883 * other threads running on this CPU then this function will return.
4885 asmlinkage long sys_sched_yield(void)
4887 struct rq *rq = this_rq_lock();
4889 schedstat_inc(rq, yld_count);
4890 current->sched_class->yield_task(rq);
4893 * Since we are going to call schedule() anyway, there's
4894 * no need to preempt or enable interrupts:
4896 __release(rq->lock);
4897 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4898 _raw_spin_unlock(&rq->lock);
4899 preempt_enable_no_resched();
4906 static void __cond_resched(void)
4908 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4909 __might_sleep(__FILE__, __LINE__);
4912 * The BKS might be reacquired before we have dropped
4913 * PREEMPT_ACTIVE, which could trigger a second
4914 * cond_resched() call.
4917 add_preempt_count(PREEMPT_ACTIVE);
4919 sub_preempt_count(PREEMPT_ACTIVE);
4920 } while (need_resched());
4923 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4924 int __sched _cond_resched(void)
4926 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4927 system_state == SYSTEM_RUNNING) {
4933 EXPORT_SYMBOL(_cond_resched);
4937 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4938 * call schedule, and on return reacquire the lock.
4940 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4941 * operations here to prevent schedule() from being called twice (once via
4942 * spin_unlock(), once by hand).
4944 int cond_resched_lock(spinlock_t *lock)
4946 int resched = need_resched() && system_state == SYSTEM_RUNNING;
4949 if (spin_needbreak(lock) || resched) {
4951 if (resched && need_resched())
4960 EXPORT_SYMBOL(cond_resched_lock);
4962 int __sched cond_resched_softirq(void)
4964 BUG_ON(!in_softirq());
4966 if (need_resched() && system_state == SYSTEM_RUNNING) {
4974 EXPORT_SYMBOL(cond_resched_softirq);
4977 * yield - yield the current processor to other threads.
4979 * This is a shortcut for kernel-space yielding - it marks the
4980 * thread runnable and calls sys_sched_yield().
4982 void __sched yield(void)
4984 set_current_state(TASK_RUNNING);
4987 EXPORT_SYMBOL(yield);
4990 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4991 * that process accounting knows that this is a task in IO wait state.
4993 * But don't do that if it is a deliberate, throttling IO wait (this task
4994 * has set its backing_dev_info: the queue against which it should throttle)
4996 void __sched io_schedule(void)
4998 struct rq *rq = &__raw_get_cpu_var(runqueues);
5000 delayacct_blkio_start();
5001 atomic_inc(&rq->nr_iowait);
5003 atomic_dec(&rq->nr_iowait);
5004 delayacct_blkio_end();
5006 EXPORT_SYMBOL(io_schedule);
5008 long __sched io_schedule_timeout(long timeout)
5010 struct rq *rq = &__raw_get_cpu_var(runqueues);
5013 delayacct_blkio_start();
5014 atomic_inc(&rq->nr_iowait);
5015 ret = schedule_timeout(timeout);
5016 atomic_dec(&rq->nr_iowait);
5017 delayacct_blkio_end();
5022 * sys_sched_get_priority_max - return maximum RT priority.
5023 * @policy: scheduling class.
5025 * this syscall returns the maximum rt_priority that can be used
5026 * by a given scheduling class.
5028 asmlinkage long sys_sched_get_priority_max(int policy)
5035 ret = MAX_USER_RT_PRIO-1;
5047 * sys_sched_get_priority_min - return minimum RT priority.
5048 * @policy: scheduling class.
5050 * this syscall returns the minimum rt_priority that can be used
5051 * by a given scheduling class.
5053 asmlinkage long sys_sched_get_priority_min(int policy)
5071 * sys_sched_rr_get_interval - return the default timeslice of a process.
5072 * @pid: pid of the process.
5073 * @interval: userspace pointer to the timeslice value.
5075 * this syscall writes the default timeslice value of a given process
5076 * into the user-space timespec buffer. A value of '0' means infinity.
5079 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5081 struct task_struct *p;
5082 unsigned int time_slice;
5090 read_lock(&tasklist_lock);
5091 p = find_process_by_pid(pid);
5095 retval = security_task_getscheduler(p);
5100 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5101 * tasks that are on an otherwise idle runqueue:
5104 if (p->policy == SCHED_RR) {
5105 time_slice = DEF_TIMESLICE;
5107 struct sched_entity *se = &p->se;
5108 unsigned long flags;
5111 rq = task_rq_lock(p, &flags);
5112 if (rq->cfs.load.weight)
5113 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5114 task_rq_unlock(rq, &flags);
5116 read_unlock(&tasklist_lock);
5117 jiffies_to_timespec(time_slice, &t);
5118 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5122 read_unlock(&tasklist_lock);
5126 static const char stat_nam[] = "RSDTtZX";
5128 void sched_show_task(struct task_struct *p)
5130 unsigned long free = 0;
5133 state = p->state ? __ffs(p->state) + 1 : 0;
5134 printk(KERN_INFO "%-13.13s %c", p->comm,
5135 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5136 #if BITS_PER_LONG == 32
5137 if (state == TASK_RUNNING)
5138 printk(KERN_CONT " running ");
5140 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5142 if (state == TASK_RUNNING)
5143 printk(KERN_CONT " running task ");
5145 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5147 #ifdef CONFIG_DEBUG_STACK_USAGE
5149 unsigned long *n = end_of_stack(p);
5152 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5155 printk(KERN_CONT "%5lu %5d %6d\n", free,
5156 task_pid_nr(p), task_pid_nr(p->real_parent));
5158 show_stack(p, NULL);
5161 void show_state_filter(unsigned long state_filter)
5163 struct task_struct *g, *p;
5165 #if BITS_PER_LONG == 32
5167 " task PC stack pid father\n");
5170 " task PC stack pid father\n");
5172 read_lock(&tasklist_lock);
5173 do_each_thread(g, p) {
5175 * reset the NMI-timeout, listing all files on a slow
5176 * console might take alot of time:
5178 touch_nmi_watchdog();
5179 if (!state_filter || (p->state & state_filter))
5181 } while_each_thread(g, p);
5183 touch_all_softlockup_watchdogs();
5185 #ifdef CONFIG_SCHED_DEBUG
5186 sysrq_sched_debug_show();
5188 read_unlock(&tasklist_lock);
5190 * Only show locks if all tasks are dumped:
5192 if (state_filter == -1)
5193 debug_show_all_locks();
5196 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5198 idle->sched_class = &idle_sched_class;
5202 * init_idle - set up an idle thread for a given CPU
5203 * @idle: task in question
5204 * @cpu: cpu the idle task belongs to
5206 * NOTE: this function does not set the idle thread's NEED_RESCHED
5207 * flag, to make booting more robust.
5209 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5211 struct rq *rq = cpu_rq(cpu);
5212 unsigned long flags;
5215 idle->se.exec_start = sched_clock();
5217 idle->prio = idle->normal_prio = MAX_PRIO;
5218 idle->cpus_allowed = cpumask_of_cpu(cpu);
5219 __set_task_cpu(idle, cpu);
5221 spin_lock_irqsave(&rq->lock, flags);
5222 rq->curr = rq->idle = idle;
5223 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5226 spin_unlock_irqrestore(&rq->lock, flags);
5228 /* Set the preempt count _outside_ the spinlocks! */
5229 task_thread_info(idle)->preempt_count = 0;
5232 * The idle tasks have their own, simple scheduling class:
5234 idle->sched_class = &idle_sched_class;
5238 * In a system that switches off the HZ timer nohz_cpu_mask
5239 * indicates which cpus entered this state. This is used
5240 * in the rcu update to wait only for active cpus. For system
5241 * which do not switch off the HZ timer nohz_cpu_mask should
5242 * always be CPU_MASK_NONE.
5244 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5247 * Increase the granularity value when there are more CPUs,
5248 * because with more CPUs the 'effective latency' as visible
5249 * to users decreases. But the relationship is not linear,
5250 * so pick a second-best guess by going with the log2 of the
5253 * This idea comes from the SD scheduler of Con Kolivas:
5255 static inline void sched_init_granularity(void)
5257 unsigned int factor = 1 + ilog2(num_online_cpus());
5258 const unsigned long limit = 200000000;
5260 sysctl_sched_min_granularity *= factor;
5261 if (sysctl_sched_min_granularity > limit)
5262 sysctl_sched_min_granularity = limit;
5264 sysctl_sched_latency *= factor;
5265 if (sysctl_sched_latency > limit)
5266 sysctl_sched_latency = limit;
5268 sysctl_sched_wakeup_granularity *= factor;
5269 sysctl_sched_batch_wakeup_granularity *= factor;
5274 * This is how migration works:
5276 * 1) we queue a struct migration_req structure in the source CPU's
5277 * runqueue and wake up that CPU's migration thread.
5278 * 2) we down() the locked semaphore => thread blocks.
5279 * 3) migration thread wakes up (implicitly it forces the migrated
5280 * thread off the CPU)
5281 * 4) it gets the migration request and checks whether the migrated
5282 * task is still in the wrong runqueue.
5283 * 5) if it's in the wrong runqueue then the migration thread removes
5284 * it and puts it into the right queue.
5285 * 6) migration thread up()s the semaphore.
5286 * 7) we wake up and the migration is done.
