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;
349 * This is the main, per-CPU runqueue data structure.
351 * Locking rule: those places that want to lock multiple runqueues
352 * (such as the load balancing or the thread migration code), lock
353 * acquire operations must be ordered by ascending &runqueue.
360 * nr_running and cpu_load should be in the same cacheline because
361 * remote CPUs use both these fields when doing load calculation.
363 unsigned long nr_running;
364 #define CPU_LOAD_IDX_MAX 5
365 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
366 unsigned char idle_at_tick;
368 unsigned char in_nohz_recently;
370 /* capture load from *all* tasks on this cpu: */
371 struct load_weight load;
372 unsigned long nr_load_updates;
376 #ifdef CONFIG_FAIR_GROUP_SCHED
377 /* list of leaf cfs_rq on this cpu: */
378 struct list_head leaf_cfs_rq_list;
383 * This is part of a global counter where only the total sum
384 * over all CPUs matters. A task can increase this counter on
385 * one CPU and if it got migrated afterwards it may decrease
386 * it on another CPU. Always updated under the runqueue lock:
388 unsigned long nr_uninterruptible;
390 struct task_struct *curr, *idle;
391 unsigned long next_balance;
392 struct mm_struct *prev_mm;
394 u64 clock, prev_clock_raw;
397 unsigned int clock_warps, clock_overflows;
399 unsigned int clock_deep_idle_events;
405 struct sched_domain *sd;
407 /* For active balancing */
410 /* cpu of this runqueue: */
413 struct task_struct *migration_thread;
414 struct list_head migration_queue;
417 #ifdef CONFIG_SCHEDSTATS
419 struct sched_info rq_sched_info;
421 /* sys_sched_yield() stats */
422 unsigned int yld_exp_empty;
423 unsigned int yld_act_empty;
424 unsigned int yld_both_empty;
425 unsigned int yld_count;
427 /* schedule() stats */
428 unsigned int sched_switch;
429 unsigned int sched_count;
430 unsigned int sched_goidle;
432 /* try_to_wake_up() stats */
433 unsigned int ttwu_count;
434 unsigned int ttwu_local;
437 unsigned int bkl_count;
439 struct lock_class_key rq_lock_key;
442 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
444 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
446 rq->curr->sched_class->check_preempt_curr(rq, p);
449 static inline int cpu_of(struct rq *rq)
459 * Update the per-runqueue clock, as finegrained as the platform can give
460 * us, but without assuming monotonicity, etc.:
462 static void __update_rq_clock(struct rq *rq)
464 u64 prev_raw = rq->prev_clock_raw;
465 u64 now = sched_clock();
466 s64 delta = now - prev_raw;
467 u64 clock = rq->clock;
469 #ifdef CONFIG_SCHED_DEBUG
470 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
473 * Protect against sched_clock() occasionally going backwards:
475 if (unlikely(delta < 0)) {
480 * Catch too large forward jumps too:
482 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
483 if (clock < rq->tick_timestamp + TICK_NSEC)
484 clock = rq->tick_timestamp + TICK_NSEC;
487 rq->clock_overflows++;
489 if (unlikely(delta > rq->clock_max_delta))
490 rq->clock_max_delta = delta;
495 rq->prev_clock_raw = now;
499 static void update_rq_clock(struct rq *rq)
501 if (likely(smp_processor_id() == cpu_of(rq)))
502 __update_rq_clock(rq);
506 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
507 * See detach_destroy_domains: synchronize_sched for details.
509 * The domain tree of any CPU may only be accessed from within
510 * preempt-disabled sections.
512 #define for_each_domain(cpu, __sd) \
513 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
515 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
516 #define this_rq() (&__get_cpu_var(runqueues))
517 #define task_rq(p) cpu_rq(task_cpu(p))
518 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
521 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
523 #ifdef CONFIG_SCHED_DEBUG
524 # define const_debug __read_mostly
526 # define const_debug static const
530 * Debugging: various feature bits
533 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
534 SCHED_FEAT_WAKEUP_PREEMPT = 2,
535 SCHED_FEAT_START_DEBIT = 4,
536 SCHED_FEAT_TREE_AVG = 8,
537 SCHED_FEAT_APPROX_AVG = 16,
540 const_debug unsigned int sysctl_sched_features =
541 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
542 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
543 SCHED_FEAT_START_DEBIT * 1 |
544 SCHED_FEAT_TREE_AVG * 0 |
545 SCHED_FEAT_APPROX_AVG * 0;
547 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
550 * Number of tasks to iterate in a single balance run.
551 * Limited because this is done with IRQs disabled.
553 const_debug unsigned int sysctl_sched_nr_migrate = 32;
556 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
557 * clock constructed from sched_clock():
559 unsigned long long cpu_clock(int cpu)
561 unsigned long long now;
565 local_irq_save(flags);
568 * Only call sched_clock() if the scheduler has already been
569 * initialized (some code might call cpu_clock() very early):
574 local_irq_restore(flags);
578 EXPORT_SYMBOL_GPL(cpu_clock);
580 #ifndef prepare_arch_switch
581 # define prepare_arch_switch(next) do { } while (0)
583 #ifndef finish_arch_switch
584 # define finish_arch_switch(prev) do { } while (0)
587 static inline int task_current(struct rq *rq, struct task_struct *p)
589 return rq->curr == p;
592 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
593 static inline int task_running(struct rq *rq, struct task_struct *p)
595 return task_current(rq, p);
598 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
602 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
604 #ifdef CONFIG_DEBUG_SPINLOCK
605 /* this is a valid case when another task releases the spinlock */
606 rq->lock.owner = current;
609 * If we are tracking spinlock dependencies then we have to
610 * fix up the runqueue lock - which gets 'carried over' from
613 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
615 spin_unlock_irq(&rq->lock);
618 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
619 static inline int task_running(struct rq *rq, struct task_struct *p)
624 return task_current(rq, p);
628 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
632 * We can optimise this out completely for !SMP, because the
633 * SMP rebalancing from interrupt is the only thing that cares
638 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
639 spin_unlock_irq(&rq->lock);
641 spin_unlock(&rq->lock);
645 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
649 * After ->oncpu is cleared, the task can be moved to a different CPU.
650 * We must ensure this doesn't happen until the switch is completely
656 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
660 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
663 * __task_rq_lock - lock the runqueue a given task resides on.
664 * Must be called interrupts disabled.
666 static inline struct rq *__task_rq_lock(struct task_struct *p)
670 struct rq *rq = task_rq(p);
671 spin_lock(&rq->lock);
672 if (likely(rq == task_rq(p)))
674 spin_unlock(&rq->lock);
679 * task_rq_lock - lock the runqueue a given task resides on and disable
680 * interrupts. Note the ordering: we can safely lookup the task_rq without
681 * explicitly disabling preemption.
683 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
689 local_irq_save(*flags);
691 spin_lock(&rq->lock);
692 if (likely(rq == task_rq(p)))
694 spin_unlock_irqrestore(&rq->lock, *flags);
698 static void __task_rq_unlock(struct rq *rq)
701 spin_unlock(&rq->lock);
704 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
707 spin_unlock_irqrestore(&rq->lock, *flags);
711 * this_rq_lock - lock this runqueue and disable interrupts.
713 static struct rq *this_rq_lock(void)
720 spin_lock(&rq->lock);
726 * We are going deep-idle (irqs are disabled):
728 void sched_clock_idle_sleep_event(void)
730 struct rq *rq = cpu_rq(smp_processor_id());
732 spin_lock(&rq->lock);
733 __update_rq_clock(rq);
734 spin_unlock(&rq->lock);
735 rq->clock_deep_idle_events++;
737 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
740 * We just idled delta nanoseconds (called with irqs disabled):
742 void sched_clock_idle_wakeup_event(u64 delta_ns)
744 struct rq *rq = cpu_rq(smp_processor_id());
745 u64 now = sched_clock();
747 touch_softlockup_watchdog();
748 rq->idle_clock += delta_ns;
750 * Override the previous timestamp and ignore all
751 * sched_clock() deltas that occured while we idled,
752 * and use the PM-provided delta_ns to advance the
755 spin_lock(&rq->lock);
756 rq->prev_clock_raw = now;
757 rq->clock += delta_ns;
758 spin_unlock(&rq->lock);
760 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
763 * resched_task - mark a task 'to be rescheduled now'.
765 * On UP this means the setting of the need_resched flag, on SMP it
766 * might also involve a cross-CPU call to trigger the scheduler on
771 #ifndef tsk_is_polling
772 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
775 static void resched_task(struct task_struct *p)
779 assert_spin_locked(&task_rq(p)->lock);
781 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
784 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
787 if (cpu == smp_processor_id())
790 /* NEED_RESCHED must be visible before we test polling */
792 if (!tsk_is_polling(p))
793 smp_send_reschedule(cpu);
796 static void resched_cpu(int cpu)
798 struct rq *rq = cpu_rq(cpu);
801 if (!spin_trylock_irqsave(&rq->lock, flags))
803 resched_task(cpu_curr(cpu));
804 spin_unlock_irqrestore(&rq->lock, flags);
807 static inline void resched_task(struct task_struct *p)
809 assert_spin_locked(&task_rq(p)->lock);
810 set_tsk_need_resched(p);
814 #if BITS_PER_LONG == 32
815 # define WMULT_CONST (~0UL)
817 # define WMULT_CONST (1UL << 32)
820 #define WMULT_SHIFT 32
823 * Shift right and round:
825 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
828 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
829 struct load_weight *lw)
833 if (unlikely(!lw->inv_weight))
834 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
836 tmp = (u64)delta_exec * weight;
838 * Check whether we'd overflow the 64-bit multiplication:
840 if (unlikely(tmp > WMULT_CONST))
841 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
844 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
846 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
849 static inline unsigned long
850 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
852 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
855 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
860 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
866 * To aid in avoiding the subversion of "niceness" due to uneven distribution
867 * of tasks with abnormal "nice" values across CPUs the contribution that
868 * each task makes to its run queue's load is weighted according to its
869 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
870 * scaled version of the new time slice allocation that they receive on time
874 #define WEIGHT_IDLEPRIO 2
875 #define WMULT_IDLEPRIO (1 << 31)
878 * Nice levels are multiplicative, with a gentle 10% change for every
879 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
880 * nice 1, it will get ~10% less CPU time than another CPU-bound task
881 * that remained on nice 0.
883 * The "10% effect" is relative and cumulative: from _any_ nice level,
884 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
885 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
886 * If a task goes up by ~10% and another task goes down by ~10% then
887 * the relative distance between them is ~25%.)
889 static const int prio_to_weight[40] = {
890 /* -20 */ 88761, 71755, 56483, 46273, 36291,
891 /* -15 */ 29154, 23254, 18705, 14949, 11916,
892 /* -10 */ 9548, 7620, 6100, 4904, 3906,
893 /* -5 */ 3121, 2501, 1991, 1586, 1277,
894 /* 0 */ 1024, 820, 655, 526, 423,
895 /* 5 */ 335, 272, 215, 172, 137,
896 /* 10 */ 110, 87, 70, 56, 45,
897 /* 15 */ 36, 29, 23, 18, 15,
901 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
903 * In cases where the weight does not change often, we can use the
904 * precalculated inverse to speed up arithmetics by turning divisions
905 * into multiplications:
907 static const u32 prio_to_wmult[40] = {
908 /* -20 */ 48388, 59856, 76040, 92818, 118348,
909 /* -15 */ 147320, 184698, 229616, 287308, 360437,
910 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
911 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
912 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
913 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
914 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
915 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
918 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
921 * runqueue iterator, to support SMP load-balancing between different
922 * scheduling classes, without having to expose their internal data
923 * structures to the load-balancing proper:
927 struct task_struct *(*start)(void *);
928 struct task_struct *(*next)(void *);
933 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
934 unsigned long max_load_move, struct sched_domain *sd,
935 enum cpu_idle_type idle, int *all_pinned,
936 int *this_best_prio, struct rq_iterator *iterator);
939 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
940 struct sched_domain *sd, enum cpu_idle_type idle,
941 struct rq_iterator *iterator);
944 #ifdef CONFIG_CGROUP_CPUACCT
945 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
947 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
950 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
952 update_load_add(&rq->load, load);
955 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
957 update_load_sub(&rq->load, load);
960 #include "sched_stats.h"
961 #include "sched_idletask.c"
962 #include "sched_fair.c"
963 #include "sched_rt.c"
964 #ifdef CONFIG_SCHED_DEBUG
965 # include "sched_debug.c"
968 #define sched_class_highest (&rt_sched_class)
970 static void inc_nr_running(struct task_struct *p, struct rq *rq)
975 static void dec_nr_running(struct task_struct *p, struct rq *rq)
980 static void set_load_weight(struct task_struct *p)
982 if (task_has_rt_policy(p)) {
983 p->se.load.weight = prio_to_weight[0] * 2;
984 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
989 * SCHED_IDLE tasks get minimal weight:
991 if (p->policy == SCHED_IDLE) {
992 p->se.load.weight = WEIGHT_IDLEPRIO;
993 p->se.load.inv_weight = WMULT_IDLEPRIO;
997 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
998 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1001 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1003 sched_info_queued(p);
1004 p->sched_class->enqueue_task(rq, p, wakeup);
1008 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1010 p->sched_class->dequeue_task(rq, p, sleep);
1015 * __normal_prio - return the priority that is based on the static prio
1017 static inline int __normal_prio(struct task_struct *p)
1019 return p->static_prio;
1023 * Calculate the expected normal priority: i.e. priority
1024 * without taking RT-inheritance into account. Might be
1025 * boosted by interactivity modifiers. Changes upon fork,
1026 * setprio syscalls, and whenever the interactivity
1027 * estimator recalculates.
1029 static inline int normal_prio(struct task_struct *p)
1033 if (task_has_rt_policy(p))
1034 prio = MAX_RT_PRIO-1 - p->rt_priority;
1036 prio = __normal_prio(p);
1041 * Calculate the current priority, i.e. the priority
1042 * taken into account by the scheduler. This value might
1043 * be boosted by RT tasks, or might be boosted by
1044 * interactivity modifiers. Will be RT if the task got
1045 * RT-boosted. If not then it returns p->normal_prio.
1047 static int effective_prio(struct task_struct *p)
1049 p->normal_prio = normal_prio(p);
1051 * If we are RT tasks or we were boosted to RT priority,
1052 * keep the priority unchanged. Otherwise, update priority
1053 * to the normal priority:
1055 if (!rt_prio(p->prio))
1056 return p->normal_prio;
1061 * activate_task - move a task to the runqueue.
1063 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1065 if (p->state == TASK_UNINTERRUPTIBLE)
1066 rq->nr_uninterruptible--;
1068 enqueue_task(rq, p, wakeup);
1069 inc_nr_running(p, rq);
1073 * deactivate_task - remove a task from the runqueue.
1075 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1077 if (p->state == TASK_UNINTERRUPTIBLE)
1078 rq->nr_uninterruptible++;
1080 dequeue_task(rq, p, sleep);
1081 dec_nr_running(p, rq);
1085 * task_curr - is this task currently executing on a CPU?
1086 * @p: the task in question.
1088 inline int task_curr(const struct task_struct *p)
1090 return cpu_curr(task_cpu(p)) == p;
1093 /* Used instead of source_load when we know the type == 0 */
1094 unsigned long weighted_cpuload(const int cpu)
1096 return cpu_rq(cpu)->load.weight;
1099 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1101 set_task_cfs_rq(p, cpu);
1104 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1105 * successfuly executed on another CPU. We must ensure that updates of
1106 * per-task data have been completed by this moment.
1109 task_thread_info(p)->cpu = cpu;
1116 * Is this task likely cache-hot:
1119 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1123 if (p->sched_class != &fair_sched_class)
1126 if (sysctl_sched_migration_cost == -1)
1128 if (sysctl_sched_migration_cost == 0)
1131 delta = now - p->se.exec_start;
1133 return delta < (s64)sysctl_sched_migration_cost;
1137 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1139 int old_cpu = task_cpu(p);
1140 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1141 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1142 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1145 clock_offset = old_rq->clock - new_rq->clock;
1147 #ifdef CONFIG_SCHEDSTATS
1148 if (p->se.wait_start)
1149 p->se.wait_start -= clock_offset;
1150 if (p->se.sleep_start)
1151 p->se.sleep_start -= clock_offset;
1152 if (p->se.block_start)
1153 p->se.block_start -= clock_offset;
1154 if (old_cpu != new_cpu) {
1155 schedstat_inc(p, se.nr_migrations);
1156 if (task_hot(p, old_rq->clock, NULL))
1157 schedstat_inc(p, se.nr_forced2_migrations);
1160 p->se.vruntime -= old_cfsrq->min_vruntime -
1161 new_cfsrq->min_vruntime;
1163 __set_task_cpu(p, new_cpu);
1166 struct migration_req {
1167 struct list_head list;
1169 struct task_struct *task;
1172 struct completion done;
1176 * The task's runqueue lock must be held.
1177 * Returns true if you have to wait for migration thread.
1180 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1182 struct rq *rq = task_rq(p);
1185 * If the task is not on a runqueue (and not running), then
1186 * it is sufficient to simply update the task's cpu field.
1188 if (!p->se.on_rq && !task_running(rq, p)) {
1189 set_task_cpu(p, dest_cpu);
1193 init_completion(&req->done);
1195 req->dest_cpu = dest_cpu;
1196 list_add(&req->list, &rq->migration_queue);
1202 * wait_task_inactive - wait for a thread to unschedule.
1204 * The caller must ensure that the task *will* unschedule sometime soon,
1205 * else this function might spin for a *long* time. This function can't
1206 * be called with interrupts off, or it may introduce deadlock with
1207 * smp_call_function() if an IPI is sent by the same process we are
1208 * waiting to become inactive.
1210 void wait_task_inactive(struct task_struct *p)
1212 unsigned long flags;
1218 * We do the initial early heuristics without holding
1219 * any task-queue locks at all. We'll only try to get
1220 * the runqueue lock when things look like they will
1226 * If the task is actively running on another CPU
1227 * still, just relax and busy-wait without holding
1230 * NOTE! Since we don't hold any locks, it's not
1231 * even sure that "rq" stays as the right runqueue!
1232 * But we don't care, since "task_running()" will
1233 * return false if the runqueue has changed and p
1234 * is actually now running somewhere else!
1236 while (task_running(rq, p))
1240 * Ok, time to look more closely! We need the rq
1241 * lock now, to be *sure*. If we're wrong, we'll
1242 * just go back and repeat.
1244 rq = task_rq_lock(p, &flags);
1245 running = task_running(rq, p);
1246 on_rq = p->se.on_rq;
1247 task_rq_unlock(rq, &flags);
1250 * Was it really running after all now that we
1251 * checked with the proper locks actually held?
1253 * Oops. Go back and try again..
1255 if (unlikely(running)) {
1261 * It's not enough that it's not actively running,
1262 * it must be off the runqueue _entirely_, and not
1265 * So if it wa still runnable (but just not actively
1266 * running right now), it's preempted, and we should
1267 * yield - it could be a while.
1269 if (unlikely(on_rq)) {
1270 schedule_timeout_uninterruptible(1);
1275 * Ahh, all good. It wasn't running, and it wasn't
1276 * runnable, which means that it will never become
1277 * running in the future either. We're all done!
1284 * kick_process - kick a running thread to enter/exit the kernel
1285 * @p: the to-be-kicked thread
1287 * Cause a process which is running on another CPU to enter
1288 * kernel-mode, without any delay. (to get signals handled.)
1290 * NOTE: this function doesnt have to take the runqueue lock,
1291 * because all it wants to ensure is that the remote task enters
1292 * the kernel. If the IPI races and the task has been migrated
1293 * to another CPU then no harm is done and the purpose has been
1296 void kick_process(struct task_struct *p)
1302 if ((cpu != smp_processor_id()) && task_curr(p))
1303 smp_send_reschedule(cpu);
1308 * Return a low guess at the load of a migration-source cpu weighted
1309 * according to the scheduling class and "nice" value.
1311 * We want to under-estimate the load of migration sources, to
1312 * balance conservatively.
1314 static unsigned long source_load(int cpu, int type)
1316 struct rq *rq = cpu_rq(cpu);
1317 unsigned long total = weighted_cpuload(cpu);
1322 return min(rq->cpu_load[type-1], total);
1326 * Return a high guess at the load of a migration-target cpu weighted
1327 * according to the scheduling class and "nice" value.
