4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/kthread.h>
56 #include <linux/seq_file.h>
57 #include <linux/sysctl.h>
58 #include <linux/syscalls.h>
59 #include <linux/times.h>
60 #include <linux/tsacct_kern.h>
61 #include <linux/kprobes.h>
62 #include <linux/delayacct.h>
63 #include <linux/reciprocal_div.h>
64 #include <linux/unistd.h>
65 #include <linux/pagemap.h>
68 #include <asm/irq_regs.h>
71 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
75 unsigned long long __attribute__((weak)) sched_clock(void)
77 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Some helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 static inline int rt_policy(int policy)
138 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
143 static inline int task_has_rt_policy(struct task_struct *p)
145 return rt_policy(p->policy);
149 * This is the priority-queue data structure of the RT scheduling class:
151 struct rt_prio_array {
152 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
153 struct list_head queue[MAX_RT_PRIO];
156 #ifdef CONFIG_FAIR_GROUP_SCHED
158 #include <linux/cgroup.h>
162 /* task group related information */
164 #ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css;
167 /* schedulable entities of this group on each cpu */
168 struct sched_entity **se;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq **cfs_rq;
173 * shares assigned to a task group governs how much of cpu bandwidth
174 * is allocated to the group. The more shares a group has, the more is
175 * the cpu bandwidth allocated to it.
177 * For ex, lets say that there are three task groups, A, B and C which
178 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
179 * cpu bandwidth allocated by the scheduler to task groups A, B and C
182 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
183 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
184 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
186 * The weight assigned to a task group's schedulable entities on every
187 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
188 * group's shares. For ex: lets say that task group A has been
189 * assigned shares of 1000 and there are two CPUs in a system. Then,
191 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
193 * Note: It's not necessary that each of a task's group schedulable
194 * entity have the same weight on all CPUs. If the group
195 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
196 * better distribution of weight could be:
198 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
199 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
201 * rebalance_shares() is responsible for distributing the shares of a
202 * task groups like this among the group's schedulable entities across
206 unsigned long shares;
211 /* Default task group's sched entity on each cpu */
212 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
213 /* Default task group's cfs_rq on each cpu */
214 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
216 static struct sched_entity *init_sched_entity_p[NR_CPUS];
217 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
219 /* task_group_mutex serializes add/remove of task groups and also changes to
220 * a task group's cpu shares.
222 static DEFINE_MUTEX(task_group_mutex);
224 /* doms_cur_mutex serializes access to doms_cur[] array */
225 static DEFINE_MUTEX(doms_cur_mutex);
228 /* kernel thread that runs rebalance_shares() periodically */
229 static struct task_struct *lb_monitor_task;
230 static int load_balance_monitor(void *unused);
233 static void set_se_shares(struct sched_entity *se, unsigned long shares);
235 /* Default task group.
236 * Every task in system belong to this group at bootup.
238 struct task_group init_task_group = {
239 .se = init_sched_entity_p,
240 .cfs_rq = init_cfs_rq_p,
243 #ifdef CONFIG_FAIR_USER_SCHED
244 # define INIT_TASK_GROUP_LOAD 2*NICE_0_LOAD
246 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
249 #define MIN_GROUP_SHARES 2
251 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
253 /* return group to which a task belongs */
254 static inline struct task_group *task_group(struct task_struct *p)
256 struct task_group *tg;
258 #ifdef CONFIG_FAIR_USER_SCHED
260 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
261 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
262 struct task_group, css);
264 tg = &init_task_group;
269 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
270 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
272 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
273 p->se.parent = task_group(p)->se[cpu];
276 static inline void lock_task_group_list(void)
278 mutex_lock(&task_group_mutex);
281 static inline void unlock_task_group_list(void)
283 mutex_unlock(&task_group_mutex);
286 static inline void lock_doms_cur(void)
288 mutex_lock(&doms_cur_mutex);
291 static inline void unlock_doms_cur(void)
293 mutex_unlock(&doms_cur_mutex);
298 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
299 static inline void lock_task_group_list(void) { }
300 static inline void unlock_task_group_list(void) { }
301 static inline void lock_doms_cur(void) { }
302 static inline void unlock_doms_cur(void) { }
304 #endif /* CONFIG_FAIR_GROUP_SCHED */
306 /* CFS-related fields in a runqueue */
308 struct load_weight load;
309 unsigned long nr_running;
314 struct rb_root tasks_timeline;
315 struct rb_node *rb_leftmost;
316 struct rb_node *rb_load_balance_curr;
317 /* 'curr' points to currently running entity on this cfs_rq.
318 * It is set to NULL otherwise (i.e when none are currently running).
320 struct sched_entity *curr;
322 unsigned long nr_spread_over;
324 #ifdef CONFIG_FAIR_GROUP_SCHED
325 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
328 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
329 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
330 * (like users, containers etc.)
332 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
333 * list is used during load balance.
335 struct list_head leaf_cfs_rq_list;
336 struct task_group *tg; /* group that "owns" this runqueue */
340 /* Real-Time classes' related field in a runqueue: */
342 struct rt_prio_array active;
343 int rt_load_balance_idx;
344 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
345 unsigned long rt_nr_running;
346 /* highest queued rt task prio */
351 * This is the main, per-CPU runqueue data structure.
353 * Locking rule: those places that want to lock multiple runqueues
354 * (such as the load balancing or the thread migration code), lock
355 * acquire operations must be ordered by ascending &runqueue.
362 * nr_running and cpu_load should be in the same cacheline because
363 * remote CPUs use both these fields when doing load calculation.
365 unsigned long nr_running;
366 #define CPU_LOAD_IDX_MAX 5
367 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
368 unsigned char idle_at_tick;
370 unsigned char in_nohz_recently;
372 /* capture load from *all* tasks on this cpu: */
373 struct load_weight load;
374 unsigned long nr_load_updates;
378 #ifdef CONFIG_FAIR_GROUP_SCHED
379 /* list of leaf cfs_rq on this cpu: */
380 struct list_head leaf_cfs_rq_list;
385 * This is part of a global counter where only the total sum
386 * over all CPUs matters. A task can increase this counter on
387 * one CPU and if it got migrated afterwards it may decrease
388 * it on another CPU. Always updated under the runqueue lock:
390 unsigned long nr_uninterruptible;
392 struct task_struct *curr, *idle;
393 unsigned long next_balance;
394 struct mm_struct *prev_mm;
396 u64 clock, prev_clock_raw;
399 unsigned int clock_warps, clock_overflows;
401 unsigned int clock_deep_idle_events;
407 struct sched_domain *sd;
409 /* For active balancing */
412 /* cpu of this runqueue: */
415 struct task_struct *migration_thread;
416 struct list_head migration_queue;
419 #ifdef CONFIG_SCHEDSTATS
421 struct sched_info rq_sched_info;
423 /* sys_sched_yield() stats */
424 unsigned int yld_exp_empty;
425 unsigned int yld_act_empty;
426 unsigned int yld_both_empty;
427 unsigned int yld_count;
429 /* schedule() stats */
430 unsigned int sched_switch;
431 unsigned int sched_count;
432 unsigned int sched_goidle;
434 /* try_to_wake_up() stats */
435 unsigned int ttwu_count;
436 unsigned int ttwu_local;
439 unsigned int bkl_count;
441 struct lock_class_key rq_lock_key;
444 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
446 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
448 rq->curr->sched_class->check_preempt_curr(rq, p);
451 static inline int cpu_of(struct rq *rq)
461 * Update the per-runqueue clock, as finegrained as the platform can give
462 * us, but without assuming monotonicity, etc.:
464 static void __update_rq_clock(struct rq *rq)
466 u64 prev_raw = rq->prev_clock_raw;
467 u64 now = sched_clock();
468 s64 delta = now - prev_raw;
469 u64 clock = rq->clock;
471 #ifdef CONFIG_SCHED_DEBUG
472 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
475 * Protect against sched_clock() occasionally going backwards:
477 if (unlikely(delta < 0)) {
482 * Catch too large forward jumps too:
484 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
485 if (clock < rq->tick_timestamp + TICK_NSEC)
486 clock = rq->tick_timestamp + TICK_NSEC;
489 rq->clock_overflows++;
491 if (unlikely(delta > rq->clock_max_delta))
492 rq->clock_max_delta = delta;
497 rq->prev_clock_raw = now;
501 static void update_rq_clock(struct rq *rq)
503 if (likely(smp_processor_id() == cpu_of(rq)))
504 __update_rq_clock(rq);
508 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
509 * See detach_destroy_domains: synchronize_sched for details.
511 * The domain tree of any CPU may only be accessed from within
512 * preempt-disabled sections.
514 #define for_each_domain(cpu, __sd) \
515 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
517 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
518 #define this_rq() (&__get_cpu_var(runqueues))
519 #define task_rq(p) cpu_rq(task_cpu(p))
520 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
523 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
525 #ifdef CONFIG_SCHED_DEBUG
526 # define const_debug __read_mostly
528 # define const_debug static const
532 * Debugging: various feature bits
535 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
536 SCHED_FEAT_WAKEUP_PREEMPT = 2,
537 SCHED_FEAT_START_DEBIT = 4,
538 SCHED_FEAT_TREE_AVG = 8,
539 SCHED_FEAT_APPROX_AVG = 16,
542 const_debug unsigned int sysctl_sched_features =
543 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
544 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
545 SCHED_FEAT_START_DEBIT * 1 |
546 SCHED_FEAT_TREE_AVG * 0 |
547 SCHED_FEAT_APPROX_AVG * 0;
549 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
552 * Number of tasks to iterate in a single balance run.
553 * Limited because this is done with IRQs disabled.
555 const_debug unsigned int sysctl_sched_nr_migrate = 32;
558 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
559 * clock constructed from sched_clock():
561 unsigned long long cpu_clock(int cpu)
563 unsigned long long now;
567 local_irq_save(flags);
570 * Only call sched_clock() if the scheduler has already been
571 * initialized (some code might call cpu_clock() very early):
576 local_irq_restore(flags);
580 EXPORT_SYMBOL_GPL(cpu_clock);
582 #ifndef prepare_arch_switch
583 # define prepare_arch_switch(next) do { } while (0)
585 #ifndef finish_arch_switch
586 # define finish_arch_switch(prev) do { } while (0)
589 static inline int task_current(struct rq *rq, struct task_struct *p)
591 return rq->curr == p;
594 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
595 static inline int task_running(struct rq *rq, struct task_struct *p)
597 return task_current(rq, p);
600 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
604 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
606 #ifdef CONFIG_DEBUG_SPINLOCK
607 /* this is a valid case when another task releases the spinlock */
608 rq->lock.owner = current;
611 * If we are tracking spinlock dependencies then we have to
612 * fix up the runqueue lock - which gets 'carried over' from
615 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
617 spin_unlock_irq(&rq->lock);
620 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
621 static inline int task_running(struct rq *rq, struct task_struct *p)
626 return task_current(rq, p);
630 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
634 * We can optimise this out completely for !SMP, because the
635 * SMP rebalancing from interrupt is the only thing that cares
640 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
641 spin_unlock_irq(&rq->lock);
643 spin_unlock(&rq->lock);
647 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
651 * After ->oncpu is cleared, the task can be moved to a different CPU.
652 * We must ensure this doesn't happen until the switch is completely
658 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
662 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
665 * __task_rq_lock - lock the runqueue a given task resides on.
666 * Must be called interrupts disabled.
668 static inline struct rq *__task_rq_lock(struct task_struct *p)
672 struct rq *rq = task_rq(p);
673 spin_lock(&rq->lock);
674 if (likely(rq == task_rq(p)))
676 spin_unlock(&rq->lock);
681 * task_rq_lock - lock the runqueue a given task resides on and disable
682 * interrupts. Note the ordering: we can safely lookup the task_rq without
683 * explicitly disabling preemption.
685 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
691 local_irq_save(*flags);
693 spin_lock(&rq->lock);
694 if (likely(rq == task_rq(p)))
696 spin_unlock_irqrestore(&rq->lock, *flags);
700 static void __task_rq_unlock(struct rq *rq)
703 spin_unlock(&rq->lock);
706 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
709 spin_unlock_irqrestore(&rq->lock, *flags);
713 * this_rq_lock - lock this runqueue and disable interrupts.
715 static struct rq *this_rq_lock(void)
722 spin_lock(&rq->lock);
728 * We are going deep-idle (irqs are disabled):
730 void sched_clock_idle_sleep_event(void)
732 struct rq *rq = cpu_rq(smp_processor_id());
734 spin_lock(&rq->lock);
735 __update_rq_clock(rq);
736 spin_unlock(&rq->lock);
737 rq->clock_deep_idle_events++;
739 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
742 * We just idled delta nanoseconds (called with irqs disabled):
744 void sched_clock_idle_wakeup_event(u64 delta_ns)
746 struct rq *rq = cpu_rq(smp_processor_id());
747 u64 now = sched_clock();
749 touch_softlockup_watchdog();
750 rq->idle_clock += delta_ns;
752 * Override the previous timestamp and ignore all
753 * sched_clock() deltas that occured while we idled,
754 * and use the PM-provided delta_ns to advance the
757 spin_lock(&rq->lock);
758 rq->prev_clock_raw = now;
759 rq->clock += delta_ns;
760 spin_unlock(&rq->lock);
762 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
765 * resched_task - mark a task 'to be rescheduled now'.
767 * On UP this means the setting of the need_resched flag, on SMP it
768 * might also involve a cross-CPU call to trigger the scheduler on
773 #ifndef tsk_is_polling
774 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
777 static void resched_task(struct task_struct *p)
781 assert_spin_locked(&task_rq(p)->lock);
783 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
786 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
789 if (cpu == smp_processor_id())
792 /* NEED_RESCHED must be visible before we test polling */
794 if (!tsk_is_polling(p))
795 smp_send_reschedule(cpu);
798 static void resched_cpu(int cpu)
800 struct rq *rq = cpu_rq(cpu);
803 if (!spin_trylock_irqsave(&rq->lock, flags))
805 resched_task(cpu_curr(cpu));
806 spin_unlock_irqrestore(&rq->lock, flags);
809 static inline void resched_task(struct task_struct *p)
811 assert_spin_locked(&task_rq(p)->lock);
812 set_tsk_need_resched(p);
816 #if BITS_PER_LONG == 32
817 # define WMULT_CONST (~0UL)
819 # define WMULT_CONST (1UL << 32)
822 #define WMULT_SHIFT 32
825 * Shift right and round:
827 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
830 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
831 struct load_weight *lw)
835 if (unlikely(!lw->inv_weight))
836 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
838 tmp = (u64)delta_exec * weight;
840 * Check whether we'd overflow the 64-bit multiplication:
842 if (unlikely(tmp > WMULT_CONST))
843 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
846 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
848 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
851 static inline unsigned long
852 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
854 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
857 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
862 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
868 * To aid in avoiding the subversion of "niceness" due to uneven distribution
869 * of tasks with abnormal "nice" values across CPUs the contribution that
870 * each task makes to its run queue's load is weighted according to its
871 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
872 * scaled version of the new time slice allocation that they receive on time
876 #define WEIGHT_IDLEPRIO 2
877 #define WMULT_IDLEPRIO (1 << 31)
880 * Nice levels are multiplicative, with a gentle 10% change for every
881 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
882 * nice 1, it will get ~10% less CPU time than another CPU-bound task
883 * that remained on nice 0.
885 * The "10% effect" is relative and cumulative: from _any_ nice level,
886 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
887 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
888 * If a task goes up by ~10% and another task goes down by ~10% then
889 * the relative distance between them is ~25%.)
891 static const int prio_to_weight[40] = {
892 /* -20 */ 88761, 71755, 56483, 46273, 36291,
893 /* -15 */ 29154, 23254, 18705, 14949, 11916,
894 /* -10 */ 9548, 7620, 6100, 4904, 3906,
895 /* -5 */ 3121, 2501, 1991, 1586, 1277,
896 /* 0 */ 1024, 820, 655, 526, 423,
897 /* 5 */ 335, 272, 215, 172, 137,
898 /* 10 */ 110, 87, 70, 56, 45,
899 /* 15 */ 36, 29, 23, 18, 15,
903 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
905 * In cases where the weight does not change often, we can use the
906 * precalculated inverse to speed up arithmetics by turning divisions
907 * into multiplications:
909 static const u32 prio_to_wmult[40] = {
910 /* -20 */ 48388, 59856, 76040, 92818, 118348,
911 /* -15 */ 147320, 184698, 229616, 287308, 360437,
912 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
913 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
914 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
915 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
916 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
917 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
920 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
923 * runqueue iterator, to support SMP load-balancing between different
924 * scheduling classes, without having to expose their internal data
925 * structures to the load-balancing proper:
929 struct task_struct *(*start)(void *);
930 struct task_struct *(*next)(void *);
935 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
936 unsigned long max_load_move, struct sched_domain *sd,
937 enum cpu_idle_type idle, int *all_pinned,
938 int *this_best_prio, struct rq_iterator *iterator);
941 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
942 struct sched_domain *sd, enum cpu_idle_type idle,
943 struct rq_iterator *iterator);
946 #ifdef CONFIG_CGROUP_CPUACCT
947 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
949 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
952 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
954 update_load_add(&rq->load, load);
957 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
959 update_load_sub(&rq->load, load);
962 #include "sched_stats.h"
963 #include "sched_idletask.c"
964 #include "sched_fair.c"
965 #include "sched_rt.c"
966 #ifdef CONFIG_SCHED_DEBUG
967 # include "sched_debug.c"
970 #define sched_class_highest (&rt_sched_class)
972 static void inc_nr_running(struct task_struct *p, struct rq *rq)
977 static void dec_nr_running(struct task_struct *p, struct rq *rq)
982 static void set_load_weight(struct task_struct *p)
984 if (task_has_rt_policy(p)) {
985 p->se.load.weight = prio_to_weight[0] * 2;
986 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
991 * SCHED_IDLE tasks get minimal weight:
993 if (p->policy == SCHED_IDLE) {
994 p->se.load.weight = WEIGHT_IDLEPRIO;
995 p->se.load.inv_weight = WMULT_IDLEPRIO;
999 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1000 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1003 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1005 sched_info_queued(p);
1006 p->sched_class->enqueue_task(rq, p, wakeup);
1010 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1012 p->sched_class->dequeue_task(rq, p, sleep);
1017 * __normal_prio - return the priority that is based on the static prio
1019 static inline int __normal_prio(struct task_struct *p)
1021 return p->static_prio;
1025 * Calculate the expected normal priority: i.e. priority
1026 * without taking RT-inheritance into account. Might be
1027 * boosted by interactivity modifiers. Changes upon fork,
1028 * setprio syscalls, and whenever the interactivity
1029 * estimator recalculates.
1031 static inline int normal_prio(struct task_struct *p)
1035 if (task_has_rt_policy(p))
1036 prio = MAX_RT_PRIO-1 - p->rt_priority;
1038 prio = __normal_prio(p);
1043 * Calculate the current priority, i.e. the priority
1044 * taken into account by the scheduler. This value might
1045 * be boosted by RT tasks, or might be boosted by
1046 * interactivity modifiers. Will be RT if the task got
1047 * RT-boosted. If not then it returns p->normal_prio.
1049 static int effective_prio(struct task_struct *p)
1051 p->normal_prio = normal_prio(p);
1053 * If we are RT tasks or we were boosted to RT priority,
1054 * keep the priority unchanged. Otherwise, update priority
1055 * to the normal priority:
1057 if (!rt_prio(p->prio))
1058 return p->normal_prio;
1063 * activate_task - move a task to the runqueue.
1065 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1067 if (p->state == TASK_UNINTERRUPTIBLE)
1068 rq->nr_uninterruptible--;
1070 enqueue_task(rq, p, wakeup);
1071 inc_nr_running(p, rq);
1075 * deactivate_task - remove a task from the runqueue.
1077 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1079 if (p->state == TASK_UNINTERRUPTIBLE)
1080 rq->nr_uninterruptible++;
1082 dequeue_task(rq, p, sleep);
1083 dec_nr_running(p, rq);
1087 * task_curr - is this task currently executing on a CPU?
1088 * @p: the task in question.
1090 inline int task_curr(const struct task_struct *p)
1092 return cpu_curr(task_cpu(p)) == p;
1095 /* Used instead of source_load when we know the type == 0 */
1096 unsigned long weighted_cpuload(const int cpu)
1098 return cpu_rq(cpu)->load.weight;
1101 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1103 set_task_cfs_rq(p, cpu);
1106 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1107 * successfuly executed on another CPU. We must ensure that updates of
1108 * per-task data have been completed by this moment.
