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 */
879 HRTICK_BLOCK
, /* stop hrtick operations */
884 * - enabled by features
885 * - hrtimer is actually high res
887 static inline int hrtick_enabled(struct rq
*rq
)
889 if (!sched_feat(HRTICK
))
891 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
893 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
897 * Called to set the hrtick timer state.
899 * called with rq->lock held and irqs disabled
901 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
903 assert_spin_locked(&rq
->lock
);
906 * preempt at: now + delay
909 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
911 * indicate we need to program the timer
913 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
915 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
918 * New slices are called from the schedule path and don't need a
922 resched_hrt(rq
->curr
);
925 static void hrtick_clear(struct rq
*rq
)
927 if (hrtimer_active(&rq
->hrtick_timer
))
928 hrtimer_cancel(&rq
->hrtick_timer
);
932 * Update the timer from the possible pending state.
934 static void hrtick_set(struct rq
*rq
)
940 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
942 spin_lock_irqsave(&rq
->lock
, flags
);
943 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
944 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
945 time
= rq
->hrtick_expire
;
946 clear_thread_flag(TIF_HRTICK_RESCHED
);
947 spin_unlock_irqrestore(&rq
->lock
, flags
);
950 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
951 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
958 * High-resolution timer tick.
959 * Runs from hardirq context with interrupts disabled.
961 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
963 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
965 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
967 spin_lock(&rq
->lock
);
968 __update_rq_clock(rq
);
969 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
970 spin_unlock(&rq
->lock
);
972 return HRTIMER_NORESTART
;
975 static void hotplug_hrtick_disable(int cpu
)
977 struct rq
*rq
= cpu_rq(cpu
);
980 spin_lock_irqsave(&rq
->lock
, flags
);
981 rq
->hrtick_flags
= 0;
982 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
983 spin_unlock_irqrestore(&rq
->lock
, flags
);
988 static void hotplug_hrtick_enable(int cpu
)
990 struct rq
*rq
= cpu_rq(cpu
);
993 spin_lock_irqsave(&rq
->lock
, flags
);
994 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
995 spin_unlock_irqrestore(&rq
->lock
, flags
);
999 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1001 int cpu
= (int)(long)hcpu
;
1004 case CPU_UP_CANCELED
:
1005 case CPU_UP_CANCELED_FROZEN
:
1006 case CPU_DOWN_PREPARE
:
1007 case CPU_DOWN_PREPARE_FROZEN
:
1009 case CPU_DEAD_FROZEN
:
1010 hotplug_hrtick_disable(cpu
);
1013 case CPU_UP_PREPARE
:
1014 case CPU_UP_PREPARE_FROZEN
:
1015 case CPU_DOWN_FAILED
:
1016 case CPU_DOWN_FAILED_FROZEN
:
1018 case CPU_ONLINE_FROZEN
:
1019 hotplug_hrtick_enable(cpu
);
1026 static void init_hrtick(void)
1028 hotcpu_notifier(hotplug_hrtick
, 0);
1031 static void init_rq_hrtick(struct rq
*rq
)
1033 rq
->hrtick_flags
= 0;
1034 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1035 rq
->hrtick_timer
.function
= hrtick
;
1036 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1039 void hrtick_resched(void)
1042 unsigned long flags
;
1044 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1047 local_irq_save(flags
);
1048 rq
= cpu_rq(smp_processor_id());
1050 local_irq_restore(flags
);
1053 static inline void hrtick_clear(struct rq
*rq
)
1057 static inline void hrtick_set(struct rq
*rq
)
1061 static inline void init_rq_hrtick(struct rq
*rq
)
1065 void hrtick_resched(void)
1069 static inline void init_hrtick(void)
1075 * resched_task - mark a task 'to be rescheduled now'.
1077 * On UP this means the setting of the need_resched flag, on SMP it
1078 * might also involve a cross-CPU call to trigger the scheduler on
1083 #ifndef tsk_is_polling
1084 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1087 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1091 assert_spin_locked(&task_rq(p
)->lock
);
1093 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1096 set_tsk_thread_flag(p
, tif_bit
);
1099 if (cpu
== smp_processor_id())
1102 /* NEED_RESCHED must be visible before we test polling */
1104 if (!tsk_is_polling(p
))
1105 smp_send_reschedule(cpu
);
1108 static void resched_cpu(int cpu
)
1110 struct rq
*rq
= cpu_rq(cpu
);
1111 unsigned long flags
;
1113 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1115 resched_task(cpu_curr(cpu
));
1116 spin_unlock_irqrestore(&rq
->lock
, flags
);
1121 * When add_timer_on() enqueues a timer into the timer wheel of an
1122 * idle CPU then this timer might expire before the next timer event
1123 * which is scheduled to wake up that CPU. In case of a completely
1124 * idle system the next event might even be infinite time into the
1125 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1126 * leaves the inner idle loop so the newly added timer is taken into
1127 * account when the CPU goes back to idle and evaluates the timer
1128 * wheel for the next timer event.
1130 void wake_up_idle_cpu(int cpu
)
1132 struct rq
*rq
= cpu_rq(cpu
);
1134 if (cpu
== smp_processor_id())
1138 * This is safe, as this function is called with the timer
1139 * wheel base lock of (cpu) held. When the CPU is on the way
1140 * to idle and has not yet set rq->curr to idle then it will
1141 * be serialized on the timer wheel base lock and take the new
1142 * timer into account automatically.
1144 if (rq
->curr
!= rq
->idle
)
1148 * We can set TIF_RESCHED on the idle task of the other CPU
1149 * lockless. The worst case is that the other CPU runs the
1150 * idle task through an additional NOOP schedule()
1152 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1154 /* NEED_RESCHED must be visible before we test polling */
1156 if (!tsk_is_polling(rq
->idle
))
1157 smp_send_reschedule(cpu
);
1162 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1164 assert_spin_locked(&task_rq(p
)->lock
);
1165 set_tsk_thread_flag(p
, tif_bit
);
1169 #if BITS_PER_LONG == 32
1170 # define WMULT_CONST (~0UL)
1172 # define WMULT_CONST (1UL << 32)
1175 #define WMULT_SHIFT 32
1178 * Shift right and round:
1180 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1182 static unsigned long
1183 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1184 struct load_weight
*lw
)
1188 if (unlikely(!lw
->inv_weight
))
1189 lw
->inv_weight
= (WMULT_CONST
-lw
->weight
/2) / (lw
->weight
+1);
1191 tmp
= (u64
)delta_exec
* weight
;
1193 * Check whether we'd overflow the 64-bit multiplication:
1195 if (unlikely(tmp
> WMULT_CONST
))
1196 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1199 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1201 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1204 static inline unsigned long
1205 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1207 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1210 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1216 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1223 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1224 * of tasks with abnormal "nice" values across CPUs the contribution that
1225 * each task makes to its run queue's load is weighted according to its
1226 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1227 * scaled version of the new time slice allocation that they receive on time
1231 #define WEIGHT_IDLEPRIO 2
1232 #define WMULT_IDLEPRIO (1 << 31)
1235 * Nice levels are multiplicative, with a gentle 10% change for every
1236 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1237 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1238 * that remained on nice 0.
1240 * The "10% effect" is relative and cumulative: from _any_ nice level,
1241 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1242 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1243 * If a task goes up by ~10% and another task goes down by ~10% then
1244 * the relative distance between them is ~25%.)
1246 static const int prio_to_weight
[40] = {
1247 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1248 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1249 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1250 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1251 /* 0 */ 1024, 820, 655, 526, 423,
1252 /* 5 */ 335, 272, 215, 172, 137,
1253 /* 10 */ 110, 87, 70, 56, 45,
1254 /* 15 */ 36, 29, 23, 18, 15,
1258 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1260 * In cases where the weight does not change often, we can use the
1261 * precalculated inverse to speed up arithmetics by turning divisions
1262 * into multiplications:
1264 static const u32 prio_to_wmult
[40] = {
1265 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1266 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1267 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1268 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1269 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1270 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1271 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1272 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1275 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1278 * runqueue iterator, to support SMP load-balancing between different
1279 * scheduling classes, without having to expose their internal data
1280 * structures to the load-balancing proper:
1282 struct rq_iterator
{
1284 struct task_struct
*(*start
)(void *);
1285 struct task_struct
*(*next
)(void *);
1289 static unsigned long
1290 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1291 unsigned long max_load_move
, struct sched_domain
*sd
,
1292 enum cpu_idle_type idle
, int *all_pinned
,
1293 int *this_best_prio
, struct rq_iterator
*iterator
);
1296 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1297 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1298 struct rq_iterator
*iterator
);
1301 #ifdef CONFIG_CGROUP_CPUACCT
1302 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1304 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1308 static unsigned long source_load(int cpu
, int type
);
1309 static unsigned long target_load(int cpu
, int type
);
1310 static unsigned long cpu_avg_load_per_task(int cpu
);
1311 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1312 #endif /* CONFIG_SMP */
1314 #include "sched_stats.h"
1315 #include "sched_idletask.c"
1316 #include "sched_fair.c"
1317 #include "sched_rt.c"
1318 #ifdef CONFIG_SCHED_DEBUG
1319 # include "sched_debug.c"
1322 #define sched_class_highest (&rt_sched_class)
1324 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
1326 update_load_add(&rq
->load
, p
->se
.load
.weight
);
1329 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
1331 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
1334 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1340 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1346 static void set_load_weight(struct task_struct
*p
)
1348 if (task_has_rt_policy(p
)) {
1349 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1350 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1355 * SCHED_IDLE tasks get minimal weight:
1357 if (p
->policy
== SCHED_IDLE
) {
1358 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1359 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1363 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1364 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1367 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1369 sched_info_queued(p
);
1370 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1374 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1376 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1381 * __normal_prio - return the priority that is based on the static prio
1383 static inline int __normal_prio(struct task_struct
*p
)
1385 return p
->static_prio
;
1389 * Calculate the expected normal priority: i.e. priority
1390 * without taking RT-inheritance into account. Might be
1391 * boosted by interactivity modifiers. Changes upon fork,
1392 * setprio syscalls, and whenever the interactivity
1393 * estimator recalculates.
1395 static inline int normal_prio(struct task_struct
*p
)
1399 if (task_has_rt_policy(p
))
1400 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1402 prio
= __normal_prio(p
);
1407 * Calculate the current priority, i.e. the priority
1408 * taken into account by the scheduler. This value might
1409 * be boosted by RT tasks, or might be boosted by
1410 * interactivity modifiers. Will be RT if the task got
1411 * RT-boosted. If not then it returns p->normal_prio.
1413 static int effective_prio(struct task_struct
*p
)
1415 p
->normal_prio
= normal_prio(p
);
1417 * If we are RT tasks or we were boosted to RT priority,
1418 * keep the priority unchanged. Otherwise, update priority
1419 * to the normal priority:
1421 if (!rt_prio(p
->prio
))
1422 return p
->normal_prio
;
1427 * activate_task - move a task to the runqueue.
1429 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1431 if (task_contributes_to_load(p
))
1432 rq
->nr_uninterruptible
--;
1434 enqueue_task(rq
, p
, wakeup
);
1435 inc_nr_running(p
, rq
);
1439 * deactivate_task - remove a task from the runqueue.
1441 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1443 if (task_contributes_to_load(p
))
1444 rq
->nr_uninterruptible
++;
1446 dequeue_task(rq
, p
, sleep
);
1447 dec_nr_running(p
, rq
);
1451 * task_curr - is this task currently executing on a CPU?
1452 * @p: the task in question.
1454 inline int task_curr(const struct task_struct
*p
)
1456 return cpu_curr(task_cpu(p
)) == p
;
1459 /* Used instead of source_load when we know the type == 0 */
1460 unsigned long weighted_cpuload(const int cpu
)
1462 return cpu_rq(cpu
)->load
.weight
;
1465 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1467 set_task_rq(p
, cpu
);
1470 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1471 * successfuly executed on another CPU. We must ensure that updates of
1472 * per-task data have been completed by this moment.
1475 task_thread_info(p
)->cpu
= cpu
;
1479 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1480 const struct sched_class
*prev_class
,
1481 int oldprio
, int running
)
1483 if (prev_class
!= p
->sched_class
) {
1484 if (prev_class
->switched_from
)
1485 prev_class
->switched_from(rq
, p
, running
);
1486 p
->sched_class
->switched_to(rq
, p
, running
);
1488 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1494 * Is this task likely cache-hot:
1497 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1502 * Buddy candidates are cache hot:
1504 if (&p
->se
== cfs_rq_of(&p
->se
)->next
)
1507 if (p
->sched_class
!= &fair_sched_class
)
1510 if (sysctl_sched_migration_cost
== -1)
1512 if (sysctl_sched_migration_cost
== 0)
1515 delta
= now
- p
->se
.exec_start
;
1517 return delta
< (s64
)sysctl_sched_migration_cost
;
1521 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1523 int old_cpu
= task_cpu(p
);
1524 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1525 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1526 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1529 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1531 #ifdef CONFIG_SCHEDSTATS
1532 if (p
->se
.wait_start
)
1533 p
->se
.wait_start
-= clock_offset
;
1534 if (p
->se
.sleep_start
)
1535 p
->se
.sleep_start
-= clock_offset
;
1536 if (p
->se
.block_start
)
1537 p
->se
.block_start
-= clock_offset
;
1538 if (old_cpu
!= new_cpu
) {
1539 schedstat_inc(p
, se
.nr_migrations
);
1540 if (task_hot(p
, old_rq
->clock
, NULL
))
1541 schedstat_inc(p
, se
.nr_forced2_migrations
);
1544 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1545 new_cfsrq
->min_vruntime
;
1547 __set_task_cpu(p
, new_cpu
);
1550 struct migration_req
{
1551 struct list_head list
;
1553 struct task_struct
*task
;
1556 struct completion done
;
1560 * The task's runqueue lock must be held.
1561 * Returns true if you have to wait for migration thread.
1564 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1566 struct rq
*rq
= task_rq(p
);
1569 * If the task is not on a runqueue (and not running), then
1570 * it is sufficient to simply update the task's cpu field.
1572 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1573 set_task_cpu(p
, dest_cpu
);
1577 init_completion(&req
->done
);
1579 req
->dest_cpu
= dest_cpu
;
1580 list_add(&req
->list
, &rq
->migration_queue
);
1586 * wait_task_inactive - wait for a thread to unschedule.
1588 * The caller must ensure that the task *will* unschedule sometime soon,
1589 * else this function might spin for a *long* time. This function can't
1590 * be called with interrupts off, or it may introduce deadlock with
1591 * smp_call_function() if an IPI is sent by the same process we are
1592 * waiting to become inactive.
1594 void wait_task_inactive(struct task_struct
*p
)
1596 unsigned long flags
;
1602 * We do the initial early heuristics without holding
1603 * any task-queue locks at all. We'll only try to get
1604 * the runqueue lock when things look like they will
1610 * If the task is actively running on another CPU
1611 * still, just relax and busy-wait without holding
1614 * NOTE! Since we don't hold any locks, it's not
1615 * even sure that "rq" stays as the right runqueue!
1616 * But we don't care, since "task_running()" will
1617 * return false if the runqueue has changed and p
1618 * is actually now running somewhere else!
1620 while (task_running(rq
, p
))
1624 * Ok, time to look more closely! We need the rq
1625 * lock now, to be *sure*. If we're wrong, we'll
1626 * just go back and repeat.
1628 rq
= task_rq_lock(p
, &flags
);
1629 running
= task_running(rq
, p
);
1630 on_rq
= p
->se
.on_rq
;
1631 task_rq_unlock(rq
, &flags
);
1634 * Was it really running after all now that we
1635 * checked with the proper locks actually held?
1637 * Oops. Go back and try again..
1639 if (unlikely(running
)) {
1645 * It's not enough that it's not actively running,
1646 * it must be off the runqueue _entirely_, and not
1649 * So if it wa still runnable (but just not actively
1650 * running right now), it's preempted, and we should
1651 * yield - it could be a while.
1653 if (unlikely(on_rq
)) {
1654 schedule_timeout_uninterruptible(1);
1659 * Ahh, all good. It wasn't running, and it wasn't
1660 * runnable, which means that it will never become
1661 * running in the future either. We're all done!
1668 * kick_process - kick a running thread to enter/exit the kernel
1669 * @p: the to-be-kicked thread
1671 * Cause a process which is running on another CPU to enter
1672 * kernel-mode, without any delay. (to get signals handled.)
1674 * NOTE: this function doesnt have to take the runqueue lock,
1675 * because all it wants to ensure is that the remote task enters
1676 * the kernel. If the IPI races and the task has been migrated
1677 * to another CPU then no harm is done and the purpose has been
1680 void kick_process(struct task_struct
*p
)
1686 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1687 smp_send_reschedule(cpu
);
1692 * Return a low guess at the load of a migration-source cpu weighted
1693 * according to the scheduling class and "nice" value.