1329 static unsigned long target_load(int cpu, int type)
1331 struct rq *rq = cpu_rq(cpu);
1332 unsigned long total = weighted_cpuload(cpu);
1337 return max(rq->cpu_load[type-1], total);
1341 * Return the average load per task on the cpu's run queue
1343 static inline unsigned long cpu_avg_load_per_task(int cpu)
1345 struct rq *rq = cpu_rq(cpu);
1346 unsigned long total = weighted_cpuload(cpu);
1347 unsigned long n = rq->nr_running;
1349 return n ? total / n : SCHED_LOAD_SCALE;
1353 * find_idlest_group finds and returns the least busy CPU group within the
1356 static struct sched_group *
1357 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1359 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1360 unsigned long min_load = ULONG_MAX, this_load = 0;
1361 int load_idx = sd->forkexec_idx;
1362 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1365 unsigned long load, avg_load;
1369 /* Skip over this group if it has no CPUs allowed */
1370 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1373 local_group = cpu_isset(this_cpu, group->cpumask);
1375 /* Tally up the load of all CPUs in the group */
1378 for_each_cpu_mask(i, group->cpumask) {
1379 /* Bias balancing toward cpus of our domain */
1381 load = source_load(i, load_idx);
1383 load = target_load(i, load_idx);
1388 /* Adjust by relative CPU power of the group */
1389 avg_load = sg_div_cpu_power(group,
1390 avg_load * SCHED_LOAD_SCALE);
1393 this_load = avg_load;
1395 } else if (avg_load < min_load) {
1396 min_load = avg_load;
1399 } while (group = group->next, group != sd->groups);
1401 if (!idlest || 100*this_load < imbalance*min_load)
1407 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1410 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1413 unsigned long load, min_load = ULONG_MAX;
1417 /* Traverse only the allowed CPUs */
1418 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1420 for_each_cpu_mask(i, tmp) {
1421 load = weighted_cpuload(i);
1423 if (load < min_load || (load == min_load && i == this_cpu)) {
1433 * sched_balance_self: balance the current task (running on cpu) in domains
1434 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1437 * Balance, ie. select the least loaded group.
1439 * Returns the target CPU number, or the same CPU if no balancing is needed.
1441 * preempt must be disabled.
1443 static int sched_balance_self(int cpu, int flag)
1445 struct task_struct *t = current;
1446 struct sched_domain *tmp, *sd = NULL;
1448 for_each_domain(cpu, tmp) {
1450 * If power savings logic is enabled for a domain, stop there.
1452 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1454 if (tmp->flags & flag)
1460 struct sched_group *group;
1461 int new_cpu, weight;
1463 if (!(sd->flags & flag)) {
1469 group = find_idlest_group(sd, t, cpu);
1475 new_cpu = find_idlest_cpu(group, t, cpu);
1476 if (new_cpu == -1 || new_cpu == cpu) {
1477 /* Now try balancing at a lower domain level of cpu */
1482 /* Now try balancing at a lower domain level of new_cpu */
1485 weight = cpus_weight(span);
1486 for_each_domain(cpu, tmp) {
1487 if (weight <= cpus_weight(tmp->span))
1489 if (tmp->flags & flag)
1492 /* while loop will break here if sd == NULL */
1498 #endif /* CONFIG_SMP */
1501 * wake_idle() will wake a task on an idle cpu if task->cpu is
1502 * not idle and an idle cpu is available. The span of cpus to
1503 * search starts with cpus closest then further out as needed,
1504 * so we always favor a closer, idle cpu.
1506 * Returns the CPU we should wake onto.
1508 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1509 static int wake_idle(int cpu, struct task_struct *p)
1512 struct sched_domain *sd;
1516 * If it is idle, then it is the best cpu to run this task.
1518 * This cpu is also the best, if it has more than one task already.
1519 * Siblings must be also busy(in most cases) as they didn't already
1520 * pickup the extra load from this cpu and hence we need not check
1521 * sibling runqueue info. This will avoid the checks and cache miss
1522 * penalities associated with that.
1524 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1527 for_each_domain(cpu, sd) {
1528 if (sd->flags & SD_WAKE_IDLE) {
1529 cpus_and(tmp, sd->span, p->cpus_allowed);
1530 for_each_cpu_mask(i, tmp) {
1532 if (i != task_cpu(p)) {
1534 se.nr_wakeups_idle);
1546 static inline int wake_idle(int cpu, struct task_struct *p)
1553 * try_to_wake_up - wake up a thread
1554 * @p: the to-be-woken-up thread
1555 * @state: the mask of task states that can be woken
1556 * @sync: do a synchronous wakeup?
1558 * Put it on the run-queue if it's not already there. The "current"
1559 * thread is always on the run-queue (except when the actual
1560 * re-schedule is in progress), and as such you're allowed to do
1561 * the simpler "current->state = TASK_RUNNING" to mark yourself
1562 * runnable without the overhead of this.
1564 * returns failure only if the task is already active.
1566 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1568 int cpu, orig_cpu, this_cpu, success = 0;
1569 unsigned long flags;
1573 struct sched_domain *sd, *this_sd = NULL;
1574 unsigned long load, this_load;
1578 rq = task_rq_lock(p, &flags);
1579 old_state = p->state;
1580 if (!(old_state & state))
1588 this_cpu = smp_processor_id();
1591 if (unlikely(task_running(rq, p)))
1596 schedstat_inc(rq, ttwu_count);
1597 if (cpu == this_cpu) {
1598 schedstat_inc(rq, ttwu_local);
1602 for_each_domain(this_cpu, sd) {
1603 if (cpu_isset(cpu, sd->span)) {
1604 schedstat_inc(sd, ttwu_wake_remote);
1610 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1614 * Check for affine wakeup and passive balancing possibilities.
1617 int idx = this_sd->wake_idx;
1618 unsigned int imbalance;
1620 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1622 load = source_load(cpu, idx);
1623 this_load = target_load(this_cpu, idx);
1625 new_cpu = this_cpu; /* Wake to this CPU if we can */
1627 if (this_sd->flags & SD_WAKE_AFFINE) {
1628 unsigned long tl = this_load;
1629 unsigned long tl_per_task;
1632 * Attract cache-cold tasks on sync wakeups:
1634 if (sync && !task_hot(p, rq->clock, this_sd))
1637 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1638 tl_per_task = cpu_avg_load_per_task(this_cpu);
1641 * If sync wakeup then subtract the (maximum possible)
1642 * effect of the currently running task from the load
1643 * of the current CPU:
1646 tl -= current->se.load.weight;
1649 tl + target_load(cpu, idx) <= tl_per_task) ||
1650 100*(tl + p->se.load.weight) <= imbalance*load) {
1652 * This domain has SD_WAKE_AFFINE and
1653 * p is cache cold in this domain, and
1654 * there is no bad imbalance.
1656 schedstat_inc(this_sd, ttwu_move_affine);
1657 schedstat_inc(p, se.nr_wakeups_affine);
1663 * Start passive balancing when half the imbalance_pct
1666 if (this_sd->flags & SD_WAKE_BALANCE) {
1667 if (imbalance*this_load <= 100*load) {
1668 schedstat_inc(this_sd, ttwu_move_balance);
1669 schedstat_inc(p, se.nr_wakeups_passive);
1675 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1677 new_cpu = wake_idle(new_cpu, p);
1678 if (new_cpu != cpu) {
1679 set_task_cpu(p, new_cpu);
1680 task_rq_unlock(rq, &flags);
1681 /* might preempt at this point */
1682 rq = task_rq_lock(p, &flags);
1683 old_state = p->state;
1684 if (!(old_state & state))
1689 this_cpu = smp_processor_id();
1694 #endif /* CONFIG_SMP */
1695 schedstat_inc(p, se.nr_wakeups);
1697 schedstat_inc(p, se.nr_wakeups_sync);
1698 if (orig_cpu != cpu)
1699 schedstat_inc(p, se.nr_wakeups_migrate);
1700 if (cpu == this_cpu)
1701 schedstat_inc(p, se.nr_wakeups_local);
1703 schedstat_inc(p, se.nr_wakeups_remote);
1704 update_rq_clock(rq);
1705 activate_task(rq, p, 1);
1706 check_preempt_curr(rq, p);
1710 p->state = TASK_RUNNING;
1712 task_rq_unlock(rq, &flags);
1717 int fastcall wake_up_process(struct task_struct *p)
1719 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1720 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1722 EXPORT_SYMBOL(wake_up_process);
1724 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1726 return try_to_wake_up(p, state, 0);
1730 * Perform scheduler related setup for a newly forked process p.
1731 * p is forked by current.
1733 * __sched_fork() is basic setup used by init_idle() too:
1735 static void __sched_fork(struct task_struct *p)
1737 p->se.exec_start = 0;
1738 p->se.sum_exec_runtime = 0;
1739 p->se.prev_sum_exec_runtime = 0;
1741 #ifdef CONFIG_SCHEDSTATS
1742 p->se.wait_start = 0;
1743 p->se.sum_sleep_runtime = 0;
1744 p->se.sleep_start = 0;
1745 p->se.block_start = 0;
1746 p->se.sleep_max = 0;
1747 p->se.block_max = 0;
1749 p->se.slice_max = 0;
1753 INIT_LIST_HEAD(&p->run_list);
1756 #ifdef CONFIG_PREEMPT_NOTIFIERS
1757 INIT_HLIST_HEAD(&p->preempt_notifiers);
1761 * We mark the process as running here, but have not actually
1762 * inserted it onto the runqueue yet. This guarantees that
1763 * nobody will actually run it, and a signal or other external
1764 * event cannot wake it up and insert it on the runqueue either.
1766 p->state = TASK_RUNNING;
1770 * fork()/clone()-time setup:
1772 void sched_fork(struct task_struct *p, int clone_flags)
1774 int cpu = get_cpu();
1779 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1781 set_task_cpu(p, cpu);
1784 * Make sure we do not leak PI boosting priority to the child:
1786 p->prio = current->normal_prio;
1787 if (!rt_prio(p->prio))
1788 p->sched_class = &fair_sched_class;
1790 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1791 if (likely(sched_info_on()))
1792 memset(&p->sched_info, 0, sizeof(p->sched_info));
1794 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1797 #ifdef CONFIG_PREEMPT
1798 /* Want to start with kernel preemption disabled. */
1799 task_thread_info(p)->preempt_count = 1;
1805 * wake_up_new_task - wake up a newly created task for the first time.
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1811 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1813 unsigned long flags;
1816 rq = task_rq_lock(p, &flags);
1817 BUG_ON(p->state != TASK_RUNNING);
1818 update_rq_clock(rq);
1820 p->prio = effective_prio(p);
1822 if (!p->sched_class->task_new || !current->se.on_rq) {
1823 activate_task(rq, p, 0);
1826 * Let the scheduling class do new task startup
1827 * management (if any):
1829 p->sched_class->task_new(rq, p);
1830 inc_nr_running(p, rq);
1832 check_preempt_curr(rq, p);
1833 task_rq_unlock(rq, &flags);
1836 #ifdef CONFIG_PREEMPT_NOTIFIERS
1839 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1840 * @notifier: notifier struct to register
1842 void preempt_notifier_register(struct preempt_notifier *notifier)
1844 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1846 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1849 * preempt_notifier_unregister - no longer interested in preemption notifications
1850 * @notifier: notifier struct to unregister
1852 * This is safe to call from within a preemption notifier.
1854 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1856 hlist_del(¬ifier->link);
1858 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1860 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1862 struct preempt_notifier *notifier;
1863 struct hlist_node *node;
1865 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1866 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1870 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1871 struct task_struct *next)
1873 struct preempt_notifier *notifier;
1874 struct hlist_node *node;
1876 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1877 notifier->ops->sched_out(notifier, next);
1882 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1887 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1888 struct task_struct *next)
1895 * prepare_task_switch - prepare to switch tasks
1896 * @rq: the runqueue preparing to switch
1897 * @prev: the current task that is being switched out
1898 * @next: the task we are going to switch to.
1900 * This is called with the rq lock held and interrupts off. It must
1901 * be paired with a subsequent finish_task_switch after the context
1904 * prepare_task_switch sets up locking and calls architecture specific
1908 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1909 struct task_struct *next)
1911 fire_sched_out_preempt_notifiers(prev, next);
1912 prepare_lock_switch(rq, next);
1913 prepare_arch_switch(next);
1917 * finish_task_switch - clean up after a task-switch
1918 * @rq: runqueue associated with task-switch
1919 * @prev: the thread we just switched away from.
1921 * finish_task_switch must be called after the context switch, paired
1922 * with a prepare_task_switch call before the context switch.
1923 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1924 * and do any other architecture-specific cleanup actions.
1926 * Note that we may have delayed dropping an mm in context_switch(). If
1927 * so, we finish that here outside of the runqueue lock. (Doing it
1928 * with the lock held can cause deadlocks; see schedule() for
1931 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1932 __releases(rq->lock)
1934 struct mm_struct *mm = rq->prev_mm;
1940 * A task struct has one reference for the use as "current".
1941 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1942 * schedule one last time. The schedule call will never return, and
1943 * the scheduled task must drop that reference.
1944 * The test for TASK_DEAD must occur while the runqueue locks are
1945 * still held, otherwise prev could be scheduled on another cpu, die
1946 * there before we look at prev->state, and then the reference would
1948 * Manfred Spraul <manfred@colorfullife.com>
1950 prev_state = prev->state;
1951 finish_arch_switch(prev);
1952 finish_lock_switch(rq, prev);
1953 fire_sched_in_preempt_notifiers(current);
1956 if (unlikely(prev_state == TASK_DEAD)) {
1958 * Remove function-return probe instances associated with this
1959 * task and put them back on the free list.
1961 kprobe_flush_task(prev);
1962 put_task_struct(prev);
1967 * schedule_tail - first thing a freshly forked thread must call.
1968 * @prev: the thread we just switched away from.
1970 asmlinkage void schedule_tail(struct task_struct *prev)
1971 __releases(rq->lock)
1973 struct rq *rq = this_rq();
1975 finish_task_switch(rq, prev);
1976 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1977 /* In this case, finish_task_switch does not reenable preemption */
1980 if (current->set_child_tid)
1981 put_user(task_pid_vnr(current), current->set_child_tid);
1985 * context_switch - switch to the new MM and the new
1986 * thread's register state.
1989 context_switch(struct rq *rq, struct task_struct *prev,
1990 struct task_struct *next)
1992 struct mm_struct *mm, *oldmm;
1994 prepare_task_switch(rq, prev, next);
1996 oldmm = prev->active_mm;
1998 * For paravirt, this is coupled with an exit in switch_to to
1999 * combine the page table reload and the switch backend into
2002 arch_enter_lazy_cpu_mode();
2004 if (unlikely(!mm)) {
2005 next->active_mm = oldmm;
2006 atomic_inc(&oldmm->mm_count);
2007 enter_lazy_tlb(oldmm, next);
2009 switch_mm(oldmm, mm, next);
2011 if (unlikely(!prev->mm)) {
2012 prev->active_mm = NULL;
2013 rq->prev_mm = oldmm;
2016 * Since the runqueue lock will be released by the next
2017 * task (which is an invalid locking op but in the case
2018 * of the scheduler it's an obvious special-case), so we
2019 * do an early lockdep release here:
2021 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2022 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2025 /* Here we just switch the register state and the stack. */
2026 switch_to(prev, next, prev);
2030 * this_rq must be evaluated again because prev may have moved
2031 * CPUs since it called schedule(), thus the 'rq' on its stack
2032 * frame will be invalid.
2034 finish_task_switch(this_rq(), prev);
2038 * nr_running, nr_uninterruptible and nr_context_switches:
2040 * externally visible scheduler statistics: current number of runnable
2041 * threads, current number of uninterruptible-sleeping threads, total
2042 * number of context switches performed since bootup.
2044 unsigned long nr_running(void)
2046 unsigned long i, sum = 0;
2048 for_each_online_cpu(i)
2049 sum += cpu_rq(i)->nr_running;
2054 unsigned long nr_uninterruptible(void)
2056 unsigned long i, sum = 0;
2058 for_each_possible_cpu(i)
2059 sum += cpu_rq(i)->nr_uninterruptible;
2062 * Since we read the counters lockless, it might be slightly
2063 * inaccurate. Do not allow it to go below zero though:
2065 if (unlikely((long)sum < 0))
2071 unsigned long long nr_context_switches(void)
2074 unsigned long long sum = 0;
2076 for_each_possible_cpu(i)
2077 sum += cpu_rq(i)->nr_switches;
2082 unsigned long nr_iowait(void)
2084 unsigned long i, sum = 0;
2086 for_each_possible_cpu(i)
2087 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2092 unsigned long nr_active(void)
2094 unsigned long i, running = 0, uninterruptible = 0;
2096 for_each_online_cpu(i) {
2097 running += cpu_rq(i)->nr_running;
2098 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2101 if (unlikely((long)uninterruptible < 0))
2102 uninterruptible = 0;
2104 return running + uninterruptible;
2108 * Update rq->cpu_load[] statistics. This function is usually called every
2109 * scheduler tick (TICK_NSEC).
2111 static void update_cpu_load(struct rq *this_rq)
2113 unsigned long this_load = this_rq->load.weight;
2116 this_rq->nr_load_updates++;
2118 /* Update our load: */
2119 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2120 unsigned long old_load, new_load;
2122 /* scale is effectively 1 << i now, and >> i divides by scale */
2124 old_load = this_rq->cpu_load[i];
2125 new_load = this_load;
2127 * Round up the averaging division if load is increasing. This
2128 * prevents us from getting stuck on 9 if the load is 10, for
2131 if (new_load > old_load)
2132 new_load += scale-1;
2133 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2140 * double_rq_lock - safely lock two runqueues
2142 * Note this does not disable interrupts like task_rq_lock,
2143 * you need to do so manually before calling.
2145 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2146 __acquires(rq1->lock)
2147 __acquires(rq2->lock)
2149 BUG_ON(!irqs_disabled());
2151 spin_lock(&rq1->lock);
2152 __acquire(rq2->lock); /* Fake it out ;) */
2155 spin_lock(&rq1->lock);
2156 spin_lock(&rq2->lock);
2158 spin_lock(&rq2->lock);
2159 spin_lock(&rq1->lock);
2162 update_rq_clock(rq1);
2163 update_rq_clock(rq2);
2167 * double_rq_unlock - safely unlock two runqueues
2169 * Note this does not restore interrupts like task_rq_unlock,
2170 * you need to do so manually after calling.
2172 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2173 __releases(rq1->lock)
2174 __releases(rq2->lock)
2176 spin_unlock(&rq1->lock);
2178 spin_unlock(&rq2->lock);
2180 __release(rq2->lock);
2184 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2186 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2187 __releases(this_rq->lock)
2188 __acquires(busiest->lock)
2189 __acquires(this_rq->lock)
2191 if (unlikely(!irqs_disabled())) {
2192 /* printk() doesn't work good under rq->lock */
2193 spin_unlock(&this_rq->lock);
2196 if (unlikely(!spin_trylock(&busiest->lock))) {
2197 if (busiest < this_rq) {
2198 spin_unlock(&this_rq->lock);
2199 spin_lock(&busiest->lock);
2200 spin_lock(&this_rq->lock);
2202 spin_lock(&busiest->lock);
2207 * If dest_cpu is allowed for this process, migrate the task to it.
2208 * This is accomplished by forcing the cpu_allowed mask to only
2209 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2210 * the cpu_allowed mask is restored.
2212 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2214 struct migration_req req;
2215 unsigned long flags;
2218 rq = task_rq_lock(p, &flags);
2219 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2220 || unlikely(cpu_is_offline(dest_cpu)))
2223 /* force the process onto the specified CPU */
2224 if (migrate_task(p, dest_cpu, &req)) {
2225 /* Need to wait for migration thread (might exit: take ref). */
2226 struct task_struct *mt = rq->migration_thread;
2228 get_task_struct(mt);
2229 task_rq_unlock(rq, &flags);
2230 wake_up_process(mt);
2231 put_task_struct(mt);
2232 wait_for_completion(&req.done);
2237 task_rq_unlock(rq, &flags);
2241 * sched_exec - execve() is a valuable balancing opportunity, because at
2242 * this point the task has the smallest effective memory and cache footprint.
2244 void sched_exec(void)
2246 int new_cpu, this_cpu = get_cpu();
2247 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2249 if (new_cpu != this_cpu)
2250 sched_migrate_task(current, new_cpu);
2254 * pull_task - move a task from a remote runqueue to the local runqueue.