1111 task_thread_info(p)->cpu = cpu;
1118 * Is this task likely cache-hot:
1121 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1125 if (p->sched_class != &fair_sched_class)
1128 if (sysctl_sched_migration_cost == -1)
1130 if (sysctl_sched_migration_cost == 0)
1133 delta = now - p->se.exec_start;
1135 return delta < (s64)sysctl_sched_migration_cost;
1139 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1141 int old_cpu = task_cpu(p);
1142 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1143 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1144 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1147 clock_offset = old_rq->clock - new_rq->clock;
1149 #ifdef CONFIG_SCHEDSTATS
1150 if (p->se.wait_start)
1151 p->se.wait_start -= clock_offset;
1152 if (p->se.sleep_start)
1153 p->se.sleep_start -= clock_offset;
1154 if (p->se.block_start)
1155 p->se.block_start -= clock_offset;
1156 if (old_cpu != new_cpu) {
1157 schedstat_inc(p, se.nr_migrations);
1158 if (task_hot(p, old_rq->clock, NULL))
1159 schedstat_inc(p, se.nr_forced2_migrations);
1162 p->se.vruntime -= old_cfsrq->min_vruntime -
1163 new_cfsrq->min_vruntime;
1165 __set_task_cpu(p, new_cpu);
1168 struct migration_req {
1169 struct list_head list;
1171 struct task_struct *task;
1174 struct completion done;
1178 * The task's runqueue lock must be held.
1179 * Returns true if you have to wait for migration thread.
1182 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1184 struct rq *rq = task_rq(p);
1187 * If the task is not on a runqueue (and not running), then
1188 * it is sufficient to simply update the task's cpu field.
1190 if (!p->se.on_rq && !task_running(rq, p)) {
1191 set_task_cpu(p, dest_cpu);
1195 init_completion(&req->done);
1197 req->dest_cpu = dest_cpu;
1198 list_add(&req->list, &rq->migration_queue);
1204 * wait_task_inactive - wait for a thread to unschedule.
1206 * The caller must ensure that the task *will* unschedule sometime soon,
1207 * else this function might spin for a *long* time. This function can't
1208 * be called with interrupts off, or it may introduce deadlock with
1209 * smp_call_function() if an IPI is sent by the same process we are
1210 * waiting to become inactive.
1212 void wait_task_inactive(struct task_struct *p)
1214 unsigned long flags;
1220 * We do the initial early heuristics without holding
1221 * any task-queue locks at all. We'll only try to get
1222 * the runqueue lock when things look like they will
1228 * If the task is actively running on another CPU
1229 * still, just relax and busy-wait without holding
1232 * NOTE! Since we don't hold any locks, it's not
1233 * even sure that "rq" stays as the right runqueue!
1234 * But we don't care, since "task_running()" will
1235 * return false if the runqueue has changed and p
1236 * is actually now running somewhere else!
1238 while (task_running(rq, p))
1242 * Ok, time to look more closely! We need the rq
1243 * lock now, to be *sure*. If we're wrong, we'll
1244 * just go back and repeat.
1246 rq = task_rq_lock(p, &flags);
1247 running = task_running(rq, p);
1248 on_rq = p->se.on_rq;
1249 task_rq_unlock(rq, &flags);
1252 * Was it really running after all now that we
1253 * checked with the proper locks actually held?
1255 * Oops. Go back and try again..
1257 if (unlikely(running)) {
1263 * It's not enough that it's not actively running,
1264 * it must be off the runqueue _entirely_, and not
1267 * So if it wa still runnable (but just not actively
1268 * running right now), it's preempted, and we should
1269 * yield - it could be a while.
1271 if (unlikely(on_rq)) {
1272 schedule_timeout_uninterruptible(1);
1277 * Ahh, all good. It wasn't running, and it wasn't
1278 * runnable, which means that it will never become
1279 * running in the future either. We're all done!
1286 * kick_process - kick a running thread to enter/exit the kernel
1287 * @p: the to-be-kicked thread
1289 * Cause a process which is running on another CPU to enter
1290 * kernel-mode, without any delay. (to get signals handled.)
1292 * NOTE: this function doesnt have to take the runqueue lock,
1293 * because all it wants to ensure is that the remote task enters
1294 * the kernel. If the IPI races and the task has been migrated
1295 * to another CPU then no harm is done and the purpose has been
1298 void kick_process(struct task_struct *p)
1304 if ((cpu != smp_processor_id()) && task_curr(p))
1305 smp_send_reschedule(cpu);
1310 * Return a low guess at the load of a migration-source cpu weighted
1311 * according to the scheduling class and "nice" value.
1313 * We want to under-estimate the load of migration sources, to
1314 * balance conservatively.
1316 static unsigned long source_load(int cpu, int type)
1318 struct rq *rq = cpu_rq(cpu);
1319 unsigned long total = weighted_cpuload(cpu);
1324 return min(rq->cpu_load[type-1], total);
1328 * Return a high guess at the load of a migration-target cpu weighted
1329 * according to the scheduling class and "nice" value.
1331 static unsigned long target_load(int cpu, int type)
1333 struct rq *rq = cpu_rq(cpu);
1334 unsigned long total = weighted_cpuload(cpu);
1339 return max(rq->cpu_load[type-1], total);
1343 * Return the average load per task on the cpu's run queue
1345 static inline unsigned long cpu_avg_load_per_task(int cpu)
1347 struct rq *rq = cpu_rq(cpu);
1348 unsigned long total = weighted_cpuload(cpu);
1349 unsigned long n = rq->nr_running;
1351 return n ? total / n : SCHED_LOAD_SCALE;
1355 * find_idlest_group finds and returns the least busy CPU group within the
1358 static struct sched_group *
1359 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1361 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1362 unsigned long min_load = ULONG_MAX, this_load = 0;
1363 int load_idx = sd->forkexec_idx;
1364 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1367 unsigned long load, avg_load;
1371 /* Skip over this group if it has no CPUs allowed */
1372 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1375 local_group = cpu_isset(this_cpu, group->cpumask);
1377 /* Tally up the load of all CPUs in the group */
1380 for_each_cpu_mask(i, group->cpumask) {
1381 /* Bias balancing toward cpus of our domain */
1383 load = source_load(i, load_idx);
1385 load = target_load(i, load_idx);
1390 /* Adjust by relative CPU power of the group */
1391 avg_load = sg_div_cpu_power(group,
1392 avg_load * SCHED_LOAD_SCALE);
1395 this_load = avg_load;
1397 } else if (avg_load < min_load) {
1398 min_load = avg_load;
1401 } while (group = group->next, group != sd->groups);
1403 if (!idlest || 100*this_load < imbalance*min_load)
1409 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1412 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1415 unsigned long load, min_load = ULONG_MAX;
1419 /* Traverse only the allowed CPUs */
1420 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1422 for_each_cpu_mask(i, tmp) {
1423 load = weighted_cpuload(i);
1425 if (load < min_load || (load == min_load && i == this_cpu)) {
1435 * sched_balance_self: balance the current task (running on cpu) in domains
1436 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1439 * Balance, ie. select the least loaded group.
1441 * Returns the target CPU number, or the same CPU if no balancing is needed.
1443 * preempt must be disabled.
1445 static int sched_balance_self(int cpu, int flag)
1447 struct task_struct *t = current;
1448 struct sched_domain *tmp, *sd = NULL;
1450 for_each_domain(cpu, tmp) {
1452 * If power savings logic is enabled for a domain, stop there.
1454 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1456 if (tmp->flags & flag)
1462 struct sched_group *group;
1463 int new_cpu, weight;
1465 if (!(sd->flags & flag)) {
1471 group = find_idlest_group(sd, t, cpu);
1477 new_cpu = find_idlest_cpu(group, t, cpu);
1478 if (new_cpu == -1 || new_cpu == cpu) {
1479 /* Now try balancing at a lower domain level of cpu */
1484 /* Now try balancing at a lower domain level of new_cpu */
1487 weight = cpus_weight(span);
1488 for_each_domain(cpu, tmp) {
1489 if (weight <= cpus_weight(tmp->span))
1491 if (tmp->flags & flag)
1494 /* while loop will break here if sd == NULL */
1500 #endif /* CONFIG_SMP */
1503 * wake_idle() will wake a task on an idle cpu if task->cpu is
1504 * not idle and an idle cpu is available. The span of cpus to
1505 * search starts with cpus closest then further out as needed,
1506 * so we always favor a closer, idle cpu.
1508 * Returns the CPU we should wake onto.
1510 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1511 static int wake_idle(int cpu, struct task_struct *p)
1514 struct sched_domain *sd;
1518 * If it is idle, then it is the best cpu to run this task.
1520 * This cpu is also the best, if it has more than one task already.
1521 * Siblings must be also busy(in most cases) as they didn't already
1522 * pickup the extra load from this cpu and hence we need not check
1523 * sibling runqueue info. This will avoid the checks and cache miss
1524 * penalities associated with that.
1526 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1529 for_each_domain(cpu, sd) {
1530 if (sd->flags & SD_WAKE_IDLE) {
1531 cpus_and(tmp, sd->span, p->cpus_allowed);
1532 for_each_cpu_mask(i, tmp) {
1534 if (i != task_cpu(p)) {
1536 se.nr_wakeups_idle);
1548 static inline int wake_idle(int cpu, struct task_struct *p)
1555 * try_to_wake_up - wake up a thread
1556 * @p: the to-be-woken-up thread
1557 * @state: the mask of task states that can be woken
1558 * @sync: do a synchronous wakeup?
1560 * Put it on the run-queue if it's not already there. The "current"
1561 * thread is always on the run-queue (except when the actual
1562 * re-schedule is in progress), and as such you're allowed to do
1563 * the simpler "current->state = TASK_RUNNING" to mark yourself
1564 * runnable without the overhead of this.
1566 * returns failure only if the task is already active.
1568 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1570 int cpu, orig_cpu, this_cpu, success = 0;
1571 unsigned long flags;
1575 struct sched_domain *sd, *this_sd = NULL;
1576 unsigned long load, this_load;
1580 rq = task_rq_lock(p, &flags);
1581 old_state = p->state;
1582 if (!(old_state & state))
1590 this_cpu = smp_processor_id();
1593 if (unlikely(task_running(rq, p)))
1598 schedstat_inc(rq, ttwu_count);
1599 if (cpu == this_cpu) {
1600 schedstat_inc(rq, ttwu_local);
1604 for_each_domain(this_cpu, sd) {
1605 if (cpu_isset(cpu, sd->span)) {
1606 schedstat_inc(sd, ttwu_wake_remote);
1612 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1616 * Check for affine wakeup and passive balancing possibilities.
1619 int idx = this_sd->wake_idx;
1620 unsigned int imbalance;
1622 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1624 load = source_load(cpu, idx);
1625 this_load = target_load(this_cpu, idx);
1627 new_cpu = this_cpu; /* Wake to this CPU if we can */
1629 if (this_sd->flags & SD_WAKE_AFFINE) {
1630 unsigned long tl = this_load;
1631 unsigned long tl_per_task;
1634 * Attract cache-cold tasks on sync wakeups:
1636 if (sync && !task_hot(p, rq->clock, this_sd))
1639 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1640 tl_per_task = cpu_avg_load_per_task(this_cpu);
1643 * If sync wakeup then subtract the (maximum possible)
1644 * effect of the currently running task from the load
1645 * of the current CPU:
1648 tl -= current->se.load.weight;
1651 tl + target_load(cpu, idx) <= tl_per_task) ||
1652 100*(tl + p->se.load.weight) <= imbalance*load) {
1654 * This domain has SD_WAKE_AFFINE and
1655 * p is cache cold in this domain, and
1656 * there is no bad imbalance.
1658 schedstat_inc(this_sd, ttwu_move_affine);
1659 schedstat_inc(p, se.nr_wakeups_affine);
1665 * Start passive balancing when half the imbalance_pct
1668 if (this_sd->flags & SD_WAKE_BALANCE) {
1669 if (imbalance*this_load <= 100*load) {
1670 schedstat_inc(this_sd, ttwu_move_balance);
1671 schedstat_inc(p, se.nr_wakeups_passive);
1677 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1679 new_cpu = wake_idle(new_cpu, p);
1680 if (new_cpu != cpu) {
1681 set_task_cpu(p, new_cpu);
1682 task_rq_unlock(rq, &flags);
1683 /* might preempt at this point */
1684 rq = task_rq_lock(p, &flags);
1685 old_state = p->state;
1686 if (!(old_state & state))
1691 this_cpu = smp_processor_id();
1696 #endif /* CONFIG_SMP */
1697 schedstat_inc(p, se.nr_wakeups);
1699 schedstat_inc(p, se.nr_wakeups_sync);
1700 if (orig_cpu != cpu)
1701 schedstat_inc(p, se.nr_wakeups_migrate);
1702 if (cpu == this_cpu)
1703 schedstat_inc(p, se.nr_wakeups_local);
1705 schedstat_inc(p, se.nr_wakeups_remote);
1706 update_rq_clock(rq);
1707 activate_task(rq, p, 1);
1708 check_preempt_curr(rq, p);
1712 p->state = TASK_RUNNING;
1714 task_rq_unlock(rq, &flags);
1719 int fastcall wake_up_process(struct task_struct *p)
1721 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1722 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1724 EXPORT_SYMBOL(wake_up_process);
1726 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1728 return try_to_wake_up(p, state, 0);
1732 * Perform scheduler related setup for a newly forked process p.
1733 * p is forked by current.
1735 * __sched_fork() is basic setup used by init_idle() too:
1737 static void __sched_fork(struct task_struct *p)
1739 p->se.exec_start = 0;
1740 p->se.sum_exec_runtime = 0;
1741 p->se.prev_sum_exec_runtime = 0;
1743 #ifdef CONFIG_SCHEDSTATS
1744 p->se.wait_start = 0;
1745 p->se.sum_sleep_runtime = 0;
1746 p->se.sleep_start = 0;
1747 p->se.block_start = 0;
1748 p->se.sleep_max = 0;
1749 p->se.block_max = 0;
1751 p->se.slice_max = 0;
1755 INIT_LIST_HEAD(&p->run_list);
1758 #ifdef CONFIG_PREEMPT_NOTIFIERS
1759 INIT_HLIST_HEAD(&p->preempt_notifiers);
1763 * We mark the process as running here, but have not actually
1764 * inserted it onto the runqueue yet. This guarantees that
1765 * nobody will actually run it, and a signal or other external
1766 * event cannot wake it up and insert it on the runqueue either.
1768 p->state = TASK_RUNNING;
1772 * fork()/clone()-time setup:
1774 void sched_fork(struct task_struct *p, int clone_flags)
1776 int cpu = get_cpu();
1781 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1783 set_task_cpu(p, cpu);
1786 * Make sure we do not leak PI boosting priority to the child:
1788 p->prio = current->normal_prio;
1789 if (!rt_prio(p->prio))
1790 p->sched_class = &fair_sched_class;
1792 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1793 if (likely(sched_info_on()))
1794 memset(&p->sched_info, 0, sizeof(p->sched_info));
1796 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1799 #ifdef CONFIG_PREEMPT
1800 /* Want to start with kernel preemption disabled. */
1801 task_thread_info(p)->preempt_count = 1;
1807 * wake_up_new_task - wake up a newly created task for the first time.
1809 * This function will do some initial scheduler statistics housekeeping
1810 * that must be done for every newly created context, then puts the task
1811 * on the runqueue and wakes it.
1813 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1815 unsigned long flags;
1818 rq = task_rq_lock(p, &flags);
1819 BUG_ON(p->state != TASK_RUNNING);
1820 update_rq_clock(rq);
1822 p->prio = effective_prio(p);
1824 if (!p->sched_class->task_new || !current->se.on_rq) {
1825 activate_task(rq, p, 0);
1828 * Let the scheduling class do new task startup
1829 * management (if any):
1831 p->sched_class->task_new(rq, p);
1832 inc_nr_running(p, rq);
1834 check_preempt_curr(rq, p);
1835 task_rq_unlock(rq, &flags);
1838 #ifdef CONFIG_PREEMPT_NOTIFIERS
1841 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1842 * @notifier: notifier struct to register
1844 void preempt_notifier_register(struct preempt_notifier *notifier)
1846 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1848 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1851 * preempt_notifier_unregister - no longer interested in preemption notifications
1852 * @notifier: notifier struct to unregister
1854 * This is safe to call from within a preemption notifier.
1856 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1858 hlist_del(¬ifier->link);
1860 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1862 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1864 struct preempt_notifier *notifier;
1865 struct hlist_node *node;
1867 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1868 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1872 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1873 struct task_struct *next)
1875 struct preempt_notifier *notifier;
1876 struct hlist_node *node;
1878 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1879 notifier->ops->sched_out(notifier, next);
1884 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1889 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1890 struct task_struct *next)
1897 * prepare_task_switch - prepare to switch tasks
1898 * @rq: the runqueue preparing to switch
1899 * @prev: the current task that is being switched out
1900 * @next: the task we are going to switch to.
1902 * This is called with the rq lock held and interrupts off. It must
1903 * be paired with a subsequent finish_task_switch after the context
1906 * prepare_task_switch sets up locking and calls architecture specific
1910 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1911 struct task_struct *next)
1913 fire_sched_out_preempt_notifiers(prev, next);
1914 prepare_lock_switch(rq, next);
1915 prepare_arch_switch(next);
1919 * finish_task_switch - clean up after a task-switch
1920 * @rq: runqueue associated with task-switch
1921 * @prev: the thread we just switched away from.
1923 * finish_task_switch must be called after the context switch, paired
1924 * with a prepare_task_switch call before the context switch.
1925 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1926 * and do any other architecture-specific cleanup actions.
1928 * Note that we may have delayed dropping an mm in context_switch(). If
1929 * so, we finish that here outside of the runqueue lock. (Doing it
1930 * with the lock held can cause deadlocks; see schedule() for
1933 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1934 __releases(rq->lock)
1936 struct mm_struct *mm = rq->prev_mm;
1942 * A task struct has one reference for the use as "current".
1943 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1944 * schedule one last time. The schedule call will never return, and
1945 * the scheduled task must drop that reference.
1946 * The test for TASK_DEAD must occur while the runqueue locks are
1947 * still held, otherwise prev could be scheduled on another cpu, die
1948 * there before we look at prev->state, and then the reference would
1950 * Manfred Spraul <manfred@colorfullife.com>
1952 prev_state = prev->state;
1953 finish_arch_switch(prev);
1954 finish_lock_switch(rq, prev);
1955 fire_sched_in_preempt_notifiers(current);
1958 if (unlikely(prev_state == TASK_DEAD)) {
1960 * Remove function-return probe instances associated with this
1961 * task and put them back on the free list.
1963 kprobe_flush_task(prev);
1964 put_task_struct(prev);
1969 * schedule_tail - first thing a freshly forked thread must call.
1970 * @prev: the thread we just switched away from.
1972 asmlinkage void schedule_tail(struct task_struct *prev)
1973 __releases(rq->lock)
1975 struct rq *rq = this_rq();
1977 finish_task_switch(rq, prev);
1978 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1979 /* In this case, finish_task_switch does not reenable preemption */
1982 if (current->set_child_tid)
1983 put_user(task_pid_vnr(current), current->set_child_tid);
1987 * context_switch - switch to the new MM and the new
1988 * thread's register state.
1991 context_switch(struct rq *rq, struct task_struct *prev,
1992 struct task_struct *next)
1994 struct mm_struct *mm, *oldmm;
1996 prepare_task_switch(rq, prev, next);
1998 oldmm = prev->active_mm;
2000 * For paravirt, this is coupled with an exit in switch_to to
2001 * combine the page table reload and the switch backend into
2004 arch_enter_lazy_cpu_mode();
2006 if (unlikely(!mm)) {
2007 next->active_mm = oldmm;
2008 atomic_inc(&oldmm->mm_count);
2009 enter_lazy_tlb(oldmm, next);
2011 switch_mm(oldmm, mm, next);
2013 if (unlikely(!prev->mm)) {
2014 prev->active_mm = NULL;
2015 rq->prev_mm = oldmm;
2018 * Since the runqueue lock will be released by the next
2019 * task (which is an invalid locking op but in the case
2020 * of the scheduler it's an obvious special-case), so we
2021 * do an early lockdep release here:
2023 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2024 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2027 /* Here we just switch the register state and the stack. */
2028 switch_to(prev, next, prev);
2032 * this_rq must be evaluated again because prev may have moved
2033 * CPUs since it called schedule(), thus the 'rq' on its stack
2034 * frame will be invalid.
2036 finish_task_switch(this_rq(), prev);
2040 * nr_running, nr_uninterruptible and nr_context_switches:
2042 * externally visible scheduler statistics: current number of runnable
2043 * threads, current number of uninterruptible-sleeping threads, total
2044 * number of context switches performed since bootup.
2046 unsigned long nr_running(void)
2048 unsigned long i, sum = 0;
2050 for_each_online_cpu(i)
2051 sum += cpu_rq(i)->nr_running;
2056 unsigned long nr_uninterruptible(void)
2058 unsigned long i, sum = 0;
2060 for_each_possible_cpu(i)
2061 sum += cpu_rq(i)->nr_uninterruptible;
2064 * Since we read the counters lockless, it might be slightly
2065 * inaccurate. Do not allow it to go below zero though:
2067 if (unlikely((long)sum < 0))
2073 unsigned long long nr_context_switches(void)
2076 unsigned long long sum = 0;
2078 for_each_possible_cpu(i)
2079 sum += cpu_rq(i)->nr_switches;
2084 unsigned long nr_iowait(void)
2086 unsigned long i, sum = 0;
2088 for_each_possible_cpu(i)
2089 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2094 unsigned long nr_active(void)
2096 unsigned long i, running = 0, uninterruptible = 0;
2098 for_each_online_cpu(i) {
2099 running += cpu_rq(i)->nr_running;
2100 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2103 if (unlikely((long)uninterruptible < 0))
2104 uninterruptible = 0;
2106 return running + uninterruptible;
2110 * Update rq->cpu_load[] statistics. This function is usually called every
2111 * scheduler tick (TICK_NSEC).