1695 * We want to under-estimate the load of migration sources, to
1696 * balance conservatively.
1698 static unsigned long source_load(int cpu
, int type
)
1700 struct rq
*rq
= cpu_rq(cpu
);
1701 unsigned long total
= weighted_cpuload(cpu
);
1706 return min(rq
->cpu_load
[type
-1], total
);
1710 * Return a high guess at the load of a migration-target cpu weighted
1711 * according to the scheduling class and "nice" value.
1713 static unsigned long target_load(int cpu
, int type
)
1715 struct rq
*rq
= cpu_rq(cpu
);
1716 unsigned long total
= weighted_cpuload(cpu
);
1721 return max(rq
->cpu_load
[type
-1], total
);
1725 * Return the average load per task on the cpu's run queue
1727 static unsigned long cpu_avg_load_per_task(int cpu
)
1729 struct rq
*rq
= cpu_rq(cpu
);
1730 unsigned long total
= weighted_cpuload(cpu
);
1731 unsigned long n
= rq
->nr_running
;
1733 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1737 * find_idlest_group finds and returns the least busy CPU group within the
1740 static struct sched_group
*
1741 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1743 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1744 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1745 int load_idx
= sd
->forkexec_idx
;
1746 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1749 unsigned long load
, avg_load
;
1753 /* Skip over this group if it has no CPUs allowed */
1754 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1757 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1759 /* Tally up the load of all CPUs in the group */
1762 for_each_cpu_mask(i
, group
->cpumask
) {
1763 /* Bias balancing toward cpus of our domain */
1765 load
= source_load(i
, load_idx
);
1767 load
= target_load(i
, load_idx
);
1772 /* Adjust by relative CPU power of the group */
1773 avg_load
= sg_div_cpu_power(group
,
1774 avg_load
* SCHED_LOAD_SCALE
);
1777 this_load
= avg_load
;
1779 } else if (avg_load
< min_load
) {
1780 min_load
= avg_load
;
1783 } while (group
= group
->next
, group
!= sd
->groups
);
1785 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1791 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1794 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1797 unsigned long load
, min_load
= ULONG_MAX
;
1801 /* Traverse only the allowed CPUs */
1802 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1804 for_each_cpu_mask(i
, tmp
) {
1805 load
= weighted_cpuload(i
);
1807 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1817 * sched_balance_self: balance the current task (running on cpu) in domains
1818 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1821 * Balance, ie. select the least loaded group.
1823 * Returns the target CPU number, or the same CPU if no balancing is needed.
1825 * preempt must be disabled.
1827 static int sched_balance_self(int cpu
, int flag
)
1829 struct task_struct
*t
= current
;
1830 struct sched_domain
*tmp
, *sd
= NULL
;
1832 for_each_domain(cpu
, tmp
) {
1834 * If power savings logic is enabled for a domain, stop there.
1836 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1838 if (tmp
->flags
& flag
)
1844 struct sched_group
*group
;
1845 int new_cpu
, weight
;
1847 if (!(sd
->flags
& flag
)) {
1853 group
= find_idlest_group(sd
, t
, cpu
);
1859 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1860 if (new_cpu
== -1 || new_cpu
== cpu
) {
1861 /* Now try balancing at a lower domain level of cpu */
1866 /* Now try balancing at a lower domain level of new_cpu */
1869 weight
= cpus_weight(span
);
1870 for_each_domain(cpu
, tmp
) {
1871 if (weight
<= cpus_weight(tmp
->span
))
1873 if (tmp
->flags
& flag
)
1876 /* while loop will break here if sd == NULL */
1882 #endif /* CONFIG_SMP */
1885 * try_to_wake_up - wake up a thread
1886 * @p: the to-be-woken-up thread
1887 * @state: the mask of task states that can be woken
1888 * @sync: do a synchronous wakeup?
1890 * Put it on the run-queue if it's not already there. The "current"
1891 * thread is always on the run-queue (except when the actual
1892 * re-schedule is in progress), and as such you're allowed to do
1893 * the simpler "current->state = TASK_RUNNING" to mark yourself
1894 * runnable without the overhead of this.
1896 * returns failure only if the task is already active.
1898 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1900 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1901 unsigned long flags
;
1906 rq
= task_rq_lock(p
, &flags
);
1907 old_state
= p
->state
;
1908 if (!(old_state
& state
))
1916 this_cpu
= smp_processor_id();
1919 if (unlikely(task_running(rq
, p
)))
1922 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1923 if (cpu
!= orig_cpu
) {
1924 set_task_cpu(p
, cpu
);
1925 task_rq_unlock(rq
, &flags
);
1926 /* might preempt at this point */
1927 rq
= task_rq_lock(p
, &flags
);
1928 old_state
= p
->state
;
1929 if (!(old_state
& state
))
1934 this_cpu
= smp_processor_id();
1938 #ifdef CONFIG_SCHEDSTATS
1939 schedstat_inc(rq
, ttwu_count
);
1940 if (cpu
== this_cpu
)
1941 schedstat_inc(rq
, ttwu_local
);
1943 struct sched_domain
*sd
;
1944 for_each_domain(this_cpu
, sd
) {
1945 if (cpu_isset(cpu
, sd
->span
)) {
1946 schedstat_inc(sd
, ttwu_wake_remote
);
1954 #endif /* CONFIG_SMP */
1955 schedstat_inc(p
, se
.nr_wakeups
);
1957 schedstat_inc(p
, se
.nr_wakeups_sync
);
1958 if (orig_cpu
!= cpu
)
1959 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1960 if (cpu
== this_cpu
)
1961 schedstat_inc(p
, se
.nr_wakeups_local
);
1963 schedstat_inc(p
, se
.nr_wakeups_remote
);
1964 update_rq_clock(rq
);
1965 activate_task(rq
, p
, 1);
1969 check_preempt_curr(rq
, p
);
1971 p
->state
= TASK_RUNNING
;
1973 if (p
->sched_class
->task_wake_up
)
1974 p
->sched_class
->task_wake_up(rq
, p
);
1977 task_rq_unlock(rq
, &flags
);
1982 int wake_up_process(struct task_struct
*p
)
1984 return try_to_wake_up(p
, TASK_ALL
, 0);
1986 EXPORT_SYMBOL(wake_up_process
);
1988 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1990 return try_to_wake_up(p
, state
, 0);
1994 * Perform scheduler related setup for a newly forked process p.
1995 * p is forked by current.
1997 * __sched_fork() is basic setup used by init_idle() too:
1999 static void __sched_fork(struct task_struct
*p
)
2001 p
->se
.exec_start
= 0;
2002 p
->se
.sum_exec_runtime
= 0;
2003 p
->se
.prev_sum_exec_runtime
= 0;
2004 p
->se
.last_wakeup
= 0;
2005 p
->se
.avg_overlap
= 0;
2007 #ifdef CONFIG_SCHEDSTATS
2008 p
->se
.wait_start
= 0;
2009 p
->se
.sum_sleep_runtime
= 0;
2010 p
->se
.sleep_start
= 0;
2011 p
->se
.block_start
= 0;
2012 p
->se
.sleep_max
= 0;
2013 p
->se
.block_max
= 0;
2015 p
->se
.slice_max
= 0;
2019 INIT_LIST_HEAD(&p
->rt
.run_list
);
2022 #ifdef CONFIG_PREEMPT_NOTIFIERS
2023 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2027 * We mark the process as running here, but have not actually
2028 * inserted it onto the runqueue yet. This guarantees that
2029 * nobody will actually run it, and a signal or other external
2030 * event cannot wake it up and insert it on the runqueue either.
2032 p
->state
= TASK_RUNNING
;
2036 * fork()/clone()-time setup:
2038 void sched_fork(struct task_struct
*p
, int clone_flags
)
2040 int cpu
= get_cpu();
2045 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2047 set_task_cpu(p
, cpu
);
2050 * Make sure we do not leak PI boosting priority to the child:
2052 p
->prio
= current
->normal_prio
;
2053 if (!rt_prio(p
->prio
))
2054 p
->sched_class
= &fair_sched_class
;
2056 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2057 if (likely(sched_info_on()))
2058 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2060 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2063 #ifdef CONFIG_PREEMPT
2064 /* Want to start with kernel preemption disabled. */
2065 task_thread_info(p
)->preempt_count
= 1;
2071 * wake_up_new_task - wake up a newly created task for the first time.
2073 * This function will do some initial scheduler statistics housekeeping
2074 * that must be done for every newly created context, then puts the task
2075 * on the runqueue and wakes it.
2077 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2079 unsigned long flags
;
2082 rq
= task_rq_lock(p
, &flags
);
2083 BUG_ON(p
->state
!= TASK_RUNNING
);
2084 update_rq_clock(rq
);
2086 p
->prio
= effective_prio(p
);
2088 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2089 activate_task(rq
, p
, 0);
2092 * Let the scheduling class do new task startup
2093 * management (if any):
2095 p
->sched_class
->task_new(rq
, p
);
2096 inc_nr_running(p
, rq
);
2098 check_preempt_curr(rq
, p
);
2100 if (p
->sched_class
->task_wake_up
)
2101 p
->sched_class
->task_wake_up(rq
, p
);
2103 task_rq_unlock(rq
, &flags
);
2106 #ifdef CONFIG_PREEMPT_NOTIFIERS
2109 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2110 * @notifier: notifier struct to register
2112 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2114 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2116 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2119 * preempt_notifier_unregister - no longer interested in preemption notifications
2120 * @notifier: notifier struct to unregister
2122 * This is safe to call from within a preemption notifier.
2124 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2126 hlist_del(¬ifier
->link
);
2128 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2130 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2132 struct preempt_notifier
*notifier
;
2133 struct hlist_node
*node
;
2135 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2136 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2140 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2141 struct task_struct
*next
)
2143 struct preempt_notifier
*notifier
;
2144 struct hlist_node
*node
;
2146 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2147 notifier
->ops
->sched_out(notifier
, next
);
2152 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2157 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2158 struct task_struct
*next
)
2165 * prepare_task_switch - prepare to switch tasks
2166 * @rq: the runqueue preparing to switch
2167 * @prev: the current task that is being switched out
2168 * @next: the task we are going to switch to.
2170 * This is called with the rq lock held and interrupts off. It must
2171 * be paired with a subsequent finish_task_switch after the context
2174 * prepare_task_switch sets up locking and calls architecture specific
2178 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2179 struct task_struct
*next
)
2181 fire_sched_out_preempt_notifiers(prev
, next
);
2182 prepare_lock_switch(rq
, next
);
2183 prepare_arch_switch(next
);
2187 * finish_task_switch - clean up after a task-switch
2188 * @rq: runqueue associated with task-switch
2189 * @prev: the thread we just switched away from.
2191 * finish_task_switch must be called after the context switch, paired
2192 * with a prepare_task_switch call before the context switch.
2193 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2194 * and do any other architecture-specific cleanup actions.
2196 * Note that we may have delayed dropping an mm in context_switch(). If
2197 * so, we finish that here outside of the runqueue lock. (Doing it
2198 * with the lock held can cause deadlocks; see schedule() for
2201 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2202 __releases(rq
->lock
)
2204 struct mm_struct
*mm
= rq
->prev_mm
;
2210 * A task struct has one reference for the use as "current".
2211 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2212 * schedule one last time. The schedule call will never return, and
2213 * the scheduled task must drop that reference.
2214 * The test for TASK_DEAD must occur while the runqueue locks are
2215 * still held, otherwise prev could be scheduled on another cpu, die
2216 * there before we look at prev->state, and then the reference would
2218 * Manfred Spraul <manfred@colorfullife.com>
2220 prev_state
= prev
->state
;
2221 finish_arch_switch(prev
);
2222 finish_lock_switch(rq
, prev
);
2224 if (current
->sched_class
->post_schedule
)
2225 current
->sched_class
->post_schedule(rq
);
2228 fire_sched_in_preempt_notifiers(current
);
2231 if (unlikely(prev_state
== TASK_DEAD
)) {
2233 * Remove function-return probe instances associated with this
2234 * task and put them back on the free list.
2236 kprobe_flush_task(prev
);
2237 put_task_struct(prev
);
2242 * schedule_tail - first thing a freshly forked thread must call.
2243 * @prev: the thread we just switched away from.
2245 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2246 __releases(rq
->lock
)
2248 struct rq
*rq
= this_rq();
2250 finish_task_switch(rq
, prev
);
2251 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2252 /* In this case, finish_task_switch does not reenable preemption */
2255 if (current
->set_child_tid
)
2256 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2260 * context_switch - switch to the new MM and the new
2261 * thread's register state.
2264 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2265 struct task_struct
*next
)
2267 struct mm_struct
*mm
, *oldmm
;
2269 prepare_task_switch(rq
, prev
, next
);
2271 oldmm
= prev
->active_mm
;
2273 * For paravirt, this is coupled with an exit in switch_to to
2274 * combine the page table reload and the switch backend into
2277 arch_enter_lazy_cpu_mode();
2279 if (unlikely(!mm
)) {
2280 next
->active_mm
= oldmm
;
2281 atomic_inc(&oldmm
->mm_count
);
2282 enter_lazy_tlb(oldmm
, next
);
2284 switch_mm(oldmm
, mm
, next
);
2286 if (unlikely(!prev
->mm
)) {
2287 prev
->active_mm
= NULL
;
2288 rq
->prev_mm
= oldmm
;
2291 * Since the runqueue lock will be released by the next
2292 * task (which is an invalid locking op but in the case
2293 * of the scheduler it's an obvious special-case), so we
2294 * do an early lockdep release here:
2296 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2297 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2300 /* Here we just switch the register state and the stack. */
2301 switch_to(prev
, next
, prev
);
2305 * this_rq must be evaluated again because prev may have moved
2306 * CPUs since it called schedule(), thus the 'rq' on its stack
2307 * frame will be invalid.
2309 finish_task_switch(this_rq(), prev
);
2313 * nr_running, nr_uninterruptible and nr_context_switches:
2315 * externally visible scheduler statistics: current number of runnable
2316 * threads, current number of uninterruptible-sleeping threads, total
2317 * number of context switches performed since bootup.
2319 unsigned long nr_running(void)
2321 unsigned long i
, sum
= 0;
2323 for_each_online_cpu(i
)
2324 sum
+= cpu_rq(i
)->nr_running
;
2329 unsigned long nr_uninterruptible(void)
2331 unsigned long i
, sum
= 0;
2333 for_each_possible_cpu(i
)
2334 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2337 * Since we read the counters lockless, it might be slightly
2338 * inaccurate. Do not allow it to go below zero though:
2340 if (unlikely((long)sum
< 0))
2346 unsigned long long nr_context_switches(void)
2349 unsigned long long sum
= 0;
2351 for_each_possible_cpu(i
)
2352 sum
+= cpu_rq(i
)->nr_switches
;
2357 unsigned long nr_iowait(void)
2359 unsigned long i
, sum
= 0;
2361 for_each_possible_cpu(i
)
2362 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2367 unsigned long nr_active(void)
2369 unsigned long i
, running
= 0, uninterruptible
= 0;
2371 for_each_online_cpu(i
) {
2372 running
+= cpu_rq(i
)->nr_running
;
2373 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2376 if (unlikely((long)uninterruptible
< 0))
2377 uninterruptible
= 0;
2379 return running
+ uninterruptible
;
2383 * Update rq->cpu_load[] statistics. This function is usually called every
2384 * scheduler tick (TICK_NSEC).
2386 static void update_cpu_load(struct rq
*this_rq
)
2388 unsigned long this_load
= this_rq
->load
.weight
;
2391 this_rq
->nr_load_updates
++;
2393 /* Update our load: */
2394 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2395 unsigned long old_load
, new_load
;
2397 /* scale is effectively 1 << i now, and >> i divides by scale */
2399 old_load
= this_rq
->cpu_load
[i
];
2400 new_load
= this_load
;
2402 * Round up the averaging division if load is increasing. This
2403 * prevents us from getting stuck on 9 if the load is 10, for
2406 if (new_load
> old_load
)
2407 new_load
+= scale
-1;
2408 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2415 * double_rq_lock - safely lock two runqueues
2417 * Note this does not disable interrupts like task_rq_lock,
2418 * you need to do so manually before calling.
2420 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2421 __acquires(rq1
->lock
)
2422 __acquires(rq2
->lock
)
2424 BUG_ON(!irqs_disabled());
2426 spin_lock(&rq1
->lock
);
2427 __acquire(rq2
->lock
); /* Fake it out ;) */
2430 spin_lock(&rq1
->lock
);
2431 spin_lock(&rq2
->lock
);
2433 spin_lock(&rq2
->lock
);
2434 spin_lock(&rq1
->lock
);
2437 update_rq_clock(rq1
);
2438 update_rq_clock(rq2
);
2442 * double_rq_unlock - safely unlock two runqueues
2444 * Note this does not restore interrupts like task_rq_unlock,
2445 * you need to do so manually after calling.