2255 * Both runqueues must be locked.
2257 static void pull_task(struct rq *src_rq, struct task_struct *p,
2258 struct rq *this_rq, int this_cpu)
2260 deactivate_task(src_rq, p, 0);
2261 set_task_cpu(p, this_cpu);
2262 activate_task(this_rq, p, 0);
2264 * Note that idle threads have a prio of MAX_PRIO, for this test
2265 * to be always true for them.
2267 check_preempt_curr(this_rq, p);
2271 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2274 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2275 struct sched_domain *sd, enum cpu_idle_type idle,
2279 * We do not migrate tasks that are:
2280 * 1) running (obviously), or
2281 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2282 * 3) are cache-hot on their current CPU.
2284 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2285 schedstat_inc(p, se.nr_failed_migrations_affine);
2290 if (task_running(rq, p)) {
2291 schedstat_inc(p, se.nr_failed_migrations_running);
2296 * Aggressive migration if:
2297 * 1) task is cache cold, or
2298 * 2) too many balance attempts have failed.
2301 if (!task_hot(p, rq->clock, sd) ||
2302 sd->nr_balance_failed > sd->cache_nice_tries) {
2303 #ifdef CONFIG_SCHEDSTATS
2304 if (task_hot(p, rq->clock, sd)) {
2305 schedstat_inc(sd, lb_hot_gained[idle]);
2306 schedstat_inc(p, se.nr_forced_migrations);
2312 if (task_hot(p, rq->clock, sd)) {
2313 schedstat_inc(p, se.nr_failed_migrations_hot);
2319 static unsigned long
2320 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2321 unsigned long max_load_move, struct sched_domain *sd,
2322 enum cpu_idle_type idle, int *all_pinned,
2323 int *this_best_prio, struct rq_iterator *iterator)
2325 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2326 struct task_struct *p;
2327 long rem_load_move = max_load_move;
2329 if (max_load_move == 0)
2335 * Start the load-balancing iterator:
2337 p = iterator->start(iterator->arg);
2339 if (!p || loops++ > sysctl_sched_nr_migrate)
2342 * To help distribute high priority tasks across CPUs we don't
2343 * skip a task if it will be the highest priority task (i.e. smallest
2344 * prio value) on its new queue regardless of its load weight
2346 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2347 SCHED_LOAD_SCALE_FUZZ;
2348 if ((skip_for_load && p->prio >= *this_best_prio) ||
2349 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2350 p = iterator->next(iterator->arg);
2354 pull_task(busiest, p, this_rq, this_cpu);
2356 rem_load_move -= p->se.load.weight;
2359 * We only want to steal up to the prescribed amount of weighted load.
2361 if (rem_load_move > 0) {
2362 if (p->prio < *this_best_prio)
2363 *this_best_prio = p->prio;
2364 p = iterator->next(iterator->arg);
2369 * Right now, this is one of only two places pull_task() is called,
2370 * so we can safely collect pull_task() stats here rather than
2371 * inside pull_task().
2373 schedstat_add(sd, lb_gained[idle], pulled);
2376 *all_pinned = pinned;
2378 return max_load_move - rem_load_move;
2382 * move_tasks tries to move up to max_load_move weighted load from busiest to
2383 * this_rq, as part of a balancing operation within domain "sd".
2384 * Returns 1 if successful and 0 otherwise.
2386 * Called with both runqueues locked.
2388 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2389 unsigned long max_load_move,
2390 struct sched_domain *sd, enum cpu_idle_type idle,
2393 const struct sched_class *class = sched_class_highest;
2394 unsigned long total_load_moved = 0;
2395 int this_best_prio = this_rq->curr->prio;
2399 class->load_balance(this_rq, this_cpu, busiest,
2400 max_load_move - total_load_moved,
2401 sd, idle, all_pinned, &this_best_prio);
2402 class = class->next;
2403 } while (class && max_load_move > total_load_moved);
2405 return total_load_moved > 0;
2409 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2410 struct sched_domain *sd, enum cpu_idle_type idle,
2411 struct rq_iterator *iterator)
2413 struct task_struct *p = iterator->start(iterator->arg);
2417 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2418 pull_task(busiest, p, this_rq, this_cpu);
2420 * Right now, this is only the second place pull_task()
2421 * is called, so we can safely collect pull_task()
2422 * stats here rather than inside pull_task().
2424 schedstat_inc(sd, lb_gained[idle]);
2428 p = iterator->next(iterator->arg);
2435 * move_one_task tries to move exactly one task from busiest to this_rq, as
2436 * part of active balancing operations within "domain".
2437 * Returns 1 if successful and 0 otherwise.
2439 * Called with both runqueues locked.
2441 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2442 struct sched_domain *sd, enum cpu_idle_type idle)
2444 const struct sched_class *class;
2446 for (class = sched_class_highest; class; class = class->next)
2447 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2454 * find_busiest_group finds and returns the busiest CPU group within the
2455 * domain. It calculates and returns the amount of weighted load which
2456 * should be moved to restore balance via the imbalance parameter.
2458 static struct sched_group *
2459 find_busiest_group(struct sched_domain *sd, int this_cpu,
2460 unsigned long *imbalance, enum cpu_idle_type idle,
2461 int *sd_idle, cpumask_t *cpus, int *balance)
2463 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2464 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2465 unsigned long max_pull;
2466 unsigned long busiest_load_per_task, busiest_nr_running;
2467 unsigned long this_load_per_task, this_nr_running;
2468 int load_idx, group_imb = 0;
2469 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2470 int power_savings_balance = 1;
2471 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2472 unsigned long min_nr_running = ULONG_MAX;
2473 struct sched_group *group_min = NULL, *group_leader = NULL;
2476 max_load = this_load = total_load = total_pwr = 0;
2477 busiest_load_per_task = busiest_nr_running = 0;
2478 this_load_per_task = this_nr_running = 0;
2479 if (idle == CPU_NOT_IDLE)
2480 load_idx = sd->busy_idx;
2481 else if (idle == CPU_NEWLY_IDLE)
2482 load_idx = sd->newidle_idx;
2484 load_idx = sd->idle_idx;
2487 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2490 int __group_imb = 0;
2491 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2492 unsigned long sum_nr_running, sum_weighted_load;
2494 local_group = cpu_isset(this_cpu, group->cpumask);
2497 balance_cpu = first_cpu(group->cpumask);
2499 /* Tally up the load of all CPUs in the group */
2500 sum_weighted_load = sum_nr_running = avg_load = 0;
2502 min_cpu_load = ~0UL;
2504 for_each_cpu_mask(i, group->cpumask) {
2507 if (!cpu_isset(i, *cpus))
2512 if (*sd_idle && rq->nr_running)
2515 /* Bias balancing toward cpus of our domain */
2517 if (idle_cpu(i) && !first_idle_cpu) {
2522 load = target_load(i, load_idx);
2524 load = source_load(i, load_idx);
2525 if (load > max_cpu_load)
2526 max_cpu_load = load;
2527 if (min_cpu_load > load)
2528 min_cpu_load = load;
2532 sum_nr_running += rq->nr_running;
2533 sum_weighted_load += weighted_cpuload(i);
2537 * First idle cpu or the first cpu(busiest) in this sched group
2538 * is eligible for doing load balancing at this and above
2539 * domains. In the newly idle case, we will allow all the cpu's
2540 * to do the newly idle load balance.
2542 if (idle != CPU_NEWLY_IDLE && local_group &&
2543 balance_cpu != this_cpu && balance) {
2548 total_load += avg_load;
2549 total_pwr += group->__cpu_power;
2551 /* Adjust by relative CPU power of the group */
2552 avg_load = sg_div_cpu_power(group,
2553 avg_load * SCHED_LOAD_SCALE);
2555 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2558 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2561 this_load = avg_load;
2563 this_nr_running = sum_nr_running;
2564 this_load_per_task = sum_weighted_load;
2565 } else if (avg_load > max_load &&
2566 (sum_nr_running > group_capacity || __group_imb)) {
2567 max_load = avg_load;
2569 busiest_nr_running = sum_nr_running;
2570 busiest_load_per_task = sum_weighted_load;
2571 group_imb = __group_imb;
2574 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2576 * Busy processors will not participate in power savings
2579 if (idle == CPU_NOT_IDLE ||
2580 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2584 * If the local group is idle or completely loaded
2585 * no need to do power savings balance at this domain
2587 if (local_group && (this_nr_running >= group_capacity ||
2589 power_savings_balance = 0;
2592 * If a group is already running at full capacity or idle,
2593 * don't include that group in power savings calculations
2595 if (!power_savings_balance || sum_nr_running >= group_capacity
2600 * Calculate the group which has the least non-idle load.
2601 * This is the group from where we need to pick up the load
2604 if ((sum_nr_running < min_nr_running) ||
2605 (sum_nr_running == min_nr_running &&
2606 first_cpu(group->cpumask) <
2607 first_cpu(group_min->cpumask))) {
2609 min_nr_running = sum_nr_running;
2610 min_load_per_task = sum_weighted_load /
2615 * Calculate the group which is almost near its
2616 * capacity but still has some space to pick up some load
2617 * from other group and save more power
2619 if (sum_nr_running <= group_capacity - 1) {
2620 if (sum_nr_running > leader_nr_running ||
2621 (sum_nr_running == leader_nr_running &&
2622 first_cpu(group->cpumask) >
2623 first_cpu(group_leader->cpumask))) {
2624 group_leader = group;
2625 leader_nr_running = sum_nr_running;
2630 group = group->next;
2631 } while (group != sd->groups);
2633 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2636 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2638 if (this_load >= avg_load ||
2639 100*max_load <= sd->imbalance_pct*this_load)
2642 busiest_load_per_task /= busiest_nr_running;
2644 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2647 * We're trying to get all the cpus to the average_load, so we don't
2648 * want to push ourselves above the average load, nor do we wish to
2649 * reduce the max loaded cpu below the average load, as either of these
2650 * actions would just result in more rebalancing later, and ping-pong
2651 * tasks around. Thus we look for the minimum possible imbalance.
2652 * Negative imbalances (*we* are more loaded than anyone else) will
2653 * be counted as no imbalance for these purposes -- we can't fix that
2654 * by pulling tasks to us. Be careful of negative numbers as they'll
2655 * appear as very large values with unsigned longs.
2657 if (max_load <= busiest_load_per_task)
2661 * In the presence of smp nice balancing, certain scenarios can have
2662 * max load less than avg load(as we skip the groups at or below
2663 * its cpu_power, while calculating max_load..)
2665 if (max_load < avg_load) {
2667 goto small_imbalance;
2670 /* Don't want to pull so many tasks that a group would go idle */
2671 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2673 /* How much load to actually move to equalise the imbalance */
2674 *imbalance = min(max_pull * busiest->__cpu_power,
2675 (avg_load - this_load) * this->__cpu_power)
2679 * if *imbalance is less than the average load per runnable task
2680 * there is no gaurantee that any tasks will be moved so we'll have
2681 * a think about bumping its value to force at least one task to be
2684 if (*imbalance < busiest_load_per_task) {
2685 unsigned long tmp, pwr_now, pwr_move;
2689 pwr_move = pwr_now = 0;
2691 if (this_nr_running) {
2692 this_load_per_task /= this_nr_running;
2693 if (busiest_load_per_task > this_load_per_task)
2696 this_load_per_task = SCHED_LOAD_SCALE;
2698 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2699 busiest_load_per_task * imbn) {
2700 *imbalance = busiest_load_per_task;
2705 * OK, we don't have enough imbalance to justify moving tasks,
2706 * however we may be able to increase total CPU power used by
2710 pwr_now += busiest->__cpu_power *
2711 min(busiest_load_per_task, max_load);
2712 pwr_now += this->__cpu_power *
2713 min(this_load_per_task, this_load);
2714 pwr_now /= SCHED_LOAD_SCALE;
2716 /* Amount of load we'd subtract */
2717 tmp = sg_div_cpu_power(busiest,
2718 busiest_load_per_task * SCHED_LOAD_SCALE);
2720 pwr_move += busiest->__cpu_power *
2721 min(busiest_load_per_task, max_load - tmp);
2723 /* Amount of load we'd add */
2724 if (max_load * busiest->__cpu_power <
2725 busiest_load_per_task * SCHED_LOAD_SCALE)
2726 tmp = sg_div_cpu_power(this,
2727 max_load * busiest->__cpu_power);
2729 tmp = sg_div_cpu_power(this,
2730 busiest_load_per_task * SCHED_LOAD_SCALE);
2731 pwr_move += this->__cpu_power *
2732 min(this_load_per_task, this_load + tmp);
2733 pwr_move /= SCHED_LOAD_SCALE;
2735 /* Move if we gain throughput */
2736 if (pwr_move > pwr_now)
2737 *imbalance = busiest_load_per_task;
2743 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2744 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2747 if (this == group_leader && group_leader != group_min) {
2748 *imbalance = min_load_per_task;
2758 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2761 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2762 unsigned long imbalance, cpumask_t *cpus)
2764 struct rq *busiest = NULL, *rq;
2765 unsigned long max_load = 0;
2768 for_each_cpu_mask(i, group->cpumask) {
2771 if (!cpu_isset(i, *cpus))
2775 wl = weighted_cpuload(i);
2777 if (rq->nr_running == 1 && wl > imbalance)
2780 if (wl > max_load) {
2790 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2791 * so long as it is large enough.
2793 #define MAX_PINNED_INTERVAL 512
2796 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2797 * tasks if there is an imbalance.
2799 static int load_balance(int this_cpu, struct rq *this_rq,
2800 struct sched_domain *sd, enum cpu_idle_type idle,
2803 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2804 struct sched_group *group;
2805 unsigned long imbalance;
2807 cpumask_t cpus = CPU_MASK_ALL;
2808 unsigned long flags;
2811 * When power savings policy is enabled for the parent domain, idle
2812 * sibling can pick up load irrespective of busy siblings. In this case,
2813 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2814 * portraying it as CPU_NOT_IDLE.
2816 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2817 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2820 schedstat_inc(sd, lb_count[idle]);
2823 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2830 schedstat_inc(sd, lb_nobusyg[idle]);
2834 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2836 schedstat_inc(sd, lb_nobusyq[idle]);
2840 BUG_ON(busiest == this_rq);
2842 schedstat_add(sd, lb_imbalance[idle], imbalance);
2845 if (busiest->nr_running > 1) {
2847 * Attempt to move tasks. If find_busiest_group has found
2848 * an imbalance but busiest->nr_running <= 1, the group is
2849 * still unbalanced. ld_moved simply stays zero, so it is
2850 * correctly treated as an imbalance.
2852 local_irq_save(flags);
2853 double_rq_lock(this_rq, busiest);
2854 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2855 imbalance, sd, idle, &all_pinned);
2856 double_rq_unlock(this_rq, busiest);
2857 local_irq_restore(flags);
2860 * some other cpu did the load balance for us.
2862 if (ld_moved && this_cpu != smp_processor_id())
2863 resched_cpu(this_cpu);
2865 /* All tasks on this runqueue were pinned by CPU affinity */
2866 if (unlikely(all_pinned)) {
2867 cpu_clear(cpu_of(busiest), cpus);
2868 if (!cpus_empty(cpus))
2875 schedstat_inc(sd, lb_failed[idle]);
2876 sd->nr_balance_failed++;
2878 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2880 spin_lock_irqsave(&busiest->lock, flags);
2882 /* don't kick the migration_thread, if the curr
2883 * task on busiest cpu can't be moved to this_cpu
2885 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2886 spin_unlock_irqrestore(&busiest->lock, flags);
2888 goto out_one_pinned;
2891 if (!busiest->active_balance) {
2892 busiest->active_balance = 1;
2893 busiest->push_cpu = this_cpu;
2896 spin_unlock_irqrestore(&busiest->lock, flags);
2898 wake_up_process(busiest->migration_thread);
2901 * We've kicked active balancing, reset the failure
2904 sd->nr_balance_failed = sd->cache_nice_tries+1;
2907 sd->nr_balance_failed = 0;
2909 if (likely(!active_balance)) {
2910 /* We were unbalanced, so reset the balancing interval */
2911 sd->balance_interval = sd->min_interval;
2914 * If we've begun active balancing, start to back off. This
2915 * case may not be covered by the all_pinned logic if there
2916 * is only 1 task on the busy runqueue (because we don't call
2919 if (sd->balance_interval < sd->max_interval)
2920 sd->balance_interval *= 2;
2923 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2924 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2929 schedstat_inc(sd, lb_balanced[idle]);
2931 sd->nr_balance_failed = 0;
2934 /* tune up the balancing interval */
2935 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2936 (sd->balance_interval < sd->max_interval))
2937 sd->balance_interval *= 2;
2939 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2940 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2946 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2947 * tasks if there is an imbalance.
2949 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2950 * this_rq is locked.
2953 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2955 struct sched_group *group;
2956 struct rq *busiest = NULL;
2957 unsigned long imbalance;
2961 cpumask_t cpus = CPU_MASK_ALL;
2964 * When power savings policy is enabled for the parent domain, idle
2965 * sibling can pick up load irrespective of busy siblings. In this case,
2966 * let the state of idle sibling percolate up as IDLE, instead of
2967 * portraying it as CPU_NOT_IDLE.
2969 if (sd->flags & SD_SHARE_CPUPOWER &&
2970 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2973 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2975 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2976 &sd_idle, &cpus, NULL);
2978 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2982 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2985 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2989 BUG_ON(busiest == this_rq);
2991 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2994 if (busiest->nr_running > 1) {
2995 /* Attempt to move tasks */
2996 double_lock_balance(this_rq, busiest);
2997 /* this_rq->clock is already updated */
2998 update_rq_clock(busiest);
2999 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3000 imbalance, sd, CPU_NEWLY_IDLE,
3002 spin_unlock(&busiest->lock);
3004 if (unlikely(all_pinned)) {
3005 cpu_clear(cpu_of(busiest), cpus);
3006 if (!cpus_empty(cpus))
3012 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3013 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3014 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3017 sd->nr_balance_failed = 0;
3022 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3023 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3024 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3026 sd->nr_balance_failed = 0;
3032 * idle_balance is called by schedule() if this_cpu is about to become
3033 * idle. Attempts to pull tasks from other CPUs.
3035 static void idle_balance(int this_cpu, struct rq *this_rq)
3037 struct sched_domain *sd;
3038 int pulled_task = -1;
3039 unsigned long next_balance = jiffies + HZ;
3041 for_each_domain(this_cpu, sd) {
3042 unsigned long interval;
3044 if (!(sd->flags & SD_LOAD_BALANCE))
3047 if (sd->flags & SD_BALANCE_NEWIDLE)
3048 /* If we've pulled tasks over stop searching: */
3049 pulled_task = load_balance_newidle(this_cpu,
3052 interval = msecs_to_jiffies(sd->balance_interval);
3053 if (time_after(next_balance, sd->last_balance + interval))
3054 next_balance = sd->last_balance + interval;
3058 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3060 * We are going idle. next_balance may be set based on
3061 * a busy processor. So reset next_balance.
3063 this_rq->next_balance = next_balance;
3068 * active_load_balance is run by migration threads. It pushes running tasks
3069 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3070 * running on each physical CPU where possible, and avoids physical /
3071 * logical imbalances.
3073 * Called with busiest_rq locked.
3075 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3077 int target_cpu = busiest_rq->push_cpu;
3078 struct sched_domain *sd;
3079 struct rq *target_rq;
3081 /* Is there any task to move? */
3082 if (busiest_rq->nr_running <= 1)
3085 target_rq = cpu_rq(target_cpu);
3088 * This condition is "impossible", if it occurs
3089 * we need to fix it. Originally reported by
3090 * Bjorn Helgaas on a 128-cpu setup.
