4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
71 #include <asm/irq_regs.h>
74 * Scheduler clock - returns current time in nanosec units.
75 * This is default implementation.
76 * Architectures and sub-architectures can override this.
78 unsigned long long __attribute__((weak
)) sched_clock(void)
80 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
124 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
133 sg
->__cpu_power
+= val
;
134 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
138 static inline int rt_policy(int policy
)
140 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
145 static inline int task_has_rt_policy(struct task_struct
*p
)
147 return rt_policy(p
->policy
);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array
{
154 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
155 struct list_head queue
[MAX_RT_PRIO
];
158 #ifdef CONFIG_GROUP_SCHED
160 #include <linux/cgroup.h>
164 static LIST_HEAD(task_groups
);
166 /* task group related information */
168 #ifdef CONFIG_CGROUP_SCHED
169 struct cgroup_subsys_state css
;
172 #ifdef CONFIG_FAIR_GROUP_SCHED
173 /* schedulable entities of this group on each cpu */
174 struct sched_entity
**se
;
175 /* runqueue "owned" by this group on each cpu */
176 struct cfs_rq
**cfs_rq
;
177 unsigned long shares
;
180 #ifdef CONFIG_RT_GROUP_SCHED
181 struct sched_rt_entity
**rt_se
;
182 struct rt_rq
**rt_rq
;
188 struct list_head list
;
191 #ifdef CONFIG_FAIR_GROUP_SCHED
192 /* Default task group's sched entity on each cpu */
193 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
194 /* Default task group's cfs_rq on each cpu */
195 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
197 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
198 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
201 #ifdef CONFIG_RT_GROUP_SCHED
202 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
203 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
205 static struct sched_rt_entity
*init_sched_rt_entity_p
[NR_CPUS
];
206 static struct rt_rq
*init_rt_rq_p
[NR_CPUS
];
209 /* task_group_lock serializes add/remove of task groups and also changes to
210 * a task group's cpu shares.
212 static DEFINE_SPINLOCK(task_group_lock
);
214 /* doms_cur_mutex serializes access to doms_cur[] array */
215 static DEFINE_MUTEX(doms_cur_mutex
);
217 #ifdef CONFIG_FAIR_GROUP_SCHED
218 #ifdef CONFIG_USER_SCHED
219 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
221 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
224 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
227 /* Default task group.
228 * Every task in system belong to this group at bootup.
230 struct task_group init_task_group
= {
231 #ifdef CONFIG_FAIR_GROUP_SCHED
232 .se
= init_sched_entity_p
,
233 .cfs_rq
= init_cfs_rq_p
,
236 #ifdef CONFIG_RT_GROUP_SCHED
237 .rt_se
= init_sched_rt_entity_p
,
238 .rt_rq
= init_rt_rq_p
,
242 /* return group to which a task belongs */
243 static inline struct task_group
*task_group(struct task_struct
*p
)
245 struct task_group
*tg
;
247 #ifdef CONFIG_USER_SCHED
249 #elif defined(CONFIG_CGROUP_SCHED)
250 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
251 struct task_group
, css
);
253 tg
= &init_task_group
;
258 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
259 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
261 #ifdef CONFIG_FAIR_GROUP_SCHED
262 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
263 p
->se
.parent
= task_group(p
)->se
[cpu
];
266 #ifdef CONFIG_RT_GROUP_SCHED
267 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
268 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
272 static inline void lock_doms_cur(void)
274 mutex_lock(&doms_cur_mutex
);
277 static inline void unlock_doms_cur(void)
279 mutex_unlock(&doms_cur_mutex
);
284 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
285 static inline void lock_doms_cur(void) { }
286 static inline void unlock_doms_cur(void) { }
288 #endif /* CONFIG_GROUP_SCHED */
290 /* CFS-related fields in a runqueue */
292 struct load_weight load
;
293 unsigned long nr_running
;
298 struct rb_root tasks_timeline
;
299 struct rb_node
*rb_leftmost
;
300 struct rb_node
*rb_load_balance_curr
;
301 /* 'curr' points to currently running entity on this cfs_rq.
302 * It is set to NULL otherwise (i.e when none are currently running).
304 struct sched_entity
*curr
, *next
;
306 unsigned long nr_spread_over
;
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
312 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
313 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
314 * (like users, containers etc.)
316 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
317 * list is used during load balance.
319 struct list_head leaf_cfs_rq_list
;
320 struct task_group
*tg
; /* group that "owns" this runqueue */
324 /* Real-Time classes' related field in a runqueue: */
326 struct rt_prio_array active
;
327 unsigned long rt_nr_running
;
328 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
329 int highest_prio
; /* highest queued rt task prio */
332 unsigned long rt_nr_migratory
;
338 #ifdef CONFIG_RT_GROUP_SCHED
339 unsigned long rt_nr_boosted
;
342 struct list_head leaf_rt_rq_list
;
343 struct task_group
*tg
;
344 struct sched_rt_entity
*rt_se
;
351 * We add the notion of a root-domain which will be used to define per-domain
352 * variables. Each exclusive cpuset essentially defines an island domain by
353 * fully partitioning the member cpus from any other cpuset. Whenever a new
354 * exclusive cpuset is created, we also create and attach a new root-domain
364 * The "RT overload" flag: it gets set if a CPU has more than
365 * one runnable RT task.
372 * By default the system creates a single root-domain with all cpus as
373 * members (mimicking the global state we have today).
375 static struct root_domain def_root_domain
;
380 * This is the main, per-CPU runqueue data structure.
382 * Locking rule: those places that want to lock multiple runqueues
383 * (such as the load balancing or the thread migration code), lock
384 * acquire operations must be ordered by ascending &runqueue.
391 * nr_running and cpu_load should be in the same cacheline because
392 * remote CPUs use both these fields when doing load calculation.
394 unsigned long nr_running
;
395 #define CPU_LOAD_IDX_MAX 5
396 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
397 unsigned char idle_at_tick
;
399 unsigned char in_nohz_recently
;
401 /* capture load from *all* tasks on this cpu: */
402 struct load_weight load
;
403 unsigned long nr_load_updates
;
408 u64 rt_period_expire
;
411 #ifdef CONFIG_FAIR_GROUP_SCHED
412 /* list of leaf cfs_rq on this cpu: */
413 struct list_head leaf_cfs_rq_list
;
415 #ifdef CONFIG_RT_GROUP_SCHED
416 struct list_head leaf_rt_rq_list
;
420 * This is part of a global counter where only the total sum
421 * over all CPUs matters. A task can increase this counter on
422 * one CPU and if it got migrated afterwards it may decrease
423 * it on another CPU. Always updated under the runqueue lock:
425 unsigned long nr_uninterruptible
;
427 struct task_struct
*curr
, *idle
;
428 unsigned long next_balance
;
429 struct mm_struct
*prev_mm
;
431 u64 clock
, prev_clock_raw
;
434 unsigned int clock_warps
, clock_overflows
, clock_underflows
;
436 unsigned int clock_deep_idle_events
;
442 struct root_domain
*rd
;
443 struct sched_domain
*sd
;
445 /* For active balancing */
448 /* cpu of this runqueue: */
451 struct task_struct
*migration_thread
;
452 struct list_head migration_queue
;
455 #ifdef CONFIG_SCHED_HRTICK
456 unsigned long hrtick_flags
;
457 ktime_t hrtick_expire
;
458 struct hrtimer hrtick_timer
;
461 #ifdef CONFIG_SCHEDSTATS
463 struct sched_info rq_sched_info
;
465 /* sys_sched_yield() stats */
466 unsigned int yld_exp_empty
;
467 unsigned int yld_act_empty
;
468 unsigned int yld_both_empty
;
469 unsigned int yld_count
;
471 /* schedule() stats */
472 unsigned int sched_switch
;
473 unsigned int sched_count
;
474 unsigned int sched_goidle
;
476 /* try_to_wake_up() stats */
477 unsigned int ttwu_count
;
478 unsigned int ttwu_local
;
481 unsigned int bkl_count
;
483 struct lock_class_key rq_lock_key
;
486 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
488 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
490 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
493 static inline int cpu_of(struct rq
*rq
)
503 * Update the per-runqueue clock, as finegrained as the platform can give
504 * us, but without assuming monotonicity, etc.:
506 static void __update_rq_clock(struct rq
*rq
)
508 u64 prev_raw
= rq
->prev_clock_raw
;
509 u64 now
= sched_clock();
510 s64 delta
= now
- prev_raw
;
511 u64 clock
= rq
->clock
;
513 #ifdef CONFIG_SCHED_DEBUG
514 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
517 * Protect against sched_clock() occasionally going backwards:
519 if (unlikely(delta
< 0)) {
524 * Catch too large forward jumps too:
526 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
527 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
528 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
531 rq
->clock_overflows
++;
533 if (unlikely(delta
> rq
->clock_max_delta
))
534 rq
->clock_max_delta
= delta
;
539 rq
->prev_clock_raw
= now
;
543 static void update_rq_clock(struct rq
*rq
)
545 if (likely(smp_processor_id() == cpu_of(rq
)))
546 __update_rq_clock(rq
);
550 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
551 * See detach_destroy_domains: synchronize_sched for details.
553 * The domain tree of any CPU may only be accessed from within
554 * preempt-disabled sections.
556 #define for_each_domain(cpu, __sd) \
557 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
559 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
560 #define this_rq() (&__get_cpu_var(runqueues))
561 #define task_rq(p) cpu_rq(task_cpu(p))
562 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
564 unsigned long rt_needs_cpu(int cpu
)
566 struct rq
*rq
= cpu_rq(cpu
);
569 if (!rq
->rt_throttled
)
572 if (rq
->clock
> rq
->rt_period_expire
)
575 delta
= rq
->rt_period_expire
- rq
->clock
;
576 do_div(delta
, NSEC_PER_SEC
/ HZ
);
578 return (unsigned long)delta
;
582 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
584 #ifdef CONFIG_SCHED_DEBUG
585 # define const_debug __read_mostly
587 # define const_debug static const
591 * Debugging: various feature bits
594 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
595 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
596 SCHED_FEAT_START_DEBIT
= 4,
597 SCHED_FEAT_HRTICK
= 8,
598 SCHED_FEAT_DOUBLE_TICK
= 16,
601 const_debug
unsigned int sysctl_sched_features
=
602 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
603 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
604 SCHED_FEAT_START_DEBIT
* 1 |
605 SCHED_FEAT_HRTICK
* 1 |
606 SCHED_FEAT_DOUBLE_TICK
* 0;
608 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
611 * Number of tasks to iterate in a single balance run.
612 * Limited because this is done with IRQs disabled.
614 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
617 * period over which we measure -rt task cpu usage in us.
620 unsigned int sysctl_sched_rt_period
= 1000000;
622 static __read_mostly
int scheduler_running
;
625 * part of the period that we allow rt tasks to run in us.
628 int sysctl_sched_rt_runtime
= 950000;
631 * single value that denotes runtime == period, ie unlimited time.
633 #define RUNTIME_INF ((u64)~0ULL)
636 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
637 * clock constructed from sched_clock():
639 unsigned long long cpu_clock(int cpu
)
641 unsigned long long now
;
646 * Only call sched_clock() if the scheduler has already been
647 * initialized (some code might call cpu_clock() very early):
649 if (unlikely(!scheduler_running
))
652 local_irq_save(flags
);
656 local_irq_restore(flags
);
660 EXPORT_SYMBOL_GPL(cpu_clock
);
662 #ifndef prepare_arch_switch
663 # define prepare_arch_switch(next) do { } while (0)
665 #ifndef finish_arch_switch
666 # define finish_arch_switch(prev) do { } while (0)
669 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
671 return rq
->curr
== p
;
674 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
675 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
677 return task_current(rq
, p
);
680 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
684 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
686 #ifdef CONFIG_DEBUG_SPINLOCK
687 /* this is a valid case when another task releases the spinlock */
688 rq
->lock
.owner
= current
;
691 * If we are tracking spinlock dependencies then we have to
692 * fix up the runqueue lock - which gets 'carried over' from
695 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
697 spin_unlock_irq(&rq
->lock
);
700 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
701 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
706 return task_current(rq
, p
);
710 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
714 * We can optimise this out completely for !SMP, because the
715 * SMP rebalancing from interrupt is the only thing that cares
720 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
721 spin_unlock_irq(&rq
->lock
);
723 spin_unlock(&rq
->lock
);
727 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
731 * After ->oncpu is cleared, the task can be moved to a different CPU.
732 * We must ensure this doesn't happen until the switch is completely
738 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
742 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
745 * __task_rq_lock - lock the runqueue a given task resides on.
746 * Must be called interrupts disabled.
748 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
752 struct rq
*rq
= task_rq(p
);
753 spin_lock(&rq
->lock
);
754 if (likely(rq
== task_rq(p
)))
756 spin_unlock(&rq
->lock
);
761 * task_rq_lock - lock the runqueue a given task resides on and disable
762 * interrupts. Note the ordering: we can safely lookup the task_rq without
763 * explicitly disabling preemption.
765 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
771 local_irq_save(*flags
);
773 spin_lock(&rq
->lock
);
774 if (likely(rq
== task_rq(p
)))
776 spin_unlock_irqrestore(&rq
->lock
, *flags
);
780 static void __task_rq_unlock(struct rq
*rq
)
783 spin_unlock(&rq
->lock
);
786 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
789 spin_unlock_irqrestore(&rq
->lock
, *flags
);
793 * this_rq_lock - lock this runqueue and disable interrupts.
795 static struct rq
*this_rq_lock(void)
802 spin_lock(&rq
->lock
);
808 * We are going deep-idle (irqs are disabled):
810 void sched_clock_idle_sleep_event(void)
812 struct rq
*rq
= cpu_rq(smp_processor_id());
814 spin_lock(&rq
->lock
);
815 __update_rq_clock(rq
);
816 spin_unlock(&rq
->lock
);
817 rq
->clock_deep_idle_events
++;
819 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
822 * We just idled delta nanoseconds (called with irqs disabled):
824 void sched_clock_idle_wakeup_event(u64 delta_ns
)
826 struct rq
*rq
= cpu_rq(smp_processor_id());
827 u64 now
= sched_clock();
829 rq
->idle_clock
+= delta_ns
;
831 * Override the previous timestamp and ignore all
832 * sched_clock() deltas that occured while we idled,
833 * and use the PM-provided delta_ns to advance the
836 spin_lock(&rq
->lock
);
837 rq
->prev_clock_raw
= now
;
838 rq
->clock
+= delta_ns
;
839 spin_unlock(&rq
->lock
);
840 touch_softlockup_watchdog();
842 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
844 static void __resched_task(struct task_struct
*p
, int tif_bit
);
846 static inline void resched_task(struct task_struct
*p
)
848 __resched_task(p
, TIF_NEED_RESCHED
);
851 #ifdef CONFIG_SCHED_HRTICK
853 * Use HR-timers to deliver accurate preemption points.
855 * Its all a bit involved since we cannot program an hrt while holding the
856 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
859 * When we get rescheduled we reprogram the hrtick_timer outside of the
862 static inline void resched_hrt(struct task_struct
*p
)
864 __resched_task(p
, TIF_HRTICK_RESCHED
);
867 static inline void resched_rq(struct rq
*rq
)
871 spin_lock_irqsave(&rq
->lock
, flags
);
872 resched_task(rq
->curr
);
873 spin_unlock_irqrestore(&rq
->lock
, flags
);
877 HRTICK_SET
, /* re-programm hrtick_timer */
878 HRTICK_RESET
, /* not a new slice */
883 * - enabled by features
884 * - hrtimer is actually high res
886 static inline int hrtick_enabled(struct rq
*rq
)
888 if (!sched_feat(HRTICK
))
890 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
894 * Called to set the hrtick timer state.
896 * called with rq->lock held and irqs disabled
898 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
900 assert_spin_locked(&rq
->lock
);
903 * preempt at: now + delay
906 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
908 * indicate we need to program the timer
910 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
912 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
915 * New slices are called from the schedule path and don't need a
919 resched_hrt(rq
->curr
);
922 static void hrtick_clear(struct rq
*rq
)
924 if (hrtimer_active(&rq
->hrtick_timer
))
925 hrtimer_cancel(&rq
->hrtick_timer
);
929 * Update the timer from the possible pending state.
931 static void hrtick_set(struct rq
*rq
)
937 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
939 spin_lock_irqsave(&rq
->lock
, flags
);
940 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
941 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
942 time
= rq
->hrtick_expire
;
943 clear_thread_flag(TIF_HRTICK_RESCHED
);
944 spin_unlock_irqrestore(&rq
->lock
, flags
);
947 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
948 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
955 * High-resolution timer tick.
956 * Runs from hardirq context with interrupts disabled.
958 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
960 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
962 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
964 spin_lock(&rq
->lock
);
965 __update_rq_clock(rq
);
966 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
967 spin_unlock(&rq
->lock
);
969 return HRTIMER_NORESTART
;
972 static inline void init_rq_hrtick(struct rq
*rq
)
974 rq
->hrtick_flags
= 0;
975 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
976 rq
->hrtick_timer
.function
= hrtick
;
977 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
980 void hrtick_resched(void)
985 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
988 local_irq_save(flags
);
989 rq
= cpu_rq(smp_processor_id());
991 local_irq_restore(flags
);
994 static inline void hrtick_clear(struct rq
*rq
)
998 static inline void hrtick_set(struct rq
*rq
)
1002 static inline void init_rq_hrtick(struct rq
*rq
)
1006 void hrtick_resched(void)
1012 * resched_task - mark a task 'to be rescheduled now'.
1014 * On UP this means the setting of the need_resched flag, on SMP it
1015 * might also involve a cross-CPU call to trigger the scheduler on
1020 #ifndef tsk_is_polling
1021 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1024 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1028 assert_spin_locked(&task_rq(p
)->lock
);
1030 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1033 set_tsk_thread_flag(p
, tif_bit
);
1036 if (cpu
== smp_processor_id())
1039 /* NEED_RESCHED must be visible before we test polling */
1041 if (!tsk_is_polling(p
))
1042 smp_send_reschedule(cpu
);
1045 static void resched_cpu(int cpu
)
1047 struct rq
*rq
= cpu_rq(cpu
);
1048 unsigned long flags
;
1050 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1052 resched_task(cpu_curr(cpu
));
1053 spin_unlock_irqrestore(&rq
->lock
, flags
);
1056 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1058 assert_spin_locked(&task_rq(p
)->lock
);
1059 set_tsk_thread_flag(p
, tif_bit
);
1063 #if BITS_PER_LONG == 32
1064 # define WMULT_CONST (~0UL)
1066 # define WMULT_CONST (1UL << 32)
1069 #define WMULT_SHIFT 32
1072 * Shift right and round:
1074 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1076 static unsigned long
1077 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1078 struct load_weight
*lw
)
1082 if (unlikely(!lw
->inv_weight
))
1083 lw
->inv_weight
= (WMULT_CONST
-lw
->weight
/2) / (lw
->weight
+1);
1085 tmp
= (u64
)delta_exec
* weight
;
1087 * Check whether we'd overflow the 64-bit multiplication:
1089 if (unlikely(tmp
> WMULT_CONST
))
1090 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1093 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1095 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1098 static inline unsigned long
1099 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1101 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1104 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1110 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1117 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1118 * of tasks with abnormal "nice" values across CPUs the contribution that
1119 * each task makes to its run queue's load is weighted according to its
1120 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1121 * scaled version of the new time slice allocation that they receive on time
1125 #define WEIGHT_IDLEPRIO 2
1126 #define WMULT_IDLEPRIO (1 << 31)
1129 * Nice levels are multiplicative, with a gentle 10% change for every
1130 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1131 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1132 * that remained on nice 0.
1134 * The "10% effect" is relative and cumulative: from _any_ nice level,
1135 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1136 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1137 * If a task goes up by ~10% and another task goes down by ~10% then
1138 * the relative distance between them is ~25%.)
1140 static const int prio_to_weight
[40] = {
1141 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1142 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1143 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1144 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1145 /* 0 */ 1024, 820, 655, 526, 423,
1146 /* 5 */ 335, 272, 215, 172, 137,
1147 /* 10 */ 110, 87, 70, 56, 45,
1148 /* 15 */ 36, 29, 23, 18, 15,
1152 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1154 * In cases where the weight does not change often, we can use the
1155 * precalculated inverse to speed up arithmetics by turning divisions
1156 * into multiplications:
1158 static const u32 prio_to_wmult
[40] = {
1159 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1160 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1161 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1162 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1163 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1164 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1165 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1166 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1169 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1172 * runqueue iterator, to support SMP load-balancing between different
1173 * scheduling classes, without having to expose their internal data
1174 * structures to the load-balancing proper:
1176 struct rq_iterator
{
1178 struct task_struct
*(*start
)(void *);
1179 struct task_struct
*(*next
)(void *);
1183 static unsigned long
1184 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1185 unsigned long max_load_move
, struct sched_domain
*sd
,
1186 enum cpu_idle_type idle
, int *all_pinned
,
1187 int *this_best_prio
, struct rq_iterator
*iterator
);
1190 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1191 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1192 struct rq_iterator
*iterator
);
1195 #ifdef CONFIG_CGROUP_CPUACCT
1196 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1198 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1202 static unsigned long source_load(int cpu
, int type
);
1203 static unsigned long target_load(int cpu
, int type
);
1204 static unsigned long cpu_avg_load_per_task(int cpu
);
1205 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1206 #endif /* CONFIG_SMP */
1208 #include "sched_stats.h"
1209 #include "sched_idletask.c"
1210 #include "sched_fair.c"
1211 #include "sched_rt.c"
1212 #ifdef CONFIG_SCHED_DEBUG
1213 # include "sched_debug.c"
1216 #define sched_class_highest (&rt_sched_class)
1218 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
1220 update_load_add(&rq
->load
, p
->se
.load
.weight
);
1223 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
1225 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
1228 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1234 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1240 static void set_load_weight(struct task_struct
*p
)
1242 if (task_has_rt_policy(p
)) {
1243 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1244 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1249 * SCHED_IDLE tasks get minimal weight:
1251 if (p
->policy
== SCHED_IDLE
) {
1252 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1253 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1257 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1258 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1261 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1263 sched_info_queued(p
);
1264 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1268 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1270 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1275 * __normal_prio - return the priority that is based on the static prio
1277 static inline int __normal_prio(struct task_struct
*p
)
1279 return p
->static_prio
;
1283 * Calculate the expected normal priority: i.e. priority
1284 * without taking RT-inheritance into account. Might be
1285 * boosted by interactivity modifiers. Changes upon fork,
1286 * setprio syscalls, and whenever the interactivity
1287 * estimator recalculates.