2113 static void update_cpu_load(struct rq *this_rq)
2115 unsigned long this_load = this_rq->load.weight;
2118 this_rq->nr_load_updates++;
2120 /* Update our load: */
2121 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2122 unsigned long old_load, new_load;
2124 /* scale is effectively 1 << i now, and >> i divides by scale */
2126 old_load = this_rq->cpu_load[i];
2127 new_load = this_load;
2129 * Round up the averaging division if load is increasing. This
2130 * prevents us from getting stuck on 9 if the load is 10, for
2133 if (new_load > old_load)
2134 new_load += scale-1;
2135 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2142 * double_rq_lock - safely lock two runqueues
2144 * Note this does not disable interrupts like task_rq_lock,
2145 * you need to do so manually before calling.
2147 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2148 __acquires(rq1->lock)
2149 __acquires(rq2->lock)
2151 BUG_ON(!irqs_disabled());
2153 spin_lock(&rq1->lock);
2154 __acquire(rq2->lock); /* Fake it out ;) */
2157 spin_lock(&rq1->lock);
2158 spin_lock(&rq2->lock);
2160 spin_lock(&rq2->lock);
2161 spin_lock(&rq1->lock);
2164 update_rq_clock(rq1);
2165 update_rq_clock(rq2);
2169 * double_rq_unlock - safely unlock two runqueues
2171 * Note this does not restore interrupts like task_rq_unlock,
2172 * you need to do so manually after calling.
2174 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2175 __releases(rq1->lock)
2176 __releases(rq2->lock)
2178 spin_unlock(&rq1->lock);
2180 spin_unlock(&rq2->lock);
2182 __release(rq2->lock);
2186 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2188 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2189 __releases(this_rq->lock)
2190 __acquires(busiest->lock)
2191 __acquires(this_rq->lock)
2193 if (unlikely(!irqs_disabled())) {
2194 /* printk() doesn't work good under rq->lock */
2195 spin_unlock(&this_rq->lock);
2198 if (unlikely(!spin_trylock(&busiest->lock))) {
2199 if (busiest < this_rq) {
2200 spin_unlock(&this_rq->lock);
2201 spin_lock(&busiest->lock);
2202 spin_lock(&this_rq->lock);
2204 spin_lock(&busiest->lock);
2209 * If dest_cpu is allowed for this process, migrate the task to it.
2210 * This is accomplished by forcing the cpu_allowed mask to only
2211 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2212 * the cpu_allowed mask is restored.
2214 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2216 struct migration_req req;
2217 unsigned long flags;
2220 rq = task_rq_lock(p, &flags);
2221 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2222 || unlikely(cpu_is_offline(dest_cpu)))
2225 /* force the process onto the specified CPU */
2226 if (migrate_task(p, dest_cpu, &req)) {
2227 /* Need to wait for migration thread (might exit: take ref). */
2228 struct task_struct *mt = rq->migration_thread;
2230 get_task_struct(mt);
2231 task_rq_unlock(rq, &flags);
2232 wake_up_process(mt);
2233 put_task_struct(mt);
2234 wait_for_completion(&req.done);
2239 task_rq_unlock(rq, &flags);
2243 * sched_exec - execve() is a valuable balancing opportunity, because at
2244 * this point the task has the smallest effective memory and cache footprint.
2246 void sched_exec(void)
2248 int new_cpu, this_cpu = get_cpu();
2249 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2251 if (new_cpu != this_cpu)
2252 sched_migrate_task(current, new_cpu);
2256 * pull_task - move a task from a remote runqueue to the local runqueue.
2257 * Both runqueues must be locked.
2259 static void pull_task(struct rq *src_rq, struct task_struct *p,
2260 struct rq *this_rq, int this_cpu)
2262 deactivate_task(src_rq, p, 0);
2263 set_task_cpu(p, this_cpu);
2264 activate_task(this_rq, p, 0);
2266 * Note that idle threads have a prio of MAX_PRIO, for this test
2267 * to be always true for them.
2269 check_preempt_curr(this_rq, p);
2273 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2276 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2277 struct sched_domain *sd, enum cpu_idle_type idle,
2281 * We do not migrate tasks that are:
2282 * 1) running (obviously), or
2283 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2284 * 3) are cache-hot on their current CPU.
2286 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2287 schedstat_inc(p, se.nr_failed_migrations_affine);
2292 if (task_running(rq, p)) {
2293 schedstat_inc(p, se.nr_failed_migrations_running);
2298 * Aggressive migration if:
2299 * 1) task is cache cold, or
2300 * 2) too many balance attempts have failed.
2303 if (!task_hot(p, rq->clock, sd) ||
2304 sd->nr_balance_failed > sd->cache_nice_tries) {
2305 #ifdef CONFIG_SCHEDSTATS
2306 if (task_hot(p, rq->clock, sd)) {
2307 schedstat_inc(sd, lb_hot_gained[idle]);
2308 schedstat_inc(p, se.nr_forced_migrations);
2314 if (task_hot(p, rq->clock, sd)) {
2315 schedstat_inc(p, se.nr_failed_migrations_hot);
2321 static unsigned long
2322 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2323 unsigned long max_load_move, struct sched_domain *sd,
2324 enum cpu_idle_type idle, int *all_pinned,
2325 int *this_best_prio, struct rq_iterator *iterator)
2327 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2328 struct task_struct *p;
2329 long rem_load_move = max_load_move;
2331 if (max_load_move == 0)
2337 * Start the load-balancing iterator:
2339 p = iterator->start(iterator->arg);
2341 if (!p || loops++ > sysctl_sched_nr_migrate)
2344 * To help distribute high priority tasks across CPUs we don't
2345 * skip a task if it will be the highest priority task (i.e. smallest
2346 * prio value) on its new queue regardless of its load weight
2348 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2349 SCHED_LOAD_SCALE_FUZZ;
2350 if ((skip_for_load && p->prio >= *this_best_prio) ||
2351 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2352 p = iterator->next(iterator->arg);
2356 pull_task(busiest, p, this_rq, this_cpu);
2358 rem_load_move -= p->se.load.weight;
2361 * We only want to steal up to the prescribed amount of weighted load.
2363 if (rem_load_move > 0) {
2364 if (p->prio < *this_best_prio)
2365 *this_best_prio = p->prio;
2366 p = iterator->next(iterator->arg);
2371 * Right now, this is one of only two places pull_task() is called,
2372 * so we can safely collect pull_task() stats here rather than
2373 * inside pull_task().
2375 schedstat_add(sd, lb_gained[idle], pulled);
2378 *all_pinned = pinned;
2380 return max_load_move - rem_load_move;
2384 * move_tasks tries to move up to max_load_move weighted load from busiest to
2385 * this_rq, as part of a balancing operation within domain "sd".
2386 * Returns 1 if successful and 0 otherwise.
2388 * Called with both runqueues locked.
2390 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2391 unsigned long max_load_move,
2392 struct sched_domain *sd, enum cpu_idle_type idle,
2395 const struct sched_class *class = sched_class_highest;
2396 unsigned long total_load_moved = 0;
2397 int this_best_prio = this_rq->curr->prio;
2401 class->load_balance(this_rq, this_cpu, busiest,
2402 max_load_move - total_load_moved,
2403 sd, idle, all_pinned, &this_best_prio);
2404 class = class->next;
2405 } while (class && max_load_move > total_load_moved);
2407 return total_load_moved > 0;
2411 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2412 struct sched_domain *sd, enum cpu_idle_type idle,
2413 struct rq_iterator *iterator)
2415 struct task_struct *p = iterator->start(iterator->arg);
2419 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2420 pull_task(busiest, p, this_rq, this_cpu);
2422 * Right now, this is only the second place pull_task()
2423 * is called, so we can safely collect pull_task()
2424 * stats here rather than inside pull_task().
2426 schedstat_inc(sd, lb_gained[idle]);
2430 p = iterator->next(iterator->arg);
2437 * move_one_task tries to move exactly one task from busiest to this_rq, as
2438 * part of active balancing operations within "domain".
2439 * Returns 1 if successful and 0 otherwise.
2441 * Called with both runqueues locked.
2443 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2444 struct sched_domain *sd, enum cpu_idle_type idle)
2446 const struct sched_class *class;
2448 for (class = sched_class_highest; class; class = class->next)
2449 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2456 * find_busiest_group finds and returns the busiest CPU group within the
2457 * domain. It calculates and returns the amount of weighted load which
2458 * should be moved to restore balance via the imbalance parameter.
2460 static struct sched_group *
2461 find_busiest_group(struct sched_domain *sd, int this_cpu,
2462 unsigned long *imbalance, enum cpu_idle_type idle,
2463 int *sd_idle, cpumask_t *cpus, int *balance)
2465 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2466 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2467 unsigned long max_pull;
2468 unsigned long busiest_load_per_task, busiest_nr_running;
2469 unsigned long this_load_per_task, this_nr_running;
2470 int load_idx, group_imb = 0;
2471 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2472 int power_savings_balance = 1;
2473 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2474 unsigned long min_nr_running = ULONG_MAX;
2475 struct sched_group *group_min = NULL, *group_leader = NULL;
2478 max_load = this_load = total_load = total_pwr = 0;
2479 busiest_load_per_task = busiest_nr_running = 0;
2480 this_load_per_task = this_nr_running = 0;
2481 if (idle == CPU_NOT_IDLE)
2482 load_idx = sd->busy_idx;
2483 else if (idle == CPU_NEWLY_IDLE)
2484 load_idx = sd->newidle_idx;
2486 load_idx = sd->idle_idx;
2489 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2492 int __group_imb = 0;
2493 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2494 unsigned long sum_nr_running, sum_weighted_load;
2496 local_group = cpu_isset(this_cpu, group->cpumask);
2499 balance_cpu = first_cpu(group->cpumask);
2501 /* Tally up the load of all CPUs in the group */
2502 sum_weighted_load = sum_nr_running = avg_load = 0;
2504 min_cpu_load = ~0UL;
2506 for_each_cpu_mask(i, group->cpumask) {
2509 if (!cpu_isset(i, *cpus))
2514 if (*sd_idle && rq->nr_running)
2517 /* Bias balancing toward cpus of our domain */
2519 if (idle_cpu(i) && !first_idle_cpu) {
2524 load = target_load(i, load_idx);
2526 load = source_load(i, load_idx);
2527 if (load > max_cpu_load)
2528 max_cpu_load = load;
2529 if (min_cpu_load > load)
2530 min_cpu_load = load;
2534 sum_nr_running += rq->nr_running;
2535 sum_weighted_load += weighted_cpuload(i);
2539 * First idle cpu or the first cpu(busiest) in this sched group
2540 * is eligible for doing load balancing at this and above
2541 * domains. In the newly idle case, we will allow all the cpu's
2542 * to do the newly idle load balance.
2544 if (idle != CPU_NEWLY_IDLE && local_group &&
2545 balance_cpu != this_cpu && balance) {
2550 total_load += avg_load;
2551 total_pwr += group->__cpu_power;
2553 /* Adjust by relative CPU power of the group */
2554 avg_load = sg_div_cpu_power(group,
2555 avg_load * SCHED_LOAD_SCALE);
2557 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2560 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2563 this_load = avg_load;
2565 this_nr_running = sum_nr_running;
2566 this_load_per_task = sum_weighted_load;
2567 } else if (avg_load > max_load &&
2568 (sum_nr_running > group_capacity || __group_imb)) {
2569 max_load = avg_load;
2571 busiest_nr_running = sum_nr_running;
2572 busiest_load_per_task = sum_weighted_load;
2573 group_imb = __group_imb;
2576 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2578 * Busy processors will not participate in power savings
2581 if (idle == CPU_NOT_IDLE ||
2582 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2586 * If the local group is idle or completely loaded
2587 * no need to do power savings balance at this domain
2589 if (local_group && (this_nr_running >= group_capacity ||
2591 power_savings_balance = 0;
2594 * If a group is already running at full capacity or idle,
2595 * don't include that group in power savings calculations
2597 if (!power_savings_balance || sum_nr_running >= group_capacity
2602 * Calculate the group which has the least non-idle load.
2603 * This is the group from where we need to pick up the load
2606 if ((sum_nr_running < min_nr_running) ||
2607 (sum_nr_running == min_nr_running &&
2608 first_cpu(group->cpumask) <
2609 first_cpu(group_min->cpumask))) {
2611 min_nr_running = sum_nr_running;
2612 min_load_per_task = sum_weighted_load /
2617 * Calculate the group which is almost near its
2618 * capacity but still has some space to pick up some load
2619 * from other group and save more power
2621 if (sum_nr_running <= group_capacity - 1) {
2622 if (sum_nr_running > leader_nr_running ||
2623 (sum_nr_running == leader_nr_running &&
2624 first_cpu(group->cpumask) >
2625 first_cpu(group_leader->cpumask))) {
2626 group_leader = group;
2627 leader_nr_running = sum_nr_running;
2632 group = group->next;
2633 } while (group != sd->groups);
2635 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2638 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2640 if (this_load >= avg_load ||
2641 100*max_load <= sd->imbalance_pct*this_load)
2644 busiest_load_per_task /= busiest_nr_running;
2646 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2649 * We're trying to get all the cpus to the average_load, so we don't
2650 * want to push ourselves above the average load, nor do we wish to
2651 * reduce the max loaded cpu below the average load, as either of these
2652 * actions would just result in more rebalancing later, and ping-pong
2653 * tasks around. Thus we look for the minimum possible imbalance.
2654 * Negative imbalances (*we* are more loaded than anyone else) will
2655 * be counted as no imbalance for these purposes -- we can't fix that
2656 * by pulling tasks to us. Be careful of negative numbers as they'll
2657 * appear as very large values with unsigned longs.
2659 if (max_load <= busiest_load_per_task)
2663 * In the presence of smp nice balancing, certain scenarios can have
2664 * max load less than avg load(as we skip the groups at or below
2665 * its cpu_power, while calculating max_load..)
2667 if (max_load < avg_load) {
2669 goto small_imbalance;
2672 /* Don't want to pull so many tasks that a group would go idle */
2673 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2675 /* How much load to actually move to equalise the imbalance */
2676 *imbalance = min(max_pull * busiest->__cpu_power,
2677 (avg_load - this_load) * this->__cpu_power)
2681 * if *imbalance is less than the average load per runnable task
2682 * there is no gaurantee that any tasks will be moved so we'll have
2683 * a think about bumping its value to force at least one task to be
2686 if (*imbalance < busiest_load_per_task) {
2687 unsigned long tmp, pwr_now, pwr_move;
2691 pwr_move = pwr_now = 0;
2693 if (this_nr_running) {
2694 this_load_per_task /= this_nr_running;
2695 if (busiest_load_per_task > this_load_per_task)
2698 this_load_per_task = SCHED_LOAD_SCALE;
2700 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2701 busiest_load_per_task * imbn) {
2702 *imbalance = busiest_load_per_task;
2707 * OK, we don't have enough imbalance to justify moving tasks,
2708 * however we may be able to increase total CPU power used by
2712 pwr_now += busiest->__cpu_power *
2713 min(busiest_load_per_task, max_load);
2714 pwr_now += this->__cpu_power *
2715 min(this_load_per_task, this_load);
2716 pwr_now /= SCHED_LOAD_SCALE;
2718 /* Amount of load we'd subtract */
2719 tmp = sg_div_cpu_power(busiest,
2720 busiest_load_per_task * SCHED_LOAD_SCALE);
2722 pwr_move += busiest->__cpu_power *
2723 min(busiest_load_per_task, max_load - tmp);
2725 /* Amount of load we'd add */
2726 if (max_load * busiest->__cpu_power <
2727 busiest_load_per_task * SCHED_LOAD_SCALE)
2728 tmp = sg_div_cpu_power(this,
2729 max_load * busiest->__cpu_power);
2731 tmp = sg_div_cpu_power(this,
2732 busiest_load_per_task * SCHED_LOAD_SCALE);
2733 pwr_move += this->__cpu_power *
2734 min(this_load_per_task, this_load + tmp);
2735 pwr_move /= SCHED_LOAD_SCALE;
2737 /* Move if we gain throughput */
2738 if (pwr_move > pwr_now)
2739 *imbalance = busiest_load_per_task;
2745 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2746 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2749 if (this == group_leader && group_leader != group_min) {
2750 *imbalance = min_load_per_task;
2760 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2763 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2764 unsigned long imbalance, cpumask_t *cpus)
2766 struct rq *busiest = NULL, *rq;
2767 unsigned long max_load = 0;
2770 for_each_cpu_mask(i, group->cpumask) {
2773 if (!cpu_isset(i, *cpus))
2777 wl = weighted_cpuload(i);
2779 if (rq->nr_running == 1 && wl > imbalance)
2782 if (wl > max_load) {
2792 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2793 * so long as it is large enough.
2795 #define MAX_PINNED_INTERVAL 512
2798 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2799 * tasks if there is an imbalance.
2801 static int load_balance(int this_cpu, struct rq *this_rq,
2802 struct sched_domain *sd, enum cpu_idle_type idle,
2805 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2806 struct sched_group *group;
2807 unsigned long imbalance;
2809 cpumask_t cpus = CPU_MASK_ALL;
2810 unsigned long flags;
2813 * When power savings policy is enabled for the parent domain, idle
2814 * sibling can pick up load irrespective of busy siblings. In this case,
2815 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2816 * portraying it as CPU_NOT_IDLE.
2818 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2819 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2822 schedstat_inc(sd, lb_count[idle]);
2825 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2832 schedstat_inc(sd, lb_nobusyg[idle]);
2836 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2838 schedstat_inc(sd, lb_nobusyq[idle]);
2842 BUG_ON(busiest == this_rq);
2844 schedstat_add(sd, lb_imbalance[idle], imbalance);
2847 if (busiest->nr_running > 1) {
2849 * Attempt to move tasks. If find_busiest_group has found
2850 * an imbalance but busiest->nr_running <= 1, the group is
2851 * still unbalanced. ld_moved simply stays zero, so it is
2852 * correctly treated as an imbalance.
2854 local_irq_save(flags);
2855 double_rq_lock(this_rq, busiest);
2856 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2857 imbalance, sd, idle, &all_pinned);
2858 double_rq_unlock(this_rq, busiest);
2859 local_irq_restore(flags);
2862 * some other cpu did the load balance for us.
2864 if (ld_moved && this_cpu != smp_processor_id())
2865 resched_cpu(this_cpu);
2867 /* All tasks on this runqueue were pinned by CPU affinity */
2868 if (unlikely(all_pinned)) {
2869 cpu_clear(cpu_of(busiest), cpus);
2870 if (!cpus_empty(cpus))
2877 schedstat_inc(sd, lb_failed[idle]);
2878 sd->nr_balance_failed++;
2880 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2882 spin_lock_irqsave(&busiest->lock, flags);
2884 /* don't kick the migration_thread, if the curr
2885 * task on busiest cpu can't be moved to this_cpu
2887 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2888 spin_unlock_irqrestore(&busiest->lock, flags);
2890 goto out_one_pinned;
2893 if (!busiest->active_balance) {
2894 busiest->active_balance = 1;
2895 busiest->push_cpu = this_cpu;
2898 spin_unlock_irqrestore(&busiest->lock, flags);
2900 wake_up_process(busiest->migration_thread);
2903 * We've kicked active balancing, reset the failure
2906 sd->nr_balance_failed = sd->cache_nice_tries+1;
2909 sd->nr_balance_failed = 0;
2911 if (likely(!active_balance)) {
2912 /* We were unbalanced, so reset the balancing interval */
2913 sd->balance_interval = sd->min_interval;
2916 * If we've begun active balancing, start to back off. This
2917 * case may not be covered by the all_pinned logic if there
2918 * is only 1 task on the busy runqueue (because we don't call
2921 if (sd->balance_interval < sd->max_interval)
2922 sd->balance_interval *= 2;
2925 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2926 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2931 schedstat_inc(sd, lb_balanced[idle]);
2933 sd->nr_balance_failed = 0;
2936 /* tune up the balancing interval */
2937 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2938 (sd->balance_interval < sd->max_interval))
2939 sd->balance_interval *= 2;
2941 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2942 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2948 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2949 * tasks if there is an imbalance.
2951 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2952 * this_rq is locked.
2955 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2957 struct sched_group *group;
2958 struct rq *busiest = NULL;
2959 unsigned long imbalance;
2963 cpumask_t cpus = CPU_MASK_ALL;
2966 * When power savings policy is enabled for the parent domain, idle
2967 * sibling can pick up load irrespective of busy siblings. In this case,
2968 * let the state of idle sibling percolate up as IDLE, instead of
2969 * portraying it as CPU_NOT_IDLE.