3092 BUG_ON(busiest_rq == target_rq);
3094 /* move a task from busiest_rq to target_rq */
3095 double_lock_balance(busiest_rq, target_rq);
3096 update_rq_clock(busiest_rq);
3097 update_rq_clock(target_rq);
3099 /* Search for an sd spanning us and the target CPU. */
3100 for_each_domain(target_cpu, sd) {
3101 if ((sd->flags & SD_LOAD_BALANCE) &&
3102 cpu_isset(busiest_cpu, sd->span))
3107 schedstat_inc(sd, alb_count);
3109 if (move_one_task(target_rq, target_cpu, busiest_rq,
3111 schedstat_inc(sd, alb_pushed);
3113 schedstat_inc(sd, alb_failed);
3115 spin_unlock(&target_rq->lock);
3120 atomic_t load_balancer;
3122 } nohz ____cacheline_aligned = {
3123 .load_balancer = ATOMIC_INIT(-1),
3124 .cpu_mask = CPU_MASK_NONE,
3128 * This routine will try to nominate the ilb (idle load balancing)
3129 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3130 * load balancing on behalf of all those cpus. If all the cpus in the system
3131 * go into this tickless mode, then there will be no ilb owner (as there is
3132 * no need for one) and all the cpus will sleep till the next wakeup event
3135 * For the ilb owner, tick is not stopped. And this tick will be used
3136 * for idle load balancing. ilb owner will still be part of
3139 * While stopping the tick, this cpu will become the ilb owner if there
3140 * is no other owner. And will be the owner till that cpu becomes busy
3141 * or if all cpus in the system stop their ticks at which point
3142 * there is no need for ilb owner.
3144 * When the ilb owner becomes busy, it nominates another owner, during the
3145 * next busy scheduler_tick()
3147 int select_nohz_load_balancer(int stop_tick)
3149 int cpu = smp_processor_id();
3152 cpu_set(cpu, nohz.cpu_mask);
3153 cpu_rq(cpu)->in_nohz_recently = 1;
3156 * If we are going offline and still the leader, give up!
3158 if (cpu_is_offline(cpu) &&
3159 atomic_read(&nohz.load_balancer) == cpu) {
3160 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3165 /* time for ilb owner also to sleep */
3166 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3167 if (atomic_read(&nohz.load_balancer) == cpu)
3168 atomic_set(&nohz.load_balancer, -1);
3172 if (atomic_read(&nohz.load_balancer) == -1) {
3173 /* make me the ilb owner */
3174 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3176 } else if (atomic_read(&nohz.load_balancer) == cpu)
3179 if (!cpu_isset(cpu, nohz.cpu_mask))
3182 cpu_clear(cpu, nohz.cpu_mask);
3184 if (atomic_read(&nohz.load_balancer) == cpu)
3185 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3192 static DEFINE_SPINLOCK(balancing);
3195 * It checks each scheduling domain to see if it is due to be balanced,
3196 * and initiates a balancing operation if so.
3198 * Balancing parameters are set up in arch_init_sched_domains.
3200 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3203 struct rq *rq = cpu_rq(cpu);
3204 unsigned long interval;
3205 struct sched_domain *sd;
3206 /* Earliest time when we have to do rebalance again */
3207 unsigned long next_balance = jiffies + 60*HZ;
3208 int update_next_balance = 0;
3210 for_each_domain(cpu, sd) {
3211 if (!(sd->flags & SD_LOAD_BALANCE))
3214 interval = sd->balance_interval;
3215 if (idle != CPU_IDLE)
3216 interval *= sd->busy_factor;
3218 /* scale ms to jiffies */
3219 interval = msecs_to_jiffies(interval);
3220 if (unlikely(!interval))
3222 if (interval > HZ*NR_CPUS/10)
3223 interval = HZ*NR_CPUS/10;
3226 if (sd->flags & SD_SERIALIZE) {
3227 if (!spin_trylock(&balancing))
3231 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3232 if (load_balance(cpu, rq, sd, idle, &balance)) {
3234 * We've pulled tasks over so either we're no
3235 * longer idle, or one of our SMT siblings is
3238 idle = CPU_NOT_IDLE;
3240 sd->last_balance = jiffies;
3242 if (sd->flags & SD_SERIALIZE)
3243 spin_unlock(&balancing);
3245 if (time_after(next_balance, sd->last_balance + interval)) {
3246 next_balance = sd->last_balance + interval;
3247 update_next_balance = 1;
3251 * Stop the load balance at this level. There is another
3252 * CPU in our sched group which is doing load balancing more
3260 * next_balance will be updated only when there is a need.
3261 * When the cpu is attached to null domain for ex, it will not be
3264 if (likely(update_next_balance))
3265 rq->next_balance = next_balance;
3269 * run_rebalance_domains is triggered when needed from the scheduler tick.
3270 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3271 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3273 static void run_rebalance_domains(struct softirq_action *h)
3275 int this_cpu = smp_processor_id();
3276 struct rq *this_rq = cpu_rq(this_cpu);
3277 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3278 CPU_IDLE : CPU_NOT_IDLE;
3280 rebalance_domains(this_cpu, idle);
3284 * If this cpu is the owner for idle load balancing, then do the
3285 * balancing on behalf of the other idle cpus whose ticks are
3288 if (this_rq->idle_at_tick &&
3289 atomic_read(&nohz.load_balancer) == this_cpu) {
3290 cpumask_t cpus = nohz.cpu_mask;
3294 cpu_clear(this_cpu, cpus);
3295 for_each_cpu_mask(balance_cpu, cpus) {
3297 * If this cpu gets work to do, stop the load balancing
3298 * work being done for other cpus. Next load
3299 * balancing owner will pick it up.
3304 rebalance_domains(balance_cpu, CPU_IDLE);
3306 rq = cpu_rq(balance_cpu);
3307 if (time_after(this_rq->next_balance, rq->next_balance))
3308 this_rq->next_balance = rq->next_balance;
3315 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3317 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3318 * idle load balancing owner or decide to stop the periodic load balancing,
3319 * if the whole system is idle.
3321 static inline void trigger_load_balance(struct rq *rq, int cpu)
3325 * If we were in the nohz mode recently and busy at the current
3326 * scheduler tick, then check if we need to nominate new idle
3329 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3330 rq->in_nohz_recently = 0;
3332 if (atomic_read(&nohz.load_balancer) == cpu) {
3333 cpu_clear(cpu, nohz.cpu_mask);
3334 atomic_set(&nohz.load_balancer, -1);
3337 if (atomic_read(&nohz.load_balancer) == -1) {
3339 * simple selection for now: Nominate the
3340 * first cpu in the nohz list to be the next
3343 * TBD: Traverse the sched domains and nominate
3344 * the nearest cpu in the nohz.cpu_mask.
3346 int ilb = first_cpu(nohz.cpu_mask);
3354 * If this cpu is idle and doing idle load balancing for all the
3355 * cpus with ticks stopped, is it time for that to stop?
3357 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3358 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3364 * If this cpu is idle and the idle load balancing is done by
3365 * someone else, then no need raise the SCHED_SOFTIRQ
3367 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3368 cpu_isset(cpu, nohz.cpu_mask))
3371 if (time_after_eq(jiffies, rq->next_balance))
3372 raise_softirq(SCHED_SOFTIRQ);
3375 #else /* CONFIG_SMP */
3378 * on UP we do not need to balance between CPUs:
3380 static inline void idle_balance(int cpu, struct rq *rq)
3386 DEFINE_PER_CPU(struct kernel_stat, kstat);
3388 EXPORT_PER_CPU_SYMBOL(kstat);
3391 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3392 * that have not yet been banked in case the task is currently running.
3394 unsigned long long task_sched_runtime(struct task_struct *p)
3396 unsigned long flags;
3400 rq = task_rq_lock(p, &flags);
3401 ns = p->se.sum_exec_runtime;
3402 if (task_current(rq, p)) {
3403 update_rq_clock(rq);
3404 delta_exec = rq->clock - p->se.exec_start;
3405 if ((s64)delta_exec > 0)
3408 task_rq_unlock(rq, &flags);
3414 * Account user cpu time to a process.
3415 * @p: the process that the cpu time gets accounted to
3416 * @cputime: the cpu time spent in user space since the last update
3418 void account_user_time(struct task_struct *p, cputime_t cputime)
3420 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3423 p->utime = cputime_add(p->utime, cputime);
3425 /* Add user time to cpustat. */
3426 tmp = cputime_to_cputime64(cputime);
3427 if (TASK_NICE(p) > 0)
3428 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3430 cpustat->user = cputime64_add(cpustat->user, tmp);
3434 * Account guest cpu time to a process.
3435 * @p: the process that the cpu time gets accounted to
3436 * @cputime: the cpu time spent in virtual machine since the last update
3438 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3441 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3443 tmp = cputime_to_cputime64(cputime);
3445 p->utime = cputime_add(p->utime, cputime);
3446 p->gtime = cputime_add(p->gtime, cputime);
3448 cpustat->user = cputime64_add(cpustat->user, tmp);
3449 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3453 * Account scaled user cpu time to a process.
3454 * @p: the process that the cpu time gets accounted to
3455 * @cputime: the cpu time spent in user space since the last update
3457 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3459 p->utimescaled = cputime_add(p->utimescaled, cputime);
3463 * Account system cpu time to a process.
3464 * @p: the process that the cpu time gets accounted to
3465 * @hardirq_offset: the offset to subtract from hardirq_count()
3466 * @cputime: the cpu time spent in kernel space since the last update
3468 void account_system_time(struct task_struct *p, int hardirq_offset,
3471 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3472 struct rq *rq = this_rq();
3475 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3476 return account_guest_time(p, cputime);
3478 p->stime = cputime_add(p->stime, cputime);
3480 /* Add system time to cpustat. */
3481 tmp = cputime_to_cputime64(cputime);
3482 if (hardirq_count() - hardirq_offset)
3483 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3484 else if (softirq_count())
3485 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3486 else if (p != rq->idle)
3487 cpustat->system = cputime64_add(cpustat->system, tmp);
3488 else if (atomic_read(&rq->nr_iowait) > 0)
3489 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3491 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3492 /* Account for system time used */
3493 acct_update_integrals(p);
3497 * Account scaled system cpu time to a process.
3498 * @p: the process that the cpu time gets accounted to
3499 * @hardirq_offset: the offset to subtract from hardirq_count()
3500 * @cputime: the cpu time spent in kernel space since the last update
3502 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3504 p->stimescaled = cputime_add(p->stimescaled, cputime);
3508 * Account for involuntary wait time.
3509 * @p: the process from which the cpu time has been stolen
3510 * @steal: the cpu time spent in involuntary wait
3512 void account_steal_time(struct task_struct *p, cputime_t steal)
3514 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3515 cputime64_t tmp = cputime_to_cputime64(steal);
3516 struct rq *rq = this_rq();
3518 if (p == rq->idle) {
3519 p->stime = cputime_add(p->stime, steal);
3520 if (atomic_read(&rq->nr_iowait) > 0)
3521 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3523 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3525 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3529 * This function gets called by the timer code, with HZ frequency.
3530 * We call it with interrupts disabled.
3532 * It also gets called by the fork code, when changing the parent's
3535 void scheduler_tick(void)
3537 int cpu = smp_processor_id();
3538 struct rq *rq = cpu_rq(cpu);
3539 struct task_struct *curr = rq->curr;
3540 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3542 spin_lock(&rq->lock);
3543 __update_rq_clock(rq);
3545 * Let rq->clock advance by at least TICK_NSEC:
3547 if (unlikely(rq->clock < next_tick))
3548 rq->clock = next_tick;
3549 rq->tick_timestamp = rq->clock;
3550 update_cpu_load(rq);
3551 if (curr != rq->idle) /* FIXME: needed? */
3552 curr->sched_class->task_tick(rq, curr);
3553 spin_unlock(&rq->lock);
3556 rq->idle_at_tick = idle_cpu(cpu);
3557 trigger_load_balance(rq, cpu);
3561 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3563 void fastcall add_preempt_count(int val)
3568 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3570 preempt_count() += val;
3572 * Spinlock count overflowing soon?
3574 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3577 EXPORT_SYMBOL(add_preempt_count);
3579 void fastcall sub_preempt_count(int val)
3584 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3587 * Is the spinlock portion underflowing?
3589 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3590 !(preempt_count() & PREEMPT_MASK)))
3593 preempt_count() -= val;
3595 EXPORT_SYMBOL(sub_preempt_count);
3600 * Print scheduling while atomic bug:
3602 static noinline void __schedule_bug(struct task_struct *prev)
3604 struct pt_regs *regs = get_irq_regs();
3606 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3607 prev->comm, prev->pid, preempt_count());
3609 debug_show_held_locks(prev);
3610 if (irqs_disabled())
3611 print_irqtrace_events(prev);
3620 * Various schedule()-time debugging checks and statistics:
3622 static inline void schedule_debug(struct task_struct *prev)
3625 * Test if we are atomic. Since do_exit() needs to call into
3626 * schedule() atomically, we ignore that path for now.
3627 * Otherwise, whine if we are scheduling when we should not be.
3629 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3630 __schedule_bug(prev);
3632 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3634 schedstat_inc(this_rq(), sched_count);
3635 #ifdef CONFIG_SCHEDSTATS
3636 if (unlikely(prev->lock_depth >= 0)) {
3637 schedstat_inc(this_rq(), bkl_count);
3638 schedstat_inc(prev, sched_info.bkl_count);
3644 * Pick up the highest-prio task:
3646 static inline struct task_struct *
3647 pick_next_task(struct rq *rq, struct task_struct *prev)
3649 const struct sched_class *class;
3650 struct task_struct *p;
3653 * Optimization: we know that if all tasks are in
3654 * the fair class we can call that function directly:
3656 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3657 p = fair_sched_class.pick_next_task(rq);
3662 class = sched_class_highest;
3664 p = class->pick_next_task(rq);
3668 * Will never be NULL as the idle class always
3669 * returns a non-NULL p:
3671 class = class->next;
3676 * schedule() is the main scheduler function.
3678 asmlinkage void __sched schedule(void)
3680 struct task_struct *prev, *next;
3687 cpu = smp_processor_id();
3691 switch_count = &prev->nivcsw;
3693 release_kernel_lock(prev);
3694 need_resched_nonpreemptible:
3696 schedule_debug(prev);
3699 * Do the rq-clock update outside the rq lock:
3701 local_irq_disable();
3702 __update_rq_clock(rq);
3703 spin_lock(&rq->lock);
3704 clear_tsk_need_resched(prev);
3706 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3707 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3708 unlikely(signal_pending(prev)))) {
3709 prev->state = TASK_RUNNING;
3711 deactivate_task(rq, prev, 1);
3713 switch_count = &prev->nvcsw;
3716 if (unlikely(!rq->nr_running))
3717 idle_balance(cpu, rq);
3719 prev->sched_class->put_prev_task(rq, prev);
3720 next = pick_next_task(rq, prev);
3722 sched_info_switch(prev, next);
3724 if (likely(prev != next)) {
3729 context_switch(rq, prev, next); /* unlocks the rq */
3731 spin_unlock_irq(&rq->lock);
3733 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3734 cpu = smp_processor_id();
3736 goto need_resched_nonpreemptible;
3738 preempt_enable_no_resched();
3739 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3742 EXPORT_SYMBOL(schedule);
3744 #ifdef CONFIG_PREEMPT
3746 * this is the entry point to schedule() from in-kernel preemption
3747 * off of preempt_enable. Kernel preemptions off return from interrupt
3748 * occur there and call schedule directly.
3750 asmlinkage void __sched preempt_schedule(void)
3752 struct thread_info *ti = current_thread_info();
3753 #ifdef CONFIG_PREEMPT_BKL
3754 struct task_struct *task = current;
3755 int saved_lock_depth;
3758 * If there is a non-zero preempt_count or interrupts are disabled,
3759 * we do not want to preempt the current task. Just return..
3761 if (likely(ti->preempt_count || irqs_disabled()))
3765 add_preempt_count(PREEMPT_ACTIVE);
3768 * We keep the big kernel semaphore locked, but we
3769 * clear ->lock_depth so that schedule() doesnt
3770 * auto-release the semaphore:
3772 #ifdef CONFIG_PREEMPT_BKL
3773 saved_lock_depth = task->lock_depth;
3774 task->lock_depth = -1;
3777 #ifdef CONFIG_PREEMPT_BKL
3778 task->lock_depth = saved_lock_depth;
3780 sub_preempt_count(PREEMPT_ACTIVE);
3783 * Check again in case we missed a preemption opportunity
3784 * between schedule and now.
3787 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3789 EXPORT_SYMBOL(preempt_schedule);
3792 * this is the entry point to schedule() from kernel preemption
3793 * off of irq context.
3794 * Note, that this is called and return with irqs disabled. This will
3795 * protect us against recursive calling from irq.
3797 asmlinkage void __sched preempt_schedule_irq(void)
3799 struct thread_info *ti = current_thread_info();
3800 #ifdef CONFIG_PREEMPT_BKL
3801 struct task_struct *task = current;
3802 int saved_lock_depth;
3804 /* Catch callers which need to be fixed */
3805 BUG_ON(ti->preempt_count || !irqs_disabled());
3808 add_preempt_count(PREEMPT_ACTIVE);
3811 * We keep the big kernel semaphore locked, but we
3812 * clear ->lock_depth so that schedule() doesnt
3813 * auto-release the semaphore:
3815 #ifdef CONFIG_PREEMPT_BKL
3816 saved_lock_depth = task->lock_depth;
3817 task->lock_depth = -1;
3821 local_irq_disable();
3822 #ifdef CONFIG_PREEMPT_BKL
3823 task->lock_depth = saved_lock_depth;
3825 sub_preempt_count(PREEMPT_ACTIVE);
3828 * Check again in case we missed a preemption opportunity
3829 * between schedule and now.
3832 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3835 #endif /* CONFIG_PREEMPT */
3837 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3840 return try_to_wake_up(curr->private, mode, sync);
3842 EXPORT_SYMBOL(default_wake_function);
3845 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3846 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3847 * number) then we wake all the non-exclusive tasks and one exclusive task.
3849 * There are circumstances in which we can try to wake a task which has already
3850 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3851 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3853 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3854 int nr_exclusive, int sync, void *key)
3856 wait_queue_t *curr, *next;
3858 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3859 unsigned flags = curr->flags;
3861 if (curr->func(curr, mode, sync, key) &&
3862 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3868 * __wake_up - wake up threads blocked on a waitqueue.
3870 * @mode: which threads
3871 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3872 * @key: is directly passed to the wakeup function
3874 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3875 int nr_exclusive, void *key)
3877 unsigned long flags;
3879 spin_lock_irqsave(&q->lock, flags);
3880 __wake_up_common(q, mode, nr_exclusive, 0, key);
3881 spin_unlock_irqrestore(&q->lock, flags);
3883 EXPORT_SYMBOL(__wake_up);
3886 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3888 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3890 __wake_up_common(q, mode, 1, 0, NULL);
3894 * __wake_up_sync - wake up threads blocked on a waitqueue.
3896 * @mode: which threads
3897 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3899 * The sync wakeup differs that the waker knows that it will schedule
3900 * away soon, so while the target thread will be woken up, it will not
3901 * be migrated to another CPU - ie. the two threads are 'synchronized'
3902 * with each other. This can prevent needless bouncing between CPUs.
3904 * On UP it can prevent extra preemption.