1289 static inline int normal_prio(struct task_struct
*p
)
1293 if (task_has_rt_policy(p
))
1294 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1296 prio
= __normal_prio(p
);
1301 * Calculate the current priority, i.e. the priority
1302 * taken into account by the scheduler. This value might
1303 * be boosted by RT tasks, or might be boosted by
1304 * interactivity modifiers. Will be RT if the task got
1305 * RT-boosted. If not then it returns p->normal_prio.
1307 static int effective_prio(struct task_struct
*p
)
1309 p
->normal_prio
= normal_prio(p
);
1311 * If we are RT tasks or we were boosted to RT priority,
1312 * keep the priority unchanged. Otherwise, update priority
1313 * to the normal priority:
1315 if (!rt_prio(p
->prio
))
1316 return p
->normal_prio
;
1321 * activate_task - move a task to the runqueue.
1323 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1325 if (task_contributes_to_load(p
))
1326 rq
->nr_uninterruptible
--;
1328 enqueue_task(rq
, p
, wakeup
);
1329 inc_nr_running(p
, rq
);
1333 * deactivate_task - remove a task from the runqueue.
1335 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1337 if (task_contributes_to_load(p
))
1338 rq
->nr_uninterruptible
++;
1340 dequeue_task(rq
, p
, sleep
);
1341 dec_nr_running(p
, rq
);
1345 * task_curr - is this task currently executing on a CPU?
1346 * @p: the task in question.
1348 inline int task_curr(const struct task_struct
*p
)
1350 return cpu_curr(task_cpu(p
)) == p
;
1353 /* Used instead of source_load when we know the type == 0 */
1354 unsigned long weighted_cpuload(const int cpu
)
1356 return cpu_rq(cpu
)->load
.weight
;
1359 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1361 set_task_rq(p
, cpu
);
1364 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1365 * successfuly executed on another CPU. We must ensure that updates of
1366 * per-task data have been completed by this moment.
1369 task_thread_info(p
)->cpu
= cpu
;
1373 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1374 const struct sched_class
*prev_class
,
1375 int oldprio
, int running
)
1377 if (prev_class
!= p
->sched_class
) {
1378 if (prev_class
->switched_from
)
1379 prev_class
->switched_from(rq
, p
, running
);
1380 p
->sched_class
->switched_to(rq
, p
, running
);
1382 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1388 * Is this task likely cache-hot:
1391 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1396 * Buddy candidates are cache hot:
1398 if (&p
->se
== cfs_rq_of(&p
->se
)->next
)
1401 if (p
->sched_class
!= &fair_sched_class
)
1404 if (sysctl_sched_migration_cost
== -1)
1406 if (sysctl_sched_migration_cost
== 0)
1409 delta
= now
- p
->se
.exec_start
;
1411 return delta
< (s64
)sysctl_sched_migration_cost
;
1415 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1417 int old_cpu
= task_cpu(p
);
1418 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1419 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1420 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1423 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1425 #ifdef CONFIG_SCHEDSTATS
1426 if (p
->se
.wait_start
)
1427 p
->se
.wait_start
-= clock_offset
;
1428 if (p
->se
.sleep_start
)
1429 p
->se
.sleep_start
-= clock_offset
;
1430 if (p
->se
.block_start
)
1431 p
->se
.block_start
-= clock_offset
;
1432 if (old_cpu
!= new_cpu
) {
1433 schedstat_inc(p
, se
.nr_migrations
);
1434 if (task_hot(p
, old_rq
->clock
, NULL
))
1435 schedstat_inc(p
, se
.nr_forced2_migrations
);
1438 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1439 new_cfsrq
->min_vruntime
;
1441 __set_task_cpu(p
, new_cpu
);
1444 struct migration_req
{
1445 struct list_head list
;
1447 struct task_struct
*task
;
1450 struct completion done
;
1454 * The task's runqueue lock must be held.
1455 * Returns true if you have to wait for migration thread.
1458 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1460 struct rq
*rq
= task_rq(p
);
1463 * If the task is not on a runqueue (and not running), then
1464 * it is sufficient to simply update the task's cpu field.
1466 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1467 set_task_cpu(p
, dest_cpu
);
1471 init_completion(&req
->done
);
1473 req
->dest_cpu
= dest_cpu
;
1474 list_add(&req
->list
, &rq
->migration_queue
);
1480 * wait_task_inactive - wait for a thread to unschedule.
1482 * The caller must ensure that the task *will* unschedule sometime soon,
1483 * else this function might spin for a *long* time. This function can't
1484 * be called with interrupts off, or it may introduce deadlock with
1485 * smp_call_function() if an IPI is sent by the same process we are
1486 * waiting to become inactive.
1488 void wait_task_inactive(struct task_struct
*p
)
1490 unsigned long flags
;
1496 * We do the initial early heuristics without holding
1497 * any task-queue locks at all. We'll only try to get
1498 * the runqueue lock when things look like they will
1504 * If the task is actively running on another CPU
1505 * still, just relax and busy-wait without holding
1508 * NOTE! Since we don't hold any locks, it's not
1509 * even sure that "rq" stays as the right runqueue!
1510 * But we don't care, since "task_running()" will
1511 * return false if the runqueue has changed and p
1512 * is actually now running somewhere else!
1514 while (task_running(rq
, p
))
1518 * Ok, time to look more closely! We need the rq
1519 * lock now, to be *sure*. If we're wrong, we'll
1520 * just go back and repeat.
1522 rq
= task_rq_lock(p
, &flags
);
1523 running
= task_running(rq
, p
);
1524 on_rq
= p
->se
.on_rq
;
1525 task_rq_unlock(rq
, &flags
);
1528 * Was it really running after all now that we
1529 * checked with the proper locks actually held?
1531 * Oops. Go back and try again..
1533 if (unlikely(running
)) {
1539 * It's not enough that it's not actively running,
1540 * it must be off the runqueue _entirely_, and not
1543 * So if it wa still runnable (but just not actively
1544 * running right now), it's preempted, and we should
1545 * yield - it could be a while.
1547 if (unlikely(on_rq
)) {
1548 schedule_timeout_uninterruptible(1);
1553 * Ahh, all good. It wasn't running, and it wasn't
1554 * runnable, which means that it will never become
1555 * running in the future either. We're all done!
1562 * kick_process - kick a running thread to enter/exit the kernel
1563 * @p: the to-be-kicked thread
1565 * Cause a process which is running on another CPU to enter
1566 * kernel-mode, without any delay. (to get signals handled.)
1568 * NOTE: this function doesnt have to take the runqueue lock,
1569 * because all it wants to ensure is that the remote task enters
1570 * the kernel. If the IPI races and the task has been migrated
1571 * to another CPU then no harm is done and the purpose has been
1574 void kick_process(struct task_struct
*p
)
1580 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1581 smp_send_reschedule(cpu
);
1586 * Return a low guess at the load of a migration-source cpu weighted
1587 * according to the scheduling class and "nice" value.
1589 * We want to under-estimate the load of migration sources, to
1590 * balance conservatively.
1592 static unsigned long source_load(int cpu
, int type
)
1594 struct rq
*rq
= cpu_rq(cpu
);
1595 unsigned long total
= weighted_cpuload(cpu
);
1600 return min(rq
->cpu_load
[type
-1], total
);
1604 * Return a high guess at the load of a migration-target cpu weighted
1605 * according to the scheduling class and "nice" value.
1607 static unsigned long target_load(int cpu
, int type
)
1609 struct rq
*rq
= cpu_rq(cpu
);
1610 unsigned long total
= weighted_cpuload(cpu
);
1615 return max(rq
->cpu_load
[type
-1], total
);
1619 * Return the average load per task on the cpu's run queue
1621 static unsigned long cpu_avg_load_per_task(int cpu
)
1623 struct rq
*rq
= cpu_rq(cpu
);
1624 unsigned long total
= weighted_cpuload(cpu
);
1625 unsigned long n
= rq
->nr_running
;
1627 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1631 * find_idlest_group finds and returns the least busy CPU group within the
1634 static struct sched_group
*
1635 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1637 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1638 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1639 int load_idx
= sd
->forkexec_idx
;
1640 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1643 unsigned long load
, avg_load
;
1647 /* Skip over this group if it has no CPUs allowed */
1648 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1651 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1653 /* Tally up the load of all CPUs in the group */
1656 for_each_cpu_mask(i
, group
->cpumask
) {
1657 /* Bias balancing toward cpus of our domain */
1659 load
= source_load(i
, load_idx
);
1661 load
= target_load(i
, load_idx
);
1666 /* Adjust by relative CPU power of the group */
1667 avg_load
= sg_div_cpu_power(group
,
1668 avg_load
* SCHED_LOAD_SCALE
);
1671 this_load
= avg_load
;
1673 } else if (avg_load
< min_load
) {
1674 min_load
= avg_load
;
1677 } while (group
= group
->next
, group
!= sd
->groups
);
1679 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1685 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1688 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1691 unsigned long load
, min_load
= ULONG_MAX
;
1695 /* Traverse only the allowed CPUs */
1696 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1698 for_each_cpu_mask(i
, tmp
) {
1699 load
= weighted_cpuload(i
);
1701 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1711 * sched_balance_self: balance the current task (running on cpu) in domains
1712 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1715 * Balance, ie. select the least loaded group.
1717 * Returns the target CPU number, or the same CPU if no balancing is needed.
1719 * preempt must be disabled.
1721 static int sched_balance_self(int cpu
, int flag
)
1723 struct task_struct
*t
= current
;
1724 struct sched_domain
*tmp
, *sd
= NULL
;
1726 for_each_domain(cpu
, tmp
) {
1728 * If power savings logic is enabled for a domain, stop there.
1730 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1732 if (tmp
->flags
& flag
)
1738 struct sched_group
*group
;
1739 int new_cpu
, weight
;
1741 if (!(sd
->flags
& flag
)) {
1747 group
= find_idlest_group(sd
, t
, cpu
);
1753 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1754 if (new_cpu
== -1 || new_cpu
== cpu
) {
1755 /* Now try balancing at a lower domain level of cpu */
1760 /* Now try balancing at a lower domain level of new_cpu */
1763 weight
= cpus_weight(span
);
1764 for_each_domain(cpu
, tmp
) {
1765 if (weight
<= cpus_weight(tmp
->span
))
1767 if (tmp
->flags
& flag
)
1770 /* while loop will break here if sd == NULL */
1776 #endif /* CONFIG_SMP */
1779 * try_to_wake_up - wake up a thread
1780 * @p: the to-be-woken-up thread
1781 * @state: the mask of task states that can be woken
1782 * @sync: do a synchronous wakeup?
1784 * Put it on the run-queue if it's not already there. The "current"
1785 * thread is always on the run-queue (except when the actual
1786 * re-schedule is in progress), and as such you're allowed to do
1787 * the simpler "current->state = TASK_RUNNING" to mark yourself
1788 * runnable without the overhead of this.
1790 * returns failure only if the task is already active.
1792 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1794 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1795 unsigned long flags
;
1800 rq
= task_rq_lock(p
, &flags
);
1801 old_state
= p
->state
;
1802 if (!(old_state
& state
))
1810 this_cpu
= smp_processor_id();
1813 if (unlikely(task_running(rq
, p
)))
1816 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1817 if (cpu
!= orig_cpu
) {
1818 set_task_cpu(p
, cpu
);
1819 task_rq_unlock(rq
, &flags
);
1820 /* might preempt at this point */
1821 rq
= task_rq_lock(p
, &flags
);
1822 old_state
= p
->state
;
1823 if (!(old_state
& state
))
1828 this_cpu
= smp_processor_id();
1832 #ifdef CONFIG_SCHEDSTATS
1833 schedstat_inc(rq
, ttwu_count
);
1834 if (cpu
== this_cpu
)
1835 schedstat_inc(rq
, ttwu_local
);
1837 struct sched_domain
*sd
;
1838 for_each_domain(this_cpu
, sd
) {
1839 if (cpu_isset(cpu
, sd
->span
)) {
1840 schedstat_inc(sd
, ttwu_wake_remote
);
1848 #endif /* CONFIG_SMP */
1849 schedstat_inc(p
, se
.nr_wakeups
);
1851 schedstat_inc(p
, se
.nr_wakeups_sync
);
1852 if (orig_cpu
!= cpu
)
1853 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1854 if (cpu
== this_cpu
)
1855 schedstat_inc(p
, se
.nr_wakeups_local
);
1857 schedstat_inc(p
, se
.nr_wakeups_remote
);
1858 update_rq_clock(rq
);
1859 activate_task(rq
, p
, 1);
1863 check_preempt_curr(rq
, p
);
1865 p
->state
= TASK_RUNNING
;
1867 if (p
->sched_class
->task_wake_up
)
1868 p
->sched_class
->task_wake_up(rq
, p
);
1871 task_rq_unlock(rq
, &flags
);
1876 int wake_up_process(struct task_struct
*p
)
1878 return try_to_wake_up(p
, TASK_ALL
, 0);
1880 EXPORT_SYMBOL(wake_up_process
);
1882 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1884 return try_to_wake_up(p
, state
, 0);
1888 * Perform scheduler related setup for a newly forked process p.
1889 * p is forked by current.
1891 * __sched_fork() is basic setup used by init_idle() too:
1893 static void __sched_fork(struct task_struct
*p
)
1895 p
->se
.exec_start
= 0;
1896 p
->se
.sum_exec_runtime
= 0;
1897 p
->se
.prev_sum_exec_runtime
= 0;
1898 p
->se
.last_wakeup
= 0;
1899 p
->se
.avg_overlap
= 0;
1901 #ifdef CONFIG_SCHEDSTATS
1902 p
->se
.wait_start
= 0;
1903 p
->se
.sum_sleep_runtime
= 0;
1904 p
->se
.sleep_start
= 0;
1905 p
->se
.block_start
= 0;
1906 p
->se
.sleep_max
= 0;
1907 p
->se
.block_max
= 0;
1909 p
->se
.slice_max
= 0;
1913 INIT_LIST_HEAD(&p
->rt
.run_list
);
1916 #ifdef CONFIG_PREEMPT_NOTIFIERS
1917 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1921 * We mark the process as running here, but have not actually
1922 * inserted it onto the runqueue yet. This guarantees that
1923 * nobody will actually run it, and a signal or other external
1924 * event cannot wake it up and insert it on the runqueue either.
1926 p
->state
= TASK_RUNNING
;
1930 * fork()/clone()-time setup:
1932 void sched_fork(struct task_struct
*p
, int clone_flags
)
1934 int cpu
= get_cpu();
1939 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1941 set_task_cpu(p
, cpu
);
1944 * Make sure we do not leak PI boosting priority to the child:
1946 p
->prio
= current
->normal_prio
;
1947 if (!rt_prio(p
->prio
))
1948 p
->sched_class
= &fair_sched_class
;
1950 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1951 if (likely(sched_info_on()))
1952 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1954 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1957 #ifdef CONFIG_PREEMPT
1958 /* Want to start with kernel preemption disabled. */
1959 task_thread_info(p
)->preempt_count
= 1;
1965 * wake_up_new_task - wake up a newly created task for the first time.
1967 * This function will do some initial scheduler statistics housekeeping
1968 * that must be done for every newly created context, then puts the task
1969 * on the runqueue and wakes it.
1971 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1973 unsigned long flags
;
1976 rq
= task_rq_lock(p
, &flags
);
1977 BUG_ON(p
->state
!= TASK_RUNNING
);
1978 update_rq_clock(rq
);
1980 p
->prio
= effective_prio(p
);
1982 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1983 activate_task(rq
, p
, 0);
1986 * Let the scheduling class do new task startup
1987 * management (if any):
1989 p
->sched_class
->task_new(rq
, p
);
1990 inc_nr_running(p
, rq
);
1992 check_preempt_curr(rq
, p
);
1994 if (p
->sched_class
->task_wake_up
)
1995 p
->sched_class
->task_wake_up(rq
, p
);
1997 task_rq_unlock(rq
, &flags
);
2000 #ifdef CONFIG_PREEMPT_NOTIFIERS
2003 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2004 * @notifier: notifier struct to register
2006 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2008 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2010 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2013 * preempt_notifier_unregister - no longer interested in preemption notifications
2014 * @notifier: notifier struct to unregister
2016 * This is safe to call from within a preemption notifier.
2018 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2020 hlist_del(¬ifier
->link
);
2022 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2024 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2026 struct preempt_notifier
*notifier
;
2027 struct hlist_node
*node
;
2029 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2030 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2034 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2035 struct task_struct
*next
)
2037 struct preempt_notifier
*notifier
;
2038 struct hlist_node
*node
;
2040 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2041 notifier
->ops
->sched_out(notifier
, next
);
2046 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2051 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2052 struct task_struct
*next
)
2059 * prepare_task_switch - prepare to switch tasks
2060 * @rq: the runqueue preparing to switch
2061 * @prev: the current task that is being switched out
2062 * @next: the task we are going to switch to.
2064 * This is called with the rq lock held and interrupts off. It must
2065 * be paired with a subsequent finish_task_switch after the context
2068 * prepare_task_switch sets up locking and calls architecture specific
2072 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2073 struct task_struct
*next
)
2075 fire_sched_out_preempt_notifiers(prev
, next
);
2076 prepare_lock_switch(rq
, next
);
2077 prepare_arch_switch(next
);
2081 * finish_task_switch - clean up after a task-switch
2082 * @rq: runqueue associated with task-switch
2083 * @prev: the thread we just switched away from.
2085 * finish_task_switch must be called after the context switch, paired
2086 * with a prepare_task_switch call before the context switch.
2087 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2088 * and do any other architecture-specific cleanup actions.
2090 * Note that we may have delayed dropping an mm in context_switch(). If
2091 * so, we finish that here outside of the runqueue lock. (Doing it
2092 * with the lock held can cause deadlocks; see schedule() for
2095 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2096 __releases(rq
->lock
)
2098 struct mm_struct
*mm
= rq
->prev_mm
;
2104 * A task struct has one reference for the use as "current".
2105 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2106 * schedule one last time. The schedule call will never return, and
2107 * the scheduled task must drop that reference.
2108 * The test for TASK_DEAD must occur while the runqueue locks are
2109 * still held, otherwise prev could be scheduled on another cpu, die
2110 * there before we look at prev->state, and then the reference would
2112 * Manfred Spraul <manfred@colorfullife.com>
2114 prev_state
= prev
->state
;
2115 finish_arch_switch(prev
);
2116 finish_lock_switch(rq
, prev
);
2118 if (current
->sched_class
->post_schedule
)
2119 current
->sched_class
->post_schedule(rq
);
2122 fire_sched_in_preempt_notifiers(current
);
2125 if (unlikely(prev_state
== TASK_DEAD
)) {
2127 * Remove function-return probe instances associated with this
2128 * task and put them back on the free list.
2130 kprobe_flush_task(prev
);
2131 put_task_struct(prev
);
2136 * schedule_tail - first thing a freshly forked thread must call.
2137 * @prev: the thread we just switched away from.
2139 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2140 __releases(rq
->lock
)
2142 struct rq
*rq
= this_rq();
2144 finish_task_switch(rq
, prev
);
2145 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2146 /* In this case, finish_task_switch does not reenable preemption */
2149 if (current
->set_child_tid
)
2150 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2154 * context_switch - switch to the new MM and the new
2155 * thread's register state.
2158 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2159 struct task_struct
*next
)
2161 struct mm_struct
*mm
, *oldmm
;
2163 prepare_task_switch(rq
, prev
, next
);
2165 oldmm
= prev
->active_mm
;
2167 * For paravirt, this is coupled with an exit in switch_to to
2168 * combine the page table reload and the switch backend into
2171 arch_enter_lazy_cpu_mode();
2173 if (unlikely(!mm
)) {
2174 next
->active_mm
= oldmm
;
2175 atomic_inc(&oldmm
->mm_count
);
2176 enter_lazy_tlb(oldmm
, next
);
2178 switch_mm(oldmm
, mm
, next
);
2180 if (unlikely(!prev
->mm
)) {
2181 prev
->active_mm
= NULL
;
2182 rq
->prev_mm
= oldmm
;
2185 * Since the runqueue lock will be released by the next
2186 * task (which is an invalid locking op but in the case
2187 * of the scheduler it's an obvious special-case), so we
2188 * do an early lockdep release here:
2190 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2191 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2194 /* Here we just switch the register state and the stack. */
2195 switch_to(prev
, next
, prev
);
2199 * this_rq must be evaluated again because prev may have moved
2200 * CPUs since it called schedule(), thus the 'rq' on its stack
2201 * frame will be invalid.