2447 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2448 __releases(rq1
->lock
)
2449 __releases(rq2
->lock
)
2451 spin_unlock(&rq1
->lock
);
2453 spin_unlock(&rq2
->lock
);
2455 __release(rq2
->lock
);
2459 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2461 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2462 __releases(this_rq
->lock
)
2463 __acquires(busiest
->lock
)
2464 __acquires(this_rq
->lock
)
2468 if (unlikely(!irqs_disabled())) {
2469 /* printk() doesn't work good under rq->lock */
2470 spin_unlock(&this_rq
->lock
);
2473 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2474 if (busiest
< this_rq
) {
2475 spin_unlock(&this_rq
->lock
);
2476 spin_lock(&busiest
->lock
);
2477 spin_lock(&this_rq
->lock
);
2480 spin_lock(&busiest
->lock
);
2486 * If dest_cpu is allowed for this process, migrate the task to it.
2487 * This is accomplished by forcing the cpu_allowed mask to only
2488 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2489 * the cpu_allowed mask is restored.
2491 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2493 struct migration_req req
;
2494 unsigned long flags
;
2497 rq
= task_rq_lock(p
, &flags
);
2498 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2499 || unlikely(cpu_is_offline(dest_cpu
)))
2502 /* force the process onto the specified CPU */
2503 if (migrate_task(p
, dest_cpu
, &req
)) {
2504 /* Need to wait for migration thread (might exit: take ref). */
2505 struct task_struct
*mt
= rq
->migration_thread
;
2507 get_task_struct(mt
);
2508 task_rq_unlock(rq
, &flags
);
2509 wake_up_process(mt
);
2510 put_task_struct(mt
);
2511 wait_for_completion(&req
.done
);
2516 task_rq_unlock(rq
, &flags
);
2520 * sched_exec - execve() is a valuable balancing opportunity, because at
2521 * this point the task has the smallest effective memory and cache footprint.
2523 void sched_exec(void)
2525 int new_cpu
, this_cpu
= get_cpu();
2526 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2528 if (new_cpu
!= this_cpu
)
2529 sched_migrate_task(current
, new_cpu
);
2533 * pull_task - move a task from a remote runqueue to the local runqueue.
2534 * Both runqueues must be locked.
2536 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2537 struct rq
*this_rq
, int this_cpu
)
2539 deactivate_task(src_rq
, p
, 0);
2540 set_task_cpu(p
, this_cpu
);
2541 activate_task(this_rq
, p
, 0);
2543 * Note that idle threads have a prio of MAX_PRIO, for this test
2544 * to be always true for them.
2546 check_preempt_curr(this_rq
, p
);
2550 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2553 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2554 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2558 * We do not migrate tasks that are:
2559 * 1) running (obviously), or
2560 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2561 * 3) are cache-hot on their current CPU.
2563 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2564 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2569 if (task_running(rq
, p
)) {
2570 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2575 * Aggressive migration if:
2576 * 1) task is cache cold, or
2577 * 2) too many balance attempts have failed.
2580 if (!task_hot(p
, rq
->clock
, sd
) ||
2581 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2582 #ifdef CONFIG_SCHEDSTATS
2583 if (task_hot(p
, rq
->clock
, sd
)) {
2584 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2585 schedstat_inc(p
, se
.nr_forced_migrations
);
2591 if (task_hot(p
, rq
->clock
, sd
)) {
2592 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2598 static unsigned long
2599 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2600 unsigned long max_load_move
, struct sched_domain
*sd
,
2601 enum cpu_idle_type idle
, int *all_pinned
,
2602 int *this_best_prio
, struct rq_iterator
*iterator
)
2604 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2605 struct task_struct
*p
;
2606 long rem_load_move
= max_load_move
;
2608 if (max_load_move
== 0)
2614 * Start the load-balancing iterator:
2616 p
= iterator
->start(iterator
->arg
);
2618 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2621 * To help distribute high priority tasks across CPUs we don't
2622 * skip a task if it will be the highest priority task (i.e. smallest
2623 * prio value) on its new queue regardless of its load weight
2625 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2626 SCHED_LOAD_SCALE_FUZZ
;
2627 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2628 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2629 p
= iterator
->next(iterator
->arg
);
2633 pull_task(busiest
, p
, this_rq
, this_cpu
);
2635 rem_load_move
-= p
->se
.load
.weight
;
2638 * We only want to steal up to the prescribed amount of weighted load.
2640 if (rem_load_move
> 0) {
2641 if (p
->prio
< *this_best_prio
)
2642 *this_best_prio
= p
->prio
;
2643 p
= iterator
->next(iterator
->arg
);
2648 * Right now, this is one of only two places pull_task() is called,
2649 * so we can safely collect pull_task() stats here rather than
2650 * inside pull_task().
2652 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2655 *all_pinned
= pinned
;
2657 return max_load_move
- rem_load_move
;
2661 * move_tasks tries to move up to max_load_move weighted load from busiest to
2662 * this_rq, as part of a balancing operation within domain "sd".
2663 * Returns 1 if successful and 0 otherwise.
2665 * Called with both runqueues locked.
2667 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2668 unsigned long max_load_move
,
2669 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2672 const struct sched_class
*class = sched_class_highest
;
2673 unsigned long total_load_moved
= 0;
2674 int this_best_prio
= this_rq
->curr
->prio
;
2678 class->load_balance(this_rq
, this_cpu
, busiest
,
2679 max_load_move
- total_load_moved
,
2680 sd
, idle
, all_pinned
, &this_best_prio
);
2681 class = class->next
;
2682 } while (class && max_load_move
> total_load_moved
);
2684 return total_load_moved
> 0;
2688 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2689 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2690 struct rq_iterator
*iterator
)
2692 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2696 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2697 pull_task(busiest
, p
, this_rq
, this_cpu
);
2699 * Right now, this is only the second place pull_task()
2700 * is called, so we can safely collect pull_task()
2701 * stats here rather than inside pull_task().
2703 schedstat_inc(sd
, lb_gained
[idle
]);
2707 p
= iterator
->next(iterator
->arg
);
2714 * move_one_task tries to move exactly one task from busiest to this_rq, as
2715 * part of active balancing operations within "domain".
2716 * Returns 1 if successful and 0 otherwise.
2718 * Called with both runqueues locked.
2720 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2721 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2723 const struct sched_class
*class;
2725 for (class = sched_class_highest
; class; class = class->next
)
2726 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2733 * find_busiest_group finds and returns the busiest CPU group within the
2734 * domain. It calculates and returns the amount of weighted load which
2735 * should be moved to restore balance via the imbalance parameter.
2737 static struct sched_group
*
2738 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2739 unsigned long *imbalance
, enum cpu_idle_type idle
,
2740 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2742 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2743 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2744 unsigned long max_pull
;
2745 unsigned long busiest_load_per_task
, busiest_nr_running
;
2746 unsigned long this_load_per_task
, this_nr_running
;
2747 int load_idx
, group_imb
= 0;
2748 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2749 int power_savings_balance
= 1;
2750 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2751 unsigned long min_nr_running
= ULONG_MAX
;
2752 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2755 max_load
= this_load
= total_load
= total_pwr
= 0;
2756 busiest_load_per_task
= busiest_nr_running
= 0;
2757 this_load_per_task
= this_nr_running
= 0;
2758 if (idle
== CPU_NOT_IDLE
)
2759 load_idx
= sd
->busy_idx
;
2760 else if (idle
== CPU_NEWLY_IDLE
)
2761 load_idx
= sd
->newidle_idx
;
2763 load_idx
= sd
->idle_idx
;
2766 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2769 int __group_imb
= 0;
2770 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2771 unsigned long sum_nr_running
, sum_weighted_load
;
2773 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2776 balance_cpu
= first_cpu(group
->cpumask
);
2778 /* Tally up the load of all CPUs in the group */
2779 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2781 min_cpu_load
= ~0UL;
2783 for_each_cpu_mask(i
, group
->cpumask
) {
2786 if (!cpu_isset(i
, *cpus
))
2791 if (*sd_idle
&& rq
->nr_running
)
2794 /* Bias balancing toward cpus of our domain */
2796 if (idle_cpu(i
) && !first_idle_cpu
) {
2801 load
= target_load(i
, load_idx
);
2803 load
= source_load(i
, load_idx
);
2804 if (load
> max_cpu_load
)
2805 max_cpu_load
= load
;
2806 if (min_cpu_load
> load
)
2807 min_cpu_load
= load
;
2811 sum_nr_running
+= rq
->nr_running
;
2812 sum_weighted_load
+= weighted_cpuload(i
);
2816 * First idle cpu or the first cpu(busiest) in this sched group
2817 * is eligible for doing load balancing at this and above
2818 * domains. In the newly idle case, we will allow all the cpu's
2819 * to do the newly idle load balance.
2821 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2822 balance_cpu
!= this_cpu
&& balance
) {
2827 total_load
+= avg_load
;
2828 total_pwr
+= group
->__cpu_power
;
2830 /* Adjust by relative CPU power of the group */
2831 avg_load
= sg_div_cpu_power(group
,
2832 avg_load
* SCHED_LOAD_SCALE
);
2834 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2837 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2840 this_load
= avg_load
;
2842 this_nr_running
= sum_nr_running
;
2843 this_load_per_task
= sum_weighted_load
;
2844 } else if (avg_load
> max_load
&&
2845 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2846 max_load
= avg_load
;
2848 busiest_nr_running
= sum_nr_running
;
2849 busiest_load_per_task
= sum_weighted_load
;
2850 group_imb
= __group_imb
;
2853 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2855 * Busy processors will not participate in power savings
2858 if (idle
== CPU_NOT_IDLE
||
2859 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2863 * If the local group is idle or completely loaded
2864 * no need to do power savings balance at this domain
2866 if (local_group
&& (this_nr_running
>= group_capacity
||
2868 power_savings_balance
= 0;
2871 * If a group is already running at full capacity or idle,
2872 * don't include that group in power savings calculations
2874 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2879 * Calculate the group which has the least non-idle load.
2880 * This is the group from where we need to pick up the load
2883 if ((sum_nr_running
< min_nr_running
) ||
2884 (sum_nr_running
== min_nr_running
&&
2885 first_cpu(group
->cpumask
) <
2886 first_cpu(group_min
->cpumask
))) {
2888 min_nr_running
= sum_nr_running
;
2889 min_load_per_task
= sum_weighted_load
/
2894 * Calculate the group which is almost near its
2895 * capacity but still has some space to pick up some load
2896 * from other group and save more power
2898 if (sum_nr_running
<= group_capacity
- 1) {
2899 if (sum_nr_running
> leader_nr_running
||
2900 (sum_nr_running
== leader_nr_running
&&
2901 first_cpu(group
->cpumask
) >
2902 first_cpu(group_leader
->cpumask
))) {
2903 group_leader
= group
;
2904 leader_nr_running
= sum_nr_running
;
2909 group
= group
->next
;
2910 } while (group
!= sd
->groups
);
2912 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2915 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2917 if (this_load
>= avg_load
||
2918 100*max_load
<= sd
->imbalance_pct
*this_load
)
2921 busiest_load_per_task
/= busiest_nr_running
;
2923 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2926 * We're trying to get all the cpus to the average_load, so we don't
2927 * want to push ourselves above the average load, nor do we wish to
2928 * reduce the max loaded cpu below the average load, as either of these
2929 * actions would just result in more rebalancing later, and ping-pong
2930 * tasks around. Thus we look for the minimum possible imbalance.
2931 * Negative imbalances (*we* are more loaded than anyone else) will
2932 * be counted as no imbalance for these purposes -- we can't fix that
2933 * by pulling tasks to us. Be careful of negative numbers as they'll
2934 * appear as very large values with unsigned longs.
2936 if (max_load
<= busiest_load_per_task
)
2940 * In the presence of smp nice balancing, certain scenarios can have
2941 * max load less than avg load(as we skip the groups at or below
2942 * its cpu_power, while calculating max_load..)
2944 if (max_load
< avg_load
) {
2946 goto small_imbalance
;
2949 /* Don't want to pull so many tasks that a group would go idle */
2950 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2952 /* How much load to actually move to equalise the imbalance */
2953 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2954 (avg_load
- this_load
) * this->__cpu_power
)
2958 * if *imbalance is less than the average load per runnable task
2959 * there is no gaurantee that any tasks will be moved so we'll have
2960 * a think about bumping its value to force at least one task to be
2963 if (*imbalance
< busiest_load_per_task
) {
2964 unsigned long tmp
, pwr_now
, pwr_move
;
2968 pwr_move
= pwr_now
= 0;
2970 if (this_nr_running
) {
2971 this_load_per_task
/= this_nr_running
;
2972 if (busiest_load_per_task
> this_load_per_task
)
2975 this_load_per_task
= SCHED_LOAD_SCALE
;
2977 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2978 busiest_load_per_task
* imbn
) {
2979 *imbalance
= busiest_load_per_task
;
2984 * OK, we don't have enough imbalance to justify moving tasks,
2985 * however we may be able to increase total CPU power used by
2989 pwr_now
+= busiest
->__cpu_power
*
2990 min(busiest_load_per_task
, max_load
);
2991 pwr_now
+= this->__cpu_power
*
2992 min(this_load_per_task
, this_load
);
2993 pwr_now
/= SCHED_LOAD_SCALE
;
2995 /* Amount of load we'd subtract */
2996 tmp
= sg_div_cpu_power(busiest
,
2997 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2999 pwr_move
+= busiest
->__cpu_power
*
3000 min(busiest_load_per_task
, max_load
- tmp
);
3002 /* Amount of load we'd add */
3003 if (max_load
* busiest
->__cpu_power
<
3004 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3005 tmp
= sg_div_cpu_power(this,
3006 max_load
* busiest
->__cpu_power
);
3008 tmp
= sg_div_cpu_power(this,
3009 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3010 pwr_move
+= this->__cpu_power
*
3011 min(this_load_per_task
, this_load
+ tmp
);
3012 pwr_move
/= SCHED_LOAD_SCALE
;
3014 /* Move if we gain throughput */
3015 if (pwr_move
> pwr_now
)
3016 *imbalance
= busiest_load_per_task
;
3022 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3023 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3026 if (this == group_leader
&& group_leader
!= group_min
) {
3027 *imbalance
= min_load_per_task
;
3037 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3040 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3041 unsigned long imbalance
, cpumask_t
*cpus
)
3043 struct rq
*busiest
= NULL
, *rq
;
3044 unsigned long max_load
= 0;
3047 for_each_cpu_mask(i
, group
->cpumask
) {
3050 if (!cpu_isset(i
, *cpus
))
3054 wl
= weighted_cpuload(i
);
3056 if (rq
->nr_running
== 1 && wl
> imbalance
)
3059 if (wl
> max_load
) {
3069 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3070 * so long as it is large enough.
3072 #define MAX_PINNED_INTERVAL 512
3075 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3076 * tasks if there is an imbalance.
3078 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3079 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3082 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3083 struct sched_group
*group
;
3084 unsigned long imbalance
;
3086 cpumask_t cpus
= CPU_MASK_ALL
;
3087 unsigned long flags
;
3090 * When power savings policy is enabled for the parent domain, idle
3091 * sibling can pick up load irrespective of busy siblings. In this case,
3092 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3093 * portraying it as CPU_NOT_IDLE.
3095 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3096 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3099 schedstat_inc(sd
, lb_count
[idle
]);
3102 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3109 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3113 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
3115 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3119 BUG_ON(busiest
== this_rq
);
3121 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3124 if (busiest
->nr_running
> 1) {
3126 * Attempt to move tasks. If find_busiest_group has found
3127 * an imbalance but busiest->nr_running <= 1, the group is
3128 * still unbalanced. ld_moved simply stays zero, so it is
3129 * correctly treated as an imbalance.
3131 local_irq_save(flags
);
3132 double_rq_lock(this_rq
, busiest
);
3133 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3134 imbalance
, sd
, idle
, &all_pinned
);
3135 double_rq_unlock(this_rq
, busiest
);
3136 local_irq_restore(flags
);
3139 * some other cpu did the load balance for us.