3907 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3909 unsigned long flags;
3915 if (unlikely(!nr_exclusive))
3918 spin_lock_irqsave(&q->lock, flags);
3919 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3920 spin_unlock_irqrestore(&q->lock, flags);
3922 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3924 void complete(struct completion *x)
3926 unsigned long flags;
3928 spin_lock_irqsave(&x->wait.lock, flags);
3930 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3932 spin_unlock_irqrestore(&x->wait.lock, flags);
3934 EXPORT_SYMBOL(complete);
3936 void complete_all(struct completion *x)
3938 unsigned long flags;
3940 spin_lock_irqsave(&x->wait.lock, flags);
3941 x->done += UINT_MAX/2;
3942 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3944 spin_unlock_irqrestore(&x->wait.lock, flags);
3946 EXPORT_SYMBOL(complete_all);
3948 static inline long __sched
3949 do_wait_for_common(struct completion *x, long timeout, int state)
3952 DECLARE_WAITQUEUE(wait, current);
3954 wait.flags |= WQ_FLAG_EXCLUSIVE;
3955 __add_wait_queue_tail(&x->wait, &wait);
3957 if (state == TASK_INTERRUPTIBLE &&
3958 signal_pending(current)) {
3959 __remove_wait_queue(&x->wait, &wait);
3960 return -ERESTARTSYS;
3962 __set_current_state(state);
3963 spin_unlock_irq(&x->wait.lock);
3964 timeout = schedule_timeout(timeout);
3965 spin_lock_irq(&x->wait.lock);
3967 __remove_wait_queue(&x->wait, &wait);
3971 __remove_wait_queue(&x->wait, &wait);
3978 wait_for_common(struct completion *x, long timeout, int state)
3982 spin_lock_irq(&x->wait.lock);
3983 timeout = do_wait_for_common(x, timeout, state);
3984 spin_unlock_irq(&x->wait.lock);
3988 void __sched wait_for_completion(struct completion *x)
3990 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3992 EXPORT_SYMBOL(wait_for_completion);
3994 unsigned long __sched
3995 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3997 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3999 EXPORT_SYMBOL(wait_for_completion_timeout);
4001 int __sched wait_for_completion_interruptible(struct completion *x)
4003 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4004 if (t == -ERESTARTSYS)
4008 EXPORT_SYMBOL(wait_for_completion_interruptible);
4010 unsigned long __sched
4011 wait_for_completion_interruptible_timeout(struct completion *x,
4012 unsigned long timeout)
4014 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4016 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4019 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4021 unsigned long flags;
4024 init_waitqueue_entry(&wait, current);
4026 __set_current_state(state);
4028 spin_lock_irqsave(&q->lock, flags);
4029 __add_wait_queue(q, &wait);
4030 spin_unlock(&q->lock);
4031 timeout = schedule_timeout(timeout);
4032 spin_lock_irq(&q->lock);
4033 __remove_wait_queue(q, &wait);
4034 spin_unlock_irqrestore(&q->lock, flags);
4039 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4041 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4043 EXPORT_SYMBOL(interruptible_sleep_on);
4046 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4048 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4050 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4052 void __sched sleep_on(wait_queue_head_t *q)
4054 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4056 EXPORT_SYMBOL(sleep_on);
4058 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4060 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4062 EXPORT_SYMBOL(sleep_on_timeout);
4064 #ifdef CONFIG_RT_MUTEXES
4067 * rt_mutex_setprio - set the current priority of a task
4069 * @prio: prio value (kernel-internal form)
4071 * This function changes the 'effective' priority of a task. It does
4072 * not touch ->normal_prio like __setscheduler().
4074 * Used by the rt_mutex code to implement priority inheritance logic.
4076 void rt_mutex_setprio(struct task_struct *p, int prio)
4078 unsigned long flags;
4079 int oldprio, on_rq, running;
4082 BUG_ON(prio < 0 || prio > MAX_PRIO);
4084 rq = task_rq_lock(p, &flags);
4085 update_rq_clock(rq);
4088 on_rq = p->se.on_rq;
4089 running = task_current(rq, p);
4091 dequeue_task(rq, p, 0);
4093 p->sched_class->put_prev_task(rq, p);
4097 p->sched_class = &rt_sched_class;
4099 p->sched_class = &fair_sched_class;
4105 p->sched_class->set_curr_task(rq);
4106 enqueue_task(rq, p, 0);
4108 * Reschedule if we are currently running on this runqueue and
4109 * our priority decreased, or if we are not currently running on
4110 * this runqueue and our priority is higher than the current's
4113 if (p->prio > oldprio)
4114 resched_task(rq->curr);
4116 check_preempt_curr(rq, p);
4119 task_rq_unlock(rq, &flags);
4124 void set_user_nice(struct task_struct *p, long nice)
4126 int old_prio, delta, on_rq;
4127 unsigned long flags;
4130 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4133 * We have to be careful, if called from sys_setpriority(),
4134 * the task might be in the middle of scheduling on another CPU.
4136 rq = task_rq_lock(p, &flags);
4137 update_rq_clock(rq);
4139 * The RT priorities are set via sched_setscheduler(), but we still
4140 * allow the 'normal' nice value to be set - but as expected
4141 * it wont have any effect on scheduling until the task is
4142 * SCHED_FIFO/SCHED_RR:
4144 if (task_has_rt_policy(p)) {
4145 p->static_prio = NICE_TO_PRIO(nice);
4148 on_rq = p->se.on_rq;
4150 dequeue_task(rq, p, 0);
4152 p->static_prio = NICE_TO_PRIO(nice);
4155 p->prio = effective_prio(p);
4156 delta = p->prio - old_prio;
4159 enqueue_task(rq, p, 0);
4161 * If the task increased its priority or is running and
4162 * lowered its priority, then reschedule its CPU:
4164 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4165 resched_task(rq->curr);
4168 task_rq_unlock(rq, &flags);
4170 EXPORT_SYMBOL(set_user_nice);
4173 * can_nice - check if a task can reduce its nice value
4177 int can_nice(const struct task_struct *p, const int nice)
4179 /* convert nice value [19,-20] to rlimit style value [1,40] */
4180 int nice_rlim = 20 - nice;
4182 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4183 capable(CAP_SYS_NICE));
4186 #ifdef __ARCH_WANT_SYS_NICE
4189 * sys_nice - change the priority of the current process.
4190 * @increment: priority increment
4192 * sys_setpriority is a more generic, but much slower function that
4193 * does similar things.
4195 asmlinkage long sys_nice(int increment)
4200 * Setpriority might change our priority at the same moment.
4201 * We don't have to worry. Conceptually one call occurs first
4202 * and we have a single winner.
4204 if (increment < -40)
4209 nice = PRIO_TO_NICE(current->static_prio) + increment;
4215 if (increment < 0 && !can_nice(current, nice))
4218 retval = security_task_setnice(current, nice);
4222 set_user_nice(current, nice);
4229 * task_prio - return the priority value of a given task.
4230 * @p: the task in question.
4232 * This is the priority value as seen by users in /proc.
4233 * RT tasks are offset by -200. Normal tasks are centered
4234 * around 0, value goes from -16 to +15.
4236 int task_prio(const struct task_struct *p)
4238 return p->prio - MAX_RT_PRIO;
4242 * task_nice - return the nice value of a given task.
4243 * @p: the task in question.
4245 int task_nice(const struct task_struct *p)
4247 return TASK_NICE(p);
4249 EXPORT_SYMBOL_GPL(task_nice);
4252 * idle_cpu - is a given cpu idle currently?
4253 * @cpu: the processor in question.
4255 int idle_cpu(int cpu)
4257 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4261 * idle_task - return the idle task for a given cpu.
4262 * @cpu: the processor in question.
4264 struct task_struct *idle_task(int cpu)
4266 return cpu_rq(cpu)->idle;
4270 * find_process_by_pid - find a process with a matching PID value.
4271 * @pid: the pid in question.
4273 static struct task_struct *find_process_by_pid(pid_t pid)
4275 return pid ? find_task_by_vpid(pid) : current;
4278 /* Actually do priority change: must hold rq lock. */
4280 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4282 BUG_ON(p->se.on_rq);
4285 switch (p->policy) {
4289 p->sched_class = &fair_sched_class;
4293 p->sched_class = &rt_sched_class;
4297 p->rt_priority = prio;
4298 p->normal_prio = normal_prio(p);
4299 /* we are holding p->pi_lock already */
4300 p->prio = rt_mutex_getprio(p);
4305 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4306 * @p: the task in question.
4307 * @policy: new policy.
4308 * @param: structure containing the new RT priority.
4310 * NOTE that the task may be already dead.
4312 int sched_setscheduler(struct task_struct *p, int policy,
4313 struct sched_param *param)
4315 int retval, oldprio, oldpolicy = -1, on_rq, running;
4316 unsigned long flags;
4319 /* may grab non-irq protected spin_locks */
4320 BUG_ON(in_interrupt());
4322 /* double check policy once rq lock held */
4324 policy = oldpolicy = p->policy;
4325 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4326 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4327 policy != SCHED_IDLE)
4330 * Valid priorities for SCHED_FIFO and SCHED_RR are
4331 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4332 * SCHED_BATCH and SCHED_IDLE is 0.
4334 if (param->sched_priority < 0 ||
4335 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4336 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4338 if (rt_policy(policy) != (param->sched_priority != 0))
4342 * Allow unprivileged RT tasks to decrease priority:
4344 if (!capable(CAP_SYS_NICE)) {
4345 if (rt_policy(policy)) {
4346 unsigned long rlim_rtprio;
4348 if (!lock_task_sighand(p, &flags))
4350 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4351 unlock_task_sighand(p, &flags);
4353 /* can't set/change the rt policy */
4354 if (policy != p->policy && !rlim_rtprio)
4357 /* can't increase priority */
4358 if (param->sched_priority > p->rt_priority &&
4359 param->sched_priority > rlim_rtprio)
4363 * Like positive nice levels, dont allow tasks to
4364 * move out of SCHED_IDLE either:
4366 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4369 /* can't change other user's priorities */
4370 if ((current->euid != p->euid) &&
4371 (current->euid != p->uid))
4375 retval = security_task_setscheduler(p, policy, param);
4379 * make sure no PI-waiters arrive (or leave) while we are
4380 * changing the priority of the task:
4382 spin_lock_irqsave(&p->pi_lock, flags);
4384 * To be able to change p->policy safely, the apropriate
4385 * runqueue lock must be held.
4387 rq = __task_rq_lock(p);
4388 /* recheck policy now with rq lock held */
4389 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4390 policy = oldpolicy = -1;
4391 __task_rq_unlock(rq);
4392 spin_unlock_irqrestore(&p->pi_lock, flags);
4395 update_rq_clock(rq);
4396 on_rq = p->se.on_rq;
4397 running = task_current(rq, p);
4399 deactivate_task(rq, p, 0);
4401 p->sched_class->put_prev_task(rq, p);
4405 __setscheduler(rq, p, policy, param->sched_priority);
4409 p->sched_class->set_curr_task(rq);
4410 activate_task(rq, p, 0);
4412 * Reschedule if we are currently running on this runqueue and
4413 * our priority decreased, or if we are not currently running on
4414 * this runqueue and our priority is higher than the current's
4417 if (p->prio > oldprio)
4418 resched_task(rq->curr);
4420 check_preempt_curr(rq, p);
4423 __task_rq_unlock(rq);
4424 spin_unlock_irqrestore(&p->pi_lock, flags);
4426 rt_mutex_adjust_pi(p);
4430 EXPORT_SYMBOL_GPL(sched_setscheduler);
4433 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4435 struct sched_param lparam;
4436 struct task_struct *p;
4439 if (!param || pid < 0)
4441 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4446 p = find_process_by_pid(pid);
4448 retval = sched_setscheduler(p, policy, &lparam);
4455 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4456 * @pid: the pid in question.
4457 * @policy: new policy.
4458 * @param: structure containing the new RT priority.
4461 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4463 /* negative values for policy are not valid */
4467 return do_sched_setscheduler(pid, policy, param);
4471 * sys_sched_setparam - set/change the RT priority of a thread
4472 * @pid: the pid in question.
4473 * @param: structure containing the new RT priority.
4475 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4477 return do_sched_setscheduler(pid, -1, param);
4481 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4482 * @pid: the pid in question.
4484 asmlinkage long sys_sched_getscheduler(pid_t pid)
4486 struct task_struct *p;
4493 read_lock(&tasklist_lock);
4494 p = find_process_by_pid(pid);
4496 retval = security_task_getscheduler(p);
4500 read_unlock(&tasklist_lock);
4505 * sys_sched_getscheduler - get the RT priority of a thread
4506 * @pid: the pid in question.
4507 * @param: structure containing the RT priority.
4509 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4511 struct sched_param lp;
4512 struct task_struct *p;
4515 if (!param || pid < 0)
4518 read_lock(&tasklist_lock);
4519 p = find_process_by_pid(pid);
4524 retval = security_task_getscheduler(p);
4528 lp.sched_priority = p->rt_priority;
4529 read_unlock(&tasklist_lock);
4532 * This one might sleep, we cannot do it with a spinlock held ...
4534 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4539 read_unlock(&tasklist_lock);
4543 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4545 cpumask_t cpus_allowed;
4546 struct task_struct *p;
4550 read_lock(&tasklist_lock);
4552 p = find_process_by_pid(pid);
4554 read_unlock(&tasklist_lock);
4560 * It is not safe to call set_cpus_allowed with the
4561 * tasklist_lock held. We will bump the task_struct's
4562 * usage count and then drop tasklist_lock.
4565 read_unlock(&tasklist_lock);
4568 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4569 !capable(CAP_SYS_NICE))
4572 retval = security_task_setscheduler(p, 0, NULL);
4576 cpus_allowed = cpuset_cpus_allowed(p);
4577 cpus_and(new_mask, new_mask, cpus_allowed);
4579 retval = set_cpus_allowed(p, new_mask);
4582 cpus_allowed = cpuset_cpus_allowed(p);
4583 if (!cpus_subset(new_mask, cpus_allowed)) {
4585 * We must have raced with a concurrent cpuset
4586 * update. Just reset the cpus_allowed to the
4587 * cpuset's cpus_allowed
4589 new_mask = cpus_allowed;
4599 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4600 cpumask_t *new_mask)
4602 if (len < sizeof(cpumask_t)) {
4603 memset(new_mask, 0, sizeof(cpumask_t));
4604 } else if (len > sizeof(cpumask_t)) {
4605 len = sizeof(cpumask_t);
4607 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4611 * sys_sched_setaffinity - set the cpu affinity of a process
4612 * @pid: pid of the process
4613 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4614 * @user_mask_ptr: user-space pointer to the new cpu mask
4616 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4617 unsigned long __user *user_mask_ptr)
4622 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4626 return sched_setaffinity(pid, new_mask);
4630 * Represents all cpu's present in the system
4631 * In systems capable of hotplug, this map could dynamically grow
4632 * as new cpu's are detected in the system via any platform specific
4633 * method, such as ACPI for e.g.
4636 cpumask_t cpu_present_map __read_mostly;
4637 EXPORT_SYMBOL(cpu_present_map);
4640 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4641 EXPORT_SYMBOL(cpu_online_map);
4643 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4644 EXPORT_SYMBOL(cpu_possible_map);
4647 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4649 struct task_struct *p;
4653 read_lock(&tasklist_lock);
4656 p = find_process_by_pid(pid);
4660 retval = security_task_getscheduler(p);
4664 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4667 read_unlock(&tasklist_lock);
4674 * sys_sched_getaffinity - get the cpu affinity of a process
4675 * @pid: pid of the process
4676 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4677 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4679 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4680 unsigned long __user *user_mask_ptr)
4685 if (len < sizeof(cpumask_t))
4688 ret = sched_getaffinity(pid, &mask);
4692 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4695 return sizeof(cpumask_t);
4699 * sys_sched_yield - yield the current processor to other threads.
4701 * This function yields the current CPU to other tasks. If there are no
4702 * other threads running on this CPU then this function will return.
4704 asmlinkage long sys_sched_yield(void)
4706 struct rq *rq = this_rq_lock();
4708 schedstat_inc(rq, yld_count);
4709 current->sched_class->yield_task(rq);
4712 * Since we are going to call schedule() anyway, there's
4713 * no need to preempt or enable interrupts:
4715 __release(rq->lock);
4716 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4717 _raw_spin_unlock(&rq->lock);
4718 preempt_enable_no_resched();
4725 static void __cond_resched(void)
4727 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4728 __might_sleep(__FILE__, __LINE__);
4731 * The BKS might be reacquired before we have dropped
4732 * PREEMPT_ACTIVE, which could trigger a second
4733 * cond_resched() call.
4736 add_preempt_count(PREEMPT_ACTIVE);
4738 sub_preempt_count(PREEMPT_ACTIVE);
4739 } while (need_resched());
4742 int __sched cond_resched(void)
4744 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4745 system_state == SYSTEM_RUNNING) {
4751 EXPORT_SYMBOL(cond_resched);
4754 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4755 * call schedule, and on return reacquire the lock.
4757 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4758 * operations here to prevent schedule() from being called twice (once via
4759 * spin_unlock(), once by hand).
4761 int cond_resched_lock(spinlock_t *lock)
4765 if (need_lockbreak(lock)) {
4771 if (need_resched() && system_state == SYSTEM_RUNNING) {
4772 spin_release(&lock->dep_map, 1, _THIS_IP_);
4773 _raw_spin_unlock(lock);
4774 preempt_enable_no_resched();
4781 EXPORT_SYMBOL(cond_resched_lock);
4783 int __sched cond_resched_softirq(void)
4785 BUG_ON(!in_softirq());
4787 if (need_resched() && system_state == SYSTEM_RUNNING) {
4795 EXPORT_SYMBOL(cond_resched_softirq);
4798 * yield - yield the current processor to other threads.
4800 * This is a shortcut for kernel-space yielding - it marks the
4801 * thread runnable and calls sys_sched_yield().
4803 void __sched yield(void)
4805 set_current_state(TASK_RUNNING);
4808 EXPORT_SYMBOL(yield);
4811 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4812 * that process accounting knows that this is a task in IO wait state.
4814 * But don't do that if it is a deliberate, throttling IO wait (this task
4815 * has set its backing_dev_info: the queue against which it should throttle)
4817 void __sched io_schedule(void)
4819 struct rq *rq = &__raw_get_cpu_var(runqueues);
4821 delayacct_blkio_start();
4822 atomic_inc(&rq->nr_iowait);
4824 atomic_dec(&rq->nr_iowait);
4825 delayacct_blkio_end();
4827 EXPORT_SYMBOL(io_schedule);
4829 long __sched io_schedule_timeout(long timeout)
4831 struct rq *rq = &__raw_get_cpu_var(runqueues);
4834 delayacct_blkio_start();
4835 atomic_inc(&rq->nr_iowait);
4836 ret = schedule_timeout(timeout);
4837 atomic_dec(&rq->nr_iowait);
4838 delayacct_blkio_end();
4843 * sys_sched_get_priority_max - return maximum RT priority.
4844 * @policy: scheduling class.
4846 * this syscall returns the maximum rt_priority that can be used
4847 * by a given scheduling class.
4849 asmlinkage long sys_sched_get_priority_max(int policy)
4856 ret = MAX_USER_RT_PRIO-1;
4868 * sys_sched_get_priority_min - return minimum RT priority.
4869 * @policy: scheduling class.
4871 * this syscall returns the minimum rt_priority that can be used
4872 * by a given scheduling class.
4874 asmlinkage long sys_sched_get_priority_min(int policy)
4892 * sys_sched_rr_get_interval - return the default timeslice of a process.
4893 * @pid: pid of the process.
4894 * @interval: userspace pointer to the timeslice value.
4896 * this syscall writes the default timeslice value of a given process
4897 * into the user-space timespec buffer. A value of '0' means infinity.
4900 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4902 struct task_struct *p;
4903 unsigned int time_slice;
4911 read_lock(&tasklist_lock);
4912 p = find_process_by_pid(pid);
4916 retval = security_task_getscheduler(p);
4921 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4922 * tasks that are on an otherwise idle runqueue:
4925 if (p->policy == SCHED_RR) {
4926 time_slice = DEF_TIMESLICE;
4928 struct sched_entity *se = &p->se;
4929 unsigned long flags;
4932 rq = task_rq_lock(p, &flags);
4933 if (rq->cfs.load.weight)
4934 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4935 task_rq_unlock(rq, &flags);
4937 read_unlock(&tasklist_lock);
4938 jiffies_to_timespec(time_slice, &t);
4939 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4943 read_unlock(&tasklist_lock);
4947 static const char stat_nam[] = "RSDTtZX";
4949 void sched_show_task(struct task_struct *p)
4951 unsigned long free = 0;
4954 state = p->state ? __ffs(p->state) + 1 : 0;
4955 printk(KERN_INFO "%-13.13s %c", p->comm,
4956 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4957 #if BITS_PER_LONG == 32
4958 if (state == TASK_RUNNING)
4959 printk(KERN_CONT " running ");
4961 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4963 if (state == TASK_RUNNING)
4964 printk(KERN_CONT " running task ");
4966 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4968 #ifdef CONFIG_DEBUG_STACK_USAGE
4970 unsigned long *n = end_of_stack(p);
4973 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4976 printk(KERN_CONT "%5lu %5d %6d\n", free,
4977 task_pid_nr(p), task_pid_nr(p->real_parent));
4979 if (state != TASK_RUNNING)
4980 show_stack(p, NULL);
4983 void show_state_filter(unsigned long state_filter)
4985 struct task_struct *g, *p;
4987 #if BITS_PER_LONG == 32
4989 " task PC stack pid father\n");
4992 " task PC stack pid father\n");
4994 read_lock(&tasklist_lock);
4995 do_each_thread(g, p) {
4997 * reset the NMI-timeout, listing all files on a slow
4998 * console might take alot of time:
5000 touch_nmi_watchdog();
5001 if (!state_filter || (p->state & state_filter))
5003 } while_each_thread(g, p);
5005 touch_all_softlockup_watchdogs();
5007 #ifdef CONFIG_SCHED_DEBUG
5008 sysrq_sched_debug_show();
5010 read_unlock(&tasklist_lock);
5012 * Only show locks if all tasks are dumped:
5014 if (state_filter == -1)
5015 debug_show_all_locks();
5018 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5020 idle->sched_class = &idle_sched_class;
5024 * init_idle - set up an idle thread for a given CPU
5025 * @idle: task in question
5026 * @cpu: cpu the idle task belongs to
5028 * NOTE: this function does not set the idle thread's NEED_RESCHED
5029 * flag, to make booting more robust.