5290 * Change a given task's CPU affinity. Migrate the thread to a
5291 * proper CPU and schedule it away if the CPU it's executing on
5292 * is removed from the allowed bitmask.
5294 * NOTE: the caller must have a valid reference to the task, the
5295 * task must not exit() & deallocate itself prematurely. The
5296 * call is not atomic; no spinlocks may be held.
5298 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5300 struct migration_req req;
5301 unsigned long flags;
5305 rq = task_rq_lock(p, &flags);
5306 if (!cpus_intersects(new_mask, cpu_online_map)) {
5311 if (p->sched_class->set_cpus_allowed)
5312 p->sched_class->set_cpus_allowed(p, &new_mask);
5314 p->cpus_allowed = new_mask;
5315 p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5318 /* Can the task run on the task's current CPU? If so, we're done */
5319 if (cpu_isset(task_cpu(p), new_mask))
5322 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5323 /* Need help from migration thread: drop lock and wait. */
5324 task_rq_unlock(rq, &flags);
5325 wake_up_process(rq->migration_thread);
5326 wait_for_completion(&req.done);
5327 tlb_migrate_finish(p->mm);
5331 task_rq_unlock(rq, &flags);
5335 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5338 * Move (not current) task off this cpu, onto dest cpu. We're doing
5339 * this because either it can't run here any more (set_cpus_allowed()
5340 * away from this CPU, or CPU going down), or because we're
5341 * attempting to rebalance this task on exec (sched_exec).
5343 * So we race with normal scheduler movements, but that's OK, as long
5344 * as the task is no longer on this CPU.
5346 * Returns non-zero if task was successfully migrated.
5348 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5350 struct rq *rq_dest, *rq_src;
5353 if (unlikely(cpu_is_offline(dest_cpu)))
5356 rq_src = cpu_rq(src_cpu);
5357 rq_dest = cpu_rq(dest_cpu);
5359 double_rq_lock(rq_src, rq_dest);
5360 /* Already moved. */
5361 if (task_cpu(p) != src_cpu)
5363 /* Affinity changed (again). */
5364 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5367 on_rq = p->se.on_rq;
5369 deactivate_task(rq_src, p, 0);
5371 set_task_cpu(p, dest_cpu);
5373 activate_task(rq_dest, p, 0);
5374 check_preempt_curr(rq_dest, p);
5378 double_rq_unlock(rq_src, rq_dest);
5383 * migration_thread - this is a highprio system thread that performs
5384 * thread migration by bumping thread off CPU then 'pushing' onto
5387 static int migration_thread(void *data)
5389 int cpu = (long)data;
5393 BUG_ON(rq->migration_thread != current);
5395 set_current_state(TASK_INTERRUPTIBLE);
5396 while (!kthread_should_stop()) {
5397 struct migration_req *req;
5398 struct list_head *head;
5400 spin_lock_irq(&rq->lock);
5402 if (cpu_is_offline(cpu)) {
5403 spin_unlock_irq(&rq->lock);
5407 if (rq->active_balance) {
5408 active_load_balance(rq, cpu);
5409 rq->active_balance = 0;
5412 head = &rq->migration_queue;
5414 if (list_empty(head)) {
5415 spin_unlock_irq(&rq->lock);
5417 set_current_state(TASK_INTERRUPTIBLE);
5420 req = list_entry(head->next, struct migration_req, list);
5421 list_del_init(head->next);
5423 spin_unlock(&rq->lock);
5424 __migrate_task(req->task, cpu, req->dest_cpu);
5427 complete(&req->done);
5429 __set_current_state(TASK_RUNNING);
5433 /* Wait for kthread_stop */
5434 set_current_state(TASK_INTERRUPTIBLE);
5435 while (!kthread_should_stop()) {
5437 set_current_state(TASK_INTERRUPTIBLE);
5439 __set_current_state(TASK_RUNNING);
5443 #ifdef CONFIG_HOTPLUG_CPU
5445 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5449 local_irq_disable();
5450 ret = __migrate_task(p, src_cpu, dest_cpu);
5456 * Figure out where task on dead CPU should go, use force if necessary.
5457 * NOTE: interrupts should be disabled by the caller
5459 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5461 unsigned long flags;
5468 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5469 cpus_and(mask, mask, p->cpus_allowed);
5470 dest_cpu = any_online_cpu(mask);
5472 /* On any allowed CPU? */
5473 if (dest_cpu == NR_CPUS)
5474 dest_cpu = any_online_cpu(p->cpus_allowed);
5476 /* No more Mr. Nice Guy. */
5477 if (dest_cpu == NR_CPUS) {
5478 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5480 * Try to stay on the same cpuset, where the
5481 * current cpuset may be a subset of all cpus.
5482 * The cpuset_cpus_allowed_locked() variant of
5483 * cpuset_cpus_allowed() will not block. It must be
5484 * called within calls to cpuset_lock/cpuset_unlock.
5486 rq = task_rq_lock(p, &flags);
5487 p->cpus_allowed = cpus_allowed;
5488 dest_cpu = any_online_cpu(p->cpus_allowed);
5489 task_rq_unlock(rq, &flags);
5492 * Don't tell them about moving exiting tasks or
5493 * kernel threads (both mm NULL), since they never
5496 if (p->mm && printk_ratelimit()) {
5497 printk(KERN_INFO "process %d (%s) no "
5498 "longer affine to cpu%d\n",
5499 task_pid_nr(p), p->comm, dead_cpu);
5502 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5506 * While a dead CPU has no uninterruptible tasks queued at this point,
5507 * it might still have a nonzero ->nr_uninterruptible counter, because
5508 * for performance reasons the counter is not stricly tracking tasks to
5509 * their home CPUs. So we just add the counter to another CPU's counter,
5510 * to keep the global sum constant after CPU-down:
5512 static void migrate_nr_uninterruptible(struct rq *rq_src)
5514 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5515 unsigned long flags;
5517 local_irq_save(flags);
5518 double_rq_lock(rq_src, rq_dest);
5519 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5520 rq_src->nr_uninterruptible = 0;
5521 double_rq_unlock(rq_src, rq_dest);
5522 local_irq_restore(flags);
5525 /* Run through task list and migrate tasks from the dead cpu. */
5526 static void migrate_live_tasks(int src_cpu)
5528 struct task_struct *p, *t;
5530 read_lock(&tasklist_lock);
5532 do_each_thread(t, p) {
5536 if (task_cpu(p) == src_cpu)
5537 move_task_off_dead_cpu(src_cpu, p);
5538 } while_each_thread(t, p);
5540 read_unlock(&tasklist_lock);
5544 * Schedules idle task to be the next runnable task on current CPU.
5545 * It does so by boosting its priority to highest possible.
5546 * Used by CPU offline code.
5548 void sched_idle_next(void)
5550 int this_cpu = smp_processor_id();
5551 struct rq *rq = cpu_rq(this_cpu);
5552 struct task_struct *p = rq->idle;
5553 unsigned long flags;
5555 /* cpu has to be offline */
5556 BUG_ON(cpu_online(this_cpu));
5559 * Strictly not necessary since rest of the CPUs are stopped by now
5560 * and interrupts disabled on the current cpu.
5562 spin_lock_irqsave(&rq->lock, flags);
5564 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5566 update_rq_clock(rq);
5567 activate_task(rq, p, 0);
5569 spin_unlock_irqrestore(&rq->lock, flags);
5573 * Ensures that the idle task is using init_mm right before its cpu goes
5576 void idle_task_exit(void)
5578 struct mm_struct *mm = current->active_mm;
5580 BUG_ON(cpu_online(smp_processor_id()));
5583 switch_mm(mm, &init_mm, current);
5587 /* called under rq->lock with disabled interrupts */
5588 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5590 struct rq *rq = cpu_rq(dead_cpu);
5592 /* Must be exiting, otherwise would be on tasklist. */
5593 BUG_ON(!p->exit_state);
5595 /* Cannot have done final schedule yet: would have vanished. */
5596 BUG_ON(p->state == TASK_DEAD);
5601 * Drop lock around migration; if someone else moves it,
5602 * that's OK. No task can be added to this CPU, so iteration is
5605 spin_unlock_irq(&rq->lock);
5606 move_task_off_dead_cpu(dead_cpu, p);
5607 spin_lock_irq(&rq->lock);
5612 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5613 static void migrate_dead_tasks(unsigned int dead_cpu)
5615 struct rq *rq = cpu_rq(dead_cpu);
5616 struct task_struct *next;
5619 if (!rq->nr_running)
5621 update_rq_clock(rq);
5622 next = pick_next_task(rq, rq->curr);
5625 migrate_dead(dead_cpu, next);
5629 #endif /* CONFIG_HOTPLUG_CPU */
5631 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5633 static struct ctl_table sd_ctl_dir[] = {
5635 .procname = "sched_domain",
5641 static struct ctl_table sd_ctl_root[] = {
5643 .ctl_name = CTL_KERN,
5644 .procname = "kernel",
5646 .child = sd_ctl_dir,
5651 static struct ctl_table *sd_alloc_ctl_entry(int n)
5653 struct ctl_table *entry =
5654 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5659 static void sd_free_ctl_entry(struct ctl_table **tablep)
5661 struct ctl_table *entry;
5664 * In the intermediate directories, both the child directory and
5665 * procname are dynamically allocated and could fail but the mode
5666 * will always be set. In the lowest directory the names are
5667 * static strings and all have proc handlers.
