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>
70 #include <asm/irq_regs.h>
73 * Scheduler clock - returns current time in nanosec units.
74 * This is default implementation.
75 * Architectures and sub-architectures can override this.
77 unsigned long long __attribute__((weak)) sched_clock(void)
79 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
119 * Since cpu_power is a 'constant', we can use a reciprocal divide.
121 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
123 return reciprocal_divide(load, sg->reciprocal_cpu_power);
127 * Each time a sched group cpu_power is changed,
128 * we must compute its reciprocal value
130 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
132 sg->__cpu_power += val;
133 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
137 static inline int rt_policy(int policy)
139 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
144 static inline int task_has_rt_policy(struct task_struct *p)
146 return rt_policy(p->policy);
150 * This is the priority-queue data structure of the RT scheduling class:
152 struct rt_prio_array {
153 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
154 struct list_head queue[MAX_RT_PRIO];
157 #ifdef CONFIG_FAIR_GROUP_SCHED
159 #include <linux/cgroup.h>
163 /* task group related information */
165 #ifdef CONFIG_FAIR_CGROUP_SCHED
166 struct cgroup_subsys_state css;
168 /* schedulable entities of this group on each cpu */
169 struct sched_entity **se;
170 /* runqueue "owned" by this group on each cpu */
171 struct cfs_rq **cfs_rq;
174 * shares assigned to a task group governs how much of cpu bandwidth
175 * is allocated to the group. The more shares a group has, the more is
176 * the cpu bandwidth allocated to it.
178 * For ex, lets say that there are three task groups, A, B and C which
179 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
180 * cpu bandwidth allocated by the scheduler to task groups A, B and C
183 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
184 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
185 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
187 * The weight assigned to a task group's schedulable entities on every
188 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
189 * group's shares. For ex: lets say that task group A has been
190 * assigned shares of 1000 and there are two CPUs in a system. Then,
192 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
194 * Note: It's not necessary that each of a task's group schedulable
195 * entity have the same weight on all CPUs. If the group
196 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
197 * better distribution of weight could be:
199 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
200 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
202 * rebalance_shares() is responsible for distributing the shares of a
203 * task groups like this among the group's schedulable entities across
207 unsigned long shares;
212 /* Default task group's sched entity on each cpu */
213 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
214 /* Default task group's cfs_rq on each cpu */
215 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
217 static struct sched_entity *init_sched_entity_p[NR_CPUS];
218 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
220 /* task_group_mutex serializes add/remove of task groups and also changes to
221 * a task group's cpu shares.
223 static DEFINE_MUTEX(task_group_mutex);
225 /* doms_cur_mutex serializes access to doms_cur[] array */
226 static DEFINE_MUTEX(doms_cur_mutex);
229 /* kernel thread that runs rebalance_shares() periodically */
230 static struct task_struct *lb_monitor_task;
231 static int load_balance_monitor(void *unused);
234 static void set_se_shares(struct sched_entity *se, unsigned long shares);
236 /* Default task group.
237 * Every task in system belong to this group at bootup.
239 struct task_group init_task_group = {
240 .se = init_sched_entity_p,
241 .cfs_rq = init_cfs_rq_p,
244 #ifdef CONFIG_FAIR_USER_SCHED
245 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
247 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
250 #define MIN_GROUP_SHARES 2
252 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
254 /* return group to which a task belongs */
255 static inline struct task_group *task_group(struct task_struct *p)
257 struct task_group *tg;
259 #ifdef CONFIG_FAIR_USER_SCHED
261 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
262 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
263 struct task_group, css);
265 tg = &init_task_group;
270 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
271 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
273 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
274 p->se.parent = task_group(p)->se[cpu];
277 static inline void lock_task_group_list(void)
279 mutex_lock(&task_group_mutex);
282 static inline void unlock_task_group_list(void)
284 mutex_unlock(&task_group_mutex);
287 static inline void lock_doms_cur(void)
289 mutex_lock(&doms_cur_mutex);
292 static inline void unlock_doms_cur(void)
294 mutex_unlock(&doms_cur_mutex);
299 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
300 static inline void lock_task_group_list(void) { }
301 static inline void unlock_task_group_list(void) { }
302 static inline void lock_doms_cur(void) { }
303 static inline void unlock_doms_cur(void) { }
305 #endif /* CONFIG_FAIR_GROUP_SCHED */
307 /* CFS-related fields in a runqueue */
309 struct load_weight load;
310 unsigned long nr_running;
315 struct rb_root tasks_timeline;
316 struct rb_node *rb_leftmost;
317 struct rb_node *rb_load_balance_curr;
318 /* 'curr' points to currently running entity on this cfs_rq.
319 * It is set to NULL otherwise (i.e when none are currently running).
321 struct sched_entity *curr;
323 unsigned long nr_spread_over;
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
329 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
330 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
331 * (like users, containers etc.)
333 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
334 * list is used during load balance.
336 struct list_head leaf_cfs_rq_list;
337 struct task_group *tg; /* group that "owns" this runqueue */
341 /* Real-Time classes' related field in a runqueue: */
343 struct rt_prio_array active;
344 int rt_load_balance_idx;
345 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
346 unsigned long rt_nr_running;
347 unsigned long rt_nr_migratory;
348 /* highest queued rt task prio */
356 * We add the notion of a root-domain which will be used to define per-domain
357 * variables. Each exclusive cpuset essentially defines an island domain by
358 * fully partitioning the member cpus from any other cpuset. Whenever a new
359 * exclusive cpuset is created, we also create and attach a new root-domain
362 * By default the system creates a single root-domain with all cpus as
363 * members (mimicking the global state we have today).
371 * The "RT overload" flag: it gets set if a CPU has more than
372 * one runnable RT task.
378 static struct root_domain def_root_domain;
383 * This is the main, per-CPU runqueue data structure.
385 * Locking rule: those places that want to lock multiple runqueues
386 * (such as the load balancing or the thread migration code), lock
387 * acquire operations must be ordered by ascending &runqueue.
394 * nr_running and cpu_load should be in the same cacheline because
395 * remote CPUs use both these fields when doing load calculation.
397 unsigned long nr_running;
398 #define CPU_LOAD_IDX_MAX 5
399 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
400 unsigned char idle_at_tick;
402 unsigned char in_nohz_recently;
404 /* capture load from *all* tasks on this cpu: */
405 struct load_weight load;
406 unsigned long nr_load_updates;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 /* list of leaf cfs_rq on this cpu: */
412 struct list_head leaf_cfs_rq_list;
417 * This is part of a global counter where only the total sum
418 * over all CPUs matters. A task can increase this counter on
419 * one CPU and if it got migrated afterwards it may decrease
420 * it on another CPU. Always updated under the runqueue lock:
422 unsigned long nr_uninterruptible;
424 struct task_struct *curr, *idle;
425 unsigned long next_balance;
426 struct mm_struct *prev_mm;
428 u64 clock, prev_clock_raw;
431 unsigned int clock_warps, clock_overflows;
433 unsigned int clock_deep_idle_events;
439 struct root_domain *rd;
440 struct sched_domain *sd;
442 /* For active balancing */
445 /* cpu of this runqueue: */
448 struct task_struct *migration_thread;
449 struct list_head migration_queue;
452 #ifdef CONFIG_SCHEDSTATS
454 struct sched_info rq_sched_info;
456 /* sys_sched_yield() stats */
457 unsigned int yld_exp_empty;
458 unsigned int yld_act_empty;
459 unsigned int yld_both_empty;
460 unsigned int yld_count;
462 /* schedule() stats */
463 unsigned int sched_switch;
464 unsigned int sched_count;
465 unsigned int sched_goidle;
467 /* try_to_wake_up() stats */
468 unsigned int ttwu_count;
469 unsigned int ttwu_local;
472 unsigned int bkl_count;
474 struct lock_class_key rq_lock_key;
477 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
479 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
481 rq->curr->sched_class->check_preempt_curr(rq, p);
484 static inline int cpu_of(struct rq *rq)
494 * Update the per-runqueue clock, as finegrained as the platform can give
495 * us, but without assuming monotonicity, etc.:
497 static void __update_rq_clock(struct rq *rq)
499 u64 prev_raw = rq->prev_clock_raw;
500 u64 now = sched_clock();
501 s64 delta = now - prev_raw;
502 u64 clock = rq->clock;
504 #ifdef CONFIG_SCHED_DEBUG
505 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
508 * Protect against sched_clock() occasionally going backwards:
510 if (unlikely(delta < 0)) {
515 * Catch too large forward jumps too:
517 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
518 if (clock < rq->tick_timestamp + TICK_NSEC)
519 clock = rq->tick_timestamp + TICK_NSEC;
522 rq->clock_overflows++;
524 if (unlikely(delta > rq->clock_max_delta))
525 rq->clock_max_delta = delta;
530 rq->prev_clock_raw = now;
534 static void update_rq_clock(struct rq *rq)
536 if (likely(smp_processor_id() == cpu_of(rq)))
537 __update_rq_clock(rq);
541 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
542 * See detach_destroy_domains: synchronize_sched for details.
544 * The domain tree of any CPU may only be accessed from within
545 * preempt-disabled sections.
547 #define for_each_domain(cpu, __sd) \
548 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
550 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
551 #define this_rq() (&__get_cpu_var(runqueues))
552 #define task_rq(p) cpu_rq(task_cpu(p))
553 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
556 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
558 #ifdef CONFIG_SCHED_DEBUG
559 # define const_debug __read_mostly
561 # define const_debug static const
565 * Debugging: various feature bits
568 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
569 SCHED_FEAT_WAKEUP_PREEMPT = 2,
570 SCHED_FEAT_START_DEBIT = 4,
571 SCHED_FEAT_TREE_AVG = 8,
572 SCHED_FEAT_APPROX_AVG = 16,
575 const_debug unsigned int sysctl_sched_features =
576 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
577 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
578 SCHED_FEAT_START_DEBIT * 1 |
579 SCHED_FEAT_TREE_AVG * 0 |
580 SCHED_FEAT_APPROX_AVG * 0;
582 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
585 * Number of tasks to iterate in a single balance run.
586 * Limited because this is done with IRQs disabled.
588 const_debug unsigned int sysctl_sched_nr_migrate = 32;
591 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
592 * clock constructed from sched_clock():
594 unsigned long long cpu_clock(int cpu)
596 unsigned long long now;
600 local_irq_save(flags);
603 * Only call sched_clock() if the scheduler has already been
604 * initialized (some code might call cpu_clock() very early):
609 local_irq_restore(flags);
613 EXPORT_SYMBOL_GPL(cpu_clock);
615 #ifndef prepare_arch_switch
616 # define prepare_arch_switch(next) do { } while (0)
618 #ifndef finish_arch_switch
619 # define finish_arch_switch(prev) do { } while (0)
622 static inline int task_current(struct rq *rq, struct task_struct *p)
624 return rq->curr == p;
627 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
628 static inline int task_running(struct rq *rq, struct task_struct *p)
630 return task_current(rq, p);
633 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
637 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
639 #ifdef CONFIG_DEBUG_SPINLOCK
640 /* this is a valid case when another task releases the spinlock */
641 rq->lock.owner = current;
644 * If we are tracking spinlock dependencies then we have to
645 * fix up the runqueue lock - which gets 'carried over' from
648 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
650 spin_unlock_irq(&rq->lock);
653 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
654 static inline int task_running(struct rq *rq, struct task_struct *p)
659 return task_current(rq, p);
663 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
667 * We can optimise this out completely for !SMP, because the
668 * SMP rebalancing from interrupt is the only thing that cares
673 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
674 spin_unlock_irq(&rq->lock);
676 spin_unlock(&rq->lock);
680 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
684 * After ->oncpu is cleared, the task can be moved to a different CPU.
685 * We must ensure this doesn't happen until the switch is completely
691 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
695 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
698 * __task_rq_lock - lock the runqueue a given task resides on.
699 * Must be called interrupts disabled.
701 static inline struct rq *__task_rq_lock(struct task_struct *p)
705 struct rq *rq = task_rq(p);
706 spin_lock(&rq->lock);
707 if (likely(rq == task_rq(p)))
709 spin_unlock(&rq->lock);
714 * task_rq_lock - lock the runqueue a given task resides on and disable
715 * interrupts. Note the ordering: we can safely lookup the task_rq without
716 * explicitly disabling preemption.
718 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
724 local_irq_save(*flags);
726 spin_lock(&rq->lock);
727 if (likely(rq == task_rq(p)))
729 spin_unlock_irqrestore(&rq->lock, *flags);
733 static void __task_rq_unlock(struct rq *rq)
736 spin_unlock(&rq->lock);
739 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
742 spin_unlock_irqrestore(&rq->lock, *flags);
746 * this_rq_lock - lock this runqueue and disable interrupts.
748 static struct rq *this_rq_lock(void)
755 spin_lock(&rq->lock);
761 * We are going deep-idle (irqs are disabled):
763 void sched_clock_idle_sleep_event(void)
765 struct rq *rq = cpu_rq(smp_processor_id());
767 spin_lock(&rq->lock);
768 __update_rq_clock(rq);
769 spin_unlock(&rq->lock);
770 rq->clock_deep_idle_events++;
772 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
775 * We just idled delta nanoseconds (called with irqs disabled):
777 void sched_clock_idle_wakeup_event(u64 delta_ns)
779 struct rq *rq = cpu_rq(smp_processor_id());
780 u64 now = sched_clock();
782 touch_softlockup_watchdog();
783 rq->idle_clock += delta_ns;
785 * Override the previous timestamp and ignore all
786 * sched_clock() deltas that occured while we idled,
787 * and use the PM-provided delta_ns to advance the
790 spin_lock(&rq->lock);
791 rq->prev_clock_raw = now;
792 rq->clock += delta_ns;
793 spin_unlock(&rq->lock);
795 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
798 * resched_task - mark a task 'to be rescheduled now'.
800 * On UP this means the setting of the need_resched flag, on SMP it
801 * might also involve a cross-CPU call to trigger the scheduler on
806 #ifndef tsk_is_polling
807 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
810 static void resched_task(struct task_struct *p)
814 assert_spin_locked(&task_rq(p)->lock);
816 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
819 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
822 if (cpu == smp_processor_id())
825 /* NEED_RESCHED must be visible before we test polling */
827 if (!tsk_is_polling(p))
828 smp_send_reschedule(cpu);
831 static void resched_cpu(int cpu)
833 struct rq *rq = cpu_rq(cpu);
836 if (!spin_trylock_irqsave(&rq->lock, flags))
838 resched_task(cpu_curr(cpu));
839 spin_unlock_irqrestore(&rq->lock, flags);
842 static inline void resched_task(struct task_struct *p)
844 assert_spin_locked(&task_rq(p)->lock);
845 set_tsk_need_resched(p);
849 #if BITS_PER_LONG == 32
850 # define WMULT_CONST (~0UL)
852 # define WMULT_CONST (1UL << 32)
855 #define WMULT_SHIFT 32
858 * Shift right and round:
860 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
863 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
864 struct load_weight *lw)
868 if (unlikely(!lw->inv_weight))
869 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
871 tmp = (u64)delta_exec * weight;
873 * Check whether we'd overflow the 64-bit multiplication:
875 if (unlikely(tmp > WMULT_CONST))
876 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
879 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
881 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
884 static inline unsigned long
885 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
887 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
890 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
895 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
901 * To aid in avoiding the subversion of "niceness" due to uneven distribution
902 * of tasks with abnormal "nice" values across CPUs the contribution that
903 * each task makes to its run queue's load is weighted according to its
904 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
905 * scaled version of the new time slice allocation that they receive on time
909 #define WEIGHT_IDLEPRIO 2
910 #define WMULT_IDLEPRIO (1 << 31)
913 * Nice levels are multiplicative, with a gentle 10% change for every
914 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
915 * nice 1, it will get ~10% less CPU time than another CPU-bound task
916 * that remained on nice 0.
918 * The "10% effect" is relative and cumulative: from _any_ nice level,
919 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
920 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
921 * If a task goes up by ~10% and another task goes down by ~10% then
922 * the relative distance between them is ~25%.)
924 static const int prio_to_weight[40] = {
925 /* -20 */ 88761, 71755, 56483, 46273, 36291,
926 /* -15 */ 29154, 23254, 18705, 14949, 11916,
927 /* -10 */ 9548, 7620, 6100, 4904, 3906,
928 /* -5 */ 3121, 2501, 1991, 1586, 1277,
929 /* 0 */ 1024, 820, 655, 526, 423,
930 /* 5 */ 335, 272, 215, 172, 137,
931 /* 10 */ 110, 87, 70, 56, 45,
932 /* 15 */ 36, 29, 23, 18, 15,
936 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
938 * In cases where the weight does not change often, we can use the
939 * precalculated inverse to speed up arithmetics by turning divisions
940 * into multiplications:
942 static const u32 prio_to_wmult[40] = {
943 /* -20 */ 48388, 59856, 76040, 92818, 118348,
944 /* -15 */ 147320, 184698, 229616, 287308, 360437,
945 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
946 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
947 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
948 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
949 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
950 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
953 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
956 * runqueue iterator, to support SMP load-balancing between different
957 * scheduling classes, without having to expose their internal data
958 * structures to the load-balancing proper:
962 struct task_struct *(*start)(void *);
963 struct task_struct *(*next)(void *);
968 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
969 unsigned long max_load_move, struct sched_domain *sd,
970 enum cpu_idle_type idle, int *all_pinned,
971 int *this_best_prio, struct rq_iterator *iterator);
974 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
975 struct sched_domain *sd, enum cpu_idle_type idle,
976 struct rq_iterator *iterator);
979 #ifdef CONFIG_CGROUP_CPUACCT
980 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
982 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
985 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
987 update_load_add(&rq->load, load);
990 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
992 update_load_sub(&rq->load, load);
996 static unsigned long source_load(int cpu, int type);
997 static unsigned long target_load(int cpu, int type);
998 static unsigned long cpu_avg_load_per_task(int cpu);
999 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1000 #endif /* CONFIG_SMP */
1002 #include "sched_stats.h"
1003 #include "sched_idletask.c"
1004 #include "sched_fair.c"
1005 #include "sched_rt.c"
1006 #ifdef CONFIG_SCHED_DEBUG
1007 # include "sched_debug.c"
1010 #define sched_class_highest (&rt_sched_class)
1012 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1017 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1022 static void set_load_weight(struct task_struct *p)
1024 if (task_has_rt_policy(p)) {
1025 p->se.load.weight = prio_to_weight[0] * 2;
1026 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1031 * SCHED_IDLE tasks get minimal weight:
1033 if (p->policy == SCHED_IDLE) {
1034 p->se.load.weight = WEIGHT_IDLEPRIO;
1035 p->se.load.inv_weight = WMULT_IDLEPRIO;
1039 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1040 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1043 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1045 sched_info_queued(p);
1046 p->sched_class->enqueue_task(rq, p, wakeup);
1050 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1052 p->sched_class->dequeue_task(rq, p, sleep);
1057 * __normal_prio - return the priority that is based on the static prio
1059 static inline int __normal_prio(struct task_struct *p)
1061 return p->static_prio;
1065 * Calculate the expected normal priority: i.e. priority
1066 * without taking RT-inheritance into account. Might be
1067 * boosted by interactivity modifiers. Changes upon fork,
1068 * setprio syscalls, and whenever the interactivity
1069 * estimator recalculates.
1071 static inline int normal_prio(struct task_struct *p)
1075 if (task_has_rt_policy(p))
1076 prio = MAX_RT_PRIO-1 - p->rt_priority;
1078 prio = __normal_prio(p);
1083 * Calculate the current priority, i.e. the priority
1084 * taken into account by the scheduler. This value might
1085 * be boosted by RT tasks, or might be boosted by
1086 * interactivity modifiers. Will be RT if the task got
1087 * RT-boosted. If not then it returns p->normal_prio.
1089 static int effective_prio(struct task_struct *p)
1091 p->normal_prio = normal_prio(p);
1093 * If we are RT tasks or we were boosted to RT priority,
1094 * keep the priority unchanged. Otherwise, update priority
1095 * to the normal priority:
1097 if (!rt_prio(p->prio))
1098 return p->normal_prio;
1103 * activate_task - move a task to the runqueue.
1105 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1107 if (p->state == TASK_UNINTERRUPTIBLE)
1108 rq->nr_uninterruptible--;
1110 enqueue_task(rq, p, wakeup);
1111 inc_nr_running(p, rq);
1115 * deactivate_task - remove a task from the runqueue.
1117 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1119 if (p->state == TASK_UNINTERRUPTIBLE)
1120 rq->nr_uninterruptible++;
1122 dequeue_task(rq, p, sleep);
1123 dec_nr_running(p, rq);
1127 * task_curr - is this task currently executing on a CPU?
1128 * @p: the task in question.
1130 inline int task_curr(const struct task_struct *p)
1132 return cpu_curr(task_cpu(p)) == p;
1135 /* Used instead of source_load when we know the type == 0 */
1136 unsigned long weighted_cpuload(const int cpu)
1138 return cpu_rq(cpu)->load.weight;
1141 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1143 set_task_cfs_rq(p, cpu);
1146 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1147 * successfuly executed on another CPU. We must ensure that updates of
1148 * per-task data have been completed by this moment.