5031 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5033 struct rq *rq = cpu_rq(cpu);
5034 unsigned long flags;
5037 idle->se.exec_start = sched_clock();
5039 idle->prio = idle->normal_prio = MAX_PRIO;
5040 idle->cpus_allowed = cpumask_of_cpu(cpu);
5041 __set_task_cpu(idle, cpu);
5043 spin_lock_irqsave(&rq->lock, flags);
5044 rq->curr = rq->idle = idle;
5045 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5048 spin_unlock_irqrestore(&rq->lock, flags);
5050 /* Set the preempt count _outside_ the spinlocks! */
5051 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5052 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5054 task_thread_info(idle)->preempt_count = 0;
5057 * The idle tasks have their own, simple scheduling class:
5059 idle->sched_class = &idle_sched_class;
5063 * In a system that switches off the HZ timer nohz_cpu_mask
5064 * indicates which cpus entered this state. This is used
5065 * in the rcu update to wait only for active cpus. For system
5066 * which do not switch off the HZ timer nohz_cpu_mask should
5067 * always be CPU_MASK_NONE.
5069 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5072 * Increase the granularity value when there are more CPUs,
5073 * because with more CPUs the 'effective latency' as visible
5074 * to users decreases. But the relationship is not linear,
5075 * so pick a second-best guess by going with the log2 of the
5078 * This idea comes from the SD scheduler of Con Kolivas:
5080 static inline void sched_init_granularity(void)
5082 unsigned int factor = 1 + ilog2(num_online_cpus());
5083 const unsigned long limit = 200000000;
5085 sysctl_sched_min_granularity *= factor;
5086 if (sysctl_sched_min_granularity > limit)
5087 sysctl_sched_min_granularity = limit;
5089 sysctl_sched_latency *= factor;
5090 if (sysctl_sched_latency > limit)
5091 sysctl_sched_latency = limit;
5093 sysctl_sched_wakeup_granularity *= factor;
5094 sysctl_sched_batch_wakeup_granularity *= factor;
5099 * This is how migration works:
5101 * 1) we queue a struct migration_req structure in the source CPU's
5102 * runqueue and wake up that CPU's migration thread.
5103 * 2) we down() the locked semaphore => thread blocks.
5104 * 3) migration thread wakes up (implicitly it forces the migrated
5105 * thread off the CPU)
5106 * 4) it gets the migration request and checks whether the migrated
5107 * task is still in the wrong runqueue.
5108 * 5) if it's in the wrong runqueue then the migration thread removes
5109 * it and puts it into the right queue.
5110 * 6) migration thread up()s the semaphore.
5111 * 7) we wake up and the migration is done.
5115 * Change a given task's CPU affinity. Migrate the thread to a
5116 * proper CPU and schedule it away if the CPU it's executing on
5117 * is removed from the allowed bitmask.
5119 * NOTE: the caller must have a valid reference to the task, the
5120 * task must not exit() & deallocate itself prematurely. The
5121 * call is not atomic; no spinlocks may be held.
5123 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5125 struct migration_req req;
5126 unsigned long flags;
5130 rq = task_rq_lock(p, &flags);
5131 if (!cpus_intersects(new_mask, cpu_online_map)) {
5136 p->cpus_allowed = new_mask;
5137 /* Can the task run on the task's current CPU? If so, we're done */
5138 if (cpu_isset(task_cpu(p), new_mask))
5141 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5142 /* Need help from migration thread: drop lock and wait. */
5143 task_rq_unlock(rq, &flags);
5144 wake_up_process(rq->migration_thread);
5145 wait_for_completion(&req.done);
5146 tlb_migrate_finish(p->mm);
5150 task_rq_unlock(rq, &flags);
5154 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5157 * Move (not current) task off this cpu, onto dest cpu. We're doing
5158 * this because either it can't run here any more (set_cpus_allowed()
5159 * away from this CPU, or CPU going down), or because we're
5160 * attempting to rebalance this task on exec (sched_exec).
5162 * So we race with normal scheduler movements, but that's OK, as long
5163 * as the task is no longer on this CPU.
5165 * Returns non-zero if task was successfully migrated.
5167 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5169 struct rq *rq_dest, *rq_src;
5172 if (unlikely(cpu_is_offline(dest_cpu)))
5175 rq_src = cpu_rq(src_cpu);
5176 rq_dest = cpu_rq(dest_cpu);
5178 double_rq_lock(rq_src, rq_dest);
5179 /* Already moved. */
5180 if (task_cpu(p) != src_cpu)
5182 /* Affinity changed (again). */
5183 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5186 on_rq = p->se.on_rq;
5188 deactivate_task(rq_src, p, 0);
5190 set_task_cpu(p, dest_cpu);
5192 activate_task(rq_dest, p, 0);
5193 check_preempt_curr(rq_dest, p);
5197 double_rq_unlock(rq_src, rq_dest);
5202 * migration_thread - this is a highprio system thread that performs
5203 * thread migration by bumping thread off CPU then 'pushing' onto
5206 static int migration_thread(void *data)
5208 int cpu = (long)data;
5212 BUG_ON(rq->migration_thread != current);
5214 set_current_state(TASK_INTERRUPTIBLE);
5215 while (!kthread_should_stop()) {
5216 struct migration_req *req;
5217 struct list_head *head;
5219 spin_lock_irq(&rq->lock);
5221 if (cpu_is_offline(cpu)) {
5222 spin_unlock_irq(&rq->lock);
5226 if (rq->active_balance) {
5227 active_load_balance(rq, cpu);
5228 rq->active_balance = 0;
5231 head = &rq->migration_queue;
5233 if (list_empty(head)) {
5234 spin_unlock_irq(&rq->lock);
5236 set_current_state(TASK_INTERRUPTIBLE);
5239 req = list_entry(head->next, struct migration_req, list);
5240 list_del_init(head->next);
5242 spin_unlock(&rq->lock);
5243 __migrate_task(req->task, cpu, req->dest_cpu);
5246 complete(&req->done);
5248 __set_current_state(TASK_RUNNING);
5252 /* Wait for kthread_stop */
5253 set_current_state(TASK_INTERRUPTIBLE);
5254 while (!kthread_should_stop()) {
5256 set_current_state(TASK_INTERRUPTIBLE);
5258 __set_current_state(TASK_RUNNING);
5262 #ifdef CONFIG_HOTPLUG_CPU
5264 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5268 local_irq_disable();
5269 ret = __migrate_task(p, src_cpu, dest_cpu);
5275 * Figure out where task on dead CPU should go, use force if necessary.
5276 * NOTE: interrupts should be disabled by the caller
5278 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5280 unsigned long flags;
5287 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5288 cpus_and(mask, mask, p->cpus_allowed);
5289 dest_cpu = any_online_cpu(mask);
5291 /* On any allowed CPU? */
5292 if (dest_cpu == NR_CPUS)
5293 dest_cpu = any_online_cpu(p->cpus_allowed);
5295 /* No more Mr. Nice Guy. */
5296 if (dest_cpu == NR_CPUS) {
5297 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5299 * Try to stay on the same cpuset, where the
5300 * current cpuset may be a subset of all cpus.
5301 * The cpuset_cpus_allowed_locked() variant of
5302 * cpuset_cpus_allowed() will not block. It must be
5303 * called within calls to cpuset_lock/cpuset_unlock.
5305 rq = task_rq_lock(p, &flags);
5306 p->cpus_allowed = cpus_allowed;
5307 dest_cpu = any_online_cpu(p->cpus_allowed);
5308 task_rq_unlock(rq, &flags);
5311 * Don't tell them about moving exiting tasks or
5312 * kernel threads (both mm NULL), since they never
5315 if (p->mm && printk_ratelimit()) {
5316 printk(KERN_INFO "process %d (%s) no "
5317 "longer affine to cpu%d\n",
5318 task_pid_nr(p), p->comm, dead_cpu);
5321 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5325 * While a dead CPU has no uninterruptible tasks queued at this point,
5326 * it might still have a nonzero ->nr_uninterruptible counter, because
5327 * for performance reasons the counter is not stricly tracking tasks to
5328 * their home CPUs. So we just add the counter to another CPU's counter,
5329 * to keep the global sum constant after CPU-down:
5331 static void migrate_nr_uninterruptible(struct rq *rq_src)
5333 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5334 unsigned long flags;
5336 local_irq_save(flags);
5337 double_rq_lock(rq_src, rq_dest);
5338 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5339 rq_src->nr_uninterruptible = 0;
5340 double_rq_unlock(rq_src, rq_dest);
5341 local_irq_restore(flags);
5344 /* Run through task list and migrate tasks from the dead cpu. */
5345 static void migrate_live_tasks(int src_cpu)
5347 struct task_struct *p, *t;
5349 read_lock(&tasklist_lock);
5351 do_each_thread(t, p) {
5355 if (task_cpu(p) == src_cpu)
5356 move_task_off_dead_cpu(src_cpu, p);
5357 } while_each_thread(t, p);
5359 read_unlock(&tasklist_lock);
5363 * Schedules idle task to be the next runnable task on current CPU.
5364 * It does so by boosting its priority to highest possible.
5365 * Used by CPU offline code.
5367 void sched_idle_next(void)
5369 int this_cpu = smp_processor_id();
5370 struct rq *rq = cpu_rq(this_cpu);
5371 struct task_struct *p = rq->idle;
5372 unsigned long flags;
5374 /* cpu has to be offline */
5375 BUG_ON(cpu_online(this_cpu));
5378 * Strictly not necessary since rest of the CPUs are stopped by now
5379 * and interrupts disabled on the current cpu.
5381 spin_lock_irqsave(&rq->lock, flags);
5383 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5385 update_rq_clock(rq);
5386 activate_task(rq, p, 0);
5388 spin_unlock_irqrestore(&rq->lock, flags);
5392 * Ensures that the idle task is using init_mm right before its cpu goes
5395 void idle_task_exit(void)
5397 struct mm_struct *mm = current->active_mm;
5399 BUG_ON(cpu_online(smp_processor_id()));
5402 switch_mm(mm, &init_mm, current);
5406 /* called under rq->lock with disabled interrupts */
5407 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5409 struct rq *rq = cpu_rq(dead_cpu);
5411 /* Must be exiting, otherwise would be on tasklist. */
5412 BUG_ON(!p->exit_state);
5414 /* Cannot have done final schedule yet: would have vanished. */
5415 BUG_ON(p->state == TASK_DEAD);
5420 * Drop lock around migration; if someone else moves it,
5421 * that's OK. No task can be added to this CPU, so iteration is
5424 spin_unlock_irq(&rq->lock);
5425 move_task_off_dead_cpu(dead_cpu, p);
5426 spin_lock_irq(&rq->lock);
5431 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5432 static void migrate_dead_tasks(unsigned int dead_cpu)
5434 struct rq *rq = cpu_rq(dead_cpu);
5435 struct task_struct *next;
5438 if (!rq->nr_running)
5440 update_rq_clock(rq);
5441 next = pick_next_task(rq, rq->curr);
5444 migrate_dead(dead_cpu, next);
5448 #endif /* CONFIG_HOTPLUG_CPU */
5450 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5452 static struct ctl_table sd_ctl_dir[] = {
5454 .procname = "sched_domain",
5460 static struct ctl_table sd_ctl_root[] = {
5462 .ctl_name = CTL_KERN,
5463 .procname = "kernel",
5465 .child = sd_ctl_dir,
5470 static struct ctl_table *sd_alloc_ctl_entry(int n)
5472 struct ctl_table *entry =
5473 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5478 static void sd_free_ctl_entry(struct ctl_table **tablep)
5480 struct ctl_table *entry;
5483 * In the intermediate directories, both the child directory and
5484 * procname are dynamically allocated and could fail but the mode
5485 * will always be set. In the lowest directory the names are
5486 * static strings and all have proc handlers.
5488 for (entry = *tablep; entry->mode; entry++) {
5490 sd_free_ctl_entry(&entry->child);
5491 if (entry->proc_handler == NULL)
5492 kfree(entry->procname);
5500 set_table_entry(struct ctl_table *entry,
5501 const char *procname, void *data, int maxlen,
5502 mode_t mode, proc_handler *proc_handler)
5504 entry->procname = procname;
5506 entry->maxlen = maxlen;
5508 entry->proc_handler = proc_handler;
5511 static struct ctl_table *
5512 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5514 struct ctl_table *table = sd_alloc_ctl_entry(12);
5519 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5520 sizeof(long), 0644, proc_doulongvec_minmax);
5521 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5522 sizeof(long), 0644, proc_doulongvec_minmax);
5523 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5524 sizeof(int), 0644, proc_dointvec_minmax);
5525 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5526 sizeof(int), 0644, proc_dointvec_minmax);
5527 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5528 sizeof(int), 0644, proc_dointvec_minmax);
5529 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5530 sizeof(int), 0644, proc_dointvec_minmax);
5531 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5532 sizeof(int), 0644, proc_dointvec_minmax);
5533 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5534 sizeof(int), 0644, proc_dointvec_minmax);
5535 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5536 sizeof(int), 0644, proc_dointvec_minmax);
5537 set_table_entry(&table[9], "cache_nice_tries",
5538 &sd->cache_nice_tries,
5539 sizeof(int), 0644, proc_dointvec_minmax);
5540 set_table_entry(&table[10], "flags", &sd->flags,
5541 sizeof(int), 0644, proc_dointvec_minmax);
5542 /* &table[11] is terminator */
5547 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5549 struct ctl_table *entry, *table;
5550 struct sched_domain *sd;
5551 int domain_num = 0, i;
5554 for_each_domain(cpu, sd)
5556 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5561 for_each_domain(cpu, sd) {
5562 snprintf(buf, 32, "domain%d", i);
5563 entry->procname = kstrdup(buf, GFP_KERNEL);
5565 entry->child = sd_alloc_ctl_domain_table(sd);
5572 static struct ctl_table_header *sd_sysctl_header;
5573 static void register_sched_domain_sysctl(void)
5575 int i, cpu_num = num_online_cpus();
5576 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5579 WARN_ON(sd_ctl_dir[0].child);
5580 sd_ctl_dir[0].child = entry;
5585 for_each_online_cpu(i) {
5586 snprintf(buf, 32, "cpu%d", i);
5587 entry->procname = kstrdup(buf, GFP_KERNEL);
5589 entry->child = sd_alloc_ctl_cpu_table(i);
5593 WARN_ON(sd_sysctl_header);
5594 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5597 /* may be called multiple times per register */
5598 static void unregister_sched_domain_sysctl(void)
5600 if (sd_sysctl_header)
5601 unregister_sysctl_table(sd_sysctl_header);
5602 sd_sysctl_header = NULL;
5603 if (sd_ctl_dir[0].child)
5604 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5607 static void register_sched_domain_sysctl(void)
5610 static void unregister_sched_domain_sysctl(void)
5616 * migration_call - callback that gets triggered when a CPU is added.
5617 * Here we can start up the necessary migration thread for the new CPU.
5619 static int __cpuinit
5620 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5622 struct task_struct *p;
5623 int cpu = (long)hcpu;
5624 unsigned long flags;
5629 case CPU_UP_PREPARE:
5630 case CPU_UP_PREPARE_FROZEN:
5631 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5634 kthread_bind(p, cpu);
5635 /* Must be high prio: stop_machine expects to yield to it. */
5636 rq = task_rq_lock(p, &flags);
5637 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5638 task_rq_unlock(rq, &flags);
5639 cpu_rq(cpu)->migration_thread = p;
5643 case CPU_ONLINE_FROZEN:
5644 /* Strictly unnecessary, as first user will wake it. */
5645 wake_up_process(cpu_rq(cpu)->migration_thread);
5648 #ifdef CONFIG_HOTPLUG_CPU
5649 case CPU_UP_CANCELED:
5650 case CPU_UP_CANCELED_FROZEN:
5651 if (!cpu_rq(cpu)->migration_thread)
5653 /* Unbind it from offline cpu so it can run. Fall thru. */
5654 kthread_bind(cpu_rq(cpu)->migration_thread,
5655 any_online_cpu(cpu_online_map));
5656 kthread_stop(cpu_rq(cpu)->migration_thread);
5657 cpu_rq(cpu)->migration_thread = NULL;
5661 case CPU_DEAD_FROZEN:
5662 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5663 migrate_live_tasks(cpu);
5665 kthread_stop(rq->migration_thread);
5666 rq->migration_thread = NULL;
5667 /* Idle task back to normal (off runqueue, low prio) */
5668 spin_lock_irq(&rq->lock);
5669 update_rq_clock(rq);
5670 deactivate_task(rq, rq->idle, 0);
5671 rq->idle->static_prio = MAX_PRIO;
5672 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5673 rq->idle->sched_class = &idle_sched_class;
5674 migrate_dead_tasks(cpu);
5675 spin_unlock_irq(&rq->lock);
5677 migrate_nr_uninterruptible(rq);
5678 BUG_ON(rq->nr_running != 0);
5681 * No need to migrate the tasks: it was best-effort if
5682 * they didn't take sched_hotcpu_mutex. Just wake up
5685 spin_lock_irq(&rq->lock);
5686 while (!list_empty(&rq->migration_queue)) {
5687 struct migration_req *req;
5689 req = list_entry(rq->migration_queue.next,
5690 struct migration_req, list);
5691 list_del_init(&req->list);
5692 complete(&req->done);
5694 spin_unlock_irq(&rq->lock);
5701 /* Register at highest priority so that task migration (migrate_all_tasks)
5702 * happens before everything else.
5704 static struct notifier_block __cpuinitdata migration_notifier = {
5705 .notifier_call = migration_call,
5709 void __init migration_init(void)
5711 void *cpu = (void *)(long)smp_processor_id();
5714 /* Start one for the boot CPU: */
5715 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5716 BUG_ON(err == NOTIFY_BAD);
5717 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5718 register_cpu_notifier(&migration_notifier);
5724 /* Number of possible processor ids */
5725 int nr_cpu_ids __read_mostly = NR_CPUS;
5726 EXPORT_SYMBOL(nr_cpu_ids);
5728 #ifdef CONFIG_SCHED_DEBUG
5730 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5732 struct sched_group *group = sd->groups;
5733 cpumask_t groupmask;
5736 cpumask_scnprintf(str, NR_CPUS, sd->span);
5737 cpus_clear(groupmask);
5739 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5741 if (!(sd->flags & SD_LOAD_BALANCE)) {
5742 printk("does not load-balance\n");
5744 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5749 printk(KERN_CONT "span %s\n", str);
5751 if (!cpu_isset(cpu, sd->span)) {
5752 printk(KERN_ERR "ERROR: domain->span does not contain "
5755 if (!cpu_isset(cpu, group->cpumask)) {
5756 printk(KERN_ERR "ERROR: domain->groups does not contain"
5760 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5764 printk(KERN_ERR "ERROR: group is NULL\n");
5768 if (!group->__cpu_power) {
5769 printk(KERN_CONT "\n");
5770 printk(KERN_ERR "ERROR: domain->cpu_power not "
5775 if (!cpus_weight(group->cpumask)) {
5776 printk(KERN_CONT "\n");
5777 printk(KERN_ERR "ERROR: empty group\n");
5781 if (cpus_intersects(groupmask, group->cpumask)) {
5782 printk(KERN_CONT "\n");
5783 printk(KERN_ERR "ERROR: repeated CPUs\n");
5787 cpus_or(groupmask, groupmask, group->cpumask);
5789 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5790 printk(KERN_CONT " %s", str);
5792 group = group->next;
5793 } while (group != sd->groups);
5794 printk(KERN_CONT "\n");
5796 if (!cpus_equal(sd->span, groupmask))
5797 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5799 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5800 printk(KERN_ERR "ERROR: parent span is not a superset "
5801 "of domain->span\n");
5805 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5810 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5814 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5817 if (sched_domain_debug_one(sd, cpu, level))
5826 # define sched_domain_debug(sd, cpu) do { } while (0)
5829 static int sd_degenerate(struct sched_domain *sd)
5831 if (cpus_weight(sd->span) == 1)
5834 /* Following flags need at least 2 groups */
5835 if (sd->flags & (SD_LOAD_BALANCE |
5836 SD_BALANCE_NEWIDLE |
5840 SD_SHARE_PKG_RESOURCES)) {
5841 if (sd->groups != sd->groups->next)
5845 /* Following flags don't use groups */
5846 if (sd->flags & (SD_WAKE_IDLE |
5855 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5857 unsigned long cflags = sd->flags, pflags = parent->flags;
5859 if (sd_degenerate(parent))
5862 if (!cpus_equal(sd->span, parent->span))
5865 /* Does parent contain flags not in child? */
5866 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5867 if (cflags & SD_WAKE_AFFINE)
5868 pflags &= ~SD_WAKE_BALANCE;
5869 /* Flags needing groups don't count if only 1 group in parent */
5870 if (parent->groups == parent->groups->next) {
5871 pflags &= ~(SD_LOAD_BALANCE |
5872 SD_BALANCE_NEWIDLE |
5876 SD_SHARE_PKG_RESOURCES);
5878 if (~cflags & pflags)
5885 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5886 * hold the hotplug lock.