5669 for (entry = *tablep; entry->mode; entry++) {
5671 sd_free_ctl_entry(&entry->child);
5672 if (entry->proc_handler == NULL)
5673 kfree(entry->procname);
5681 set_table_entry(struct ctl_table *entry,
5682 const char *procname, void *data, int maxlen,
5683 mode_t mode, proc_handler *proc_handler)
5685 entry->procname = procname;
5687 entry->maxlen = maxlen;
5689 entry->proc_handler = proc_handler;
5692 static struct ctl_table *
5693 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5695 struct ctl_table *table = sd_alloc_ctl_entry(12);
5700 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5701 sizeof(long), 0644, proc_doulongvec_minmax);
5702 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5703 sizeof(long), 0644, proc_doulongvec_minmax);
5704 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5705 sizeof(int), 0644, proc_dointvec_minmax);
5706 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5707 sizeof(int), 0644, proc_dointvec_minmax);
5708 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5709 sizeof(int), 0644, proc_dointvec_minmax);
5710 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5711 sizeof(int), 0644, proc_dointvec_minmax);
5712 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5713 sizeof(int), 0644, proc_dointvec_minmax);
5714 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5715 sizeof(int), 0644, proc_dointvec_minmax);
5716 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5717 sizeof(int), 0644, proc_dointvec_minmax);
5718 set_table_entry(&table[9], "cache_nice_tries",
5719 &sd->cache_nice_tries,
5720 sizeof(int), 0644, proc_dointvec_minmax);
5721 set_table_entry(&table[10], "flags", &sd->flags,
5722 sizeof(int), 0644, proc_dointvec_minmax);
5723 /* &table[11] is terminator */
5728 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5730 struct ctl_table *entry, *table;
5731 struct sched_domain *sd;
5732 int domain_num = 0, i;
5735 for_each_domain(cpu, sd)
5737 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5742 for_each_domain(cpu, sd) {
5743 snprintf(buf, 32, "domain%d", i);
5744 entry->procname = kstrdup(buf, GFP_KERNEL);
5746 entry->child = sd_alloc_ctl_domain_table(sd);
5753 static struct ctl_table_header *sd_sysctl_header;
5754 static void register_sched_domain_sysctl(void)
5756 int i, cpu_num = num_online_cpus();
5757 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5760 WARN_ON(sd_ctl_dir[0].child);
5761 sd_ctl_dir[0].child = entry;
5766 for_each_online_cpu(i) {
5767 snprintf(buf, 32, "cpu%d", i);
5768 entry->procname = kstrdup(buf, GFP_KERNEL);
5770 entry->child = sd_alloc_ctl_cpu_table(i);
5774 WARN_ON(sd_sysctl_header);
5775 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5778 /* may be called multiple times per register */
5779 static void unregister_sched_domain_sysctl(void)
5781 if (sd_sysctl_header)
5782 unregister_sysctl_table(sd_sysctl_header);
5783 sd_sysctl_header = NULL;
5784 if (sd_ctl_dir[0].child)
5785 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5788 static void register_sched_domain_sysctl(void)
5791 static void unregister_sched_domain_sysctl(void)
5797 * migration_call - callback that gets triggered when a CPU is added.
5798 * Here we can start up the necessary migration thread for the new CPU.
5800 static int __cpuinit
5801 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5803 struct task_struct *p;
5804 int cpu = (long)hcpu;
5805 unsigned long flags;
5810 case CPU_UP_PREPARE:
5811 case CPU_UP_PREPARE_FROZEN:
5812 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5815 kthread_bind(p, cpu);
5816 /* Must be high prio: stop_machine expects to yield to it. */
5817 rq = task_rq_lock(p, &flags);
5818 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5819 task_rq_unlock(rq, &flags);
5820 cpu_rq(cpu)->migration_thread = p;
5824 case CPU_ONLINE_FROZEN:
5825 /* Strictly unnecessary, as first user will wake it. */
5826 wake_up_process(cpu_rq(cpu)->migration_thread);
5828 /* Update our root-domain */
5830 spin_lock_irqsave(&rq->lock, flags);
5832 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5833 cpu_set(cpu, rq->rd->online);
5835 spin_unlock_irqrestore(&rq->lock, flags);
5838 #ifdef CONFIG_HOTPLUG_CPU
5839 case CPU_UP_CANCELED:
5840 case CPU_UP_CANCELED_FROZEN:
5841 if (!cpu_rq(cpu)->migration_thread)
5843 /* Unbind it from offline cpu so it can run. Fall thru. */
5844 kthread_bind(cpu_rq(cpu)->migration_thread,
5845 any_online_cpu(cpu_online_map));
5846 kthread_stop(cpu_rq(cpu)->migration_thread);
5847 cpu_rq(cpu)->migration_thread = NULL;
5851 case CPU_DEAD_FROZEN:
5852 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5853 migrate_live_tasks(cpu);
5855 kthread_stop(rq->migration_thread);
5856 rq->migration_thread = NULL;
5857 /* Idle task back to normal (off runqueue, low prio) */
5858 spin_lock_irq(&rq->lock);
5859 update_rq_clock(rq);
5860 deactivate_task(rq, rq->idle, 0);
5861 rq->idle->static_prio = MAX_PRIO;
5862 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5863 rq->idle->sched_class = &idle_sched_class;
5864 migrate_dead_tasks(cpu);
5865 spin_unlock_irq(&rq->lock);
5867 migrate_nr_uninterruptible(rq);
5868 BUG_ON(rq->nr_running != 0);
5871 * No need to migrate the tasks: it was best-effort if
5872 * they didn't take sched_hotcpu_mutex. Just wake up
5875 spin_lock_irq(&rq->lock);
5876 while (!list_empty(&rq->migration_queue)) {
5877 struct migration_req *req;
5879 req = list_entry(rq->migration_queue.next,
5880 struct migration_req, list);
5881 list_del_init(&req->list);
5882 complete(&req->done);
5884 spin_unlock_irq(&rq->lock);
5887 case CPU_DOWN_PREPARE:
5888 /* Update our root-domain */
5890 spin_lock_irqsave(&rq->lock, flags);
5892 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5893 cpu_clear(cpu, rq->rd->online);
5895 spin_unlock_irqrestore(&rq->lock, flags);
5902 /* Register at highest priority so that task migration (migrate_all_tasks)
5903 * happens before everything else.
5905 static struct notifier_block __cpuinitdata migration_notifier = {
5906 .notifier_call = migration_call,
5910 void __init migration_init(void)
5912 void *cpu = (void *)(long)smp_processor_id();
5915 /* Start one for the boot CPU: */
5916 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5917 BUG_ON(err == NOTIFY_BAD);
5918 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5919 register_cpu_notifier(&migration_notifier);
5925 /* Number of possible processor ids */
5926 int nr_cpu_ids __read_mostly = NR_CPUS;
5927 EXPORT_SYMBOL(nr_cpu_ids);
5929 #ifdef CONFIG_SCHED_DEBUG
5931 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5933 struct sched_group *group = sd->groups;
5934 cpumask_t groupmask;
5937 cpumask_scnprintf(str, NR_CPUS, sd->span);
5938 cpus_clear(groupmask);
5940 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5942 if (!(sd->flags & SD_LOAD_BALANCE)) {
5943 printk("does not load-balance\n");
5945 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5950 printk(KERN_CONT "span %s\n", str);
5952 if (!cpu_isset(cpu, sd->span)) {
5953 printk(KERN_ERR "ERROR: domain->span does not contain "
5956 if (!cpu_isset(cpu, group->cpumask)) {
5957 printk(KERN_ERR "ERROR: domain->groups does not contain"
5961 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5965 printk(KERN_ERR "ERROR: group is NULL\n");
5969 if (!group->__cpu_power) {
5970 printk(KERN_CONT "\n");
5971 printk(KERN_ERR "ERROR: domain->cpu_power not "
5976 if (!cpus_weight(group->cpumask)) {
5977 printk(KERN_CONT "\n");
5978 printk(KERN_ERR "ERROR: empty group\n");
5982 if (cpus_intersects(groupmask, group->cpumask)) {
5983 printk(KERN_CONT "\n");
5984 printk(KERN_ERR "ERROR: repeated CPUs\n");
5988 cpus_or(groupmask, groupmask, group->cpumask);
5990 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5991 printk(KERN_CONT " %s", str);
5993 group = group->next;
5994 } while (group != sd->groups);
5995 printk(KERN_CONT "\n");
5997 if (!cpus_equal(sd->span, groupmask))
5998 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6000 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
6001 printk(KERN_ERR "ERROR: parent span is not a superset "
6002 "of domain->span\n");
6006 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6011 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6015 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6018 if (sched_domain_debug_one(sd, cpu, level))
6027 # define sched_domain_debug(sd, cpu) do { } while (0)
6030 static int sd_degenerate(struct sched_domain *sd)
6032 if (cpus_weight(sd->span) == 1)
6035 /* Following flags need at least 2 groups */
6036 if (sd->flags & (SD_LOAD_BALANCE |
6037 SD_BALANCE_NEWIDLE |
6041 SD_SHARE_PKG_RESOURCES)) {
6042 if (sd->groups != sd->groups->next)
6046 /* Following flags don't use groups */
6047 if (sd->flags & (SD_WAKE_IDLE |
6056 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6058 unsigned long cflags = sd->flags, pflags = parent->flags;
6060 if (sd_degenerate(parent))
6063 if (!cpus_equal(sd->span, parent->span))
6066 /* Does parent contain flags not in child? */
6067 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6068 if (cflags & SD_WAKE_AFFINE)
6069 pflags &= ~SD_WAKE_BALANCE;
6070 /* Flags needing groups don't count if only 1 group in parent */
6071 if (parent->groups == parent->groups->next) {
6072 pflags &= ~(SD_LOAD_BALANCE |
6073 SD_BALANCE_NEWIDLE |
6077 SD_SHARE_PKG_RESOURCES);
6079 if (~cflags & pflags)
6085 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6087 unsigned long flags;
6088 const struct sched_class *class;
6090 spin_lock_irqsave(&rq->lock, flags);
6093 struct root_domain *old_rd = rq->rd;
6095 for (class = sched_class_highest; class; class = class->next) {
6096 if (class->leave_domain)
6097 class->leave_domain(rq);
6100 cpu_clear(rq->cpu, old_rd->span);
6101 cpu_clear(rq->cpu, old_rd->online);
6103 if (atomic_dec_and_test(&old_rd->refcount))
6107 atomic_inc(&rd->refcount);
6110 cpu_set(rq->cpu, rd->span);
6111 if (cpu_isset(rq->cpu, cpu_online_map))
6112 cpu_set(rq->cpu, rd->online);
6114 for (class = sched_class_highest; class; class = class->next) {
6115 if (class->join_domain)
6116 class->join_domain(rq);
6119 spin_unlock_irqrestore(&rq->lock, flags);
6122 static void init_rootdomain(struct root_domain *rd)
6124 memset(rd, 0, sizeof(*rd));
6126 cpus_clear(rd->span);
6127 cpus_clear(rd->online);
6130 static void init_defrootdomain(void)
6132 init_rootdomain(&def_root_domain);
6133 atomic_set(&def_root_domain.refcount, 1);
6136 static struct root_domain *alloc_rootdomain(void)
6138 struct root_domain *rd;
6140 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6144 init_rootdomain(rd);
6150 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6151 * hold the hotplug lock.