2971 if (sd->flags & SD_SHARE_CPUPOWER &&
2972 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2975 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2977 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2978 &sd_idle, &cpus, NULL);
2980 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2984 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2987 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2991 BUG_ON(busiest == this_rq);
2993 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2996 if (busiest->nr_running > 1) {
2997 /* Attempt to move tasks */
2998 double_lock_balance(this_rq, busiest);
2999 /* this_rq->clock is already updated */
3000 update_rq_clock(busiest);
3001 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3002 imbalance, sd, CPU_NEWLY_IDLE,
3004 spin_unlock(&busiest->lock);
3006 if (unlikely(all_pinned)) {
3007 cpu_clear(cpu_of(busiest), cpus);
3008 if (!cpus_empty(cpus))
3014 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3015 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3016 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3019 sd->nr_balance_failed = 0;
3024 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3025 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3026 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3028 sd->nr_balance_failed = 0;
3034 * idle_balance is called by schedule() if this_cpu is about to become
3035 * idle. Attempts to pull tasks from other CPUs.
3037 static void idle_balance(int this_cpu, struct rq *this_rq)
3039 struct sched_domain *sd;
3040 int pulled_task = -1;
3041 unsigned long next_balance = jiffies + HZ;
3043 for_each_domain(this_cpu, sd) {
3044 unsigned long interval;
3046 if (!(sd->flags & SD_LOAD_BALANCE))
3049 if (sd->flags & SD_BALANCE_NEWIDLE)
3050 /* If we've pulled tasks over stop searching: */
3051 pulled_task = load_balance_newidle(this_cpu,
3054 interval = msecs_to_jiffies(sd->balance_interval);
3055 if (time_after(next_balance, sd->last_balance + interval))
3056 next_balance = sd->last_balance + interval;
3060 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3062 * We are going idle. next_balance may be set based on
3063 * a busy processor. So reset next_balance.
3065 this_rq->next_balance = next_balance;
3070 * active_load_balance is run by migration threads. It pushes running tasks
3071 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3072 * running on each physical CPU where possible, and avoids physical /
3073 * logical imbalances.
3075 * Called with busiest_rq locked.
3077 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3079 int target_cpu = busiest_rq->push_cpu;
3080 struct sched_domain *sd;
3081 struct rq *target_rq;
3083 /* Is there any task to move? */
3084 if (busiest_rq->nr_running <= 1)
3087 target_rq = cpu_rq(target_cpu);
3090 * This condition is "impossible", if it occurs
3091 * we need to fix it. Originally reported by
3092 * Bjorn Helgaas on a 128-cpu setup.
3094 BUG_ON(busiest_rq == target_rq);
3096 /* move a task from busiest_rq to target_rq */
3097 double_lock_balance(busiest_rq, target_rq);
3098 update_rq_clock(busiest_rq);
3099 update_rq_clock(target_rq);
3101 /* Search for an sd spanning us and the target CPU. */
3102 for_each_domain(target_cpu, sd) {
3103 if ((sd->flags & SD_LOAD_BALANCE) &&
3104 cpu_isset(busiest_cpu, sd->span))
3109 schedstat_inc(sd, alb_count);
3111 if (move_one_task(target_rq, target_cpu, busiest_rq,
3113 schedstat_inc(sd, alb_pushed);
3115 schedstat_inc(sd, alb_failed);
3117 spin_unlock(&target_rq->lock);
3122 atomic_t load_balancer;
3124 } nohz ____cacheline_aligned = {
3125 .load_balancer = ATOMIC_INIT(-1),
3126 .cpu_mask = CPU_MASK_NONE,
3130 * This routine will try to nominate the ilb (idle load balancing)
3131 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3132 * load balancing on behalf of all those cpus. If all the cpus in the system
3133 * go into this tickless mode, then there will be no ilb owner (as there is
3134 * no need for one) and all the cpus will sleep till the next wakeup event
3137 * For the ilb owner, tick is not stopped. And this tick will be used
3138 * for idle load balancing. ilb owner will still be part of
3141 * While stopping the tick, this cpu will become the ilb owner if there
3142 * is no other owner. And will be the owner till that cpu becomes busy
3143 * or if all cpus in the system stop their ticks at which point
3144 * there is no need for ilb owner.
3146 * When the ilb owner becomes busy, it nominates another owner, during the
3147 * next busy scheduler_tick()
3149 int select_nohz_load_balancer(int stop_tick)
3151 int cpu = smp_processor_id();
3154 cpu_set(cpu, nohz.cpu_mask);
3155 cpu_rq(cpu)->in_nohz_recently = 1;
3158 * If we are going offline and still the leader, give up!
3160 if (cpu_is_offline(cpu) &&
3161 atomic_read(&nohz.load_balancer) == cpu) {
3162 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3167 /* time for ilb owner also to sleep */
3168 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3169 if (atomic_read(&nohz.load_balancer) == cpu)
3170 atomic_set(&nohz.load_balancer, -1);
3174 if (atomic_read(&nohz.load_balancer) == -1) {
3175 /* make me the ilb owner */
3176 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3178 } else if (atomic_read(&nohz.load_balancer) == cpu)
3181 if (!cpu_isset(cpu, nohz.cpu_mask))
3184 cpu_clear(cpu, nohz.cpu_mask);
3186 if (atomic_read(&nohz.load_balancer) == cpu)
3187 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3194 static DEFINE_SPINLOCK(balancing);
3197 * It checks each scheduling domain to see if it is due to be balanced,
3198 * and initiates a balancing operation if so.
3200 * Balancing parameters are set up in arch_init_sched_domains.
3202 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3205 struct rq *rq = cpu_rq(cpu);
3206 unsigned long interval;
3207 struct sched_domain *sd;
3208 /* Earliest time when we have to do rebalance again */
3209 unsigned long next_balance = jiffies + 60*HZ;
3210 int update_next_balance = 0;
3212 for_each_domain(cpu, sd) {
3213 if (!(sd->flags & SD_LOAD_BALANCE))
3216 interval = sd->balance_interval;
3217 if (idle != CPU_IDLE)
3218 interval *= sd->busy_factor;
3220 /* scale ms to jiffies */
3221 interval = msecs_to_jiffies(interval);
3222 if (unlikely(!interval))
3224 if (interval > HZ*NR_CPUS/10)
3225 interval = HZ*NR_CPUS/10;
3228 if (sd->flags & SD_SERIALIZE) {
3229 if (!spin_trylock(&balancing))
3233 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3234 if (load_balance(cpu, rq, sd, idle, &balance)) {
3236 * We've pulled tasks over so either we're no
3237 * longer idle, or one of our SMT siblings is
3240 idle = CPU_NOT_IDLE;
3242 sd->last_balance = jiffies;
3244 if (sd->flags & SD_SERIALIZE)
3245 spin_unlock(&balancing);
3247 if (time_after(next_balance, sd->last_balance + interval)) {
3248 next_balance = sd->last_balance + interval;
3249 update_next_balance = 1;
3253 * Stop the load balance at this level. There is another
3254 * CPU in our sched group which is doing load balancing more
3262 * next_balance will be updated only when there is a need.
3263 * When the cpu is attached to null domain for ex, it will not be
3266 if (likely(update_next_balance))
3267 rq->next_balance = next_balance;
3271 * run_rebalance_domains is triggered when needed from the scheduler tick.
3272 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3273 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3275 static void run_rebalance_domains(struct softirq_action *h)
3277 int this_cpu = smp_processor_id();
3278 struct rq *this_rq = cpu_rq(this_cpu);
3279 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3280 CPU_IDLE : CPU_NOT_IDLE;
3282 rebalance_domains(this_cpu, idle);
3286 * If this cpu is the owner for idle load balancing, then do the
3287 * balancing on behalf of the other idle cpus whose ticks are
3290 if (this_rq->idle_at_tick &&
3291 atomic_read(&nohz.load_balancer) == this_cpu) {
3292 cpumask_t cpus = nohz.cpu_mask;
3296 cpu_clear(this_cpu, cpus);
3297 for_each_cpu_mask(balance_cpu, cpus) {
3299 * If this cpu gets work to do, stop the load balancing
3300 * work being done for other cpus. Next load
3301 * balancing owner will pick it up.
3306 rebalance_domains(balance_cpu, CPU_IDLE);
3308 rq = cpu_rq(balance_cpu);
3309 if (time_after(this_rq->next_balance, rq->next_balance))
3310 this_rq->next_balance = rq->next_balance;
3317 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3319 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3320 * idle load balancing owner or decide to stop the periodic load balancing,
3321 * if the whole system is idle.
3323 static inline void trigger_load_balance(struct rq *rq, int cpu)
3327 * If we were in the nohz mode recently and busy at the current
3328 * scheduler tick, then check if we need to nominate new idle
3331 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3332 rq->in_nohz_recently = 0;
3334 if (atomic_read(&nohz.load_balancer) == cpu) {
3335 cpu_clear(cpu, nohz.cpu_mask);
3336 atomic_set(&nohz.load_balancer, -1);
3339 if (atomic_read(&nohz.load_balancer) == -1) {
3341 * simple selection for now: Nominate the
3342 * first cpu in the nohz list to be the next
3345 * TBD: Traverse the sched domains and nominate
3346 * the nearest cpu in the nohz.cpu_mask.
3348 int ilb = first_cpu(nohz.cpu_mask);
3356 * If this cpu is idle and doing idle load balancing for all the
3357 * cpus with ticks stopped, is it time for that to stop?
3359 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3360 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3366 * If this cpu is idle and the idle load balancing is done by
3367 * someone else, then no need raise the SCHED_SOFTIRQ
3369 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3370 cpu_isset(cpu, nohz.cpu_mask))
3373 if (time_after_eq(jiffies, rq->next_balance))
3374 raise_softirq(SCHED_SOFTIRQ);
3377 #else /* CONFIG_SMP */
3380 * on UP we do not need to balance between CPUs:
3382 static inline void idle_balance(int cpu, struct rq *rq)
3388 DEFINE_PER_CPU(struct kernel_stat, kstat);
3390 EXPORT_PER_CPU_SYMBOL(kstat);
3393 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3394 * that have not yet been banked in case the task is currently running.
3396 unsigned long long task_sched_runtime(struct task_struct *p)
3398 unsigned long flags;
3402 rq = task_rq_lock(p, &flags);
3403 ns = p->se.sum_exec_runtime;
3404 if (task_current(rq, p)) {
3405 update_rq_clock(rq);
3406 delta_exec = rq->clock - p->se.exec_start;
3407 if ((s64)delta_exec > 0)
3410 task_rq_unlock(rq, &flags);
3416 * Account user cpu time to a process.
3417 * @p: the process that the cpu time gets accounted to
3418 * @cputime: the cpu time spent in user space since the last update
3420 void account_user_time(struct task_struct *p, cputime_t cputime)
3422 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3425 p->utime = cputime_add(p->utime, cputime);
3427 /* Add user time to cpustat. */
3428 tmp = cputime_to_cputime64(cputime);
3429 if (TASK_NICE(p) > 0)
3430 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3432 cpustat->user = cputime64_add(cpustat->user, tmp);
3436 * Account guest cpu time to a process.
3437 * @p: the process that the cpu time gets accounted to
3438 * @cputime: the cpu time spent in virtual machine since the last update
3440 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3443 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3445 tmp = cputime_to_cputime64(cputime);
3447 p->utime = cputime_add(p->utime, cputime);
3448 p->gtime = cputime_add(p->gtime, cputime);
3450 cpustat->user = cputime64_add(cpustat->user, tmp);
3451 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3455 * Account scaled user cpu time to a process.
3456 * @p: the process that the cpu time gets accounted to
3457 * @cputime: the cpu time spent in user space since the last update
3459 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3461 p->utimescaled = cputime_add(p->utimescaled, cputime);
3465 * Account system cpu time to a process.
3466 * @p: the process that the cpu time gets accounted to
3467 * @hardirq_offset: the offset to subtract from hardirq_count()
3468 * @cputime: the cpu time spent in kernel space since the last update
3470 void account_system_time(struct task_struct *p, int hardirq_offset,
3473 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3474 struct rq *rq = this_rq();
3477 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3478 return account_guest_time(p, cputime);
3480 p->stime = cputime_add(p->stime, cputime);
3482 /* Add system time to cpustat. */
3483 tmp = cputime_to_cputime64(cputime);
3484 if (hardirq_count() - hardirq_offset)
3485 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3486 else if (softirq_count())
3487 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3488 else if (p != rq->idle)
3489 cpustat->system = cputime64_add(cpustat->system, tmp);
3490 else if (atomic_read(&rq->nr_iowait) > 0)
3491 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3493 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3494 /* Account for system time used */
3495 acct_update_integrals(p);
3499 * Account scaled system cpu time to a process.
3500 * @p: the process that the cpu time gets accounted to
3501 * @hardirq_offset: the offset to subtract from hardirq_count()
3502 * @cputime: the cpu time spent in kernel space since the last update
3504 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3506 p->stimescaled = cputime_add(p->stimescaled, cputime);
3510 * Account for involuntary wait time.
3511 * @p: the process from which the cpu time has been stolen
3512 * @steal: the cpu time spent in involuntary wait
3514 void account_steal_time(struct task_struct *p, cputime_t steal)
3516 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3517 cputime64_t tmp = cputime_to_cputime64(steal);
3518 struct rq *rq = this_rq();
3520 if (p == rq->idle) {
3521 p->stime = cputime_add(p->stime, steal);
3522 if (atomic_read(&rq->nr_iowait) > 0)
3523 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3525 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3527 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3531 * This function gets called by the timer code, with HZ frequency.
3532 * We call it with interrupts disabled.
3534 * It also gets called by the fork code, when changing the parent's
3537 void scheduler_tick(void)
3539 int cpu = smp_processor_id();
3540 struct rq *rq = cpu_rq(cpu);
3541 struct task_struct *curr = rq->curr;
3542 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3544 spin_lock(&rq->lock);
3545 __update_rq_clock(rq);
3547 * Let rq->clock advance by at least TICK_NSEC:
3549 if (unlikely(rq->clock < next_tick))
3550 rq->clock = next_tick;
3551 rq->tick_timestamp = rq->clock;
3552 update_cpu_load(rq);
3553 if (curr != rq->idle) /* FIXME: needed? */
3554 curr->sched_class->task_tick(rq, curr);
3555 spin_unlock(&rq->lock);
3558 rq->idle_at_tick = idle_cpu(cpu);
3559 trigger_load_balance(rq, cpu);
3563 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3565 void fastcall add_preempt_count(int val)
3570 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3572 preempt_count() += val;
3574 * Spinlock count overflowing soon?
3576 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3579 EXPORT_SYMBOL(add_preempt_count);
3581 void fastcall sub_preempt_count(int val)
3586 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3589 * Is the spinlock portion underflowing?
3591 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3592 !(preempt_count() & PREEMPT_MASK)))
3595 preempt_count() -= val;
3597 EXPORT_SYMBOL(sub_preempt_count);
3602 * Print scheduling while atomic bug:
3604 static noinline void __schedule_bug(struct task_struct *prev)
3606 struct pt_regs *regs = get_irq_regs();
3608 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3609 prev->comm, prev->pid, preempt_count());
3611 debug_show_held_locks(prev);
3612 if (irqs_disabled())
3613 print_irqtrace_events(prev);
3622 * Various schedule()-time debugging checks and statistics:
3624 static inline void schedule_debug(struct task_struct *prev)
3627 * Test if we are atomic. Since do_exit() needs to call into
3628 * schedule() atomically, we ignore that path for now.
3629 * Otherwise, whine if we are scheduling when we should not be.
3631 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3632 __schedule_bug(prev);
3634 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3636 schedstat_inc(this_rq(), sched_count);
3637 #ifdef CONFIG_SCHEDSTATS
3638 if (unlikely(prev->lock_depth >= 0)) {
3639 schedstat_inc(this_rq(), bkl_count);
3640 schedstat_inc(prev, sched_info.bkl_count);
3646 * Pick up the highest-prio task:
3648 static inline struct task_struct *
3649 pick_next_task(struct rq *rq, struct task_struct *prev)
3651 const struct sched_class *class;
3652 struct task_struct *p;
3655 * Optimization: we know that if all tasks are in
3656 * the fair class we can call that function directly:
3658 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3659 p = fair_sched_class.pick_next_task(rq);
3664 class = sched_class_highest;
3666 p = class->pick_next_task(rq);
3670 * Will never be NULL as the idle class always
3671 * returns a non-NULL p:
3673 class = class->next;
3678 * schedule() is the main scheduler function.
3680 asmlinkage void __sched schedule(void)
3682 struct task_struct *prev, *next;
3689 cpu = smp_processor_id();
3693 switch_count = &prev->nivcsw;
3695 release_kernel_lock(prev);
3696 need_resched_nonpreemptible:
3698 schedule_debug(prev);
3701 * Do the rq-clock update outside the rq lock:
3703 local_irq_disable();
3704 __update_rq_clock(rq);
3705 spin_lock(&rq->lock);
3706 clear_tsk_need_resched(prev);
3708 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3709 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3710 unlikely(signal_pending(prev)))) {
3711 prev->state = TASK_RUNNING;
3713 deactivate_task(rq, prev, 1);
3715 switch_count = &prev->nvcsw;
3718 if (unlikely(!rq->nr_running))
3719 idle_balance(cpu, rq);
3721 prev->sched_class->put_prev_task(rq, prev);
3722 next = pick_next_task(rq, prev);
3724 sched_info_switch(prev, next);
3726 if (likely(prev != next)) {
3731 context_switch(rq, prev, next); /* unlocks the rq */
3733 spin_unlock_irq(&rq->lock);
3735 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3736 cpu = smp_processor_id();
3738 goto need_resched_nonpreemptible;
3740 preempt_enable_no_resched();
3741 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3744 EXPORT_SYMBOL(schedule);
3746 #ifdef CONFIG_PREEMPT
3748 * this is the entry point to schedule() from in-kernel preemption
3749 * off of preempt_enable. Kernel preemptions off return from interrupt
3750 * occur there and call schedule directly.
3752 asmlinkage void __sched preempt_schedule(void)
3754 struct thread_info *ti = current_thread_info();
3755 #ifdef CONFIG_PREEMPT_BKL
3756 struct task_struct *task = current;
3757 int saved_lock_depth;
3760 * If there is a non-zero preempt_count or interrupts are disabled,
3761 * we do not want to preempt the current task. Just return..
3763 if (likely(ti->preempt_count || irqs_disabled()))
3767 add_preempt_count(PREEMPT_ACTIVE);
3770 * We keep the big kernel semaphore locked, but we
3771 * clear ->lock_depth so that schedule() doesnt
3772 * auto-release the semaphore:
3774 #ifdef CONFIG_PREEMPT_BKL
3775 saved_lock_depth = task->lock_depth;
3776 task->lock_depth = -1;
3779 #ifdef CONFIG_PREEMPT_BKL
3780 task->lock_depth = saved_lock_depth;
3782 sub_preempt_count(PREEMPT_ACTIVE);
3785 * Check again in case we missed a preemption opportunity
3786 * between schedule and now.
3789 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3791 EXPORT_SYMBOL(preempt_schedule);
3794 * this is the entry point to schedule() from kernel preemption
3795 * off of irq context.
3796 * Note, that this is called and return with irqs disabled. This will
3797 * protect us against recursive calling from irq.
3799 asmlinkage void __sched preempt_schedule_irq(void)
3801 struct thread_info *ti = current_thread_info();
3802 #ifdef CONFIG_PREEMPT_BKL
3803 struct task_struct *task = current;
3804 int saved_lock_depth;
3806 /* Catch callers which need to be fixed */
3807 BUG_ON(ti->preempt_count || !irqs_disabled());
3810 add_preempt_count(PREEMPT_ACTIVE);
3813 * We keep the big kernel semaphore locked, but we
3814 * clear ->lock_depth so that schedule() doesnt
3815 * auto-release the semaphore:
3817 #ifdef CONFIG_PREEMPT_BKL
3818 saved_lock_depth = task->lock_depth;
3819 task->lock_depth = -1;
3823 local_irq_disable();
3824 #ifdef CONFIG_PREEMPT_BKL
3825 task->lock_depth = saved_lock_depth;
3827 sub_preempt_count(PREEMPT_ACTIVE);
3830 * Check again in case we missed a preemption opportunity
3831 * between schedule and now.
3834 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3837 #endif /* CONFIG_PREEMPT */
3839 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3842 return try_to_wake_up(curr->private, mode, sync);
3844 EXPORT_SYMBOL(default_wake_function);
3847 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3848 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3849 * number) then we wake all the non-exclusive tasks and one exclusive task.
3851 * There are circumstances in which we can try to wake a task which has already
3852 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3853 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3855 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3856 int nr_exclusive, int sync, void *key)
3858 wait_queue_t *curr, *next;
3860 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3861 unsigned flags = curr->flags;
3863 if (curr->func(curr, mode, sync, key) &&
3864 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3870 * __wake_up - wake up threads blocked on a waitqueue.