3141 if (ld_moved
&& this_cpu
!= smp_processor_id())
3142 resched_cpu(this_cpu
);
3144 /* All tasks on this runqueue were pinned by CPU affinity */
3145 if (unlikely(all_pinned
)) {
3146 cpu_clear(cpu_of(busiest
), cpus
);
3147 if (!cpus_empty(cpus
))
3154 schedstat_inc(sd
, lb_failed
[idle
]);
3155 sd
->nr_balance_failed
++;
3157 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3159 spin_lock_irqsave(&busiest
->lock
, flags
);
3161 /* don't kick the migration_thread, if the curr
3162 * task on busiest cpu can't be moved to this_cpu
3164 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3165 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3167 goto out_one_pinned
;
3170 if (!busiest
->active_balance
) {
3171 busiest
->active_balance
= 1;
3172 busiest
->push_cpu
= this_cpu
;
3175 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3177 wake_up_process(busiest
->migration_thread
);
3180 * We've kicked active balancing, reset the failure
3183 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3186 sd
->nr_balance_failed
= 0;
3188 if (likely(!active_balance
)) {
3189 /* We were unbalanced, so reset the balancing interval */
3190 sd
->balance_interval
= sd
->min_interval
;
3193 * If we've begun active balancing, start to back off. This
3194 * case may not be covered by the all_pinned logic if there
3195 * is only 1 task on the busy runqueue (because we don't call
3198 if (sd
->balance_interval
< sd
->max_interval
)
3199 sd
->balance_interval
*= 2;
3202 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3203 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3208 schedstat_inc(sd
, lb_balanced
[idle
]);
3210 sd
->nr_balance_failed
= 0;
3213 /* tune up the balancing interval */
3214 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3215 (sd
->balance_interval
< sd
->max_interval
))
3216 sd
->balance_interval
*= 2;
3218 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3219 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3225 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3226 * tasks if there is an imbalance.
3228 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3229 * this_rq is locked.
3232 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
3234 struct sched_group
*group
;
3235 struct rq
*busiest
= NULL
;
3236 unsigned long imbalance
;
3240 cpumask_t cpus
= CPU_MASK_ALL
;
3243 * When power savings policy is enabled for the parent domain, idle
3244 * sibling can pick up load irrespective of busy siblings. In this case,
3245 * let the state of idle sibling percolate up as IDLE, instead of
3246 * portraying it as CPU_NOT_IDLE.
3248 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3249 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3252 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3254 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3255 &sd_idle
, &cpus
, NULL
);
3257 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3261 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
3264 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3268 BUG_ON(busiest
== this_rq
);
3270 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3273 if (busiest
->nr_running
> 1) {
3274 /* Attempt to move tasks */
3275 double_lock_balance(this_rq
, busiest
);
3276 /* this_rq->clock is already updated */
3277 update_rq_clock(busiest
);
3278 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3279 imbalance
, sd
, CPU_NEWLY_IDLE
,
3281 spin_unlock(&busiest
->lock
);
3283 if (unlikely(all_pinned
)) {
3284 cpu_clear(cpu_of(busiest
), cpus
);
3285 if (!cpus_empty(cpus
))
3291 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3292 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3293 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3296 sd
->nr_balance_failed
= 0;
3301 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3302 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3303 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3305 sd
->nr_balance_failed
= 0;
3311 * idle_balance is called by schedule() if this_cpu is about to become
3312 * idle. Attempts to pull tasks from other CPUs.
3314 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3316 struct sched_domain
*sd
;
3317 int pulled_task
= -1;
3318 unsigned long next_balance
= jiffies
+ HZ
;
3320 for_each_domain(this_cpu
, sd
) {
3321 unsigned long interval
;
3323 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3326 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3327 /* If we've pulled tasks over stop searching: */
3328 pulled_task
= load_balance_newidle(this_cpu
,
3331 interval
= msecs_to_jiffies(sd
->balance_interval
);
3332 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3333 next_balance
= sd
->last_balance
+ interval
;
3337 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3339 * We are going idle. next_balance may be set based on
3340 * a busy processor. So reset next_balance.
3342 this_rq
->next_balance
= next_balance
;
3347 * active_load_balance is run by migration threads. It pushes running tasks
3348 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3349 * running on each physical CPU where possible, and avoids physical /
3350 * logical imbalances.
3352 * Called with busiest_rq locked.
3354 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3356 int target_cpu
= busiest_rq
->push_cpu
;
3357 struct sched_domain
*sd
;
3358 struct rq
*target_rq
;
3360 /* Is there any task to move? */
3361 if (busiest_rq
->nr_running
<= 1)
3364 target_rq
= cpu_rq(target_cpu
);
3367 * This condition is "impossible", if it occurs
3368 * we need to fix it. Originally reported by
3369 * Bjorn Helgaas on a 128-cpu setup.
3371 BUG_ON(busiest_rq
== target_rq
);
3373 /* move a task from busiest_rq to target_rq */
3374 double_lock_balance(busiest_rq
, target_rq
);
3375 update_rq_clock(busiest_rq
);
3376 update_rq_clock(target_rq
);
3378 /* Search for an sd spanning us and the target CPU. */
3379 for_each_domain(target_cpu
, sd
) {
3380 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3381 cpu_isset(busiest_cpu
, sd
->span
))
3386 schedstat_inc(sd
, alb_count
);
3388 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3390 schedstat_inc(sd
, alb_pushed
);
3392 schedstat_inc(sd
, alb_failed
);
3394 spin_unlock(&target_rq
->lock
);
3399 atomic_t load_balancer
;
3401 } nohz ____cacheline_aligned
= {
3402 .load_balancer
= ATOMIC_INIT(-1),
3403 .cpu_mask
= CPU_MASK_NONE
,
3407 * This routine will try to nominate the ilb (idle load balancing)
3408 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3409 * load balancing on behalf of all those cpus. If all the cpus in the system
3410 * go into this tickless mode, then there will be no ilb owner (as there is
3411 * no need for one) and all the cpus will sleep till the next wakeup event
3414 * For the ilb owner, tick is not stopped. And this tick will be used
3415 * for idle load balancing. ilb owner will still be part of
3418 * While stopping the tick, this cpu will become the ilb owner if there
3419 * is no other owner. And will be the owner till that cpu becomes busy
3420 * or if all cpus in the system stop their ticks at which point
3421 * there is no need for ilb owner.
3423 * When the ilb owner becomes busy, it nominates another owner, during the
3424 * next busy scheduler_tick()
3426 int select_nohz_load_balancer(int stop_tick
)
3428 int cpu
= smp_processor_id();
3431 cpu_set(cpu
, nohz
.cpu_mask
);
3432 cpu_rq(cpu
)->in_nohz_recently
= 1;
3435 * If we are going offline and still the leader, give up!
3437 if (cpu_is_offline(cpu
) &&
3438 atomic_read(&nohz
.load_balancer
) == cpu
) {
3439 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3444 /* time for ilb owner also to sleep */
3445 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3446 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3447 atomic_set(&nohz
.load_balancer
, -1);
3451 if (atomic_read(&nohz
.load_balancer
) == -1) {
3452 /* make me the ilb owner */
3453 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3455 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3458 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3461 cpu_clear(cpu
, nohz
.cpu_mask
);
3463 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3464 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3471 static DEFINE_SPINLOCK(balancing
);
3474 * It checks each scheduling domain to see if it is due to be balanced,
3475 * and initiates a balancing operation if so.
3477 * Balancing parameters are set up in arch_init_sched_domains.
3479 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3482 struct rq
*rq
= cpu_rq(cpu
);
3483 unsigned long interval
;
3484 struct sched_domain
*sd
;
3485 /* Earliest time when we have to do rebalance again */
3486 unsigned long next_balance
= jiffies
+ 60*HZ
;
3487 int update_next_balance
= 0;
3489 for_each_domain(cpu
, sd
) {
3490 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3493 interval
= sd
->balance_interval
;
3494 if (idle
!= CPU_IDLE
)
3495 interval
*= sd
->busy_factor
;
3497 /* scale ms to jiffies */
3498 interval
= msecs_to_jiffies(interval
);
3499 if (unlikely(!interval
))
3501 if (interval
> HZ
*NR_CPUS
/10)
3502 interval
= HZ
*NR_CPUS
/10;
3505 if (sd
->flags
& SD_SERIALIZE
) {
3506 if (!spin_trylock(&balancing
))
3510 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3511 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3513 * We've pulled tasks over so either we're no
3514 * longer idle, or one of our SMT siblings is
3517 idle
= CPU_NOT_IDLE
;
3519 sd
->last_balance
= jiffies
;
3521 if (sd
->flags
& SD_SERIALIZE
)
3522 spin_unlock(&balancing
);
3524 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3525 next_balance
= sd
->last_balance
+ interval
;
3526 update_next_balance
= 1;
3530 * Stop the load balance at this level. There is another
3531 * CPU in our sched group which is doing load balancing more
3539 * next_balance will be updated only when there is a need.
3540 * When the cpu is attached to null domain for ex, it will not be
3543 if (likely(update_next_balance
))
3544 rq
->next_balance
= next_balance
;
3548 * run_rebalance_domains is triggered when needed from the scheduler tick.
3549 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3550 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3552 static void run_rebalance_domains(struct softirq_action
*h
)
3554 int this_cpu
= smp_processor_id();
3555 struct rq
*this_rq
= cpu_rq(this_cpu
);
3556 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3557 CPU_IDLE
: CPU_NOT_IDLE
;
3559 rebalance_domains(this_cpu
, idle
);
3563 * If this cpu is the owner for idle load balancing, then do the
3564 * balancing on behalf of the other idle cpus whose ticks are
3567 if (this_rq
->idle_at_tick
&&
3568 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3569 cpumask_t cpus
= nohz
.cpu_mask
;
3573 cpu_clear(this_cpu
, cpus
);
3574 for_each_cpu_mask(balance_cpu
, cpus
) {
3576 * If this cpu gets work to do, stop the load balancing
3577 * work being done for other cpus. Next load
3578 * balancing owner will pick it up.
3583 rebalance_domains(balance_cpu
, CPU_IDLE
);
3585 rq
= cpu_rq(balance_cpu
);
3586 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3587 this_rq
->next_balance
= rq
->next_balance
;
3594 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3596 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3597 * idle load balancing owner or decide to stop the periodic load balancing,
3598 * if the whole system is idle.
3600 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3604 * If we were in the nohz mode recently and busy at the current
3605 * scheduler tick, then check if we need to nominate new idle
3608 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3609 rq
->in_nohz_recently
= 0;
3611 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3612 cpu_clear(cpu
, nohz
.cpu_mask
);
3613 atomic_set(&nohz
.load_balancer
, -1);
3616 if (atomic_read(&nohz
.load_balancer
) == -1) {
3618 * simple selection for now: Nominate the
3619 * first cpu in the nohz list to be the next
3622 * TBD: Traverse the sched domains and nominate
3623 * the nearest cpu in the nohz.cpu_mask.
3625 int ilb
= first_cpu(nohz
.cpu_mask
);
3633 * If this cpu is idle and doing idle load balancing for all the
3634 * cpus with ticks stopped, is it time for that to stop?
3636 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3637 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3643 * If this cpu is idle and the idle load balancing is done by
3644 * someone else, then no need raise the SCHED_SOFTIRQ
3646 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3647 cpu_isset(cpu
, nohz
.cpu_mask
))
3650 if (time_after_eq(jiffies
, rq
->next_balance
))
3651 raise_softirq(SCHED_SOFTIRQ
);
3654 #else /* CONFIG_SMP */
3657 * on UP we do not need to balance between CPUs:
3659 static inline void idle_balance(int cpu
, struct rq
*rq
)
3665 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3667 EXPORT_PER_CPU_SYMBOL(kstat
);
3670 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3671 * that have not yet been banked in case the task is currently running.
3673 unsigned long long task_sched_runtime(struct task_struct
*p
)
3675 unsigned long flags
;
3679 rq
= task_rq_lock(p
, &flags
);
3680 ns
= p
->se
.sum_exec_runtime
;
3681 if (task_current(rq
, p
)) {
3682 update_rq_clock(rq
);
3683 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3684 if ((s64
)delta_exec
> 0)
3687 task_rq_unlock(rq
, &flags
);
3693 * Account user cpu time to a process.
3694 * @p: the process that the cpu time gets accounted to
3695 * @cputime: the cpu time spent in user space since the last update
3697 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3699 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3702 p
->utime
= cputime_add(p
->utime
, cputime
);
3704 /* Add user time to cpustat. */
3705 tmp
= cputime_to_cputime64(cputime
);
3706 if (TASK_NICE(p
) > 0)
3707 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3709 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3713 * Account guest cpu time to a process.
3714 * @p: the process that the cpu time gets accounted to
3715 * @cputime: the cpu time spent in virtual machine since the last update
3717 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3720 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3722 tmp
= cputime_to_cputime64(cputime
);
3724 p
->utime
= cputime_add(p
->utime
, cputime
);
3725 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3727 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3728 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3732 * Account scaled user cpu time to a process.
3733 * @p: the process that the cpu time gets accounted to
3734 * @cputime: the cpu time spent in user space since the last update
3736 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3738 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3742 * Account system cpu time to a process.
3743 * @p: the process that the cpu time gets accounted to
3744 * @hardirq_offset: the offset to subtract from hardirq_count()
3745 * @cputime: the cpu time spent in kernel space since the last update
3747 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3750 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3751 struct rq
*rq
= this_rq();
3754 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3755 return account_guest_time(p
, cputime
);
3757 p
->stime
= cputime_add(p
->stime
, cputime
);
3759 /* Add system time to cpustat. */
3760 tmp
= cputime_to_cputime64(cputime
);
3761 if (hardirq_count() - hardirq_offset
)
3762 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3763 else if (softirq_count())
3764 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3765 else if (p
!= rq
->idle
)
3766 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3767 else if (atomic_read(&rq
->nr_iowait
) > 0)
3768 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3770 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3771 /* Account for system time used */
3772 acct_update_integrals(p
);
3776 * Account scaled system cpu time to a process.
3777 * @p: the process that the cpu time gets accounted to
3778 * @hardirq_offset: the offset to subtract from hardirq_count()
3779 * @cputime: the cpu time spent in kernel space since the last update
3781 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3783 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3787 * Account for involuntary wait time.
3788 * @p: the process from which the cpu time has been stolen
3789 * @steal: the cpu time spent in involuntary wait
3791 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3793 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3794 cputime64_t tmp
= cputime_to_cputime64(steal
);
3795 struct rq
*rq
= this_rq();
3797 if (p
== rq
->idle
) {
3798 p
->stime
= cputime_add(p
->stime
, steal
);
3799 if (atomic_read(&rq
->nr_iowait
) > 0)
3800 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3802 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3804 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3808 * This function gets called by the timer code, with HZ frequency.
3809 * We call it with interrupts disabled.
3811 * It also gets called by the fork code, when changing the parent's
3814 void scheduler_tick(void)
3816 int cpu
= smp_processor_id();
3817 struct rq
*rq
= cpu_rq(cpu
);
3818 struct task_struct
*curr
= rq
->curr
;
3819 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3821 spin_lock(&rq
->lock
);
3822 __update_rq_clock(rq
);
3824 * Let rq->clock advance by at least TICK_NSEC:
3826 if (unlikely(rq
->clock
< next_tick
)) {
3827 rq
->clock
= next_tick
;
3828 rq
->clock_underflows
++;
3830 rq
->tick_timestamp
= rq
->clock
;
3831 update_cpu_load(rq
);
3832 curr
->sched_class
->task_tick(rq
, curr
, 0);
3833 update_sched_rt_period(rq
);
3834 spin_unlock(&rq
->lock
);
3837 rq
->idle_at_tick
= idle_cpu(cpu
);
3838 trigger_load_balance(rq
, cpu
);
3842 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3844 void __kprobes
add_preempt_count(int val
)
3849 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3851 preempt_count() += val
;
3853 * Spinlock count overflowing soon?
3855 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3858 EXPORT_SYMBOL(add_preempt_count
);
3860 void __kprobes
sub_preempt_count(int val
)
3865 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3868 * Is the spinlock portion underflowing?
3870 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3871 !(preempt_count() & PREEMPT_MASK
)))
3874 preempt_count() -= val
;
3876 EXPORT_SYMBOL(sub_preempt_count
);
3881 * Print scheduling while atomic bug:
3883 static noinline
void __schedule_bug(struct task_struct
*prev
)
3885 struct pt_regs
*regs
= get_irq_regs();
3887 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3888 prev
->comm
, prev
->pid
, preempt_count());
3890 debug_show_held_locks(prev
);
3891 if (irqs_disabled())
3892 print_irqtrace_events(prev
);
3901 * Various schedule()-time debugging checks and statistics:
3903 static inline void schedule_debug(struct task_struct
*prev
)
3906 * Test if we are atomic. Since do_exit() needs to call into
3907 * schedule() atomically, we ignore that path for now.
3908 * Otherwise, whine if we are scheduling when we should not be.