6154 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6156 struct rq *rq = cpu_rq(cpu);
6157 struct sched_domain *tmp;
6159 /* Remove the sched domains which do not contribute to scheduling. */
6160 for (tmp = sd; tmp; tmp = tmp->parent) {
6161 struct sched_domain *parent = tmp->parent;
6164 if (sd_parent_degenerate(tmp, parent)) {
6165 tmp->parent = parent->parent;
6167 parent->parent->child = tmp;
6171 if (sd && sd_degenerate(sd)) {
6177 sched_domain_debug(sd, cpu);
6179 rq_attach_root(rq, rd);
6180 rcu_assign_pointer(rq->sd, sd);
6183 /* cpus with isolated domains */
6184 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6186 /* Setup the mask of cpus configured for isolated domains */
6187 static int __init isolated_cpu_setup(char *str)
6189 int ints[NR_CPUS], i;
6191 str = get_options(str, ARRAY_SIZE(ints), ints);
6192 cpus_clear(cpu_isolated_map);
6193 for (i = 1; i <= ints[0]; i++)
6194 if (ints[i] < NR_CPUS)
6195 cpu_set(ints[i], cpu_isolated_map);
6199 __setup("isolcpus=", isolated_cpu_setup);
6202 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6203 * to a function which identifies what group(along with sched group) a CPU
6204 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6205 * (due to the fact that we keep track of groups covered with a cpumask_t).
6207 * init_sched_build_groups will build a circular linked list of the groups
6208 * covered by the given span, and will set each group's ->cpumask correctly,
6209 * and ->cpu_power to 0.
6212 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6213 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6214 struct sched_group **sg))
6216 struct sched_group *first = NULL, *last = NULL;
6217 cpumask_t covered = CPU_MASK_NONE;
6220 for_each_cpu_mask(i, span) {
6221 struct sched_group *sg;
6222 int group = group_fn(i, cpu_map, &sg);
6225 if (cpu_isset(i, covered))
6228 sg->cpumask = CPU_MASK_NONE;
6229 sg->__cpu_power = 0;
6231 for_each_cpu_mask(j, span) {
6232 if (group_fn(j, cpu_map, NULL) != group)
6235 cpu_set(j, covered);
6236 cpu_set(j, sg->cpumask);
6247 #define SD_NODES_PER_DOMAIN 16
6252 * find_next_best_node - find the next node to include in a sched_domain
6253 * @node: node whose sched_domain we're building
6254 * @used_nodes: nodes already in the sched_domain
6256 * Find the next node to include in a given scheduling domain. Simply
6257 * finds the closest node not already in the @used_nodes map.
6259 * Should use nodemask_t.
6261 static int find_next_best_node(int node, unsigned long *used_nodes)
6263 int i, n, val, min_val, best_node = 0;
6267 for (i = 0; i < MAX_NUMNODES; i++) {
6268 /* Start at @node */
6269 n = (node + i) % MAX_NUMNODES;
6271 if (!nr_cpus_node(n))
6274 /* Skip already used nodes */
6275 if (test_bit(n, used_nodes))
6278 /* Simple min distance search */
6279 val = node_distance(node, n);
6281 if (val < min_val) {
6287 set_bit(best_node, used_nodes);
6292 * sched_domain_node_span - get a cpumask for a node's sched_domain
6293 * @node: node whose cpumask we're constructing
6294 * @size: number of nodes to include in this span
6296 * Given a node, construct a good cpumask for its sched_domain to span. It
6297 * should be one that prevents unnecessary balancing, but also spreads tasks
6300 static cpumask_t sched_domain_node_span(int node)
6302 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6303 cpumask_t span, nodemask;
6307 bitmap_zero(used_nodes, MAX_NUMNODES);
6309 nodemask = node_to_cpumask(node);
6310 cpus_or(span, span, nodemask);
6311 set_bit(node, used_nodes);
6313 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6314 int next_node = find_next_best_node(node, used_nodes);
6316 nodemask = node_to_cpumask(next_node);
6317 cpus_or(span, span, nodemask);
6324 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6327 * SMT sched-domains:
6329 #ifdef CONFIG_SCHED_SMT
6330 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6331 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6334 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6337 *sg = &per_cpu(sched_group_cpus, cpu);
6343 * multi-core sched-domains:
6345 #ifdef CONFIG_SCHED_MC
6346 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6347 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6350 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6352 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6355 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6356 cpus_and(mask, mask, *cpu_map);
6357 group = first_cpu(mask);
6359 *sg = &per_cpu(sched_group_core, group);
6362 #elif defined(CONFIG_SCHED_MC)
6364 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6367 *sg = &per_cpu(sched_group_core, cpu);
6372 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6373 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6376 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6379 #ifdef CONFIG_SCHED_MC
6380 cpumask_t mask = cpu_coregroup_map(cpu);
6381 cpus_and(mask, mask, *cpu_map);
6382 group = first_cpu(mask);
6383 #elif defined(CONFIG_SCHED_SMT)
6384 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6385 cpus_and(mask, mask, *cpu_map);
6386 group = first_cpu(mask);
6391 *sg = &per_cpu(sched_group_phys, group);
6397 * The init_sched_build_groups can't handle what we want to do with node
6398 * groups, so roll our own. Now each node has its own list of groups which
6399 * gets dynamically allocated.
6401 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6402 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6404 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6405 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6407 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6408 struct sched_group **sg)
6410 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6413 cpus_and(nodemask, nodemask, *cpu_map);
6414 group = first_cpu(nodemask);
6417 *sg = &per_cpu(sched_group_allnodes, group);
6421 static void init_numa_sched_groups_power(struct sched_group *group_head)
6423 struct sched_group *sg = group_head;
6429 for_each_cpu_mask(j, sg->cpumask) {
6430 struct sched_domain *sd;
6432 sd = &per_cpu(phys_domains, j);
6433 if (j != first_cpu(sd->groups->cpumask)) {
6435 * Only add "power" once for each
6441 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6444 } while (sg != group_head);
6449 /* Free memory allocated for various sched_group structures */
6450 static void free_sched_groups(const cpumask_t *cpu_map)
6454 for_each_cpu_mask(cpu, *cpu_map) {
6455 struct sched_group **sched_group_nodes
6456 = sched_group_nodes_bycpu[cpu];
6458 if (!sched_group_nodes)
6461 for (i = 0; i < MAX_NUMNODES; i++) {
6462 cpumask_t nodemask = node_to_cpumask(i);
6463 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6465 cpus_and(nodemask, nodemask, *cpu_map);
6466 if (cpus_empty(nodemask))
6476 if (oldsg != sched_group_nodes[i])
6479 kfree(sched_group_nodes);
6480 sched_group_nodes_bycpu[cpu] = NULL;
6484 static void free_sched_groups(const cpumask_t *cpu_map)
6490 * Initialize sched groups cpu_power.
6492 * cpu_power indicates the capacity of sched group, which is used while
6493 * distributing the load between different sched groups in a sched domain.