2203 finish_task_switch(this_rq(), prev
);
2207 * nr_running, nr_uninterruptible and nr_context_switches:
2209 * externally visible scheduler statistics: current number of runnable
2210 * threads, current number of uninterruptible-sleeping threads, total
2211 * number of context switches performed since bootup.
2213 unsigned long nr_running(void)
2215 unsigned long i
, sum
= 0;
2217 for_each_online_cpu(i
)
2218 sum
+= cpu_rq(i
)->nr_running
;
2223 unsigned long nr_uninterruptible(void)
2225 unsigned long i
, sum
= 0;
2227 for_each_possible_cpu(i
)
2228 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2231 * Since we read the counters lockless, it might be slightly
2232 * inaccurate. Do not allow it to go below zero though:
2234 if (unlikely((long)sum
< 0))
2240 unsigned long long nr_context_switches(void)
2243 unsigned long long sum
= 0;
2245 for_each_possible_cpu(i
)
2246 sum
+= cpu_rq(i
)->nr_switches
;
2251 unsigned long nr_iowait(void)
2253 unsigned long i
, sum
= 0;
2255 for_each_possible_cpu(i
)
2256 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2261 unsigned long nr_active(void)
2263 unsigned long i
, running
= 0, uninterruptible
= 0;
2265 for_each_online_cpu(i
) {
2266 running
+= cpu_rq(i
)->nr_running
;
2267 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2270 if (unlikely((long)uninterruptible
< 0))
2271 uninterruptible
= 0;
2273 return running
+ uninterruptible
;
2277 * Update rq->cpu_load[] statistics. This function is usually called every
2278 * scheduler tick (TICK_NSEC).
2280 static void update_cpu_load(struct rq
*this_rq
)
2282 unsigned long this_load
= this_rq
->load
.weight
;
2285 this_rq
->nr_load_updates
++;
2287 /* Update our load: */
2288 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2289 unsigned long old_load
, new_load
;
2291 /* scale is effectively 1 << i now, and >> i divides by scale */
2293 old_load
= this_rq
->cpu_load
[i
];
2294 new_load
= this_load
;
2296 * Round up the averaging division if load is increasing. This
2297 * prevents us from getting stuck on 9 if the load is 10, for
2300 if (new_load
> old_load
)
2301 new_load
+= scale
-1;
2302 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2309 * double_rq_lock - safely lock two runqueues
2311 * Note this does not disable interrupts like task_rq_lock,
2312 * you need to do so manually before calling.
2314 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2315 __acquires(rq1
->lock
)
2316 __acquires(rq2
->lock
)
2318 BUG_ON(!irqs_disabled());
2320 spin_lock(&rq1
->lock
);
2321 __acquire(rq2
->lock
); /* Fake it out ;) */
2324 spin_lock(&rq1
->lock
);
2325 spin_lock(&rq2
->lock
);
2327 spin_lock(&rq2
->lock
);
2328 spin_lock(&rq1
->lock
);
2331 update_rq_clock(rq1
);
2332 update_rq_clock(rq2
);
2336 * double_rq_unlock - safely unlock two runqueues
2338 * Note this does not restore interrupts like task_rq_unlock,
2339 * you need to do so manually after calling.
2341 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2342 __releases(rq1
->lock
)
2343 __releases(rq2
->lock
)
2345 spin_unlock(&rq1
->lock
);
2347 spin_unlock(&rq2
->lock
);
2349 __release(rq2
->lock
);
2353 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2355 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2356 __releases(this_rq
->lock
)
2357 __acquires(busiest
->lock
)
2358 __acquires(this_rq
->lock
)
2362 if (unlikely(!irqs_disabled())) {
2363 /* printk() doesn't work good under rq->lock */
2364 spin_unlock(&this_rq
->lock
);
2367 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2368 if (busiest
< this_rq
) {
2369 spin_unlock(&this_rq
->lock
);
2370 spin_lock(&busiest
->lock
);
2371 spin_lock(&this_rq
->lock
);
2374 spin_lock(&busiest
->lock
);
2380 * If dest_cpu is allowed for this process, migrate the task to it.
2381 * This is accomplished by forcing the cpu_allowed mask to only
2382 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2383 * the cpu_allowed mask is restored.
2385 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2387 struct migration_req req
;
2388 unsigned long flags
;
2391 rq
= task_rq_lock(p
, &flags
);
2392 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2393 || unlikely(cpu_is_offline(dest_cpu
)))
2396 /* force the process onto the specified CPU */
2397 if (migrate_task(p
, dest_cpu
, &req
)) {
2398 /* Need to wait for migration thread (might exit: take ref). */
2399 struct task_struct
*mt
= rq
->migration_thread
;
2401 get_task_struct(mt
);
2402 task_rq_unlock(rq
, &flags
);
2403 wake_up_process(mt
);
2404 put_task_struct(mt
);
2405 wait_for_completion(&req
.done
);
2410 task_rq_unlock(rq
, &flags
);
2414 * sched_exec - execve() is a valuable balancing opportunity, because at
2415 * this point the task has the smallest effective memory and cache footprint.
2417 void sched_exec(void)
2419 int new_cpu
, this_cpu
= get_cpu();
2420 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2422 if (new_cpu
!= this_cpu
)
2423 sched_migrate_task(current
, new_cpu
);
2427 * pull_task - move a task from a remote runqueue to the local runqueue.
2428 * Both runqueues must be locked.
2430 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2431 struct rq
*this_rq
, int this_cpu
)
2433 deactivate_task(src_rq
, p
, 0);
2434 set_task_cpu(p
, this_cpu
);
2435 activate_task(this_rq
, p
, 0);
2437 * Note that idle threads have a prio of MAX_PRIO, for this test
2438 * to be always true for them.
2440 check_preempt_curr(this_rq
, p
);
2444 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2447 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2448 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2452 * We do not migrate tasks that are:
2453 * 1) running (obviously), or
2454 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2455 * 3) are cache-hot on their current CPU.
2457 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2458 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2463 if (task_running(rq
, p
)) {
2464 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2469 * Aggressive migration if:
2470 * 1) task is cache cold, or
2471 * 2) too many balance attempts have failed.
2474 if (!task_hot(p
, rq
->clock
, sd
) ||
2475 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2476 #ifdef CONFIG_SCHEDSTATS
2477 if (task_hot(p
, rq
->clock
, sd
)) {
2478 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2479 schedstat_inc(p
, se
.nr_forced_migrations
);
2485 if (task_hot(p
, rq
->clock
, sd
)) {
2486 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2492 static unsigned long
2493 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2494 unsigned long max_load_move
, struct sched_domain
*sd
,
2495 enum cpu_idle_type idle
, int *all_pinned
,
2496 int *this_best_prio
, struct rq_iterator
*iterator
)
2498 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2499 struct task_struct
*p
;
2500 long rem_load_move
= max_load_move
;
2502 if (max_load_move
== 0)
2508 * Start the load-balancing iterator:
2510 p
= iterator
->start(iterator
->arg
);
2512 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2515 * To help distribute high priority tasks across CPUs we don't
2516 * skip a task if it will be the highest priority task (i.e. smallest
2517 * prio value) on its new queue regardless of its load weight
2519 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2520 SCHED_LOAD_SCALE_FUZZ
;
2521 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2522 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2523 p
= iterator
->next(iterator
->arg
);
2527 pull_task(busiest
, p
, this_rq
, this_cpu
);
2529 rem_load_move
-= p
->se
.load
.weight
;
2532 * We only want to steal up to the prescribed amount of weighted load.
2534 if (rem_load_move
> 0) {
2535 if (p
->prio
< *this_best_prio
)
2536 *this_best_prio
= p
->prio
;
2537 p
= iterator
->next(iterator
->arg
);
2542 * Right now, this is one of only two places pull_task() is called,
2543 * so we can safely collect pull_task() stats here rather than
2544 * inside pull_task().
2546 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2549 *all_pinned
= pinned
;
2551 return max_load_move
- rem_load_move
;
2555 * move_tasks tries to move up to max_load_move weighted load from busiest to
2556 * this_rq, as part of a balancing operation within domain "sd".
2557 * Returns 1 if successful and 0 otherwise.
2559 * Called with both runqueues locked.
2561 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2562 unsigned long max_load_move
,
2563 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2566 const struct sched_class
*class = sched_class_highest
;
2567 unsigned long total_load_moved
= 0;
2568 int this_best_prio
= this_rq
->curr
->prio
;
2572 class->load_balance(this_rq
, this_cpu
, busiest
,
2573 max_load_move
- total_load_moved
,
2574 sd
, idle
, all_pinned
, &this_best_prio
);
2575 class = class->next
;
2576 } while (class && max_load_move
> total_load_moved
);
2578 return total_load_moved
> 0;
2582 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2583 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2584 struct rq_iterator
*iterator
)
2586 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2590 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2591 pull_task(busiest
, p
, this_rq
, this_cpu
);
2593 * Right now, this is only the second place pull_task()
2594 * is called, so we can safely collect pull_task()
2595 * stats here rather than inside pull_task().
2597 schedstat_inc(sd
, lb_gained
[idle
]);
2601 p
= iterator
->next(iterator
->arg
);
2608 * move_one_task tries to move exactly one task from busiest to this_rq, as
2609 * part of active balancing operations within "domain".
2610 * Returns 1 if successful and 0 otherwise.
2612 * Called with both runqueues locked.
2614 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2615 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2617 const struct sched_class
*class;
2619 for (class = sched_class_highest
; class; class = class->next
)
2620 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2627 * find_busiest_group finds and returns the busiest CPU group within the
2628 * domain. It calculates and returns the amount of weighted load which
2629 * should be moved to restore balance via the imbalance parameter.
2631 static struct sched_group
*
2632 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2633 unsigned long *imbalance
, enum cpu_idle_type idle
,
2634 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2636 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2637 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2638 unsigned long max_pull
;
2639 unsigned long busiest_load_per_task
, busiest_nr_running
;
2640 unsigned long this_load_per_task
, this_nr_running
;
2641 int load_idx
, group_imb
= 0;
2642 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2643 int power_savings_balance
= 1;
2644 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2645 unsigned long min_nr_running
= ULONG_MAX
;
2646 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2649 max_load
= this_load
= total_load
= total_pwr
= 0;
2650 busiest_load_per_task
= busiest_nr_running
= 0;
2651 this_load_per_task
= this_nr_running
= 0;
2652 if (idle
== CPU_NOT_IDLE
)
2653 load_idx
= sd
->busy_idx
;
2654 else if (idle
== CPU_NEWLY_IDLE
)
2655 load_idx
= sd
->newidle_idx
;
2657 load_idx
= sd
->idle_idx
;
2660 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2663 int __group_imb
= 0;
2664 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2665 unsigned long sum_nr_running
, sum_weighted_load
;
2667 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2670 balance_cpu
= first_cpu(group
->cpumask
);
2672 /* Tally up the load of all CPUs in the group */
2673 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2675 min_cpu_load
= ~0UL;
2677 for_each_cpu_mask(i
, group
->cpumask
) {
2680 if (!cpu_isset(i
, *cpus
))
2685 if (*sd_idle
&& rq
->nr_running
)
2688 /* Bias balancing toward cpus of our domain */
2690 if (idle_cpu(i
) && !first_idle_cpu
) {
2695 load
= target_load(i
, load_idx
);
2697 load
= source_load(i
, load_idx
);
2698 if (load
> max_cpu_load
)
2699 max_cpu_load
= load
;
2700 if (min_cpu_load
> load
)
2701 min_cpu_load
= load
;
2705 sum_nr_running
+= rq
->nr_running
;
2706 sum_weighted_load
+= weighted_cpuload(i
);
2710 * First idle cpu or the first cpu(busiest) in this sched group
2711 * is eligible for doing load balancing at this and above
2712 * domains. In the newly idle case, we will allow all the cpu's
2713 * to do the newly idle load balance.
2715 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2716 balance_cpu
!= this_cpu
&& balance
) {
2721 total_load
+= avg_load
;
2722 total_pwr
+= group
->__cpu_power
;
2724 /* Adjust by relative CPU power of the group */
2725 avg_load
= sg_div_cpu_power(group
,
2726 avg_load
* SCHED_LOAD_SCALE
);
2728 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2731 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2734 this_load
= avg_load
;
2736 this_nr_running
= sum_nr_running
;
2737 this_load_per_task
= sum_weighted_load
;
2738 } else if (avg_load
> max_load
&&
2739 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2740 max_load
= avg_load
;
2742 busiest_nr_running
= sum_nr_running
;
2743 busiest_load_per_task
= sum_weighted_load
;
2744 group_imb
= __group_imb
;
2747 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2749 * Busy processors will not participate in power savings
2752 if (idle
== CPU_NOT_IDLE
||
2753 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2757 * If the local group is idle or completely loaded
2758 * no need to do power savings balance at this domain
2760 if (local_group
&& (this_nr_running
>= group_capacity
||
2762 power_savings_balance
= 0;
2765 * If a group is already running at full capacity or idle,
2766 * don't include that group in power savings calculations
2768 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2773 * Calculate the group which has the least non-idle load.
2774 * This is the group from where we need to pick up the load
2777 if ((sum_nr_running
< min_nr_running
) ||
2778 (sum_nr_running
== min_nr_running
&&
2779 first_cpu(group
->cpumask
) <
2780 first_cpu(group_min
->cpumask
))) {
2782 min_nr_running
= sum_nr_running
;
2783 min_load_per_task
= sum_weighted_load
/
2788 * Calculate the group which is almost near its
2789 * capacity but still has some space to pick up some load
2790 * from other group and save more power
2792 if (sum_nr_running
<= group_capacity
- 1) {
2793 if (sum_nr_running
> leader_nr_running
||
2794 (sum_nr_running
== leader_nr_running
&&
2795 first_cpu(group
->cpumask
) >
2796 first_cpu(group_leader
->cpumask
))) {
2797 group_leader
= group
;
2798 leader_nr_running
= sum_nr_running
;
2803 group
= group
->next
;
2804 } while (group
!= sd
->groups
);
2806 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2809 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2811 if (this_load
>= avg_load
||
2812 100*max_load
<= sd
->imbalance_pct
*this_load
)
2815 busiest_load_per_task
/= busiest_nr_running
;
2817 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2820 * We're trying to get all the cpus to the average_load, so we don't
2821 * want to push ourselves above the average load, nor do we wish to
2822 * reduce the max loaded cpu below the average load, as either of these
2823 * actions would just result in more rebalancing later, and ping-pong
2824 * tasks around. Thus we look for the minimum possible imbalance.
2825 * Negative imbalances (*we* are more loaded than anyone else) will
2826 * be counted as no imbalance for these purposes -- we can't fix that
2827 * by pulling tasks to us. Be careful of negative numbers as they'll
2828 * appear as very large values with unsigned longs.
2830 if (max_load
<= busiest_load_per_task
)
2834 * In the presence of smp nice balancing, certain scenarios can have
2835 * max load less than avg load(as we skip the groups at or below
2836 * its cpu_power, while calculating max_load..)
2838 if (max_load
< avg_load
) {
2840 goto small_imbalance
;
2843 /* Don't want to pull so many tasks that a group would go idle */
2844 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2846 /* How much load to actually move to equalise the imbalance */
2847 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2848 (avg_load
- this_load
) * this->__cpu_power
)
2852 * if *imbalance is less than the average load per runnable task
2853 * there is no gaurantee that any tasks will be moved so we'll have
2854 * a think about bumping its value to force at least one task to be
2857 if (*imbalance
< busiest_load_per_task
) {
2858 unsigned long tmp
, pwr_now
, pwr_move
;
2862 pwr_move
= pwr_now
= 0;
2864 if (this_nr_running
) {
2865 this_load_per_task
/= this_nr_running
;
2866 if (busiest_load_per_task
> this_load_per_task
)
2869 this_load_per_task
= SCHED_LOAD_SCALE
;
2871 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2872 busiest_load_per_task
* imbn
) {
2873 *imbalance
= busiest_load_per_task
;
2878 * OK, we don't have enough imbalance to justify moving tasks,
2879 * however we may be able to increase total CPU power used by
2883 pwr_now
+= busiest
->__cpu_power
*
2884 min(busiest_load_per_task
, max_load
);
2885 pwr_now
+= this->__cpu_power
*
2886 min(this_load_per_task
, this_load
);
2887 pwr_now
/= SCHED_LOAD_SCALE
;
2889 /* Amount of load we'd subtract */
2890 tmp
= sg_div_cpu_power(busiest
,
2891 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2893 pwr_move
+= busiest
->__cpu_power
*
2894 min(busiest_load_per_task
, max_load
- tmp
);
2896 /* Amount of load we'd add */
2897 if (max_load
* busiest
->__cpu_power
<
2898 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2899 tmp
= sg_div_cpu_power(this,
2900 max_load
* busiest
->__cpu_power
);
2902 tmp
= sg_div_cpu_power(this,
2903 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2904 pwr_move
+= this->__cpu_power
*
2905 min(this_load_per_task
, this_load
+ tmp
);
2906 pwr_move
/= SCHED_LOAD_SCALE
;
2908 /* Move if we gain throughput */
2909 if (pwr_move
> pwr_now
)
2910 *imbalance
= busiest_load_per_task
;
2916 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2917 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2920 if (this == group_leader
&& group_leader
!= group_min
) {
2921 *imbalance
= min_load_per_task
;
2931 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2934 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2935 unsigned long imbalance
, cpumask_t
*cpus
)
2937 struct rq
*busiest
= NULL
, *rq
;
2938 unsigned long max_load
= 0;
2941 for_each_cpu_mask(i
, group
->cpumask
) {
2944 if (!cpu_isset(i
, *cpus
))
2948 wl
= weighted_cpuload(i
);
2950 if (rq
->nr_running
== 1 && wl
> imbalance
)
2953 if (wl
> max_load
) {
2963 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2964 * so long as it is large enough.
2966 #define MAX_PINNED_INTERVAL 512
2969 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2970 * tasks if there is an imbalance.
2972 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2973 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2976 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2977 struct sched_group
*group
;
2978 unsigned long imbalance
;
2980 cpumask_t cpus
= CPU_MASK_ALL
;
2981 unsigned long flags
;
2984 * When power savings policy is enabled for the parent domain, idle
2985 * sibling can pick up load irrespective of busy siblings. In this case,
2986 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2987 * portraying it as CPU_NOT_IDLE.
2989 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2990 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2993 schedstat_inc(sd
, lb_count
[idle
]);
2996 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3003 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3007 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
3009 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3013 BUG_ON(busiest
== this_rq
);
3015 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3018 if (busiest
->nr_running
> 1) {
3020 * Attempt to move tasks. If find_busiest_group has found
3021 * an imbalance but busiest->nr_running <= 1, the group is
3022 * still unbalanced. ld_moved simply stays zero, so it is
3023 * correctly treated as an imbalance.
3025 local_irq_save(flags
);
3026 double_rq_lock(this_rq
, busiest
);
3027 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3028 imbalance
, sd
, idle
, &all_pinned
);
3029 double_rq_unlock(this_rq
, busiest
);
3030 local_irq_restore(flags
);
3033 * some other cpu did the load balance for us.
3035 if (ld_moved
&& this_cpu
!= smp_processor_id())
3036 resched_cpu(this_cpu
);
3038 /* All tasks on this runqueue were pinned by CPU affinity */
3039 if (unlikely(all_pinned
)) {
3040 cpu_clear(cpu_of(busiest
), cpus
);
3041 if (!cpus_empty(cpus
))
3048 schedstat_inc(sd
, lb_failed
[idle
]);
3049 sd
->nr_balance_failed
++;
3051 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3053 spin_lock_irqsave(&busiest
->lock
, flags
);
3055 /* don't kick the migration_thread, if the curr
3056 * task on busiest cpu can't be moved to this_cpu
3058 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3059 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3061 goto out_one_pinned
;
3064 if (!busiest
->active_balance
) {
3065 busiest
->active_balance
= 1;
3066 busiest
->push_cpu
= this_cpu
;
3069 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3071 wake_up_process(busiest
->migration_thread
);
3074 * We've kicked active balancing, reset the failure
3077 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3080 sd
->nr_balance_failed
= 0;
3082 if (likely(!active_balance
)) {
3083 /* We were unbalanced, so reset the balancing interval */
3084 sd
->balance_interval
= sd
->min_interval
;
3087 * If we've begun active balancing, start to back off. This
3088 * case may not be covered by the all_pinned logic if there
3089 * is only 1 task on the busy runqueue (because we don't call
3092 if (sd
->balance_interval
< sd
->max_interval
)
3093 sd
->balance_interval
*= 2;
3096 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3097 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3102 schedstat_inc(sd
, lb_balanced
[idle
]);
3104 sd
->nr_balance_failed
= 0;
3107 /* tune up the balancing interval */
3108 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3109 (sd
->balance_interval
< sd
->max_interval
))
3110 sd
->balance_interval
*= 2;
3112 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3113 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3119 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3120 * tasks if there is an imbalance.
3122 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3123 * this_rq is locked.
3126 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
3128 struct sched_group
*group
;
3129 struct rq
*busiest
= NULL
;
3130 unsigned long imbalance
;
3134 cpumask_t cpus
= CPU_MASK_ALL
;
3137 * When power savings policy is enabled for the parent domain, idle
3138 * sibling can pick up load irrespective of busy siblings. In this case,
3139 * let the state of idle sibling percolate up as IDLE, instead of
3140 * portraying it as CPU_NOT_IDLE.