1151 task_thread_info(p)->cpu = cpu;
1158 * Is this task likely cache-hot:
1161 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1165 if (p->sched_class != &fair_sched_class)
1168 if (sysctl_sched_migration_cost == -1)
1170 if (sysctl_sched_migration_cost == 0)
1173 delta = now - p->se.exec_start;
1175 return delta < (s64)sysctl_sched_migration_cost;
1179 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1181 int old_cpu = task_cpu(p);
1182 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1183 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1184 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1187 clock_offset = old_rq->clock - new_rq->clock;
1189 #ifdef CONFIG_SCHEDSTATS
1190 if (p->se.wait_start)
1191 p->se.wait_start -= clock_offset;
1192 if (p->se.sleep_start)
1193 p->se.sleep_start -= clock_offset;
1194 if (p->se.block_start)
1195 p->se.block_start -= clock_offset;
1196 if (old_cpu != new_cpu) {
1197 schedstat_inc(p, se.nr_migrations);
1198 if (task_hot(p, old_rq->clock, NULL))
1199 schedstat_inc(p, se.nr_forced2_migrations);
1202 p->se.vruntime -= old_cfsrq->min_vruntime -
1203 new_cfsrq->min_vruntime;
1205 __set_task_cpu(p, new_cpu);
1208 struct migration_req {
1209 struct list_head list;
1211 struct task_struct *task;
1214 struct completion done;
1218 * The task's runqueue lock must be held.
1219 * Returns true if you have to wait for migration thread.
1222 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1224 struct rq *rq = task_rq(p);
1227 * If the task is not on a runqueue (and not running), then
1228 * it is sufficient to simply update the task's cpu field.
1230 if (!p->se.on_rq && !task_running(rq, p)) {
1231 set_task_cpu(p, dest_cpu);
1235 init_completion(&req->done);
1237 req->dest_cpu = dest_cpu;
1238 list_add(&req->list, &rq->migration_queue);
1244 * wait_task_inactive - wait for a thread to unschedule.
1246 * The caller must ensure that the task *will* unschedule sometime soon,
1247 * else this function might spin for a *long* time. This function can't
1248 * be called with interrupts off, or it may introduce deadlock with
1249 * smp_call_function() if an IPI is sent by the same process we are
1250 * waiting to become inactive.
1252 void wait_task_inactive(struct task_struct *p)
1254 unsigned long flags;
1260 * We do the initial early heuristics without holding
1261 * any task-queue locks at all. We'll only try to get
1262 * the runqueue lock when things look like they will
1268 * If the task is actively running on another CPU
1269 * still, just relax and busy-wait without holding
1272 * NOTE! Since we don't hold any locks, it's not
1273 * even sure that "rq" stays as the right runqueue!
1274 * But we don't care, since "task_running()" will
1275 * return false if the runqueue has changed and p
1276 * is actually now running somewhere else!
1278 while (task_running(rq, p))
1282 * Ok, time to look more closely! We need the rq
1283 * lock now, to be *sure*. If we're wrong, we'll
1284 * just go back and repeat.
1286 rq = task_rq_lock(p, &flags);
1287 running = task_running(rq, p);
1288 on_rq = p->se.on_rq;
1289 task_rq_unlock(rq, &flags);
1292 * Was it really running after all now that we
1293 * checked with the proper locks actually held?
1295 * Oops. Go back and try again..
1297 if (unlikely(running)) {
1303 * It's not enough that it's not actively running,
1304 * it must be off the runqueue _entirely_, and not
1307 * So if it wa still runnable (but just not actively
1308 * running right now), it's preempted, and we should
1309 * yield - it could be a while.
1311 if (unlikely(on_rq)) {
1312 schedule_timeout_uninterruptible(1);
1317 * Ahh, all good. It wasn't running, and it wasn't
1318 * runnable, which means that it will never become
1319 * running in the future either. We're all done!
1326 * kick_process - kick a running thread to enter/exit the kernel
1327 * @p: the to-be-kicked thread
1329 * Cause a process which is running on another CPU to enter
1330 * kernel-mode, without any delay. (to get signals handled.)
1332 * NOTE: this function doesnt have to take the runqueue lock,
1333 * because all it wants to ensure is that the remote task enters
1334 * the kernel. If the IPI races and the task has been migrated
1335 * to another CPU then no harm is done and the purpose has been
1338 void kick_process(struct task_struct *p)
1344 if ((cpu != smp_processor_id()) && task_curr(p))
1345 smp_send_reschedule(cpu);
1350 * Return a low guess at the load of a migration-source cpu weighted
1351 * according to the scheduling class and "nice" value.
1353 * We want to under-estimate the load of migration sources, to
1354 * balance conservatively.
1356 static unsigned long source_load(int cpu, int type)
1358 struct rq *rq = cpu_rq(cpu);
1359 unsigned long total = weighted_cpuload(cpu);
1364 return min(rq->cpu_load[type-1], total);
1368 * Return a high guess at the load of a migration-target cpu weighted
1369 * according to the scheduling class and "nice" value.
1371 static unsigned long target_load(int cpu, int type)
1373 struct rq *rq = cpu_rq(cpu);
1374 unsigned long total = weighted_cpuload(cpu);
1379 return max(rq->cpu_load[type-1], total);
1383 * Return the average load per task on the cpu's run queue
1385 static unsigned long cpu_avg_load_per_task(int cpu)
1387 struct rq *rq = cpu_rq(cpu);
1388 unsigned long total = weighted_cpuload(cpu);
1389 unsigned long n = rq->nr_running;
1391 return n ? total / n : SCHED_LOAD_SCALE;
1395 * find_idlest_group finds and returns the least busy CPU group within the
1398 static struct sched_group *
1399 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1401 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1402 unsigned long min_load = ULONG_MAX, this_load = 0;
1403 int load_idx = sd->forkexec_idx;
1404 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1407 unsigned long load, avg_load;
1411 /* Skip over this group if it has no CPUs allowed */
1412 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1415 local_group = cpu_isset(this_cpu, group->cpumask);
1417 /* Tally up the load of all CPUs in the group */
1420 for_each_cpu_mask(i, group->cpumask) {
1421 /* Bias balancing toward cpus of our domain */
1423 load = source_load(i, load_idx);
1425 load = target_load(i, load_idx);
1430 /* Adjust by relative CPU power of the group */
1431 avg_load = sg_div_cpu_power(group,
1432 avg_load * SCHED_LOAD_SCALE);
1435 this_load = avg_load;
1437 } else if (avg_load < min_load) {
1438 min_load = avg_load;
1441 } while (group = group->next, group != sd->groups);
1443 if (!idlest || 100*this_load < imbalance*min_load)
1449 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1452 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1455 unsigned long load, min_load = ULONG_MAX;
1459 /* Traverse only the allowed CPUs */
1460 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1462 for_each_cpu_mask(i, tmp) {
1463 load = weighted_cpuload(i);
1465 if (load < min_load || (load == min_load && i == this_cpu)) {
1475 * sched_balance_self: balance the current task (running on cpu) in domains
1476 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1479 * Balance, ie. select the least loaded group.
1481 * Returns the target CPU number, or the same CPU if no balancing is needed.
1483 * preempt must be disabled.
1485 static int sched_balance_self(int cpu, int flag)
1487 struct task_struct *t = current;
1488 struct sched_domain *tmp, *sd = NULL;
1490 for_each_domain(cpu, tmp) {
1492 * If power savings logic is enabled for a domain, stop there.
1494 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1496 if (tmp->flags & flag)
1502 struct sched_group *group;
1503 int new_cpu, weight;
1505 if (!(sd->flags & flag)) {
1511 group = find_idlest_group(sd, t, cpu);
1517 new_cpu = find_idlest_cpu(group, t, cpu);
1518 if (new_cpu == -1 || new_cpu == cpu) {
1519 /* Now try balancing at a lower domain level of cpu */
1524 /* Now try balancing at a lower domain level of new_cpu */
1527 weight = cpus_weight(span);
1528 for_each_domain(cpu, tmp) {
1529 if (weight <= cpus_weight(tmp->span))
1531 if (tmp->flags & flag)
1534 /* while loop will break here if sd == NULL */
1540 #endif /* CONFIG_SMP */
1543 * try_to_wake_up - wake up a thread
1544 * @p: the to-be-woken-up thread
1545 * @state: the mask of task states that can be woken
1546 * @sync: do a synchronous wakeup?
1548 * Put it on the run-queue if it's not already there. The "current"
1549 * thread is always on the run-queue (except when the actual
1550 * re-schedule is in progress), and as such you're allowed to do
1551 * the simpler "current->state = TASK_RUNNING" to mark yourself
1552 * runnable without the overhead of this.
1554 * returns failure only if the task is already active.
1556 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1558 int cpu, orig_cpu, this_cpu, success = 0;
1559 unsigned long flags;
1563 rq = task_rq_lock(p, &flags);
1564 old_state = p->state;
1565 if (!(old_state & state))
1573 this_cpu = smp_processor_id();
1576 if (unlikely(task_running(rq, p)))
1579 cpu = p->sched_class->select_task_rq(p, sync);
1580 if (cpu != orig_cpu) {
1581 set_task_cpu(p, cpu);
1582 task_rq_unlock(rq, &flags);
1583 /* might preempt at this point */
1584 rq = task_rq_lock(p, &flags);
1585 old_state = p->state;
1586 if (!(old_state & state))
1591 this_cpu = smp_processor_id();
1595 #ifdef CONFIG_SCHEDSTATS
1596 schedstat_inc(rq, ttwu_count);
1597 if (cpu == this_cpu)
1598 schedstat_inc(rq, ttwu_local);
1600 struct sched_domain *sd;
1601 for_each_domain(this_cpu, sd) {
1602 if (cpu_isset(cpu, sd->span)) {
1603 schedstat_inc(sd, ttwu_wake_remote);
1611 #endif /* CONFIG_SMP */
1612 schedstat_inc(p, se.nr_wakeups);
1614 schedstat_inc(p, se.nr_wakeups_sync);
1615 if (orig_cpu != cpu)
1616 schedstat_inc(p, se.nr_wakeups_migrate);
1617 if (cpu == this_cpu)
1618 schedstat_inc(p, se.nr_wakeups_local);
1620 schedstat_inc(p, se.nr_wakeups_remote);
1621 update_rq_clock(rq);
1622 activate_task(rq, p, 1);
1623 check_preempt_curr(rq, p);
1627 p->state = TASK_RUNNING;
1629 if (p->sched_class->task_wake_up)
1630 p->sched_class->task_wake_up(rq, p);
1633 task_rq_unlock(rq, &flags);
1638 int fastcall wake_up_process(struct task_struct *p)
1640 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1641 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1643 EXPORT_SYMBOL(wake_up_process);
1645 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1647 return try_to_wake_up(p, state, 0);
1651 * Perform scheduler related setup for a newly forked process p.
1652 * p is forked by current.
1654 * __sched_fork() is basic setup used by init_idle() too:
1656 static void __sched_fork(struct task_struct *p)
1658 p->se.exec_start = 0;
1659 p->se.sum_exec_runtime = 0;
1660 p->se.prev_sum_exec_runtime = 0;
1662 #ifdef CONFIG_SCHEDSTATS
1663 p->se.wait_start = 0;
1664 p->se.sum_sleep_runtime = 0;
1665 p->se.sleep_start = 0;
1666 p->se.block_start = 0;
1667 p->se.sleep_max = 0;
1668 p->se.block_max = 0;
1670 p->se.slice_max = 0;
1674 INIT_LIST_HEAD(&p->run_list);
1677 #ifdef CONFIG_PREEMPT_NOTIFIERS
1678 INIT_HLIST_HEAD(&p->preempt_notifiers);
1682 * We mark the process as running here, but have not actually
1683 * inserted it onto the runqueue yet. This guarantees that
1684 * nobody will actually run it, and a signal or other external
1685 * event cannot wake it up and insert it on the runqueue either.
1687 p->state = TASK_RUNNING;
1691 * fork()/clone()-time setup:
1693 void sched_fork(struct task_struct *p, int clone_flags)
1695 int cpu = get_cpu();
1700 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1702 set_task_cpu(p, cpu);
1705 * Make sure we do not leak PI boosting priority to the child:
1707 p->prio = current->normal_prio;
1708 if (!rt_prio(p->prio))
1709 p->sched_class = &fair_sched_class;
1711 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1712 if (likely(sched_info_on()))
1713 memset(&p->sched_info, 0, sizeof(p->sched_info));
1715 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1718 #ifdef CONFIG_PREEMPT
1719 /* Want to start with kernel preemption disabled. */
1720 task_thread_info(p)->preempt_count = 1;
1726 * wake_up_new_task - wake up a newly created task for the first time.
1728 * This function will do some initial scheduler statistics housekeeping
1729 * that must be done for every newly created context, then puts the task
1730 * on the runqueue and wakes it.
1732 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1734 unsigned long flags;
1737 rq = task_rq_lock(p, &flags);
1738 BUG_ON(p->state != TASK_RUNNING);
1739 update_rq_clock(rq);
1741 p->prio = effective_prio(p);
1743 if (!p->sched_class->task_new || !current->se.on_rq) {
1744 activate_task(rq, p, 0);
1747 * Let the scheduling class do new task startup
1748 * management (if any):
1750 p->sched_class->task_new(rq, p);
1751 inc_nr_running(p, rq);
1753 check_preempt_curr(rq, p);
1755 if (p->sched_class->task_wake_up)
1756 p->sched_class->task_wake_up(rq, p);
1758 task_rq_unlock(rq, &flags);
1761 #ifdef CONFIG_PREEMPT_NOTIFIERS
1764 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1765 * @notifier: notifier struct to register
1767 void preempt_notifier_register(struct preempt_notifier *notifier)
1769 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1771 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1774 * preempt_notifier_unregister - no longer interested in preemption notifications
1775 * @notifier: notifier struct to unregister
1777 * This is safe to call from within a preemption notifier.
1779 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1781 hlist_del(¬ifier->link);
1783 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1785 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1787 struct preempt_notifier *notifier;
1788 struct hlist_node *node;
1790 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1791 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1795 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1796 struct task_struct *next)
1798 struct preempt_notifier *notifier;
1799 struct hlist_node *node;
1801 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1802 notifier->ops->sched_out(notifier, next);
1807 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1812 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1813 struct task_struct *next)
1820 * prepare_task_switch - prepare to switch tasks
1821 * @rq: the runqueue preparing to switch
1822 * @prev: the current task that is being switched out
1823 * @next: the task we are going to switch to.
1825 * This is called with the rq lock held and interrupts off. It must
1826 * be paired with a subsequent finish_task_switch after the context
1829 * prepare_task_switch sets up locking and calls architecture specific
1833 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1834 struct task_struct *next)
1836 fire_sched_out_preempt_notifiers(prev, next);
1837 prepare_lock_switch(rq, next);
1838 prepare_arch_switch(next);
1842 * finish_task_switch - clean up after a task-switch
1843 * @rq: runqueue associated with task-switch
1844 * @prev: the thread we just switched away from.
1846 * finish_task_switch must be called after the context switch, paired
1847 * with a prepare_task_switch call before the context switch.
1848 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1849 * and do any other architecture-specific cleanup actions.
1851 * Note that we may have delayed dropping an mm in context_switch(). If
1852 * so, we finish that here outside of the runqueue lock. (Doing it
1853 * with the lock held can cause deadlocks; see schedule() for
1856 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1857 __releases(rq->lock)
1859 struct mm_struct *mm = rq->prev_mm;
1865 * A task struct has one reference for the use as "current".
1866 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1867 * schedule one last time. The schedule call will never return, and
1868 * the scheduled task must drop that reference.
1869 * The test for TASK_DEAD must occur while the runqueue locks are
1870 * still held, otherwise prev could be scheduled on another cpu, die
1871 * there before we look at prev->state, and then the reference would
1873 * Manfred Spraul <manfred@colorfullife.com>
1875 prev_state = prev->state;
1876 finish_arch_switch(prev);
1877 finish_lock_switch(rq, prev);
1879 if (current->sched_class->post_schedule)
1880 current->sched_class->post_schedule(rq);
1883 fire_sched_in_preempt_notifiers(current);
1886 if (unlikely(prev_state == TASK_DEAD)) {
1888 * Remove function-return probe instances associated with this
1889 * task and put them back on the free list.
1891 kprobe_flush_task(prev);
1892 put_task_struct(prev);
1897 * schedule_tail - first thing a freshly forked thread must call.
1898 * @prev: the thread we just switched away from.
1900 asmlinkage void schedule_tail(struct task_struct *prev)
1901 __releases(rq->lock)
1903 struct rq *rq = this_rq();
1905 finish_task_switch(rq, prev);
1906 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1907 /* In this case, finish_task_switch does not reenable preemption */
1910 if (current->set_child_tid)
1911 put_user(task_pid_vnr(current), current->set_child_tid);
1915 * context_switch - switch to the new MM and the new
1916 * thread's register state.
1919 context_switch(struct rq *rq, struct task_struct *prev,
1920 struct task_struct *next)
1922 struct mm_struct *mm, *oldmm;
1924 prepare_task_switch(rq, prev, next);
1926 oldmm = prev->active_mm;
1928 * For paravirt, this is coupled with an exit in switch_to to
1929 * combine the page table reload and the switch backend into
1932 arch_enter_lazy_cpu_mode();
1934 if (unlikely(!mm)) {
1935 next->active_mm = oldmm;
1936 atomic_inc(&oldmm->mm_count);
1937 enter_lazy_tlb(oldmm, next);
1939 switch_mm(oldmm, mm, next);
1941 if (unlikely(!prev->mm)) {
1942 prev->active_mm = NULL;
1943 rq->prev_mm = oldmm;
1946 * Since the runqueue lock will be released by the next
1947 * task (which is an invalid locking op but in the case
1948 * of the scheduler it's an obvious special-case), so we
1949 * do an early lockdep release here:
1951 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1952 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1955 /* Here we just switch the register state and the stack. */
1956 switch_to(prev, next, prev);
1960 * this_rq must be evaluated again because prev may have moved
1961 * CPUs since it called schedule(), thus the 'rq' on its stack
1962 * frame will be invalid.
1964 finish_task_switch(this_rq(), prev);
1968 * nr_running, nr_uninterruptible and nr_context_switches:
1970 * externally visible scheduler statistics: current number of runnable
1971 * threads, current number of uninterruptible-sleeping threads, total
1972 * number of context switches performed since bootup.
1974 unsigned long nr_running(void)
1976 unsigned long i, sum = 0;
1978 for_each_online_cpu(i)
1979 sum += cpu_rq(i)->nr_running;
1984 unsigned long nr_uninterruptible(void)
1986 unsigned long i, sum = 0;
1988 for_each_possible_cpu(i)
1989 sum += cpu_rq(i)->nr_uninterruptible;
1992 * Since we read the counters lockless, it might be slightly
1993 * inaccurate. Do not allow it to go below zero though:
1995 if (unlikely((long)sum < 0))
2001 unsigned long long nr_context_switches(void)
2004 unsigned long long sum = 0;
2006 for_each_possible_cpu(i)
2007 sum += cpu_rq(i)->nr_switches;
2012 unsigned long nr_iowait(void)
2014 unsigned long i, sum = 0;
2016 for_each_possible_cpu(i)
2017 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2022 unsigned long nr_active(void)
2024 unsigned long i, running = 0, uninterruptible = 0;
2026 for_each_online_cpu(i) {
2027 running += cpu_rq(i)->nr_running;
2028 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2031 if (unlikely((long)uninterruptible < 0))
2032 uninterruptible = 0;
2034 return running + uninterruptible;
2038 * Update rq->cpu_load[] statistics. This function is usually called every
2039 * scheduler tick (TICK_NSEC).
2041 static void update_cpu_load(struct rq *this_rq)
2043 unsigned long this_load = this_rq->load.weight;
2046 this_rq->nr_load_updates++;
2048 /* Update our load: */
2049 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2050 unsigned long old_load, new_load;
2052 /* scale is effectively 1 << i now, and >> i divides by scale */
2054 old_load = this_rq->cpu_load[i];
2055 new_load = this_load;
2057 * Round up the averaging division if load is increasing. This
2058 * prevents us from getting stuck on 9 if the load is 10, for
2061 if (new_load > old_load)
2062 new_load += scale-1;
2063 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2070 * double_rq_lock - safely lock two runqueues
2072 * Note this does not disable interrupts like task_rq_lock,
2073 * you need to do so manually before calling.
2075 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2076 __acquires(rq1->lock)
2077 __acquires(rq2->lock)
2079 BUG_ON(!irqs_disabled());
2081 spin_lock(&rq1->lock);
2082 __acquire(rq2->lock); /* Fake it out ;) */
2085 spin_lock(&rq1->lock);
2086 spin_lock(&rq2->lock);
2088 spin_lock(&rq2->lock);
2089 spin_lock(&rq1->lock);
2092 update_rq_clock(rq1);
2093 update_rq_clock(rq2);
2097 * double_rq_unlock - safely unlock two runqueues
2099 * Note this does not restore interrupts like task_rq_unlock,
2100 * you need to do so manually after calling.