6494 * Typically cpu_power for all the groups in a sched domain will be same unless
6495 * there are asymmetries in the topology. If there are asymmetries, group
6496 * having more cpu_power will pickup more load compared to the group having
6499 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6500 * the maximum number of tasks a group can handle in the presence of other idle
6501 * or lightly loaded groups in the same sched domain.
6503 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6505 struct sched_domain *child;
6506 struct sched_group *group;
6508 WARN_ON(!sd || !sd->groups);
6510 if (cpu != first_cpu(sd->groups->cpumask))
6515 sd->groups->__cpu_power = 0;
6518 * For perf policy, if the groups in child domain share resources
6519 * (for example cores sharing some portions of the cache hierarchy
6520 * or SMT), then set this domain groups cpu_power such that each group
6521 * can handle only one task, when there are other idle groups in the
6522 * same sched domain.
6524 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6526 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6527 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6532 * add cpu_power of each child group to this groups cpu_power
6534 group = child->groups;
6536 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6537 group = group->next;
6538 } while (group != child->groups);
6542 * Build sched domains for a given set of cpus and attach the sched domains
6543 * to the individual cpus
6545 static int build_sched_domains(const cpumask_t *cpu_map)
6548 struct root_domain *rd;
6550 struct sched_group **sched_group_nodes = NULL;
6551 int sd_allnodes = 0;
6554 * Allocate the per-node list of sched groups
6556 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6558 if (!sched_group_nodes) {
6559 printk(KERN_WARNING "Can not alloc sched group node list\n");
6562 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6565 rd = alloc_rootdomain();
6567 printk(KERN_WARNING "Cannot alloc root domain\n");
6572 * Set up domains for cpus specified by the cpu_map.
6574 for_each_cpu_mask(i, *cpu_map) {
6575 struct sched_domain *sd = NULL, *p;
6576 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6578 cpus_and(nodemask, nodemask, *cpu_map);
6581 if (cpus_weight(*cpu_map) >
6582 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6583 sd = &per_cpu(allnodes_domains, i);
6584 *sd = SD_ALLNODES_INIT;
6585 sd->span = *cpu_map;
6586 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6592 sd = &per_cpu(node_domains, i);
6594 sd->span = sched_domain_node_span(cpu_to_node(i));
6598 cpus_and(sd->span, sd->span, *cpu_map);
6602 sd = &per_cpu(phys_domains, i);
6604 sd->span = nodemask;
6608 cpu_to_phys_group(i, cpu_map, &sd->groups);
6610 #ifdef CONFIG_SCHED_MC
6612 sd = &per_cpu(core_domains, i);
6614 sd->span = cpu_coregroup_map(i);
6615 cpus_and(sd->span, sd->span, *cpu_map);
6618 cpu_to_core_group(i, cpu_map, &sd->groups);
6621 #ifdef CONFIG_SCHED_SMT
6623 sd = &per_cpu(cpu_domains, i);
6624 *sd = SD_SIBLING_INIT;
6625 sd->span = per_cpu(cpu_sibling_map, i);
6626 cpus_and(sd->span, sd->span, *cpu_map);
6629 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6633 #ifdef CONFIG_SCHED_SMT
6634 /* Set up CPU (sibling) groups */
6635 for_each_cpu_mask(i, *cpu_map) {
6636 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6637 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6638 if (i != first_cpu(this_sibling_map))
6641 init_sched_build_groups(this_sibling_map, cpu_map,
6646 #ifdef CONFIG_SCHED_MC
6647 /* Set up multi-core groups */
6648 for_each_cpu_mask(i, *cpu_map) {
6649 cpumask_t this_core_map = cpu_coregroup_map(i);
6650 cpus_and(this_core_map, this_core_map, *cpu_map);
6651 if (i != first_cpu(this_core_map))
6653 init_sched_build_groups(this_core_map, cpu_map,
6654 &cpu_to_core_group);
6658 /* Set up physical groups */
6659 for (i = 0; i < MAX_NUMNODES; i++) {
6660 cpumask_t nodemask = node_to_cpumask(i);
6662 cpus_and(nodemask, nodemask, *cpu_map);
6663 if (cpus_empty(nodemask))
6666 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6670 /* Set up node groups */
6672 init_sched_build_groups(*cpu_map, cpu_map,
6673 &cpu_to_allnodes_group);
6675 for (i = 0; i < MAX_NUMNODES; i++) {
6676 /* Set up node groups */
6677 struct sched_group *sg, *prev;
6678 cpumask_t nodemask = node_to_cpumask(i);
6679 cpumask_t domainspan;
6680 cpumask_t covered = CPU_MASK_NONE;
6683 cpus_and(nodemask, nodemask, *cpu_map);
6684 if (cpus_empty(nodemask)) {
6685 sched_group_nodes[i] = NULL;
6689 domainspan = sched_domain_node_span(i);
6690 cpus_and(domainspan, domainspan, *cpu_map);
6692 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6694 printk(KERN_WARNING "Can not alloc domain group for "
6698 sched_group_nodes[i] = sg;
6699 for_each_cpu_mask(j, nodemask) {
6700 struct sched_domain *sd;
6702 sd = &per_cpu(node_domains, j);
6705 sg->__cpu_power = 0;
6706 sg->cpumask = nodemask;
6708 cpus_or(covered, covered, nodemask);
6711 for (j = 0; j < MAX_NUMNODES; j++) {
6712 cpumask_t tmp, notcovered;
6713 int n = (i + j) % MAX_NUMNODES;
6715 cpus_complement(notcovered, covered);
6716 cpus_and(tmp, notcovered, *cpu_map);
6717 cpus_and(tmp, tmp, domainspan);
6718 if (cpus_empty(tmp))
6721 nodemask = node_to_cpumask(n);
6722 cpus_and(tmp, tmp, nodemask);
6723 if (cpus_empty(tmp))
6726 sg = kmalloc_node(sizeof(struct sched_group),
6730 "Can not alloc domain group for node %d\n", j);
6733 sg->__cpu_power = 0;
6735 sg->next = prev->next;
6736 cpus_or(covered, covered, tmp);
6743 /* Calculate CPU power for physical packages and nodes */
6744 #ifdef CONFIG_SCHED_SMT
6745 for_each_cpu_mask(i, *cpu_map) {
6746 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6748 init_sched_groups_power(i, sd);
6751 #ifdef CONFIG_SCHED_MC
6752 for_each_cpu_mask(i, *cpu_map) {
6753 struct sched_domain *sd = &per_cpu(core_domains, i);
6755 init_sched_groups_power(i, sd);
6759 for_each_cpu_mask(i, *cpu_map) {
6760 struct sched_domain *sd = &per_cpu(phys_domains, i);
6762 init_sched_groups_power(i, sd);
6766 for (i = 0; i < MAX_NUMNODES; i++)
6767 init_numa_sched_groups_power(sched_group_nodes[i]);
6770 struct sched_group *sg;
6772 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6773 init_numa_sched_groups_power(sg);
6777 /* Attach the domains */
6778 for_each_cpu_mask(i, *cpu_map) {
6779 struct sched_domain *sd;
6780 #ifdef CONFIG_SCHED_SMT
6781 sd = &per_cpu(cpu_domains, i);
6782 #elif defined(CONFIG_SCHED_MC)
6783 sd = &per_cpu(core_domains, i);
6785 sd = &per_cpu(phys_domains, i);
6787 cpu_attach_domain(sd, rd, i);
6794 free_sched_groups(cpu_map);
6799 static cpumask_t *doms_cur; /* current sched domains */
6800 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6803 * Special case: If a kmalloc of a doms_cur partition (array of
6804 * cpumask_t) fails, then fallback to a single sched domain,
6805 * as determined by the single cpumask_t fallback_doms.
6807 static cpumask_t fallback_doms;
6810 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6811 * For now this just excludes isolated cpus, but could be used to
6812 * exclude other special cases in the future.
6814 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6819 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6821 doms_cur = &fallback_doms;
6822 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6823 err = build_sched_domains(doms_cur);
6824 register_sched_domain_sysctl();
6829 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6831 free_sched_groups(cpu_map);
6835 * Detach sched domains from a group of cpus specified in cpu_map
6836 * These cpus will now be attached to the NULL domain
6838 static void detach_destroy_domains(const cpumask_t *cpu_map)
6842 unregister_sched_domain_sysctl();
6844 for_each_cpu_mask(i, *cpu_map)
6845 cpu_attach_domain(NULL, &def_root_domain, i);
6846 synchronize_sched();
6847 arch_destroy_sched_domains(cpu_map);
6851 * Partition sched domains as specified by the 'ndoms_new'
6852 * cpumasks in the array doms_new[] of cpumasks. This compares
6853 * doms_new[] to the current sched domain partitioning, doms_cur[].
6854 * It destroys each deleted domain and builds each new domain.