3142 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3143 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3146 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3148 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3149 &sd_idle
, &cpus
, NULL
);
3151 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3155 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
3158 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3162 BUG_ON(busiest
== this_rq
);
3164 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3167 if (busiest
->nr_running
> 1) {
3168 /* Attempt to move tasks */
3169 double_lock_balance(this_rq
, busiest
);
3170 /* this_rq->clock is already updated */
3171 update_rq_clock(busiest
);
3172 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3173 imbalance
, sd
, CPU_NEWLY_IDLE
,
3175 spin_unlock(&busiest
->lock
);
3177 if (unlikely(all_pinned
)) {
3178 cpu_clear(cpu_of(busiest
), cpus
);
3179 if (!cpus_empty(cpus
))
3185 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3186 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3187 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3190 sd
->nr_balance_failed
= 0;
3195 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3196 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3197 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3199 sd
->nr_balance_failed
= 0;
3205 * idle_balance is called by schedule() if this_cpu is about to become
3206 * idle. Attempts to pull tasks from other CPUs.
3208 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3210 struct sched_domain
*sd
;
3211 int pulled_task
= -1;
3212 unsigned long next_balance
= jiffies
+ HZ
;
3214 for_each_domain(this_cpu
, sd
) {
3215 unsigned long interval
;
3217 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3220 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3221 /* If we've pulled tasks over stop searching: */
3222 pulled_task
= load_balance_newidle(this_cpu
,
3225 interval
= msecs_to_jiffies(sd
->balance_interval
);
3226 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3227 next_balance
= sd
->last_balance
+ interval
;
3231 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3233 * We are going idle. next_balance may be set based on
3234 * a busy processor. So reset next_balance.
3236 this_rq
->next_balance
= next_balance
;
3241 * active_load_balance is run by migration threads. It pushes running tasks
3242 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3243 * running on each physical CPU where possible, and avoids physical /
3244 * logical imbalances.
3246 * Called with busiest_rq locked.
3248 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3250 int target_cpu
= busiest_rq
->push_cpu
;
3251 struct sched_domain
*sd
;
3252 struct rq
*target_rq
;
3254 /* Is there any task to move? */
3255 if (busiest_rq
->nr_running
<= 1)
3258 target_rq
= cpu_rq(target_cpu
);
3261 * This condition is "impossible", if it occurs
3262 * we need to fix it. Originally reported by
3263 * Bjorn Helgaas on a 128-cpu setup.
3265 BUG_ON(busiest_rq
== target_rq
);
3267 /* move a task from busiest_rq to target_rq */
3268 double_lock_balance(busiest_rq
, target_rq
);
3269 update_rq_clock(busiest_rq
);
3270 update_rq_clock(target_rq
);
3272 /* Search for an sd spanning us and the target CPU. */
3273 for_each_domain(target_cpu
, sd
) {
3274 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3275 cpu_isset(busiest_cpu
, sd
->span
))
3280 schedstat_inc(sd
, alb_count
);
3282 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3284 schedstat_inc(sd
, alb_pushed
);
3286 schedstat_inc(sd
, alb_failed
);
3288 spin_unlock(&target_rq
->lock
);
3293 atomic_t load_balancer
;
3295 } nohz ____cacheline_aligned
= {
3296 .load_balancer
= ATOMIC_INIT(-1),
3297 .cpu_mask
= CPU_MASK_NONE
,
3301 * This routine will try to nominate the ilb (idle load balancing)
3302 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3303 * load balancing on behalf of all those cpus. If all the cpus in the system
3304 * go into this tickless mode, then there will be no ilb owner (as there is
3305 * no need for one) and all the cpus will sleep till the next wakeup event
3308 * For the ilb owner, tick is not stopped. And this tick will be used
3309 * for idle load balancing. ilb owner will still be part of
3312 * While stopping the tick, this cpu will become the ilb owner if there
3313 * is no other owner. And will be the owner till that cpu becomes busy
3314 * or if all cpus in the system stop their ticks at which point
3315 * there is no need for ilb owner.
3317 * When the ilb owner becomes busy, it nominates another owner, during the
3318 * next busy scheduler_tick()
3320 int select_nohz_load_balancer(int stop_tick
)
3322 int cpu
= smp_processor_id();
3325 cpu_set(cpu
, nohz
.cpu_mask
);
3326 cpu_rq(cpu
)->in_nohz_recently
= 1;
3329 * If we are going offline and still the leader, give up!
3331 if (cpu_is_offline(cpu
) &&
3332 atomic_read(&nohz
.load_balancer
) == cpu
) {
3333 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3338 /* time for ilb owner also to sleep */
3339 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3340 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3341 atomic_set(&nohz
.load_balancer
, -1);
3345 if (atomic_read(&nohz
.load_balancer
) == -1) {
3346 /* make me the ilb owner */
3347 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3349 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3352 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3355 cpu_clear(cpu
, nohz
.cpu_mask
);
3357 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3358 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3365 static DEFINE_SPINLOCK(balancing
);
3368 * It checks each scheduling domain to see if it is due to be balanced,
3369 * and initiates a balancing operation if so.
3371 * Balancing parameters are set up in arch_init_sched_domains.
3373 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3376 struct rq
*rq
= cpu_rq(cpu
);
3377 unsigned long interval
;
3378 struct sched_domain
*sd
;
3379 /* Earliest time when we have to do rebalance again */
3380 unsigned long next_balance
= jiffies
+ 60*HZ
;
3381 int update_next_balance
= 0;
3383 for_each_domain(cpu
, sd
) {
3384 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3387 interval
= sd
->balance_interval
;
3388 if (idle
!= CPU_IDLE
)
3389 interval
*= sd
->busy_factor
;
3391 /* scale ms to jiffies */
3392 interval
= msecs_to_jiffies(interval
);
3393 if (unlikely(!interval
))
3395 if (interval
> HZ
*NR_CPUS
/10)
3396 interval
= HZ
*NR_CPUS
/10;
3399 if (sd
->flags
& SD_SERIALIZE
) {
3400 if (!spin_trylock(&balancing
))
3404 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3405 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3407 * We've pulled tasks over so either we're no
3408 * longer idle, or one of our SMT siblings is
3411 idle
= CPU_NOT_IDLE
;
3413 sd
->last_balance
= jiffies
;
3415 if (sd
->flags
& SD_SERIALIZE
)
3416 spin_unlock(&balancing
);
3418 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3419 next_balance
= sd
->last_balance
+ interval
;
3420 update_next_balance
= 1;
3424 * Stop the load balance at this level. There is another
3425 * CPU in our sched group which is doing load balancing more
3433 * next_balance will be updated only when there is a need.
3434 * When the cpu is attached to null domain for ex, it will not be
3437 if (likely(update_next_balance
))
3438 rq
->next_balance
= next_balance
;
3442 * run_rebalance_domains is triggered when needed from the scheduler tick.
3443 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3444 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3446 static void run_rebalance_domains(struct softirq_action
*h
)
3448 int this_cpu
= smp_processor_id();
3449 struct rq
*this_rq
= cpu_rq(this_cpu
);
3450 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3451 CPU_IDLE
: CPU_NOT_IDLE
;
3453 rebalance_domains(this_cpu
, idle
);
3457 * If this cpu is the owner for idle load balancing, then do the
3458 * balancing on behalf of the other idle cpus whose ticks are
3461 if (this_rq
->idle_at_tick
&&
3462 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3463 cpumask_t cpus
= nohz
.cpu_mask
;
3467 cpu_clear(this_cpu
, cpus
);
3468 for_each_cpu_mask(balance_cpu
, cpus
) {
3470 * If this cpu gets work to do, stop the load balancing
3471 * work being done for other cpus. Next load
3472 * balancing owner will pick it up.
3477 rebalance_domains(balance_cpu
, CPU_IDLE
);
3479 rq
= cpu_rq(balance_cpu
);
3480 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3481 this_rq
->next_balance
= rq
->next_balance
;
3488 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3490 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3491 * idle load balancing owner or decide to stop the periodic load balancing,
3492 * if the whole system is idle.
3494 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3498 * If we were in the nohz mode recently and busy at the current
3499 * scheduler tick, then check if we need to nominate new idle
3502 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3503 rq
->in_nohz_recently
= 0;
3505 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3506 cpu_clear(cpu
, nohz
.cpu_mask
);
3507 atomic_set(&nohz
.load_balancer
, -1);
3510 if (atomic_read(&nohz
.load_balancer
) == -1) {
3512 * simple selection for now: Nominate the
3513 * first cpu in the nohz list to be the next
3516 * TBD: Traverse the sched domains and nominate
3517 * the nearest cpu in the nohz.cpu_mask.
3519 int ilb
= first_cpu(nohz
.cpu_mask
);
3527 * If this cpu is idle and doing idle load balancing for all the
3528 * cpus with ticks stopped, is it time for that to stop?
3530 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3531 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3537 * If this cpu is idle and the idle load balancing is done by
3538 * someone else, then no need raise the SCHED_SOFTIRQ
3540 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3541 cpu_isset(cpu
, nohz
.cpu_mask
))
3544 if (time_after_eq(jiffies
, rq
->next_balance
))
3545 raise_softirq(SCHED_SOFTIRQ
);
3548 #else /* CONFIG_SMP */
3551 * on UP we do not need to balance between CPUs:
3553 static inline void idle_balance(int cpu
, struct rq
*rq
)
3559 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3561 EXPORT_PER_CPU_SYMBOL(kstat
);
3564 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3565 * that have not yet been banked in case the task is currently running.
3567 unsigned long long task_sched_runtime(struct task_struct
*p
)
3569 unsigned long flags
;
3573 rq
= task_rq_lock(p
, &flags
);
3574 ns
= p
->se
.sum_exec_runtime
;
3575 if (task_current(rq
, p
)) {
3576 update_rq_clock(rq
);
3577 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3578 if ((s64
)delta_exec
> 0)
3581 task_rq_unlock(rq
, &flags
);
3587 * Account user cpu time to a process.
3588 * @p: the process that the cpu time gets accounted to
3589 * @cputime: the cpu time spent in user space since the last update
3591 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3593 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3596 p
->utime
= cputime_add(p
->utime
, cputime
);
3598 /* Add user time to cpustat. */
3599 tmp
= cputime_to_cputime64(cputime
);
3600 if (TASK_NICE(p
) > 0)
3601 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3603 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3607 * Account guest cpu time to a process.
3608 * @p: the process that the cpu time gets accounted to
3609 * @cputime: the cpu time spent in virtual machine since the last update
3611 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3614 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3616 tmp
= cputime_to_cputime64(cputime
);
3618 p
->utime
= cputime_add(p
->utime
, cputime
);
3619 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3621 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3622 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3626 * Account scaled user cpu time to a process.
3627 * @p: the process that the cpu time gets accounted to
3628 * @cputime: the cpu time spent in user space since the last update
3630 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3632 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3636 * Account system cpu time to a process.
3637 * @p: the process that the cpu time gets accounted to
3638 * @hardirq_offset: the offset to subtract from hardirq_count()
3639 * @cputime: the cpu time spent in kernel space since the last update
3641 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3644 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3645 struct rq
*rq
= this_rq();
3648 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3649 return account_guest_time(p
, cputime
);
3651 p
->stime
= cputime_add(p
->stime
, cputime
);
3653 /* Add system time to cpustat. */
3654 tmp
= cputime_to_cputime64(cputime
);
3655 if (hardirq_count() - hardirq_offset
)
3656 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3657 else if (softirq_count())
3658 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3659 else if (p
!= rq
->idle
)
3660 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3661 else if (atomic_read(&rq
->nr_iowait
) > 0)
3662 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3664 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3665 /* Account for system time used */
3666 acct_update_integrals(p
);
3670 * Account scaled system cpu time to a process.
3671 * @p: the process that the cpu time gets accounted to
3672 * @hardirq_offset: the offset to subtract from hardirq_count()
3673 * @cputime: the cpu time spent in kernel space since the last update
3675 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3677 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3681 * Account for involuntary wait time.
3682 * @p: the process from which the cpu time has been stolen
3683 * @steal: the cpu time spent in involuntary wait
3685 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3687 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3688 cputime64_t tmp
= cputime_to_cputime64(steal
);
3689 struct rq
*rq
= this_rq();
3691 if (p
== rq
->idle
) {
3692 p
->stime
= cputime_add(p
->stime
, steal
);
3693 if (atomic_read(&rq
->nr_iowait
) > 0)
3694 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3696 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3698 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3702 * This function gets called by the timer code, with HZ frequency.
3703 * We call it with interrupts disabled.
3705 * It also gets called by the fork code, when changing the parent's
3708 void scheduler_tick(void)
3710 int cpu
= smp_processor_id();
3711 struct rq
*rq
= cpu_rq(cpu
);
3712 struct task_struct
*curr
= rq
->curr
;
3713 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3715 spin_lock(&rq
->lock
);
3716 __update_rq_clock(rq
);
3718 * Let rq->clock advance by at least TICK_NSEC:
3720 if (unlikely(rq
->clock
< next_tick
)) {
3721 rq
->clock
= next_tick
;
3722 rq
->clock_underflows
++;
3724 rq
->tick_timestamp
= rq
->clock
;
3725 update_cpu_load(rq
);
3726 curr
->sched_class
->task_tick(rq
, curr
, 0);
3727 update_sched_rt_period(rq
);
3728 spin_unlock(&rq
->lock
);
3731 rq
->idle_at_tick
= idle_cpu(cpu
);
3732 trigger_load_balance(rq
, cpu
);
3736 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3738 void __kprobes
add_preempt_count(int val
)
3743 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3745 preempt_count() += val
;
3747 * Spinlock count overflowing soon?
3749 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3752 EXPORT_SYMBOL(add_preempt_count
);
3754 void __kprobes
sub_preempt_count(int val
)
3759 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3762 * Is the spinlock portion underflowing?
3764 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3765 !(preempt_count() & PREEMPT_MASK
)))
3768 preempt_count() -= val
;
3770 EXPORT_SYMBOL(sub_preempt_count
);
3775 * Print scheduling while atomic bug:
3777 static noinline
void __schedule_bug(struct task_struct
*prev
)
3779 struct pt_regs
*regs
= get_irq_regs();
3781 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3782 prev
->comm
, prev
->pid
, preempt_count());
3784 debug_show_held_locks(prev
);
3785 if (irqs_disabled())
3786 print_irqtrace_events(prev
);
3795 * Various schedule()-time debugging checks and statistics:
3797 static inline void schedule_debug(struct task_struct
*prev
)
3800 * Test if we are atomic. Since do_exit() needs to call into
3801 * schedule() atomically, we ignore that path for now.
3802 * Otherwise, whine if we are scheduling when we should not be.
3804 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3805 __schedule_bug(prev
);
3807 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3809 schedstat_inc(this_rq(), sched_count
);
3810 #ifdef CONFIG_SCHEDSTATS
3811 if (unlikely(prev
->lock_depth
>= 0)) {
3812 schedstat_inc(this_rq(), bkl_count
);
3813 schedstat_inc(prev
, sched_info
.bkl_count
);
3819 * Pick up the highest-prio task:
3821 static inline struct task_struct
*
3822 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3824 const struct sched_class
*class;
3825 struct task_struct
*p
;
3828 * Optimization: we know that if all tasks are in
3829 * the fair class we can call that function directly:
3831 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3832 p
= fair_sched_class
.pick_next_task(rq
);
3837 class = sched_class_highest
;
3839 p
= class->pick_next_task(rq
);
3843 * Will never be NULL as the idle class always
3844 * returns a non-NULL p:
3846 class = class->next
;
3851 * schedule() is the main scheduler function.
3853 asmlinkage
void __sched
schedule(void)
3855 struct task_struct
*prev
, *next
;
3856 unsigned long *switch_count
;
3862 cpu
= smp_processor_id();
3866 switch_count
= &prev
->nivcsw
;
3868 release_kernel_lock(prev
);
3869 need_resched_nonpreemptible
:
3871 schedule_debug(prev
);
3876 * Do the rq-clock update outside the rq lock:
3878 local_irq_disable();
3879 __update_rq_clock(rq
);
3880 spin_lock(&rq
->lock
);
3881 clear_tsk_need_resched(prev
);
3883 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3884 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3885 signal_pending(prev
))) {
3886 prev
->state
= TASK_RUNNING
;
3888 deactivate_task(rq
, prev
, 1);
3890 switch_count
= &prev
->nvcsw
;
3894 if (prev
->sched_class
->pre_schedule
)
3895 prev
->sched_class
->pre_schedule(rq
, prev
);
3898 if (unlikely(!rq
->nr_running
))
3899 idle_balance(cpu
, rq
);
3901 prev
->sched_class
->put_prev_task(rq
, prev
);
3902 next
= pick_next_task(rq
, prev
);
3904 sched_info_switch(prev
, next
);
3906 if (likely(prev
!= next
)) {
3911 context_switch(rq
, prev
, next
); /* unlocks the rq */
3913 * the context switch might have flipped the stack from under
3914 * us, hence refresh the local variables.
3916 cpu
= smp_processor_id();
3919 spin_unlock_irq(&rq
->lock
);
3923 if (unlikely(reacquire_kernel_lock(current
) < 0))
3924 goto need_resched_nonpreemptible
;
3926 preempt_enable_no_resched();
3927 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3930 EXPORT_SYMBOL(schedule
);
3932 #ifdef CONFIG_PREEMPT
3934 * this is the entry point to schedule() from in-kernel preemption
3935 * off of preempt_enable. Kernel preemptions off return from interrupt
3936 * occur there and call schedule directly.
3938 asmlinkage
void __sched
preempt_schedule(void)
3940 struct thread_info
*ti
= current_thread_info();
3941 struct task_struct
*task
= current
;
3942 int saved_lock_depth
;
3945 * If there is a non-zero preempt_count or interrupts are disabled,
3946 * we do not want to preempt the current task. Just return..
3948 if (likely(ti
->preempt_count
|| irqs_disabled()))
3952 add_preempt_count(PREEMPT_ACTIVE
);
3955 * We keep the big kernel semaphore locked, but we
3956 * clear ->lock_depth so that schedule() doesnt
3957 * auto-release the semaphore:
3959 saved_lock_depth
= task
->lock_depth
;
3960 task
->lock_depth
= -1;
3962 task
->lock_depth
= saved_lock_depth
;
3963 sub_preempt_count(PREEMPT_ACTIVE
);
3966 * Check again in case we missed a preemption opportunity
3967 * between schedule and now.
3970 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3972 EXPORT_SYMBOL(preempt_schedule
);
3975 * this is the entry point to schedule() from kernel preemption
3976 * off of irq context.
3977 * Note, that this is called and return with irqs disabled. This will
3978 * protect us against recursive calling from irq.
3980 asmlinkage
void __sched
preempt_schedule_irq(void)
3982 struct thread_info
*ti
= current_thread_info();
3983 struct task_struct
*task
= current
;
3984 int saved_lock_depth
;
3986 /* Catch callers which need to be fixed */
3987 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3990 add_preempt_count(PREEMPT_ACTIVE
);
3993 * We keep the big kernel semaphore locked, but we
3994 * clear ->lock_depth so that schedule() doesnt
3995 * auto-release the semaphore:
3997 saved_lock_depth
= task
->lock_depth
;
3998 task
->lock_depth
= -1;
4001 local_irq_disable();
4002 task
->lock_depth
= saved_lock_depth
;
4003 sub_preempt_count(PREEMPT_ACTIVE
);
4006 * Check again in case we missed a preemption opportunity
4007 * between schedule and now.
4010 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4013 #endif /* CONFIG_PREEMPT */
4015 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4018 return try_to_wake_up(curr
->private, mode
, sync
);
4020 EXPORT_SYMBOL(default_wake_function
);
4023 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4024 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4025 * number) then we wake all the non-exclusive tasks and one exclusive task.
4027 * There are circumstances in which we can try to wake a task which has already
4028 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4029 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4031 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4032 int nr_exclusive
, int sync
, void *key
)
4034 wait_queue_t
*curr
, *next
;
4036 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4037 unsigned flags
= curr
->flags
;
4039 if (curr
->func(curr
, mode
, sync
, key
) &&
4040 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4046 * __wake_up - wake up threads blocked on a waitqueue.
4048 * @mode: which threads
4049 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4050 * @key: is directly passed to the wakeup function
4052 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4053 int nr_exclusive
, void *key
)
4055 unsigned long flags
;
4057 spin_lock_irqsave(&q
->lock
, flags
);
4058 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4059 spin_unlock_irqrestore(&q
->lock
, flags
);
4061 EXPORT_SYMBOL(__wake_up
);
4064 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4066 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4068 __wake_up_common(q
, mode
, 1, 0, NULL
);
4072 * __wake_up_sync - wake up threads blocked on a waitqueue.
4074 * @mode: which threads
4075 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4077 * The sync wakeup differs that the waker knows that it will schedule
4078 * away soon, so while the target thread will be woken up, it will not
4079 * be migrated to another CPU - ie. the two threads are 'synchronized'
4080 * with each other. This can prevent needless bouncing between CPUs.
4082 * On UP it can prevent extra preemption.