2102 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2103 __releases(rq1->lock)
2104 __releases(rq2->lock)
2106 spin_unlock(&rq1->lock);
2108 spin_unlock(&rq2->lock);
2110 __release(rq2->lock);
2114 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2116 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2117 __releases(this_rq->lock)
2118 __acquires(busiest->lock)
2119 __acquires(this_rq->lock)
2123 if (unlikely(!irqs_disabled())) {
2124 /* printk() doesn't work good under rq->lock */
2125 spin_unlock(&this_rq->lock);
2128 if (unlikely(!spin_trylock(&busiest->lock))) {
2129 if (busiest < this_rq) {
2130 spin_unlock(&this_rq->lock);
2131 spin_lock(&busiest->lock);
2132 spin_lock(&this_rq->lock);
2135 spin_lock(&busiest->lock);
2141 * If dest_cpu is allowed for this process, migrate the task to it.
2142 * This is accomplished by forcing the cpu_allowed mask to only
2143 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2144 * the cpu_allowed mask is restored.
2146 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2148 struct migration_req req;
2149 unsigned long flags;
2152 rq = task_rq_lock(p, &flags);
2153 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2154 || unlikely(cpu_is_offline(dest_cpu)))
2157 /* force the process onto the specified CPU */
2158 if (migrate_task(p, dest_cpu, &req)) {
2159 /* Need to wait for migration thread (might exit: take ref). */
2160 struct task_struct *mt = rq->migration_thread;
2162 get_task_struct(mt);
2163 task_rq_unlock(rq, &flags);
2164 wake_up_process(mt);
2165 put_task_struct(mt);
2166 wait_for_completion(&req.done);
2171 task_rq_unlock(rq, &flags);
2175 * sched_exec - execve() is a valuable balancing opportunity, because at
2176 * this point the task has the smallest effective memory and cache footprint.
2178 void sched_exec(void)
2180 int new_cpu, this_cpu = get_cpu();
2181 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2183 if (new_cpu != this_cpu)
2184 sched_migrate_task(current, new_cpu);
2188 * pull_task - move a task from a remote runqueue to the local runqueue.
2189 * Both runqueues must be locked.
2191 static void pull_task(struct rq *src_rq, struct task_struct *p,
2192 struct rq *this_rq, int this_cpu)
2194 deactivate_task(src_rq, p, 0);
2195 set_task_cpu(p, this_cpu);
2196 activate_task(this_rq, p, 0);
2198 * Note that idle threads have a prio of MAX_PRIO, for this test
2199 * to be always true for them.
2201 check_preempt_curr(this_rq, p);
2205 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2208 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2209 struct sched_domain *sd, enum cpu_idle_type idle,
2213 * We do not migrate tasks that are:
2214 * 1) running (obviously), or
2215 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2216 * 3) are cache-hot on their current CPU.
2218 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2219 schedstat_inc(p, se.nr_failed_migrations_affine);
2224 if (task_running(rq, p)) {
2225 schedstat_inc(p, se.nr_failed_migrations_running);
2230 * Aggressive migration if:
2231 * 1) task is cache cold, or
2232 * 2) too many balance attempts have failed.
2235 if (!task_hot(p, rq->clock, sd) ||
2236 sd->nr_balance_failed > sd->cache_nice_tries) {
2237 #ifdef CONFIG_SCHEDSTATS
2238 if (task_hot(p, rq->clock, sd)) {
2239 schedstat_inc(sd, lb_hot_gained[idle]);
2240 schedstat_inc(p, se.nr_forced_migrations);
2246 if (task_hot(p, rq->clock, sd)) {
2247 schedstat_inc(p, se.nr_failed_migrations_hot);
2253 static unsigned long
2254 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2255 unsigned long max_load_move, struct sched_domain *sd,
2256 enum cpu_idle_type idle, int *all_pinned,
2257 int *this_best_prio, struct rq_iterator *iterator)
2259 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2260 struct task_struct *p;
2261 long rem_load_move = max_load_move;
2263 if (max_load_move == 0)
2269 * Start the load-balancing iterator:
2271 p = iterator->start(iterator->arg);
2273 if (!p || loops++ > sysctl_sched_nr_migrate)
2276 * To help distribute high priority tasks across CPUs we don't
2277 * skip a task if it will be the highest priority task (i.e. smallest
2278 * prio value) on its new queue regardless of its load weight
2280 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2281 SCHED_LOAD_SCALE_FUZZ;
2282 if ((skip_for_load && p->prio >= *this_best_prio) ||
2283 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2284 p = iterator->next(iterator->arg);
2288 pull_task(busiest, p, this_rq, this_cpu);
2290 rem_load_move -= p->se.load.weight;
2293 * We only want to steal up to the prescribed amount of weighted load.
2295 if (rem_load_move > 0) {
2296 if (p->prio < *this_best_prio)
2297 *this_best_prio = p->prio;
2298 p = iterator->next(iterator->arg);
2303 * Right now, this is one of only two places pull_task() is called,
2304 * so we can safely collect pull_task() stats here rather than
2305 * inside pull_task().
2307 schedstat_add(sd, lb_gained[idle], pulled);
2310 *all_pinned = pinned;
2312 return max_load_move - rem_load_move;
2316 * move_tasks tries to move up to max_load_move weighted load from busiest to
2317 * this_rq, as part of a balancing operation within domain "sd".
2318 * Returns 1 if successful and 0 otherwise.
2320 * Called with both runqueues locked.
2322 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2323 unsigned long max_load_move,
2324 struct sched_domain *sd, enum cpu_idle_type idle,
2327 const struct sched_class *class = sched_class_highest;
2328 unsigned long total_load_moved = 0;
2329 int this_best_prio = this_rq->curr->prio;
2333 class->load_balance(this_rq, this_cpu, busiest,
2334 max_load_move - total_load_moved,
2335 sd, idle, all_pinned, &this_best_prio);
2336 class = class->next;
2337 } while (class && max_load_move > total_load_moved);
2339 return total_load_moved > 0;
2343 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2344 struct sched_domain *sd, enum cpu_idle_type idle,
2345 struct rq_iterator *iterator)
2347 struct task_struct *p = iterator->start(iterator->arg);
2351 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2352 pull_task(busiest, p, this_rq, this_cpu);
2354 * Right now, this is only the second place pull_task()
2355 * is called, so we can safely collect pull_task()
2356 * stats here rather than inside pull_task().
2358 schedstat_inc(sd, lb_gained[idle]);
2362 p = iterator->next(iterator->arg);
2369 * move_one_task tries to move exactly one task from busiest to this_rq, as
2370 * part of active balancing operations within "domain".
2371 * Returns 1 if successful and 0 otherwise.
2373 * Called with both runqueues locked.
2375 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2376 struct sched_domain *sd, enum cpu_idle_type idle)
2378 const struct sched_class *class;
2380 for (class = sched_class_highest; class; class = class->next)
2381 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2388 * find_busiest_group finds and returns the busiest CPU group within the
2389 * domain. It calculates and returns the amount of weighted load which
2390 * should be moved to restore balance via the imbalance parameter.
2392 static struct sched_group *
2393 find_busiest_group(struct sched_domain *sd, int this_cpu,
2394 unsigned long *imbalance, enum cpu_idle_type idle,
2395 int *sd_idle, cpumask_t *cpus, int *balance)
2397 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2398 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2399 unsigned long max_pull;
2400 unsigned long busiest_load_per_task, busiest_nr_running;
2401 unsigned long this_load_per_task, this_nr_running;
2402 int load_idx, group_imb = 0;
2403 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2404 int power_savings_balance = 1;
2405 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2406 unsigned long min_nr_running = ULONG_MAX;
2407 struct sched_group *group_min = NULL, *group_leader = NULL;
2410 max_load = this_load = total_load = total_pwr = 0;
2411 busiest_load_per_task = busiest_nr_running = 0;
2412 this_load_per_task = this_nr_running = 0;
2413 if (idle == CPU_NOT_IDLE)
2414 load_idx = sd->busy_idx;
2415 else if (idle == CPU_NEWLY_IDLE)
2416 load_idx = sd->newidle_idx;
2418 load_idx = sd->idle_idx;
2421 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2424 int __group_imb = 0;
2425 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2426 unsigned long sum_nr_running, sum_weighted_load;
2428 local_group = cpu_isset(this_cpu, group->cpumask);
2431 balance_cpu = first_cpu(group->cpumask);
2433 /* Tally up the load of all CPUs in the group */
2434 sum_weighted_load = sum_nr_running = avg_load = 0;
2436 min_cpu_load = ~0UL;
2438 for_each_cpu_mask(i, group->cpumask) {
2441 if (!cpu_isset(i, *cpus))
2446 if (*sd_idle && rq->nr_running)
2449 /* Bias balancing toward cpus of our domain */
2451 if (idle_cpu(i) && !first_idle_cpu) {
2456 load = target_load(i, load_idx);
2458 load = source_load(i, load_idx);
2459 if (load > max_cpu_load)
2460 max_cpu_load = load;
2461 if (min_cpu_load > load)
2462 min_cpu_load = load;
2466 sum_nr_running += rq->nr_running;
2467 sum_weighted_load += weighted_cpuload(i);
2471 * First idle cpu or the first cpu(busiest) in this sched group
2472 * is eligible for doing load balancing at this and above
2473 * domains. In the newly idle case, we will allow all the cpu's
2474 * to do the newly idle load balance.
2476 if (idle != CPU_NEWLY_IDLE && local_group &&
2477 balance_cpu != this_cpu && balance) {
2482 total_load += avg_load;
2483 total_pwr += group->__cpu_power;
2485 /* Adjust by relative CPU power of the group */
2486 avg_load = sg_div_cpu_power(group,
2487 avg_load * SCHED_LOAD_SCALE);
2489 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2492 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2495 this_load = avg_load;
2497 this_nr_running = sum_nr_running;
2498 this_load_per_task = sum_weighted_load;
2499 } else if (avg_load > max_load &&
2500 (sum_nr_running > group_capacity || __group_imb)) {
2501 max_load = avg_load;
2503 busiest_nr_running = sum_nr_running;
2504 busiest_load_per_task = sum_weighted_load;
2505 group_imb = __group_imb;
2508 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2510 * Busy processors will not participate in power savings
2513 if (idle == CPU_NOT_IDLE ||
2514 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2518 * If the local group is idle or completely loaded
2519 * no need to do power savings balance at this domain
2521 if (local_group && (this_nr_running >= group_capacity ||
2523 power_savings_balance = 0;
2526 * If a group is already running at full capacity or idle,
2527 * don't include that group in power savings calculations
2529 if (!power_savings_balance || sum_nr_running >= group_capacity
2534 * Calculate the group which has the least non-idle load.
2535 * This is the group from where we need to pick up the load
2538 if ((sum_nr_running < min_nr_running) ||
2539 (sum_nr_running == min_nr_running &&
2540 first_cpu(group->cpumask) <
2541 first_cpu(group_min->cpumask))) {
2543 min_nr_running = sum_nr_running;
2544 min_load_per_task = sum_weighted_load /
2549 * Calculate the group which is almost near its
2550 * capacity but still has some space to pick up some load
2551 * from other group and save more power
2553 if (sum_nr_running <= group_capacity - 1) {
2554 if (sum_nr_running > leader_nr_running ||
2555 (sum_nr_running == leader_nr_running &&
2556 first_cpu(group->cpumask) >
2557 first_cpu(group_leader->cpumask))) {
2558 group_leader = group;
2559 leader_nr_running = sum_nr_running;
2564 group = group->next;
2565 } while (group != sd->groups);
2567 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2570 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2572 if (this_load >= avg_load ||
2573 100*max_load <= sd->imbalance_pct*this_load)
2576 busiest_load_per_task /= busiest_nr_running;
2578 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2581 * We're trying to get all the cpus to the average_load, so we don't
2582 * want to push ourselves above the average load, nor do we wish to
2583 * reduce the max loaded cpu below the average load, as either of these
2584 * actions would just result in more rebalancing later, and ping-pong
2585 * tasks around. Thus we look for the minimum possible imbalance.
2586 * Negative imbalances (*we* are more loaded than anyone else) will
2587 * be counted as no imbalance for these purposes -- we can't fix that
2588 * by pulling tasks to us. Be careful of negative numbers as they'll
2589 * appear as very large values with unsigned longs.
2591 if (max_load <= busiest_load_per_task)
2595 * In the presence of smp nice balancing, certain scenarios can have
2596 * max load less than avg load(as we skip the groups at or below
2597 * its cpu_power, while calculating max_load..)
2599 if (max_load < avg_load) {
2601 goto small_imbalance;
2604 /* Don't want to pull so many tasks that a group would go idle */
2605 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2607 /* How much load to actually move to equalise the imbalance */
2608 *imbalance = min(max_pull * busiest->__cpu_power,
2609 (avg_load - this_load) * this->__cpu_power)
2613 * if *imbalance is less than the average load per runnable task
2614 * there is no gaurantee that any tasks will be moved so we'll have
2615 * a think about bumping its value to force at least one task to be
2618 if (*imbalance < busiest_load_per_task) {
2619 unsigned long tmp, pwr_now, pwr_move;
2623 pwr_move = pwr_now = 0;
2625 if (this_nr_running) {
2626 this_load_per_task /= this_nr_running;
2627 if (busiest_load_per_task > this_load_per_task)
2630 this_load_per_task = SCHED_LOAD_SCALE;
2632 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2633 busiest_load_per_task * imbn) {
2634 *imbalance = busiest_load_per_task;
2639 * OK, we don't have enough imbalance to justify moving tasks,
2640 * however we may be able to increase total CPU power used by
2644 pwr_now += busiest->__cpu_power *
2645 min(busiest_load_per_task, max_load);
2646 pwr_now += this->__cpu_power *
2647 min(this_load_per_task, this_load);
2648 pwr_now /= SCHED_LOAD_SCALE;
2650 /* Amount of load we'd subtract */
2651 tmp = sg_div_cpu_power(busiest,
2652 busiest_load_per_task * SCHED_LOAD_SCALE);
2654 pwr_move += busiest->__cpu_power *
2655 min(busiest_load_per_task, max_load - tmp);
2657 /* Amount of load we'd add */
2658 if (max_load * busiest->__cpu_power <
2659 busiest_load_per_task * SCHED_LOAD_SCALE)
2660 tmp = sg_div_cpu_power(this,
2661 max_load * busiest->__cpu_power);
2663 tmp = sg_div_cpu_power(this,
2664 busiest_load_per_task * SCHED_LOAD_SCALE);
2665 pwr_move += this->__cpu_power *
2666 min(this_load_per_task, this_load + tmp);
2667 pwr_move /= SCHED_LOAD_SCALE;
2669 /* Move if we gain throughput */
2670 if (pwr_move > pwr_now)
2671 *imbalance = busiest_load_per_task;
2677 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2678 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2681 if (this == group_leader && group_leader != group_min) {
2682 *imbalance = min_load_per_task;
2692 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2695 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2696 unsigned long imbalance, cpumask_t *cpus)
2698 struct rq *busiest = NULL, *rq;
2699 unsigned long max_load = 0;
2702 for_each_cpu_mask(i, group->cpumask) {
2705 if (!cpu_isset(i, *cpus))
2709 wl = weighted_cpuload(i);
2711 if (rq->nr_running == 1 && wl > imbalance)
2714 if (wl > max_load) {
2724 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2725 * so long as it is large enough.
2727 #define MAX_PINNED_INTERVAL 512
2730 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2731 * tasks if there is an imbalance.
2733 static int load_balance(int this_cpu, struct rq *this_rq,
2734 struct sched_domain *sd, enum cpu_idle_type idle,
2737 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2738 struct sched_group *group;
2739 unsigned long imbalance;
2741 cpumask_t cpus = CPU_MASK_ALL;
2742 unsigned long flags;
2745 * When power savings policy is enabled for the parent domain, idle
2746 * sibling can pick up load irrespective of busy siblings. In this case,
2747 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2748 * portraying it as CPU_NOT_IDLE.
2750 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2751 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2754 schedstat_inc(sd, lb_count[idle]);
2757 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2764 schedstat_inc(sd, lb_nobusyg[idle]);
2768 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2770 schedstat_inc(sd, lb_nobusyq[idle]);
2774 BUG_ON(busiest == this_rq);
2776 schedstat_add(sd, lb_imbalance[idle], imbalance);
2779 if (busiest->nr_running > 1) {
2781 * Attempt to move tasks. If find_busiest_group has found
2782 * an imbalance but busiest->nr_running <= 1, the group is
2783 * still unbalanced. ld_moved simply stays zero, so it is
2784 * correctly treated as an imbalance.
2786 local_irq_save(flags);
2787 double_rq_lock(this_rq, busiest);
2788 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2789 imbalance, sd, idle, &all_pinned);
2790 double_rq_unlock(this_rq, busiest);
2791 local_irq_restore(flags);
2794 * some other cpu did the load balance for us.
2796 if (ld_moved && this_cpu != smp_processor_id())
2797 resched_cpu(this_cpu);
2799 /* All tasks on this runqueue were pinned by CPU affinity */
2800 if (unlikely(all_pinned)) {
2801 cpu_clear(cpu_of(busiest), cpus);
2802 if (!cpus_empty(cpus))
2809 schedstat_inc(sd, lb_failed[idle]);
2810 sd->nr_balance_failed++;
2812 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2814 spin_lock_irqsave(&busiest->lock, flags);
2816 /* don't kick the migration_thread, if the curr
2817 * task on busiest cpu can't be moved to this_cpu
2819 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2820 spin_unlock_irqrestore(&busiest->lock, flags);
2822 goto out_one_pinned;
2825 if (!busiest->active_balance) {
2826 busiest->active_balance = 1;
2827 busiest->push_cpu = this_cpu;
2830 spin_unlock_irqrestore(&busiest->lock, flags);
2832 wake_up_process(busiest->migration_thread);
2835 * We've kicked active balancing, reset the failure
2838 sd->nr_balance_failed = sd->cache_nice_tries+1;
2841 sd->nr_balance_failed = 0;
2843 if (likely(!active_balance)) {
2844 /* We were unbalanced, so reset the balancing interval */
2845 sd->balance_interval = sd->min_interval;
2848 * If we've begun active balancing, start to back off. This
2849 * case may not be covered by the all_pinned logic if there
2850 * is only 1 task on the busy runqueue (because we don't call
2853 if (sd->balance_interval < sd->max_interval)
2854 sd->balance_interval *= 2;
2857 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2858 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2863 schedstat_inc(sd, lb_balanced[idle]);
2865 sd->nr_balance_failed = 0;
2868 /* tune up the balancing interval */
2869 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2870 (sd->balance_interval < sd->max_interval))
2871 sd->balance_interval *= 2;
2873 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2874 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2880 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2881 * tasks if there is an imbalance.
2883 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2884 * this_rq is locked.
2887 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2889 struct sched_group *group;
2890 struct rq *busiest = NULL;
2891 unsigned long imbalance;
2895 cpumask_t cpus = CPU_MASK_ALL;
2898 * When power savings policy is enabled for the parent domain, idle
2899 * sibling can pick up load irrespective of busy siblings. In this case,
2900 * let the state of idle sibling percolate up as IDLE, instead of
2901 * portraying it as CPU_NOT_IDLE.
2903 if (sd->flags & SD_SHARE_CPUPOWER &&
2904 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2907 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2909 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2910 &sd_idle, &cpus, NULL);
2912 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2916 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2919 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2923 BUG_ON(busiest == this_rq);
2925 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2928 if (busiest->nr_running > 1) {
2929 /* Attempt to move tasks */
2930 double_lock_balance(this_rq, busiest);
2931 /* this_rq->clock is already updated */
2932 update_rq_clock(busiest);
2933 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2934 imbalance, sd, CPU_NEWLY_IDLE,
2936 spin_unlock(&busiest->lock);
2938 if (unlikely(all_pinned)) {
2939 cpu_clear(cpu_of(busiest), cpus);
2940 if (!cpus_empty(cpus))
2946 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2947 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2948 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2951 sd->nr_balance_failed = 0;
2956 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2957 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2958 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2960 sd->nr_balance_failed = 0;
2966 * idle_balance is called by schedule() if this_cpu is about to become
2967 * idle. Attempts to pull tasks from other CPUs.
2969 static void idle_balance(int this_cpu, struct rq *this_rq)
2971 struct sched_domain *sd;
2972 int pulled_task = -1;
2973 unsigned long next_balance = jiffies + HZ;
2975 for_each_domain(this_cpu, sd) {
2976 unsigned long interval;
2978 if (!(sd->flags & SD_LOAD_BALANCE))
2981 if (sd->flags & SD_BALANCE_NEWIDLE)
2982 /* If we've pulled tasks over stop searching: */
2983 pulled_task = load_balance_newidle(this_cpu,
2986 interval = msecs_to_jiffies(sd->balance_interval);
2987 if (time_after(next_balance, sd->last_balance + interval))
2988 next_balance = sd->last_balance + interval;
2992 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2994 * We are going idle. next_balance may be set based on
2995 * a busy processor. So reset next_balance.