3872 * @mode: which threads
3873 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3874 * @key: is directly passed to the wakeup function
3876 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3877 int nr_exclusive, void *key)
3879 unsigned long flags;
3881 spin_lock_irqsave(&q->lock, flags);
3882 __wake_up_common(q, mode, nr_exclusive, 0, key);
3883 spin_unlock_irqrestore(&q->lock, flags);
3885 EXPORT_SYMBOL(__wake_up);
3888 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3890 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3892 __wake_up_common(q, mode, 1, 0, NULL);
3896 * __wake_up_sync - wake up threads blocked on a waitqueue.
3898 * @mode: which threads
3899 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3901 * The sync wakeup differs that the waker knows that it will schedule
3902 * away soon, so while the target thread will be woken up, it will not
3903 * be migrated to another CPU - ie. the two threads are 'synchronized'
3904 * with each other. This can prevent needless bouncing between CPUs.
3906 * On UP it can prevent extra preemption.
3909 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3911 unsigned long flags;
3917 if (unlikely(!nr_exclusive))
3920 spin_lock_irqsave(&q->lock, flags);
3921 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3922 spin_unlock_irqrestore(&q->lock, flags);
3924 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3926 void complete(struct completion *x)
3928 unsigned long flags;
3930 spin_lock_irqsave(&x->wait.lock, flags);
3932 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3934 spin_unlock_irqrestore(&x->wait.lock, flags);
3936 EXPORT_SYMBOL(complete);
3938 void complete_all(struct completion *x)
3940 unsigned long flags;
3942 spin_lock_irqsave(&x->wait.lock, flags);
3943 x->done += UINT_MAX/2;
3944 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3946 spin_unlock_irqrestore(&x->wait.lock, flags);
3948 EXPORT_SYMBOL(complete_all);
3950 static inline long __sched
3951 do_wait_for_common(struct completion *x, long timeout, int state)
3954 DECLARE_WAITQUEUE(wait, current);
3956 wait.flags |= WQ_FLAG_EXCLUSIVE;
3957 __add_wait_queue_tail(&x->wait, &wait);
3959 if (state == TASK_INTERRUPTIBLE &&
3960 signal_pending(current)) {
3961 __remove_wait_queue(&x->wait, &wait);
3962 return -ERESTARTSYS;
3964 __set_current_state(state);
3965 spin_unlock_irq(&x->wait.lock);
3966 timeout = schedule_timeout(timeout);
3967 spin_lock_irq(&x->wait.lock);
3969 __remove_wait_queue(&x->wait, &wait);
3973 __remove_wait_queue(&x->wait, &wait);
3980 wait_for_common(struct completion *x, long timeout, int state)
3984 spin_lock_irq(&x->wait.lock);
3985 timeout = do_wait_for_common(x, timeout, state);
3986 spin_unlock_irq(&x->wait.lock);
3990 void __sched wait_for_completion(struct completion *x)
3992 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3994 EXPORT_SYMBOL(wait_for_completion);
3996 unsigned long __sched
3997 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3999 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4001 EXPORT_SYMBOL(wait_for_completion_timeout);
4003 int __sched wait_for_completion_interruptible(struct completion *x)
4005 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4006 if (t == -ERESTARTSYS)
4010 EXPORT_SYMBOL(wait_for_completion_interruptible);
4012 unsigned long __sched
4013 wait_for_completion_interruptible_timeout(struct completion *x,
4014 unsigned long timeout)
4016 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4018 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4021 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4023 unsigned long flags;
4026 init_waitqueue_entry(&wait, current);
4028 __set_current_state(state);
4030 spin_lock_irqsave(&q->lock, flags);
4031 __add_wait_queue(q, &wait);
4032 spin_unlock(&q->lock);
4033 timeout = schedule_timeout(timeout);
4034 spin_lock_irq(&q->lock);
4035 __remove_wait_queue(q, &wait);
4036 spin_unlock_irqrestore(&q->lock, flags);
4041 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4043 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4045 EXPORT_SYMBOL(interruptible_sleep_on);
4048 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4050 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4052 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4054 void __sched sleep_on(wait_queue_head_t *q)
4056 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4058 EXPORT_SYMBOL(sleep_on);
4060 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4062 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4064 EXPORT_SYMBOL(sleep_on_timeout);
4066 #ifdef CONFIG_RT_MUTEXES
4069 * rt_mutex_setprio - set the current priority of a task
4071 * @prio: prio value (kernel-internal form)
4073 * This function changes the 'effective' priority of a task. It does
4074 * not touch ->normal_prio like __setscheduler().
4076 * Used by the rt_mutex code to implement priority inheritance logic.
4078 void rt_mutex_setprio(struct task_struct *p, int prio)
4080 unsigned long flags;
4081 int oldprio, on_rq, running;
4084 BUG_ON(prio < 0 || prio > MAX_PRIO);
4086 rq = task_rq_lock(p, &flags);
4087 update_rq_clock(rq);
4090 on_rq = p->se.on_rq;
4091 running = task_current(rq, p);
4093 dequeue_task(rq, p, 0);
4095 p->sched_class->put_prev_task(rq, p);
4099 p->sched_class = &rt_sched_class;
4101 p->sched_class = &fair_sched_class;
4107 p->sched_class->set_curr_task(rq);
4108 enqueue_task(rq, p, 0);
4110 * Reschedule if we are currently running on this runqueue and
4111 * our priority decreased, or if we are not currently running on
4112 * this runqueue and our priority is higher than the current's
4115 if (p->prio > oldprio)
4116 resched_task(rq->curr);
4118 check_preempt_curr(rq, p);
4121 task_rq_unlock(rq, &flags);
4126 void set_user_nice(struct task_struct *p, long nice)
4128 int old_prio, delta, on_rq;
4129 unsigned long flags;
4132 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4135 * We have to be careful, if called from sys_setpriority(),
4136 * the task might be in the middle of scheduling on another CPU.
4138 rq = task_rq_lock(p, &flags);
4139 update_rq_clock(rq);
4141 * The RT priorities are set via sched_setscheduler(), but we still
4142 * allow the 'normal' nice value to be set - but as expected
4143 * it wont have any effect on scheduling until the task is
4144 * SCHED_FIFO/SCHED_RR:
4146 if (task_has_rt_policy(p)) {
4147 p->static_prio = NICE_TO_PRIO(nice);
4150 on_rq = p->se.on_rq;
4152 dequeue_task(rq, p, 0);
4154 p->static_prio = NICE_TO_PRIO(nice);
4157 p->prio = effective_prio(p);
4158 delta = p->prio - old_prio;
4161 enqueue_task(rq, p, 0);
4163 * If the task increased its priority or is running and
4164 * lowered its priority, then reschedule its CPU:
4166 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4167 resched_task(rq->curr);
4170 task_rq_unlock(rq, &flags);
4172 EXPORT_SYMBOL(set_user_nice);
4175 * can_nice - check if a task can reduce its nice value
4179 int can_nice(const struct task_struct *p, const int nice)
4181 /* convert nice value [19,-20] to rlimit style value [1,40] */
4182 int nice_rlim = 20 - nice;
4184 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4185 capable(CAP_SYS_NICE));
4188 #ifdef __ARCH_WANT_SYS_NICE
4191 * sys_nice - change the priority of the current process.
4192 * @increment: priority increment
4194 * sys_setpriority is a more generic, but much slower function that
4195 * does similar things.
4197 asmlinkage long sys_nice(int increment)
4202 * Setpriority might change our priority at the same moment.
4203 * We don't have to worry. Conceptually one call occurs first
4204 * and we have a single winner.
4206 if (increment < -40)
4211 nice = PRIO_TO_NICE(current->static_prio) + increment;
4217 if (increment < 0 && !can_nice(current, nice))
4220 retval = security_task_setnice(current, nice);
4224 set_user_nice(current, nice);
4231 * task_prio - return the priority value of a given task.
4232 * @p: the task in question.
4234 * This is the priority value as seen by users in /proc.
4235 * RT tasks are offset by -200. Normal tasks are centered
4236 * around 0, value goes from -16 to +15.
4238 int task_prio(const struct task_struct *p)
4240 return p->prio - MAX_RT_PRIO;
4244 * task_nice - return the nice value of a given task.
4245 * @p: the task in question.
4247 int task_nice(const struct task_struct *p)
4249 return TASK_NICE(p);
4251 EXPORT_SYMBOL_GPL(task_nice);
4254 * idle_cpu - is a given cpu idle currently?
4255 * @cpu: the processor in question.
4257 int idle_cpu(int cpu)
4259 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4263 * idle_task - return the idle task for a given cpu.
4264 * @cpu: the processor in question.
4266 struct task_struct *idle_task(int cpu)
4268 return cpu_rq(cpu)->idle;
4272 * find_process_by_pid - find a process with a matching PID value.
4273 * @pid: the pid in question.
4275 static struct task_struct *find_process_by_pid(pid_t pid)
4277 return pid ? find_task_by_vpid(pid) : current;
4280 /* Actually do priority change: must hold rq lock. */
4282 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4284 BUG_ON(p->se.on_rq);
4287 switch (p->policy) {
4291 p->sched_class = &fair_sched_class;
4295 p->sched_class = &rt_sched_class;
4299 p->rt_priority = prio;
4300 p->normal_prio = normal_prio(p);
4301 /* we are holding p->pi_lock already */
4302 p->prio = rt_mutex_getprio(p);
4307 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4308 * @p: the task in question.
4309 * @policy: new policy.
4310 * @param: structure containing the new RT priority.
4312 * NOTE that the task may be already dead.
4314 int sched_setscheduler(struct task_struct *p, int policy,
4315 struct sched_param *param)
4317 int retval, oldprio, oldpolicy = -1, on_rq, running;
4318 unsigned long flags;
4321 /* may grab non-irq protected spin_locks */
4322 BUG_ON(in_interrupt());
4324 /* double check policy once rq lock held */
4326 policy = oldpolicy = p->policy;
4327 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4328 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4329 policy != SCHED_IDLE)
4332 * Valid priorities for SCHED_FIFO and SCHED_RR are
4333 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4334 * SCHED_BATCH and SCHED_IDLE is 0.
4336 if (param->sched_priority < 0 ||
4337 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4338 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4340 if (rt_policy(policy) != (param->sched_priority != 0))
4344 * Allow unprivileged RT tasks to decrease priority:
4346 if (!capable(CAP_SYS_NICE)) {
4347 if (rt_policy(policy)) {
4348 unsigned long rlim_rtprio;
4350 if (!lock_task_sighand(p, &flags))
4352 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4353 unlock_task_sighand(p, &flags);
4355 /* can't set/change the rt policy */
4356 if (policy != p->policy && !rlim_rtprio)
4359 /* can't increase priority */
4360 if (param->sched_priority > p->rt_priority &&
4361 param->sched_priority > rlim_rtprio)
4365 * Like positive nice levels, dont allow tasks to
4366 * move out of SCHED_IDLE either:
4368 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4371 /* can't change other user's priorities */
4372 if ((current->euid != p->euid) &&
4373 (current->euid != p->uid))
4377 retval = security_task_setscheduler(p, policy, param);
4381 * make sure no PI-waiters arrive (or leave) while we are
4382 * changing the priority of the task:
4384 spin_lock_irqsave(&p->pi_lock, flags);
4386 * To be able to change p->policy safely, the apropriate
4387 * runqueue lock must be held.
4389 rq = __task_rq_lock(p);
4390 /* recheck policy now with rq lock held */
4391 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4392 policy = oldpolicy = -1;
4393 __task_rq_unlock(rq);
4394 spin_unlock_irqrestore(&p->pi_lock, flags);
4397 update_rq_clock(rq);
4398 on_rq = p->se.on_rq;
4399 running = task_current(rq, p);
4401 deactivate_task(rq, p, 0);
4403 p->sched_class->put_prev_task(rq, p);
4407 __setscheduler(rq, p, policy, param->sched_priority);
4411 p->sched_class->set_curr_task(rq);
4412 activate_task(rq, p, 0);
4414 * Reschedule if we are currently running on this runqueue and
4415 * our priority decreased, or if we are not currently running on
4416 * this runqueue and our priority is higher than the current's
4419 if (p->prio > oldprio)
4420 resched_task(rq->curr);
4422 check_preempt_curr(rq, p);
4425 __task_rq_unlock(rq);
4426 spin_unlock_irqrestore(&p->pi_lock, flags);
4428 rt_mutex_adjust_pi(p);
4432 EXPORT_SYMBOL_GPL(sched_setscheduler);
4435 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4437 struct sched_param lparam;
4438 struct task_struct *p;
4441 if (!param || pid < 0)
4443 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4448 p = find_process_by_pid(pid);
4450 retval = sched_setscheduler(p, policy, &lparam);
4457 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4458 * @pid: the pid in question.
4459 * @policy: new policy.
4460 * @param: structure containing the new RT priority.
4463 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4465 /* negative values for policy are not valid */
4469 return do_sched_setscheduler(pid, policy, param);
4473 * sys_sched_setparam - set/change the RT priority of a thread
4474 * @pid: the pid in question.
4475 * @param: structure containing the new RT priority.
4477 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4479 return do_sched_setscheduler(pid, -1, param);
4483 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4484 * @pid: the pid in question.
4486 asmlinkage long sys_sched_getscheduler(pid_t pid)
4488 struct task_struct *p;
4495 read_lock(&tasklist_lock);
4496 p = find_process_by_pid(pid);
4498 retval = security_task_getscheduler(p);
4502 read_unlock(&tasklist_lock);
4507 * sys_sched_getscheduler - get the RT priority of a thread
4508 * @pid: the pid in question.
4509 * @param: structure containing the RT priority.
4511 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4513 struct sched_param lp;
4514 struct task_struct *p;
4517 if (!param || pid < 0)
4520 read_lock(&tasklist_lock);
4521 p = find_process_by_pid(pid);
4526 retval = security_task_getscheduler(p);
4530 lp.sched_priority = p->rt_priority;
4531 read_unlock(&tasklist_lock);
4534 * This one might sleep, we cannot do it with a spinlock held ...
4536 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4541 read_unlock(&tasklist_lock);
4545 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4547 cpumask_t cpus_allowed;
4548 struct task_struct *p;
4552 read_lock(&tasklist_lock);
4554 p = find_process_by_pid(pid);
4556 read_unlock(&tasklist_lock);
4562 * It is not safe to call set_cpus_allowed with the
4563 * tasklist_lock held. We will bump the task_struct's
4564 * usage count and then drop tasklist_lock.
4567 read_unlock(&tasklist_lock);
4570 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4571 !capable(CAP_SYS_NICE))
4574 retval = security_task_setscheduler(p, 0, NULL);
4578 cpus_allowed = cpuset_cpus_allowed(p);
4579 cpus_and(new_mask, new_mask, cpus_allowed);
4581 retval = set_cpus_allowed(p, new_mask);
4584 cpus_allowed = cpuset_cpus_allowed(p);
4585 if (!cpus_subset(new_mask, cpus_allowed)) {
4587 * We must have raced with a concurrent cpuset
4588 * update. Just reset the cpus_allowed to the
4589 * cpuset's cpus_allowed
4591 new_mask = cpus_allowed;
4601 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4602 cpumask_t *new_mask)
4604 if (len < sizeof(cpumask_t)) {
4605 memset(new_mask, 0, sizeof(cpumask_t));
4606 } else if (len > sizeof(cpumask_t)) {
4607 len = sizeof(cpumask_t);
4609 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4613 * sys_sched_setaffinity - set 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 the new cpu mask
4618 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4619 unsigned long __user *user_mask_ptr)
4624 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4628 return sched_setaffinity(pid, new_mask);
4632 * Represents all cpu's present in the system
4633 * In systems capable of hotplug, this map could dynamically grow
4634 * as new cpu's are detected in the system via any platform specific
4635 * method, such as ACPI for e.g.
4638 cpumask_t cpu_present_map __read_mostly;
4639 EXPORT_SYMBOL(cpu_present_map);
4642 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4643 EXPORT_SYMBOL(cpu_online_map);
4645 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4646 EXPORT_SYMBOL(cpu_possible_map);
4649 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4651 struct task_struct *p;
4655 read_lock(&tasklist_lock);
4658 p = find_process_by_pid(pid);
4662 retval = security_task_getscheduler(p);
4666 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4669 read_unlock(&tasklist_lock);
4676 * sys_sched_getaffinity - get the cpu affinity of a process
4677 * @pid: pid of the process
4678 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4679 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4681 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4682 unsigned long __user *user_mask_ptr)
4687 if (len < sizeof(cpumask_t))
4690 ret = sched_getaffinity(pid, &mask);
4694 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4697 return sizeof(cpumask_t);
4701 * sys_sched_yield - yield the current processor to other threads.
4703 * This function yields the current CPU to other tasks. If there are no
4704 * other threads running on this CPU then this function will return.
4706 asmlinkage long sys_sched_yield(void)
4708 struct rq *rq = this_rq_lock();
4710 schedstat_inc(rq, yld_count);
4711 current->sched_class->yield_task(rq);
4714 * Since we are going to call schedule() anyway, there's
4715 * no need to preempt or enable interrupts:
4717 __release(rq->lock);
4718 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4719 _raw_spin_unlock(&rq->lock);
4720 preempt_enable_no_resched();
4727 static void __cond_resched(void)
4729 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4730 __might_sleep(__FILE__, __LINE__);
4733 * The BKS might be reacquired before we have dropped
4734 * PREEMPT_ACTIVE, which could trigger a second
4735 * cond_resched() call.
4738 add_preempt_count(PREEMPT_ACTIVE);
4740 sub_preempt_count(PREEMPT_ACTIVE);
4741 } while (need_resched());
4744 int __sched cond_resched(void)
4746 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4747 system_state == SYSTEM_RUNNING) {
4753 EXPORT_SYMBOL(cond_resched);
4756 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4757 * call schedule, and on return reacquire the lock.
4759 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4760 * operations here to prevent schedule() from being called twice (once via
4761 * spin_unlock(), once by hand).
4763 int cond_resched_lock(spinlock_t *lock)
4767 if (need_lockbreak(lock)) {
4773 if (need_resched() && system_state == SYSTEM_RUNNING) {
4774 spin_release(&lock->dep_map, 1, _THIS_IP_);
4775 _raw_spin_unlock(lock);
4776 preempt_enable_no_resched();
4783 EXPORT_SYMBOL(cond_resched_lock);
4785 int __sched cond_resched_softirq(void)
4787 BUG_ON(!in_softirq());
4789 if (need_resched() && system_state == SYSTEM_RUNNING) {
4797 EXPORT_SYMBOL(cond_resched_softirq);
4800 * yield - yield the current processor to other threads.
4802 * This is a shortcut for kernel-space yielding - it marks the
4803 * thread runnable and calls sys_sched_yield().
4805 void __sched yield(void)
4807 set_current_state(TASK_RUNNING);
4810 EXPORT_SYMBOL(yield);
4813 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4814 * that process accounting knows that this is a task in IO wait state.
4816 * But don't do that if it is a deliberate, throttling IO wait (this task
4817 * has set its backing_dev_info: the queue against which it should throttle)
4819 void __sched io_schedule(void)
4821 struct rq *rq = &__raw_get_cpu_var(runqueues);
4823 delayacct_blkio_start();
4824 atomic_inc(&rq->nr_iowait);
4826 atomic_dec(&rq->nr_iowait);
4827 delayacct_blkio_end();
4829 EXPORT_SYMBOL(io_schedule);
4831 long __sched io_schedule_timeout(long timeout)
4833 struct rq *rq = &__raw_get_cpu_var(runqueues);
4836 delayacct_blkio_start();
4837 atomic_inc(&rq->nr_iowait);
4838 ret = schedule_timeout(timeout);
4839 atomic_dec(&rq->nr_iowait);
4840 delayacct_blkio_end();
4845 * sys_sched_get_priority_max - return maximum RT priority.
4846 * @policy: scheduling class.
4848 * this syscall returns the maximum rt_priority that can be used
4849 * by a given scheduling class.
4851 asmlinkage long sys_sched_get_priority_max(int policy)
4858 ret = MAX_USER_RT_PRIO-1;
4870 * sys_sched_get_priority_min - return minimum RT priority.
4871 * @policy: scheduling class.
4873 * this syscall returns the minimum rt_priority that can be used
4874 * by a given scheduling class.
4876 asmlinkage long sys_sched_get_priority_min(int policy)
4894 * sys_sched_rr_get_interval - return the default timeslice of a process.
4895 * @pid: pid of the process.
4896 * @interval: userspace pointer to the timeslice value.
4898 * this syscall writes the default timeslice value of a given process
4899 * into the user-space timespec buffer. A value of '0' means infinity.