4085 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4087 unsigned long flags
;
4093 if (unlikely(!nr_exclusive
))
4096 spin_lock_irqsave(&q
->lock
, flags
);
4097 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4098 spin_unlock_irqrestore(&q
->lock
, flags
);
4100 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4102 void complete(struct completion
*x
)
4104 unsigned long flags
;
4106 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4108 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4109 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4111 EXPORT_SYMBOL(complete
);
4113 void complete_all(struct completion
*x
)
4115 unsigned long flags
;
4117 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4118 x
->done
+= UINT_MAX
/2;
4119 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4120 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4122 EXPORT_SYMBOL(complete_all
);
4124 static inline long __sched
4125 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4128 DECLARE_WAITQUEUE(wait
, current
);
4130 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4131 __add_wait_queue_tail(&x
->wait
, &wait
);
4133 if ((state
== TASK_INTERRUPTIBLE
&&
4134 signal_pending(current
)) ||
4135 (state
== TASK_KILLABLE
&&
4136 fatal_signal_pending(current
))) {
4137 __remove_wait_queue(&x
->wait
, &wait
);
4138 return -ERESTARTSYS
;
4140 __set_current_state(state
);
4141 spin_unlock_irq(&x
->wait
.lock
);
4142 timeout
= schedule_timeout(timeout
);
4143 spin_lock_irq(&x
->wait
.lock
);
4145 __remove_wait_queue(&x
->wait
, &wait
);
4149 __remove_wait_queue(&x
->wait
, &wait
);
4156 wait_for_common(struct completion
*x
, long timeout
, int state
)
4160 spin_lock_irq(&x
->wait
.lock
);
4161 timeout
= do_wait_for_common(x
, timeout
, state
);
4162 spin_unlock_irq(&x
->wait
.lock
);
4166 void __sched
wait_for_completion(struct completion
*x
)
4168 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4170 EXPORT_SYMBOL(wait_for_completion
);
4172 unsigned long __sched
4173 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4175 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4177 EXPORT_SYMBOL(wait_for_completion_timeout
);
4179 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4181 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4182 if (t
== -ERESTARTSYS
)
4186 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4188 unsigned long __sched
4189 wait_for_completion_interruptible_timeout(struct completion
*x
,
4190 unsigned long timeout
)
4192 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4194 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4196 int __sched
wait_for_completion_killable(struct completion
*x
)
4198 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4199 if (t
== -ERESTARTSYS
)
4203 EXPORT_SYMBOL(wait_for_completion_killable
);
4206 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4208 unsigned long flags
;
4211 init_waitqueue_entry(&wait
, current
);
4213 __set_current_state(state
);
4215 spin_lock_irqsave(&q
->lock
, flags
);
4216 __add_wait_queue(q
, &wait
);
4217 spin_unlock(&q
->lock
);
4218 timeout
= schedule_timeout(timeout
);
4219 spin_lock_irq(&q
->lock
);
4220 __remove_wait_queue(q
, &wait
);
4221 spin_unlock_irqrestore(&q
->lock
, flags
);
4226 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4228 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4230 EXPORT_SYMBOL(interruptible_sleep_on
);
4233 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4235 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4237 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4239 void __sched
sleep_on(wait_queue_head_t
*q
)
4241 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4243 EXPORT_SYMBOL(sleep_on
);
4245 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4247 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4249 EXPORT_SYMBOL(sleep_on_timeout
);
4251 #ifdef CONFIG_RT_MUTEXES
4254 * rt_mutex_setprio - set the current priority of a task
4256 * @prio: prio value (kernel-internal form)
4258 * This function changes the 'effective' priority of a task. It does
4259 * not touch ->normal_prio like __setscheduler().
4261 * Used by the rt_mutex code to implement priority inheritance logic.
4263 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4265 unsigned long flags
;
4266 int oldprio
, on_rq
, running
;
4268 const struct sched_class
*prev_class
= p
->sched_class
;
4270 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4272 rq
= task_rq_lock(p
, &flags
);
4273 update_rq_clock(rq
);
4276 on_rq
= p
->se
.on_rq
;
4277 running
= task_current(rq
, p
);
4279 dequeue_task(rq
, p
, 0);
4281 p
->sched_class
->put_prev_task(rq
, p
);
4284 p
->sched_class
= &rt_sched_class
;
4286 p
->sched_class
= &fair_sched_class
;
4291 p
->sched_class
->set_curr_task(rq
);
4293 enqueue_task(rq
, p
, 0);
4295 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4297 task_rq_unlock(rq
, &flags
);
4302 void set_user_nice(struct task_struct
*p
, long nice
)
4304 int old_prio
, delta
, on_rq
;
4305 unsigned long flags
;
4308 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4311 * We have to be careful, if called from sys_setpriority(),
4312 * the task might be in the middle of scheduling on another CPU.
4314 rq
= task_rq_lock(p
, &flags
);
4315 update_rq_clock(rq
);
4317 * The RT priorities are set via sched_setscheduler(), but we still
4318 * allow the 'normal' nice value to be set - but as expected
4319 * it wont have any effect on scheduling until the task is
4320 * SCHED_FIFO/SCHED_RR:
4322 if (task_has_rt_policy(p
)) {
4323 p
->static_prio
= NICE_TO_PRIO(nice
);
4326 on_rq
= p
->se
.on_rq
;
4328 dequeue_task(rq
, p
, 0);
4332 p
->static_prio
= NICE_TO_PRIO(nice
);
4335 p
->prio
= effective_prio(p
);
4336 delta
= p
->prio
- old_prio
;
4339 enqueue_task(rq
, p
, 0);
4342 * If the task increased its priority or is running and
4343 * lowered its priority, then reschedule its CPU:
4345 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4346 resched_task(rq
->curr
);
4349 task_rq_unlock(rq
, &flags
);
4351 EXPORT_SYMBOL(set_user_nice
);
4354 * can_nice - check if a task can reduce its nice value
4358 int can_nice(const struct task_struct
*p
, const int nice
)
4360 /* convert nice value [19,-20] to rlimit style value [1,40] */
4361 int nice_rlim
= 20 - nice
;
4363 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4364 capable(CAP_SYS_NICE
));
4367 #ifdef __ARCH_WANT_SYS_NICE
4370 * sys_nice - change the priority of the current process.
4371 * @increment: priority increment
4373 * sys_setpriority is a more generic, but much slower function that
4374 * does similar things.
4376 asmlinkage
long sys_nice(int increment
)
4381 * Setpriority might change our priority at the same moment.
4382 * We don't have to worry. Conceptually one call occurs first
4383 * and we have a single winner.
4385 if (increment
< -40)
4390 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4396 if (increment
< 0 && !can_nice(current
, nice
))
4399 retval
= security_task_setnice(current
, nice
);
4403 set_user_nice(current
, nice
);
4410 * task_prio - return the priority value of a given task.
4411 * @p: the task in question.
4413 * This is the priority value as seen by users in /proc.
4414 * RT tasks are offset by -200. Normal tasks are centered
4415 * around 0, value goes from -16 to +15.
4417 int task_prio(const struct task_struct
*p
)
4419 return p
->prio
- MAX_RT_PRIO
;
4423 * task_nice - return the nice value of a given task.
4424 * @p: the task in question.
4426 int task_nice(const struct task_struct
*p
)
4428 return TASK_NICE(p
);
4430 EXPORT_SYMBOL(task_nice
);
4433 * idle_cpu - is a given cpu idle currently?
4434 * @cpu: the processor in question.
4436 int idle_cpu(int cpu
)
4438 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4442 * idle_task - return the idle task for a given cpu.
4443 * @cpu: the processor in question.
4445 struct task_struct
*idle_task(int cpu
)
4447 return cpu_rq(cpu
)->idle
;
4451 * find_process_by_pid - find a process with a matching PID value.
4452 * @pid: the pid in question.
4454 static struct task_struct
*find_process_by_pid(pid_t pid
)
4456 return pid
? find_task_by_vpid(pid
) : current
;
4459 /* Actually do priority change: must hold rq lock. */
4461 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4463 BUG_ON(p
->se
.on_rq
);
4466 switch (p
->policy
) {
4470 p
->sched_class
= &fair_sched_class
;
4474 p
->sched_class
= &rt_sched_class
;
4478 p
->rt_priority
= prio
;
4479 p
->normal_prio
= normal_prio(p
);
4480 /* we are holding p->pi_lock already */
4481 p
->prio
= rt_mutex_getprio(p
);
4486 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4487 * @p: the task in question.
4488 * @policy: new policy.
4489 * @param: structure containing the new RT priority.
4491 * NOTE that the task may be already dead.
4493 int sched_setscheduler(struct task_struct
*p
, int policy
,
4494 struct sched_param
*param
)
4496 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4497 unsigned long flags
;
4498 const struct sched_class
*prev_class
= p
->sched_class
;
4501 /* may grab non-irq protected spin_locks */
4502 BUG_ON(in_interrupt());
4504 /* double check policy once rq lock held */
4506 policy
= oldpolicy
= p
->policy
;
4507 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4508 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4509 policy
!= SCHED_IDLE
)
4512 * Valid priorities for SCHED_FIFO and SCHED_RR are
4513 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4514 * SCHED_BATCH and SCHED_IDLE is 0.
4516 if (param
->sched_priority
< 0 ||
4517 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4518 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4520 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4524 * Allow unprivileged RT tasks to decrease priority:
4526 if (!capable(CAP_SYS_NICE
)) {
4527 if (rt_policy(policy
)) {
4528 unsigned long rlim_rtprio
;
4530 if (!lock_task_sighand(p
, &flags
))
4532 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4533 unlock_task_sighand(p
, &flags
);
4535 /* can't set/change the rt policy */
4536 if (policy
!= p
->policy
&& !rlim_rtprio
)
4539 /* can't increase priority */
4540 if (param
->sched_priority
> p
->rt_priority
&&
4541 param
->sched_priority
> rlim_rtprio
)
4545 * Like positive nice levels, dont allow tasks to
4546 * move out of SCHED_IDLE either:
4548 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4551 /* can't change other user's priorities */
4552 if ((current
->euid
!= p
->euid
) &&
4553 (current
->euid
!= p
->uid
))
4557 #ifdef CONFIG_RT_GROUP_SCHED
4559 * Do not allow realtime tasks into groups that have no runtime
4562 if (rt_policy(policy
) && task_group(p
)->rt_runtime
== 0)
4566 retval
= security_task_setscheduler(p
, policy
, param
);
4570 * make sure no PI-waiters arrive (or leave) while we are
4571 * changing the priority of the task:
4573 spin_lock_irqsave(&p
->pi_lock
, flags
);
4575 * To be able to change p->policy safely, the apropriate
4576 * runqueue lock must be held.
4578 rq
= __task_rq_lock(p
);
4579 /* recheck policy now with rq lock held */
4580 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4581 policy
= oldpolicy
= -1;
4582 __task_rq_unlock(rq
);
4583 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4586 update_rq_clock(rq
);
4587 on_rq
= p
->se
.on_rq
;
4588 running
= task_current(rq
, p
);
4590 deactivate_task(rq
, p
, 0);
4592 p
->sched_class
->put_prev_task(rq
, p
);
4595 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4598 p
->sched_class
->set_curr_task(rq
);
4600 activate_task(rq
, p
, 0);
4602 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4604 __task_rq_unlock(rq
);
4605 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4607 rt_mutex_adjust_pi(p
);
4611 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4614 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4616 struct sched_param lparam
;
4617 struct task_struct
*p
;
4620 if (!param
|| pid
< 0)
4622 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4627 p
= find_process_by_pid(pid
);
4629 retval
= sched_setscheduler(p
, policy
, &lparam
);
4636 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4637 * @pid: the pid in question.
4638 * @policy: new policy.
4639 * @param: structure containing the new RT priority.
4642 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4644 /* negative values for policy are not valid */
4648 return do_sched_setscheduler(pid
, policy
, param
);
4652 * sys_sched_setparam - set/change the RT priority of a thread
4653 * @pid: the pid in question.
4654 * @param: structure containing the new RT priority.
4656 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4658 return do_sched_setscheduler(pid
, -1, param
);
4662 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4663 * @pid: the pid in question.
4665 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4667 struct task_struct
*p
;
4674 read_lock(&tasklist_lock
);
4675 p
= find_process_by_pid(pid
);
4677 retval
= security_task_getscheduler(p
);
4681 read_unlock(&tasklist_lock
);
4686 * sys_sched_getscheduler - get the RT priority of a thread
4687 * @pid: the pid in question.
4688 * @param: structure containing the RT priority.
4690 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4692 struct sched_param lp
;
4693 struct task_struct
*p
;
4696 if (!param
|| pid
< 0)
4699 read_lock(&tasklist_lock
);
4700 p
= find_process_by_pid(pid
);
4705 retval
= security_task_getscheduler(p
);
4709 lp
.sched_priority
= p
->rt_priority
;
4710 read_unlock(&tasklist_lock
);
4713 * This one might sleep, we cannot do it with a spinlock held ...
4715 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4720 read_unlock(&tasklist_lock
);
4724 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4726 cpumask_t cpus_allowed
;
4727 struct task_struct
*p
;
4731 read_lock(&tasklist_lock
);
4733 p
= find_process_by_pid(pid
);
4735 read_unlock(&tasklist_lock
);
4741 * It is not safe to call set_cpus_allowed with the
4742 * tasklist_lock held. We will bump the task_struct's
4743 * usage count and then drop tasklist_lock.
4746 read_unlock(&tasklist_lock
);
4749 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4750 !capable(CAP_SYS_NICE
))
4753 retval
= security_task_setscheduler(p
, 0, NULL
);
4757 cpus_allowed
= cpuset_cpus_allowed(p
);
4758 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4760 retval
= set_cpus_allowed(p
, new_mask
);
4763 cpus_allowed
= cpuset_cpus_allowed(p
);
4764 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4766 * We must have raced with a concurrent cpuset
4767 * update. Just reset the cpus_allowed to the
4768 * cpuset's cpus_allowed
4770 new_mask
= cpus_allowed
;
4780 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4781 cpumask_t
*new_mask
)
4783 if (len
< sizeof(cpumask_t
)) {
4784 memset(new_mask
, 0, sizeof(cpumask_t
));
4785 } else if (len
> sizeof(cpumask_t
)) {
4786 len
= sizeof(cpumask_t
);
4788 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4792 * sys_sched_setaffinity - set the cpu affinity of a process
4793 * @pid: pid of the process
4794 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4795 * @user_mask_ptr: user-space pointer to the new cpu mask
4797 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4798 unsigned long __user
*user_mask_ptr
)
4803 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4807 return sched_setaffinity(pid
, new_mask
);
4811 * Represents all cpu's present in the system
4812 * In systems capable of hotplug, this map could dynamically grow
4813 * as new cpu's are detected in the system via any platform specific
4814 * method, such as ACPI for e.g.
4817 cpumask_t cpu_present_map __read_mostly
;
4818 EXPORT_SYMBOL(cpu_present_map
);
4821 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4822 EXPORT_SYMBOL(cpu_online_map
);
4824 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4825 EXPORT_SYMBOL(cpu_possible_map
);
4828 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4830 struct task_struct
*p
;
4834 read_lock(&tasklist_lock
);
4837 p
= find_process_by_pid(pid
);
4841 retval
= security_task_getscheduler(p
);
4845 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4848 read_unlock(&tasklist_lock
);
4855 * sys_sched_getaffinity - get the cpu affinity of a process
4856 * @pid: pid of the process
4857 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4858 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4860 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4861 unsigned long __user
*user_mask_ptr
)
4866 if (len
< sizeof(cpumask_t
))
4869 ret
= sched_getaffinity(pid
, &mask
);
4873 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4876 return sizeof(cpumask_t
);
4880 * sys_sched_yield - yield the current processor to other threads.
4882 * This function yields the current CPU to other tasks. If there are no
4883 * other threads running on this CPU then this function will return.
4885 asmlinkage
long sys_sched_yield(void)
4887 struct rq
*rq
= this_rq_lock();
4889 schedstat_inc(rq
, yld_count
);
4890 current
->sched_class
->yield_task(rq
);
4893 * Since we are going to call schedule() anyway, there's
4894 * no need to preempt or enable interrupts:
4896 __release(rq
->lock
);
4897 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4898 _raw_spin_unlock(&rq
->lock
);
4899 preempt_enable_no_resched();
4906 static void __cond_resched(void)
4908 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4909 __might_sleep(__FILE__
, __LINE__
);
4912 * The BKS might be reacquired before we have dropped
4913 * PREEMPT_ACTIVE, which could trigger a second
4914 * cond_resched() call.
4917 add_preempt_count(PREEMPT_ACTIVE
);
4919 sub_preempt_count(PREEMPT_ACTIVE
);
4920 } while (need_resched());
4923 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4924 int __sched
_cond_resched(void)
4926 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4927 system_state
== SYSTEM_RUNNING
) {
4933 EXPORT_SYMBOL(_cond_resched
);
4937 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4938 * call schedule, and on return reacquire the lock.
4940 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4941 * operations here to prevent schedule() from being called twice (once via
4942 * spin_unlock(), once by hand).
4944 int cond_resched_lock(spinlock_t
*lock
)
4946 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
4949 if (spin_needbreak(lock
) || resched
) {
4951 if (resched
&& need_resched())
4960 EXPORT_SYMBOL(cond_resched_lock
);
4962 int __sched
cond_resched_softirq(void)
4964 BUG_ON(!in_softirq());
4966 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4974 EXPORT_SYMBOL(cond_resched_softirq
);
4977 * yield - yield the current processor to other threads.
4979 * This is a shortcut for kernel-space yielding - it marks the
4980 * thread runnable and calls sys_sched_yield().
4982 void __sched
yield(void)
4984 set_current_state(TASK_RUNNING
);
4987 EXPORT_SYMBOL(yield
);
4990 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4991 * that process accounting knows that this is a task in IO wait state.
4993 * But don't do that if it is a deliberate, throttling IO wait (this task
4994 * has set its backing_dev_info: the queue against which it should throttle)
4996 void __sched
io_schedule(void)
4998 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5000 delayacct_blkio_start();
5001 atomic_inc(&rq
->nr_iowait
);
5003 atomic_dec(&rq
->nr_iowait
);
5004 delayacct_blkio_end();
5006 EXPORT_SYMBOL(io_schedule
);
5008 long __sched
io_schedule_timeout(long timeout
)
5010 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5013 delayacct_blkio_start();
5014 atomic_inc(&rq
->nr_iowait
);
5015 ret
= schedule_timeout(timeout
);
5016 atomic_dec(&rq
->nr_iowait
);
5017 delayacct_blkio_end();
5022 * sys_sched_get_priority_max - return maximum RT priority.
5023 * @policy: scheduling class.
5025 * this syscall returns the maximum rt_priority that can be used
5026 * by a given scheduling class.
5028 asmlinkage
long sys_sched_get_priority_max(int policy
)
5035 ret
= MAX_USER_RT_PRIO
-1;
5047 * sys_sched_get_priority_min - return minimum RT priority.
5048 * @policy: scheduling class.
5050 * this syscall returns the minimum rt_priority that can be used
5051 * by a given scheduling class.
5053 asmlinkage
long sys_sched_get_priority_min(int policy
)
5071 * sys_sched_rr_get_interval - return the default timeslice of a process.
5072 * @pid: pid of the process.
5073 * @interval: userspace pointer to the timeslice value.
5075 * this syscall writes the default timeslice value of a given process
5076 * into the user-space timespec buffer. A value of '0' means infinity.
5079 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5081 struct task_struct
*p
;
5082 unsigned int time_slice
;
5090 read_lock(&tasklist_lock
);
5091 p
= find_process_by_pid(pid
);
5095 retval
= security_task_getscheduler(p
);
5100 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5101 * tasks that are on an otherwise idle runqueue:
5104 if (p
->policy
== SCHED_RR
) {
5105 time_slice
= DEF_TIMESLICE
;
5106 } else if (p
->policy
!= SCHED_FIFO
) {
5107 struct sched_entity
*se
= &p
->se
;
5108 unsigned long flags
;
5111 rq
= task_rq_lock(p
, &flags
);
5112 if (rq
->cfs
.load
.weight
)
5113 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5114 task_rq_unlock(rq
, &flags
);
5116 read_unlock(&tasklist_lock
);
5117 jiffies_to_timespec(time_slice
, &t
);
5118 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5122 read_unlock(&tasklist_lock
);
5126 static const char stat_nam
[] = "RSDTtZX";
5128 void sched_show_task(struct task_struct
*p
)
5130 unsigned long free
= 0;
5133 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5134 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5135 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5136 #if BITS_PER_LONG == 32
5137 if (state
== TASK_RUNNING
)
5138 printk(KERN_CONT
" running ");
5140 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5142 if (state
== TASK_RUNNING
)
5143 printk(KERN_CONT
" running task ");
5145 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5147 #ifdef CONFIG_DEBUG_STACK_USAGE
5149 unsigned long *n
= end_of_stack(p
);
5152 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5155 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5156 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5158 show_stack(p
, NULL
);
5161 void show_state_filter(unsigned long state_filter
)
5163 struct task_struct
*g
, *p
;
5165 #if BITS_PER_LONG == 32
5167 " task PC stack pid father\n");
5170 " task PC stack pid father\n");
5172 read_lock(&tasklist_lock
);
5173 do_each_thread(g
, p
) {
5175 * reset the NMI-timeout, listing all files on a slow
5176 * console might take alot of time:
5178 touch_nmi_watchdog();
5179 if (!state_filter
|| (p
->state
& state_filter
))
5181 } while_each_thread(g
, p
);
5183 touch_all_softlockup_watchdogs();
5185 #ifdef CONFIG_SCHED_DEBUG
5186 sysrq_sched_debug_show();
5188 read_unlock(&tasklist_lock
);
5190 * Only show locks if all tasks are dumped:
5192 if (state_filter
== -1)
5193 debug_show_all_locks();
5196 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5198 idle
->sched_class
= &idle_sched_class
;
5202 * init_idle - set up an idle thread for a given CPU
5203 * @idle: task in question
5204 * @cpu: cpu the idle task belongs to
5206 * NOTE: this function does not set the idle thread's NEED_RESCHED
5207 * flag, to make booting more robust.