6856 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6857 * The masks don't intersect (don't overlap.) We should setup one
6858 * sched domain for each mask. CPUs not in any of the cpumasks will
6859 * not be load balanced. If the same cpumask appears both in the
6860 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6863 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6864 * ownership of it and will kfree it when done with it. If the caller
6865 * failed the kmalloc call, then it can pass in doms_new == NULL,
6866 * and partition_sched_domains() will fallback to the single partition
6869 * Call with hotplug lock held
6871 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6877 /* always unregister in case we don't destroy any domains */
6878 unregister_sched_domain_sysctl();
6880 if (doms_new == NULL) {
6882 doms_new = &fallback_doms;
6883 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6886 /* Destroy deleted domains */
6887 for (i = 0; i < ndoms_cur; i++) {
6888 for (j = 0; j < ndoms_new; j++) {
6889 if (cpus_equal(doms_cur[i], doms_new[j]))
6892 /* no match - a current sched domain not in new doms_new[] */
6893 detach_destroy_domains(doms_cur + i);
6898 /* Build new domains */
6899 for (i = 0; i < ndoms_new; i++) {
6900 for (j = 0; j < ndoms_cur; j++) {
6901 if (cpus_equal(doms_new[i], doms_cur[j]))
6904 /* no match - add a new doms_new */
6905 build_sched_domains(doms_new + i);
6910 /* Remember the new sched domains */
6911 if (doms_cur != &fallback_doms)
6913 doms_cur = doms_new;
6914 ndoms_cur = ndoms_new;
6916 register_sched_domain_sysctl();
6921 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6922 static int arch_reinit_sched_domains(void)
6927 detach_destroy_domains(&cpu_online_map);
6928 err = arch_init_sched_domains(&cpu_online_map);
6934 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6938 if (buf[0] != '0' && buf[0] != '1')
6942 sched_smt_power_savings = (buf[0] == '1');
6944 sched_mc_power_savings = (buf[0] == '1');
6946 ret = arch_reinit_sched_domains();
6948 return ret ? ret : count;
6951 #ifdef CONFIG_SCHED_MC
6952 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6954 return sprintf(page, "%u\n", sched_mc_power_savings);
6956 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6957 const char *buf, size_t count)
6959 return sched_power_savings_store(buf, count, 0);
6961 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6962 sched_mc_power_savings_store);
6965 #ifdef CONFIG_SCHED_SMT
6966 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6968 return sprintf(page, "%u\n", sched_smt_power_savings);
6970 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6971 const char *buf, size_t count)
6973 return sched_power_savings_store(buf, count, 1);
6975 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6976 sched_smt_power_savings_store);
6979 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6983 #ifdef CONFIG_SCHED_SMT
6985 err = sysfs_create_file(&cls->kset.kobj,
6986 &attr_sched_smt_power_savings.attr);
6988 #ifdef CONFIG_SCHED_MC
6989 if (!err && mc_capable())
6990 err = sysfs_create_file(&cls->kset.kobj,
6991 &attr_sched_mc_power_savings.attr);
6998 * Force a reinitialization of the sched domains hierarchy. The domains
6999 * and groups cannot be updated in place without racing with the balancing
7000 * code, so we temporarily attach all running cpus to the NULL domain
7001 * which will prevent rebalancing while the sched domains are recalculated.
7003 static int update_sched_domains(struct notifier_block *nfb,
7004 unsigned long action, void *hcpu)
7007 case CPU_UP_PREPARE:
7008 case CPU_UP_PREPARE_FROZEN:
7009 case CPU_DOWN_PREPARE:
7010 case CPU_DOWN_PREPARE_FROZEN:
7011 detach_destroy_domains(&cpu_online_map);
7014 case CPU_UP_CANCELED:
7015 case CPU_UP_CANCELED_FROZEN:
7016 case CPU_DOWN_FAILED:
7017 case CPU_DOWN_FAILED_FROZEN:
7019 case CPU_ONLINE_FROZEN:
7021 case CPU_DEAD_FROZEN:
7023 * Fall through and re-initialise the domains.
7030 /* The hotplug lock is already held by cpu_up/cpu_down */
7031 arch_init_sched_domains(&cpu_online_map);
7036 void __init sched_init_smp(void)
7038 cpumask_t non_isolated_cpus;
7041 arch_init_sched_domains(&cpu_online_map);
7042 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7043 if (cpus_empty(non_isolated_cpus))
7044 cpu_set(smp_processor_id(), non_isolated_cpus);
7046 /* XXX: Theoretical race here - CPU may be hotplugged now */
7047 hotcpu_notifier(update_sched_domains, 0);
7049 /* Move init over to a non-isolated CPU */
7050 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7052 sched_init_granularity();
7054 #ifdef CONFIG_FAIR_GROUP_SCHED
7055 if (nr_cpu_ids == 1)
7058 lb_monitor_task = kthread_create(load_balance_monitor, NULL,
7060 if (!IS_ERR(lb_monitor_task)) {
7061 lb_monitor_task->flags |= PF_NOFREEZE;
7062 wake_up_process(lb_monitor_task);
7064 printk(KERN_ERR "Could not create load balance monitor thread"
7065 "(error = %ld) \n", PTR_ERR(lb_monitor_task));
7070 void __init sched_init_smp(void)
7072 sched_init_granularity();
7074 #endif /* CONFIG_SMP */
7076 int in_sched_functions(unsigned long addr)
7078 return in_lock_functions(addr) ||
7079 (addr >= (unsigned long)__sched_text_start
7080 && addr < (unsigned long)__sched_text_end);
7083 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7085 cfs_rq->tasks_timeline = RB_ROOT;
7086 #ifdef CONFIG_FAIR_GROUP_SCHED
7089 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7092 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7094 struct rt_prio_array *array;
7097 array = &rt_rq->active;
7098 for (i = 0; i < MAX_RT_PRIO; i++) {
7099 INIT_LIST_HEAD(array->queue + i);
7100 __clear_bit(i, array->bitmap);
7102 /* delimiter for bitsearch: */
7103 __set_bit(MAX_RT_PRIO, array->bitmap);
7105 #if defined CONFIG_SMP || defined CONFIG_FAIR_GROUP_SCHED
7106 rt_rq->highest_prio = MAX_RT_PRIO;
7109 rt_rq->rt_nr_migratory = 0;
7110 rt_rq->overloaded = 0;
7114 rt_rq->rt_throttled = 0;
7116 #ifdef CONFIG_FAIR_GROUP_SCHED
7117 rt_rq->rt_nr_boosted = 0;
7122 #ifdef CONFIG_FAIR_GROUP_SCHED
7123 static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7124 struct cfs_rq *cfs_rq, struct sched_entity *se,
7127 tg->cfs_rq[cpu] = cfs_rq;
7128 init_cfs_rq(cfs_rq, rq);
7131 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7134 se->cfs_rq = &rq->cfs;
7136 se->load.weight = tg->shares;
7137 se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7141 static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7142 struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7145 tg->rt_rq[cpu] = rt_rq;
7146 init_rt_rq(rt_rq, rq);
7148 rt_rq->rt_se = rt_se;
7150 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7152 tg->rt_se[cpu] = rt_se;
7153 rt_se->rt_rq = &rq->rt;
7154 rt_se->my_q = rt_rq;
7155 rt_se->parent = NULL;
7156 INIT_LIST_HEAD(&rt_se->run_list);
7160 void __init sched_init(void)
7162 int highest_cpu = 0;
7166 init_defrootdomain();
7169 #ifdef CONFIG_FAIR_GROUP_SCHED
7170 list_add(&init_task_group.list, &task_groups);
7173 for_each_possible_cpu(i) {
7177 spin_lock_init(&rq->lock);
7178 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7181 init_cfs_rq(&rq->cfs, rq);
7182 init_rt_rq(&rq->rt, rq);
7183 #ifdef CONFIG_FAIR_GROUP_SCHED
7184 init_task_group.shares = init_task_group_load;
7185 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7186 init_tg_cfs_entry(rq, &init_task_group,
7187 &per_cpu(init_cfs_rq, i),
7188 &per_cpu(init_sched_entity, i), i, 1);
7190 init_task_group.rt_ratio = sysctl_sched_rt_ratio; /* XXX */
7191 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7192 init_tg_rt_entry(rq, &init_task_group,
7193 &per_cpu(init_rt_rq, i),
7194 &per_cpu(init_sched_rt_entity, i), i, 1);
7196 rq->rt_period_expire = 0;
7197 rq->rt_throttled = 0;
7199 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7200 rq->cpu_load[j] = 0;
7204 rq->active_balance = 0;
7205 rq->next_balance = jiffies;
7208 rq->migration_thread = NULL;
7209 INIT_LIST_HEAD(&rq->migration_queue);
7210 rq_attach_root(rq, &def_root_domain);
7213 atomic_set(&rq->nr_iowait, 0);
7217 set_load_weight(&init_task);
7219 #ifdef CONFIG_PREEMPT_NOTIFIERS
7220 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7224 nr_cpu_ids = highest_cpu + 1;
7225 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7228 #ifdef CONFIG_RT_MUTEXES
7229 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7233 * The boot idle thread does lazy MMU switching as well:
7235 atomic_inc(&init_mm.mm_count);
7236 enter_lazy_tlb(&init_mm, current);
7239 * Make us the idle thread. Technically, schedule() should not be
7240 * called from this thread, however somewhere below it might be,
7241 * but because we are the idle thread, we just pick up running again
7242 * when this runqueue becomes "idle".