3910 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3911 __schedule_bug(prev
);
3913 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3915 schedstat_inc(this_rq(), sched_count
);
3916 #ifdef CONFIG_SCHEDSTATS
3917 if (unlikely(prev
->lock_depth
>= 0)) {
3918 schedstat_inc(this_rq(), bkl_count
);
3919 schedstat_inc(prev
, sched_info
.bkl_count
);
3925 * Pick up the highest-prio task:
3927 static inline struct task_struct
*
3928 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3930 const struct sched_class
*class;
3931 struct task_struct
*p
;
3934 * Optimization: we know that if all tasks are in
3935 * the fair class we can call that function directly:
3937 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3938 p
= fair_sched_class
.pick_next_task(rq
);
3943 class = sched_class_highest
;
3945 p
= class->pick_next_task(rq
);
3949 * Will never be NULL as the idle class always
3950 * returns a non-NULL p:
3952 class = class->next
;
3957 * schedule() is the main scheduler function.
3959 asmlinkage
void __sched
schedule(void)
3961 struct task_struct
*prev
, *next
;
3962 unsigned long *switch_count
;
3968 cpu
= smp_processor_id();
3972 switch_count
= &prev
->nivcsw
;
3974 release_kernel_lock(prev
);
3975 need_resched_nonpreemptible
:
3977 schedule_debug(prev
);
3982 * Do the rq-clock update outside the rq lock:
3984 local_irq_disable();
3985 __update_rq_clock(rq
);
3986 spin_lock(&rq
->lock
);
3987 clear_tsk_need_resched(prev
);
3989 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3990 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3991 signal_pending(prev
))) {
3992 prev
->state
= TASK_RUNNING
;
3994 deactivate_task(rq
, prev
, 1);
3996 switch_count
= &prev
->nvcsw
;
4000 if (prev
->sched_class
->pre_schedule
)
4001 prev
->sched_class
->pre_schedule(rq
, prev
);
4004 if (unlikely(!rq
->nr_running
))
4005 idle_balance(cpu
, rq
);
4007 prev
->sched_class
->put_prev_task(rq
, prev
);
4008 next
= pick_next_task(rq
, prev
);
4010 sched_info_switch(prev
, next
);
4012 if (likely(prev
!= next
)) {
4017 context_switch(rq
, prev
, next
); /* unlocks the rq */
4019 * the context switch might have flipped the stack from under
4020 * us, hence refresh the local variables.
4022 cpu
= smp_processor_id();
4025 spin_unlock_irq(&rq
->lock
);
4029 if (unlikely(reacquire_kernel_lock(current
) < 0))
4030 goto need_resched_nonpreemptible
;
4032 preempt_enable_no_resched();
4033 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4036 EXPORT_SYMBOL(schedule
);
4038 #ifdef CONFIG_PREEMPT
4040 * this is the entry point to schedule() from in-kernel preemption
4041 * off of preempt_enable. Kernel preemptions off return from interrupt
4042 * occur there and call schedule directly.
4044 asmlinkage
void __sched
preempt_schedule(void)
4046 struct thread_info
*ti
= current_thread_info();
4047 struct task_struct
*task
= current
;
4048 int saved_lock_depth
;
4051 * If there is a non-zero preempt_count or interrupts are disabled,
4052 * we do not want to preempt the current task. Just return..
4054 if (likely(ti
->preempt_count
|| irqs_disabled()))
4058 add_preempt_count(PREEMPT_ACTIVE
);
4061 * We keep the big kernel semaphore locked, but we
4062 * clear ->lock_depth so that schedule() doesnt
4063 * auto-release the semaphore:
4065 saved_lock_depth
= task
->lock_depth
;
4066 task
->lock_depth
= -1;
4068 task
->lock_depth
= saved_lock_depth
;
4069 sub_preempt_count(PREEMPT_ACTIVE
);
4072 * Check again in case we missed a preemption opportunity
4073 * between schedule and now.
4076 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4078 EXPORT_SYMBOL(preempt_schedule
);
4081 * this is the entry point to schedule() from kernel preemption
4082 * off of irq context.
4083 * Note, that this is called and return with irqs disabled. This will
4084 * protect us against recursive calling from irq.
4086 asmlinkage
void __sched
preempt_schedule_irq(void)
4088 struct thread_info
*ti
= current_thread_info();
4089 struct task_struct
*task
= current
;
4090 int saved_lock_depth
;
4092 /* Catch callers which need to be fixed */
4093 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4096 add_preempt_count(PREEMPT_ACTIVE
);
4099 * We keep the big kernel semaphore locked, but we
4100 * clear ->lock_depth so that schedule() doesnt
4101 * auto-release the semaphore:
4103 saved_lock_depth
= task
->lock_depth
;
4104 task
->lock_depth
= -1;
4107 local_irq_disable();
4108 task
->lock_depth
= saved_lock_depth
;
4109 sub_preempt_count(PREEMPT_ACTIVE
);
4112 * Check again in case we missed a preemption opportunity
4113 * between schedule and now.
4116 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4119 #endif /* CONFIG_PREEMPT */
4121 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4124 return try_to_wake_up(curr
->private, mode
, sync
);
4126 EXPORT_SYMBOL(default_wake_function
);
4129 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4130 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4131 * number) then we wake all the non-exclusive tasks and one exclusive task.
4133 * There are circumstances in which we can try to wake a task which has already
4134 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4135 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4137 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4138 int nr_exclusive
, int sync
, void *key
)
4140 wait_queue_t
*curr
, *next
;
4142 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4143 unsigned flags
= curr
->flags
;
4145 if (curr
->func(curr
, mode
, sync
, key
) &&
4146 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4152 * __wake_up - wake up threads blocked on a waitqueue.
4154 * @mode: which threads
4155 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4156 * @key: is directly passed to the wakeup function
4158 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4159 int nr_exclusive
, void *key
)
4161 unsigned long flags
;
4163 spin_lock_irqsave(&q
->lock
, flags
);
4164 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4165 spin_unlock_irqrestore(&q
->lock
, flags
);
4167 EXPORT_SYMBOL(__wake_up
);
4170 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4172 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4174 __wake_up_common(q
, mode
, 1, 0, NULL
);
4178 * __wake_up_sync - wake up threads blocked on a waitqueue.
4180 * @mode: which threads
4181 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4183 * The sync wakeup differs that the waker knows that it will schedule
4184 * away soon, so while the target thread will be woken up, it will not
4185 * be migrated to another CPU - ie. the two threads are 'synchronized'
4186 * with each other. This can prevent needless bouncing between CPUs.
4188 * On UP it can prevent extra preemption.
4191 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4193 unsigned long flags
;
4199 if (unlikely(!nr_exclusive
))
4202 spin_lock_irqsave(&q
->lock
, flags
);
4203 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4204 spin_unlock_irqrestore(&q
->lock
, flags
);
4206 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4208 void complete(struct completion
*x
)
4210 unsigned long flags
;
4212 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4214 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4215 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4217 EXPORT_SYMBOL(complete
);
4219 void complete_all(struct completion
*x
)
4221 unsigned long flags
;
4223 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4224 x
->done
+= UINT_MAX
/2;
4225 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4226 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4228 EXPORT_SYMBOL(complete_all
);
4230 static inline long __sched
4231 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4234 DECLARE_WAITQUEUE(wait
, current
);
4236 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4237 __add_wait_queue_tail(&x
->wait
, &wait
);
4239 if ((state
== TASK_INTERRUPTIBLE
&&
4240 signal_pending(current
)) ||
4241 (state
== TASK_KILLABLE
&&
4242 fatal_signal_pending(current
))) {
4243 __remove_wait_queue(&x
->wait
, &wait
);
4244 return -ERESTARTSYS
;
4246 __set_current_state(state
);
4247 spin_unlock_irq(&x
->wait
.lock
);
4248 timeout
= schedule_timeout(timeout
);
4249 spin_lock_irq(&x
->wait
.lock
);
4251 __remove_wait_queue(&x
->wait
, &wait
);
4255 __remove_wait_queue(&x
->wait
, &wait
);
4262 wait_for_common(struct completion
*x
, long timeout
, int state
)
4266 spin_lock_irq(&x
->wait
.lock
);
4267 timeout
= do_wait_for_common(x
, timeout
, state
);
4268 spin_unlock_irq(&x
->wait
.lock
);
4272 void __sched
wait_for_completion(struct completion
*x
)
4274 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4276 EXPORT_SYMBOL(wait_for_completion
);
4278 unsigned long __sched
4279 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4281 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4283 EXPORT_SYMBOL(wait_for_completion_timeout
);
4285 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4287 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4288 if (t
== -ERESTARTSYS
)
4292 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4294 unsigned long __sched
4295 wait_for_completion_interruptible_timeout(struct completion
*x
,
4296 unsigned long timeout
)
4298 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4300 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4302 int __sched
wait_for_completion_killable(struct completion
*x
)
4304 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4305 if (t
== -ERESTARTSYS
)
4309 EXPORT_SYMBOL(wait_for_completion_killable
);
4312 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4314 unsigned long flags
;
4317 init_waitqueue_entry(&wait
, current
);
4319 __set_current_state(state
);
4321 spin_lock_irqsave(&q
->lock
, flags
);
4322 __add_wait_queue(q
, &wait
);
4323 spin_unlock(&q
->lock
);
4324 timeout
= schedule_timeout(timeout
);
4325 spin_lock_irq(&q
->lock
);
4326 __remove_wait_queue(q
, &wait
);
4327 spin_unlock_irqrestore(&q
->lock
, flags
);
4332 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4334 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4336 EXPORT_SYMBOL(interruptible_sleep_on
);
4339 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4341 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4343 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4345 void __sched
sleep_on(wait_queue_head_t
*q
)
4347 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4349 EXPORT_SYMBOL(sleep_on
);
4351 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4353 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4355 EXPORT_SYMBOL(sleep_on_timeout
);
4357 #ifdef CONFIG_RT_MUTEXES
4360 * rt_mutex_setprio - set the current priority of a task
4362 * @prio: prio value (kernel-internal form)
4364 * This function changes the 'effective' priority of a task. It does
4365 * not touch ->normal_prio like __setscheduler().
4367 * Used by the rt_mutex code to implement priority inheritance logic.
4369 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4371 unsigned long flags
;
4372 int oldprio
, on_rq
, running
;
4374 const struct sched_class
*prev_class
= p
->sched_class
;
4376 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4378 rq
= task_rq_lock(p
, &flags
);
4379 update_rq_clock(rq
);
4382 on_rq
= p
->se
.on_rq
;
4383 running
= task_current(rq
, p
);
4385 dequeue_task(rq
, p
, 0);
4387 p
->sched_class
->put_prev_task(rq
, p
);
4390 p
->sched_class
= &rt_sched_class
;
4392 p
->sched_class
= &fair_sched_class
;
4397 p
->sched_class
->set_curr_task(rq
);
4399 enqueue_task(rq
, p
, 0);
4401 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4403 task_rq_unlock(rq
, &flags
);
4408 void set_user_nice(struct task_struct
*p
, long nice
)
4410 int old_prio
, delta
, on_rq
;
4411 unsigned long flags
;
4414 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4417 * We have to be careful, if called from sys_setpriority(),
4418 * the task might be in the middle of scheduling on another CPU.
4420 rq
= task_rq_lock(p
, &flags
);
4421 update_rq_clock(rq
);
4423 * The RT priorities are set via sched_setscheduler(), but we still
4424 * allow the 'normal' nice value to be set - but as expected
4425 * it wont have any effect on scheduling until the task is
4426 * SCHED_FIFO/SCHED_RR:
4428 if (task_has_rt_policy(p
)) {
4429 p
->static_prio
= NICE_TO_PRIO(nice
);
4432 on_rq
= p
->se
.on_rq
;
4434 dequeue_task(rq
, p
, 0);
4438 p
->static_prio
= NICE_TO_PRIO(nice
);
4441 p
->prio
= effective_prio(p
);
4442 delta
= p
->prio
- old_prio
;
4445 enqueue_task(rq
, p
, 0);
4448 * If the task increased its priority or is running and
4449 * lowered its priority, then reschedule its CPU:
4451 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4452 resched_task(rq
->curr
);
4455 task_rq_unlock(rq
, &flags
);
4457 EXPORT_SYMBOL(set_user_nice
);
4460 * can_nice - check if a task can reduce its nice value
4464 int can_nice(const struct task_struct
*p
, const int nice
)
4466 /* convert nice value [19,-20] to rlimit style value [1,40] */
4467 int nice_rlim
= 20 - nice
;
4469 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4470 capable(CAP_SYS_NICE
));
4473 #ifdef __ARCH_WANT_SYS_NICE
4476 * sys_nice - change the priority of the current process.
4477 * @increment: priority increment
4479 * sys_setpriority is a more generic, but much slower function that
4480 * does similar things.
4482 asmlinkage
long sys_nice(int increment
)
4487 * Setpriority might change our priority at the same moment.
4488 * We don't have to worry. Conceptually one call occurs first
4489 * and we have a single winner.
4491 if (increment
< -40)
4496 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4502 if (increment
< 0 && !can_nice(current
, nice
))
4505 retval
= security_task_setnice(current
, nice
);
4509 set_user_nice(current
, nice
);
4516 * task_prio - return the priority value of a given task.
4517 * @p: the task in question.
4519 * This is the priority value as seen by users in /proc.
4520 * RT tasks are offset by -200. Normal tasks are centered
4521 * around 0, value goes from -16 to +15.
4523 int task_prio(const struct task_struct
*p
)
4525 return p
->prio
- MAX_RT_PRIO
;
4529 * task_nice - return the nice value of a given task.
4530 * @p: the task in question.
4532 int task_nice(const struct task_struct
*p
)
4534 return TASK_NICE(p
);
4536 EXPORT_SYMBOL(task_nice
);
4539 * idle_cpu - is a given cpu idle currently?
4540 * @cpu: the processor in question.
4542 int idle_cpu(int cpu
)
4544 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4548 * idle_task - return the idle task for a given cpu.
4549 * @cpu: the processor in question.
4551 struct task_struct
*idle_task(int cpu
)
4553 return cpu_rq(cpu
)->idle
;
4557 * find_process_by_pid - find a process with a matching PID value.
4558 * @pid: the pid in question.
4560 static struct task_struct
*find_process_by_pid(pid_t pid
)
4562 return pid
? find_task_by_vpid(pid
) : current
;
4565 /* Actually do priority change: must hold rq lock. */
4567 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4569 BUG_ON(p
->se
.on_rq
);
4572 switch (p
->policy
) {
4576 p
->sched_class
= &fair_sched_class
;
4580 p
->sched_class
= &rt_sched_class
;
4584 p
->rt_priority
= prio
;
4585 p
->normal_prio
= normal_prio(p
);
4586 /* we are holding p->pi_lock already */
4587 p
->prio
= rt_mutex_getprio(p
);
4592 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4593 * @p: the task in question.
4594 * @policy: new policy.
4595 * @param: structure containing the new RT priority.
4597 * NOTE that the task may be already dead.
4599 int sched_setscheduler(struct task_struct
*p
, int policy
,
4600 struct sched_param
*param
)
4602 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4603 unsigned long flags
;
4604 const struct sched_class
*prev_class
= p
->sched_class
;
4607 /* may grab non-irq protected spin_locks */
4608 BUG_ON(in_interrupt());
4610 /* double check policy once rq lock held */
4612 policy
= oldpolicy
= p
->policy
;
4613 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4614 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4615 policy
!= SCHED_IDLE
)
4618 * Valid priorities for SCHED_FIFO and SCHED_RR are
4619 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4620 * SCHED_BATCH and SCHED_IDLE is 0.
4622 if (param
->sched_priority
< 0 ||
4623 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4624 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4626 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4630 * Allow unprivileged RT tasks to decrease priority:
4632 if (!capable(CAP_SYS_NICE
)) {
4633 if (rt_policy(policy
)) {
4634 unsigned long rlim_rtprio
;
4636 if (!lock_task_sighand(p
, &flags
))
4638 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4639 unlock_task_sighand(p
, &flags
);
4641 /* can't set/change the rt policy */
4642 if (policy
!= p
->policy
&& !rlim_rtprio
)
4645 /* can't increase priority */
4646 if (param
->sched_priority
> p
->rt_priority
&&
4647 param
->sched_priority
> rlim_rtprio
)
4651 * Like positive nice levels, dont allow tasks to
4652 * move out of SCHED_IDLE either:
4654 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4657 /* can't change other user's priorities */
4658 if ((current
->euid
!= p
->euid
) &&
4659 (current
->euid
!= p
->uid
))
4663 #ifdef CONFIG_RT_GROUP_SCHED
4665 * Do not allow realtime tasks into groups that have no runtime
4668 if (rt_policy(policy
) && task_group(p
)->rt_runtime
== 0)
4672 retval
= security_task_setscheduler(p
, policy
, param
);
4676 * make sure no PI-waiters arrive (or leave) while we are
4677 * changing the priority of the task:
4679 spin_lock_irqsave(&p
->pi_lock
, flags
);
4681 * To be able to change p->policy safely, the apropriate
4682 * runqueue lock must be held.