5209 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5211 struct rq
*rq
= cpu_rq(cpu
);
5212 unsigned long flags
;
5215 idle
->se
.exec_start
= sched_clock();
5217 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5218 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5219 __set_task_cpu(idle
, cpu
);
5221 spin_lock_irqsave(&rq
->lock
, flags
);
5222 rq
->curr
= rq
->idle
= idle
;
5223 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5226 spin_unlock_irqrestore(&rq
->lock
, flags
);
5228 /* Set the preempt count _outside_ the spinlocks! */
5229 task_thread_info(idle
)->preempt_count
= 0;
5232 * The idle tasks have their own, simple scheduling class:
5234 idle
->sched_class
= &idle_sched_class
;
5238 * In a system that switches off the HZ timer nohz_cpu_mask
5239 * indicates which cpus entered this state. This is used
5240 * in the rcu update to wait only for active cpus. For system
5241 * which do not switch off the HZ timer nohz_cpu_mask should
5242 * always be CPU_MASK_NONE.
5244 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5247 * Increase the granularity value when there are more CPUs,
5248 * because with more CPUs the 'effective latency' as visible
5249 * to users decreases. But the relationship is not linear,
5250 * so pick a second-best guess by going with the log2 of the
5253 * This idea comes from the SD scheduler of Con Kolivas:
5255 static inline void sched_init_granularity(void)
5257 unsigned int factor
= 1 + ilog2(num_online_cpus());
5258 const unsigned long limit
= 200000000;
5260 sysctl_sched_min_granularity
*= factor
;
5261 if (sysctl_sched_min_granularity
> limit
)
5262 sysctl_sched_min_granularity
= limit
;
5264 sysctl_sched_latency
*= factor
;
5265 if (sysctl_sched_latency
> limit
)
5266 sysctl_sched_latency
= limit
;
5268 sysctl_sched_wakeup_granularity
*= factor
;
5269 sysctl_sched_batch_wakeup_granularity
*= factor
;
5274 * This is how migration works:
5276 * 1) we queue a struct migration_req structure in the source CPU's
5277 * runqueue and wake up that CPU's migration thread.
5278 * 2) we down() the locked semaphore => thread blocks.
5279 * 3) migration thread wakes up (implicitly it forces the migrated
5280 * thread off the CPU)
5281 * 4) it gets the migration request and checks whether the migrated
5282 * task is still in the wrong runqueue.
5283 * 5) if it's in the wrong runqueue then the migration thread removes
5284 * it and puts it into the right queue.
5285 * 6) migration thread up()s the semaphore.
5286 * 7) we wake up and the migration is done.
5290 * Change a given task's CPU affinity. Migrate the thread to a
5291 * proper CPU and schedule it away if the CPU it's executing on
5292 * is removed from the allowed bitmask.
5294 * NOTE: the caller must have a valid reference to the task, the
5295 * task must not exit() & deallocate itself prematurely. The
5296 * call is not atomic; no spinlocks may be held.
5298 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5300 struct migration_req req
;
5301 unsigned long flags
;
5305 rq
= task_rq_lock(p
, &flags
);
5306 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5311 if (p
->sched_class
->set_cpus_allowed
)
5312 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5314 p
->cpus_allowed
= new_mask
;
5315 p
->rt
.nr_cpus_allowed
= cpus_weight(new_mask
);
5318 /* Can the task run on the task's current CPU? If so, we're done */
5319 if (cpu_isset(task_cpu(p
), new_mask
))
5322 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5323 /* Need help from migration thread: drop lock and wait. */
5324 task_rq_unlock(rq
, &flags
);
5325 wake_up_process(rq
->migration_thread
);
5326 wait_for_completion(&req
.done
);
5327 tlb_migrate_finish(p
->mm
);
5331 task_rq_unlock(rq
, &flags
);
5335 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5338 * Move (not current) task off this cpu, onto dest cpu. We're doing
5339 * this because either it can't run here any more (set_cpus_allowed()
5340 * away from this CPU, or CPU going down), or because we're
5341 * attempting to rebalance this task on exec (sched_exec).
5343 * So we race with normal scheduler movements, but that's OK, as long
5344 * as the task is no longer on this CPU.
5346 * Returns non-zero if task was successfully migrated.
5348 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5350 struct rq
*rq_dest
, *rq_src
;
5353 if (unlikely(cpu_is_offline(dest_cpu
)))
5356 rq_src
= cpu_rq(src_cpu
);
5357 rq_dest
= cpu_rq(dest_cpu
);
5359 double_rq_lock(rq_src
, rq_dest
);
5360 /* Already moved. */
5361 if (task_cpu(p
) != src_cpu
)
5363 /* Affinity changed (again). */
5364 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5367 on_rq
= p
->se
.on_rq
;
5369 deactivate_task(rq_src
, p
, 0);
5371 set_task_cpu(p
, dest_cpu
);
5373 activate_task(rq_dest
, p
, 0);
5374 check_preempt_curr(rq_dest
, p
);
5378 double_rq_unlock(rq_src
, rq_dest
);
5383 * migration_thread - this is a highprio system thread that performs
5384 * thread migration by bumping thread off CPU then 'pushing' onto
5387 static int migration_thread(void *data
)
5389 int cpu
= (long)data
;
5393 BUG_ON(rq
->migration_thread
!= current
);
5395 set_current_state(TASK_INTERRUPTIBLE
);
5396 while (!kthread_should_stop()) {
5397 struct migration_req
*req
;
5398 struct list_head
*head
;
5400 spin_lock_irq(&rq
->lock
);
5402 if (cpu_is_offline(cpu
)) {
5403 spin_unlock_irq(&rq
->lock
);
5407 if (rq
->active_balance
) {
5408 active_load_balance(rq
, cpu
);
5409 rq
->active_balance
= 0;
5412 head
= &rq
->migration_queue
;
5414 if (list_empty(head
)) {
5415 spin_unlock_irq(&rq
->lock
);
5417 set_current_state(TASK_INTERRUPTIBLE
);
5420 req
= list_entry(head
->next
, struct migration_req
, list
);
5421 list_del_init(head
->next
);
5423 spin_unlock(&rq
->lock
);
5424 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5427 complete(&req
->done
);
5429 __set_current_state(TASK_RUNNING
);
5433 /* Wait for kthread_stop */
5434 set_current_state(TASK_INTERRUPTIBLE
);
5435 while (!kthread_should_stop()) {
5437 set_current_state(TASK_INTERRUPTIBLE
);
5439 __set_current_state(TASK_RUNNING
);
5443 #ifdef CONFIG_HOTPLUG_CPU
5445 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5449 local_irq_disable();
5450 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5456 * Figure out where task on dead CPU should go, use force if necessary.
5457 * NOTE: interrupts should be disabled by the caller
5459 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5461 unsigned long flags
;
5468 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5469 cpus_and(mask
, mask
, p
->cpus_allowed
);
5470 dest_cpu
= any_online_cpu(mask
);
5472 /* On any allowed CPU? */
5473 if (dest_cpu
== NR_CPUS
)
5474 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5476 /* No more Mr. Nice Guy. */
5477 if (dest_cpu
== NR_CPUS
) {
5478 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5480 * Try to stay on the same cpuset, where the
5481 * current cpuset may be a subset of all cpus.
5482 * The cpuset_cpus_allowed_locked() variant of
5483 * cpuset_cpus_allowed() will not block. It must be
5484 * called within calls to cpuset_lock/cpuset_unlock.
5486 rq
= task_rq_lock(p
, &flags
);
5487 p
->cpus_allowed
= cpus_allowed
;
5488 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5489 task_rq_unlock(rq
, &flags
);
5492 * Don't tell them about moving exiting tasks or
5493 * kernel threads (both mm NULL), since they never
5496 if (p
->mm
&& printk_ratelimit()) {
5497 printk(KERN_INFO
"process %d (%s) no "
5498 "longer affine to cpu%d\n",
5499 task_pid_nr(p
), p
->comm
, dead_cpu
);
5502 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5506 * While a dead CPU has no uninterruptible tasks queued at this point,
5507 * it might still have a nonzero ->nr_uninterruptible counter, because
5508 * for performance reasons the counter is not stricly tracking tasks to
5509 * their home CPUs. So we just add the counter to another CPU's counter,
5510 * to keep the global sum constant after CPU-down:
5512 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5514 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5515 unsigned long flags
;
5517 local_irq_save(flags
);
5518 double_rq_lock(rq_src
, rq_dest
);
5519 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5520 rq_src
->nr_uninterruptible
= 0;
5521 double_rq_unlock(rq_src
, rq_dest
);
5522 local_irq_restore(flags
);
5525 /* Run through task list and migrate tasks from the dead cpu. */
5526 static void migrate_live_tasks(int src_cpu
)
5528 struct task_struct
*p
, *t
;
5530 read_lock(&tasklist_lock
);
5532 do_each_thread(t
, p
) {
5536 if (task_cpu(p
) == src_cpu
)
5537 move_task_off_dead_cpu(src_cpu
, p
);
5538 } while_each_thread(t
, p
);
5540 read_unlock(&tasklist_lock
);
5544 * Schedules idle task to be the next runnable task on current CPU.
5545 * It does so by boosting its priority to highest possible.
5546 * Used by CPU offline code.
5548 void sched_idle_next(void)
5550 int this_cpu
= smp_processor_id();
5551 struct rq
*rq
= cpu_rq(this_cpu
);
5552 struct task_struct
*p
= rq
->idle
;
5553 unsigned long flags
;
5555 /* cpu has to be offline */
5556 BUG_ON(cpu_online(this_cpu
));
5559 * Strictly not necessary since rest of the CPUs are stopped by now
5560 * and interrupts disabled on the current cpu.
5562 spin_lock_irqsave(&rq
->lock
, flags
);
5564 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5566 update_rq_clock(rq
);
5567 activate_task(rq
, p
, 0);
5569 spin_unlock_irqrestore(&rq
->lock
, flags
);
5573 * Ensures that the idle task is using init_mm right before its cpu goes
5576 void idle_task_exit(void)
5578 struct mm_struct
*mm
= current
->active_mm
;
5580 BUG_ON(cpu_online(smp_processor_id()));
5583 switch_mm(mm
, &init_mm
, current
);
5587 /* called under rq->lock with disabled interrupts */
5588 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5590 struct rq
*rq
= cpu_rq(dead_cpu
);
5592 /* Must be exiting, otherwise would be on tasklist. */
5593 BUG_ON(!p
->exit_state
);
5595 /* Cannot have done final schedule yet: would have vanished. */
5596 BUG_ON(p
->state
== TASK_DEAD
);
5601 * Drop lock around migration; if someone else moves it,
5602 * that's OK. No task can be added to this CPU, so iteration is
5605 spin_unlock_irq(&rq
->lock
);
5606 move_task_off_dead_cpu(dead_cpu
, p
);
5607 spin_lock_irq(&rq
->lock
);
5612 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5613 static void migrate_dead_tasks(unsigned int dead_cpu
)
5615 struct rq
*rq
= cpu_rq(dead_cpu
);
5616 struct task_struct
*next
;
5619 if (!rq
->nr_running
)
5621 update_rq_clock(rq
);
5622 next
= pick_next_task(rq
, rq
->curr
);
5625 migrate_dead(dead_cpu
, next
);
5629 #endif /* CONFIG_HOTPLUG_CPU */
5631 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5633 static struct ctl_table sd_ctl_dir
[] = {
5635 .procname
= "sched_domain",
5641 static struct ctl_table sd_ctl_root
[] = {
5643 .ctl_name
= CTL_KERN
,
5644 .procname
= "kernel",
5646 .child
= sd_ctl_dir
,
5651 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5653 struct ctl_table
*entry
=
5654 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5659 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5661 struct ctl_table
*entry
;
5664 * In the intermediate directories, both the child directory and
5665 * procname are dynamically allocated and could fail but the mode
5666 * will always be set. In the lowest directory the names are
5667 * static strings and all have proc handlers.
5669 for (entry
= *tablep
; entry
->mode
; entry
++) {
5671 sd_free_ctl_entry(&entry
->child
);
5672 if (entry
->proc_handler
== NULL
)
5673 kfree(entry
->procname
);
5681 set_table_entry(struct ctl_table
*entry
,
5682 const char *procname
, void *data
, int maxlen
,
5683 mode_t mode
, proc_handler
*proc_handler
)
5685 entry
->procname
= procname
;
5687 entry
->maxlen
= maxlen
;
5689 entry
->proc_handler
= proc_handler
;
5692 static struct ctl_table
*
5693 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5695 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5700 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5701 sizeof(long), 0644, proc_doulongvec_minmax
);
5702 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5703 sizeof(long), 0644, proc_doulongvec_minmax
);
5704 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5705 sizeof(int), 0644, proc_dointvec_minmax
);
5706 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5707 sizeof(int), 0644, proc_dointvec_minmax
);
5708 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5709 sizeof(int), 0644, proc_dointvec_minmax
);
5710 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5711 sizeof(int), 0644, proc_dointvec_minmax
);
5712 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5713 sizeof(int), 0644, proc_dointvec_minmax
);
5714 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5715 sizeof(int), 0644, proc_dointvec_minmax
);
5716 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5717 sizeof(int), 0644, proc_dointvec_minmax
);
5718 set_table_entry(&table
[9], "cache_nice_tries",
5719 &sd
->cache_nice_tries
,
5720 sizeof(int), 0644, proc_dointvec_minmax
);
5721 set_table_entry(&table
[10], "flags", &sd
->flags
,
5722 sizeof(int), 0644, proc_dointvec_minmax
);
5723 /* &table[11] is terminator */
5728 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5730 struct ctl_table
*entry
, *table
;
5731 struct sched_domain
*sd
;
5732 int domain_num
= 0, i
;
5735 for_each_domain(cpu
, sd
)
5737 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5742 for_each_domain(cpu
, sd
) {
5743 snprintf(buf
, 32, "domain%d", i
);
5744 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5746 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5753 static struct ctl_table_header
*sd_sysctl_header
;
5754 static void register_sched_domain_sysctl(void)
5756 int i
, cpu_num
= num_online_cpus();
5757 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5760 WARN_ON(sd_ctl_dir
[0].child
);
5761 sd_ctl_dir
[0].child
= entry
;
5766 for_each_online_cpu(i
) {
5767 snprintf(buf
, 32, "cpu%d", i
);
5768 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5770 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5774 WARN_ON(sd_sysctl_header
);
5775 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5778 /* may be called multiple times per register */
5779 static void unregister_sched_domain_sysctl(void)
5781 if (sd_sysctl_header
)
5782 unregister_sysctl_table(sd_sysctl_header
);
5783 sd_sysctl_header
= NULL
;
5784 if (sd_ctl_dir
[0].child
)
5785 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5788 static void register_sched_domain_sysctl(void)
5791 static void unregister_sched_domain_sysctl(void)
5797 * migration_call - callback that gets triggered when a CPU is added.
5798 * Here we can start up the necessary migration thread for the new CPU.
5800 static int __cpuinit
5801 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5803 struct task_struct
*p
;
5804 int cpu
= (long)hcpu
;
5805 unsigned long flags
;
5810 case CPU_UP_PREPARE
:
5811 case CPU_UP_PREPARE_FROZEN
:
5812 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5815 kthread_bind(p
, cpu
);
5816 /* Must be high prio: stop_machine expects to yield to it. */
5817 rq
= task_rq_lock(p
, &flags
);
5818 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5819 task_rq_unlock(rq
, &flags
);
5820 cpu_rq(cpu
)->migration_thread
= p
;
5824 case CPU_ONLINE_FROZEN
:
5825 /* Strictly unnecessary, as first user will wake it. */
5826 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5828 /* Update our root-domain */
5830 spin_lock_irqsave(&rq
->lock
, flags
);
5832 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5833 cpu_set(cpu
, rq
->rd
->online
);
5835 spin_unlock_irqrestore(&rq
->lock
, flags
);
5838 #ifdef CONFIG_HOTPLUG_CPU
5839 case CPU_UP_CANCELED
:
5840 case CPU_UP_CANCELED_FROZEN
:
5841 if (!cpu_rq(cpu
)->migration_thread
)
5843 /* Unbind it from offline cpu so it can run. Fall thru. */
5844 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5845 any_online_cpu(cpu_online_map
));
5846 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5847 cpu_rq(cpu
)->migration_thread
= NULL
;
5851 case CPU_DEAD_FROZEN
:
5852 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5853 migrate_live_tasks(cpu
);
5855 kthread_stop(rq
->migration_thread
);
5856 rq
->migration_thread
= NULL
;
5857 /* Idle task back to normal (off runqueue, low prio) */
5858 spin_lock_irq(&rq
->lock
);
5859 update_rq_clock(rq
);
5860 deactivate_task(rq
, rq
->idle
, 0);
5861 rq
->idle
->static_prio
= MAX_PRIO
;
5862 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5863 rq
->idle
->sched_class
= &idle_sched_class
;
5864 migrate_dead_tasks(cpu
);
5865 spin_unlock_irq(&rq
->lock
);
5867 migrate_nr_uninterruptible(rq
);
5868 BUG_ON(rq
->nr_running
!= 0);
5871 * No need to migrate the tasks: it was best-effort if
5872 * they didn't take sched_hotcpu_mutex. Just wake up
5875 spin_lock_irq(&rq
->lock
);
5876 while (!list_empty(&rq
->migration_queue
)) {
5877 struct migration_req
*req
;
5879 req
= list_entry(rq
->migration_queue
.next
,
5880 struct migration_req
, list
);
5881 list_del_init(&req
->list
);
5882 complete(&req
->done
);
5884 spin_unlock_irq(&rq
->lock
);
5888 case CPU_DYING_FROZEN
:
5889 /* Update our root-domain */
5891 spin_lock_irqsave(&rq
->lock
, flags
);
5893 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5894 cpu_clear(cpu
, rq
->rd
->online
);
5896 spin_unlock_irqrestore(&rq
->lock
, flags
);
5903 /* Register at highest priority so that task migration (migrate_all_tasks)
5904 * happens before everything else.
5906 static struct notifier_block __cpuinitdata migration_notifier
= {
5907 .notifier_call
= migration_call
,
5911 void __init
migration_init(void)
5913 void *cpu
= (void *)(long)smp_processor_id();
5916 /* Start one for the boot CPU: */
5917 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5918 BUG_ON(err
== NOTIFY_BAD
);
5919 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5920 register_cpu_notifier(&migration_notifier
);
5926 /* Number of possible processor ids */
5927 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5928 EXPORT_SYMBOL(nr_cpu_ids
);
5930 #ifdef CONFIG_SCHED_DEBUG
5932 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5934 struct sched_group
*group
= sd
->groups
;
5935 cpumask_t groupmask
;
5938 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5939 cpus_clear(groupmask
);
5941 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5943 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5944 printk("does not load-balance\n");
5946 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5951 printk(KERN_CONT
"span %s\n", str
);
5953 if (!cpu_isset(cpu
, sd
->span
)) {
5954 printk(KERN_ERR
"ERROR: domain->span does not contain "
5957 if (!cpu_isset(cpu
, group
->cpumask
)) {
5958 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5962 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5966 printk(KERN_ERR
"ERROR: group is NULL\n");
5970 if (!group
->__cpu_power
) {
5971 printk(KERN_CONT
"\n");
5972 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5977 if (!cpus_weight(group
->cpumask
)) {
5978 printk(KERN_CONT
"\n");
5979 printk(KERN_ERR
"ERROR: empty group\n");
5983 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5984 printk(KERN_CONT
"\n");
5985 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5989 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5991 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5992 printk(KERN_CONT
" %s", str
);
5994 group
= group
->next
;
5995 } while (group
!= sd
->groups
);
5996 printk(KERN_CONT
"\n");
5998 if (!cpus_equal(sd
->span
, groupmask
))
5999 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6001 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
6002 printk(KERN_ERR
"ERROR: parent span is not a superset "
6003 "of domain->span\n");
6007 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6012 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6016 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6019 if (sched_domain_debug_one(sd
, cpu
, level
))
6028 # define sched_domain_debug(sd, cpu) do { } while (0)
6031 static int sd_degenerate(struct sched_domain
*sd
)
6033 if (cpus_weight(sd
->span
) == 1)
6036 /* Following flags need at least 2 groups */
6037 if (sd
->flags
& (SD_LOAD_BALANCE
|
6038 SD_BALANCE_NEWIDLE
|
6042 SD_SHARE_PKG_RESOURCES
)) {
6043 if (sd
->groups
!= sd
->groups
->next
)
6047 /* Following flags don't use groups */
6048 if (sd
->flags
& (SD_WAKE_IDLE
|
6057 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6059 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6061 if (sd_degenerate(parent
))
6064 if (!cpus_equal(sd
->span
, parent
->span
))
6067 /* Does parent contain flags not in child? */
6068 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6069 if (cflags
& SD_WAKE_AFFINE
)
6070 pflags
&= ~SD_WAKE_BALANCE
;
6071 /* Flags needing groups don't count if only 1 group in parent */
6072 if (parent
->groups
== parent
->groups
->next
) {
6073 pflags
&= ~(SD_LOAD_BALANCE
|
6074 SD_BALANCE_NEWIDLE
|
6078 SD_SHARE_PKG_RESOURCES
);
6080 if (~cflags
& pflags
)
6086 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6088 unsigned long flags
;
6089 const struct sched_class
*class;
6091 spin_lock_irqsave(&rq
->lock
, flags
);
6094 struct root_domain
*old_rd
= rq
->rd
;
6096 for (class = sched_class_highest
; class; class = class->next
) {
6097 if (class->leave_domain
)
6098 class->leave_domain(rq
);
6101 cpu_clear(rq
->cpu
, old_rd
->span
);
6102 cpu_clear(rq
->cpu
, old_rd
->online
);
6104 if (atomic_dec_and_test(&old_rd
->refcount
))
6108 atomic_inc(&rd
->refcount
);
6111 cpu_set(rq
->cpu
, rd
->span
);
6112 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6113 cpu_set(rq
->cpu
, rd
->online
);
6115 for (class = sched_class_highest
; class; class = class->next
) {
6116 if (class->join_domain
)
6117 class->join_domain(rq
);
6120 spin_unlock_irqrestore(&rq
->lock
, flags
);
6123 static void init_rootdomain(struct root_domain
*rd
)
6125 memset(rd
, 0, sizeof(*rd
));
6127 cpus_clear(rd
->span
);
6128 cpus_clear(rd
->online
);
6131 static void init_defrootdomain(void)
6133 init_rootdomain(&def_root_domain
);
6134 atomic_set(&def_root_domain
.refcount
, 1);
6137 static struct root_domain
*alloc_rootdomain(void)
6139 struct root_domain
*rd
;
6141 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6145 init_rootdomain(rd
);
6151 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6152 * hold the hotplug lock.