4902 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4904 struct task_struct *p;
4905 unsigned int time_slice;
4913 read_lock(&tasklist_lock);
4914 p = find_process_by_pid(pid);
4918 retval = security_task_getscheduler(p);
4923 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4924 * tasks that are on an otherwise idle runqueue:
4927 if (p->policy == SCHED_RR) {
4928 time_slice = DEF_TIMESLICE;
4930 struct sched_entity *se = &p->se;
4931 unsigned long flags;
4934 rq = task_rq_lock(p, &flags);
4935 if (rq->cfs.load.weight)
4936 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4937 task_rq_unlock(rq, &flags);
4939 read_unlock(&tasklist_lock);
4940 jiffies_to_timespec(time_slice, &t);
4941 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4945 read_unlock(&tasklist_lock);
4949 static const char stat_nam[] = "RSDTtZX";
4951 void sched_show_task(struct task_struct *p)
4953 unsigned long free = 0;
4956 state = p->state ? __ffs(p->state) + 1 : 0;
4957 printk(KERN_INFO "%-13.13s %c", p->comm,
4958 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4959 #if BITS_PER_LONG == 32
4960 if (state == TASK_RUNNING)
4961 printk(KERN_CONT " running ");
4963 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4965 if (state == TASK_RUNNING)
4966 printk(KERN_CONT " running task ");
4968 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4970 #ifdef CONFIG_DEBUG_STACK_USAGE
4972 unsigned long *n = end_of_stack(p);
4975 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4978 printk(KERN_CONT "%5lu %5d %6d\n", free,
4979 task_pid_nr(p), task_pid_nr(p->real_parent));
4981 if (state != TASK_RUNNING)
4982 show_stack(p, NULL);
4985 void show_state_filter(unsigned long state_filter)
4987 struct task_struct *g, *p;
4989 #if BITS_PER_LONG == 32
4991 " task PC stack pid father\n");
4994 " task PC stack pid father\n");
4996 read_lock(&tasklist_lock);
4997 do_each_thread(g, p) {
4999 * reset the NMI-timeout, listing all files on a slow
5000 * console might take alot of time:
5002 touch_nmi_watchdog();
5003 if (!state_filter || (p->state & state_filter))
5005 } while_each_thread(g, p);
5007 touch_all_softlockup_watchdogs();
5009 #ifdef CONFIG_SCHED_DEBUG
5010 sysrq_sched_debug_show();
5012 read_unlock(&tasklist_lock);
5014 * Only show locks if all tasks are dumped:
5016 if (state_filter == -1)
5017 debug_show_all_locks();
5020 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5022 idle->sched_class = &idle_sched_class;
5026 * init_idle - set up an idle thread for a given CPU
5027 * @idle: task in question
5028 * @cpu: cpu the idle task belongs to
5030 * NOTE: this function does not set the idle thread's NEED_RESCHED
5031 * flag, to make booting more robust.
5033 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5035 struct rq *rq = cpu_rq(cpu);
5036 unsigned long flags;
5039 idle->se.exec_start = sched_clock();
5041 idle->prio = idle->normal_prio = MAX_PRIO;
5042 idle->cpus_allowed = cpumask_of_cpu(cpu);
5043 __set_task_cpu(idle, cpu);
5045 spin_lock_irqsave(&rq->lock, flags);
5046 rq->curr = rq->idle = idle;
5047 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5050 spin_unlock_irqrestore(&rq->lock, flags);
5052 /* Set the preempt count _outside_ the spinlocks! */
5053 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5054 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5056 task_thread_info(idle)->preempt_count = 0;
5059 * The idle tasks have their own, simple scheduling class:
5061 idle->sched_class = &idle_sched_class;
5065 * In a system that switches off the HZ timer nohz_cpu_mask
5066 * indicates which cpus entered this state. This is used
5067 * in the rcu update to wait only for active cpus. For system
5068 * which do not switch off the HZ timer nohz_cpu_mask should
5069 * always be CPU_MASK_NONE.
5071 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5074 * Increase the granularity value when there are more CPUs,
5075 * because with more CPUs the 'effective latency' as visible
5076 * to users decreases. But the relationship is not linear,
5077 * so pick a second-best guess by going with the log2 of the
5080 * This idea comes from the SD scheduler of Con Kolivas:
5082 static inline void sched_init_granularity(void)
5084 unsigned int factor = 1 + ilog2(num_online_cpus());
5085 const unsigned long limit = 200000000;
5087 sysctl_sched_min_granularity *= factor;
5088 if (sysctl_sched_min_granularity > limit)
5089 sysctl_sched_min_granularity = limit;
5091 sysctl_sched_latency *= factor;
5092 if (sysctl_sched_latency > limit)
5093 sysctl_sched_latency = limit;
5095 sysctl_sched_wakeup_granularity *= factor;
5096 sysctl_sched_batch_wakeup_granularity *= factor;
5101 * This is how migration works:
5103 * 1) we queue a struct migration_req structure in the source CPU's
5104 * runqueue and wake up that CPU's migration thread.
5105 * 2) we down() the locked semaphore => thread blocks.
5106 * 3) migration thread wakes up (implicitly it forces the migrated
5107 * thread off the CPU)
5108 * 4) it gets the migration request and checks whether the migrated
5109 * task is still in the wrong runqueue.
5110 * 5) if it's in the wrong runqueue then the migration thread removes
5111 * it and puts it into the right queue.
5112 * 6) migration thread up()s the semaphore.
5113 * 7) we wake up and the migration is done.
5117 * Change a given task's CPU affinity. Migrate the thread to a
5118 * proper CPU and schedule it away if the CPU it's executing on
5119 * is removed from the allowed bitmask.
5121 * NOTE: the caller must have a valid reference to the task, the
5122 * task must not exit() & deallocate itself prematurely. The
5123 * call is not atomic; no spinlocks may be held.
5125 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5127 struct migration_req req;
5128 unsigned long flags;
5132 rq = task_rq_lock(p, &flags);
5133 if (!cpus_intersects(new_mask, cpu_online_map)) {
5138 p->cpus_allowed = new_mask;
5139 /* Can the task run on the task's current CPU? If so, we're done */
5140 if (cpu_isset(task_cpu(p), new_mask))
5143 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5144 /* Need help from migration thread: drop lock and wait. */
5145 task_rq_unlock(rq, &flags);
5146 wake_up_process(rq->migration_thread);
5147 wait_for_completion(&req.done);
5148 tlb_migrate_finish(p->mm);
5152 task_rq_unlock(rq, &flags);
5156 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5159 * Move (not current) task off this cpu, onto dest cpu. We're doing
5160 * this because either it can't run here any more (set_cpus_allowed()
5161 * away from this CPU, or CPU going down), or because we're
5162 * attempting to rebalance this task on exec (sched_exec).
5164 * So we race with normal scheduler movements, but that's OK, as long
5165 * as the task is no longer on this CPU.
5167 * Returns non-zero if task was successfully migrated.
5169 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5171 struct rq *rq_dest, *rq_src;
5174 if (unlikely(cpu_is_offline(dest_cpu)))
5177 rq_src = cpu_rq(src_cpu);
5178 rq_dest = cpu_rq(dest_cpu);
5180 double_rq_lock(rq_src, rq_dest);
5181 /* Already moved. */
5182 if (task_cpu(p) != src_cpu)
5184 /* Affinity changed (again). */
5185 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5188 on_rq = p->se.on_rq;
5190 deactivate_task(rq_src, p, 0);
5192 set_task_cpu(p, dest_cpu);
5194 activate_task(rq_dest, p, 0);
5195 check_preempt_curr(rq_dest, p);
5199 double_rq_unlock(rq_src, rq_dest);
5204 * migration_thread - this is a highprio system thread that performs
5205 * thread migration by bumping thread off CPU then 'pushing' onto
5208 static int migration_thread(void *data)
5210 int cpu = (long)data;
5214 BUG_ON(rq->migration_thread != current);
5216 set_current_state(TASK_INTERRUPTIBLE);
5217 while (!kthread_should_stop()) {
5218 struct migration_req *req;
5219 struct list_head *head;
5221 spin_lock_irq(&rq->lock);
5223 if (cpu_is_offline(cpu)) {
5224 spin_unlock_irq(&rq->lock);
5228 if (rq->active_balance) {
5229 active_load_balance(rq, cpu);
5230 rq->active_balance = 0;
5233 head = &rq->migration_queue;
5235 if (list_empty(head)) {
5236 spin_unlock_irq(&rq->lock);
5238 set_current_state(TASK_INTERRUPTIBLE);
5241 req = list_entry(head->next, struct migration_req, list);
5242 list_del_init(head->next);
5244 spin_unlock(&rq->lock);
5245 __migrate_task(req->task, cpu, req->dest_cpu);
5248 complete(&req->done);
5250 __set_current_state(TASK_RUNNING);
5254 /* Wait for kthread_stop */
5255 set_current_state(TASK_INTERRUPTIBLE);
5256 while (!kthread_should_stop()) {
5258 set_current_state(TASK_INTERRUPTIBLE);
5260 __set_current_state(TASK_RUNNING);
5264 #ifdef CONFIG_HOTPLUG_CPU
5266 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5270 local_irq_disable();
5271 ret = __migrate_task(p, src_cpu, dest_cpu);
5277 * Figure out where task on dead CPU should go, use force if necessary.
5278 * NOTE: interrupts should be disabled by the caller
5280 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5282 unsigned long flags;
5289 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5290 cpus_and(mask, mask, p->cpus_allowed);
5291 dest_cpu = any_online_cpu(mask);
5293 /* On any allowed CPU? */
5294 if (dest_cpu == NR_CPUS)
5295 dest_cpu = any_online_cpu(p->cpus_allowed);
5297 /* No more Mr. Nice Guy. */
5298 if (dest_cpu == NR_CPUS) {
5299 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5301 * Try to stay on the same cpuset, where the
5302 * current cpuset may be a subset of all cpus.
5303 * The cpuset_cpus_allowed_locked() variant of
5304 * cpuset_cpus_allowed() will not block. It must be
5305 * called within calls to cpuset_lock/cpuset_unlock.
5307 rq = task_rq_lock(p, &flags);
5308 p->cpus_allowed = cpus_allowed;
5309 dest_cpu = any_online_cpu(p->cpus_allowed);
5310 task_rq_unlock(rq, &flags);
5313 * Don't tell them about moving exiting tasks or
5314 * kernel threads (both mm NULL), since they never
5317 if (p->mm && printk_ratelimit()) {
5318 printk(KERN_INFO "process %d (%s) no "
5319 "longer affine to cpu%d\n",
5320 task_pid_nr(p), p->comm, dead_cpu);
5323 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5327 * While a dead CPU has no uninterruptible tasks queued at this point,
5328 * it might still have a nonzero ->nr_uninterruptible counter, because
5329 * for performance reasons the counter is not stricly tracking tasks to
5330 * their home CPUs. So we just add the counter to another CPU's counter,
5331 * to keep the global sum constant after CPU-down:
5333 static void migrate_nr_uninterruptible(struct rq *rq_src)
5335 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5336 unsigned long flags;
5338 local_irq_save(flags);
5339 double_rq_lock(rq_src, rq_dest);
5340 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5341 rq_src->nr_uninterruptible = 0;
5342 double_rq_unlock(rq_src, rq_dest);
5343 local_irq_restore(flags);
5346 /* Run through task list and migrate tasks from the dead cpu. */
5347 static void migrate_live_tasks(int src_cpu)
5349 struct task_struct *p, *t;
5351 read_lock(&tasklist_lock);
5353 do_each_thread(t, p) {
5357 if (task_cpu(p) == src_cpu)
5358 move_task_off_dead_cpu(src_cpu, p);
5359 } while_each_thread(t, p);
5361 read_unlock(&tasklist_lock);
5365 * Schedules idle task to be the next runnable task on current CPU.
5366 * It does so by boosting its priority to highest possible.
5367 * Used by CPU offline code.
5369 void sched_idle_next(void)
5371 int this_cpu = smp_processor_id();
5372 struct rq *rq = cpu_rq(this_cpu);
5373 struct task_struct *p = rq->idle;
5374 unsigned long flags;
5376 /* cpu has to be offline */
5377 BUG_ON(cpu_online(this_cpu));
5380 * Strictly not necessary since rest of the CPUs are stopped by now
5381 * and interrupts disabled on the current cpu.
5383 spin_lock_irqsave(&rq->lock, flags);
5385 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5387 update_rq_clock(rq);
5388 activate_task(rq, p, 0);
5390 spin_unlock_irqrestore(&rq->lock, flags);
5394 * Ensures that the idle task is using init_mm right before its cpu goes
5397 void idle_task_exit(void)
5399 struct mm_struct *mm = current->active_mm;
5401 BUG_ON(cpu_online(smp_processor_id()));
5404 switch_mm(mm, &init_mm, current);
5408 /* called under rq->lock with disabled interrupts */
5409 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5411 struct rq *rq = cpu_rq(dead_cpu);
5413 /* Must be exiting, otherwise would be on tasklist. */
5414 BUG_ON(!p->exit_state);
5416 /* Cannot have done final schedule yet: would have vanished. */
5417 BUG_ON(p->state == TASK_DEAD);
5422 * Drop lock around migration; if someone else moves it,
5423 * that's OK. No task can be added to this CPU, so iteration is
5426 spin_unlock_irq(&rq->lock);
5427 move_task_off_dead_cpu(dead_cpu, p);
5428 spin_lock_irq(&rq->lock);
5433 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5434 static void migrate_dead_tasks(unsigned int dead_cpu)
5436 struct rq *rq = cpu_rq(dead_cpu);
5437 struct task_struct *next;
5440 if (!rq->nr_running)
5442 update_rq_clock(rq);
5443 next = pick_next_task(rq, rq->curr);
5446 migrate_dead(dead_cpu, next);
5450 #endif /* CONFIG_HOTPLUG_CPU */
5452 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5454 static struct ctl_table sd_ctl_dir[] = {
5456 .procname = "sched_domain",
5462 static struct ctl_table sd_ctl_root[] = {
5464 .ctl_name = CTL_KERN,
5465 .procname = "kernel",
5467 .child = sd_ctl_dir,
5472 static struct ctl_table *sd_alloc_ctl_entry(int n)
5474 struct ctl_table *entry =
5475 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5480 static void sd_free_ctl_entry(struct ctl_table **tablep)
5482 struct ctl_table *entry;
5485 * In the intermediate directories, both the child directory and
5486 * procname are dynamically allocated and could fail but the mode
5487 * will always be set. In the lowest directory the names are
5488 * static strings and all have proc handlers.
5490 for (entry = *tablep; entry->mode; entry++) {
5492 sd_free_ctl_entry(&entry->child);
5493 if (entry->proc_handler == NULL)
5494 kfree(entry->procname);
5502 set_table_entry(struct ctl_table *entry,
5503 const char *procname, void *data, int maxlen,
5504 mode_t mode, proc_handler *proc_handler)
5506 entry->procname = procname;
5508 entry->maxlen = maxlen;
5510 entry->proc_handler = proc_handler;
5513 static struct ctl_table *
5514 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5516 struct ctl_table *table = sd_alloc_ctl_entry(12);
5521 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5522 sizeof(long), 0644, proc_doulongvec_minmax);
5523 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5524 sizeof(long), 0644, proc_doulongvec_minmax);
5525 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5526 sizeof(int), 0644, proc_dointvec_minmax);
5527 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5528 sizeof(int), 0644, proc_dointvec_minmax);
5529 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5530 sizeof(int), 0644, proc_dointvec_minmax);
5531 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5532 sizeof(int), 0644, proc_dointvec_minmax);
5533 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5534 sizeof(int), 0644, proc_dointvec_minmax);
5535 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5536 sizeof(int), 0644, proc_dointvec_minmax);
5537 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5538 sizeof(int), 0644, proc_dointvec_minmax);
5539 set_table_entry(&table[9], "cache_nice_tries",
5540 &sd->cache_nice_tries,
5541 sizeof(int), 0644, proc_dointvec_minmax);
5542 set_table_entry(&table[10], "flags", &sd->flags,
5543 sizeof(int), 0644, proc_dointvec_minmax);
5544 /* &table[11] is terminator */
5549 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5551 struct ctl_table *entry, *table;
5552 struct sched_domain *sd;
5553 int domain_num = 0, i;
5556 for_each_domain(cpu, sd)
5558 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5563 for_each_domain(cpu, sd) {
5564 snprintf(buf, 32, "domain%d", i);
5565 entry->procname = kstrdup(buf, GFP_KERNEL);
5567 entry->child = sd_alloc_ctl_domain_table(sd);
5574 static struct ctl_table_header *sd_sysctl_header;
5575 static void register_sched_domain_sysctl(void)
5577 int i, cpu_num = num_online_cpus();
5578 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5581 WARN_ON(sd_ctl_dir[0].child);
5582 sd_ctl_dir[0].child = entry;
5587 for_each_online_cpu(i) {
5588 snprintf(buf, 32, "cpu%d", i);
5589 entry->procname = kstrdup(buf, GFP_KERNEL);
5591 entry->child = sd_alloc_ctl_cpu_table(i);
5595 WARN_ON(sd_sysctl_header);
5596 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5599 /* may be called multiple times per register */
5600 static void unregister_sched_domain_sysctl(void)
5602 if (sd_sysctl_header)
5603 unregister_sysctl_table(sd_sysctl_header);
5604 sd_sysctl_header = NULL;
5605 if (sd_ctl_dir[0].child)
5606 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5609 static void register_sched_domain_sysctl(void)
5612 static void unregister_sched_domain_sysctl(void)
5618 * migration_call - callback that gets triggered when a CPU is added.
5619 * Here we can start up the necessary migration thread for the new CPU.
5621 static int __cpuinit
5622 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5624 struct task_struct *p;
5625 int cpu = (long)hcpu;
5626 unsigned long flags;
5631 case CPU_UP_PREPARE:
5632 case CPU_UP_PREPARE_FROZEN:
5633 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5636 kthread_bind(p, cpu);
5637 /* Must be high prio: stop_machine expects to yield to it. */
5638 rq = task_rq_lock(p, &flags);
5639 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5640 task_rq_unlock(rq, &flags);
5641 cpu_rq(cpu)->migration_thread = p;
5645 case CPU_ONLINE_FROZEN:
5646 /* Strictly unnecessary, as first user will wake it. */
5647 wake_up_process(cpu_rq(cpu)->migration_thread);
5650 #ifdef CONFIG_HOTPLUG_CPU
5651 case CPU_UP_CANCELED:
5652 case CPU_UP_CANCELED_FROZEN:
5653 if (!cpu_rq(cpu)->migration_thread)
5655 /* Unbind it from offline cpu so it can run. Fall thru. */
5656 kthread_bind(cpu_rq(cpu)->migration_thread,
5657 any_online_cpu(cpu_online_map));
5658 kthread_stop(cpu_rq(cpu)->migration_thread);
5659 cpu_rq(cpu)->migration_thread = NULL;
5663 case CPU_DEAD_FROZEN:
5664 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5665 migrate_live_tasks(cpu);
5667 kthread_stop(rq->migration_thread);
5668 rq->migration_thread = NULL;
5669 /* Idle task back to normal (off runqueue, low prio) */
5670 spin_lock_irq(&rq->lock);
5671 update_rq_clock(rq);
5672 deactivate_task(rq, rq->idle, 0);
5673 rq->idle->static_prio = MAX_PRIO;
5674 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5675 rq->idle->sched_class = &idle_sched_class;
5676 migrate_dead_tasks(cpu);
5677 spin_unlock_irq(&rq->lock);
5679 migrate_nr_uninterruptible(rq);
5680 BUG_ON(rq->nr_running != 0);
5683 * No need to migrate the tasks: it was best-effort if
5684 * they didn't take sched_hotcpu_mutex. Just wake up
5687 spin_lock_irq(&rq->lock);
5688 while (!list_empty(&rq->migration_queue)) {
5689 struct migration_req *req;
5691 req = list_entry(rq->migration_queue.next,
5692 struct migration_req, list);
5693 list_del_init(&req->list);
5694 complete(&req->done);
5696 spin_unlock_irq(&rq->lock);
5703 /* Register at highest priority so that task migration (migrate_all_tasks)
5704 * happens before everything else.