2997 this_rq->next_balance = next_balance;
3002 * active_load_balance is run by migration threads. It pushes running tasks
3003 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3004 * running on each physical CPU where possible, and avoids physical /
3005 * logical imbalances.
3007 * Called with busiest_rq locked.
3009 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3011 int target_cpu = busiest_rq->push_cpu;
3012 struct sched_domain *sd;
3013 struct rq *target_rq;
3015 /* Is there any task to move? */
3016 if (busiest_rq->nr_running <= 1)
3019 target_rq = cpu_rq(target_cpu);
3022 * This condition is "impossible", if it occurs
3023 * we need to fix it. Originally reported by
3024 * Bjorn Helgaas on a 128-cpu setup.
3026 BUG_ON(busiest_rq == target_rq);
3028 /* move a task from busiest_rq to target_rq */
3029 double_lock_balance(busiest_rq, target_rq);
3030 update_rq_clock(busiest_rq);
3031 update_rq_clock(target_rq);
3033 /* Search for an sd spanning us and the target CPU. */
3034 for_each_domain(target_cpu, sd) {
3035 if ((sd->flags & SD_LOAD_BALANCE) &&
3036 cpu_isset(busiest_cpu, sd->span))
3041 schedstat_inc(sd, alb_count);
3043 if (move_one_task(target_rq, target_cpu, busiest_rq,
3045 schedstat_inc(sd, alb_pushed);
3047 schedstat_inc(sd, alb_failed);
3049 spin_unlock(&target_rq->lock);
3054 atomic_t load_balancer;
3056 } nohz ____cacheline_aligned = {
3057 .load_balancer = ATOMIC_INIT(-1),
3058 .cpu_mask = CPU_MASK_NONE,
3062 * This routine will try to nominate the ilb (idle load balancing)
3063 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3064 * load balancing on behalf of all those cpus. If all the cpus in the system
3065 * go into this tickless mode, then there will be no ilb owner (as there is
3066 * no need for one) and all the cpus will sleep till the next wakeup event
3069 * For the ilb owner, tick is not stopped. And this tick will be used
3070 * for idle load balancing. ilb owner will still be part of
3073 * While stopping the tick, this cpu will become the ilb owner if there
3074 * is no other owner. And will be the owner till that cpu becomes busy
3075 * or if all cpus in the system stop their ticks at which point
3076 * there is no need for ilb owner.
3078 * When the ilb owner becomes busy, it nominates another owner, during the
3079 * next busy scheduler_tick()
3081 int select_nohz_load_balancer(int stop_tick)
3083 int cpu = smp_processor_id();
3086 cpu_set(cpu, nohz.cpu_mask);
3087 cpu_rq(cpu)->in_nohz_recently = 1;
3090 * If we are going offline and still the leader, give up!
3092 if (cpu_is_offline(cpu) &&
3093 atomic_read(&nohz.load_balancer) == cpu) {
3094 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3099 /* time for ilb owner also to sleep */
3100 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3101 if (atomic_read(&nohz.load_balancer) == cpu)
3102 atomic_set(&nohz.load_balancer, -1);
3106 if (atomic_read(&nohz.load_balancer) == -1) {
3107 /* make me the ilb owner */
3108 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3110 } else if (atomic_read(&nohz.load_balancer) == cpu)
3113 if (!cpu_isset(cpu, nohz.cpu_mask))
3116 cpu_clear(cpu, nohz.cpu_mask);
3118 if (atomic_read(&nohz.load_balancer) == cpu)
3119 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3126 static DEFINE_SPINLOCK(balancing);
3129 * It checks each scheduling domain to see if it is due to be balanced,
3130 * and initiates a balancing operation if so.
3132 * Balancing parameters are set up in arch_init_sched_domains.
3134 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3137 struct rq *rq = cpu_rq(cpu);
3138 unsigned long interval;
3139 struct sched_domain *sd;
3140 /* Earliest time when we have to do rebalance again */
3141 unsigned long next_balance = jiffies + 60*HZ;
3142 int update_next_balance = 0;
3144 for_each_domain(cpu, sd) {
3145 if (!(sd->flags & SD_LOAD_BALANCE))
3148 interval = sd->balance_interval;
3149 if (idle != CPU_IDLE)
3150 interval *= sd->busy_factor;
3152 /* scale ms to jiffies */
3153 interval = msecs_to_jiffies(interval);
3154 if (unlikely(!interval))
3156 if (interval > HZ*NR_CPUS/10)
3157 interval = HZ*NR_CPUS/10;
3160 if (sd->flags & SD_SERIALIZE) {
3161 if (!spin_trylock(&balancing))
3165 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3166 if (load_balance(cpu, rq, sd, idle, &balance)) {
3168 * We've pulled tasks over so either we're no
3169 * longer idle, or one of our SMT siblings is
3172 idle = CPU_NOT_IDLE;
3174 sd->last_balance = jiffies;
3176 if (sd->flags & SD_SERIALIZE)
3177 spin_unlock(&balancing);
3179 if (time_after(next_balance, sd->last_balance + interval)) {
3180 next_balance = sd->last_balance + interval;
3181 update_next_balance = 1;
3185 * Stop the load balance at this level. There is another
3186 * CPU in our sched group which is doing load balancing more
3194 * next_balance will be updated only when there is a need.
3195 * When the cpu is attached to null domain for ex, it will not be
3198 if (likely(update_next_balance))
3199 rq->next_balance = next_balance;
3203 * run_rebalance_domains is triggered when needed from the scheduler tick.
3204 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3205 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3207 static void run_rebalance_domains(struct softirq_action *h)
3209 int this_cpu = smp_processor_id();
3210 struct rq *this_rq = cpu_rq(this_cpu);
3211 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3212 CPU_IDLE : CPU_NOT_IDLE;
3214 rebalance_domains(this_cpu, idle);
3218 * If this cpu is the owner for idle load balancing, then do the
3219 * balancing on behalf of the other idle cpus whose ticks are
3222 if (this_rq->idle_at_tick &&
3223 atomic_read(&nohz.load_balancer) == this_cpu) {
3224 cpumask_t cpus = nohz.cpu_mask;
3228 cpu_clear(this_cpu, cpus);
3229 for_each_cpu_mask(balance_cpu, cpus) {
3231 * If this cpu gets work to do, stop the load balancing
3232 * work being done for other cpus. Next load
3233 * balancing owner will pick it up.
3238 rebalance_domains(balance_cpu, CPU_IDLE);
3240 rq = cpu_rq(balance_cpu);
3241 if (time_after(this_rq->next_balance, rq->next_balance))
3242 this_rq->next_balance = rq->next_balance;
3249 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3251 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3252 * idle load balancing owner or decide to stop the periodic load balancing,
3253 * if the whole system is idle.
3255 static inline void trigger_load_balance(struct rq *rq, int cpu)
3259 * If we were in the nohz mode recently and busy at the current
3260 * scheduler tick, then check if we need to nominate new idle
3263 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3264 rq->in_nohz_recently = 0;
3266 if (atomic_read(&nohz.load_balancer) == cpu) {
3267 cpu_clear(cpu, nohz.cpu_mask);
3268 atomic_set(&nohz.load_balancer, -1);
3271 if (atomic_read(&nohz.load_balancer) == -1) {
3273 * simple selection for now: Nominate the
3274 * first cpu in the nohz list to be the next
3277 * TBD: Traverse the sched domains and nominate
3278 * the nearest cpu in the nohz.cpu_mask.
3280 int ilb = first_cpu(nohz.cpu_mask);
3288 * If this cpu is idle and doing idle load balancing for all the
3289 * cpus with ticks stopped, is it time for that to stop?
3291 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3292 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3298 * If this cpu is idle and the idle load balancing is done by
3299 * someone else, then no need raise the SCHED_SOFTIRQ
3301 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3302 cpu_isset(cpu, nohz.cpu_mask))
3305 if (time_after_eq(jiffies, rq->next_balance))
3306 raise_softirq(SCHED_SOFTIRQ);
3309 #else /* CONFIG_SMP */
3312 * on UP we do not need to balance between CPUs:
3314 static inline void idle_balance(int cpu, struct rq *rq)
3320 DEFINE_PER_CPU(struct kernel_stat, kstat);
3322 EXPORT_PER_CPU_SYMBOL(kstat);
3325 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3326 * that have not yet been banked in case the task is currently running.
3328 unsigned long long task_sched_runtime(struct task_struct *p)
3330 unsigned long flags;
3334 rq = task_rq_lock(p, &flags);
3335 ns = p->se.sum_exec_runtime;
3336 if (task_current(rq, p)) {
3337 update_rq_clock(rq);
3338 delta_exec = rq->clock - p->se.exec_start;
3339 if ((s64)delta_exec > 0)
3342 task_rq_unlock(rq, &flags);
3348 * Account user cpu time to a process.
3349 * @p: the process that the cpu time gets accounted to
3350 * @cputime: the cpu time spent in user space since the last update
3352 void account_user_time(struct task_struct *p, cputime_t cputime)
3354 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3357 p->utime = cputime_add(p->utime, cputime);
3359 /* Add user time to cpustat. */
3360 tmp = cputime_to_cputime64(cputime);
3361 if (TASK_NICE(p) > 0)
3362 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3364 cpustat->user = cputime64_add(cpustat->user, tmp);
3368 * Account guest cpu time to a process.
3369 * @p: the process that the cpu time gets accounted to
3370 * @cputime: the cpu time spent in virtual machine since the last update
3372 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3375 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3377 tmp = cputime_to_cputime64(cputime);
3379 p->utime = cputime_add(p->utime, cputime);
3380 p->gtime = cputime_add(p->gtime, cputime);
3382 cpustat->user = cputime64_add(cpustat->user, tmp);
3383 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3387 * Account scaled user cpu time to a process.
3388 * @p: the process that the cpu time gets accounted to
3389 * @cputime: the cpu time spent in user space since the last update
3391 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3393 p->utimescaled = cputime_add(p->utimescaled, cputime);
3397 * Account system cpu time to a process.
3398 * @p: the process that the cpu time gets accounted to
3399 * @hardirq_offset: the offset to subtract from hardirq_count()
3400 * @cputime: the cpu time spent in kernel space since the last update
3402 void account_system_time(struct task_struct *p, int hardirq_offset,
3405 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3406 struct rq *rq = this_rq();
3409 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3410 return account_guest_time(p, cputime);
3412 p->stime = cputime_add(p->stime, cputime);
3414 /* Add system time to cpustat. */
3415 tmp = cputime_to_cputime64(cputime);
3416 if (hardirq_count() - hardirq_offset)
3417 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3418 else if (softirq_count())
3419 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3420 else if (p != rq->idle)
3421 cpustat->system = cputime64_add(cpustat->system, tmp);
3422 else if (atomic_read(&rq->nr_iowait) > 0)
3423 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3425 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3426 /* Account for system time used */
3427 acct_update_integrals(p);
3431 * Account scaled system cpu time to a process.
3432 * @p: the process that the cpu time gets accounted to
3433 * @hardirq_offset: the offset to subtract from hardirq_count()
3434 * @cputime: the cpu time spent in kernel space since the last update
3436 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3438 p->stimescaled = cputime_add(p->stimescaled, cputime);
3442 * Account for involuntary wait time.
3443 * @p: the process from which the cpu time has been stolen
3444 * @steal: the cpu time spent in involuntary wait
3446 void account_steal_time(struct task_struct *p, cputime_t steal)
3448 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3449 cputime64_t tmp = cputime_to_cputime64(steal);
3450 struct rq *rq = this_rq();
3452 if (p == rq->idle) {
3453 p->stime = cputime_add(p->stime, steal);
3454 if (atomic_read(&rq->nr_iowait) > 0)
3455 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3457 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3459 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3463 * This function gets called by the timer code, with HZ frequency.
3464 * We call it with interrupts disabled.
3466 * It also gets called by the fork code, when changing the parent's
3469 void scheduler_tick(void)
3471 int cpu = smp_processor_id();
3472 struct rq *rq = cpu_rq(cpu);
3473 struct task_struct *curr = rq->curr;
3474 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3476 spin_lock(&rq->lock);
3477 __update_rq_clock(rq);
3479 * Let rq->clock advance by at least TICK_NSEC:
3481 if (unlikely(rq->clock < next_tick))
3482 rq->clock = next_tick;
3483 rq->tick_timestamp = rq->clock;
3484 update_cpu_load(rq);
3485 if (curr != rq->idle) /* FIXME: needed? */
3486 curr->sched_class->task_tick(rq, curr);
3487 spin_unlock(&rq->lock);
3490 rq->idle_at_tick = idle_cpu(cpu);
3491 trigger_load_balance(rq, cpu);
3495 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3497 void fastcall add_preempt_count(int val)
3502 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3504 preempt_count() += val;
3506 * Spinlock count overflowing soon?
3508 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3511 EXPORT_SYMBOL(add_preempt_count);
3513 void fastcall sub_preempt_count(int val)
3518 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3521 * Is the spinlock portion underflowing?
3523 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3524 !(preempt_count() & PREEMPT_MASK)))
3527 preempt_count() -= val;
3529 EXPORT_SYMBOL(sub_preempt_count);
3534 * Print scheduling while atomic bug:
3536 static noinline void __schedule_bug(struct task_struct *prev)
3538 struct pt_regs *regs = get_irq_regs();
3540 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3541 prev->comm, prev->pid, preempt_count());
3543 debug_show_held_locks(prev);
3544 if (irqs_disabled())
3545 print_irqtrace_events(prev);
3554 * Various schedule()-time debugging checks and statistics:
3556 static inline void schedule_debug(struct task_struct *prev)
3559 * Test if we are atomic. Since do_exit() needs to call into
3560 * schedule() atomically, we ignore that path for now.
3561 * Otherwise, whine if we are scheduling when we should not be.
3563 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3564 __schedule_bug(prev);
3566 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3568 schedstat_inc(this_rq(), sched_count);
3569 #ifdef CONFIG_SCHEDSTATS
3570 if (unlikely(prev->lock_depth >= 0)) {
3571 schedstat_inc(this_rq(), bkl_count);
3572 schedstat_inc(prev, sched_info.bkl_count);
3578 * Pick up the highest-prio task:
3580 static inline struct task_struct *
3581 pick_next_task(struct rq *rq, struct task_struct *prev)
3583 const struct sched_class *class;
3584 struct task_struct *p;
3587 * Optimization: we know that if all tasks are in
3588 * the fair class we can call that function directly:
3590 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3591 p = fair_sched_class.pick_next_task(rq);
3596 class = sched_class_highest;
3598 p = class->pick_next_task(rq);
3602 * Will never be NULL as the idle class always
3603 * returns a non-NULL p:
3605 class = class->next;
3610 * schedule() is the main scheduler function.
3612 asmlinkage void __sched schedule(void)
3614 struct task_struct *prev, *next;
3621 cpu = smp_processor_id();
3625 switch_count = &prev->nivcsw;
3627 release_kernel_lock(prev);
3628 need_resched_nonpreemptible:
3630 schedule_debug(prev);
3633 * Do the rq-clock update outside the rq lock:
3635 local_irq_disable();
3636 __update_rq_clock(rq);
3637 spin_lock(&rq->lock);
3638 clear_tsk_need_resched(prev);
3640 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3641 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3642 unlikely(signal_pending(prev)))) {
3643 prev->state = TASK_RUNNING;
3645 deactivate_task(rq, prev, 1);
3647 switch_count = &prev->nvcsw;
3651 if (prev->sched_class->pre_schedule)
3652 prev->sched_class->pre_schedule(rq, prev);
3655 if (unlikely(!rq->nr_running))
3656 idle_balance(cpu, rq);
3658 prev->sched_class->put_prev_task(rq, prev);
3659 next = pick_next_task(rq, prev);
3661 sched_info_switch(prev, next);
3663 if (likely(prev != next)) {
3668 context_switch(rq, prev, next); /* unlocks the rq */
3670 spin_unlock_irq(&rq->lock);
3672 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3673 cpu = smp_processor_id();
3675 goto need_resched_nonpreemptible;
3677 preempt_enable_no_resched();
3678 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3681 EXPORT_SYMBOL(schedule);
3683 #ifdef CONFIG_PREEMPT
3685 * this is the entry point to schedule() from in-kernel preemption
3686 * off of preempt_enable. Kernel preemptions off return from interrupt
3687 * occur there and call schedule directly.
3689 asmlinkage void __sched preempt_schedule(void)
3691 struct thread_info *ti = current_thread_info();
3692 #ifdef CONFIG_PREEMPT_BKL
3693 struct task_struct *task = current;
3694 int saved_lock_depth;
3697 * If there is a non-zero preempt_count or interrupts are disabled,
3698 * we do not want to preempt the current task. Just return..
3700 if (likely(ti->preempt_count || irqs_disabled()))
3704 add_preempt_count(PREEMPT_ACTIVE);
3707 * We keep the big kernel semaphore locked, but we
3708 * clear ->lock_depth so that schedule() doesnt
3709 * auto-release the semaphore:
3711 #ifdef CONFIG_PREEMPT_BKL
3712 saved_lock_depth = task->lock_depth;
3713 task->lock_depth = -1;
3716 #ifdef CONFIG_PREEMPT_BKL
3717 task->lock_depth = saved_lock_depth;
3719 sub_preempt_count(PREEMPT_ACTIVE);
3722 * Check again in case we missed a preemption opportunity
3723 * between schedule and now.
3726 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3728 EXPORT_SYMBOL(preempt_schedule);
3731 * this is the entry point to schedule() from kernel preemption
3732 * off of irq context.
3733 * Note, that this is called and return with irqs disabled. This will
3734 * protect us against recursive calling from irq.
3736 asmlinkage void __sched preempt_schedule_irq(void)
3738 struct thread_info *ti = current_thread_info();
3739 #ifdef CONFIG_PREEMPT_BKL
3740 struct task_struct *task = current;
3741 int saved_lock_depth;
3743 /* Catch callers which need to be fixed */
3744 BUG_ON(ti->preempt_count || !irqs_disabled());
3747 add_preempt_count(PREEMPT_ACTIVE);
3750 * We keep the big kernel semaphore locked, but we
3751 * clear ->lock_depth so that schedule() doesnt
3752 * auto-release the semaphore:
3754 #ifdef CONFIG_PREEMPT_BKL
3755 saved_lock_depth = task->lock_depth;
3756 task->lock_depth = -1;
3760 local_irq_disable();
3761 #ifdef CONFIG_PREEMPT_BKL
3762 task->lock_depth = saved_lock_depth;
3764 sub_preempt_count(PREEMPT_ACTIVE);
3767 * Check again in case we missed a preemption opportunity
3768 * between schedule and now.
3771 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3774 #endif /* CONFIG_PREEMPT */
3776 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3779 return try_to_wake_up(curr->private, mode, sync);
3781 EXPORT_SYMBOL(default_wake_function);
3784 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3785 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3786 * number) then we wake all the non-exclusive tasks and one exclusive task.
3788 * There are circumstances in which we can try to wake a task which has already
3789 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3790 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3792 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3793 int nr_exclusive, int sync, void *key)
3795 wait_queue_t *curr, *next;
3797 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3798 unsigned flags = curr->flags;
3800 if (curr->func(curr, mode, sync, key) &&
3801 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3807 * __wake_up - wake up threads blocked on a waitqueue.
3809 * @mode: which threads
3810 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3811 * @key: is directly passed to the wakeup function
3813 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3814 int nr_exclusive, void *key)
3816 unsigned long flags;
3818 spin_lock_irqsave(&q->lock, flags);
3819 __wake_up_common(q, mode, nr_exclusive, 0, key);
3820 spin_unlock_irqrestore(&q->lock, flags);
3822 EXPORT_SYMBOL(__wake_up);
3825 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3827 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3829 __wake_up_common(q, mode, 1, 0, NULL);
3833 * __wake_up_sync - wake up threads blocked on a waitqueue.
3835 * @mode: which threads
3836 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3838 * The sync wakeup differs that the waker knows that it will schedule
3839 * away soon, so while the target thread will be woken up, it will not
3840 * be migrated to another CPU - ie. the two threads are 'synchronized'
3841 * with each other. This can prevent needless bouncing between CPUs.
3843 * On UP it can prevent extra preemption.