6155 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6157 struct rq
*rq
= cpu_rq(cpu
);
6158 struct sched_domain
*tmp
;
6160 /* Remove the sched domains which do not contribute to scheduling. */
6161 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6162 struct sched_domain
*parent
= tmp
->parent
;
6165 if (sd_parent_degenerate(tmp
, parent
)) {
6166 tmp
->parent
= parent
->parent
;
6168 parent
->parent
->child
= tmp
;
6172 if (sd
&& sd_degenerate(sd
)) {
6178 sched_domain_debug(sd
, cpu
);
6180 rq_attach_root(rq
, rd
);
6181 rcu_assign_pointer(rq
->sd
, sd
);
6184 /* cpus with isolated domains */
6185 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6187 /* Setup the mask of cpus configured for isolated domains */
6188 static int __init
isolated_cpu_setup(char *str
)
6190 int ints
[NR_CPUS
], i
;
6192 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6193 cpus_clear(cpu_isolated_map
);
6194 for (i
= 1; i
<= ints
[0]; i
++)
6195 if (ints
[i
] < NR_CPUS
)
6196 cpu_set(ints
[i
], cpu_isolated_map
);
6200 __setup("isolcpus=", isolated_cpu_setup
);
6203 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6204 * to a function which identifies what group(along with sched group) a CPU
6205 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6206 * (due to the fact that we keep track of groups covered with a cpumask_t).
6208 * init_sched_build_groups will build a circular linked list of the groups
6209 * covered by the given span, and will set each group's ->cpumask correctly,
6210 * and ->cpu_power to 0.
6213 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
6214 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6215 struct sched_group
**sg
))
6217 struct sched_group
*first
= NULL
, *last
= NULL
;
6218 cpumask_t covered
= CPU_MASK_NONE
;
6221 for_each_cpu_mask(i
, span
) {
6222 struct sched_group
*sg
;
6223 int group
= group_fn(i
, cpu_map
, &sg
);
6226 if (cpu_isset(i
, covered
))
6229 sg
->cpumask
= CPU_MASK_NONE
;
6230 sg
->__cpu_power
= 0;
6232 for_each_cpu_mask(j
, span
) {
6233 if (group_fn(j
, cpu_map
, NULL
) != group
)
6236 cpu_set(j
, covered
);
6237 cpu_set(j
, sg
->cpumask
);
6248 #define SD_NODES_PER_DOMAIN 16
6253 * find_next_best_node - find the next node to include in a sched_domain
6254 * @node: node whose sched_domain we're building
6255 * @used_nodes: nodes already in the sched_domain
6257 * Find the next node to include in a given scheduling domain. Simply
6258 * finds the closest node not already in the @used_nodes map.
6260 * Should use nodemask_t.
6262 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6264 int i
, n
, val
, min_val
, best_node
= 0;
6268 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6269 /* Start at @node */
6270 n
= (node
+ i
) % MAX_NUMNODES
;
6272 if (!nr_cpus_node(n
))
6275 /* Skip already used nodes */
6276 if (test_bit(n
, used_nodes
))
6279 /* Simple min distance search */
6280 val
= node_distance(node
, n
);
6282 if (val
< min_val
) {
6288 set_bit(best_node
, used_nodes
);
6293 * sched_domain_node_span - get a cpumask for a node's sched_domain
6294 * @node: node whose cpumask we're constructing
6295 * @size: number of nodes to include in this span
6297 * Given a node, construct a good cpumask for its sched_domain to span. It
6298 * should be one that prevents unnecessary balancing, but also spreads tasks
6301 static cpumask_t
sched_domain_node_span(int node
)
6303 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6304 cpumask_t span
, nodemask
;
6308 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6310 nodemask
= node_to_cpumask(node
);
6311 cpus_or(span
, span
, nodemask
);
6312 set_bit(node
, used_nodes
);
6314 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6315 int next_node
= find_next_best_node(node
, used_nodes
);
6317 nodemask
= node_to_cpumask(next_node
);
6318 cpus_or(span
, span
, nodemask
);
6325 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6328 * SMT sched-domains:
6330 #ifdef CONFIG_SCHED_SMT
6331 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6332 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6335 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6338 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6344 * multi-core sched-domains:
6346 #ifdef CONFIG_SCHED_MC
6347 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6348 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6351 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6353 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6356 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6357 cpus_and(mask
, mask
, *cpu_map
);
6358 group
= first_cpu(mask
);
6360 *sg
= &per_cpu(sched_group_core
, group
);
6363 #elif defined(CONFIG_SCHED_MC)
6365 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6368 *sg
= &per_cpu(sched_group_core
, cpu
);
6373 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6374 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6377 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6380 #ifdef CONFIG_SCHED_MC
6381 cpumask_t mask
= cpu_coregroup_map(cpu
);
6382 cpus_and(mask
, mask
, *cpu_map
);
6383 group
= first_cpu(mask
);
6384 #elif defined(CONFIG_SCHED_SMT)
6385 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6386 cpus_and(mask
, mask
, *cpu_map
);
6387 group
= first_cpu(mask
);
6392 *sg
= &per_cpu(sched_group_phys
, group
);
6398 * The init_sched_build_groups can't handle what we want to do with node
6399 * groups, so roll our own. Now each node has its own list of groups which
6400 * gets dynamically allocated.
6402 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6403 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6405 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6406 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6408 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6409 struct sched_group
**sg
)
6411 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6414 cpus_and(nodemask
, nodemask
, *cpu_map
);
6415 group
= first_cpu(nodemask
);
6418 *sg
= &per_cpu(sched_group_allnodes
, group
);
6422 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6424 struct sched_group
*sg
= group_head
;
6430 for_each_cpu_mask(j
, sg
->cpumask
) {
6431 struct sched_domain
*sd
;
6433 sd
= &per_cpu(phys_domains
, j
);
6434 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6436 * Only add "power" once for each
6442 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6445 } while (sg
!= group_head
);
6450 /* Free memory allocated for various sched_group structures */
6451 static void free_sched_groups(const cpumask_t
*cpu_map
)
6455 for_each_cpu_mask(cpu
, *cpu_map
) {
6456 struct sched_group
**sched_group_nodes
6457 = sched_group_nodes_bycpu
[cpu
];
6459 if (!sched_group_nodes
)
6462 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6463 cpumask_t nodemask
= node_to_cpumask(i
);
6464 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6466 cpus_and(nodemask
, nodemask
, *cpu_map
);
6467 if (cpus_empty(nodemask
))
6477 if (oldsg
!= sched_group_nodes
[i
])
6480 kfree(sched_group_nodes
);
6481 sched_group_nodes_bycpu
[cpu
] = NULL
;
6485 static void free_sched_groups(const cpumask_t
*cpu_map
)
6491 * Initialize sched groups cpu_power.
6493 * cpu_power indicates the capacity of sched group, which is used while
6494 * distributing the load between different sched groups in a sched domain.
6495 * Typically cpu_power for all the groups in a sched domain will be same unless
6496 * there are asymmetries in the topology. If there are asymmetries, group
6497 * having more cpu_power will pickup more load compared to the group having
6500 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6501 * the maximum number of tasks a group can handle in the presence of other idle
6502 * or lightly loaded groups in the same sched domain.
6504 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6506 struct sched_domain
*child
;
6507 struct sched_group
*group
;
6509 WARN_ON(!sd
|| !sd
->groups
);
6511 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6516 sd
->groups
->__cpu_power
= 0;
6519 * For perf policy, if the groups in child domain share resources
6520 * (for example cores sharing some portions of the cache hierarchy
6521 * or SMT), then set this domain groups cpu_power such that each group
6522 * can handle only one task, when there are other idle groups in the
6523 * same sched domain.
6525 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6527 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6528 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6533 * add cpu_power of each child group to this groups cpu_power
6535 group
= child
->groups
;
6537 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6538 group
= group
->next
;
6539 } while (group
!= child
->groups
);
6543 * Build sched domains for a given set of cpus and attach the sched domains
6544 * to the individual cpus
6546 static int build_sched_domains(const cpumask_t
*cpu_map
)
6549 struct root_domain
*rd
;
6551 struct sched_group
**sched_group_nodes
= NULL
;
6552 int sd_allnodes
= 0;
6555 * Allocate the per-node list of sched groups
6557 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6559 if (!sched_group_nodes
) {
6560 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6563 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6566 rd
= alloc_rootdomain();
6568 printk(KERN_WARNING
"Cannot alloc root domain\n");
6573 * Set up domains for cpus specified by the cpu_map.
6575 for_each_cpu_mask(i
, *cpu_map
) {
6576 struct sched_domain
*sd
= NULL
, *p
;
6577 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6579 cpus_and(nodemask
, nodemask
, *cpu_map
);
6582 if (cpus_weight(*cpu_map
) >
6583 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6584 sd
= &per_cpu(allnodes_domains
, i
);
6585 *sd
= SD_ALLNODES_INIT
;
6586 sd
->span
= *cpu_map
;
6587 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6593 sd
= &per_cpu(node_domains
, i
);
6595 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6599 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6603 sd
= &per_cpu(phys_domains
, i
);
6605 sd
->span
= nodemask
;
6609 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6611 #ifdef CONFIG_SCHED_MC
6613 sd
= &per_cpu(core_domains
, i
);
6615 sd
->span
= cpu_coregroup_map(i
);
6616 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6619 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6622 #ifdef CONFIG_SCHED_SMT
6624 sd
= &per_cpu(cpu_domains
, i
);
6625 *sd
= SD_SIBLING_INIT
;
6626 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6627 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6630 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6634 #ifdef CONFIG_SCHED_SMT
6635 /* Set up CPU (sibling) groups */
6636 for_each_cpu_mask(i
, *cpu_map
) {
6637 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6638 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6639 if (i
!= first_cpu(this_sibling_map
))
6642 init_sched_build_groups(this_sibling_map
, cpu_map
,
6647 #ifdef CONFIG_SCHED_MC
6648 /* Set up multi-core groups */
6649 for_each_cpu_mask(i
, *cpu_map
) {
6650 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6651 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6652 if (i
!= first_cpu(this_core_map
))
6654 init_sched_build_groups(this_core_map
, cpu_map
,
6655 &cpu_to_core_group
);
6659 /* Set up physical groups */
6660 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6661 cpumask_t nodemask
= node_to_cpumask(i
);
6663 cpus_and(nodemask
, nodemask
, *cpu_map
);
6664 if (cpus_empty(nodemask
))
6667 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6671 /* Set up node groups */
6673 init_sched_build_groups(*cpu_map
, cpu_map
,
6674 &cpu_to_allnodes_group
);
6676 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6677 /* Set up node groups */
6678 struct sched_group
*sg
, *prev
;
6679 cpumask_t nodemask
= node_to_cpumask(i
);
6680 cpumask_t domainspan
;
6681 cpumask_t covered
= CPU_MASK_NONE
;
6684 cpus_and(nodemask
, nodemask
, *cpu_map
);
6685 if (cpus_empty(nodemask
)) {
6686 sched_group_nodes
[i
] = NULL
;
6690 domainspan
= sched_domain_node_span(i
);
6691 cpus_and(domainspan
, domainspan
, *cpu_map
);
6693 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6695 printk(KERN_WARNING
"Can not alloc domain group for "
6699 sched_group_nodes
[i
] = sg
;
6700 for_each_cpu_mask(j
, nodemask
) {
6701 struct sched_domain
*sd
;
6703 sd
= &per_cpu(node_domains
, j
);
6706 sg
->__cpu_power
= 0;
6707 sg
->cpumask
= nodemask
;
6709 cpus_or(covered
, covered
, nodemask
);
6712 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6713 cpumask_t tmp
, notcovered
;
6714 int n
= (i
+ j
) % MAX_NUMNODES
;
6716 cpus_complement(notcovered
, covered
);
6717 cpus_and(tmp
, notcovered
, *cpu_map
);
6718 cpus_and(tmp
, tmp
, domainspan
);
6719 if (cpus_empty(tmp
))
6722 nodemask
= node_to_cpumask(n
);
6723 cpus_and(tmp
, tmp
, nodemask
);
6724 if (cpus_empty(tmp
))
6727 sg
= kmalloc_node(sizeof(struct sched_group
),
6731 "Can not alloc domain group for node %d\n", j
);
6734 sg
->__cpu_power
= 0;
6736 sg
->next
= prev
->next
;
6737 cpus_or(covered
, covered
, tmp
);
6744 /* Calculate CPU power for physical packages and nodes */
6745 #ifdef CONFIG_SCHED_SMT
6746 for_each_cpu_mask(i
, *cpu_map
) {
6747 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6749 init_sched_groups_power(i
, sd
);
6752 #ifdef CONFIG_SCHED_MC
6753 for_each_cpu_mask(i
, *cpu_map
) {
6754 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6756 init_sched_groups_power(i
, sd
);
6760 for_each_cpu_mask(i
, *cpu_map
) {
6761 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6763 init_sched_groups_power(i
, sd
);
6767 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6768 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6771 struct sched_group
*sg
;
6773 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6774 init_numa_sched_groups_power(sg
);
6778 /* Attach the domains */
6779 for_each_cpu_mask(i
, *cpu_map
) {
6780 struct sched_domain
*sd
;
6781 #ifdef CONFIG_SCHED_SMT
6782 sd
= &per_cpu(cpu_domains
, i
);
6783 #elif defined(CONFIG_SCHED_MC)
6784 sd
= &per_cpu(core_domains
, i
);
6786 sd
= &per_cpu(phys_domains
, i
);
6788 cpu_attach_domain(sd
, rd
, i
);
6795 free_sched_groups(cpu_map
);
6800 static cpumask_t
*doms_cur
; /* current sched domains */
6801 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6804 * Special case: If a kmalloc of a doms_cur partition (array of
6805 * cpumask_t) fails, then fallback to a single sched domain,
6806 * as determined by the single cpumask_t fallback_doms.
6808 static cpumask_t fallback_doms
;
6811 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6812 * For now this just excludes isolated cpus, but could be used to
6813 * exclude other special cases in the future.
6815 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6820 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6822 doms_cur
= &fallback_doms
;
6823 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6824 err
= build_sched_domains(doms_cur
);
6825 register_sched_domain_sysctl();
6830 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6832 free_sched_groups(cpu_map
);
6836 * Detach sched domains from a group of cpus specified in cpu_map
6837 * These cpus will now be attached to the NULL domain
6839 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6843 unregister_sched_domain_sysctl();
6845 for_each_cpu_mask(i
, *cpu_map
)
6846 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6847 synchronize_sched();
6848 arch_destroy_sched_domains(cpu_map
);
6852 * Partition sched domains as specified by the 'ndoms_new'
6853 * cpumasks in the array doms_new[] of cpumasks. This compares
6854 * doms_new[] to the current sched domain partitioning, doms_cur[].
6855 * It destroys each deleted domain and builds each new domain.
6857 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6858 * The masks don't intersect (don't overlap.) We should setup one
6859 * sched domain for each mask. CPUs not in any of the cpumasks will
6860 * not be load balanced. If the same cpumask appears both in the
6861 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6864 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6865 * ownership of it and will kfree it when done with it. If the caller
6866 * failed the kmalloc call, then it can pass in doms_new == NULL,
6867 * and partition_sched_domains() will fallback to the single partition
6870 * Call with hotplug lock held
6872 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6878 /* always unregister in case we don't destroy any domains */
6879 unregister_sched_domain_sysctl();
6881 if (doms_new
== NULL
) {
6883 doms_new
= &fallback_doms
;
6884 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6887 /* Destroy deleted domains */
6888 for (i
= 0; i
< ndoms_cur
; i
++) {
6889 for (j
= 0; j
< ndoms_new
; j
++) {
6890 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6893 /* no match - a current sched domain not in new doms_new[] */
6894 detach_destroy_domains(doms_cur
+ i
);
6899 /* Build new domains */
6900 for (i
= 0; i
< ndoms_new
; i
++) {
6901 for (j
= 0; j
< ndoms_cur
; j
++) {
6902 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6905 /* no match - add a new doms_new */
6906 build_sched_domains(doms_new
+ i
);
6911 /* Remember the new sched domains */
6912 if (doms_cur
!= &fallback_doms
)
6914 doms_cur
= doms_new
;
6915 ndoms_cur
= ndoms_new
;
6917 register_sched_domain_sysctl();
6922 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6923 static int arch_reinit_sched_domains(void)
6928 detach_destroy_domains(&cpu_online_map
);
6929 err
= arch_init_sched_domains(&cpu_online_map
);
6935 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6939 if (buf
[0] != '0' && buf
[0] != '1')
6943 sched_smt_power_savings
= (buf
[0] == '1');
6945 sched_mc_power_savings
= (buf
[0] == '1');
6947 ret
= arch_reinit_sched_domains();
6949 return ret
? ret
: count
;
6952 #ifdef CONFIG_SCHED_MC
6953 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6955 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6957 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6958 const char *buf
, size_t count
)
6960 return sched_power_savings_store(buf
, count
, 0);
6962 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6963 sched_mc_power_savings_store
);
6966 #ifdef CONFIG_SCHED_SMT
6967 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6969 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6971 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6972 const char *buf
, size_t count
)
6974 return sched_power_savings_store(buf
, count
, 1);
6976 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6977 sched_smt_power_savings_store
);
6980 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6984 #ifdef CONFIG_SCHED_SMT
6986 err
= sysfs_create_file(&cls
->kset
.kobj
,
6987 &attr_sched_smt_power_savings
.attr
);
6989 #ifdef CONFIG_SCHED_MC
6990 if (!err
&& mc_capable())
6991 err
= sysfs_create_file(&cls
->kset
.kobj
,
6992 &attr_sched_mc_power_savings
.attr
);
6999 * Force a reinitialization of the sched domains hierarchy. The domains
7000 * and groups cannot be updated in place without racing with the balancing
7001 * code, so we temporarily attach all running cpus to the NULL domain
7002 * which will prevent rebalancing while the sched domains are recalculated.
7004 static int update_sched_domains(struct notifier_block
*nfb
,
7005 unsigned long action
, void *hcpu
)
7008 case CPU_UP_PREPARE
:
7009 case CPU_UP_PREPARE_FROZEN
:
7010 case CPU_DOWN_PREPARE
:
7011 case CPU_DOWN_PREPARE_FROZEN
:
7012 detach_destroy_domains(&cpu_online_map
);
7015 case CPU_UP_CANCELED
:
7016 case CPU_UP_CANCELED_FROZEN
:
7017 case CPU_DOWN_FAILED
:
7018 case CPU_DOWN_FAILED_FROZEN
:
7020 case CPU_ONLINE_FROZEN
:
7022 case CPU_DEAD_FROZEN
:
7024 * Fall through and re-initialise the domains.