5888 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5890 struct rq *rq = cpu_rq(cpu);
5891 struct sched_domain *tmp;
5893 /* Remove the sched domains which do not contribute to scheduling. */
5894 for (tmp = sd; tmp; tmp = tmp->parent) {
5895 struct sched_domain *parent = tmp->parent;
5898 if (sd_parent_degenerate(tmp, parent)) {
5899 tmp->parent = parent->parent;
5901 parent->parent->child = tmp;
5905 if (sd && sd_degenerate(sd)) {
5911 sched_domain_debug(sd, cpu);
5913 rcu_assign_pointer(rq->sd, sd);
5916 /* cpus with isolated domains */
5917 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5919 /* Setup the mask of cpus configured for isolated domains */
5920 static int __init isolated_cpu_setup(char *str)
5922 int ints[NR_CPUS], i;
5924 str = get_options(str, ARRAY_SIZE(ints), ints);
5925 cpus_clear(cpu_isolated_map);
5926 for (i = 1; i <= ints[0]; i++)
5927 if (ints[i] < NR_CPUS)
5928 cpu_set(ints[i], cpu_isolated_map);
5932 __setup("isolcpus=", isolated_cpu_setup);
5935 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5936 * to a function which identifies what group(along with sched group) a CPU
5937 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5938 * (due to the fact that we keep track of groups covered with a cpumask_t).
5940 * init_sched_build_groups will build a circular linked list of the groups
5941 * covered by the given span, and will set each group's ->cpumask correctly,
5942 * and ->cpu_power to 0.
5945 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5946 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5947 struct sched_group **sg))
5949 struct sched_group *first = NULL, *last = NULL;
5950 cpumask_t covered = CPU_MASK_NONE;
5953 for_each_cpu_mask(i, span) {
5954 struct sched_group *sg;
5955 int group = group_fn(i, cpu_map, &sg);
5958 if (cpu_isset(i, covered))
5961 sg->cpumask = CPU_MASK_NONE;
5962 sg->__cpu_power = 0;
5964 for_each_cpu_mask(j, span) {
5965 if (group_fn(j, cpu_map, NULL) != group)
5968 cpu_set(j, covered);
5969 cpu_set(j, sg->cpumask);
5980 #define SD_NODES_PER_DOMAIN 16
5985 * find_next_best_node - find the next node to include in a sched_domain
5986 * @node: node whose sched_domain we're building
5987 * @used_nodes: nodes already in the sched_domain
5989 * Find the next node to include in a given scheduling domain. Simply
5990 * finds the closest node not already in the @used_nodes map.
5992 * Should use nodemask_t.
5994 static int find_next_best_node(int node, unsigned long *used_nodes)
5996 int i, n, val, min_val, best_node = 0;
6000 for (i = 0; i < MAX_NUMNODES; i++) {
6001 /* Start at @node */
6002 n = (node + i) % MAX_NUMNODES;
6004 if (!nr_cpus_node(n))
6007 /* Skip already used nodes */
6008 if (test_bit(n, used_nodes))
6011 /* Simple min distance search */
6012 val = node_distance(node, n);
6014 if (val < min_val) {
6020 set_bit(best_node, used_nodes);
6025 * sched_domain_node_span - get a cpumask for a node's sched_domain
6026 * @node: node whose cpumask we're constructing
6027 * @size: number of nodes to include in this span
6029 * Given a node, construct a good cpumask for its sched_domain to span. It
6030 * should be one that prevents unnecessary balancing, but also spreads tasks
6033 static cpumask_t sched_domain_node_span(int node)
6035 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6036 cpumask_t span, nodemask;
6040 bitmap_zero(used_nodes, MAX_NUMNODES);
6042 nodemask = node_to_cpumask(node);
6043 cpus_or(span, span, nodemask);
6044 set_bit(node, used_nodes);
6046 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6047 int next_node = find_next_best_node(node, used_nodes);
6049 nodemask = node_to_cpumask(next_node);
6050 cpus_or(span, span, nodemask);
6057 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6060 * SMT sched-domains:
6062 #ifdef CONFIG_SCHED_SMT
6063 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6064 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6067 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6070 *sg = &per_cpu(sched_group_cpus, cpu);
6076 * multi-core sched-domains:
6078 #ifdef CONFIG_SCHED_MC
6079 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6080 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6083 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6085 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6088 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6089 cpus_and(mask, mask, *cpu_map);
6090 group = first_cpu(mask);
6092 *sg = &per_cpu(sched_group_core, group);
6095 #elif defined(CONFIG_SCHED_MC)
6097 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6100 *sg = &per_cpu(sched_group_core, cpu);
6105 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6106 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6109 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6112 #ifdef CONFIG_SCHED_MC
6113 cpumask_t mask = cpu_coregroup_map(cpu);
6114 cpus_and(mask, mask, *cpu_map);
6115 group = first_cpu(mask);
6116 #elif defined(CONFIG_SCHED_SMT)
6117 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6118 cpus_and(mask, mask, *cpu_map);
6119 group = first_cpu(mask);
6124 *sg = &per_cpu(sched_group_phys, group);
6130 * The init_sched_build_groups can't handle what we want to do with node
6131 * groups, so roll our own. Now each node has its own list of groups which
6132 * gets dynamically allocated.
6134 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6135 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6137 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6138 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6140 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6141 struct sched_group **sg)
6143 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6146 cpus_and(nodemask, nodemask, *cpu_map);
6147 group = first_cpu(nodemask);
6150 *sg = &per_cpu(sched_group_allnodes, group);
6154 static void init_numa_sched_groups_power(struct sched_group *group_head)
6156 struct sched_group *sg = group_head;
6162 for_each_cpu_mask(j, sg->cpumask) {
6163 struct sched_domain *sd;
6165 sd = &per_cpu(phys_domains, j);
6166 if (j != first_cpu(sd->groups->cpumask)) {
6168 * Only add "power" once for each
6174 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6177 } while (sg != group_head);
6182 /* Free memory allocated for various sched_group structures */
6183 static void free_sched_groups(const cpumask_t *cpu_map)
6187 for_each_cpu_mask(cpu, *cpu_map) {
6188 struct sched_group **sched_group_nodes
6189 = sched_group_nodes_bycpu[cpu];
6191 if (!sched_group_nodes)
6194 for (i = 0; i < MAX_NUMNODES; i++) {
6195 cpumask_t nodemask = node_to_cpumask(i);
6196 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6198 cpus_and(nodemask, nodemask, *cpu_map);
6199 if (cpus_empty(nodemask))
6209 if (oldsg != sched_group_nodes[i])
6212 kfree(sched_group_nodes);
6213 sched_group_nodes_bycpu[cpu] = NULL;
6217 static void free_sched_groups(const cpumask_t *cpu_map)
6223 * Initialize sched groups cpu_power.
6225 * cpu_power indicates the capacity of sched group, which is used while
6226 * distributing the load between different sched groups in a sched domain.
6227 * Typically cpu_power for all the groups in a sched domain will be same unless
6228 * there are asymmetries in the topology. If there are asymmetries, group
6229 * having more cpu_power will pickup more load compared to the group having
6232 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6233 * the maximum number of tasks a group can handle in the presence of other idle
6234 * or lightly loaded groups in the same sched domain.
6236 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6238 struct sched_domain *child;
6239 struct sched_group *group;
6241 WARN_ON(!sd || !sd->groups);
6243 if (cpu != first_cpu(sd->groups->cpumask))
6248 sd->groups->__cpu_power = 0;
6251 * For perf policy, if the groups in child domain share resources
6252 * (for example cores sharing some portions of the cache hierarchy
6253 * or SMT), then set this domain groups cpu_power such that each group
6254 * can handle only one task, when there are other idle groups in the
6255 * same sched domain.
6257 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6259 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6260 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6265 * add cpu_power of each child group to this groups cpu_power
6267 group = child->groups;
6269 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6270 group = group->next;
6271 } while (group != child->groups);
6275 * Build sched domains for a given set of cpus and attach the sched domains
6276 * to the individual cpus
6278 static int build_sched_domains(const cpumask_t *cpu_map)
6282 struct sched_group **sched_group_nodes = NULL;
6283 int sd_allnodes = 0;
6286 * Allocate the per-node list of sched groups
6288 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6290 if (!sched_group_nodes) {
6291 printk(KERN_WARNING "Can not alloc sched group node list\n");
6294 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6298 * Set up domains for cpus specified by the cpu_map.
6300 for_each_cpu_mask(i, *cpu_map) {
6301 struct sched_domain *sd = NULL, *p;
6302 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6304 cpus_and(nodemask, nodemask, *cpu_map);
6307 if (cpus_weight(*cpu_map) >
6308 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6309 sd = &per_cpu(allnodes_domains, i);
6310 *sd = SD_ALLNODES_INIT;
6311 sd->span = *cpu_map;
6312 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6318 sd = &per_cpu(node_domains, i);
6320 sd->span = sched_domain_node_span(cpu_to_node(i));
6324 cpus_and(sd->span, sd->span, *cpu_map);
6328 sd = &per_cpu(phys_domains, i);
6330 sd->span = nodemask;
6334 cpu_to_phys_group(i, cpu_map, &sd->groups);
6336 #ifdef CONFIG_SCHED_MC
6338 sd = &per_cpu(core_domains, i);
6340 sd->span = cpu_coregroup_map(i);
6341 cpus_and(sd->span, sd->span, *cpu_map);
6344 cpu_to_core_group(i, cpu_map, &sd->groups);
6347 #ifdef CONFIG_SCHED_SMT
6349 sd = &per_cpu(cpu_domains, i);
6350 *sd = SD_SIBLING_INIT;
6351 sd->span = per_cpu(cpu_sibling_map, i);
6352 cpus_and(sd->span, sd->span, *cpu_map);
6355 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6359 #ifdef CONFIG_SCHED_SMT
6360 /* Set up CPU (sibling) groups */
6361 for_each_cpu_mask(i, *cpu_map) {
6362 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6363 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6364 if (i != first_cpu(this_sibling_map))
6367 init_sched_build_groups(this_sibling_map, cpu_map,
6372 #ifdef CONFIG_SCHED_MC
6373 /* Set up multi-core groups */
6374 for_each_cpu_mask(i, *cpu_map) {
6375 cpumask_t this_core_map = cpu_coregroup_map(i);
6376 cpus_and(this_core_map, this_core_map, *cpu_map);
6377 if (i != first_cpu(this_core_map))
6379 init_sched_build_groups(this_core_map, cpu_map,
6380 &cpu_to_core_group);
6384 /* Set up physical groups */
6385 for (i = 0; i < MAX_NUMNODES; i++) {
6386 cpumask_t nodemask = node_to_cpumask(i);
6388 cpus_and(nodemask, nodemask, *cpu_map);
6389 if (cpus_empty(nodemask))
6392 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6396 /* Set up node groups */
6398 init_sched_build_groups(*cpu_map, cpu_map,
6399 &cpu_to_allnodes_group);
6401 for (i = 0; i < MAX_NUMNODES; i++) {
6402 /* Set up node groups */
6403 struct sched_group *sg, *prev;
6404 cpumask_t nodemask = node_to_cpumask(i);
6405 cpumask_t domainspan;
6406 cpumask_t covered = CPU_MASK_NONE;
6409 cpus_and(nodemask, nodemask, *cpu_map);
6410 if (cpus_empty(nodemask)) {
6411 sched_group_nodes[i] = NULL;
6415 domainspan = sched_domain_node_span(i);
6416 cpus_and(domainspan, domainspan, *cpu_map);
6418 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6420 printk(KERN_WARNING "Can not alloc domain group for "
6424 sched_group_nodes[i] = sg;
6425 for_each_cpu_mask(j, nodemask) {
6426 struct sched_domain *sd;
6428 sd = &per_cpu(node_domains, j);
6431 sg->__cpu_power = 0;
6432 sg->cpumask = nodemask;
6434 cpus_or(covered, covered, nodemask);
6437 for (j = 0; j < MAX_NUMNODES; j++) {
6438 cpumask_t tmp, notcovered;
6439 int n = (i + j) % MAX_NUMNODES;
6441 cpus_complement(notcovered, covered);
6442 cpus_and(tmp, notcovered, *cpu_map);
6443 cpus_and(tmp, tmp, domainspan);
6444 if (cpus_empty(tmp))
6447 nodemask = node_to_cpumask(n);
6448 cpus_and(tmp, tmp, nodemask);
6449 if (cpus_empty(tmp))
6452 sg = kmalloc_node(sizeof(struct sched_group),
6456 "Can not alloc domain group for node %d\n", j);
6459 sg->__cpu_power = 0;
6461 sg->next = prev->next;
6462 cpus_or(covered, covered, tmp);
6469 /* Calculate CPU power for physical packages and nodes */
6470 #ifdef CONFIG_SCHED_SMT
6471 for_each_cpu_mask(i, *cpu_map) {
6472 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6474 init_sched_groups_power(i, sd);
6477 #ifdef CONFIG_SCHED_MC
6478 for_each_cpu_mask(i, *cpu_map) {
6479 struct sched_domain *sd = &per_cpu(core_domains, i);
6481 init_sched_groups_power(i, sd);
6485 for_each_cpu_mask(i, *cpu_map) {
6486 struct sched_domain *sd = &per_cpu(phys_domains, i);
6488 init_sched_groups_power(i, sd);
6492 for (i = 0; i < MAX_NUMNODES; i++)
6493 init_numa_sched_groups_power(sched_group_nodes[i]);
6496 struct sched_group *sg;
6498 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6499 init_numa_sched_groups_power(sg);
6503 /* Attach the domains */
6504 for_each_cpu_mask(i, *cpu_map) {
6505 struct sched_domain *sd;
6506 #ifdef CONFIG_SCHED_SMT
6507 sd = &per_cpu(cpu_domains, i);
6508 #elif defined(CONFIG_SCHED_MC)
6509 sd = &per_cpu(core_domains, i);
6511 sd = &per_cpu(phys_domains, i);
6513 cpu_attach_domain(sd, i);
6520 free_sched_groups(cpu_map);
6525 static cpumask_t *doms_cur; /* current sched domains */
6526 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6529 * Special case: If a kmalloc of a doms_cur partition (array of
6530 * cpumask_t) fails, then fallback to a single sched domain,
6531 * as determined by the single cpumask_t fallback_doms.
6533 static cpumask_t fallback_doms;
6536 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6537 * For now this just excludes isolated cpus, but could be used to
6538 * exclude other special cases in the future.
6540 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6545 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6547 doms_cur = &fallback_doms;
6548 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6549 err = build_sched_domains(doms_cur);
6550 register_sched_domain_sysctl();
6555 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6557 free_sched_groups(cpu_map);
6561 * Detach sched domains from a group of cpus specified in cpu_map
6562 * These cpus will now be attached to the NULL domain
6564 static void detach_destroy_domains(const cpumask_t *cpu_map)
6568 unregister_sched_domain_sysctl();
6570 for_each_cpu_mask(i, *cpu_map)
6571 cpu_attach_domain(NULL, i);
6572 synchronize_sched();
6573 arch_destroy_sched_domains(cpu_map);
6577 * Partition sched domains as specified by the 'ndoms_new'
6578 * cpumasks in the array doms_new[] of cpumasks. This compares
6579 * doms_new[] to the current sched domain partitioning, doms_cur[].
6580 * It destroys each deleted domain and builds each new domain.
6582 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6583 * The masks don't intersect (don't overlap.) We should setup one
6584 * sched domain for each mask. CPUs not in any of the cpumasks will
6585 * not be load balanced. If the same cpumask appears both in the
6586 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6589 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6590 * ownership of it and will kfree it when done with it. If the caller
6591 * failed the kmalloc call, then it can pass in doms_new == NULL,
6592 * and partition_sched_domains() will fallback to the single partition
6595 * Call with hotplug lock held
6597 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6603 /* always unregister in case we don't destroy any domains */
6604 unregister_sched_domain_sysctl();
6606 if (doms_new == NULL) {
6608 doms_new = &fallback_doms;
6609 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6612 /* Destroy deleted domains */
6613 for (i = 0; i < ndoms_cur; i++) {
6614 for (j = 0; j < ndoms_new; j++) {
6615 if (cpus_equal(doms_cur[i], doms_new[j]))
6618 /* no match - a current sched domain not in new doms_new[] */
6619 detach_destroy_domains(doms_cur + i);
6624 /* Build new domains */
6625 for (i = 0; i < ndoms_new; i++) {
6626 for (j = 0; j < ndoms_cur; j++) {
6627 if (cpus_equal(doms_new[i], doms_cur[j]))
6630 /* no match - add a new doms_new */
6631 build_sched_domains(doms_new + i);
6636 /* Remember the new sched domains */
6637 if (doms_cur != &fallback_doms)
6639 doms_cur = doms_new;
6640 ndoms_cur = ndoms_new;
6642 register_sched_domain_sysctl();
6647 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6648 static int arch_reinit_sched_domains(void)
6653 detach_destroy_domains(&cpu_online_map);
6654 err = arch_init_sched_domains(&cpu_online_map);
6660 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6664 if (buf[0] != '0' && buf[0] != '1')
6668 sched_smt_power_savings = (buf[0] == '1');
6670 sched_mc_power_savings = (buf[0] == '1');
6672 ret = arch_reinit_sched_domains();
6674 return ret ? ret : count;
6677 #ifdef CONFIG_SCHED_MC
6678 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6680 return sprintf(page, "%u\n", sched_mc_power_savings);
6682 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6683 const char *buf, size_t count)
6685 return sched_power_savings_store(buf, count, 0);
6687 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6688 sched_mc_power_savings_store);
6691 #ifdef CONFIG_SCHED_SMT
6692 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6694 return sprintf(page, "%u\n", sched_smt_power_savings);
6696 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6697 const char *buf, size_t count)
6699 return sched_power_savings_store(buf, count, 1);
6701 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6702 sched_smt_power_savings_store);
6705 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6709 #ifdef CONFIG_SCHED_SMT
6711 err = sysfs_create_file(&cls->kset.kobj,
6712 &attr_sched_smt_power_savings.attr);
6714 #ifdef CONFIG_SCHED_MC
6715 if (!err && mc_capable())
6716 err = sysfs_create_file(&cls->kset.kobj,
6717 &attr_sched_mc_power_savings.attr);
6724 * Force a reinitialization of the sched domains hierarchy. The domains
6725 * and groups cannot be updated in place without racing with the balancing
6726 * code, so we temporarily attach all running cpus to the NULL domain
6727 * which will prevent rebalancing while the sched domains are recalculated.
6729 static int update_sched_domains(struct notifier_block *nfb,
6730 unsigned long action, void *hcpu)
6733 case CPU_UP_PREPARE:
6734 case CPU_UP_PREPARE_FROZEN:
6735 case CPU_DOWN_PREPARE:
6736 case CPU_DOWN_PREPARE_FROZEN:
6737 detach_destroy_domains(&cpu_online_map);
6740 case CPU_UP_CANCELED:
6741 case CPU_UP_CANCELED_FROZEN:
6742 case CPU_DOWN_FAILED:
6743 case CPU_DOWN_FAILED_FROZEN:
6745 case CPU_ONLINE_FROZEN:
6747 case CPU_DEAD_FROZEN:
6749 * Fall through and re-initialise the domains.