3846 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3848 unsigned long flags;
3854 if (unlikely(!nr_exclusive))
3857 spin_lock_irqsave(&q->lock, flags);
3858 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3859 spin_unlock_irqrestore(&q->lock, flags);
3861 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3863 void complete(struct completion *x)
3865 unsigned long flags;
3867 spin_lock_irqsave(&x->wait.lock, flags);
3869 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3871 spin_unlock_irqrestore(&x->wait.lock, flags);
3873 EXPORT_SYMBOL(complete);
3875 void complete_all(struct completion *x)
3877 unsigned long flags;
3879 spin_lock_irqsave(&x->wait.lock, flags);
3880 x->done += UINT_MAX/2;
3881 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3883 spin_unlock_irqrestore(&x->wait.lock, flags);
3885 EXPORT_SYMBOL(complete_all);
3887 static inline long __sched
3888 do_wait_for_common(struct completion *x, long timeout, int state)
3891 DECLARE_WAITQUEUE(wait, current);
3893 wait.flags |= WQ_FLAG_EXCLUSIVE;
3894 __add_wait_queue_tail(&x->wait, &wait);
3896 if (state == TASK_INTERRUPTIBLE &&
3897 signal_pending(current)) {
3898 __remove_wait_queue(&x->wait, &wait);
3899 return -ERESTARTSYS;
3901 __set_current_state(state);
3902 spin_unlock_irq(&x->wait.lock);
3903 timeout = schedule_timeout(timeout);
3904 spin_lock_irq(&x->wait.lock);
3906 __remove_wait_queue(&x->wait, &wait);
3910 __remove_wait_queue(&x->wait, &wait);
3917 wait_for_common(struct completion *x, long timeout, int state)
3921 spin_lock_irq(&x->wait.lock);
3922 timeout = do_wait_for_common(x, timeout, state);
3923 spin_unlock_irq(&x->wait.lock);
3927 void __sched wait_for_completion(struct completion *x)
3929 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3931 EXPORT_SYMBOL(wait_for_completion);
3933 unsigned long __sched
3934 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3936 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3938 EXPORT_SYMBOL(wait_for_completion_timeout);
3940 int __sched wait_for_completion_interruptible(struct completion *x)
3942 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3943 if (t == -ERESTARTSYS)
3947 EXPORT_SYMBOL(wait_for_completion_interruptible);
3949 unsigned long __sched
3950 wait_for_completion_interruptible_timeout(struct completion *x,
3951 unsigned long timeout)
3953 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3955 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3958 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3960 unsigned long flags;
3963 init_waitqueue_entry(&wait, current);
3965 __set_current_state(state);
3967 spin_lock_irqsave(&q->lock, flags);
3968 __add_wait_queue(q, &wait);
3969 spin_unlock(&q->lock);
3970 timeout = schedule_timeout(timeout);
3971 spin_lock_irq(&q->lock);
3972 __remove_wait_queue(q, &wait);
3973 spin_unlock_irqrestore(&q->lock, flags);
3978 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3980 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3982 EXPORT_SYMBOL(interruptible_sleep_on);
3985 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3987 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3989 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3991 void __sched sleep_on(wait_queue_head_t *q)
3993 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3995 EXPORT_SYMBOL(sleep_on);
3997 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3999 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4001 EXPORT_SYMBOL(sleep_on_timeout);
4003 #ifdef CONFIG_RT_MUTEXES
4006 * rt_mutex_setprio - set the current priority of a task
4008 * @prio: prio value (kernel-internal form)
4010 * This function changes the 'effective' priority of a task. It does
4011 * not touch ->normal_prio like __setscheduler().
4013 * Used by the rt_mutex code to implement priority inheritance logic.
4015 void rt_mutex_setprio(struct task_struct *p, int prio)
4017 unsigned long flags;
4018 int oldprio, on_rq, running;
4021 BUG_ON(prio < 0 || prio > MAX_PRIO);
4023 rq = task_rq_lock(p, &flags);
4024 update_rq_clock(rq);
4027 on_rq = p->se.on_rq;
4028 running = task_current(rq, p);
4030 dequeue_task(rq, p, 0);
4032 p->sched_class->put_prev_task(rq, p);
4036 p->sched_class = &rt_sched_class;
4038 p->sched_class = &fair_sched_class;
4044 p->sched_class->set_curr_task(rq);
4045 enqueue_task(rq, p, 0);
4047 * Reschedule if we are currently running on this runqueue and
4048 * our priority decreased, or if we are not currently running on
4049 * this runqueue and our priority is higher than the current's
4052 if (p->prio > oldprio)
4053 resched_task(rq->curr);
4055 check_preempt_curr(rq, p);
4058 task_rq_unlock(rq, &flags);
4063 void set_user_nice(struct task_struct *p, long nice)
4065 int old_prio, delta, on_rq;
4066 unsigned long flags;
4069 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4072 * We have to be careful, if called from sys_setpriority(),
4073 * the task might be in the middle of scheduling on another CPU.
4075 rq = task_rq_lock(p, &flags);
4076 update_rq_clock(rq);
4078 * The RT priorities are set via sched_setscheduler(), but we still
4079 * allow the 'normal' nice value to be set - but as expected
4080 * it wont have any effect on scheduling until the task is
4081 * SCHED_FIFO/SCHED_RR:
4083 if (task_has_rt_policy(p)) {
4084 p->static_prio = NICE_TO_PRIO(nice);
4087 on_rq = p->se.on_rq;
4089 dequeue_task(rq, p, 0);
4091 p->static_prio = NICE_TO_PRIO(nice);
4094 p->prio = effective_prio(p);
4095 delta = p->prio - old_prio;
4098 enqueue_task(rq, p, 0);
4100 * If the task increased its priority or is running and
4101 * lowered its priority, then reschedule its CPU:
4103 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4104 resched_task(rq->curr);
4107 task_rq_unlock(rq, &flags);
4109 EXPORT_SYMBOL(set_user_nice);
4112 * can_nice - check if a task can reduce its nice value
4116 int can_nice(const struct task_struct *p, const int nice)
4118 /* convert nice value [19,-20] to rlimit style value [1,40] */
4119 int nice_rlim = 20 - nice;
4121 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4122 capable(CAP_SYS_NICE));
4125 #ifdef __ARCH_WANT_SYS_NICE
4128 * sys_nice - change the priority of the current process.
4129 * @increment: priority increment
4131 * sys_setpriority is a more generic, but much slower function that
4132 * does similar things.
4134 asmlinkage long sys_nice(int increment)
4139 * Setpriority might change our priority at the same moment.
4140 * We don't have to worry. Conceptually one call occurs first
4141 * and we have a single winner.
4143 if (increment < -40)
4148 nice = PRIO_TO_NICE(current->static_prio) + increment;
4154 if (increment < 0 && !can_nice(current, nice))
4157 retval = security_task_setnice(current, nice);
4161 set_user_nice(current, nice);
4168 * task_prio - return the priority value of a given task.
4169 * @p: the task in question.
4171 * This is the priority value as seen by users in /proc.
4172 * RT tasks are offset by -200. Normal tasks are centered
4173 * around 0, value goes from -16 to +15.
4175 int task_prio(const struct task_struct *p)
4177 return p->prio - MAX_RT_PRIO;
4181 * task_nice - return the nice value of a given task.
4182 * @p: the task in question.
4184 int task_nice(const struct task_struct *p)
4186 return TASK_NICE(p);
4188 EXPORT_SYMBOL_GPL(task_nice);
4191 * idle_cpu - is a given cpu idle currently?
4192 * @cpu: the processor in question.
4194 int idle_cpu(int cpu)
4196 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4200 * idle_task - return the idle task for a given cpu.
4201 * @cpu: the processor in question.
4203 struct task_struct *idle_task(int cpu)
4205 return cpu_rq(cpu)->idle;
4209 * find_process_by_pid - find a process with a matching PID value.
4210 * @pid: the pid in question.
4212 static struct task_struct *find_process_by_pid(pid_t pid)
4214 return pid ? find_task_by_vpid(pid) : current;
4217 /* Actually do priority change: must hold rq lock. */
4219 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4221 BUG_ON(p->se.on_rq);
4224 switch (p->policy) {
4228 p->sched_class = &fair_sched_class;
4232 p->sched_class = &rt_sched_class;
4236 p->rt_priority = prio;
4237 p->normal_prio = normal_prio(p);
4238 /* we are holding p->pi_lock already */
4239 p->prio = rt_mutex_getprio(p);
4244 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4245 * @p: the task in question.
4246 * @policy: new policy.
4247 * @param: structure containing the new RT priority.
4249 * NOTE that the task may be already dead.
4251 int sched_setscheduler(struct task_struct *p, int policy,
4252 struct sched_param *param)
4254 int retval, oldprio, oldpolicy = -1, on_rq, running;
4255 unsigned long flags;
4258 /* may grab non-irq protected spin_locks */
4259 BUG_ON(in_interrupt());
4261 /* double check policy once rq lock held */
4263 policy = oldpolicy = p->policy;
4264 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4265 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4266 policy != SCHED_IDLE)
4269 * Valid priorities for SCHED_FIFO and SCHED_RR are
4270 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4271 * SCHED_BATCH and SCHED_IDLE is 0.
4273 if (param->sched_priority < 0 ||
4274 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4275 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4277 if (rt_policy(policy) != (param->sched_priority != 0))
4281 * Allow unprivileged RT tasks to decrease priority:
4283 if (!capable(CAP_SYS_NICE)) {
4284 if (rt_policy(policy)) {
4285 unsigned long rlim_rtprio;
4287 if (!lock_task_sighand(p, &flags))
4289 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4290 unlock_task_sighand(p, &flags);
4292 /* can't set/change the rt policy */
4293 if (policy != p->policy && !rlim_rtprio)
4296 /* can't increase priority */
4297 if (param->sched_priority > p->rt_priority &&
4298 param->sched_priority > rlim_rtprio)
4302 * Like positive nice levels, dont allow tasks to
4303 * move out of SCHED_IDLE either:
4305 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4308 /* can't change other user's priorities */
4309 if ((current->euid != p->euid) &&
4310 (current->euid != p->uid))
4314 retval = security_task_setscheduler(p, policy, param);
4318 * make sure no PI-waiters arrive (or leave) while we are
4319 * changing the priority of the task:
4321 spin_lock_irqsave(&p->pi_lock, flags);
4323 * To be able to change p->policy safely, the apropriate
4324 * runqueue lock must be held.
4326 rq = __task_rq_lock(p);
4327 /* recheck policy now with rq lock held */
4328 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4329 policy = oldpolicy = -1;
4330 __task_rq_unlock(rq);
4331 spin_unlock_irqrestore(&p->pi_lock, flags);
4334 update_rq_clock(rq);
4335 on_rq = p->se.on_rq;
4336 running = task_current(rq, p);
4338 deactivate_task(rq, p, 0);
4340 p->sched_class->put_prev_task(rq, p);
4344 __setscheduler(rq, p, policy, param->sched_priority);
4348 p->sched_class->set_curr_task(rq);
4349 activate_task(rq, p, 0);
4351 * Reschedule if we are currently running on this runqueue and
4352 * our priority decreased, or if we are not currently running on
4353 * this runqueue and our priority is higher than the current's
4356 if (p->prio > oldprio)
4357 resched_task(rq->curr);
4359 check_preempt_curr(rq, p);
4362 __task_rq_unlock(rq);
4363 spin_unlock_irqrestore(&p->pi_lock, flags);
4365 rt_mutex_adjust_pi(p);
4369 EXPORT_SYMBOL_GPL(sched_setscheduler);
4372 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4374 struct sched_param lparam;
4375 struct task_struct *p;
4378 if (!param || pid < 0)
4380 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4385 p = find_process_by_pid(pid);
4387 retval = sched_setscheduler(p, policy, &lparam);
4394 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4395 * @pid: the pid in question.
4396 * @policy: new policy.
4397 * @param: structure containing the new RT priority.
4400 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4402 /* negative values for policy are not valid */
4406 return do_sched_setscheduler(pid, policy, param);
4410 * sys_sched_setparam - set/change the RT priority of a thread
4411 * @pid: the pid in question.
4412 * @param: structure containing the new RT priority.
4414 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4416 return do_sched_setscheduler(pid, -1, param);
4420 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4421 * @pid: the pid in question.
4423 asmlinkage long sys_sched_getscheduler(pid_t pid)
4425 struct task_struct *p;
4432 read_lock(&tasklist_lock);
4433 p = find_process_by_pid(pid);
4435 retval = security_task_getscheduler(p);
4439 read_unlock(&tasklist_lock);
4444 * sys_sched_getscheduler - get the RT priority of a thread
4445 * @pid: the pid in question.
4446 * @param: structure containing the RT priority.
4448 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4450 struct sched_param lp;
4451 struct task_struct *p;
4454 if (!param || pid < 0)
4457 read_lock(&tasklist_lock);
4458 p = find_process_by_pid(pid);
4463 retval = security_task_getscheduler(p);
4467 lp.sched_priority = p->rt_priority;
4468 read_unlock(&tasklist_lock);
4471 * This one might sleep, we cannot do it with a spinlock held ...
4473 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4478 read_unlock(&tasklist_lock);
4482 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4484 cpumask_t cpus_allowed;
4485 struct task_struct *p;
4489 read_lock(&tasklist_lock);
4491 p = find_process_by_pid(pid);
4493 read_unlock(&tasklist_lock);
4499 * It is not safe to call set_cpus_allowed with the
4500 * tasklist_lock held. We will bump the task_struct's
4501 * usage count and then drop tasklist_lock.
4504 read_unlock(&tasklist_lock);
4507 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4508 !capable(CAP_SYS_NICE))
4511 retval = security_task_setscheduler(p, 0, NULL);
4515 cpus_allowed = cpuset_cpus_allowed(p);
4516 cpus_and(new_mask, new_mask, cpus_allowed);
4518 retval = set_cpus_allowed(p, new_mask);
4521 cpus_allowed = cpuset_cpus_allowed(p);
4522 if (!cpus_subset(new_mask, cpus_allowed)) {
4524 * We must have raced with a concurrent cpuset
4525 * update. Just reset the cpus_allowed to the
4526 * cpuset's cpus_allowed
4528 new_mask = cpus_allowed;
4538 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4539 cpumask_t *new_mask)
4541 if (len < sizeof(cpumask_t)) {
4542 memset(new_mask, 0, sizeof(cpumask_t));
4543 } else if (len > sizeof(cpumask_t)) {
4544 len = sizeof(cpumask_t);
4546 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4550 * sys_sched_setaffinity - set the cpu affinity of a process
4551 * @pid: pid of the process
4552 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4553 * @user_mask_ptr: user-space pointer to the new cpu mask
4555 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4556 unsigned long __user *user_mask_ptr)
4561 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4565 return sched_setaffinity(pid, new_mask);
4569 * Represents all cpu's present in the system
4570 * In systems capable of hotplug, this map could dynamically grow
4571 * as new cpu's are detected in the system via any platform specific
4572 * method, such as ACPI for e.g.
4575 cpumask_t cpu_present_map __read_mostly;
4576 EXPORT_SYMBOL(cpu_present_map);
4579 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4580 EXPORT_SYMBOL(cpu_online_map);
4582 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4583 EXPORT_SYMBOL(cpu_possible_map);
4586 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4588 struct task_struct *p;
4592 read_lock(&tasklist_lock);
4595 p = find_process_by_pid(pid);
4599 retval = security_task_getscheduler(p);
4603 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4606 read_unlock(&tasklist_lock);
4613 * sys_sched_getaffinity - get the cpu affinity of a process
4614 * @pid: pid of the process
4615 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4616 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4618 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4619 unsigned long __user *user_mask_ptr)
4624 if (len < sizeof(cpumask_t))
4627 ret = sched_getaffinity(pid, &mask);
4631 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4634 return sizeof(cpumask_t);
4638 * sys_sched_yield - yield the current processor to other threads.
4640 * This function yields the current CPU to other tasks. If there are no
4641 * other threads running on this CPU then this function will return.
4643 asmlinkage long sys_sched_yield(void)
4645 struct rq *rq = this_rq_lock();
4647 schedstat_inc(rq, yld_count);
4648 current->sched_class->yield_task(rq);
4651 * Since we are going to call schedule() anyway, there's
4652 * no need to preempt or enable interrupts:
4654 __release(rq->lock);
4655 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4656 _raw_spin_unlock(&rq->lock);
4657 preempt_enable_no_resched();
4664 static void __cond_resched(void)
4666 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4667 __might_sleep(__FILE__, __LINE__);
4670 * The BKS might be reacquired before we have dropped
4671 * PREEMPT_ACTIVE, which could trigger a second
4672 * cond_resched() call.
4675 add_preempt_count(PREEMPT_ACTIVE);
4677 sub_preempt_count(PREEMPT_ACTIVE);
4678 } while (need_resched());
4681 int __sched cond_resched(void)
4683 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4684 system_state == SYSTEM_RUNNING) {
4690 EXPORT_SYMBOL(cond_resched);
4693 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4694 * call schedule, and on return reacquire the lock.
4696 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4697 * operations here to prevent schedule() from being called twice (once via
4698 * spin_unlock(), once by hand).
4700 int cond_resched_lock(spinlock_t *lock)
4704 if (need_lockbreak(lock)) {
4710 if (need_resched() && system_state == SYSTEM_RUNNING) {
4711 spin_release(&lock->dep_map, 1, _THIS_IP_);
4712 _raw_spin_unlock(lock);
4713 preempt_enable_no_resched();
4720 EXPORT_SYMBOL(cond_resched_lock);
4722 int __sched cond_resched_softirq(void)
4724 BUG_ON(!in_softirq());
4726 if (need_resched() && system_state == SYSTEM_RUNNING) {
4734 EXPORT_SYMBOL(cond_resched_softirq);
4737 * yield - yield the current processor to other threads.
4739 * This is a shortcut for kernel-space yielding - it marks the
4740 * thread runnable and calls sys_sched_yield().
4742 void __sched yield(void)
4744 set_current_state(TASK_RUNNING);
4747 EXPORT_SYMBOL(yield);
4750 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4751 * that process accounting knows that this is a task in IO wait state.
4753 * But don't do that if it is a deliberate, throttling IO wait (this task
4754 * has set its backing_dev_info: the queue against which it should throttle)
4756 void __sched io_schedule(void)
4758 struct rq *rq = &__raw_get_cpu_var(runqueues);
4760 delayacct_blkio_start();
4761 atomic_inc(&rq->nr_iowait);
4763 atomic_dec(&rq->nr_iowait);
4764 delayacct_blkio_end();
4766 EXPORT_SYMBOL(io_schedule);
4768 long __sched io_schedule_timeout(long timeout)
4770 struct rq *rq = &__raw_get_cpu_var(runqueues);
4773 delayacct_blkio_start();
4774 atomic_inc(&rq->nr_iowait);
4775 ret = schedule_timeout(timeout);
4776 atomic_dec(&rq->nr_iowait);
4777 delayacct_blkio_end();
4782 * sys_sched_get_priority_max - return maximum RT priority.
4783 * @policy: scheduling class.
4785 * this syscall returns the maximum rt_priority that can be used
4786 * by a given scheduling class.
4788 asmlinkage long sys_sched_get_priority_max(int policy)
4795 ret = MAX_USER_RT_PRIO-1;
4807 * sys_sched_get_priority_min - return minimum RT priority.
4808 * @policy: scheduling class.
4810 * this syscall returns the minimum rt_priority that can be used
4811 * by a given scheduling class.
4813 asmlinkage long sys_sched_get_priority_min(int policy)
4831 * sys_sched_rr_get_interval - return the default timeslice of a process.
4832 * @pid: pid of the process.
4833 * @interval: userspace pointer to the timeslice value.
4835 * this syscall writes the default timeslice value of a given process
4836 * into the user-space timespec buffer. A value of '0' means infinity.
4839 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4841 struct task_struct *p;
4842 unsigned int time_slice;
4850 read_lock(&tasklist_lock);
4851 p = find_process_by_pid(pid);
4855 retval = security_task_getscheduler(p);
4860 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4861 * tasks that are on an otherwise idle runqueue:
4864 if (p->policy == SCHED_RR) {
4865 time_slice = DEF_TIMESLICE;
4867 struct sched_entity *se = &p->se;
4868 unsigned long flags;
4871 rq = task_rq_lock(p, &flags);
4872 if (rq->cfs.load.weight)
4873 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4874 task_rq_unlock(rq, &flags);
4876 read_unlock(&tasklist_lock);
4877 jiffies_to_timespec(time_slice, &t);
4878 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4882 read_unlock(&tasklist_lock);
4886 static const char stat_nam[] = "RSDTtZX";
4888 void sched_show_task(struct task_struct *p)
4890 unsigned long free = 0;
4893 state = p->state ? __ffs(p->state) + 1 : 0;
4894 printk(KERN_INFO "%-13.13s %c", p->comm,
4895 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4896 #if BITS_PER_LONG == 32
4897 if (state == TASK_RUNNING)
4898 printk(KERN_CONT " running ");
4900 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4902 if (state == TASK_RUNNING)
4903 printk(KERN_CONT " running task ");
4905 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4907 #ifdef CONFIG_DEBUG_STACK_USAGE
4909 unsigned long *n = end_of_stack(p);
4912 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4915 printk(KERN_CONT "%5lu %5d %6d\n", free,
4916 task_pid_nr(p), task_pid_nr(p->real_parent));
4918 if (state != TASK_RUNNING)
4919 show_stack(p, NULL);
4922 void show_state_filter(unsigned long state_filter)
4924 struct task_struct *g, *p;
4926 #if BITS_PER_LONG == 32
4928 " task PC stack pid father\n");
4931 " task PC stack pid father\n");
4933 read_lock(&tasklist_lock);
4934 do_each_thread(g, p) {
4936 * reset the NMI-timeout, listing all files on a slow
4937 * console might take alot of time:
4939 touch_nmi_watchdog();
4940 if (!state_filter || (p->state & state_filter))
4942 } while_each_thread(g, p);
4944 touch_all_softlockup_watchdogs();
4946 #ifdef CONFIG_SCHED_DEBUG
4947 sysrq_sched_debug_show();
4949 read_unlock(&tasklist_lock);
4951 * Only show locks if all tasks are dumped:
4953 if (state_filter == -1)
4954 debug_show_all_locks();
4957 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4959 idle->sched_class = &idle_sched_class;
4963 * init_idle - set up an idle thread for a given CPU
4964 * @idle: task in question
4965 * @cpu: cpu the idle task belongs to
4967 * NOTE: this function does not set the idle thread's NEED_RESCHED
4968 * flag, to make booting more robust.