7031 /* The hotplug lock is already held by cpu_up/cpu_down */
7032 arch_init_sched_domains(&cpu_online_map
);
7037 void __init
sched_init_smp(void)
7039 cpumask_t non_isolated_cpus
;
7042 arch_init_sched_domains(&cpu_online_map
);
7043 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7044 if (cpus_empty(non_isolated_cpus
))
7045 cpu_set(smp_processor_id(), non_isolated_cpus
);
7047 /* XXX: Theoretical race here - CPU may be hotplugged now */
7048 hotcpu_notifier(update_sched_domains
, 0);
7050 /* Move init over to a non-isolated CPU */
7051 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
7053 sched_init_granularity();
7056 void __init
sched_init_smp(void)
7058 sched_init_granularity();
7060 #endif /* CONFIG_SMP */
7062 int in_sched_functions(unsigned long addr
)
7064 return in_lock_functions(addr
) ||
7065 (addr
>= (unsigned long)__sched_text_start
7066 && addr
< (unsigned long)__sched_text_end
);
7069 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7071 cfs_rq
->tasks_timeline
= RB_ROOT
;
7072 #ifdef CONFIG_FAIR_GROUP_SCHED
7075 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7078 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7080 struct rt_prio_array
*array
;
7083 array
= &rt_rq
->active
;
7084 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7085 INIT_LIST_HEAD(array
->queue
+ i
);
7086 __clear_bit(i
, array
->bitmap
);
7088 /* delimiter for bitsearch: */
7089 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7091 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7092 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7095 rt_rq
->rt_nr_migratory
= 0;
7096 rt_rq
->overloaded
= 0;
7100 rt_rq
->rt_throttled
= 0;
7102 #ifdef CONFIG_RT_GROUP_SCHED
7103 rt_rq
->rt_nr_boosted
= 0;
7108 #ifdef CONFIG_FAIR_GROUP_SCHED
7109 static void init_tg_cfs_entry(struct rq
*rq
, struct task_group
*tg
,
7110 struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
7113 tg
->cfs_rq
[cpu
] = cfs_rq
;
7114 init_cfs_rq(cfs_rq
, rq
);
7117 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7120 se
->cfs_rq
= &rq
->cfs
;
7122 se
->load
.weight
= tg
->shares
;
7123 se
->load
.inv_weight
= div64_64(1ULL<<32, se
->load
.weight
);
7128 #ifdef CONFIG_RT_GROUP_SCHED
7129 static void init_tg_rt_entry(struct rq
*rq
, struct task_group
*tg
,
7130 struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
,
7133 tg
->rt_rq
[cpu
] = rt_rq
;
7134 init_rt_rq(rt_rq
, rq
);
7136 rt_rq
->rt_se
= rt_se
;
7138 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7140 tg
->rt_se
[cpu
] = rt_se
;
7141 rt_se
->rt_rq
= &rq
->rt
;
7142 rt_se
->my_q
= rt_rq
;
7143 rt_se
->parent
= NULL
;
7144 INIT_LIST_HEAD(&rt_se
->run_list
);
7148 void __init
sched_init(void)
7150 int highest_cpu
= 0;
7154 init_defrootdomain();
7157 #ifdef CONFIG_GROUP_SCHED
7158 list_add(&init_task_group
.list
, &task_groups
);
7161 for_each_possible_cpu(i
) {
7165 spin_lock_init(&rq
->lock
);
7166 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7169 init_cfs_rq(&rq
->cfs
, rq
);
7170 init_rt_rq(&rq
->rt
, rq
);
7171 #ifdef CONFIG_FAIR_GROUP_SCHED
7172 init_task_group
.shares
= init_task_group_load
;
7173 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7174 init_tg_cfs_entry(rq
, &init_task_group
,
7175 &per_cpu(init_cfs_rq
, i
),
7176 &per_cpu(init_sched_entity
, i
), i
, 1);
7179 #ifdef CONFIG_RT_GROUP_SCHED
7180 init_task_group
.rt_runtime
=
7181 sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
7182 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7183 init_tg_rt_entry(rq
, &init_task_group
,
7184 &per_cpu(init_rt_rq
, i
),
7185 &per_cpu(init_sched_rt_entity
, i
), i
, 1);
7187 rq
->rt_period_expire
= 0;
7188 rq
->rt_throttled
= 0;
7190 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7191 rq
->cpu_load
[j
] = 0;
7195 rq
->active_balance
= 0;
7196 rq
->next_balance
= jiffies
;
7199 rq
->migration_thread
= NULL
;
7200 INIT_LIST_HEAD(&rq
->migration_queue
);
7201 rq_attach_root(rq
, &def_root_domain
);
7204 atomic_set(&rq
->nr_iowait
, 0);
7208 set_load_weight(&init_task
);
7210 #ifdef CONFIG_PREEMPT_NOTIFIERS
7211 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7215 nr_cpu_ids
= highest_cpu
+ 1;
7216 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7219 #ifdef CONFIG_RT_MUTEXES
7220 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7224 * The boot idle thread does lazy MMU switching as well:
7226 atomic_inc(&init_mm
.mm_count
);
7227 enter_lazy_tlb(&init_mm
, current
);
7230 * Make us the idle thread. Technically, schedule() should not be
7231 * called from this thread, however somewhere below it might be,
7232 * but because we are the idle thread, we just pick up running again
7233 * when this runqueue becomes "idle".
7235 init_idle(current
, smp_processor_id());
7237 * During early bootup we pretend to be a normal task:
7239 current
->sched_class
= &fair_sched_class
;
7241 scheduler_running
= 1;
7244 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7245 void __might_sleep(char *file
, int line
)
7248 static unsigned long prev_jiffy
; /* ratelimiting */
7250 if ((in_atomic() || irqs_disabled()) &&
7251 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7252 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7254 prev_jiffy
= jiffies
;
7255 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7256 " context at %s:%d\n", file
, line
);
7257 printk("in_atomic():%d, irqs_disabled():%d\n",
7258 in_atomic(), irqs_disabled());
7259 debug_show_held_locks(current
);
7260 if (irqs_disabled())
7261 print_irqtrace_events(current
);
7266 EXPORT_SYMBOL(__might_sleep
);
7269 #ifdef CONFIG_MAGIC_SYSRQ
7270 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7273 update_rq_clock(rq
);
7274 on_rq
= p
->se
.on_rq
;
7276 deactivate_task(rq
, p
, 0);
7277 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7279 activate_task(rq
, p
, 0);
7280 resched_task(rq
->curr
);
7284 void normalize_rt_tasks(void)
7286 struct task_struct
*g
, *p
;
7287 unsigned long flags
;
7290 read_lock_irqsave(&tasklist_lock
, flags
);
7291 do_each_thread(g
, p
) {
7293 * Only normalize user tasks:
7298 p
->se
.exec_start
= 0;
7299 #ifdef CONFIG_SCHEDSTATS
7300 p
->se
.wait_start
= 0;
7301 p
->se
.sleep_start
= 0;
7302 p
->se
.block_start
= 0;
7304 task_rq(p
)->clock
= 0;
7308 * Renice negative nice level userspace
7311 if (TASK_NICE(p
) < 0 && p
->mm
)
7312 set_user_nice(p
, 0);
7316 spin_lock(&p
->pi_lock
);
7317 rq
= __task_rq_lock(p
);
7319 normalize_task(rq
, p
);
7321 __task_rq_unlock(rq
);
7322 spin_unlock(&p
->pi_lock
);
7323 } while_each_thread(g
, p
);
7325 read_unlock_irqrestore(&tasklist_lock
, flags
);
7328 #endif /* CONFIG_MAGIC_SYSRQ */
7332 * These functions are only useful for the IA64 MCA handling.
7334 * They can only be called when the whole system has been
7335 * stopped - every CPU needs to be quiescent, and no scheduling
7336 * activity can take place. Using them for anything else would
7337 * be a serious bug, and as a result, they aren't even visible
7338 * under any other configuration.
7342 * curr_task - return the current task for a given cpu.
7343 * @cpu: the processor in question.
7345 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7347 struct task_struct
*curr_task(int cpu
)
7349 return cpu_curr(cpu
);
7353 * set_curr_task - set the current task for a given cpu.
7354 * @cpu: the processor in question.
7355 * @p: the task pointer to set.
7357 * Description: This function must only be used when non-maskable interrupts
7358 * are serviced on a separate stack. It allows the architecture to switch the
7359 * notion of the current task on a cpu in a non-blocking manner. This function
7360 * must be called with all CPU's synchronized, and interrupts disabled, the
7361 * and caller must save the original value of the current task (see
7362 * curr_task() above) and restore that value before reenabling interrupts and
7363 * re-starting the system.
7365 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7367 void set_curr_task(int cpu
, struct task_struct
*p
)
7374 #ifdef CONFIG_GROUP_SCHED
7376 #ifdef CONFIG_FAIR_GROUP_SCHED
7377 static void free_fair_sched_group(struct task_group
*tg
)
7381 for_each_possible_cpu(i
) {
7383 kfree(tg
->cfs_rq
[i
]);
7392 static int alloc_fair_sched_group(struct task_group
*tg
)
7394 struct cfs_rq
*cfs_rq
;
7395 struct sched_entity
*se
;
7399 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7402 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7406 tg
->shares
= NICE_0_LOAD
;
7408 for_each_possible_cpu(i
) {
7411 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
7412 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7416 se
= kmalloc_node(sizeof(struct sched_entity
),
7417 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7421 init_tg_cfs_entry(rq
, tg
, cfs_rq
, se
, i
, 0);
7430 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7432 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
7433 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
7436 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7438 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
7441 static inline void free_fair_sched_group(struct task_group
*tg
)
7445 static inline int alloc_fair_sched_group(struct task_group
*tg
)
7450 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7454 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7459 #ifdef CONFIG_RT_GROUP_SCHED
7460 static void free_rt_sched_group(struct task_group
*tg
)
7464 for_each_possible_cpu(i
) {
7466 kfree(tg
->rt_rq
[i
]);
7468 kfree(tg
->rt_se
[i
]);
7475 static int alloc_rt_sched_group(struct task_group
*tg
)
7477 struct rt_rq
*rt_rq
;
7478 struct sched_rt_entity
*rt_se
;
7482 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * NR_CPUS
, GFP_KERNEL
);
7485 tg
->rt_se
= kzalloc(sizeof(rt_se
) * NR_CPUS
, GFP_KERNEL
);
7491 for_each_possible_cpu(i
) {
7494 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
7495 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7499 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
7500 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7504 init_tg_rt_entry(rq
, tg
, rt_rq
, rt_se
, i
, 0);
7513 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7515 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
7516 &cpu_rq(cpu
)->leaf_rt_rq_list
);
7519 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7521 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
7524 static inline void free_rt_sched_group(struct task_group
*tg
)
7528 static inline int alloc_rt_sched_group(struct task_group
*tg
)
7533 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7537 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7542 static void free_sched_group(struct task_group
*tg
)
7544 free_fair_sched_group(tg
);
7545 free_rt_sched_group(tg
);
7549 /* allocate runqueue etc for a new task group */
7550 struct task_group
*sched_create_group(void)
7552 struct task_group
*tg
;
7553 unsigned long flags
;
7556 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7558 return ERR_PTR(-ENOMEM
);
7560 if (!alloc_fair_sched_group(tg
))
7563 if (!alloc_rt_sched_group(tg
))
7566 spin_lock_irqsave(&task_group_lock
, flags
);
7567 for_each_possible_cpu(i
) {
7568 register_fair_sched_group(tg
, i
);
7569 register_rt_sched_group(tg
, i
);
7571 list_add_rcu(&tg
->list
, &task_groups
);
7572 spin_unlock_irqrestore(&task_group_lock
, flags
);
7577 free_sched_group(tg
);
7578 return ERR_PTR(-ENOMEM
);
7581 /* rcu callback to free various structures associated with a task group */
7582 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7584 /* now it should be safe to free those cfs_rqs */
7585 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7588 /* Destroy runqueue etc associated with a task group */
7589 void sched_destroy_group(struct task_group
*tg
)
7591 unsigned long flags
;
7594 spin_lock_irqsave(&task_group_lock
, flags
);
7595 for_each_possible_cpu(i
) {
7596 unregister_fair_sched_group(tg
, i
);
7597 unregister_rt_sched_group(tg
, i
);
7599 list_del_rcu(&tg
->list
);
7600 spin_unlock_irqrestore(&task_group_lock
, flags
);
7602 /* wait for possible concurrent references to cfs_rqs complete */
7603 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7606 /* change task's runqueue when it moves between groups.
7607 * The caller of this function should have put the task in its new group
7608 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7609 * reflect its new group.
7611 void sched_move_task(struct task_struct
*tsk
)
7614 unsigned long flags
;
7617 rq
= task_rq_lock(tsk
, &flags
);
7619 update_rq_clock(rq
);
7621 running
= task_current(rq
, tsk
);
7622 on_rq
= tsk
->se
.on_rq
;
7625 dequeue_task(rq
, tsk
, 0);
7626 if (unlikely(running
))
7627 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7629 set_task_rq(tsk
, task_cpu(tsk
));
7631 #ifdef CONFIG_FAIR_GROUP_SCHED
7632 if (tsk
->sched_class
->moved_group
)
7633 tsk
->sched_class
->moved_group(tsk
);
7636 if (unlikely(running
))
7637 tsk
->sched_class
->set_curr_task(rq
);
7639 enqueue_task(rq
, tsk
, 0);
7641 task_rq_unlock(rq
, &flags
);
7644 #ifdef CONFIG_FAIR_GROUP_SCHED
7645 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7647 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7648 struct rq
*rq
= cfs_rq
->rq
;
7651 spin_lock_irq(&rq
->lock
);
7655 dequeue_entity(cfs_rq
, se
, 0);
7657 se
->load
.weight
= shares
;
7658 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7661 enqueue_entity(cfs_rq
, se
, 0);
7663 spin_unlock_irq(&rq
->lock
);
7666 static DEFINE_MUTEX(shares_mutex
);
7668 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7671 unsigned long flags
;
7674 * A weight of 0 or 1 can cause arithmetics problems.
7675 * (The default weight is 1024 - so there's no practical
7676 * limitation from this.)
7681 mutex_lock(&shares_mutex
);
7682 if (tg
->shares
== shares
)
7685 spin_lock_irqsave(&task_group_lock
, flags
);
7686 for_each_possible_cpu(i
)
7687 unregister_fair_sched_group(tg
, i
);
7688 spin_unlock_irqrestore(&task_group_lock
, flags
);
7690 /* wait for any ongoing reference to this group to finish */
7691 synchronize_sched();
7694 * Now we are free to modify the group's share on each cpu
7695 * w/o tripping rebalance_share or load_balance_fair.
7697 tg
->shares
= shares
;
7698 for_each_possible_cpu(i
)
7699 set_se_shares(tg
->se
[i
], shares
);
7702 * Enable load balance activity on this group, by inserting it back on
7703 * each cpu's rq->leaf_cfs_rq_list.
7705 spin_lock_irqsave(&task_group_lock
, flags
);
7706 for_each_possible_cpu(i
)
7707 register_fair_sched_group(tg
, i
);
7708 spin_unlock_irqrestore(&task_group_lock
, flags
);
7710 mutex_unlock(&shares_mutex
);
7714 unsigned long sched_group_shares(struct task_group
*tg
)
7720 #ifdef CONFIG_RT_GROUP_SCHED
7722 * Ensure that the real time constraints are schedulable.
7724 static DEFINE_MUTEX(rt_constraints_mutex
);
7726 static unsigned long to_ratio(u64 period
, u64 runtime
)
7728 if (runtime
== RUNTIME_INF
)
7731 return div64_64(runtime
<< 16, period
);
7734 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7736 struct task_group
*tgi
;
7737 unsigned long total
= 0;
7738 unsigned long global_ratio
=
7739 to_ratio(sysctl_sched_rt_period
,
7740 sysctl_sched_rt_runtime
< 0 ?
7741 RUNTIME_INF
: sysctl_sched_rt_runtime
);
7744 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
7748 total
+= to_ratio(period
, tgi
->rt_runtime
);
7752 return total
+ to_ratio(period
, runtime
) < global_ratio
;
7755 /* Must be called with tasklist_lock held */
7756 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7758 struct task_struct
*g
, *p
;
7759 do_each_thread(g
, p
) {
7760 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
7762 } while_each_thread(g
, p
);
7766 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7768 u64 rt_runtime
, rt_period
;
7771 rt_period
= (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
7772 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7773 if (rt_runtime_us
== -1)
7774 rt_runtime
= RUNTIME_INF
;
7776 mutex_lock(&rt_constraints_mutex
);
7777 read_lock(&tasklist_lock
);
7778 if (rt_runtime_us
== 0 && tg_has_rt_tasks(tg
)) {
7782 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
7786 tg
->rt_runtime
= rt_runtime
;
7788 read_unlock(&tasklist_lock
);
7789 mutex_unlock(&rt_constraints_mutex
);
7794 long sched_group_rt_runtime(struct task_group
*tg
)
7798 if (tg
->rt_runtime
== RUNTIME_INF
)
7801 rt_runtime_us
= tg
->rt_runtime
;
7802 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7803 return rt_runtime_us
;
7806 #endif /* CONFIG_GROUP_SCHED */
7808 #ifdef CONFIG_CGROUP_SCHED
7810 /* return corresponding task_group object of a cgroup */
7811 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7813 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7814 struct task_group
, css
);
7817 static struct cgroup_subsys_state
*
7818 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7820 struct task_group
*tg
;
7822 if (!cgrp
->parent
) {
7823 /* This is early initialization for the top cgroup */
7824 init_task_group
.css
.cgroup
= cgrp
;
7825 return &init_task_group
.css
;
7828 /* we support only 1-level deep hierarchical scheduler atm */
7829 if (cgrp
->parent
->parent
)
7830 return ERR_PTR(-EINVAL
);
7832 tg
= sched_create_group();
7834 return ERR_PTR(-ENOMEM
);
7836 /* Bind the cgroup to task_group object we just created */
7837 tg
->css
.cgroup
= cgrp
;
7843 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7845 struct task_group
*tg
= cgroup_tg(cgrp
);
7847 sched_destroy_group(tg
);
7851 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7852 struct task_struct
*tsk
)
7854 #ifdef CONFIG_RT_GROUP_SCHED
7855 /* Don't accept realtime tasks when there is no way for them to run */
7856 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_runtime
== 0)
7859 /* We don't support RT-tasks being in separate groups */
7860 if (tsk
->sched_class
!= &fair_sched_class
)
7868 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7869 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7871 sched_move_task(tsk
);
7874 #ifdef CONFIG_FAIR_GROUP_SCHED
7875 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7878 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7881 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7883 struct task_group
*tg
= cgroup_tg(cgrp
);
7885 return (u64
) tg
->shares
;
7889 #ifdef CONFIG_RT_GROUP_SCHED
7890 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7892 const char __user
*userbuf
,
7893 size_t nbytes
, loff_t
*unused_ppos
)
7902 if (nbytes
>= sizeof(buffer
))
7904 if (copy_from_user(buffer
, userbuf
, nbytes
))
7907 buffer
[nbytes
] = 0; /* nul-terminate */
7909 /* strip newline if necessary */
7910 if (nbytes
&& (buffer
[nbytes
-1] == '\n'))
7911 buffer
[nbytes
-1] = 0;
7912 val
= simple_strtoll(buffer
, &end
, 0);
7916 /* Pass to subsystem */
7917 retval
= sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7923 static ssize_t
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
,
7925 char __user
*buf
, size_t nbytes
,
7929 long val
= sched_group_rt_runtime(cgroup_tg(cgrp
));
7930 int len
= sprintf(tmp
, "%ld\n", val
);
7932 return simple_read_from_buffer(buf
, nbytes
, ppos
, tmp
, len
);
7936 static struct cftype cpu_files
[] = {
7937 #ifdef CONFIG_FAIR_GROUP_SCHED
7940 .read_uint
= cpu_shares_read_uint
,
7941 .write_uint
= cpu_shares_write_uint
,
7944 #ifdef CONFIG_RT_GROUP_SCHED
7946 .name
= "rt_runtime_us",
7947 .read
= cpu_rt_runtime_read
,
7948 .write
= cpu_rt_runtime_write
,
7953 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7955 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7958 struct cgroup_subsys cpu_cgroup_subsys
= {
7960 .create
= cpu_cgroup_create
,
7961 .destroy
= cpu_cgroup_destroy
,
7962 .can_attach
= cpu_cgroup_can_attach
,
7963 .attach
= cpu_cgroup_attach
,
7964 .populate
= cpu_cgroup_populate
,
7965 .subsys_id
= cpu_cgroup_subsys_id
,
7969 #endif /* CONFIG_CGROUP_SCHED */
7971 #ifdef CONFIG_CGROUP_CPUACCT
7974 * CPU accounting code for task groups.
7976 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7977 * (balbir@in.ibm.com).
7980 /* track cpu usage of a group of tasks */
7982 struct cgroup_subsys_state css
;
7983 /* cpuusage holds pointer to a u64-type object on every cpu */
7987 struct cgroup_subsys cpuacct_subsys
;
7989 /* return cpu accounting group corresponding to this container */
7990 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7992 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7993 struct cpuacct
, css
);
7996 /* return cpu accounting group to which this task belongs */
7997 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7999 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8000 struct cpuacct
, css
);
8003 /* create a new cpu accounting group */
8004 static struct cgroup_subsys_state
*cpuacct_create(
8005 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8007 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8010 return ERR_PTR(-ENOMEM
);
8012 ca
->cpuusage
= alloc_percpu(u64
);
8013 if (!ca
->cpuusage
) {
8015 return ERR_PTR(-ENOMEM
);
8021 /* destroy an existing cpu accounting group */
8023 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8025 struct cpuacct
*ca
= cgroup_ca(cont
);
8027 free_percpu(ca
->cpuusage
);
8031 /* return total cpu usage (in nanoseconds) of a group */
8032 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
8034 struct cpuacct
*ca
= cgroup_ca(cont
);
8035 u64 totalcpuusage
= 0;
8038 for_each_possible_cpu(i
) {
8039 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8042 * Take rq->lock to make 64-bit addition safe on 32-bit
8045 spin_lock_irq(&cpu_rq(i
)->lock
);
8046 totalcpuusage
+= *cpuusage
;
8047 spin_unlock_irq(&cpu_rq(i
)->lock
);
8050 return totalcpuusage
;
8053 static struct cftype files
[] = {
8056 .read_uint
= cpuusage_read
,
8060 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8062 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
8066 * charge this task's execution time to its accounting group.
8068 * called with rq->lock held.
8070 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8074 if (!cpuacct_subsys
.active
)
8079 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
8081 *cpuusage
+= cputime
;
8085 struct cgroup_subsys cpuacct_subsys
= {
8087 .create
= cpuacct_create
,
8088 .destroy
= cpuacct_destroy
,
8089 .populate
= cpuacct_populate
,
8090 .subsys_id
= cpuacct_subsys_id
,
8092 #endif /* CONFIG_CGROUP_CPUACCT */