4684 rq
= __task_rq_lock(p
);
4685 /* recheck policy now with rq lock held */
4686 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4687 policy
= oldpolicy
= -1;
4688 __task_rq_unlock(rq
);
4689 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4692 update_rq_clock(rq
);
4693 on_rq
= p
->se
.on_rq
;
4694 running
= task_current(rq
, p
);
4696 deactivate_task(rq
, p
, 0);
4698 p
->sched_class
->put_prev_task(rq
, p
);
4701 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4704 p
->sched_class
->set_curr_task(rq
);
4706 activate_task(rq
, p
, 0);
4708 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4710 __task_rq_unlock(rq
);
4711 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4713 rt_mutex_adjust_pi(p
);
4717 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4720 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4722 struct sched_param lparam
;
4723 struct task_struct
*p
;
4726 if (!param
|| pid
< 0)
4728 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4733 p
= find_process_by_pid(pid
);
4735 retval
= sched_setscheduler(p
, policy
, &lparam
);
4742 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4743 * @pid: the pid in question.
4744 * @policy: new policy.
4745 * @param: structure containing the new RT priority.
4748 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4750 /* negative values for policy are not valid */
4754 return do_sched_setscheduler(pid
, policy
, param
);
4758 * sys_sched_setparam - set/change the RT priority of a thread
4759 * @pid: the pid in question.
4760 * @param: structure containing the new RT priority.
4762 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4764 return do_sched_setscheduler(pid
, -1, param
);
4768 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4769 * @pid: the pid in question.
4771 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4773 struct task_struct
*p
;
4780 read_lock(&tasklist_lock
);
4781 p
= find_process_by_pid(pid
);
4783 retval
= security_task_getscheduler(p
);
4787 read_unlock(&tasklist_lock
);
4792 * sys_sched_getscheduler - get the RT priority of a thread
4793 * @pid: the pid in question.
4794 * @param: structure containing the RT priority.
4796 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4798 struct sched_param lp
;
4799 struct task_struct
*p
;
4802 if (!param
|| pid
< 0)
4805 read_lock(&tasklist_lock
);
4806 p
= find_process_by_pid(pid
);
4811 retval
= security_task_getscheduler(p
);
4815 lp
.sched_priority
= p
->rt_priority
;
4816 read_unlock(&tasklist_lock
);
4819 * This one might sleep, we cannot do it with a spinlock held ...
4821 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4826 read_unlock(&tasklist_lock
);
4830 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4832 cpumask_t cpus_allowed
;
4833 struct task_struct
*p
;
4837 read_lock(&tasklist_lock
);
4839 p
= find_process_by_pid(pid
);
4841 read_unlock(&tasklist_lock
);
4847 * It is not safe to call set_cpus_allowed with the
4848 * tasklist_lock held. We will bump the task_struct's
4849 * usage count and then drop tasklist_lock.
4852 read_unlock(&tasklist_lock
);
4855 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4856 !capable(CAP_SYS_NICE
))
4859 retval
= security_task_setscheduler(p
, 0, NULL
);
4863 cpus_allowed
= cpuset_cpus_allowed(p
);
4864 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4866 retval
= set_cpus_allowed(p
, new_mask
);
4869 cpus_allowed
= cpuset_cpus_allowed(p
);
4870 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4872 * We must have raced with a concurrent cpuset
4873 * update. Just reset the cpus_allowed to the
4874 * cpuset's cpus_allowed
4876 new_mask
= cpus_allowed
;
4886 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4887 cpumask_t
*new_mask
)
4889 if (len
< sizeof(cpumask_t
)) {
4890 memset(new_mask
, 0, sizeof(cpumask_t
));
4891 } else if (len
> sizeof(cpumask_t
)) {
4892 len
= sizeof(cpumask_t
);
4894 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4898 * sys_sched_setaffinity - set the cpu affinity of a process
4899 * @pid: pid of the process
4900 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4901 * @user_mask_ptr: user-space pointer to the new cpu mask
4903 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4904 unsigned long __user
*user_mask_ptr
)
4909 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4913 return sched_setaffinity(pid
, new_mask
);
4917 * Represents all cpu's present in the system
4918 * In systems capable of hotplug, this map could dynamically grow
4919 * as new cpu's are detected in the system via any platform specific
4920 * method, such as ACPI for e.g.
4923 cpumask_t cpu_present_map __read_mostly
;
4924 EXPORT_SYMBOL(cpu_present_map
);
4927 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4928 EXPORT_SYMBOL(cpu_online_map
);
4930 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4931 EXPORT_SYMBOL(cpu_possible_map
);
4934 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4936 struct task_struct
*p
;
4940 read_lock(&tasklist_lock
);
4943 p
= find_process_by_pid(pid
);
4947 retval
= security_task_getscheduler(p
);
4951 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4954 read_unlock(&tasklist_lock
);
4961 * sys_sched_getaffinity - get the cpu affinity of a process
4962 * @pid: pid of the process
4963 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4964 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4966 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4967 unsigned long __user
*user_mask_ptr
)
4972 if (len
< sizeof(cpumask_t
))
4975 ret
= sched_getaffinity(pid
, &mask
);
4979 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4982 return sizeof(cpumask_t
);
4986 * sys_sched_yield - yield the current processor to other threads.
4988 * This function yields the current CPU to other tasks. If there are no
4989 * other threads running on this CPU then this function will return.
4991 asmlinkage
long sys_sched_yield(void)
4993 struct rq
*rq
= this_rq_lock();
4995 schedstat_inc(rq
, yld_count
);
4996 current
->sched_class
->yield_task(rq
);
4999 * Since we are going to call schedule() anyway, there's
5000 * no need to preempt or enable interrupts:
5002 __release(rq
->lock
);
5003 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5004 _raw_spin_unlock(&rq
->lock
);
5005 preempt_enable_no_resched();
5012 static void __cond_resched(void)
5014 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5015 __might_sleep(__FILE__
, __LINE__
);
5018 * The BKS might be reacquired before we have dropped
5019 * PREEMPT_ACTIVE, which could trigger a second
5020 * cond_resched() call.
5023 add_preempt_count(PREEMPT_ACTIVE
);
5025 sub_preempt_count(PREEMPT_ACTIVE
);
5026 } while (need_resched());
5029 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5030 int __sched
_cond_resched(void)
5032 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5033 system_state
== SYSTEM_RUNNING
) {
5039 EXPORT_SYMBOL(_cond_resched
);
5043 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5044 * call schedule, and on return reacquire the lock.
5046 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5047 * operations here to prevent schedule() from being called twice (once via
5048 * spin_unlock(), once by hand).
5050 int cond_resched_lock(spinlock_t
*lock
)
5052 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5055 if (spin_needbreak(lock
) || resched
) {
5057 if (resched
&& need_resched())
5066 EXPORT_SYMBOL(cond_resched_lock
);
5068 int __sched
cond_resched_softirq(void)
5070 BUG_ON(!in_softirq());
5072 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5080 EXPORT_SYMBOL(cond_resched_softirq
);
5083 * yield - yield the current processor to other threads.
5085 * This is a shortcut for kernel-space yielding - it marks the
5086 * thread runnable and calls sys_sched_yield().
5088 void __sched
yield(void)
5090 set_current_state(TASK_RUNNING
);
5093 EXPORT_SYMBOL(yield
);
5096 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5097 * that process accounting knows that this is a task in IO wait state.
5099 * But don't do that if it is a deliberate, throttling IO wait (this task
5100 * has set its backing_dev_info: the queue against which it should throttle)
5102 void __sched
io_schedule(void)
5104 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5106 delayacct_blkio_start();
5107 atomic_inc(&rq
->nr_iowait
);
5109 atomic_dec(&rq
->nr_iowait
);
5110 delayacct_blkio_end();
5112 EXPORT_SYMBOL(io_schedule
);
5114 long __sched
io_schedule_timeout(long timeout
)
5116 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5119 delayacct_blkio_start();
5120 atomic_inc(&rq
->nr_iowait
);
5121 ret
= schedule_timeout(timeout
);
5122 atomic_dec(&rq
->nr_iowait
);
5123 delayacct_blkio_end();
5128 * sys_sched_get_priority_max - return maximum RT priority.
5129 * @policy: scheduling class.
5131 * this syscall returns the maximum rt_priority that can be used
5132 * by a given scheduling class.
5134 asmlinkage
long sys_sched_get_priority_max(int policy
)
5141 ret
= MAX_USER_RT_PRIO
-1;
5153 * sys_sched_get_priority_min - return minimum RT priority.
5154 * @policy: scheduling class.
5156 * this syscall returns the minimum rt_priority that can be used
5157 * by a given scheduling class.
5159 asmlinkage
long sys_sched_get_priority_min(int policy
)
5177 * sys_sched_rr_get_interval - return the default timeslice of a process.
5178 * @pid: pid of the process.
5179 * @interval: userspace pointer to the timeslice value.
5181 * this syscall writes the default timeslice value of a given process
5182 * into the user-space timespec buffer. A value of '0' means infinity.
5185 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5187 struct task_struct
*p
;
5188 unsigned int time_slice
;
5196 read_lock(&tasklist_lock
);
5197 p
= find_process_by_pid(pid
);
5201 retval
= security_task_getscheduler(p
);
5206 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5207 * tasks that are on an otherwise idle runqueue:
5210 if (p
->policy
== SCHED_RR
) {
5211 time_slice
= DEF_TIMESLICE
;
5212 } else if (p
->policy
!= SCHED_FIFO
) {
5213 struct sched_entity
*se
= &p
->se
;
5214 unsigned long flags
;
5217 rq
= task_rq_lock(p
, &flags
);
5218 if (rq
->cfs
.load
.weight
)
5219 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5220 task_rq_unlock(rq
, &flags
);
5222 read_unlock(&tasklist_lock
);
5223 jiffies_to_timespec(time_slice
, &t
);
5224 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5228 read_unlock(&tasklist_lock
);
5232 static const char stat_nam
[] = "RSDTtZX";
5234 void sched_show_task(struct task_struct
*p
)
5236 unsigned long free
= 0;
5239 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5240 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5241 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5242 #if BITS_PER_LONG == 32
5243 if (state
== TASK_RUNNING
)
5244 printk(KERN_CONT
" running ");
5246 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5248 if (state
== TASK_RUNNING
)
5249 printk(KERN_CONT
" running task ");
5251 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5253 #ifdef CONFIG_DEBUG_STACK_USAGE
5255 unsigned long *n
= end_of_stack(p
);
5258 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5261 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5262 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5264 show_stack(p
, NULL
);
5267 void show_state_filter(unsigned long state_filter
)
5269 struct task_struct
*g
, *p
;
5271 #if BITS_PER_LONG == 32
5273 " task PC stack pid father\n");
5276 " task PC stack pid father\n");
5278 read_lock(&tasklist_lock
);
5279 do_each_thread(g
, p
) {
5281 * reset the NMI-timeout, listing all files on a slow
5282 * console might take alot of time:
5284 touch_nmi_watchdog();
5285 if (!state_filter
|| (p
->state
& state_filter
))
5287 } while_each_thread(g
, p
);
5289 touch_all_softlockup_watchdogs();
5291 #ifdef CONFIG_SCHED_DEBUG
5292 sysrq_sched_debug_show();
5294 read_unlock(&tasklist_lock
);
5296 * Only show locks if all tasks are dumped:
5298 if (state_filter
== -1)
5299 debug_show_all_locks();
5302 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5304 idle
->sched_class
= &idle_sched_class
;
5308 * init_idle - set up an idle thread for a given CPU
5309 * @idle: task in question
5310 * @cpu: cpu the idle task belongs to
5312 * NOTE: this function does not set the idle thread's NEED_RESCHED
5313 * flag, to make booting more robust.
5315 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5317 struct rq
*rq
= cpu_rq(cpu
);
5318 unsigned long flags
;
5321 idle
->se
.exec_start
= sched_clock();
5323 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5324 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5325 __set_task_cpu(idle
, cpu
);
5327 spin_lock_irqsave(&rq
->lock
, flags
);
5328 rq
->curr
= rq
->idle
= idle
;
5329 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5332 spin_unlock_irqrestore(&rq
->lock
, flags
);
5334 /* Set the preempt count _outside_ the spinlocks! */
5335 task_thread_info(idle
)->preempt_count
= 0;
5338 * The idle tasks have their own, simple scheduling class:
5340 idle
->sched_class
= &idle_sched_class
;
5344 * In a system that switches off the HZ timer nohz_cpu_mask
5345 * indicates which cpus entered this state. This is used
5346 * in the rcu update to wait only for active cpus. For system
5347 * which do not switch off the HZ timer nohz_cpu_mask should
5348 * always be CPU_MASK_NONE.
5350 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5353 * Increase the granularity value when there are more CPUs,
5354 * because with more CPUs the 'effective latency' as visible
5355 * to users decreases. But the relationship is not linear,
5356 * so pick a second-best guess by going with the log2 of the
5359 * This idea comes from the SD scheduler of Con Kolivas:
5361 static inline void sched_init_granularity(void)
5363 unsigned int factor
= 1 + ilog2(num_online_cpus());
5364 const unsigned long limit
= 200000000;
5366 sysctl_sched_min_granularity
*= factor
;
5367 if (sysctl_sched_min_granularity
> limit
)
5368 sysctl_sched_min_granularity
= limit
;
5370 sysctl_sched_latency
*= factor
;
5371 if (sysctl_sched_latency
> limit
)
5372 sysctl_sched_latency
= limit
;
5374 sysctl_sched_wakeup_granularity
*= factor
;
5375 sysctl_sched_batch_wakeup_granularity
*= factor
;
5380 * This is how migration works:
5382 * 1) we queue a struct migration_req structure in the source CPU's
5383 * runqueue and wake up that CPU's migration thread.
5384 * 2) we down() the locked semaphore => thread blocks.
5385 * 3) migration thread wakes up (implicitly it forces the migrated
5386 * thread off the CPU)
5387 * 4) it gets the migration request and checks whether the migrated
5388 * task is still in the wrong runqueue.
5389 * 5) if it's in the wrong runqueue then the migration thread removes
5390 * it and puts it into the right queue.
5391 * 6) migration thread up()s the semaphore.
5392 * 7) we wake up and the migration is done.
5396 * Change a given task's CPU affinity. Migrate the thread to a
5397 * proper CPU and schedule it away if the CPU it's executing on
5398 * is removed from the allowed bitmask.
5400 * NOTE: the caller must have a valid reference to the task, the
5401 * task must not exit() & deallocate itself prematurely. The
5402 * call is not atomic; no spinlocks may be held.
5404 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5406 struct migration_req req
;
5407 unsigned long flags
;
5411 rq
= task_rq_lock(p
, &flags
);
5412 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5417 if (p
->sched_class
->set_cpus_allowed
)
5418 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5420 p
->cpus_allowed
= new_mask
;
5421 p
->rt
.nr_cpus_allowed
= cpus_weight(new_mask
);
5424 /* Can the task run on the task's current CPU? If so, we're done */
5425 if (cpu_isset(task_cpu(p
), new_mask
))
5428 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5429 /* Need help from migration thread: drop lock and wait. */
5430 task_rq_unlock(rq
, &flags
);
5431 wake_up_process(rq
->migration_thread
);
5432 wait_for_completion(&req
.done
);
5433 tlb_migrate_finish(p
->mm
);
5437 task_rq_unlock(rq
, &flags
);
5441 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5444 * Move (not current) task off this cpu, onto dest cpu. We're doing
5445 * this because either it can't run here any more (set_cpus_allowed()
5446 * away from this CPU, or CPU going down), or because we're
5447 * attempting to rebalance this task on exec (sched_exec).
5449 * So we race with normal scheduler movements, but that's OK, as long
5450 * as the task is no longer on this CPU.
5452 * Returns non-zero if task was successfully migrated.