4970 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4972 struct rq *rq = cpu_rq(cpu);
4973 unsigned long flags;
4976 idle->se.exec_start = sched_clock();
4978 idle->prio = idle->normal_prio = MAX_PRIO;
4979 idle->cpus_allowed = cpumask_of_cpu(cpu);
4980 __set_task_cpu(idle, cpu);
4982 spin_lock_irqsave(&rq->lock, flags);
4983 rq->curr = rq->idle = idle;
4984 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4987 spin_unlock_irqrestore(&rq->lock, flags);
4989 /* Set the preempt count _outside_ the spinlocks! */
4990 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4991 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4993 task_thread_info(idle)->preempt_count = 0;
4996 * The idle tasks have their own, simple scheduling class:
4998 idle->sched_class = &idle_sched_class;
5002 * In a system that switches off the HZ timer nohz_cpu_mask
5003 * indicates which cpus entered this state. This is used
5004 * in the rcu update to wait only for active cpus. For system
5005 * which do not switch off the HZ timer nohz_cpu_mask should
5006 * always be CPU_MASK_NONE.
5008 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5011 * Increase the granularity value when there are more CPUs,
5012 * because with more CPUs the 'effective latency' as visible
5013 * to users decreases. But the relationship is not linear,
5014 * so pick a second-best guess by going with the log2 of the
5017 * This idea comes from the SD scheduler of Con Kolivas:
5019 static inline void sched_init_granularity(void)
5021 unsigned int factor = 1 + ilog2(num_online_cpus());
5022 const unsigned long limit = 200000000;
5024 sysctl_sched_min_granularity *= factor;
5025 if (sysctl_sched_min_granularity > limit)
5026 sysctl_sched_min_granularity = limit;
5028 sysctl_sched_latency *= factor;
5029 if (sysctl_sched_latency > limit)
5030 sysctl_sched_latency = limit;
5032 sysctl_sched_wakeup_granularity *= factor;
5033 sysctl_sched_batch_wakeup_granularity *= factor;
5038 * This is how migration works:
5040 * 1) we queue a struct migration_req structure in the source CPU's
5041 * runqueue and wake up that CPU's migration thread.
5042 * 2) we down() the locked semaphore => thread blocks.
5043 * 3) migration thread wakes up (implicitly it forces the migrated
5044 * thread off the CPU)
5045 * 4) it gets the migration request and checks whether the migrated
5046 * task is still in the wrong runqueue.
5047 * 5) if it's in the wrong runqueue then the migration thread removes
5048 * it and puts it into the right queue.
5049 * 6) migration thread up()s the semaphore.
5050 * 7) we wake up and the migration is done.
5054 * Change a given task's CPU affinity. Migrate the thread to a
5055 * proper CPU and schedule it away if the CPU it's executing on
5056 * is removed from the allowed bitmask.
5058 * NOTE: the caller must have a valid reference to the task, the
5059 * task must not exit() & deallocate itself prematurely. The
5060 * call is not atomic; no spinlocks may be held.
5062 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5064 struct migration_req req;
5065 unsigned long flags;
5069 rq = task_rq_lock(p, &flags);
5070 if (!cpus_intersects(new_mask, cpu_online_map)) {
5075 if (p->sched_class->set_cpus_allowed)
5076 p->sched_class->set_cpus_allowed(p, &new_mask);
5078 p->cpus_allowed = new_mask;
5079 p->nr_cpus_allowed = cpus_weight(new_mask);
5082 /* Can the task run on the task's current CPU? If so, we're done */
5083 if (cpu_isset(task_cpu(p), new_mask))
5086 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5087 /* Need help from migration thread: drop lock and wait. */
5088 task_rq_unlock(rq, &flags);
5089 wake_up_process(rq->migration_thread);
5090 wait_for_completion(&req.done);
5091 tlb_migrate_finish(p->mm);
5095 task_rq_unlock(rq, &flags);
5099 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5102 * Move (not current) task off this cpu, onto dest cpu. We're doing
5103 * this because either it can't run here any more (set_cpus_allowed()
5104 * away from this CPU, or CPU going down), or because we're
5105 * attempting to rebalance this task on exec (sched_exec).
5107 * So we race with normal scheduler movements, but that's OK, as long
5108 * as the task is no longer on this CPU.
5110 * Returns non-zero if task was successfully migrated.
5112 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5114 struct rq *rq_dest, *rq_src;
5117 if (unlikely(cpu_is_offline(dest_cpu)))
5120 rq_src = cpu_rq(src_cpu);
5121 rq_dest = cpu_rq(dest_cpu);
5123 double_rq_lock(rq_src, rq_dest);
5124 /* Already moved. */
5125 if (task_cpu(p) != src_cpu)
5127 /* Affinity changed (again). */
5128 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5131 on_rq = p->se.on_rq;
5133 deactivate_task(rq_src, p, 0);
5135 set_task_cpu(p, dest_cpu);
5137 activate_task(rq_dest, p, 0);
5138 check_preempt_curr(rq_dest, p);
5142 double_rq_unlock(rq_src, rq_dest);
5147 * migration_thread - this is a highprio system thread that performs
5148 * thread migration by bumping thread off CPU then 'pushing' onto
5151 static int migration_thread(void *data)
5153 int cpu = (long)data;
5157 BUG_ON(rq->migration_thread != current);
5159 set_current_state(TASK_INTERRUPTIBLE);
5160 while (!kthread_should_stop()) {
5161 struct migration_req *req;
5162 struct list_head *head;
5164 spin_lock_irq(&rq->lock);
5166 if (cpu_is_offline(cpu)) {
5167 spin_unlock_irq(&rq->lock);
5171 if (rq->active_balance) {
5172 active_load_balance(rq, cpu);
5173 rq->active_balance = 0;
5176 head = &rq->migration_queue;
5178 if (list_empty(head)) {
5179 spin_unlock_irq(&rq->lock);
5181 set_current_state(TASK_INTERRUPTIBLE);
5184 req = list_entry(head->next, struct migration_req, list);
5185 list_del_init(head->next);
5187 spin_unlock(&rq->lock);
5188 __migrate_task(req->task, cpu, req->dest_cpu);
5191 complete(&req->done);
5193 __set_current_state(TASK_RUNNING);
5197 /* Wait for kthread_stop */
5198 set_current_state(TASK_INTERRUPTIBLE);
5199 while (!kthread_should_stop()) {
5201 set_current_state(TASK_INTERRUPTIBLE);
5203 __set_current_state(TASK_RUNNING);
5207 #ifdef CONFIG_HOTPLUG_CPU
5209 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5213 local_irq_disable();
5214 ret = __migrate_task(p, src_cpu, dest_cpu);
5220 * Figure out where task on dead CPU should go, use force if necessary.
5221 * NOTE: interrupts should be disabled by the caller
5223 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5225 unsigned long flags;
5232 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5233 cpus_and(mask, mask, p->cpus_allowed);
5234 dest_cpu = any_online_cpu(mask);
5236 /* On any allowed CPU? */
5237 if (dest_cpu == NR_CPUS)
5238 dest_cpu = any_online_cpu(p->cpus_allowed);
5240 /* No more Mr. Nice Guy. */
5241 if (dest_cpu == NR_CPUS) {
5242 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5244 * Try to stay on the same cpuset, where the
5245 * current cpuset may be a subset of all cpus.
5246 * The cpuset_cpus_allowed_locked() variant of
5247 * cpuset_cpus_allowed() will not block. It must be
5248 * called within calls to cpuset_lock/cpuset_unlock.
5250 rq = task_rq_lock(p, &flags);
5251 p->cpus_allowed = cpus_allowed;
5252 dest_cpu = any_online_cpu(p->cpus_allowed);
5253 task_rq_unlock(rq, &flags);
5256 * Don't tell them about moving exiting tasks or
5257 * kernel threads (both mm NULL), since they never
5260 if (p->mm && printk_ratelimit()) {
5261 printk(KERN_INFO "process %d (%s) no "
5262 "longer affine to cpu%d\n",
5263 task_pid_nr(p), p->comm, dead_cpu);
5266 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5270 * While a dead CPU has no uninterruptible tasks queued at this point,
5271 * it might still have a nonzero ->nr_uninterruptible counter, because
5272 * for performance reasons the counter is not stricly tracking tasks to
5273 * their home CPUs. So we just add the counter to another CPU's counter,
5274 * to keep the global sum constant after CPU-down:
5276 static void migrate_nr_uninterruptible(struct rq *rq_src)
5278 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5279 unsigned long flags;
5281 local_irq_save(flags);
5282 double_rq_lock(rq_src, rq_dest);
5283 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5284 rq_src->nr_uninterruptible = 0;
5285 double_rq_unlock(rq_src, rq_dest);
5286 local_irq_restore(flags);
5289 /* Run through task list and migrate tasks from the dead cpu. */
5290 static void migrate_live_tasks(int src_cpu)
5292 struct task_struct *p, *t;
5294 read_lock(&tasklist_lock);
5296 do_each_thread(t, p) {
5300 if (task_cpu(p) == src_cpu)
5301 move_task_off_dead_cpu(src_cpu, p);
5302 } while_each_thread(t, p);
5304 read_unlock(&tasklist_lock);
5308 * Schedules idle task to be the next runnable task on current CPU.
5309 * It does so by boosting its priority to highest possible.
5310 * Used by CPU offline code.
5312 void sched_idle_next(void)
5314 int this_cpu = smp_processor_id();
5315 struct rq *rq = cpu_rq(this_cpu);
5316 struct task_struct *p = rq->idle;
5317 unsigned long flags;
5319 /* cpu has to be offline */
5320 BUG_ON(cpu_online(this_cpu));
5323 * Strictly not necessary since rest of the CPUs are stopped by now
5324 * and interrupts disabled on the current cpu.
5326 spin_lock_irqsave(&rq->lock, flags);
5328 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5330 update_rq_clock(rq);
5331 activate_task(rq, p, 0);
5333 spin_unlock_irqrestore(&rq->lock, flags);
5337 * Ensures that the idle task is using init_mm right before its cpu goes
5340 void idle_task_exit(void)
5342 struct mm_struct *mm = current->active_mm;
5344 BUG_ON(cpu_online(smp_processor_id()));
5347 switch_mm(mm, &init_mm, current);
5351 /* called under rq->lock with disabled interrupts */
5352 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5354 struct rq *rq = cpu_rq(dead_cpu);
5356 /* Must be exiting, otherwise would be on tasklist. */
5357 BUG_ON(!p->exit_state);
5359 /* Cannot have done final schedule yet: would have vanished. */
5360 BUG_ON(p->state == TASK_DEAD);
5365 * Drop lock around migration; if someone else moves it,
5366 * that's OK. No task can be added to this CPU, so iteration is
5369 spin_unlock_irq(&rq->lock);
5370 move_task_off_dead_cpu(dead_cpu, p);
5371 spin_lock_irq(&rq->lock);
5376 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5377 static void migrate_dead_tasks(unsigned int dead_cpu)
5379 struct rq *rq = cpu_rq(dead_cpu);
5380 struct task_struct *next;
5383 if (!rq->nr_running)
5385 update_rq_clock(rq);
5386 next = pick_next_task(rq, rq->curr);
5389 migrate_dead(dead_cpu, next);
5393 #endif /* CONFIG_HOTPLUG_CPU */
5395 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5397 static struct ctl_table sd_ctl_dir[] = {
5399 .procname = "sched_domain",
5405 static struct ctl_table sd_ctl_root[] = {
5407 .ctl_name = CTL_KERN,
5408 .procname = "kernel",
5410 .child = sd_ctl_dir,
5415 static struct ctl_table *sd_alloc_ctl_entry(int n)
5417 struct ctl_table *entry =
5418 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5423 static void sd_free_ctl_entry(struct ctl_table **tablep)
5425 struct ctl_table *entry;
5428 * In the intermediate directories, both the child directory and
5429 * procname are dynamically allocated and could fail but the mode
5430 * will always be set. In the lowest directory the names are
5431 * static strings and all have proc handlers.
5433 for (entry = *tablep; entry->mode; entry++) {
5435 sd_free_ctl_entry(&entry->child);
5436 if (entry->proc_handler == NULL)
5437 kfree(entry->procname);
5445 set_table_entry(struct ctl_table *entry,
5446 const char *procname, void *data, int maxlen,
5447 mode_t mode, proc_handler *proc_handler)
5449 entry->procname = procname;
5451 entry->maxlen = maxlen;
5453 entry->proc_handler = proc_handler;
5456 static struct ctl_table *
5457 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5459 struct ctl_table *table = sd_alloc_ctl_entry(12);
5464 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5465 sizeof(long), 0644, proc_doulongvec_minmax);
5466 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5467 sizeof(long), 0644, proc_doulongvec_minmax);
5468 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5473 sizeof(int), 0644, proc_dointvec_minmax);
5474 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5475 sizeof(int), 0644, proc_dointvec_minmax);
5476 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5477 sizeof(int), 0644, proc_dointvec_minmax);
5478 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5479 sizeof(int), 0644, proc_dointvec_minmax);
5480 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5481 sizeof(int), 0644, proc_dointvec_minmax);
5482 set_table_entry(&table[9], "cache_nice_tries",
5483 &sd->cache_nice_tries,
5484 sizeof(int), 0644, proc_dointvec_minmax);
5485 set_table_entry(&table[10], "flags", &sd->flags,
5486 sizeof(int), 0644, proc_dointvec_minmax);
5487 /* &table[11] is terminator */
5492 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5494 struct ctl_table *entry, *table;
5495 struct sched_domain *sd;
5496 int domain_num = 0, i;
5499 for_each_domain(cpu, sd)
5501 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5506 for_each_domain(cpu, sd) {
5507 snprintf(buf, 32, "domain%d", i);
5508 entry->procname = kstrdup(buf, GFP_KERNEL);
5510 entry->child = sd_alloc_ctl_domain_table(sd);
5517 static struct ctl_table_header *sd_sysctl_header;
5518 static void register_sched_domain_sysctl(void)
5520 int i, cpu_num = num_online_cpus();
5521 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5524 WARN_ON(sd_ctl_dir[0].child);
5525 sd_ctl_dir[0].child = entry;
5530 for_each_online_cpu(i) {
5531 snprintf(buf, 32, "cpu%d", i);
5532 entry->procname = kstrdup(buf, GFP_KERNEL);
5534 entry->child = sd_alloc_ctl_cpu_table(i);
5538 WARN_ON(sd_sysctl_header);
5539 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5542 /* may be called multiple times per register */
5543 static void unregister_sched_domain_sysctl(void)
5545 if (sd_sysctl_header)
5546 unregister_sysctl_table(sd_sysctl_header);
5547 sd_sysctl_header = NULL;
5548 if (sd_ctl_dir[0].child)
5549 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5552 static void register_sched_domain_sysctl(void)
5555 static void unregister_sched_domain_sysctl(void)
5561 * migration_call - callback that gets triggered when a CPU is added.
5562 * Here we can start up the necessary migration thread for the new CPU.
5564 static int __cpuinit
5565 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5567 struct task_struct *p;
5568 int cpu = (long)hcpu;
5569 unsigned long flags;
5574 case CPU_UP_PREPARE:
5575 case CPU_UP_PREPARE_FROZEN:
5576 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5579 kthread_bind(p, cpu);
5580 /* Must be high prio: stop_machine expects to yield to it. */
5581 rq = task_rq_lock(p, &flags);
5582 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5583 task_rq_unlock(rq, &flags);
5584 cpu_rq(cpu)->migration_thread = p;
5588 case CPU_ONLINE_FROZEN:
5589 /* Strictly unnecessary, as first user will wake it. */
5590 wake_up_process(cpu_rq(cpu)->migration_thread);
5592 /* Update our root-domain */
5594 spin_lock_irqsave(&rq->lock, flags);
5596 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5597 cpu_set(cpu, rq->rd->online);
5599 spin_unlock_irqrestore(&rq->lock, flags);
5602 #ifdef CONFIG_HOTPLUG_CPU
5603 case CPU_UP_CANCELED:
5604 case CPU_UP_CANCELED_FROZEN:
5605 if (!cpu_rq(cpu)->migration_thread)
5607 /* Unbind it from offline cpu so it can run. Fall thru. */
5608 kthread_bind(cpu_rq(cpu)->migration_thread,
5609 any_online_cpu(cpu_online_map));
5610 kthread_stop(cpu_rq(cpu)->migration_thread);
5611 cpu_rq(cpu)->migration_thread = NULL;
5615 case CPU_DEAD_FROZEN:
5616 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5617 migrate_live_tasks(cpu);
5619 kthread_stop(rq->migration_thread);
5620 rq->migration_thread = NULL;
5621 /* Idle task back to normal (off runqueue, low prio) */
5622 spin_lock_irq(&rq->lock);
5623 update_rq_clock(rq);
5624 deactivate_task(rq, rq->idle, 0);
5625 rq->idle->static_prio = MAX_PRIO;
5626 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5627 rq->idle->sched_class = &idle_sched_class;
5628 migrate_dead_tasks(cpu);
5629 spin_unlock_irq(&rq->lock);
5631 migrate_nr_uninterruptible(rq);
5632 BUG_ON(rq->nr_running != 0);
5635 * No need to migrate the tasks: it was best-effort if
5636 * they didn't take sched_hotcpu_mutex. Just wake up
5639 spin_lock_irq(&rq->lock);
5640 while (!list_empty(&rq->migration_queue)) {
5641 struct migration_req *req;
5643 req = list_entry(rq->migration_queue.next,
5644 struct migration_req, list);
5645 list_del_init(&req->list);
5646 complete(&req->done);
5648 spin_unlock_irq(&rq->lock);
5651 case CPU_DOWN_PREPARE:
5652 /* Update our root-domain */
5654 spin_lock_irqsave(&rq->lock, flags);
5656 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5657 cpu_clear(cpu, rq->rd->online);
5659 spin_unlock_irqrestore(&rq->lock, flags);
5666 /* Register at highest priority so that task migration (migrate_all_tasks)
5667 * happens before everything else.
5669 static struct notifier_block __cpuinitdata migration_notifier = {
5670 .notifier_call = migration_call,
5674 void __init migration_init(void)
5676 void *cpu = (void *)(long)smp_processor_id();
5679 /* Start one for the boot CPU: */
5680 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5681 BUG_ON(err == NOTIFY_BAD);
5682 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5683 register_cpu_notifier(&migration_notifier);
5689 /* Number of possible processor ids */
5690 int nr_cpu_ids __read_mostly = NR_CPUS;
5691 EXPORT_SYMBOL(nr_cpu_ids);
5693 #ifdef CONFIG_SCHED_DEBUG
5695 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5697 struct sched_group *group = sd->groups;
5698 cpumask_t groupmask;
5701 cpumask_scnprintf(str, NR_CPUS, sd->span);
5702 cpus_clear(groupmask);
5704 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5706 if (!(sd->flags & SD_LOAD_BALANCE)) {
5707 printk("does not load-balance\n");
5709 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5714 printk(KERN_CONT "span %s\n", str);
5716 if (!cpu_isset(cpu, sd->span)) {
5717 printk(KERN_ERR "ERROR: domain->span does not contain "
5720 if (!cpu_isset(cpu, group->cpumask)) {
5721 printk(KERN_ERR "ERROR: domain->groups does not contain"
5725 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5729 printk(KERN_ERR "ERROR: group is NULL\n");
5733 if (!group->__cpu_power) {
5734 printk(KERN_CONT "\n");
5735 printk(KERN_ERR "ERROR: domain->cpu_power not "
5740 if (!cpus_weight(group->cpumask)) {
5741 printk(KERN_CONT "\n");
5742 printk(KERN_ERR "ERROR: empty group\n");
5746 if (cpus_intersects(groupmask, group->cpumask)) {
5747 printk(KERN_CONT "\n");
5748 printk(KERN_ERR "ERROR: repeated CPUs\n");
5752 cpus_or(groupmask, groupmask, group->cpumask);
5754 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5755 printk(KERN_CONT " %s", str);
5757 group = group->next;
5758 } while (group != sd->groups);
5759 printk(KERN_CONT "\n");
5761 if (!cpus_equal(sd->span, groupmask))
5762 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5764 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5765 printk(KERN_ERR "ERROR: parent span is not a superset "
5766 "of domain->span\n");
5770 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5775 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5779 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5782 if (sched_domain_debug_one(sd, cpu, level))
5791 # define sched_domain_debug(sd, cpu) do { } while (0)
5794 static int sd_degenerate(struct sched_domain *sd)
5796 if (cpus_weight(sd->span) == 1)
5799 /* Following flags need at least 2 groups */
5800 if (sd->flags & (SD_LOAD_BALANCE |
5801 SD_BALANCE_NEWIDLE |
5805 SD_SHARE_PKG_RESOURCES)) {
5806 if (sd->groups != sd->groups->next)
5810 /* Following flags don't use groups */
5811 if (sd->flags & (SD_WAKE_IDLE |
5820 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5822 unsigned long cflags = sd->flags, pflags = parent->flags;
5824 if (sd_degenerate(parent))
5827 if (!cpus_equal(sd->span, parent->span))
5830 /* Does parent contain flags not in child? */
5831 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5832 if (cflags & SD_WAKE_AFFINE)
5833 pflags &= ~SD_WAKE_BALANCE;
5834 /* Flags needing groups don't count if only 1 group in parent */
5835 if (parent->groups == parent->groups->next) {
5836 pflags &= ~(SD_LOAD_BALANCE |
5837 SD_BALANCE_NEWIDLE |
5841 SD_SHARE_PKG_RESOURCES);
5843 if (~cflags & pflags)
5849 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5851 unsigned long flags;
5852 const struct sched_class *class;
5854 spin_lock_irqsave(&rq->lock, flags);
5857 struct root_domain *old_rd = rq->rd;
5859 for (class = sched_class_highest; class; class = class->next) {
5860 if (class->leave_domain)
5861 class->leave_domain(rq);
5864 if (atomic_dec_and_test(&old_rd->refcount))
5868 atomic_inc(&rd->refcount);
5871 for (class = sched_class_highest; class; class = class->next) {
5872 if (class->join_domain)
5873 class->join_domain(rq);
5876 spin_unlock_irqrestore(&rq->lock, flags);
5879 static void init_rootdomain(struct root_domain *rd, const cpumask_t *map)
5881 memset(rd, 0, sizeof(*rd));
5884 cpus_and(rd->online, rd->span, cpu_online_map);
5887 static void init_defrootdomain(void)
5889 cpumask_t cpus = CPU_MASK_ALL;
5891 init_rootdomain(&def_root_domain, &cpus);
5892 atomic_set(&def_root_domain.refcount, 1);
5895 static struct root_domain *alloc_rootdomain(const cpumask_t *map)
5897 struct root_domain *rd;
5899 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5903 init_rootdomain(rd, map);
5909 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5910 * hold the hotplug lock.