7244 init_idle(current, smp_processor_id());
7246 * During early bootup we pretend to be a normal task:
7248 current->sched_class = &fair_sched_class;
7251 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7252 void __might_sleep(char *file, int line)
7255 static unsigned long prev_jiffy; /* ratelimiting */
7257 if ((in_atomic() || irqs_disabled()) &&
7258 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7259 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7261 prev_jiffy = jiffies;
7262 printk(KERN_ERR "BUG: sleeping function called from invalid"
7263 " context at %s:%d\n", file, line);
7264 printk("in_atomic():%d, irqs_disabled():%d\n",
7265 in_atomic(), irqs_disabled());
7266 debug_show_held_locks(current);
7267 if (irqs_disabled())
7268 print_irqtrace_events(current);
7273 EXPORT_SYMBOL(__might_sleep);
7276 #ifdef CONFIG_MAGIC_SYSRQ
7277 static void normalize_task(struct rq *rq, struct task_struct *p)
7280 update_rq_clock(rq);
7281 on_rq = p->se.on_rq;
7283 deactivate_task(rq, p, 0);
7284 __setscheduler(rq, p, SCHED_NORMAL, 0);
7286 activate_task(rq, p, 0);
7287 resched_task(rq->curr);
7291 void normalize_rt_tasks(void)
7293 struct task_struct *g, *p;
7294 unsigned long flags;
7297 read_lock_irqsave(&tasklist_lock, flags);
7298 do_each_thread(g, p) {
7300 * Only normalize user tasks:
7305 p->se.exec_start = 0;
7306 #ifdef CONFIG_SCHEDSTATS
7307 p->se.wait_start = 0;
7308 p->se.sleep_start = 0;
7309 p->se.block_start = 0;
7311 task_rq(p)->clock = 0;
7315 * Renice negative nice level userspace
7318 if (TASK_NICE(p) < 0 && p->mm)
7319 set_user_nice(p, 0);
7323 spin_lock(&p->pi_lock);
7324 rq = __task_rq_lock(p);
7326 normalize_task(rq, p);
7328 __task_rq_unlock(rq);
7329 spin_unlock(&p->pi_lock);
7330 } while_each_thread(g, p);
7332 read_unlock_irqrestore(&tasklist_lock, flags);
7335 #endif /* CONFIG_MAGIC_SYSRQ */
7339 * These functions are only useful for the IA64 MCA handling.
7341 * They can only be called when the whole system has been
7342 * stopped - every CPU needs to be quiescent, and no scheduling
7343 * activity can take place. Using them for anything else would
7344 * be a serious bug, and as a result, they aren't even visible
7345 * under any other configuration.
7349 * curr_task - return the current task for a given cpu.
7350 * @cpu: the processor in question.
7352 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7354 struct task_struct *curr_task(int cpu)
7356 return cpu_curr(cpu);
7360 * set_curr_task - set the current task for a given cpu.
7361 * @cpu: the processor in question.
7362 * @p: the task pointer to set.
7364 * Description: This function must only be used when non-maskable interrupts
7365 * are serviced on a separate stack. It allows the architecture to switch the
7366 * notion of the current task on a cpu in a non-blocking manner. This function
7367 * must be called with all CPU's synchronized, and interrupts disabled, the
7368 * and caller must save the original value of the current task (see
7369 * curr_task() above) and restore that value before reenabling interrupts and
7370 * re-starting the system.
7372 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7374 void set_curr_task(int cpu, struct task_struct *p)
7381 #ifdef CONFIG_FAIR_GROUP_SCHED
7385 * distribute shares of all task groups among their schedulable entities,
7386 * to reflect load distribution across cpus.
7388 static int rebalance_shares(struct sched_domain *sd, int this_cpu)
7390 struct cfs_rq *cfs_rq;
7391 struct rq *rq = cpu_rq(this_cpu);
7392 cpumask_t sdspan = sd->span;
7395 /* Walk thr' all the task groups that we have */
7396 for_each_leaf_cfs_rq(rq, cfs_rq) {
7398 unsigned long total_load = 0, total_shares;
7399 struct task_group *tg = cfs_rq->tg;
7401 /* Gather total task load of this group across cpus */
7402 for_each_cpu_mask(i, sdspan)
7403 total_load += tg->cfs_rq[i]->load.weight;
7405 /* Nothing to do if this group has no load */
7410 * tg->shares represents the number of cpu shares the task group
7411 * is eligible to hold on a single cpu. On N cpus, it is
7412 * eligible to hold (N * tg->shares) number of cpu shares.
7414 total_shares = tg->shares * cpus_weight(sdspan);
7417 * redistribute total_shares across cpus as per the task load
7420 for_each_cpu_mask(i, sdspan) {
7421 unsigned long local_load, local_shares;
7423 local_load = tg->cfs_rq[i]->load.weight;
7424 local_shares = (local_load * total_shares) / total_load;
7426 local_shares = MIN_GROUP_SHARES;
7427 if (local_shares == tg->se[i]->load.weight)
7430 spin_lock_irq(&cpu_rq(i)->lock);
7431 set_se_shares(tg->se[i], local_shares);
7432 spin_unlock_irq(&cpu_rq(i)->lock);
7441 * How frequently should we rebalance_shares() across cpus?
7443 * The more frequently we rebalance shares, the more accurate is the fairness
7444 * of cpu bandwidth distribution between task groups. However higher frequency
7445 * also implies increased scheduling overhead.
7447 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7448 * consecutive calls to rebalance_shares() in the same sched domain.
7450 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7451 * consecutive calls to rebalance_shares() in the same sched domain.
7453 * These settings allows for the appropriate trade-off between accuracy of
7454 * fairness and the associated overhead.
7458 /* default: 8ms, units: milliseconds */
7459 const_debug unsigned int sysctl_sched_min_bal_int_shares = 8;
7461 /* default: 128ms, units: milliseconds */
7462 const_debug unsigned int sysctl_sched_max_bal_int_shares = 128;
7464 /* kernel thread that runs rebalance_shares() periodically */
7465 static int load_balance_monitor(void *unused)
7467 unsigned int timeout = sysctl_sched_min_bal_int_shares;
7468 struct sched_param schedparm;
7472 * We don't want this thread's execution to be limited by the shares
7473 * assigned to default group (init_task_group). Hence make it run
7474 * as a SCHED_RR RT task at the lowest priority.
7476 schedparm.sched_priority = 1;
7477 ret = sched_setscheduler(current, SCHED_RR, &schedparm);
7479 printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance"
7480 " monitor thread (error = %d) \n", ret);
7482 while (!kthread_should_stop()) {
7483 int i, cpu, balanced = 1;
7485 /* Prevent cpus going down or coming up */
7487 /* lockout changes to doms_cur[] array */
7490 * Enter a rcu read-side critical section to safely walk rq->sd
7491 * chain on various cpus and to walk task group list
7492 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7496 for (i = 0; i < ndoms_cur; i++) {
7497 cpumask_t cpumap = doms_cur[i];
7498 struct sched_domain *sd = NULL, *sd_prev = NULL;
7500 cpu = first_cpu(cpumap);
7502 /* Find the highest domain at which to balance shares */
7503 for_each_domain(cpu, sd) {
7504 if (!(sd->flags & SD_LOAD_BALANCE))
7510 /* sd == NULL? No load balance reqd in this domain */
7514 balanced &= rebalance_shares(sd, cpu);
7523 timeout = sysctl_sched_min_bal_int_shares;
7524 else if (timeout < sysctl_sched_max_bal_int_shares)
7527 msleep_interruptible(timeout);
7532 #endif /* CONFIG_SMP */
7534 static void free_sched_group(struct task_group *tg)
7538 for_each_possible_cpu(i) {
7540 kfree(tg->cfs_rq[i]);
7544 kfree(tg->rt_rq[i]);
7546 kfree(tg->rt_se[i]);
7556 /* allocate runqueue etc for a new task group */
7557 struct task_group *sched_create_group(void)
7559 struct task_group *tg;
7560 struct cfs_rq *cfs_rq;
7561 struct sched_entity *se;
7562 struct rt_rq *rt_rq;
7563 struct sched_rt_entity *rt_se;
7565 unsigned long flags;
7568 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7570 return ERR_PTR(-ENOMEM);
7572 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7575 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7578 tg->rt_rq = kzalloc(sizeof(rt_rq) * NR_CPUS, GFP_KERNEL);
7581 tg->rt_se = kzalloc(sizeof(rt_se) * NR_CPUS, GFP_KERNEL);
7585 tg->shares = NICE_0_LOAD;
7586 tg->rt_ratio = 0; /* XXX */
7588 for_each_possible_cpu(i) {
7591 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7592 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7596 se = kmalloc_node(sizeof(struct sched_entity),
7597 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7601 rt_rq = kmalloc_node(sizeof(struct rt_rq),
7602 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7606 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7607 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7611 init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7612 init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7615 spin_lock_irqsave(&task_group_lock, flags);
7616 for_each_possible_cpu(i) {
7618 cfs_rq = tg->cfs_rq[i];
7619 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7620 rt_rq = tg->rt_rq[i];
7621 list_add_rcu(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7623 list_add_rcu(&tg->list, &task_groups);
7624 spin_unlock_irqrestore(&task_group_lock, flags);
7629 free_sched_group(tg);
7630 return ERR_PTR(-ENOMEM);
7633 /* rcu callback to free various structures associated with a task group */
7634 static void free_sched_group_rcu(struct rcu_head *rhp)
7636 /* now it should be safe to free those cfs_rqs */
7637 free_sched_group(container_of(rhp, struct task_group, rcu));
7640 /* Destroy runqueue etc associated with a task group */
7641 void sched_destroy_group(struct task_group *tg)
7643 struct cfs_rq *cfs_rq = NULL;
7644 struct rt_rq *rt_rq = NULL;
7645 unsigned long flags;
7648 spin_lock_irqsave(&task_group_lock, flags);
7649 for_each_possible_cpu(i) {
7650 cfs_rq = tg->cfs_rq[i];
7651 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7652 rt_rq = tg->rt_rq[i];
7653 list_del_rcu(&rt_rq->leaf_rt_rq_list);
7655 list_del_rcu(&tg->list);
7656 spin_unlock_irqrestore(&task_group_lock, flags);
7660 /* wait for possible concurrent references to cfs_rqs complete */
7661 call_rcu(&tg->rcu, free_sched_group_rcu);
7664 /* change task's runqueue when it moves between groups.