5454 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5456 struct rq
*rq_dest
, *rq_src
;
5459 if (unlikely(cpu_is_offline(dest_cpu
)))
5462 rq_src
= cpu_rq(src_cpu
);
5463 rq_dest
= cpu_rq(dest_cpu
);
5465 double_rq_lock(rq_src
, rq_dest
);
5466 /* Already moved. */
5467 if (task_cpu(p
) != src_cpu
)
5469 /* Affinity changed (again). */
5470 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5473 on_rq
= p
->se
.on_rq
;
5475 deactivate_task(rq_src
, p
, 0);
5477 set_task_cpu(p
, dest_cpu
);
5479 activate_task(rq_dest
, p
, 0);
5480 check_preempt_curr(rq_dest
, p
);
5484 double_rq_unlock(rq_src
, rq_dest
);
5489 * migration_thread - this is a highprio system thread that performs
5490 * thread migration by bumping thread off CPU then 'pushing' onto
5493 static int migration_thread(void *data
)
5495 int cpu
= (long)data
;
5499 BUG_ON(rq
->migration_thread
!= current
);
5501 set_current_state(TASK_INTERRUPTIBLE
);
5502 while (!kthread_should_stop()) {
5503 struct migration_req
*req
;
5504 struct list_head
*head
;
5506 spin_lock_irq(&rq
->lock
);
5508 if (cpu_is_offline(cpu
)) {
5509 spin_unlock_irq(&rq
->lock
);
5513 if (rq
->active_balance
) {
5514 active_load_balance(rq
, cpu
);
5515 rq
->active_balance
= 0;
5518 head
= &rq
->migration_queue
;
5520 if (list_empty(head
)) {
5521 spin_unlock_irq(&rq
->lock
);
5523 set_current_state(TASK_INTERRUPTIBLE
);
5526 req
= list_entry(head
->next
, struct migration_req
, list
);
5527 list_del_init(head
->next
);
5529 spin_unlock(&rq
->lock
);
5530 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5533 complete(&req
->done
);
5535 __set_current_state(TASK_RUNNING
);
5539 /* Wait for kthread_stop */
5540 set_current_state(TASK_INTERRUPTIBLE
);
5541 while (!kthread_should_stop()) {
5543 set_current_state(TASK_INTERRUPTIBLE
);
5545 __set_current_state(TASK_RUNNING
);
5549 #ifdef CONFIG_HOTPLUG_CPU
5551 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5555 local_irq_disable();
5556 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5562 * Figure out where task on dead CPU should go, use force if necessary.
5563 * NOTE: interrupts should be disabled by the caller
5565 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5567 unsigned long flags
;
5574 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5575 cpus_and(mask
, mask
, p
->cpus_allowed
);
5576 dest_cpu
= any_online_cpu(mask
);
5578 /* On any allowed CPU? */
5579 if (dest_cpu
== NR_CPUS
)
5580 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5582 /* No more Mr. Nice Guy. */
5583 if (dest_cpu
== NR_CPUS
) {
5584 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5586 * Try to stay on the same cpuset, where the
5587 * current cpuset may be a subset of all cpus.
5588 * The cpuset_cpus_allowed_locked() variant of
5589 * cpuset_cpus_allowed() will not block. It must be
5590 * called within calls to cpuset_lock/cpuset_unlock.
5592 rq
= task_rq_lock(p
, &flags
);
5593 p
->cpus_allowed
= cpus_allowed
;
5594 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5595 task_rq_unlock(rq
, &flags
);
5598 * Don't tell them about moving exiting tasks or
5599 * kernel threads (both mm NULL), since they never
5602 if (p
->mm
&& printk_ratelimit()) {
5603 printk(KERN_INFO
"process %d (%s) no "
5604 "longer affine to cpu%d\n",
5605 task_pid_nr(p
), p
->comm
, dead_cpu
);
5608 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5612 * While a dead CPU has no uninterruptible tasks queued at this point,
5613 * it might still have a nonzero ->nr_uninterruptible counter, because
5614 * for performance reasons the counter is not stricly tracking tasks to
5615 * their home CPUs. So we just add the counter to another CPU's counter,
5616 * to keep the global sum constant after CPU-down:
5618 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5620 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5621 unsigned long flags
;
5623 local_irq_save(flags
);
5624 double_rq_lock(rq_src
, rq_dest
);
5625 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5626 rq_src
->nr_uninterruptible
= 0;
5627 double_rq_unlock(rq_src
, rq_dest
);
5628 local_irq_restore(flags
);
5631 /* Run through task list and migrate tasks from the dead cpu. */
5632 static void migrate_live_tasks(int src_cpu
)
5634 struct task_struct
*p
, *t
;
5636 read_lock(&tasklist_lock
);
5638 do_each_thread(t
, p
) {
5642 if (task_cpu(p
) == src_cpu
)
5643 move_task_off_dead_cpu(src_cpu
, p
);
5644 } while_each_thread(t
, p
);
5646 read_unlock(&tasklist_lock
);
5650 * Schedules idle task to be the next runnable task on current CPU.
5651 * It does so by boosting its priority to highest possible.
5652 * Used by CPU offline code.
5654 void sched_idle_next(void)
5656 int this_cpu
= smp_processor_id();
5657 struct rq
*rq
= cpu_rq(this_cpu
);
5658 struct task_struct
*p
= rq
->idle
;
5659 unsigned long flags
;
5661 /* cpu has to be offline */
5662 BUG_ON(cpu_online(this_cpu
));
5665 * Strictly not necessary since rest of the CPUs are stopped by now
5666 * and interrupts disabled on the current cpu.
5668 spin_lock_irqsave(&rq
->lock
, flags
);
5670 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5672 update_rq_clock(rq
);
5673 activate_task(rq
, p
, 0);
5675 spin_unlock_irqrestore(&rq
->lock
, flags
);
5679 * Ensures that the idle task is using init_mm right before its cpu goes
5682 void idle_task_exit(void)
5684 struct mm_struct
*mm
= current
->active_mm
;
5686 BUG_ON(cpu_online(smp_processor_id()));
5689 switch_mm(mm
, &init_mm
, current
);
5693 /* called under rq->lock with disabled interrupts */
5694 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5696 struct rq
*rq
= cpu_rq(dead_cpu
);
5698 /* Must be exiting, otherwise would be on tasklist. */
5699 BUG_ON(!p
->exit_state
);
5701 /* Cannot have done final schedule yet: would have vanished. */
5702 BUG_ON(p
->state
== TASK_DEAD
);
5707 * Drop lock around migration; if someone else moves it,
5708 * that's OK. No task can be added to this CPU, so iteration is
5711 spin_unlock_irq(&rq
->lock
);
5712 move_task_off_dead_cpu(dead_cpu
, p
);
5713 spin_lock_irq(&rq
->lock
);
5718 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5719 static void migrate_dead_tasks(unsigned int dead_cpu
)
5721 struct rq
*rq
= cpu_rq(dead_cpu
);
5722 struct task_struct
*next
;
5725 if (!rq
->nr_running
)
5727 update_rq_clock(rq
);
5728 next
= pick_next_task(rq
, rq
->curr
);
5731 migrate_dead(dead_cpu
, next
);
5735 #endif /* CONFIG_HOTPLUG_CPU */
5737 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5739 static struct ctl_table sd_ctl_dir
[] = {
5741 .procname
= "sched_domain",
5747 static struct ctl_table sd_ctl_root
[] = {
5749 .ctl_name
= CTL_KERN
,
5750 .procname
= "kernel",
5752 .child
= sd_ctl_dir
,
5757 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5759 struct ctl_table
*entry
=
5760 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5765 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5767 struct ctl_table
*entry
;
5770 * In the intermediate directories, both the child directory and
5771 * procname are dynamically allocated and could fail but the mode
5772 * will always be set. In the lowest directory the names are
5773 * static strings and all have proc handlers.
5775 for (entry
= *tablep
; entry
->mode
; entry
++) {
5777 sd_free_ctl_entry(&entry
->child
);
5778 if (entry
->proc_handler
== NULL
)
5779 kfree(entry
->procname
);
5787 set_table_entry(struct ctl_table
*entry
,
5788 const char *procname
, void *data
, int maxlen
,
5789 mode_t mode
, proc_handler
*proc_handler
)
5791 entry
->procname
= procname
;
5793 entry
->maxlen
= maxlen
;
5795 entry
->proc_handler
= proc_handler
;
5798 static struct ctl_table
*
5799 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5801 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5806 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5807 sizeof(long), 0644, proc_doulongvec_minmax
);
5808 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5809 sizeof(long), 0644, proc_doulongvec_minmax
);
5810 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5811 sizeof(int), 0644, proc_dointvec_minmax
);
5812 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5813 sizeof(int), 0644, proc_dointvec_minmax
);
5814 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5815 sizeof(int), 0644, proc_dointvec_minmax
);
5816 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5817 sizeof(int), 0644, proc_dointvec_minmax
);
5818 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5819 sizeof(int), 0644, proc_dointvec_minmax
);
5820 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5821 sizeof(int), 0644, proc_dointvec_minmax
);
5822 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5823 sizeof(int), 0644, proc_dointvec_minmax
);
5824 set_table_entry(&table
[9], "cache_nice_tries",
5825 &sd
->cache_nice_tries
,
5826 sizeof(int), 0644, proc_dointvec_minmax
);
5827 set_table_entry(&table
[10], "flags", &sd
->flags
,
5828 sizeof(int), 0644, proc_dointvec_minmax
);
5829 /* &table[11] is terminator */
5834 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5836 struct ctl_table
*entry
, *table
;
5837 struct sched_domain
*sd
;
5838 int domain_num
= 0, i
;
5841 for_each_domain(cpu
, sd
)
5843 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5848 for_each_domain(cpu
, sd
) {
5849 snprintf(buf
, 32, "domain%d", i
);
5850 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5852 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5859 static struct ctl_table_header
*sd_sysctl_header
;
5860 static void register_sched_domain_sysctl(void)
5862 int i
, cpu_num
= num_online_cpus();
5863 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5866 WARN_ON(sd_ctl_dir
[0].child
);
5867 sd_ctl_dir
[0].child
= entry
;
5872 for_each_online_cpu(i
) {
5873 snprintf(buf
, 32, "cpu%d", i
);
5874 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5876 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5880 WARN_ON(sd_sysctl_header
);
5881 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5884 /* may be called multiple times per register */
5885 static void unregister_sched_domain_sysctl(void)
5887 if (sd_sysctl_header
)
5888 unregister_sysctl_table(sd_sysctl_header
);
5889 sd_sysctl_header
= NULL
;
5890 if (sd_ctl_dir
[0].child
)
5891 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5894 static void register_sched_domain_sysctl(void)
5897 static void unregister_sched_domain_sysctl(void)
5903 * migration_call - callback that gets triggered when a CPU is added.
5904 * Here we can start up the necessary migration thread for the new CPU.
5906 static int __cpuinit
5907 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5909 struct task_struct
*p
;
5910 int cpu
= (long)hcpu
;
5911 unsigned long flags
;
5916 case CPU_UP_PREPARE
:
5917 case CPU_UP_PREPARE_FROZEN
:
5918 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5921 kthread_bind(p
, cpu
);
5922 /* Must be high prio: stop_machine expects to yield to it. */
5923 rq
= task_rq_lock(p
, &flags
);
5924 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5925 task_rq_unlock(rq
, &flags
);
5926 cpu_rq(cpu
)->migration_thread
= p
;
5930 case CPU_ONLINE_FROZEN
:
5931 /* Strictly unnecessary, as first user will wake it. */
5932 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5934 /* Update our root-domain */
5936 spin_lock_irqsave(&rq
->lock
, flags
);
5938 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5939 cpu_set(cpu
, rq
->rd
->online
);
5941 spin_unlock_irqrestore(&rq
->lock
, flags
);
5944 #ifdef CONFIG_HOTPLUG_CPU
5945 case CPU_UP_CANCELED
:
5946 case CPU_UP_CANCELED_FROZEN
:
5947 if (!cpu_rq(cpu
)->migration_thread
)
5949 /* Unbind it from offline cpu so it can run. Fall thru. */
5950 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5951 any_online_cpu(cpu_online_map
));
5952 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5953 cpu_rq(cpu
)->migration_thread
= NULL
;
5957 case CPU_DEAD_FROZEN
:
5958 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5959 migrate_live_tasks(cpu
);
5961 kthread_stop(rq
->migration_thread
);
5962 rq
->migration_thread
= NULL
;
5963 /* Idle task back to normal (off runqueue, low prio) */
5964 spin_lock_irq(&rq
->lock
);
5965 update_rq_clock(rq
);
5966 deactivate_task(rq
, rq
->idle
, 0);
5967 rq
->idle
->static_prio
= MAX_PRIO
;
5968 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5969 rq
->idle
->sched_class
= &idle_sched_class
;
5970 migrate_dead_tasks(cpu
);
5971 spin_unlock_irq(&rq
->lock
);
5973 migrate_nr_uninterruptible(rq
);
5974 BUG_ON(rq
->nr_running
!= 0);
5977 * No need to migrate the tasks: it was best-effort if
5978 * they didn't take sched_hotcpu_mutex. Just wake up
5981 spin_lock_irq(&rq
->lock
);
5982 while (!list_empty(&rq
->migration_queue
)) {
5983 struct migration_req
*req
;
5985 req
= list_entry(rq
->migration_queue
.next
,
5986 struct migration_req
, list
);
5987 list_del_init(&req
->list
);
5988 complete(&req
->done
);
5990 spin_unlock_irq(&rq
->lock
);
5994 case CPU_DYING_FROZEN
:
5995 /* Update our root-domain */
5997 spin_lock_irqsave(&rq
->lock
, flags
);
5999 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6000 cpu_clear(cpu
, rq
->rd
->online
);
6002 spin_unlock_irqrestore(&rq
->lock
, flags
);
6009 /* Register at highest priority so that task migration (migrate_all_tasks)
6010 * happens before everything else.
6012 static struct notifier_block __cpuinitdata migration_notifier
= {
6013 .notifier_call
= migration_call
,
6017 void __init
migration_init(void)
6019 void *cpu
= (void *)(long)smp_processor_id();
6022 /* Start one for the boot CPU: */
6023 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6024 BUG_ON(err
== NOTIFY_BAD
);
6025 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6026 register_cpu_notifier(&migration_notifier
);
6032 /* Number of possible processor ids */
6033 int nr_cpu_ids __read_mostly
= NR_CPUS
;
6034 EXPORT_SYMBOL(nr_cpu_ids
);
6036 #ifdef CONFIG_SCHED_DEBUG
6038 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
6040 struct sched_group
*group
= sd
->groups
;
6041 cpumask_t groupmask
;
6044 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
6045 cpus_clear(groupmask
);
6047 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6049 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6050 printk("does not load-balance\n");
6052 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6057 printk(KERN_CONT
"span %s\n", str
);
6059 if (!cpu_isset(cpu
, sd
->span
)) {
6060 printk(KERN_ERR
"ERROR: domain->span does not contain "
6063 if (!cpu_isset(cpu
, group
->cpumask
)) {
6064 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6068 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6072 printk(KERN_ERR
"ERROR: group is NULL\n");
6076 if (!group
->__cpu_power
) {
6077 printk(KERN_CONT
"\n");
6078 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6083 if (!cpus_weight(group
->cpumask
)) {
6084 printk(KERN_CONT
"\n");
6085 printk(KERN_ERR
"ERROR: empty group\n");
6089 if (cpus_intersects(groupmask
, group
->cpumask
)) {
6090 printk(KERN_CONT
"\n");
6091 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6095 cpus_or(groupmask
, groupmask
, group
->cpumask
);
6097 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
6098 printk(KERN_CONT
" %s", str
);
6100 group
= group
->next
;
6101 } while (group
!= sd
->groups
);
6102 printk(KERN_CONT
"\n");
6104 if (!cpus_equal(sd
->span
, groupmask
))
6105 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6107 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
6108 printk(KERN_ERR
"ERROR: parent span is not a superset "
6109 "of domain->span\n");
6113 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6118 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6122 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6125 if (sched_domain_debug_one(sd
, cpu
, level
))
6134 # define sched_domain_debug(sd, cpu) do { } while (0)
6137 static int sd_degenerate(struct sched_domain
*sd
)
6139 if (cpus_weight(sd
->span
) == 1)
6142 /* Following flags need at least 2 groups */
6143 if (sd
->flags
& (SD_LOAD_BALANCE
|
6144 SD_BALANCE_NEWIDLE
|
6148 SD_SHARE_PKG_RESOURCES
)) {
6149 if (sd
->groups
!= sd
->groups
->next
)
6153 /* Following flags don't use groups */
6154 if (sd
->flags
& (SD_WAKE_IDLE
|
6163 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6165 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6167 if (sd_degenerate(parent
))
6170 if (!cpus_equal(sd
->span
, parent
->span
))
6173 /* Does parent contain flags not in child? */
6174 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6175 if (cflags
& SD_WAKE_AFFINE
)
6176 pflags
&= ~SD_WAKE_BALANCE
;
6177 /* Flags needing groups don't count if only 1 group in parent */
6178 if (parent
->groups
== parent
->groups
->next
) {
6179 pflags
&= ~(SD_LOAD_BALANCE
|
6180 SD_BALANCE_NEWIDLE
|
6184 SD_SHARE_PKG_RESOURCES
);
6186 if (~cflags
& pflags
)
6192 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6194 unsigned long flags
;
6195 const struct sched_class
*class;
6197 spin_lock_irqsave(&rq
->lock
, flags
);
6200 struct root_domain
*old_rd
= rq
->rd
;
6202 for (class = sched_class_highest
; class; class = class->next
) {
6203 if (class->leave_domain
)
6204 class->leave_domain(rq
);
6207 cpu_clear(rq
->cpu
, old_rd
->span
);
6208 cpu_clear(rq
->cpu
, old_rd
->online
);
6210 if (atomic_dec_and_test(&old_rd
->refcount
))
6214 atomic_inc(&rd
->refcount
);
6217 cpu_set(rq
->cpu
, rd
->span
);
6218 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6219 cpu_set(rq
->cpu
, rd
->online
);
6221 for (class = sched_class_highest
; class; class = class->next
) {
6222 if (class->join_domain
)
6223 class->join_domain(rq
);
6226 spin_unlock_irqrestore(&rq
->lock
, flags
);
6229 static void init_rootdomain(struct root_domain
*rd
)
6231 memset(rd
, 0, sizeof(*rd
));
6233 cpus_clear(rd
->span
);
6234 cpus_clear(rd
->online
);
6237 static void init_defrootdomain(void)
6239 init_rootdomain(&def_root_domain
);
6240 atomic_set(&def_root_domain
.refcount
, 1);
6243 static struct root_domain
*alloc_rootdomain(void)
6245 struct root_domain
*rd
;
6247 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6251 init_rootdomain(rd
);
6257 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6258 * hold the hotplug lock.