6756 /* The hotplug lock is already held by cpu_up/cpu_down */
6757 arch_init_sched_domains(&cpu_online_map);
6762 void __init sched_init_smp(void)
6764 cpumask_t non_isolated_cpus;
6767 arch_init_sched_domains(&cpu_online_map);
6768 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6769 if (cpus_empty(non_isolated_cpus))
6770 cpu_set(smp_processor_id(), non_isolated_cpus);
6772 /* XXX: Theoretical race here - CPU may be hotplugged now */
6773 hotcpu_notifier(update_sched_domains, 0);
6775 /* Move init over to a non-isolated CPU */
6776 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6778 sched_init_granularity();
6780 #ifdef CONFIG_FAIR_GROUP_SCHED
6781 if (nr_cpu_ids == 1)
6784 lb_monitor_task = kthread_create(load_balance_monitor, NULL,
6786 if (!IS_ERR(lb_monitor_task)) {
6787 lb_monitor_task->flags |= PF_NOFREEZE;
6788 wake_up_process(lb_monitor_task);
6790 printk(KERN_ERR "Could not create load balance monitor thread"
6791 "(error = %ld) \n", PTR_ERR(lb_monitor_task));
6796 void __init sched_init_smp(void)
6798 sched_init_granularity();
6800 #endif /* CONFIG_SMP */
6802 int in_sched_functions(unsigned long addr)
6804 return in_lock_functions(addr) ||
6805 (addr >= (unsigned long)__sched_text_start
6806 && addr < (unsigned long)__sched_text_end);
6809 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6811 cfs_rq->tasks_timeline = RB_ROOT;
6812 #ifdef CONFIG_FAIR_GROUP_SCHED
6815 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6818 void __init sched_init(void)
6820 int highest_cpu = 0;
6823 for_each_possible_cpu(i) {
6824 struct rt_prio_array *array;
6828 spin_lock_init(&rq->lock);
6829 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6832 init_cfs_rq(&rq->cfs, rq);
6833 #ifdef CONFIG_FAIR_GROUP_SCHED
6834 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6836 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6837 struct sched_entity *se =
6838 &per_cpu(init_sched_entity, i);
6840 init_cfs_rq_p[i] = cfs_rq;
6841 init_cfs_rq(cfs_rq, rq);
6842 cfs_rq->tg = &init_task_group;
6843 list_add(&cfs_rq->leaf_cfs_rq_list,
6844 &rq->leaf_cfs_rq_list);
6846 init_sched_entity_p[i] = se;
6847 se->cfs_rq = &rq->cfs;
6849 se->load.weight = init_task_group_load;
6850 se->load.inv_weight =
6851 div64_64(1ULL<<32, init_task_group_load);
6854 init_task_group.shares = init_task_group_load;
6857 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6858 rq->cpu_load[j] = 0;
6861 rq->active_balance = 0;
6862 rq->next_balance = jiffies;
6865 rq->migration_thread = NULL;
6866 INIT_LIST_HEAD(&rq->migration_queue);
6868 atomic_set(&rq->nr_iowait, 0);
6870 array = &rq->rt.active;
6871 for (j = 0; j < MAX_RT_PRIO; j++) {
6872 INIT_LIST_HEAD(array->queue + j);
6873 __clear_bit(j, array->bitmap);
6876 /* delimiter for bitsearch: */
6877 __set_bit(MAX_RT_PRIO, array->bitmap);
6880 set_load_weight(&init_task);
6882 #ifdef CONFIG_PREEMPT_NOTIFIERS
6883 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6887 nr_cpu_ids = highest_cpu + 1;
6888 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6891 #ifdef CONFIG_RT_MUTEXES
6892 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6896 * The boot idle thread does lazy MMU switching as well:
6898 atomic_inc(&init_mm.mm_count);
6899 enter_lazy_tlb(&init_mm, current);
6902 * Make us the idle thread. Technically, schedule() should not be
6903 * called from this thread, however somewhere below it might be,
6904 * but because we are the idle thread, we just pick up running again
6905 * when this runqueue becomes "idle".
6907 init_idle(current, smp_processor_id());
6909 * During early bootup we pretend to be a normal task:
6911 current->sched_class = &fair_sched_class;
6914 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6915 void __might_sleep(char *file, int line)
6918 static unsigned long prev_jiffy; /* ratelimiting */
6920 if ((in_atomic() || irqs_disabled()) &&
6921 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6922 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6924 prev_jiffy = jiffies;
6925 printk(KERN_ERR "BUG: sleeping function called from invalid"
6926 " context at %s:%d\n", file, line);
6927 printk("in_atomic():%d, irqs_disabled():%d\n",
6928 in_atomic(), irqs_disabled());
6929 debug_show_held_locks(current);
6930 if (irqs_disabled())
6931 print_irqtrace_events(current);
6936 EXPORT_SYMBOL(__might_sleep);
6939 #ifdef CONFIG_MAGIC_SYSRQ
6940 static void normalize_task(struct rq *rq, struct task_struct *p)
6943 update_rq_clock(rq);
6944 on_rq = p->se.on_rq;
6946 deactivate_task(rq, p, 0);
6947 __setscheduler(rq, p, SCHED_NORMAL, 0);
6949 activate_task(rq, p, 0);
6950 resched_task(rq->curr);
6954 void normalize_rt_tasks(void)
6956 struct task_struct *g, *p;
6957 unsigned long flags;
6960 read_lock_irq(&tasklist_lock);
6961 do_each_thread(g, p) {
6963 * Only normalize user tasks:
6968 p->se.exec_start = 0;
6969 #ifdef CONFIG_SCHEDSTATS
6970 p->se.wait_start = 0;
6971 p->se.sleep_start = 0;
6972 p->se.block_start = 0;
6974 task_rq(p)->clock = 0;
6978 * Renice negative nice level userspace
6981 if (TASK_NICE(p) < 0 && p->mm)
6982 set_user_nice(p, 0);
6986 spin_lock_irqsave(&p->pi_lock, flags);
6987 rq = __task_rq_lock(p);
6989 normalize_task(rq, p);
6991 __task_rq_unlock(rq);
6992 spin_unlock_irqrestore(&p->pi_lock, flags);
6993 } while_each_thread(g, p);
6995 read_unlock_irq(&tasklist_lock);
6998 #endif /* CONFIG_MAGIC_SYSRQ */
7002 * These functions are only useful for the IA64 MCA handling.
7004 * They can only be called when the whole system has been
7005 * stopped - every CPU needs to be quiescent, and no scheduling
7006 * activity can take place. Using them for anything else would
7007 * be a serious bug, and as a result, they aren't even visible
7008 * under any other configuration.
7012 * curr_task - return the current task for a given cpu.
7013 * @cpu: the processor in question.
7015 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7017 struct task_struct *curr_task(int cpu)
7019 return cpu_curr(cpu);
7023 * set_curr_task - set the current task for a given cpu.
7024 * @cpu: the processor in question.
7025 * @p: the task pointer to set.
7027 * Description: This function must only be used when non-maskable interrupts
7028 * are serviced on a separate stack. It allows the architecture to switch the
7029 * notion of the current task on a cpu in a non-blocking manner. This function
7030 * must be called with all CPU's synchronized, and interrupts disabled, the
7031 * and caller must save the original value of the current task (see
7032 * curr_task() above) and restore that value before reenabling interrupts and
7033 * re-starting the system.
7035 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7037 void set_curr_task(int cpu, struct task_struct *p)
7044 #ifdef CONFIG_FAIR_GROUP_SCHED
7048 * distribute shares of all task groups among their schedulable entities,
7049 * to reflect load distrbution across cpus.
7051 static int rebalance_shares(struct sched_domain *sd, int this_cpu)
7053 struct cfs_rq *cfs_rq;
7054 struct rq *rq = cpu_rq(this_cpu);
7055 cpumask_t sdspan = sd->span;
7058 /* Walk thr' all the task groups that we have */
7059 for_each_leaf_cfs_rq(rq, cfs_rq) {
7061 unsigned long total_load = 0, total_shares;
7062 struct task_group *tg = cfs_rq->tg;
7064 /* Gather total task load of this group across cpus */
7065 for_each_cpu_mask(i, sdspan)
7066 total_load += tg->cfs_rq[i]->load.weight;
7068 /* Nothing to do if this group has no load */
7073 * tg->shares represents the number of cpu shares the task group
7074 * is eligible to hold on a single cpu. On N cpus, it is
7075 * eligible to hold (N * tg->shares) number of cpu shares.
7077 total_shares = tg->shares * cpus_weight(sdspan);
7080 * redistribute total_shares across cpus as per the task load
7083 for_each_cpu_mask(i, sdspan) {
7084 unsigned long local_load, local_shares;
7086 local_load = tg->cfs_rq[i]->load.weight;
7087 local_shares = (local_load * total_shares) / total_load;
7089 local_shares = MIN_GROUP_SHARES;
7090 if (local_shares == tg->se[i]->load.weight)
7093 spin_lock_irq(&cpu_rq(i)->lock);
7094 set_se_shares(tg->se[i], local_shares);
7095 spin_unlock_irq(&cpu_rq(i)->lock);
7104 * How frequently should we rebalance_shares() across cpus?
7106 * The more frequently we rebalance shares, the more accurate is the fairness
7107 * of cpu bandwidth distribution between task groups. However higher frequency
7108 * also implies increased scheduling overhead.
7110 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7111 * consecutive calls to rebalance_shares() in the same sched domain.
7113 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7114 * consecutive calls to rebalance_shares() in the same sched domain.
7116 * These settings allows for the appropriate tradeoff between accuracy of
7117 * fairness and the associated overhead.
7121 /* default: 8ms, units: milliseconds */
7122 const_debug unsigned int sysctl_sched_min_bal_int_shares = 8;
7124 /* default: 128ms, units: milliseconds */
7125 const_debug unsigned int sysctl_sched_max_bal_int_shares = 128;
7127 /* kernel thread that runs rebalance_shares() periodically */
7128 static int load_balance_monitor(void *unused)
7130 unsigned int timeout = sysctl_sched_min_bal_int_shares;
7131 struct sched_param schedparm;
7135 * We don't want this thread's execution to be limited by the shares
7136 * assigned to default group (init_task_group). Hence make it run
7137 * as a SCHED_RR RT task at the lowest priority.
7139 schedparm.sched_priority = 1;
7140 ret = sched_setscheduler(current, SCHED_RR, &schedparm);
7142 printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance"
7143 " monitor thread (error = %d) \n", ret);
7145 while (!kthread_should_stop()) {
7146 int i, cpu, balanced = 1;
7148 /* Prevent cpus going down or coming up */
7150 /* lockout changes to doms_cur[] array */
7153 * Enter a rcu read-side critical section to safely walk rq->sd
7154 * chain on various cpus and to walk task group list
7155 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7159 for (i = 0; i < ndoms_cur; i++) {
7160 cpumask_t cpumap = doms_cur[i];
7161 struct sched_domain *sd = NULL, *sd_prev = NULL;
7163 cpu = first_cpu(cpumap);
7165 /* Find the highest domain at which to balance shares */
7166 for_each_domain(cpu, sd) {
7167 if (!(sd->flags & SD_LOAD_BALANCE))
7173 /* sd == NULL? No load balance reqd in this domain */
7177 balanced &= rebalance_shares(sd, cpu);
7186 timeout = sysctl_sched_min_bal_int_shares;
7187 else if (timeout < sysctl_sched_max_bal_int_shares)
7190 msleep_interruptible(timeout);
7195 #endif /* CONFIG_SMP */
7197 /* allocate runqueue etc for a new task group */
7198 struct task_group *sched_create_group(void)
7200 struct task_group *tg;
7201 struct cfs_rq *cfs_rq;
7202 struct sched_entity *se;
7206 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7208 return ERR_PTR(-ENOMEM);
7210 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7213 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7217 for_each_possible_cpu(i) {
7220 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7225 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7230 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7231 memset(se, 0, sizeof(struct sched_entity));
7233 tg->cfs_rq[i] = cfs_rq;
7234 init_cfs_rq(cfs_rq, rq);
7238 se->cfs_rq = &rq->cfs;
7240 se->load.weight = NICE_0_LOAD;
7241 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7245 tg->shares = NICE_0_LOAD;
7247 lock_task_group_list();
7248 for_each_possible_cpu(i) {
7250 cfs_rq = tg->cfs_rq[i];
7251 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7253 unlock_task_group_list();
7258 for_each_possible_cpu(i) {
7260 kfree(tg->cfs_rq[i]);
7268 return ERR_PTR(-ENOMEM);
7271 /* rcu callback to free various structures associated with a task group */
7272 static void free_sched_group(struct rcu_head *rhp)
7274 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7275 struct cfs_rq *cfs_rq;
7276 struct sched_entity *se;
7279 /* now it should be safe to free those cfs_rqs */
7280 for_each_possible_cpu(i) {
7281 cfs_rq = tg->cfs_rq[i];
7293 /* Destroy runqueue etc associated with a task group */
7294 void sched_destroy_group(struct task_group *tg)
7296 struct cfs_rq *cfs_rq = NULL;
7299 lock_task_group_list();
7300 for_each_possible_cpu(i) {
7301 cfs_rq = tg->cfs_rq[i];
7302 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7304 unlock_task_group_list();
7308 /* wait for possible concurrent references to cfs_rqs complete */
7309 call_rcu(&tg->rcu, free_sched_group);
7312 /* change task's runqueue when it moves between groups.
7313 * The caller of this function should have put the task in its new group
7314 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7315 * reflect its new group.
7317 void sched_move_task(struct task_struct *tsk)
7320 unsigned long flags;
7323 rq = task_rq_lock(tsk, &flags);
7325 if (tsk->sched_class != &fair_sched_class) {
7326 set_task_cfs_rq(tsk, task_cpu(tsk));
7330 update_rq_clock(rq);
7332 running = task_current(rq, tsk);
7333 on_rq = tsk->se.on_rq;
7336 dequeue_task(rq, tsk, 0);
7337 if (unlikely(running))
7338 tsk->sched_class->put_prev_task(rq, tsk);
7341 set_task_cfs_rq(tsk, task_cpu(tsk));
7344 if (unlikely(running))
7345 tsk->sched_class->set_curr_task(rq);
7346 enqueue_task(rq, tsk, 0);
7350 task_rq_unlock(rq, &flags);
7353 /* rq->lock to be locked by caller */
7354 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7356 struct cfs_rq *cfs_rq = se->cfs_rq;
7357 struct rq *rq = cfs_rq->rq;
7361 shares = MIN_GROUP_SHARES;
7365 dequeue_entity(cfs_rq, se, 0);
7366 dec_cpu_load(rq, se->load.weight);
7369 se->load.weight = shares;
7370 se->load.inv_weight = div64_64((1ULL<<32), shares);
7373 enqueue_entity(cfs_rq, se, 0);
7374 inc_cpu_load(rq, se->load.weight);
7378 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7381 struct cfs_rq *cfs_rq;
7384 lock_task_group_list();
7385 if (tg->shares == shares)
7388 if (shares < MIN_GROUP_SHARES)
7389 shares = MIN_GROUP_SHARES;
7392 * Prevent any load balance activity (rebalance_shares,
7393 * load_balance_fair) from referring to this group first,
7394 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7396 for_each_possible_cpu(i) {
7397 cfs_rq = tg->cfs_rq[i];
7398 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7401 /* wait for any ongoing reference to this group to finish */
7402 synchronize_sched();
7405 * Now we are free to modify the group's share on each cpu
7406 * w/o tripping rebalance_share or load_balance_fair.
7408 tg->shares = shares;
7409 for_each_possible_cpu(i) {
7410 spin_lock_irq(&cpu_rq(i)->lock);
7411 set_se_shares(tg->se[i], shares);
7412 spin_unlock_irq(&cpu_rq(i)->lock);
7416 * Enable load balance activity on this group, by inserting it back on
7417 * each cpu's rq->leaf_cfs_rq_list.
7419 for_each_possible_cpu(i) {
7421 cfs_rq = tg->cfs_rq[i];
7422 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7425 unlock_task_group_list();
7429 unsigned long sched_group_shares(struct task_group *tg)
7434 #endif /* CONFIG_FAIR_GROUP_SCHED */
7436 #ifdef CONFIG_FAIR_CGROUP_SCHED
7438 /* return corresponding task_group object of a cgroup */
7439 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7441 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7442 struct task_group, css);
7445 static struct cgroup_subsys_state *
7446 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7448 struct task_group *tg;
7450 if (!cgrp->parent) {
7451 /* This is early initialization for the top cgroup */
7452 init_task_group.css.cgroup = cgrp;
7453 return &init_task_group.css;
7456 /* we support only 1-level deep hierarchical scheduler atm */
7457 if (cgrp->parent->parent)
7458 return ERR_PTR(-EINVAL);
7460 tg = sched_create_group();
7462 return ERR_PTR(-ENOMEM);
7464 /* Bind the cgroup to task_group object we just created */
7465 tg->css.cgroup = cgrp;
7471 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7473 struct task_group *tg = cgroup_tg(cgrp);
7475 sched_destroy_group(tg);
7479 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7480 struct task_struct *tsk)
7482 /* We don't support RT-tasks being in separate groups */
7483 if (tsk->sched_class != &fair_sched_class)
7490 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7491 struct cgroup *old_cont, struct task_struct *tsk)
7493 sched_move_task(tsk);
7496 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7499 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7502 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7504 struct task_group *tg = cgroup_tg(cgrp);
7506 return (u64) tg->shares;
7509 static struct cftype cpu_files[] = {
7512 .read_uint = cpu_shares_read_uint,
7513 .write_uint = cpu_shares_write_uint,
7517 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7519 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7522 struct cgroup_subsys cpu_cgroup_subsys = {
7524 .create = cpu_cgroup_create,
7525 .destroy = cpu_cgroup_destroy,
7526 .can_attach = cpu_cgroup_can_attach,
7527 .attach = cpu_cgroup_attach,
7528 .populate = cpu_cgroup_populate,
7529 .subsys_id = cpu_cgroup_subsys_id,
7533 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7535 #ifdef CONFIG_CGROUP_CPUACCT
7538 * CPU accounting code for task groups.
7540 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7541 * (balbir@in.ibm.com).
7544 /* track cpu usage of a group of tasks */
7546 struct cgroup_subsys_state css;
7547 /* cpuusage holds pointer to a u64-type object on every cpu */
7551 struct cgroup_subsys cpuacct_subsys;
7553 /* return cpu accounting group corresponding to this container */
7554 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7556 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7557 struct cpuacct, css);
7560 /* return cpu accounting group to which this task belongs */
7561 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7563 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7564 struct cpuacct, css);
7567 /* create a new cpu accounting group */
7568 static struct cgroup_subsys_state *cpuacct_create(
7569 struct cgroup_subsys *ss, struct cgroup *cont)
7571 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7574 return ERR_PTR(-ENOMEM);
7576 ca->cpuusage = alloc_percpu(u64);
7577 if (!ca->cpuusage) {
7579 return ERR_PTR(-ENOMEM);
7585 /* destroy an existing cpu accounting group */
7587 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
7589 struct cpuacct *ca = cgroup_ca(cont);
7591 free_percpu(ca->cpuusage);
7595 /* return total cpu usage (in nanoseconds) of a group */
7596 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7598 struct cpuacct *ca = cgroup_ca(cont);
7599 u64 totalcpuusage = 0;
7602 for_each_possible_cpu(i) {
7603 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7606 * Take rq->lock to make 64-bit addition safe on 32-bit
7609 spin_lock_irq(&cpu_rq(i)->lock);
7610 totalcpuusage += *cpuusage;
7611 spin_unlock_irq(&cpu_rq(i)->lock);
7614 return totalcpuusage;
7617 static struct cftype files[] = {
7620 .read_uint = cpuusage_read,
7624 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7626 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7630 * charge this task's execution time to its accounting group.
7632 * called with rq->lock held.
7634 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7638 if (!cpuacct_subsys.active)
7643 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7645 *cpuusage += cputime;
7649 struct cgroup_subsys cpuacct_subsys = {
7651 .create = cpuacct_create,
7652 .destroy = cpuacct_destroy,
7653 .populate = cpuacct_populate,
7654 .subsys_id = cpuacct_subsys_id,
7656 #endif /* CONFIG_CGROUP_CPUACCT */