7665 * The caller of this function should have put the task in its new group
7666 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7667 * reflect its new group.
7669 void sched_move_task(struct task_struct *tsk)
7672 unsigned long flags;
7675 rq = task_rq_lock(tsk, &flags);
7677 update_rq_clock(rq);
7679 running = task_current(rq, tsk);
7680 on_rq = tsk->se.on_rq;
7683 dequeue_task(rq, tsk, 0);
7684 if (unlikely(running))
7685 tsk->sched_class->put_prev_task(rq, tsk);
7688 set_task_rq(tsk, task_cpu(tsk));
7691 if (unlikely(running))
7692 tsk->sched_class->set_curr_task(rq);
7693 enqueue_task(rq, tsk, 0);
7696 task_rq_unlock(rq, &flags);
7699 /* rq->lock to be locked by caller */
7700 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7702 struct cfs_rq *cfs_rq = se->cfs_rq;
7703 struct rq *rq = cfs_rq->rq;
7707 shares = MIN_GROUP_SHARES;
7711 dequeue_entity(cfs_rq, se, 0);
7712 dec_cpu_load(rq, se->load.weight);
7715 se->load.weight = shares;
7716 se->load.inv_weight = div64_64((1ULL<<32), shares);
7719 enqueue_entity(cfs_rq, se, 0);
7720 inc_cpu_load(rq, se->load.weight);
7724 static DEFINE_MUTEX(shares_mutex);
7726 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7729 struct cfs_rq *cfs_rq;
7731 unsigned long flags;
7733 mutex_lock(&shares_mutex);
7734 if (tg->shares == shares)
7737 if (shares < MIN_GROUP_SHARES)
7738 shares = MIN_GROUP_SHARES;
7741 * Prevent any load balance activity (rebalance_shares,
7742 * load_balance_fair) from referring to this group first,
7743 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7745 spin_lock_irqsave(&task_group_lock, flags);
7746 for_each_possible_cpu(i) {
7747 cfs_rq = tg->cfs_rq[i];
7748 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7750 spin_unlock_irqrestore(&task_group_lock, flags);
7752 /* wait for any ongoing reference to this group to finish */
7753 synchronize_sched();
7756 * Now we are free to modify the group's share on each cpu
7757 * w/o tripping rebalance_share or load_balance_fair.
7759 tg->shares = shares;
7760 for_each_possible_cpu(i) {
7761 spin_lock_irq(&cpu_rq(i)->lock);
7762 set_se_shares(tg->se[i], shares);
7763 spin_unlock_irq(&cpu_rq(i)->lock);
7767 * Enable load balance activity on this group, by inserting it back on
7768 * each cpu's rq->leaf_cfs_rq_list.
7770 spin_lock_irqsave(&task_group_lock, flags);
7771 for_each_possible_cpu(i) {
7773 cfs_rq = tg->cfs_rq[i];
7774 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7776 spin_unlock_irqrestore(&task_group_lock, flags);
7778 mutex_unlock(&shares_mutex);
7782 unsigned long sched_group_shares(struct task_group *tg)
7788 * Ensure the total rt_ratio <= sysctl_sched_rt_ratio
7790 int sched_group_set_rt_ratio(struct task_group *tg, unsigned long rt_ratio)
7792 struct task_group *tgi;
7793 unsigned long total = 0;
7796 list_for_each_entry_rcu(tgi, &task_groups, list)
7797 total += tgi->rt_ratio;
7800 if (total + rt_ratio - tg->rt_ratio > sysctl_sched_rt_ratio)
7803 tg->rt_ratio = rt_ratio;
7807 unsigned long sched_group_rt_ratio(struct task_group *tg)
7809 return tg->rt_ratio;
7812 #endif /* CONFIG_FAIR_GROUP_SCHED */
7814 #ifdef CONFIG_FAIR_CGROUP_SCHED
7816 /* return corresponding task_group object of a cgroup */
7817 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7819 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7820 struct task_group, css);
7823 static struct cgroup_subsys_state *
7824 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7826 struct task_group *tg;
7828 if (!cgrp->parent) {
7829 /* This is early initialization for the top cgroup */
7830 init_task_group.css.cgroup = cgrp;
7831 return &init_task_group.css;
7834 /* we support only 1-level deep hierarchical scheduler atm */
7835 if (cgrp->parent->parent)
7836 return ERR_PTR(-EINVAL);
7838 tg = sched_create_group();
7840 return ERR_PTR(-ENOMEM);
7842 /* Bind the cgroup to task_group object we just created */
7843 tg->css.cgroup = cgrp;
7849 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7851 struct task_group *tg = cgroup_tg(cgrp);
7853 sched_destroy_group(tg);
7857 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7858 struct task_struct *tsk)
7860 /* We don't support RT-tasks being in separate groups */
7861 if (tsk->sched_class != &fair_sched_class)
7868 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7869 struct cgroup *old_cont, struct task_struct *tsk)
7871 sched_move_task(tsk);
7874 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7877 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7880 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7882 struct task_group *tg = cgroup_tg(cgrp);
7884 return (u64) tg->shares;
7887 static int cpu_rt_ratio_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7890 return sched_group_set_rt_ratio(cgroup_tg(cgrp), rt_ratio_val);
7893 static u64 cpu_rt_ratio_read_uint(struct cgroup *cgrp, struct cftype *cft)
7895 struct task_group *tg = cgroup_tg(cgrp);
7897 return (u64) tg->rt_ratio;
7900 static struct cftype cpu_files[] = {
7903 .read_uint = cpu_shares_read_uint,
7904 .write_uint = cpu_shares_write_uint,
7908 .read_uint = cpu_rt_ratio_read_uint,
7909 .write_uint = cpu_rt_ratio_write_uint,
7913 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7915 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7918 struct cgroup_subsys cpu_cgroup_subsys = {
7920 .create = cpu_cgroup_create,
7921 .destroy = cpu_cgroup_destroy,
7922 .can_attach = cpu_cgroup_can_attach,
7923 .attach = cpu_cgroup_attach,
7924 .populate = cpu_cgroup_populate,
7925 .subsys_id = cpu_cgroup_subsys_id,
7929 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7931 #ifdef CONFIG_CGROUP_CPUACCT
7934 * CPU accounting code for task groups.
7936 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7937 * (balbir@in.ibm.com).
7940 /* track cpu usage of a group of tasks */
7942 struct cgroup_subsys_state css;
7943 /* cpuusage holds pointer to a u64-type object on every cpu */
7947 struct cgroup_subsys cpuacct_subsys;
7949 /* return cpu accounting group corresponding to this container */
7950 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7952 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7953 struct cpuacct, css);
7956 /* return cpu accounting group to which this task belongs */
7957 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7959 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7960 struct cpuacct, css);
7963 /* create a new cpu accounting group */
7964 static struct cgroup_subsys_state *cpuacct_create(
7965 struct cgroup_subsys *ss, struct cgroup *cont)
7967 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7970 return ERR_PTR(-ENOMEM);
7972 ca->cpuusage = alloc_percpu(u64);
7973 if (!ca->cpuusage) {
7975 return ERR_PTR(-ENOMEM);
7981 /* destroy an existing cpu accounting group */
7983 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
7985 struct cpuacct *ca = cgroup_ca(cont);
7987 free_percpu(ca->cpuusage);
7991 /* return total cpu usage (in nanoseconds) of a group */
7992 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7994 struct cpuacct *ca = cgroup_ca(cont);
7995 u64 totalcpuusage = 0;
7998 for_each_possible_cpu(i) {
7999 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8002 * Take rq->lock to make 64-bit addition safe on 32-bit
8005 spin_lock_irq(&cpu_rq(i)->lock);
8006 totalcpuusage += *cpuusage;
8007 spin_unlock_irq(&cpu_rq(i)->lock);
8010 return totalcpuusage;
8013 static struct cftype files[] = {
8016 .read_uint = cpuusage_read,
8020 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8022 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
8026 * charge this task's execution time to its accounting group.
8028 * called with rq->lock held.
8030 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8034 if (!cpuacct_subsys.active)
8039 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8041 *cpuusage += cputime;
8045 struct cgroup_subsys cpuacct_subsys = {
8047 .create = cpuacct_create,
8048 .destroy = cpuacct_destroy,
8049 .populate = cpuacct_populate,
8050 .subsys_id = cpuacct_subsys_id,
8052 #endif /* CONFIG_CGROUP_CPUACCT */