5706 static struct notifier_block __cpuinitdata migration_notifier = {
5707 .notifier_call = migration_call,
5711 void __init migration_init(void)
5713 void *cpu = (void *)(long)smp_processor_id();
5716 /* Start one for the boot CPU: */
5717 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5718 BUG_ON(err == NOTIFY_BAD);
5719 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5720 register_cpu_notifier(&migration_notifier);
5726 /* Number of possible processor ids */
5727 int nr_cpu_ids __read_mostly = NR_CPUS;
5728 EXPORT_SYMBOL(nr_cpu_ids);
5730 #ifdef CONFIG_SCHED_DEBUG
5732 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5734 struct sched_group *group = sd->groups;
5735 cpumask_t groupmask;
5738 cpumask_scnprintf(str, NR_CPUS, sd->span);
5739 cpus_clear(groupmask);
5741 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5743 if (!(sd->flags & SD_LOAD_BALANCE)) {
5744 printk("does not load-balance\n");
5746 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5751 printk(KERN_CONT "span %s\n", str);
5753 if (!cpu_isset(cpu, sd->span)) {
5754 printk(KERN_ERR "ERROR: domain->span does not contain "
5757 if (!cpu_isset(cpu, group->cpumask)) {
5758 printk(KERN_ERR "ERROR: domain->groups does not contain"
5762 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5766 printk(KERN_ERR "ERROR: group is NULL\n");
5770 if (!group->__cpu_power) {
5771 printk(KERN_CONT "\n");
5772 printk(KERN_ERR "ERROR: domain->cpu_power not "
5777 if (!cpus_weight(group->cpumask)) {
5778 printk(KERN_CONT "\n");
5779 printk(KERN_ERR "ERROR: empty group\n");
5783 if (cpus_intersects(groupmask, group->cpumask)) {
5784 printk(KERN_CONT "\n");
5785 printk(KERN_ERR "ERROR: repeated CPUs\n");
5789 cpus_or(groupmask, groupmask, group->cpumask);
5791 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5792 printk(KERN_CONT " %s", str);
5794 group = group->next;
5795 } while (group != sd->groups);
5796 printk(KERN_CONT "\n");
5798 if (!cpus_equal(sd->span, groupmask))
5799 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5801 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5802 printk(KERN_ERR "ERROR: parent span is not a superset "
5803 "of domain->span\n");
5807 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5812 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5816 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5819 if (sched_domain_debug_one(sd, cpu, level))
5828 # define sched_domain_debug(sd, cpu) do { } while (0)
5831 static int sd_degenerate(struct sched_domain *sd)
5833 if (cpus_weight(sd->span) == 1)
5836 /* Following flags need at least 2 groups */
5837 if (sd->flags & (SD_LOAD_BALANCE |
5838 SD_BALANCE_NEWIDLE |
5842 SD_SHARE_PKG_RESOURCES)) {
5843 if (sd->groups != sd->groups->next)
5847 /* Following flags don't use groups */
5848 if (sd->flags & (SD_WAKE_IDLE |
5857 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5859 unsigned long cflags = sd->flags, pflags = parent->flags;
5861 if (sd_degenerate(parent))
5864 if (!cpus_equal(sd->span, parent->span))
5867 /* Does parent contain flags not in child? */
5868 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5869 if (cflags & SD_WAKE_AFFINE)
5870 pflags &= ~SD_WAKE_BALANCE;
5871 /* Flags needing groups don't count if only 1 group in parent */
5872 if (parent->groups == parent->groups->next) {
5873 pflags &= ~(SD_LOAD_BALANCE |
5874 SD_BALANCE_NEWIDLE |
5878 SD_SHARE_PKG_RESOURCES);
5880 if (~cflags & pflags)
5887 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5888 * hold the hotplug lock.
5890 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5892 struct rq *rq = cpu_rq(cpu);
5893 struct sched_domain *tmp;
5895 /* Remove the sched domains which do not contribute to scheduling. */
5896 for (tmp = sd; tmp; tmp = tmp->parent) {
5897 struct sched_domain *parent = tmp->parent;
5900 if (sd_parent_degenerate(tmp, parent)) {
5901 tmp->parent = parent->parent;
5903 parent->parent->child = tmp;
5907 if (sd && sd_degenerate(sd)) {
5913 sched_domain_debug(sd, cpu);
5915 rcu_assign_pointer(rq->sd, sd);
5918 /* cpus with isolated domains */
5919 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5921 /* Setup the mask of cpus configured for isolated domains */
5922 static int __init isolated_cpu_setup(char *str)
5924 int ints[NR_CPUS], i;
5926 str = get_options(str, ARRAY_SIZE(ints), ints);
5927 cpus_clear(cpu_isolated_map);
5928 for (i = 1; i <= ints[0]; i++)
5929 if (ints[i] < NR_CPUS)
5930 cpu_set(ints[i], cpu_isolated_map);
5934 __setup("isolcpus=", isolated_cpu_setup);
5937 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5938 * to a function which identifies what group(along with sched group) a CPU
5939 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5940 * (due to the fact that we keep track of groups covered with a cpumask_t).
5942 * init_sched_build_groups will build a circular linked list of the groups
5943 * covered by the given span, and will set each group's ->cpumask correctly,
5944 * and ->cpu_power to 0.
5947 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5948 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5949 struct sched_group **sg))
5951 struct sched_group *first = NULL, *last = NULL;
5952 cpumask_t covered = CPU_MASK_NONE;
5955 for_each_cpu_mask(i, span) {
5956 struct sched_group *sg;
5957 int group = group_fn(i, cpu_map, &sg);
5960 if (cpu_isset(i, covered))
5963 sg->cpumask = CPU_MASK_NONE;
5964 sg->__cpu_power = 0;
5966 for_each_cpu_mask(j, span) {
5967 if (group_fn(j, cpu_map, NULL) != group)
5970 cpu_set(j, covered);
5971 cpu_set(j, sg->cpumask);
5982 #define SD_NODES_PER_DOMAIN 16
5987 * find_next_best_node - find the next node to include in a sched_domain
5988 * @node: node whose sched_domain we're building
5989 * @used_nodes: nodes already in the sched_domain
5991 * Find the next node to include in a given scheduling domain. Simply
5992 * finds the closest node not already in the @used_nodes map.
5994 * Should use nodemask_t.
5996 static int find_next_best_node(int node, unsigned long *used_nodes)
5998 int i, n, val, min_val, best_node = 0;
6002 for (i = 0; i < MAX_NUMNODES; i++) {
6003 /* Start at @node */
6004 n = (node + i) % MAX_NUMNODES;
6006 if (!nr_cpus_node(n))
6009 /* Skip already used nodes */
6010 if (test_bit(n, used_nodes))
6013 /* Simple min distance search */
6014 val = node_distance(node, n);
6016 if (val < min_val) {
6022 set_bit(best_node, used_nodes);
6027 * sched_domain_node_span - get a cpumask for a node's sched_domain
6028 * @node: node whose cpumask we're constructing
6029 * @size: number of nodes to include in this span
6031 * Given a node, construct a good cpumask for its sched_domain to span. It
6032 * should be one that prevents unnecessary balancing, but also spreads tasks
6035 static cpumask_t sched_domain_node_span(int node)
6037 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6038 cpumask_t span, nodemask;
6042 bitmap_zero(used_nodes, MAX_NUMNODES);
6044 nodemask = node_to_cpumask(node);
6045 cpus_or(span, span, nodemask);
6046 set_bit(node, used_nodes);
6048 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6049 int next_node = find_next_best_node(node, used_nodes);
6051 nodemask = node_to_cpumask(next_node);
6052 cpus_or(span, span, nodemask);
6059 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6062 * SMT sched-domains:
6064 #ifdef CONFIG_SCHED_SMT
6065 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6066 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6069 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6072 *sg = &per_cpu(sched_group_cpus, cpu);
6078 * multi-core sched-domains:
6080 #ifdef CONFIG_SCHED_MC
6081 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6082 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6085 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6087 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6090 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6091 cpus_and(mask, mask, *cpu_map);
6092 group = first_cpu(mask);
6094 *sg = &per_cpu(sched_group_core, group);
6097 #elif defined(CONFIG_SCHED_MC)
6099 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6102 *sg = &per_cpu(sched_group_core, cpu);
6107 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6108 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6111 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6114 #ifdef CONFIG_SCHED_MC
6115 cpumask_t mask = cpu_coregroup_map(cpu);
6116 cpus_and(mask, mask, *cpu_map);
6117 group = first_cpu(mask);
6118 #elif defined(CONFIG_SCHED_SMT)
6119 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6120 cpus_and(mask, mask, *cpu_map);
6121 group = first_cpu(mask);
6126 *sg = &per_cpu(sched_group_phys, group);
6132 * The init_sched_build_groups can't handle what we want to do with node
6133 * groups, so roll our own. Now each node has its own list of groups which
6134 * gets dynamically allocated.
6136 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6137 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6139 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6140 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6142 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6143 struct sched_group **sg)
6145 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6148 cpus_and(nodemask, nodemask, *cpu_map);
6149 group = first_cpu(nodemask);
6152 *sg = &per_cpu(sched_group_allnodes, group);
6156 static void init_numa_sched_groups_power(struct sched_group *group_head)
6158 struct sched_group *sg = group_head;
6164 for_each_cpu_mask(j, sg->cpumask) {
6165 struct sched_domain *sd;
6167 sd = &per_cpu(phys_domains, j);
6168 if (j != first_cpu(sd->groups->cpumask)) {
6170 * Only add "power" once for each
6176 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6179 } while (sg != group_head);
6184 /* Free memory allocated for various sched_group structures */
6185 static void free_sched_groups(const cpumask_t *cpu_map)
6189 for_each_cpu_mask(cpu, *cpu_map) {
6190 struct sched_group **sched_group_nodes
6191 = sched_group_nodes_bycpu[cpu];
6193 if (!sched_group_nodes)
6196 for (i = 0; i < MAX_NUMNODES; i++) {
6197 cpumask_t nodemask = node_to_cpumask(i);
6198 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6200 cpus_and(nodemask, nodemask, *cpu_map);
6201 if (cpus_empty(nodemask))
6211 if (oldsg != sched_group_nodes[i])
6214 kfree(sched_group_nodes);
6215 sched_group_nodes_bycpu[cpu] = NULL;
6219 static void free_sched_groups(const cpumask_t *cpu_map)
6225 * Initialize sched groups cpu_power.
6227 * cpu_power indicates the capacity of sched group, which is used while
6228 * distributing the load between different sched groups in a sched domain.
6229 * Typically cpu_power for all the groups in a sched domain will be same unless
6230 * there are asymmetries in the topology. If there are asymmetries, group
6231 * having more cpu_power will pickup more load compared to the group having
6234 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6235 * the maximum number of tasks a group can handle in the presence of other idle
6236 * or lightly loaded groups in the same sched domain.
6238 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6240 struct sched_domain *child;
6241 struct sched_group *group;
6243 WARN_ON(!sd || !sd->groups);
6245 if (cpu != first_cpu(sd->groups->cpumask))
6250 sd->groups->__cpu_power = 0;
6253 * For perf policy, if the groups in child domain share resources
6254 * (for example cores sharing some portions of the cache hierarchy
6255 * or SMT), then set this domain groups cpu_power such that each group
6256 * can handle only one task, when there are other idle groups in the
6257 * same sched domain.
6259 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6261 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6262 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6267 * add cpu_power of each child group to this groups cpu_power
6269 group = child->groups;
6271 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6272 group = group->next;
6273 } while (group != child->groups);
6277 * Build sched domains for a given set of cpus and attach the sched domains
6278 * to the individual cpus
6280 static int build_sched_domains(const cpumask_t *cpu_map)
6284 struct sched_group **sched_group_nodes = NULL;
6285 int sd_allnodes = 0;
6288 * Allocate the per-node list of sched groups
6290 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6292 if (!sched_group_nodes) {
6293 printk(KERN_WARNING "Can not alloc sched group node list\n");
6296 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6300 * Set up domains for cpus specified by the cpu_map.
6302 for_each_cpu_mask(i, *cpu_map) {
6303 struct sched_domain *sd = NULL, *p;
6304 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6306 cpus_and(nodemask, nodemask, *cpu_map);
6309 if (cpus_weight(*cpu_map) >
6310 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6311 sd = &per_cpu(allnodes_domains, i);
6312 *sd = SD_ALLNODES_INIT;
6313 sd->span = *cpu_map;
6314 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6320 sd = &per_cpu(node_domains, i);
6322 sd->span = sched_domain_node_span(cpu_to_node(i));
6326 cpus_and(sd->span, sd->span, *cpu_map);
6330 sd = &per_cpu(phys_domains, i);
6332 sd->span = nodemask;
6336 cpu_to_phys_group(i, cpu_map, &sd->groups);
6338 #ifdef CONFIG_SCHED_MC
6340 sd = &per_cpu(core_domains, i);
6342 sd->span = cpu_coregroup_map(i);
6343 cpus_and(sd->span, sd->span, *cpu_map);
6346 cpu_to_core_group(i, cpu_map, &sd->groups);
6349 #ifdef CONFIG_SCHED_SMT
6351 sd = &per_cpu(cpu_domains, i);
6352 *sd = SD_SIBLING_INIT;
6353 sd->span = per_cpu(cpu_sibling_map, i);
6354 cpus_and(sd->span, sd->span, *cpu_map);
6357 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6361 #ifdef CONFIG_SCHED_SMT
6362 /* Set up CPU (sibling) groups */
6363 for_each_cpu_mask(i, *cpu_map) {
6364 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6365 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6366 if (i != first_cpu(this_sibling_map))
6369 init_sched_build_groups(this_sibling_map, cpu_map,
6374 #ifdef CONFIG_SCHED_MC
6375 /* Set up multi-core groups */
6376 for_each_cpu_mask(i, *cpu_map) {
6377 cpumask_t this_core_map = cpu_coregroup_map(i);
6378 cpus_and(this_core_map, this_core_map, *cpu_map);
6379 if (i != first_cpu(this_core_map))
6381 init_sched_build_groups(this_core_map, cpu_map,
6382 &cpu_to_core_group);
6386 /* Set up physical groups */
6387 for (i = 0; i < MAX_NUMNODES; i++) {
6388 cpumask_t nodemask = node_to_cpumask(i);
6390 cpus_and(nodemask, nodemask, *cpu_map);
6391 if (cpus_empty(nodemask))
6394 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6398 /* Set up node groups */
6400 init_sched_build_groups(*cpu_map, cpu_map,
6401 &cpu_to_allnodes_group);
6403 for (i = 0; i < MAX_NUMNODES; i++) {
6404 /* Set up node groups */
6405 struct sched_group *sg, *prev;
6406 cpumask_t nodemask = node_to_cpumask(i);
6407 cpumask_t domainspan;
6408 cpumask_t covered = CPU_MASK_NONE;
6411 cpus_and(nodemask, nodemask, *cpu_map);
6412 if (cpus_empty(nodemask)) {
6413 sched_group_nodes[i] = NULL;
6417 domainspan = sched_domain_node_span(i);
6418 cpus_and(domainspan, domainspan, *cpu_map);
6420 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6422 printk(KERN_WARNING "Can not alloc domain group for "
6426 sched_group_nodes[i] = sg;
6427 for_each_cpu_mask(j, nodemask) {
6428 struct sched_domain *sd;
6430 sd = &per_cpu(node_domains, j);
6433 sg->__cpu_power = 0;
6434 sg->cpumask = nodemask;
6436 cpus_or(covered, covered, nodemask);
6439 for (j = 0; j < MAX_NUMNODES; j++) {
6440 cpumask_t tmp, notcovered;
6441 int n = (i + j) % MAX_NUMNODES;
6443 cpus_complement(notcovered, covered);
6444 cpus_and(tmp, notcovered, *cpu_map);
6445 cpus_and(tmp, tmp, domainspan);
6446 if (cpus_empty(tmp))
6449 nodemask = node_to_cpumask(n);
6450 cpus_and(tmp, tmp, nodemask);
6451 if (cpus_empty(tmp))
6454 sg = kmalloc_node(sizeof(struct sched_group),
6458 "Can not alloc domain group for node %d\n", j);
6461 sg->__cpu_power = 0;
6463 sg->next = prev->next;
6464 cpus_or(covered, covered, tmp);
6471 /* Calculate CPU power for physical packages and nodes */
6472 #ifdef CONFIG_SCHED_SMT
6473 for_each_cpu_mask(i, *cpu_map) {
6474 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6476 init_sched_groups_power(i, sd);
6479 #ifdef CONFIG_SCHED_MC
6480 for_each_cpu_mask(i, *cpu_map) {
6481 struct sched_domain *sd = &per_cpu(core_domains, i);
6483 init_sched_groups_power(i, sd);
6487 for_each_cpu_mask(i, *cpu_map) {
6488 struct sched_domain *sd = &per_cpu(phys_domains, i);
6490 init_sched_groups_power(i, sd);
6494 for (i = 0; i < MAX_NUMNODES; i++)
6495 init_numa_sched_groups_power(sched_group_nodes[i]);
6498 struct sched_group *sg;
6500 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6501 init_numa_sched_groups_power(sg);
6505 /* Attach the domains */
6506 for_each_cpu_mask(i, *cpu_map) {
6507 struct sched_domain *sd;
6508 #ifdef CONFIG_SCHED_SMT
6509 sd = &per_cpu(cpu_domains, i);
6510 #elif defined(CONFIG_SCHED_MC)
6511 sd = &per_cpu(core_domains, i);
6513 sd = &per_cpu(phys_domains, i);
6515 cpu_attach_domain(sd, i);
6522 free_sched_groups(cpu_map);
6527 static cpumask_t *doms_cur; /* current sched domains */
6528 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6531 * Special case: If a kmalloc of a doms_cur partition (array of
6532 * cpumask_t) fails, then fallback to a single sched domain,
6533 * as determined by the single cpumask_t fallback_doms.
6535 static cpumask_t fallback_doms;
6538 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6539 * For now this just excludes isolated cpus, but could be used to
6540 * exclude other special cases in the future.
6542 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6547 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6549 doms_cur = &fallback_doms;
6550 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6551 err = build_sched_domains(doms_cur);
6552 register_sched_domain_sysctl();
6557 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6559 free_sched_groups(cpu_map);
6563 * Detach sched domains from a group of cpus specified in cpu_map
6564 * These cpus will now be attached to the NULL domain
6566 static void detach_destroy_domains(const cpumask_t *cpu_map)
6570 unregister_sched_domain_sysctl();
6572 for_each_cpu_mask(i, *cpu_map)
6573 cpu_attach_domain(NULL, i);
6574 synchronize_sched();
6575 arch_destroy_sched_domains(cpu_map);
6579 * Partition sched domains as specified by the 'ndoms_new'
6580 * cpumasks in the array doms_new[] of cpumasks. This compares
6581 * doms_new[] to the current sched domain partitioning, doms_cur[].
6582 * It destroys each deleted domain and builds each new domain.
6584 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6585 * The masks don't intersect (don't overlap.) We should setup one
6586 * sched domain for each mask. CPUs not in any of the cpumasks will
6587 * not be load balanced. If the same cpumask appears both in the
6588 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6591 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6592 * ownership of it and will kfree it when done with it. If the caller
6593 * failed the kmalloc call, then it can pass in doms_new == NULL,
6594 * and partition_sched_domains() will fallback to the single partition
6597 * Call with hotplug lock held
6599 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6605 /* always unregister in case we don't destroy any domains */
6606 unregister_sched_domain_sysctl();
6608 if (doms_new == NULL) {
6610 doms_new = &fallback_doms;
6611 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6614 /* Destroy deleted domains */
6615 for (i = 0; i < ndoms_cur; i++) {
6616 for (j = 0; j < ndoms_new; j++) {
6617 if (cpus_equal(doms_cur[i], doms_new[j]))
6620 /* no match - a current sched domain not in new doms_new[] */
6621 detach_destroy_domains(doms_cur + i);
6626 /* Build new domains */
6627 for (i = 0; i < ndoms_new; i++) {
6628 for (j = 0; j < ndoms_cur; j++) {
6629 if (cpus_equal(doms_new[i], doms_cur[j]))
6632 /* no match - add a new doms_new */
6633 build_sched_domains(doms_new + i);
6638 /* Remember the new sched domains */
6639 if (doms_cur != &fallback_doms)
6641 doms_cur = doms_new;
6642 ndoms_cur = ndoms_new;
6644 register_sched_domain_sysctl();
6649 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6650 static int arch_reinit_sched_domains(void)
6655 detach_destroy_domains(&cpu_online_map);
6656 err = arch_init_sched_domains(&cpu_online_map);
6662 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6666 if (buf[0] != '0' && buf[0] != '1')
6670 sched_smt_power_savings = (buf[0] == '1');
6672 sched_mc_power_savings = (buf[0] == '1');
6674 ret = arch_reinit_sched_domains();
6676 return ret ? ret : count;
6679 #ifdef CONFIG_SCHED_MC
6680 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6682 return sprintf(page, "%u\n", sched_mc_power_savings);
6684 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6685 const char *buf, size_t count)
6687 return sched_power_savings_store(buf, count, 0);
6689 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6690 sched_mc_power_savings_store);
6693 #ifdef CONFIG_SCHED_SMT
6694 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6696 return sprintf(page, "%u\n", sched_smt_power_savings);
6698 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6699 const char *buf, size_t count)
6701 return sched_power_savings_store(buf, count, 1);
6703 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6704 sched_smt_power_savings_store);
6707 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6711 #ifdef CONFIG_SCHED_SMT
6713 err = sysfs_create_file(&cls->kset.kobj,
6714 &attr_sched_smt_power_savings.attr);
6716 #ifdef CONFIG_SCHED_MC
6717 if (!err && mc_capable())
6718 err = sysfs_create_file(&cls->kset.kobj,
6719 &attr_sched_mc_power_savings.attr);
6726 * Force a reinitialization of the sched domains hierarchy. The domains
6727 * and groups cannot be updated in place without racing with the balancing
6728 * code, so we temporarily attach all running cpus to the NULL domain
6729 * which will prevent rebalancing while the sched domains are recalculated.