5913 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5915 struct rq *rq = cpu_rq(cpu);
5916 struct sched_domain *tmp;
5918 /* Remove the sched domains which do not contribute to scheduling. */
5919 for (tmp = sd; tmp; tmp = tmp->parent) {
5920 struct sched_domain *parent = tmp->parent;
5923 if (sd_parent_degenerate(tmp, parent)) {
5924 tmp->parent = parent->parent;
5926 parent->parent->child = tmp;
5930 if (sd && sd_degenerate(sd)) {
5936 sched_domain_debug(sd, cpu);
5938 rq_attach_root(rq, rd);
5939 rcu_assign_pointer(rq->sd, sd);
5942 /* cpus with isolated domains */
5943 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5945 /* Setup the mask of cpus configured for isolated domains */
5946 static int __init isolated_cpu_setup(char *str)
5948 int ints[NR_CPUS], i;
5950 str = get_options(str, ARRAY_SIZE(ints), ints);
5951 cpus_clear(cpu_isolated_map);
5952 for (i = 1; i <= ints[0]; i++)
5953 if (ints[i] < NR_CPUS)
5954 cpu_set(ints[i], cpu_isolated_map);
5958 __setup("isolcpus=", isolated_cpu_setup);
5961 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5962 * to a function which identifies what group(along with sched group) a CPU
5963 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5964 * (due to the fact that we keep track of groups covered with a cpumask_t).
5966 * init_sched_build_groups will build a circular linked list of the groups
5967 * covered by the given span, and will set each group's ->cpumask correctly,
5968 * and ->cpu_power to 0.
5971 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5972 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5973 struct sched_group **sg))
5975 struct sched_group *first = NULL, *last = NULL;
5976 cpumask_t covered = CPU_MASK_NONE;
5979 for_each_cpu_mask(i, span) {
5980 struct sched_group *sg;
5981 int group = group_fn(i, cpu_map, &sg);
5984 if (cpu_isset(i, covered))
5987 sg->cpumask = CPU_MASK_NONE;
5988 sg->__cpu_power = 0;
5990 for_each_cpu_mask(j, span) {
5991 if (group_fn(j, cpu_map, NULL) != group)
5994 cpu_set(j, covered);
5995 cpu_set(j, sg->cpumask);
6006 #define SD_NODES_PER_DOMAIN 16
6011 * find_next_best_node - find the next node to include in a sched_domain
6012 * @node: node whose sched_domain we're building
6013 * @used_nodes: nodes already in the sched_domain
6015 * Find the next node to include in a given scheduling domain. Simply
6016 * finds the closest node not already in the @used_nodes map.
6018 * Should use nodemask_t.
6020 static int find_next_best_node(int node, unsigned long *used_nodes)
6022 int i, n, val, min_val, best_node = 0;
6026 for (i = 0; i < MAX_NUMNODES; i++) {
6027 /* Start at @node */
6028 n = (node + i) % MAX_NUMNODES;
6030 if (!nr_cpus_node(n))
6033 /* Skip already used nodes */
6034 if (test_bit(n, used_nodes))
6037 /* Simple min distance search */
6038 val = node_distance(node, n);
6040 if (val < min_val) {
6046 set_bit(best_node, used_nodes);
6051 * sched_domain_node_span - get a cpumask for a node's sched_domain
6052 * @node: node whose cpumask we're constructing
6053 * @size: number of nodes to include in this span
6055 * Given a node, construct a good cpumask for its sched_domain to span. It
6056 * should be one that prevents unnecessary balancing, but also spreads tasks
6059 static cpumask_t sched_domain_node_span(int node)
6061 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6062 cpumask_t span, nodemask;
6066 bitmap_zero(used_nodes, MAX_NUMNODES);
6068 nodemask = node_to_cpumask(node);
6069 cpus_or(span, span, nodemask);
6070 set_bit(node, used_nodes);
6072 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6073 int next_node = find_next_best_node(node, used_nodes);
6075 nodemask = node_to_cpumask(next_node);
6076 cpus_or(span, span, nodemask);
6083 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6086 * SMT sched-domains:
6088 #ifdef CONFIG_SCHED_SMT
6089 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6090 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6093 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6096 *sg = &per_cpu(sched_group_cpus, cpu);
6102 * multi-core sched-domains:
6104 #ifdef CONFIG_SCHED_MC
6105 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6106 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6109 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6111 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6114 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6115 cpus_and(mask, mask, *cpu_map);
6116 group = first_cpu(mask);
6118 *sg = &per_cpu(sched_group_core, group);
6121 #elif defined(CONFIG_SCHED_MC)
6123 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6126 *sg = &per_cpu(sched_group_core, cpu);
6131 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6132 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6135 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6138 #ifdef CONFIG_SCHED_MC
6139 cpumask_t mask = cpu_coregroup_map(cpu);
6140 cpus_and(mask, mask, *cpu_map);
6141 group = first_cpu(mask);
6142 #elif defined(CONFIG_SCHED_SMT)
6143 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6144 cpus_and(mask, mask, *cpu_map);
6145 group = first_cpu(mask);
6150 *sg = &per_cpu(sched_group_phys, group);
6156 * The init_sched_build_groups can't handle what we want to do with node
6157 * groups, so roll our own. Now each node has its own list of groups which
6158 * gets dynamically allocated.
6160 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6161 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6163 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6164 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6166 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6167 struct sched_group **sg)
6169 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6172 cpus_and(nodemask, nodemask, *cpu_map);
6173 group = first_cpu(nodemask);
6176 *sg = &per_cpu(sched_group_allnodes, group);
6180 static void init_numa_sched_groups_power(struct sched_group *group_head)
6182 struct sched_group *sg = group_head;
6188 for_each_cpu_mask(j, sg->cpumask) {
6189 struct sched_domain *sd;
6191 sd = &per_cpu(phys_domains, j);
6192 if (j != first_cpu(sd->groups->cpumask)) {
6194 * Only add "power" once for each
6200 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6203 } while (sg != group_head);
6208 /* Free memory allocated for various sched_group structures */
6209 static void free_sched_groups(const cpumask_t *cpu_map)
6213 for_each_cpu_mask(cpu, *cpu_map) {
6214 struct sched_group **sched_group_nodes
6215 = sched_group_nodes_bycpu[cpu];
6217 if (!sched_group_nodes)
6220 for (i = 0; i < MAX_NUMNODES; i++) {
6221 cpumask_t nodemask = node_to_cpumask(i);
6222 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6224 cpus_and(nodemask, nodemask, *cpu_map);
6225 if (cpus_empty(nodemask))
6235 if (oldsg != sched_group_nodes[i])
6238 kfree(sched_group_nodes);
6239 sched_group_nodes_bycpu[cpu] = NULL;
6243 static void free_sched_groups(const cpumask_t *cpu_map)
6249 * Initialize sched groups cpu_power.
6251 * cpu_power indicates the capacity of sched group, which is used while
6252 * distributing the load between different sched groups in a sched domain.
6253 * Typically cpu_power for all the groups in a sched domain will be same unless
6254 * there are asymmetries in the topology. If there are asymmetries, group
6255 * having more cpu_power will pickup more load compared to the group having
6258 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6259 * the maximum number of tasks a group can handle in the presence of other idle
6260 * or lightly loaded groups in the same sched domain.
6262 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6264 struct sched_domain *child;
6265 struct sched_group *group;
6267 WARN_ON(!sd || !sd->groups);
6269 if (cpu != first_cpu(sd->groups->cpumask))
6274 sd->groups->__cpu_power = 0;
6277 * For perf policy, if the groups in child domain share resources
6278 * (for example cores sharing some portions of the cache hierarchy
6279 * or SMT), then set this domain groups cpu_power such that each group
6280 * can handle only one task, when there are other idle groups in the
6281 * same sched domain.
6283 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6285 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6286 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6291 * add cpu_power of each child group to this groups cpu_power
6293 group = child->groups;
6295 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6296 group = group->next;
6297 } while (group != child->groups);
6301 * Build sched domains for a given set of cpus and attach the sched domains
6302 * to the individual cpus
6304 static int build_sched_domains(const cpumask_t *cpu_map)
6307 struct root_domain *rd;
6309 struct sched_group **sched_group_nodes = NULL;
6310 int sd_allnodes = 0;
6313 * Allocate the per-node list of sched groups
6315 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6317 if (!sched_group_nodes) {
6318 printk(KERN_WARNING "Can not alloc sched group node list\n");
6321 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6324 rd = alloc_rootdomain(cpu_map);
6326 printk(KERN_WARNING "Cannot alloc root domain\n");
6331 * Set up domains for cpus specified by the cpu_map.
6333 for_each_cpu_mask(i, *cpu_map) {
6334 struct sched_domain *sd = NULL, *p;
6335 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6337 cpus_and(nodemask, nodemask, *cpu_map);
6340 if (cpus_weight(*cpu_map) >
6341 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6342 sd = &per_cpu(allnodes_domains, i);
6343 *sd = SD_ALLNODES_INIT;
6344 sd->span = *cpu_map;
6345 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6351 sd = &per_cpu(node_domains, i);
6353 sd->span = sched_domain_node_span(cpu_to_node(i));
6357 cpus_and(sd->span, sd->span, *cpu_map);
6361 sd = &per_cpu(phys_domains, i);
6363 sd->span = nodemask;
6367 cpu_to_phys_group(i, cpu_map, &sd->groups);
6369 #ifdef CONFIG_SCHED_MC
6371 sd = &per_cpu(core_domains, i);
6373 sd->span = cpu_coregroup_map(i);
6374 cpus_and(sd->span, sd->span, *cpu_map);
6377 cpu_to_core_group(i, cpu_map, &sd->groups);
6380 #ifdef CONFIG_SCHED_SMT
6382 sd = &per_cpu(cpu_domains, i);
6383 *sd = SD_SIBLING_INIT;
6384 sd->span = per_cpu(cpu_sibling_map, i);
6385 cpus_and(sd->span, sd->span, *cpu_map);
6388 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6392 #ifdef CONFIG_SCHED_SMT
6393 /* Set up CPU (sibling) groups */
6394 for_each_cpu_mask(i, *cpu_map) {
6395 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6396 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6397 if (i != first_cpu(this_sibling_map))
6400 init_sched_build_groups(this_sibling_map, cpu_map,
6405 #ifdef CONFIG_SCHED_MC
6406 /* Set up multi-core groups */
6407 for_each_cpu_mask(i, *cpu_map) {
6408 cpumask_t this_core_map = cpu_coregroup_map(i);
6409 cpus_and(this_core_map, this_core_map, *cpu_map);
6410 if (i != first_cpu(this_core_map))
6412 init_sched_build_groups(this_core_map, cpu_map,
6413 &cpu_to_core_group);
6417 /* Set up physical groups */
6418 for (i = 0; i < MAX_NUMNODES; i++) {
6419 cpumask_t nodemask = node_to_cpumask(i);
6421 cpus_and(nodemask, nodemask, *cpu_map);
6422 if (cpus_empty(nodemask))
6425 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6429 /* Set up node groups */
6431 init_sched_build_groups(*cpu_map, cpu_map,
6432 &cpu_to_allnodes_group);
6434 for (i = 0; i < MAX_NUMNODES; i++) {
6435 /* Set up node groups */
6436 struct sched_group *sg, *prev;
6437 cpumask_t nodemask = node_to_cpumask(i);
6438 cpumask_t domainspan;
6439 cpumask_t covered = CPU_MASK_NONE;
6442 cpus_and(nodemask, nodemask, *cpu_map);
6443 if (cpus_empty(nodemask)) {
6444 sched_group_nodes[i] = NULL;
6448 domainspan = sched_domain_node_span(i);
6449 cpus_and(domainspan, domainspan, *cpu_map);
6451 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6453 printk(KERN_WARNING "Can not alloc domain group for "
6457 sched_group_nodes[i] = sg;
6458 for_each_cpu_mask(j, nodemask) {
6459 struct sched_domain *sd;
6461 sd = &per_cpu(node_domains, j);
6464 sg->__cpu_power = 0;
6465 sg->cpumask = nodemask;
6467 cpus_or(covered, covered, nodemask);
6470 for (j = 0; j < MAX_NUMNODES; j++) {
6471 cpumask_t tmp, notcovered;
6472 int n = (i + j) % MAX_NUMNODES;
6474 cpus_complement(notcovered, covered);
6475 cpus_and(tmp, notcovered, *cpu_map);
6476 cpus_and(tmp, tmp, domainspan);
6477 if (cpus_empty(tmp))
6480 nodemask = node_to_cpumask(n);
6481 cpus_and(tmp, tmp, nodemask);
6482 if (cpus_empty(tmp))
6485 sg = kmalloc_node(sizeof(struct sched_group),
6489 "Can not alloc domain group for node %d\n", j);
6492 sg->__cpu_power = 0;
6494 sg->next = prev->next;
6495 cpus_or(covered, covered, tmp);
6502 /* Calculate CPU power for physical packages and nodes */
6503 #ifdef CONFIG_SCHED_SMT
6504 for_each_cpu_mask(i, *cpu_map) {
6505 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6507 init_sched_groups_power(i, sd);
6510 #ifdef CONFIG_SCHED_MC
6511 for_each_cpu_mask(i, *cpu_map) {
6512 struct sched_domain *sd = &per_cpu(core_domains, i);
6514 init_sched_groups_power(i, sd);
6518 for_each_cpu_mask(i, *cpu_map) {
6519 struct sched_domain *sd = &per_cpu(phys_domains, i);
6521 init_sched_groups_power(i, sd);
6525 for (i = 0; i < MAX_NUMNODES; i++)
6526 init_numa_sched_groups_power(sched_group_nodes[i]);
6529 struct sched_group *sg;
6531 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6532 init_numa_sched_groups_power(sg);
6536 /* Attach the domains */
6537 for_each_cpu_mask(i, *cpu_map) {
6538 struct sched_domain *sd;
6539 #ifdef CONFIG_SCHED_SMT
6540 sd = &per_cpu(cpu_domains, i);
6541 #elif defined(CONFIG_SCHED_MC)
6542 sd = &per_cpu(core_domains, i);
6544 sd = &per_cpu(phys_domains, i);
6546 cpu_attach_domain(sd, rd, i);
6553 free_sched_groups(cpu_map);
6558 static cpumask_t *doms_cur; /* current sched domains */
6559 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6562 * Special case: If a kmalloc of a doms_cur partition (array of
6563 * cpumask_t) fails, then fallback to a single sched domain,
6564 * as determined by the single cpumask_t fallback_doms.
6566 static cpumask_t fallback_doms;
6569 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6570 * For now this just excludes isolated cpus, but could be used to
6571 * exclude other special cases in the future.
6573 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6578 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6580 doms_cur = &fallback_doms;
6581 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6582 err = build_sched_domains(doms_cur);
6583 register_sched_domain_sysctl();
6588 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6590 free_sched_groups(cpu_map);
6594 * Detach sched domains from a group of cpus specified in cpu_map
6595 * These cpus will now be attached to the NULL domain
6597 static void detach_destroy_domains(const cpumask_t *cpu_map)
6601 unregister_sched_domain_sysctl();
6603 for_each_cpu_mask(i, *cpu_map)
6604 cpu_attach_domain(NULL, &def_root_domain, i);
6605 synchronize_sched();
6606 arch_destroy_sched_domains(cpu_map);
6610 * Partition sched domains as specified by the 'ndoms_new'
6611 * cpumasks in the array doms_new[] of cpumasks. This compares
6612 * doms_new[] to the current sched domain partitioning, doms_cur[].
6613 * It destroys each deleted domain and builds each new domain.
6615 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6616 * The masks don't intersect (don't overlap.) We should setup one
6617 * sched domain for each mask. CPUs not in any of the cpumasks will
6618 * not be load balanced. If the same cpumask appears both in the
6619 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6622 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6623 * ownership of it and will kfree it when done with it. If the caller
6624 * failed the kmalloc call, then it can pass in doms_new == NULL,
6625 * and partition_sched_domains() will fallback to the single partition
6628 * Call with hotplug lock held
6630 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6636 /* always unregister in case we don't destroy any domains */
6637 unregister_sched_domain_sysctl();
6639 if (doms_new == NULL) {
6641 doms_new = &fallback_doms;
6642 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6645 /* Destroy deleted domains */
6646 for (i = 0; i < ndoms_cur; i++) {
6647 for (j = 0; j < ndoms_new; j++) {
6648 if (cpus_equal(doms_cur[i], doms_new[j]))
6651 /* no match - a current sched domain not in new doms_new[] */
6652 detach_destroy_domains(doms_cur + i);
6657 /* Build new domains */
6658 for (i = 0; i < ndoms_new; i++) {
6659 for (j = 0; j < ndoms_cur; j++) {
6660 if (cpus_equal(doms_new[i], doms_cur[j]))
6663 /* no match - add a new doms_new */
6664 build_sched_domains(doms_new + i);
6669 /* Remember the new sched domains */
6670 if (doms_cur != &fallback_doms)
6672 doms_cur = doms_new;
6673 ndoms_cur = ndoms_new;
6675 register_sched_domain_sysctl();
6680 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6681 static int arch_reinit_sched_domains(void)
6686 detach_destroy_domains(&cpu_online_map);
6687 err = arch_init_sched_domains(&cpu_online_map);
6693 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6697 if (buf[0] != '0' && buf[0] != '1')
6701 sched_smt_power_savings = (buf[0] == '1');
6703 sched_mc_power_savings = (buf[0] == '1');
6705 ret = arch_reinit_sched_domains();
6707 return ret ? ret : count;
6710 #ifdef CONFIG_SCHED_MC
6711 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6713 return sprintf(page, "%u\n", sched_mc_power_savings);
6715 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6716 const char *buf, size_t count)
6718 return sched_power_savings_store(buf, count, 0);
6720 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6721 sched_mc_power_savings_store);
6724 #ifdef CONFIG_SCHED_SMT
6725 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6727 return sprintf(page, "%u\n", sched_smt_power_savings);
6729 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6730 const char *buf, size_t count)
6732 return sched_power_savings_store(buf, count, 1);
6734 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6735 sched_smt_power_savings_store);
6738 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6742 #ifdef CONFIG_SCHED_SMT
6744 err = sysfs_create_file(&cls->kset.kobj,
6745 &attr_sched_smt_power_savings.attr);
6747 #ifdef CONFIG_SCHED_MC
6748 if (!err && mc_capable())
6749 err = sysfs_create_file(&cls->kset.kobj,
6750 &attr_sched_mc_power_savings.attr);
6757 * Force a reinitialization of the sched domains hierarchy. The domains
6758 * and groups cannot be updated in place without racing with the balancing
6759 * code, so we temporarily attach all running cpus to the NULL domain
6760 * which will prevent rebalancing while the sched domains are recalculated.