6731 static int update_sched_domains(struct notifier_block *nfb,
6732 unsigned long action, void *hcpu)
6735 case CPU_UP_PREPARE:
6736 case CPU_UP_PREPARE_FROZEN:
6737 case CPU_DOWN_PREPARE:
6738 case CPU_DOWN_PREPARE_FROZEN:
6739 detach_destroy_domains(&cpu_online_map);
6742 case CPU_UP_CANCELED:
6743 case CPU_UP_CANCELED_FROZEN:
6744 case CPU_DOWN_FAILED:
6745 case CPU_DOWN_FAILED_FROZEN:
6747 case CPU_ONLINE_FROZEN:
6749 case CPU_DEAD_FROZEN:
6751 * Fall through and re-initialise the domains.
6758 /* The hotplug lock is already held by cpu_up/cpu_down */
6759 arch_init_sched_domains(&cpu_online_map);
6764 void __init sched_init_smp(void)
6766 cpumask_t non_isolated_cpus;
6769 arch_init_sched_domains(&cpu_online_map);
6770 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6771 if (cpus_empty(non_isolated_cpus))
6772 cpu_set(smp_processor_id(), non_isolated_cpus);
6774 /* XXX: Theoretical race here - CPU may be hotplugged now */
6775 hotcpu_notifier(update_sched_domains, 0);
6777 /* Move init over to a non-isolated CPU */
6778 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6780 sched_init_granularity();
6782 #ifdef CONFIG_FAIR_GROUP_SCHED
6783 if (nr_cpu_ids == 1)
6786 lb_monitor_task = kthread_create(load_balance_monitor, NULL,
6788 if (!IS_ERR(lb_monitor_task)) {
6789 lb_monitor_task->flags |= PF_NOFREEZE;
6790 wake_up_process(lb_monitor_task);
6792 printk(KERN_ERR "Could not create load balance monitor thread"
6793 "(error = %ld) \n", PTR_ERR(lb_monitor_task));
6798 void __init sched_init_smp(void)
6800 sched_init_granularity();
6802 #endif /* CONFIG_SMP */
6804 int in_sched_functions(unsigned long addr)
6806 return in_lock_functions(addr) ||
6807 (addr >= (unsigned long)__sched_text_start
6808 && addr < (unsigned long)__sched_text_end);
6811 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6813 cfs_rq->tasks_timeline = RB_ROOT;
6814 #ifdef CONFIG_FAIR_GROUP_SCHED
6817 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6820 void __init sched_init(void)
6822 int highest_cpu = 0;
6825 for_each_possible_cpu(i) {
6826 struct rt_prio_array *array;
6830 spin_lock_init(&rq->lock);
6831 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6834 init_cfs_rq(&rq->cfs, rq);
6835 #ifdef CONFIG_FAIR_GROUP_SCHED
6836 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6838 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6839 struct sched_entity *se =
6840 &per_cpu(init_sched_entity, i);
6842 init_cfs_rq_p[i] = cfs_rq;
6843 init_cfs_rq(cfs_rq, rq);
6844 cfs_rq->tg = &init_task_group;
6845 list_add(&cfs_rq->leaf_cfs_rq_list,
6846 &rq->leaf_cfs_rq_list);
6848 init_sched_entity_p[i] = se;
6849 se->cfs_rq = &rq->cfs;
6851 se->load.weight = init_task_group_load;
6852 se->load.inv_weight =
6853 div64_64(1ULL<<32, init_task_group_load);
6856 init_task_group.shares = init_task_group_load;
6859 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6860 rq->cpu_load[j] = 0;
6863 rq->active_balance = 0;
6864 rq->next_balance = jiffies;
6867 rq->migration_thread = NULL;
6868 INIT_LIST_HEAD(&rq->migration_queue);
6869 rq->rt.highest_prio = MAX_RT_PRIO;
6871 atomic_set(&rq->nr_iowait, 0);
6873 array = &rq->rt.active;
6874 for (j = 0; j < MAX_RT_PRIO; j++) {
6875 INIT_LIST_HEAD(array->queue + j);
6876 __clear_bit(j, array->bitmap);
6879 /* delimiter for bitsearch: */
6880 __set_bit(MAX_RT_PRIO, array->bitmap);
6883 set_load_weight(&init_task);
6885 #ifdef CONFIG_PREEMPT_NOTIFIERS
6886 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6890 nr_cpu_ids = highest_cpu + 1;
6891 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6894 #ifdef CONFIG_RT_MUTEXES
6895 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6899 * The boot idle thread does lazy MMU switching as well:
6901 atomic_inc(&init_mm.mm_count);
6902 enter_lazy_tlb(&init_mm, current);
6905 * Make us the idle thread. Technically, schedule() should not be
6906 * called from this thread, however somewhere below it might be,
6907 * but because we are the idle thread, we just pick up running again
6908 * when this runqueue becomes "idle".
6910 init_idle(current, smp_processor_id());
6912 * During early bootup we pretend to be a normal task:
6914 current->sched_class = &fair_sched_class;
6917 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6918 void __might_sleep(char *file, int line)
6921 static unsigned long prev_jiffy; /* ratelimiting */
6923 if ((in_atomic() || irqs_disabled()) &&
6924 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6925 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6927 prev_jiffy = jiffies;
6928 printk(KERN_ERR "BUG: sleeping function called from invalid"
6929 " context at %s:%d\n", file, line);
6930 printk("in_atomic():%d, irqs_disabled():%d\n",
6931 in_atomic(), irqs_disabled());
6932 debug_show_held_locks(current);
6933 if (irqs_disabled())
6934 print_irqtrace_events(current);
6939 EXPORT_SYMBOL(__might_sleep);
6942 #ifdef CONFIG_MAGIC_SYSRQ
6943 static void normalize_task(struct rq *rq, struct task_struct *p)
6946 update_rq_clock(rq);
6947 on_rq = p->se.on_rq;
6949 deactivate_task(rq, p, 0);
6950 __setscheduler(rq, p, SCHED_NORMAL, 0);
6952 activate_task(rq, p, 0);
6953 resched_task(rq->curr);
6957 void normalize_rt_tasks(void)
6959 struct task_struct *g, *p;
6960 unsigned long flags;
6963 read_lock_irq(&tasklist_lock);
6964 do_each_thread(g, p) {
6966 * Only normalize user tasks:
6971 p->se.exec_start = 0;
6972 #ifdef CONFIG_SCHEDSTATS
6973 p->se.wait_start = 0;
6974 p->se.sleep_start = 0;
6975 p->se.block_start = 0;
6977 task_rq(p)->clock = 0;
6981 * Renice negative nice level userspace
6984 if (TASK_NICE(p) < 0 && p->mm)
6985 set_user_nice(p, 0);
6989 spin_lock_irqsave(&p->pi_lock, flags);
6990 rq = __task_rq_lock(p);
6992 normalize_task(rq, p);
6994 __task_rq_unlock(rq);
6995 spin_unlock_irqrestore(&p->pi_lock, flags);
6996 } while_each_thread(g, p);
6998 read_unlock_irq(&tasklist_lock);
7001 #endif /* CONFIG_MAGIC_SYSRQ */
7005 * These functions are only useful for the IA64 MCA handling.
7007 * They can only be called when the whole system has been
7008 * stopped - every CPU needs to be quiescent, and no scheduling
7009 * activity can take place. Using them for anything else would
7010 * be a serious bug, and as a result, they aren't even visible
7011 * under any other configuration.
7015 * curr_task - return the current task for a given cpu.
7016 * @cpu: the processor in question.
7018 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7020 struct task_struct *curr_task(int cpu)
7022 return cpu_curr(cpu);
7026 * set_curr_task - set the current task for a given cpu.
7027 * @cpu: the processor in question.
7028 * @p: the task pointer to set.
7030 * Description: This function must only be used when non-maskable interrupts
7031 * are serviced on a separate stack. It allows the architecture to switch the
7032 * notion of the current task on a cpu in a non-blocking manner. This function
7033 * must be called with all CPU's synchronized, and interrupts disabled, the
7034 * and caller must save the original value of the current task (see
7035 * curr_task() above) and restore that value before reenabling interrupts and
7036 * re-starting the system.
7038 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7040 void set_curr_task(int cpu, struct task_struct *p)
7047 #ifdef CONFIG_FAIR_GROUP_SCHED
7051 * distribute shares of all task groups among their schedulable entities,
7052 * to reflect load distrbution across cpus.
7054 static int rebalance_shares(struct sched_domain *sd, int this_cpu)
7056 struct cfs_rq *cfs_rq;
7057 struct rq *rq = cpu_rq(this_cpu);
7058 cpumask_t sdspan = sd->span;
7061 /* Walk thr' all the task groups that we have */
7062 for_each_leaf_cfs_rq(rq, cfs_rq) {
7064 unsigned long total_load = 0, total_shares;
7065 struct task_group *tg = cfs_rq->tg;
7067 /* Gather total task load of this group across cpus */
7068 for_each_cpu_mask(i, sdspan)
7069 total_load += tg->cfs_rq[i]->load.weight;
7071 /* Nothing to do if this group has no load */
7076 * tg->shares represents the number of cpu shares the task group
7077 * is eligible to hold on a single cpu. On N cpus, it is
7078 * eligible to hold (N * tg->shares) number of cpu shares.
7080 total_shares = tg->shares * cpus_weight(sdspan);
7083 * redistribute total_shares across cpus as per the task load
7086 for_each_cpu_mask(i, sdspan) {
7087 unsigned long local_load, local_shares;
7089 local_load = tg->cfs_rq[i]->load.weight;
7090 local_shares = (local_load * total_shares) / total_load;
7092 local_shares = MIN_GROUP_SHARES;
7093 if (local_shares == tg->se[i]->load.weight)
7096 spin_lock_irq(&cpu_rq(i)->lock);
7097 set_se_shares(tg->se[i], local_shares);
7098 spin_unlock_irq(&cpu_rq(i)->lock);
7107 * How frequently should we rebalance_shares() across cpus?
7109 * The more frequently we rebalance shares, the more accurate is the fairness
7110 * of cpu bandwidth distribution between task groups. However higher frequency
7111 * also implies increased scheduling overhead.
7113 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7114 * consecutive calls to rebalance_shares() in the same sched domain.
7116 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7117 * consecutive calls to rebalance_shares() in the same sched domain.
7119 * These settings allows for the appropriate tradeoff between accuracy of
7120 * fairness and the associated overhead.
7124 /* default: 8ms, units: milliseconds */
7125 const_debug unsigned int sysctl_sched_min_bal_int_shares = 8;
7127 /* default: 128ms, units: milliseconds */
7128 const_debug unsigned int sysctl_sched_max_bal_int_shares = 128;
7130 /* kernel thread that runs rebalance_shares() periodically */
7131 static int load_balance_monitor(void *unused)
7133 unsigned int timeout = sysctl_sched_min_bal_int_shares;
7134 struct sched_param schedparm;
7138 * We don't want this thread's execution to be limited by the shares
7139 * assigned to default group (init_task_group). Hence make it run
7140 * as a SCHED_RR RT task at the lowest priority.
7142 schedparm.sched_priority = 1;
7143 ret = sched_setscheduler(current, SCHED_RR, &schedparm);
7145 printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance"
7146 " monitor thread (error = %d) \n", ret);
7148 while (!kthread_should_stop()) {
7149 int i, cpu, balanced = 1;
7151 /* Prevent cpus going down or coming up */
7153 /* lockout changes to doms_cur[] array */
7156 * Enter a rcu read-side critical section to safely walk rq->sd
7157 * chain on various cpus and to walk task group list
7158 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7162 for (i = 0; i < ndoms_cur; i++) {
7163 cpumask_t cpumap = doms_cur[i];
7164 struct sched_domain *sd = NULL, *sd_prev = NULL;
7166 cpu = first_cpu(cpumap);
7168 /* Find the highest domain at which to balance shares */
7169 for_each_domain(cpu, sd) {
7170 if (!(sd->flags & SD_LOAD_BALANCE))
7176 /* sd == NULL? No load balance reqd in this domain */
7180 balanced &= rebalance_shares(sd, cpu);
7189 timeout = sysctl_sched_min_bal_int_shares;
7190 else if (timeout < sysctl_sched_max_bal_int_shares)
7193 msleep_interruptible(timeout);
7198 #endif /* CONFIG_SMP */
7200 /* allocate runqueue etc for a new task group */
7201 struct task_group *sched_create_group(void)
7203 struct task_group *tg;
7204 struct cfs_rq *cfs_rq;
7205 struct sched_entity *se;
7209 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7211 return ERR_PTR(-ENOMEM);
7213 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7216 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7220 for_each_possible_cpu(i) {
7223 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7228 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7233 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7234 memset(se, 0, sizeof(struct sched_entity));
7236 tg->cfs_rq[i] = cfs_rq;
7237 init_cfs_rq(cfs_rq, rq);
7241 se->cfs_rq = &rq->cfs;
7243 se->load.weight = NICE_0_LOAD;
7244 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7248 tg->shares = NICE_0_LOAD;
7250 lock_task_group_list();
7251 for_each_possible_cpu(i) {
7253 cfs_rq = tg->cfs_rq[i];
7254 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7256 unlock_task_group_list();
7261 for_each_possible_cpu(i) {
7263 kfree(tg->cfs_rq[i]);
7271 return ERR_PTR(-ENOMEM);
7274 /* rcu callback to free various structures associated with a task group */
7275 static void free_sched_group(struct rcu_head *rhp)
7277 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7278 struct cfs_rq *cfs_rq;
7279 struct sched_entity *se;
7282 /* now it should be safe to free those cfs_rqs */
7283 for_each_possible_cpu(i) {
7284 cfs_rq = tg->cfs_rq[i];
7296 /* Destroy runqueue etc associated with a task group */
7297 void sched_destroy_group(struct task_group *tg)
7299 struct cfs_rq *cfs_rq = NULL;
7302 lock_task_group_list();
7303 for_each_possible_cpu(i) {
7304 cfs_rq = tg->cfs_rq[i];
7305 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7307 unlock_task_group_list();
7311 /* wait for possible concurrent references to cfs_rqs complete */
7312 call_rcu(&tg->rcu, free_sched_group);
7315 /* change task's runqueue when it moves between groups.
7316 * The caller of this function should have put the task in its new group
7317 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7318 * reflect its new group.
7320 void sched_move_task(struct task_struct *tsk)
7323 unsigned long flags;
7326 rq = task_rq_lock(tsk, &flags);
7328 if (tsk->sched_class != &fair_sched_class) {
7329 set_task_cfs_rq(tsk, task_cpu(tsk));
7333 update_rq_clock(rq);
7335 running = task_current(rq, tsk);
7336 on_rq = tsk->se.on_rq;
7339 dequeue_task(rq, tsk, 0);
7340 if (unlikely(running))
7341 tsk->sched_class->put_prev_task(rq, tsk);
7344 set_task_cfs_rq(tsk, task_cpu(tsk));
7347 if (unlikely(running))
7348 tsk->sched_class->set_curr_task(rq);
7349 enqueue_task(rq, tsk, 0);
7353 task_rq_unlock(rq, &flags);
7356 /* rq->lock to be locked by caller */
7357 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7359 struct cfs_rq *cfs_rq = se->cfs_rq;
7360 struct rq *rq = cfs_rq->rq;
7364 shares = MIN_GROUP_SHARES;
7368 dequeue_entity(cfs_rq, se, 0);
7369 dec_cpu_load(rq, se->load.weight);
7372 se->load.weight = shares;
7373 se->load.inv_weight = div64_64((1ULL<<32), shares);
7376 enqueue_entity(cfs_rq, se, 0);
7377 inc_cpu_load(rq, se->load.weight);
7381 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7384 struct cfs_rq *cfs_rq;
7387 lock_task_group_list();
7388 if (tg->shares == shares)
7391 if (shares < MIN_GROUP_SHARES)
7392 shares = MIN_GROUP_SHARES;
7395 * Prevent any load balance activity (rebalance_shares,
7396 * load_balance_fair) from referring to this group first,
7397 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7399 for_each_possible_cpu(i) {
7400 cfs_rq = tg->cfs_rq[i];
7401 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7404 /* wait for any ongoing reference to this group to finish */
7405 synchronize_sched();
7408 * Now we are free to modify the group's share on each cpu
7409 * w/o tripping rebalance_share or load_balance_fair.
7411 tg->shares = shares;
7412 for_each_possible_cpu(i) {
7413 spin_lock_irq(&cpu_rq(i)->lock);
7414 set_se_shares(tg->se[i], shares);
7415 spin_unlock_irq(&cpu_rq(i)->lock);
7419 * Enable load balance activity on this group, by inserting it back on
7420 * each cpu's rq->leaf_cfs_rq_list.
7422 for_each_possible_cpu(i) {
7424 cfs_rq = tg->cfs_rq[i];
7425 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7428 unlock_task_group_list();
7432 unsigned long sched_group_shares(struct task_group *tg)
7437 #endif /* CONFIG_FAIR_GROUP_SCHED */
7439 #ifdef CONFIG_FAIR_CGROUP_SCHED
7441 /* return corresponding task_group object of a cgroup */
7442 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7444 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7445 struct task_group, css);
7448 static struct cgroup_subsys_state *
7449 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7451 struct task_group *tg;
7453 if (!cgrp->parent) {
7454 /* This is early initialization for the top cgroup */
7455 init_task_group.css.cgroup = cgrp;
7456 return &init_task_group.css;
7459 /* we support only 1-level deep hierarchical scheduler atm */
7460 if (cgrp->parent->parent)
7461 return ERR_PTR(-EINVAL);
7463 tg = sched_create_group();
7465 return ERR_PTR(-ENOMEM);
7467 /* Bind the cgroup to task_group object we just created */
7468 tg->css.cgroup = cgrp;
7474 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7476 struct task_group *tg = cgroup_tg(cgrp);
7478 sched_destroy_group(tg);
7482 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7483 struct task_struct *tsk)
7485 /* We don't support RT-tasks being in separate groups */
7486 if (tsk->sched_class != &fair_sched_class)
7493 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7494 struct cgroup *old_cont, struct task_struct *tsk)
7496 sched_move_task(tsk);
7499 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7502 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7505 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7507 struct task_group *tg = cgroup_tg(cgrp);
7509 return (u64) tg->shares;
7512 static struct cftype cpu_files[] = {
7515 .read_uint = cpu_shares_read_uint,
7516 .write_uint = cpu_shares_write_uint,
7520 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7522 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7525 struct cgroup_subsys cpu_cgroup_subsys = {
7527 .create = cpu_cgroup_create,
7528 .destroy = cpu_cgroup_destroy,
7529 .can_attach = cpu_cgroup_can_attach,
7530 .attach = cpu_cgroup_attach,
7531 .populate = cpu_cgroup_populate,
7532 .subsys_id = cpu_cgroup_subsys_id,
7536 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7538 #ifdef CONFIG_CGROUP_CPUACCT
7541 * CPU accounting code for task groups.
7543 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7544 * (balbir@in.ibm.com).
7547 /* track cpu usage of a group of tasks */
7549 struct cgroup_subsys_state css;
7550 /* cpuusage holds pointer to a u64-type object on every cpu */
7554 struct cgroup_subsys cpuacct_subsys;
7556 /* return cpu accounting group corresponding to this container */
7557 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7559 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7560 struct cpuacct, css);
7563 /* return cpu accounting group to which this task belongs */
7564 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7566 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7567 struct cpuacct, css);
7570 /* create a new cpu accounting group */
7571 static struct cgroup_subsys_state *cpuacct_create(
7572 struct cgroup_subsys *ss, struct cgroup *cont)
7574 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7577 return ERR_PTR(-ENOMEM);
7579 ca->cpuusage = alloc_percpu(u64);
7580 if (!ca->cpuusage) {
7582 return ERR_PTR(-ENOMEM);
7588 /* destroy an existing cpu accounting group */
7590 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
7592 struct cpuacct *ca = cgroup_ca(cont);
7594 free_percpu(ca->cpuusage);
7598 /* return total cpu usage (in nanoseconds) of a group */
7599 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7601 struct cpuacct *ca = cgroup_ca(cont);
7602 u64 totalcpuusage = 0;
7605 for_each_possible_cpu(i) {
7606 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7609 * Take rq->lock to make 64-bit addition safe on 32-bit
7612 spin_lock_irq(&cpu_rq(i)->lock);
7613 totalcpuusage += *cpuusage;
7614 spin_unlock_irq(&cpu_rq(i)->lock);
7617 return totalcpuusage;
7620 static struct cftype files[] = {
7623 .read_uint = cpuusage_read,
7627 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7629 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7633 * charge this task's execution time to its accounting group.
7635 * called with rq->lock held.
7637 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7641 if (!cpuacct_subsys.active)
7646 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7648 *cpuusage += cputime;
7652 struct cgroup_subsys cpuacct_subsys = {
7654 .create = cpuacct_create,
7655 .destroy = cpuacct_destroy,
7656 .populate = cpuacct_populate,
7657 .subsys_id = cpuacct_subsys_id,
7659 #endif /* CONFIG_CGROUP_CPUACCT */