6762 static int update_sched_domains(struct notifier_block *nfb,
6763 unsigned long action, void *hcpu)
6766 case CPU_UP_PREPARE:
6767 case CPU_UP_PREPARE_FROZEN:
6768 case CPU_DOWN_PREPARE:
6769 case CPU_DOWN_PREPARE_FROZEN:
6770 detach_destroy_domains(&cpu_online_map);
6773 case CPU_UP_CANCELED:
6774 case CPU_UP_CANCELED_FROZEN:
6775 case CPU_DOWN_FAILED:
6776 case CPU_DOWN_FAILED_FROZEN:
6778 case CPU_ONLINE_FROZEN:
6780 case CPU_DEAD_FROZEN:
6782 * Fall through and re-initialise the domains.
6789 /* The hotplug lock is already held by cpu_up/cpu_down */
6790 arch_init_sched_domains(&cpu_online_map);
6795 void __init sched_init_smp(void)
6797 cpumask_t non_isolated_cpus;
6800 arch_init_sched_domains(&cpu_online_map);
6801 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6802 if (cpus_empty(non_isolated_cpus))
6803 cpu_set(smp_processor_id(), non_isolated_cpus);
6805 /* XXX: Theoretical race here - CPU may be hotplugged now */
6806 hotcpu_notifier(update_sched_domains, 0);
6808 /* Move init over to a non-isolated CPU */
6809 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6811 sched_init_granularity();
6813 #ifdef CONFIG_FAIR_GROUP_SCHED
6814 if (nr_cpu_ids == 1)
6817 lb_monitor_task = kthread_create(load_balance_monitor, NULL,
6819 if (!IS_ERR(lb_monitor_task)) {
6820 lb_monitor_task->flags |= PF_NOFREEZE;
6821 wake_up_process(lb_monitor_task);
6823 printk(KERN_ERR "Could not create load balance monitor thread"
6824 "(error = %ld) \n", PTR_ERR(lb_monitor_task));
6829 void __init sched_init_smp(void)
6831 sched_init_granularity();
6833 #endif /* CONFIG_SMP */
6835 int in_sched_functions(unsigned long addr)
6837 return in_lock_functions(addr) ||
6838 (addr >= (unsigned long)__sched_text_start
6839 && addr < (unsigned long)__sched_text_end);
6842 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6844 cfs_rq->tasks_timeline = RB_ROOT;
6845 #ifdef CONFIG_FAIR_GROUP_SCHED
6848 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6851 void __init sched_init(void)
6853 int highest_cpu = 0;
6857 init_defrootdomain();
6860 for_each_possible_cpu(i) {
6861 struct rt_prio_array *array;
6865 spin_lock_init(&rq->lock);
6866 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6869 init_cfs_rq(&rq->cfs, rq);
6870 #ifdef CONFIG_FAIR_GROUP_SCHED
6871 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6873 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6874 struct sched_entity *se =
6875 &per_cpu(init_sched_entity, i);
6877 init_cfs_rq_p[i] = cfs_rq;
6878 init_cfs_rq(cfs_rq, rq);
6879 cfs_rq->tg = &init_task_group;
6880 list_add(&cfs_rq->leaf_cfs_rq_list,
6881 &rq->leaf_cfs_rq_list);
6883 init_sched_entity_p[i] = se;
6884 se->cfs_rq = &rq->cfs;
6886 se->load.weight = init_task_group_load;
6887 se->load.inv_weight =
6888 div64_64(1ULL<<32, init_task_group_load);
6891 init_task_group.shares = init_task_group_load;
6894 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6895 rq->cpu_load[j] = 0;
6899 rq_attach_root(rq, &def_root_domain);
6900 rq->active_balance = 0;
6901 rq->next_balance = jiffies;
6904 rq->migration_thread = NULL;
6905 INIT_LIST_HEAD(&rq->migration_queue);
6906 rq->rt.highest_prio = MAX_RT_PRIO;
6907 rq->rt.overloaded = 0;
6909 atomic_set(&rq->nr_iowait, 0);
6911 array = &rq->rt.active;
6912 for (j = 0; j < MAX_RT_PRIO; j++) {
6913 INIT_LIST_HEAD(array->queue + j);
6914 __clear_bit(j, array->bitmap);
6917 /* delimiter for bitsearch: */
6918 __set_bit(MAX_RT_PRIO, array->bitmap);
6921 set_load_weight(&init_task);
6923 #ifdef CONFIG_PREEMPT_NOTIFIERS
6924 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6928 nr_cpu_ids = highest_cpu + 1;
6929 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6932 #ifdef CONFIG_RT_MUTEXES
6933 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6937 * The boot idle thread does lazy MMU switching as well:
6939 atomic_inc(&init_mm.mm_count);
6940 enter_lazy_tlb(&init_mm, current);
6943 * Make us the idle thread. Technically, schedule() should not be
6944 * called from this thread, however somewhere below it might be,
6945 * but because we are the idle thread, we just pick up running again
6946 * when this runqueue becomes "idle".
6948 init_idle(current, smp_processor_id());
6950 * During early bootup we pretend to be a normal task:
6952 current->sched_class = &fair_sched_class;
6955 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6956 void __might_sleep(char *file, int line)
6959 static unsigned long prev_jiffy; /* ratelimiting */
6961 if ((in_atomic() || irqs_disabled()) &&
6962 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6963 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6965 prev_jiffy = jiffies;
6966 printk(KERN_ERR "BUG: sleeping function called from invalid"
6967 " context at %s:%d\n", file, line);
6968 printk("in_atomic():%d, irqs_disabled():%d\n",
6969 in_atomic(), irqs_disabled());
6970 debug_show_held_locks(current);
6971 if (irqs_disabled())
6972 print_irqtrace_events(current);
6977 EXPORT_SYMBOL(__might_sleep);
6980 #ifdef CONFIG_MAGIC_SYSRQ
6981 static void normalize_task(struct rq *rq, struct task_struct *p)
6984 update_rq_clock(rq);
6985 on_rq = p->se.on_rq;
6987 deactivate_task(rq, p, 0);
6988 __setscheduler(rq, p, SCHED_NORMAL, 0);
6990 activate_task(rq, p, 0);
6991 resched_task(rq->curr);
6995 void normalize_rt_tasks(void)
6997 struct task_struct *g, *p;
6998 unsigned long flags;
7001 read_lock_irq(&tasklist_lock);
7002 do_each_thread(g, p) {
7004 * Only normalize user tasks:
7009 p->se.exec_start = 0;
7010 #ifdef CONFIG_SCHEDSTATS
7011 p->se.wait_start = 0;
7012 p->se.sleep_start = 0;
7013 p->se.block_start = 0;
7015 task_rq(p)->clock = 0;
7019 * Renice negative nice level userspace
7022 if (TASK_NICE(p) < 0 && p->mm)
7023 set_user_nice(p, 0);
7027 spin_lock_irqsave(&p->pi_lock, flags);
7028 rq = __task_rq_lock(p);
7030 normalize_task(rq, p);
7032 __task_rq_unlock(rq);
7033 spin_unlock_irqrestore(&p->pi_lock, flags);
7034 } while_each_thread(g, p);
7036 read_unlock_irq(&tasklist_lock);
7039 #endif /* CONFIG_MAGIC_SYSRQ */
7043 * These functions are only useful for the IA64 MCA handling.
7045 * They can only be called when the whole system has been
7046 * stopped - every CPU needs to be quiescent, and no scheduling
7047 * activity can take place. Using them for anything else would
7048 * be a serious bug, and as a result, they aren't even visible
7049 * under any other configuration.
7053 * curr_task - return the current task for a given cpu.
7054 * @cpu: the processor in question.
7056 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7058 struct task_struct *curr_task(int cpu)
7060 return cpu_curr(cpu);
7064 * set_curr_task - set the current task for a given cpu.
7065 * @cpu: the processor in question.
7066 * @p: the task pointer to set.
7068 * Description: This function must only be used when non-maskable interrupts
7069 * are serviced on a separate stack. It allows the architecture to switch the
7070 * notion of the current task on a cpu in a non-blocking manner. This function
7071 * must be called with all CPU's synchronized, and interrupts disabled, the
7072 * and caller must save the original value of the current task (see
7073 * curr_task() above) and restore that value before reenabling interrupts and
7074 * re-starting the system.
7076 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7078 void set_curr_task(int cpu, struct task_struct *p)
7085 #ifdef CONFIG_FAIR_GROUP_SCHED
7089 * distribute shares of all task groups among their schedulable entities,
7090 * to reflect load distrbution across cpus.
7092 static int rebalance_shares(struct sched_domain *sd, int this_cpu)
7094 struct cfs_rq *cfs_rq;
7095 struct rq *rq = cpu_rq(this_cpu);
7096 cpumask_t sdspan = sd->span;
7099 /* Walk thr' all the task groups that we have */
7100 for_each_leaf_cfs_rq(rq, cfs_rq) {
7102 unsigned long total_load = 0, total_shares;
7103 struct task_group *tg = cfs_rq->tg;
7105 /* Gather total task load of this group across cpus */
7106 for_each_cpu_mask(i, sdspan)
7107 total_load += tg->cfs_rq[i]->load.weight;
7109 /* Nothing to do if this group has no load */
7114 * tg->shares represents the number of cpu shares the task group
7115 * is eligible to hold on a single cpu. On N cpus, it is
7116 * eligible to hold (N * tg->shares) number of cpu shares.
7118 total_shares = tg->shares * cpus_weight(sdspan);
7121 * redistribute total_shares across cpus as per the task load
7124 for_each_cpu_mask(i, sdspan) {
7125 unsigned long local_load, local_shares;
7127 local_load = tg->cfs_rq[i]->load.weight;
7128 local_shares = (local_load * total_shares) / total_load;
7130 local_shares = MIN_GROUP_SHARES;
7131 if (local_shares == tg->se[i]->load.weight)
7134 spin_lock_irq(&cpu_rq(i)->lock);
7135 set_se_shares(tg->se[i], local_shares);
7136 spin_unlock_irq(&cpu_rq(i)->lock);
7145 * How frequently should we rebalance_shares() across cpus?
7147 * The more frequently we rebalance shares, the more accurate is the fairness
7148 * of cpu bandwidth distribution between task groups. However higher frequency
7149 * also implies increased scheduling overhead.
7151 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7152 * consecutive calls to rebalance_shares() in the same sched domain.
7154 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7155 * consecutive calls to rebalance_shares() in the same sched domain.
7157 * These settings allows for the appropriate tradeoff between accuracy of
7158 * fairness and the associated overhead.
7162 /* default: 8ms, units: milliseconds */
7163 const_debug unsigned int sysctl_sched_min_bal_int_shares = 8;
7165 /* default: 128ms, units: milliseconds */
7166 const_debug unsigned int sysctl_sched_max_bal_int_shares = 128;
7168 /* kernel thread that runs rebalance_shares() periodically */
7169 static int load_balance_monitor(void *unused)
7171 unsigned int timeout = sysctl_sched_min_bal_int_shares;
7172 struct sched_param schedparm;
7176 * We don't want this thread's execution to be limited by the shares
7177 * assigned to default group (init_task_group). Hence make it run
7178 * as a SCHED_RR RT task at the lowest priority.
7180 schedparm.sched_priority = 1;
7181 ret = sched_setscheduler(current, SCHED_RR, &schedparm);
7183 printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance"
7184 " monitor thread (error = %d) \n", ret);
7186 while (!kthread_should_stop()) {
7187 int i, cpu, balanced = 1;
7189 /* Prevent cpus going down or coming up */
7191 /* lockout changes to doms_cur[] array */
7194 * Enter a rcu read-side critical section to safely walk rq->sd
7195 * chain on various cpus and to walk task group list
7196 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7200 for (i = 0; i < ndoms_cur; i++) {
7201 cpumask_t cpumap = doms_cur[i];
7202 struct sched_domain *sd = NULL, *sd_prev = NULL;
7204 cpu = first_cpu(cpumap);
7206 /* Find the highest domain at which to balance shares */
7207 for_each_domain(cpu, sd) {
7208 if (!(sd->flags & SD_LOAD_BALANCE))
7214 /* sd == NULL? No load balance reqd in this domain */
7218 balanced &= rebalance_shares(sd, cpu);
7227 timeout = sysctl_sched_min_bal_int_shares;
7228 else if (timeout < sysctl_sched_max_bal_int_shares)
7231 msleep_interruptible(timeout);
7236 #endif /* CONFIG_SMP */
7238 /* allocate runqueue etc for a new task group */
7239 struct task_group *sched_create_group(void)
7241 struct task_group *tg;
7242 struct cfs_rq *cfs_rq;
7243 struct sched_entity *se;
7247 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7249 return ERR_PTR(-ENOMEM);
7251 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7254 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7258 for_each_possible_cpu(i) {
7261 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7266 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7271 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7272 memset(se, 0, sizeof(struct sched_entity));
7274 tg->cfs_rq[i] = cfs_rq;
7275 init_cfs_rq(cfs_rq, rq);
7279 se->cfs_rq = &rq->cfs;
7281 se->load.weight = NICE_0_LOAD;
7282 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7286 tg->shares = NICE_0_LOAD;
7288 lock_task_group_list();
7289 for_each_possible_cpu(i) {
7291 cfs_rq = tg->cfs_rq[i];
7292 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7294 unlock_task_group_list();
7299 for_each_possible_cpu(i) {
7301 kfree(tg->cfs_rq[i]);
7309 return ERR_PTR(-ENOMEM);
7312 /* rcu callback to free various structures associated with a task group */
7313 static void free_sched_group(struct rcu_head *rhp)
7315 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7316 struct cfs_rq *cfs_rq;
7317 struct sched_entity *se;
7320 /* now it should be safe to free those cfs_rqs */
7321 for_each_possible_cpu(i) {
7322 cfs_rq = tg->cfs_rq[i];
7334 /* Destroy runqueue etc associated with a task group */
7335 void sched_destroy_group(struct task_group *tg)
7337 struct cfs_rq *cfs_rq = NULL;
7340 lock_task_group_list();
7341 for_each_possible_cpu(i) {
7342 cfs_rq = tg->cfs_rq[i];
7343 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7345 unlock_task_group_list();
7349 /* wait for possible concurrent references to cfs_rqs complete */
7350 call_rcu(&tg->rcu, free_sched_group);
7353 /* change task's runqueue when it moves between groups.
7354 * The caller of this function should have put the task in its new group
7355 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7356 * reflect its new group.
7358 void sched_move_task(struct task_struct *tsk)
7361 unsigned long flags;
7364 rq = task_rq_lock(tsk, &flags);
7366 if (tsk->sched_class != &fair_sched_class) {
7367 set_task_cfs_rq(tsk, task_cpu(tsk));
7371 update_rq_clock(rq);
7373 running = task_current(rq, tsk);
7374 on_rq = tsk->se.on_rq;
7377 dequeue_task(rq, tsk, 0);
7378 if (unlikely(running))
7379 tsk->sched_class->put_prev_task(rq, tsk);
7382 set_task_cfs_rq(tsk, task_cpu(tsk));
7385 if (unlikely(running))
7386 tsk->sched_class->set_curr_task(rq);
7387 enqueue_task(rq, tsk, 0);
7391 task_rq_unlock(rq, &flags);
7394 /* rq->lock to be locked by caller */
7395 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7397 struct cfs_rq *cfs_rq = se->cfs_rq;
7398 struct rq *rq = cfs_rq->rq;
7402 shares = MIN_GROUP_SHARES;
7406 dequeue_entity(cfs_rq, se, 0);
7407 dec_cpu_load(rq, se->load.weight);
7410 se->load.weight = shares;
7411 se->load.inv_weight = div64_64((1ULL<<32), shares);
7414 enqueue_entity(cfs_rq, se, 0);
7415 inc_cpu_load(rq, se->load.weight);
7419 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7422 struct cfs_rq *cfs_rq;
7425 lock_task_group_list();
7426 if (tg->shares == shares)
7429 if (shares < MIN_GROUP_SHARES)
7430 shares = MIN_GROUP_SHARES;
7433 * Prevent any load balance activity (rebalance_shares,
7434 * load_balance_fair) from referring to this group first,
7435 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7437 for_each_possible_cpu(i) {
7438 cfs_rq = tg->cfs_rq[i];
7439 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7442 /* wait for any ongoing reference to this group to finish */
7443 synchronize_sched();
7446 * Now we are free to modify the group's share on each cpu
7447 * w/o tripping rebalance_share or load_balance_fair.
7449 tg->shares = shares;
7450 for_each_possible_cpu(i) {
7451 spin_lock_irq(&cpu_rq(i)->lock);
7452 set_se_shares(tg->se[i], shares);
7453 spin_unlock_irq(&cpu_rq(i)->lock);
7457 * Enable load balance activity on this group, by inserting it back on
7458 * each cpu's rq->leaf_cfs_rq_list.
7460 for_each_possible_cpu(i) {
7462 cfs_rq = tg->cfs_rq[i];
7463 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7466 unlock_task_group_list();
7470 unsigned long sched_group_shares(struct task_group *tg)
7475 #endif /* CONFIG_FAIR_GROUP_SCHED */
7477 #ifdef CONFIG_FAIR_CGROUP_SCHED
7479 /* return corresponding task_group object of a cgroup */
7480 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7482 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7483 struct task_group, css);
7486 static struct cgroup_subsys_state *
7487 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7489 struct task_group *tg;
7491 if (!cgrp->parent) {
7492 /* This is early initialization for the top cgroup */
7493 init_task_group.css.cgroup = cgrp;
7494 return &init_task_group.css;
7497 /* we support only 1-level deep hierarchical scheduler atm */
7498 if (cgrp->parent->parent)
7499 return ERR_PTR(-EINVAL);
7501 tg = sched_create_group();
7503 return ERR_PTR(-ENOMEM);
7505 /* Bind the cgroup to task_group object we just created */
7506 tg->css.cgroup = cgrp;
7512 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7514 struct task_group *tg = cgroup_tg(cgrp);
7516 sched_destroy_group(tg);
7520 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7521 struct task_struct *tsk)
7523 /* We don't support RT-tasks being in separate groups */
7524 if (tsk->sched_class != &fair_sched_class)
7531 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7532 struct cgroup *old_cont, struct task_struct *tsk)
7534 sched_move_task(tsk);
7537 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7540 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7543 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7545 struct task_group *tg = cgroup_tg(cgrp);
7547 return (u64) tg->shares;
7550 static struct cftype cpu_files[] = {
7553 .read_uint = cpu_shares_read_uint,
7554 .write_uint = cpu_shares_write_uint,
7558 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7560 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7563 struct cgroup_subsys cpu_cgroup_subsys = {
7565 .create = cpu_cgroup_create,
7566 .destroy = cpu_cgroup_destroy,
7567 .can_attach = cpu_cgroup_can_attach,
7568 .attach = cpu_cgroup_attach,
7569 .populate = cpu_cgroup_populate,
7570 .subsys_id = cpu_cgroup_subsys_id,
7574 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7576 #ifdef CONFIG_CGROUP_CPUACCT
7579 * CPU accounting code for task groups.
7581 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7582 * (balbir@in.ibm.com).
7585 /* track cpu usage of a group of tasks */
7587 struct cgroup_subsys_state css;
7588 /* cpuusage holds pointer to a u64-type object on every cpu */
7592 struct cgroup_subsys cpuacct_subsys;
7594 /* return cpu accounting group corresponding to this container */
7595 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7597 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7598 struct cpuacct, css);
7601 /* return cpu accounting group to which this task belongs */
7602 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7604 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7605 struct cpuacct, css);
7608 /* create a new cpu accounting group */
7609 static struct cgroup_subsys_state *cpuacct_create(
7610 struct cgroup_subsys *ss, struct cgroup *cont)
7612 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7615 return ERR_PTR(-ENOMEM);
7617 ca->cpuusage = alloc_percpu(u64);
7618 if (!ca->cpuusage) {
7620 return ERR_PTR(-ENOMEM);
7626 /* destroy an existing cpu accounting group */
7628 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
7630 struct cpuacct *ca = cgroup_ca(cont);
7632 free_percpu(ca->cpuusage);
7636 /* return total cpu usage (in nanoseconds) of a group */
7637 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7639 struct cpuacct *ca = cgroup_ca(cont);
7640 u64 totalcpuusage = 0;
7643 for_each_possible_cpu(i) {
7644 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7647 * Take rq->lock to make 64-bit addition safe on 32-bit
7650 spin_lock_irq(&cpu_rq(i)->lock);
7651 totalcpuusage += *cpuusage;
7652 spin_unlock_irq(&cpu_rq(i)->lock);
7655 return totalcpuusage;
7658 static struct cftype files[] = {
7661 .read_uint = cpuusage_read,
7665 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7667 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7671 * charge this task's execution time to its accounting group.
7673 * called with rq->lock held.
7675 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7679 if (!cpuacct_subsys.active)
7684 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7686 *cpuusage += cputime;
7690 struct cgroup_subsys cpuacct_subsys = {
7692 .create = cpuacct_create,
7693 .destroy = cpuacct_destroy,
7694 .populate = cpuacct_populate,
7695 .subsys_id = cpuacct_subsys_id,
7697 #endif /* CONFIG_CGROUP_CPUACCT */