6261 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6263 struct rq
*rq
= cpu_rq(cpu
);
6264 struct sched_domain
*tmp
;
6266 /* Remove the sched domains which do not contribute to scheduling. */
6267 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6268 struct sched_domain
*parent
= tmp
->parent
;
6271 if (sd_parent_degenerate(tmp
, parent
)) {
6272 tmp
->parent
= parent
->parent
;
6274 parent
->parent
->child
= tmp
;
6278 if (sd
&& sd_degenerate(sd
)) {
6284 sched_domain_debug(sd
, cpu
);
6286 rq_attach_root(rq
, rd
);
6287 rcu_assign_pointer(rq
->sd
, sd
);
6290 /* cpus with isolated domains */
6291 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6293 /* Setup the mask of cpus configured for isolated domains */
6294 static int __init
isolated_cpu_setup(char *str
)
6296 int ints
[NR_CPUS
], i
;
6298 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6299 cpus_clear(cpu_isolated_map
);
6300 for (i
= 1; i
<= ints
[0]; i
++)
6301 if (ints
[i
] < NR_CPUS
)
6302 cpu_set(ints
[i
], cpu_isolated_map
);
6306 __setup("isolcpus=", isolated_cpu_setup
);
6309 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6310 * to a function which identifies what group(along with sched group) a CPU
6311 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6312 * (due to the fact that we keep track of groups covered with a cpumask_t).
6314 * init_sched_build_groups will build a circular linked list of the groups
6315 * covered by the given span, and will set each group's ->cpumask correctly,
6316 * and ->cpu_power to 0.
6319 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
6320 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6321 struct sched_group
**sg
))
6323 struct sched_group
*first
= NULL
, *last
= NULL
;
6324 cpumask_t covered
= CPU_MASK_NONE
;
6327 for_each_cpu_mask(i
, span
) {
6328 struct sched_group
*sg
;
6329 int group
= group_fn(i
, cpu_map
, &sg
);
6332 if (cpu_isset(i
, covered
))
6335 sg
->cpumask
= CPU_MASK_NONE
;
6336 sg
->__cpu_power
= 0;
6338 for_each_cpu_mask(j
, span
) {
6339 if (group_fn(j
, cpu_map
, NULL
) != group
)
6342 cpu_set(j
, covered
);
6343 cpu_set(j
, sg
->cpumask
);
6354 #define SD_NODES_PER_DOMAIN 16
6359 * find_next_best_node - find the next node to include in a sched_domain
6360 * @node: node whose sched_domain we're building
6361 * @used_nodes: nodes already in the sched_domain
6363 * Find the next node to include in a given scheduling domain. Simply
6364 * finds the closest node not already in the @used_nodes map.
6366 * Should use nodemask_t.
6368 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6370 int i
, n
, val
, min_val
, best_node
= 0;
6374 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6375 /* Start at @node */
6376 n
= (node
+ i
) % MAX_NUMNODES
;
6378 if (!nr_cpus_node(n
))
6381 /* Skip already used nodes */
6382 if (test_bit(n
, used_nodes
))
6385 /* Simple min distance search */
6386 val
= node_distance(node
, n
);
6388 if (val
< min_val
) {
6394 set_bit(best_node
, used_nodes
);
6399 * sched_domain_node_span - get a cpumask for a node's sched_domain
6400 * @node: node whose cpumask we're constructing
6401 * @size: number of nodes to include in this span
6403 * Given a node, construct a good cpumask for its sched_domain to span. It
6404 * should be one that prevents unnecessary balancing, but also spreads tasks
6407 static cpumask_t
sched_domain_node_span(int node
)
6409 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6410 cpumask_t span
, nodemask
;
6414 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6416 nodemask
= node_to_cpumask(node
);
6417 cpus_or(span
, span
, nodemask
);
6418 set_bit(node
, used_nodes
);
6420 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6421 int next_node
= find_next_best_node(node
, used_nodes
);
6423 nodemask
= node_to_cpumask(next_node
);
6424 cpus_or(span
, span
, nodemask
);
6431 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6434 * SMT sched-domains:
6436 #ifdef CONFIG_SCHED_SMT
6437 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6438 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6441 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6444 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6450 * multi-core sched-domains:
6452 #ifdef CONFIG_SCHED_MC
6453 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6454 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6457 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6459 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6462 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6463 cpus_and(mask
, mask
, *cpu_map
);
6464 group
= first_cpu(mask
);
6466 *sg
= &per_cpu(sched_group_core
, group
);
6469 #elif defined(CONFIG_SCHED_MC)
6471 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6474 *sg
= &per_cpu(sched_group_core
, cpu
);
6479 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6480 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6483 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6486 #ifdef CONFIG_SCHED_MC
6487 cpumask_t mask
= cpu_coregroup_map(cpu
);
6488 cpus_and(mask
, mask
, *cpu_map
);
6489 group
= first_cpu(mask
);
6490 #elif defined(CONFIG_SCHED_SMT)
6491 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6492 cpus_and(mask
, mask
, *cpu_map
);
6493 group
= first_cpu(mask
);
6498 *sg
= &per_cpu(sched_group_phys
, group
);
6504 * The init_sched_build_groups can't handle what we want to do with node
6505 * groups, so roll our own. Now each node has its own list of groups which
6506 * gets dynamically allocated.
6508 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6509 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6511 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6512 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6514 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6515 struct sched_group
**sg
)
6517 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6520 cpus_and(nodemask
, nodemask
, *cpu_map
);
6521 group
= first_cpu(nodemask
);
6524 *sg
= &per_cpu(sched_group_allnodes
, group
);
6528 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6530 struct sched_group
*sg
= group_head
;
6536 for_each_cpu_mask(j
, sg
->cpumask
) {
6537 struct sched_domain
*sd
;
6539 sd
= &per_cpu(phys_domains
, j
);
6540 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6542 * Only add "power" once for each
6548 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6551 } while (sg
!= group_head
);
6556 /* Free memory allocated for various sched_group structures */
6557 static void free_sched_groups(const cpumask_t
*cpu_map
)
6561 for_each_cpu_mask(cpu
, *cpu_map
) {
6562 struct sched_group
**sched_group_nodes
6563 = sched_group_nodes_bycpu
[cpu
];
6565 if (!sched_group_nodes
)
6568 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6569 cpumask_t nodemask
= node_to_cpumask(i
);
6570 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6572 cpus_and(nodemask
, nodemask
, *cpu_map
);
6573 if (cpus_empty(nodemask
))
6583 if (oldsg
!= sched_group_nodes
[i
])
6586 kfree(sched_group_nodes
);
6587 sched_group_nodes_bycpu
[cpu
] = NULL
;
6591 static void free_sched_groups(const cpumask_t
*cpu_map
)
6597 * Initialize sched groups cpu_power.
6599 * cpu_power indicates the capacity of sched group, which is used while
6600 * distributing the load between different sched groups in a sched domain.
6601 * Typically cpu_power for all the groups in a sched domain will be same unless
6602 * there are asymmetries in the topology. If there are asymmetries, group
6603 * having more cpu_power will pickup more load compared to the group having
6606 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6607 * the maximum number of tasks a group can handle in the presence of other idle
6608 * or lightly loaded groups in the same sched domain.
6610 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6612 struct sched_domain
*child
;
6613 struct sched_group
*group
;
6615 WARN_ON(!sd
|| !sd
->groups
);
6617 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6622 sd
->groups
->__cpu_power
= 0;
6625 * For perf policy, if the groups in child domain share resources
6626 * (for example cores sharing some portions of the cache hierarchy
6627 * or SMT), then set this domain groups cpu_power such that each group
6628 * can handle only one task, when there are other idle groups in the
6629 * same sched domain.
6631 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6633 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6634 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6639 * add cpu_power of each child group to this groups cpu_power
6641 group
= child
->groups
;
6643 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6644 group
= group
->next
;
6645 } while (group
!= child
->groups
);
6649 * Build sched domains for a given set of cpus and attach the sched domains
6650 * to the individual cpus
6652 static int build_sched_domains(const cpumask_t
*cpu_map
)
6655 struct root_domain
*rd
;
6657 struct sched_group
**sched_group_nodes
= NULL
;
6658 int sd_allnodes
= 0;
6661 * Allocate the per-node list of sched groups
6663 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6665 if (!sched_group_nodes
) {
6666 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6669 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6672 rd
= alloc_rootdomain();
6674 printk(KERN_WARNING
"Cannot alloc root domain\n");
6679 * Set up domains for cpus specified by the cpu_map.
6681 for_each_cpu_mask(i
, *cpu_map
) {
6682 struct sched_domain
*sd
= NULL
, *p
;
6683 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6685 cpus_and(nodemask
, nodemask
, *cpu_map
);
6688 if (cpus_weight(*cpu_map
) >
6689 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6690 sd
= &per_cpu(allnodes_domains
, i
);
6691 *sd
= SD_ALLNODES_INIT
;
6692 sd
->span
= *cpu_map
;
6693 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6699 sd
= &per_cpu(node_domains
, i
);
6701 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6705 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6709 sd
= &per_cpu(phys_domains
, i
);
6711 sd
->span
= nodemask
;
6715 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6717 #ifdef CONFIG_SCHED_MC
6719 sd
= &per_cpu(core_domains
, i
);
6721 sd
->span
= cpu_coregroup_map(i
);
6722 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6725 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6728 #ifdef CONFIG_SCHED_SMT
6730 sd
= &per_cpu(cpu_domains
, i
);
6731 *sd
= SD_SIBLING_INIT
;
6732 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6733 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6736 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6740 #ifdef CONFIG_SCHED_SMT
6741 /* Set up CPU (sibling) groups */
6742 for_each_cpu_mask(i
, *cpu_map
) {
6743 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6744 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6745 if (i
!= first_cpu(this_sibling_map
))
6748 init_sched_build_groups(this_sibling_map
, cpu_map
,
6753 #ifdef CONFIG_SCHED_MC
6754 /* Set up multi-core groups */
6755 for_each_cpu_mask(i
, *cpu_map
) {
6756 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6757 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6758 if (i
!= first_cpu(this_core_map
))
6760 init_sched_build_groups(this_core_map
, cpu_map
,
6761 &cpu_to_core_group
);
6765 /* Set up physical groups */
6766 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6767 cpumask_t nodemask
= node_to_cpumask(i
);
6769 cpus_and(nodemask
, nodemask
, *cpu_map
);
6770 if (cpus_empty(nodemask
))
6773 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6777 /* Set up node groups */
6779 init_sched_build_groups(*cpu_map
, cpu_map
,
6780 &cpu_to_allnodes_group
);
6782 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6783 /* Set up node groups */
6784 struct sched_group
*sg
, *prev
;
6785 cpumask_t nodemask
= node_to_cpumask(i
);
6786 cpumask_t domainspan
;
6787 cpumask_t covered
= CPU_MASK_NONE
;
6790 cpus_and(nodemask
, nodemask
, *cpu_map
);
6791 if (cpus_empty(nodemask
)) {
6792 sched_group_nodes
[i
] = NULL
;
6796 domainspan
= sched_domain_node_span(i
);
6797 cpus_and(domainspan
, domainspan
, *cpu_map
);
6799 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6801 printk(KERN_WARNING
"Can not alloc domain group for "
6805 sched_group_nodes
[i
] = sg
;
6806 for_each_cpu_mask(j
, nodemask
) {
6807 struct sched_domain
*sd
;
6809 sd
= &per_cpu(node_domains
, j
);
6812 sg
->__cpu_power
= 0;
6813 sg
->cpumask
= nodemask
;
6815 cpus_or(covered
, covered
, nodemask
);
6818 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6819 cpumask_t tmp
, notcovered
;
6820 int n
= (i
+ j
) % MAX_NUMNODES
;
6822 cpus_complement(notcovered
, covered
);
6823 cpus_and(tmp
, notcovered
, *cpu_map
);
6824 cpus_and(tmp
, tmp
, domainspan
);
6825 if (cpus_empty(tmp
))
6828 nodemask
= node_to_cpumask(n
);
6829 cpus_and(tmp
, tmp
, nodemask
);
6830 if (cpus_empty(tmp
))
6833 sg
= kmalloc_node(sizeof(struct sched_group
),
6837 "Can not alloc domain group for node %d\n", j
);
6840 sg
->__cpu_power
= 0;
6842 sg
->next
= prev
->next
;
6843 cpus_or(covered
, covered
, tmp
);
6850 /* Calculate CPU power for physical packages and nodes */
6851 #ifdef CONFIG_SCHED_SMT
6852 for_each_cpu_mask(i
, *cpu_map
) {
6853 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6855 init_sched_groups_power(i
, sd
);
6858 #ifdef CONFIG_SCHED_MC
6859 for_each_cpu_mask(i
, *cpu_map
) {
6860 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6862 init_sched_groups_power(i
, sd
);
6866 for_each_cpu_mask(i
, *cpu_map
) {
6867 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6869 init_sched_groups_power(i
, sd
);
6873 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6874 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6877 struct sched_group
*sg
;
6879 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6880 init_numa_sched_groups_power(sg
);
6884 /* Attach the domains */
6885 for_each_cpu_mask(i
, *cpu_map
) {
6886 struct sched_domain
*sd
;
6887 #ifdef CONFIG_SCHED_SMT
6888 sd
= &per_cpu(cpu_domains
, i
);
6889 #elif defined(CONFIG_SCHED_MC)
6890 sd
= &per_cpu(core_domains
, i
);
6892 sd
= &per_cpu(phys_domains
, i
);
6894 cpu_attach_domain(sd
, rd
, i
);
6901 free_sched_groups(cpu_map
);
6906 static cpumask_t
*doms_cur
; /* current sched domains */
6907 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6910 * Special case: If a kmalloc of a doms_cur partition (array of
6911 * cpumask_t) fails, then fallback to a single sched domain,
6912 * as determined by the single cpumask_t fallback_doms.
6914 static cpumask_t fallback_doms
;
6916 void __attribute__((weak
)) arch_update_cpu_topology(void)
6921 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6922 * For now this just excludes isolated cpus, but could be used to
6923 * exclude other special cases in the future.
6925 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6929 arch_update_cpu_topology();
6931 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6933 doms_cur
= &fallback_doms
;
6934 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6935 err
= build_sched_domains(doms_cur
);
6936 register_sched_domain_sysctl();
6941 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6943 free_sched_groups(cpu_map
);
6947 * Detach sched domains from a group of cpus specified in cpu_map
6948 * These cpus will now be attached to the NULL domain
6950 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6954 unregister_sched_domain_sysctl();
6956 for_each_cpu_mask(i
, *cpu_map
)
6957 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6958 synchronize_sched();
6959 arch_destroy_sched_domains(cpu_map
);
6963 * Partition sched domains as specified by the 'ndoms_new'
6964 * cpumasks in the array doms_new[] of cpumasks. This compares
6965 * doms_new[] to the current sched domain partitioning, doms_cur[].
6966 * It destroys each deleted domain and builds each new domain.
6968 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6969 * The masks don't intersect (don't overlap.) We should setup one
6970 * sched domain for each mask. CPUs not in any of the cpumasks will
6971 * not be load balanced. If the same cpumask appears both in the
6972 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6975 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6976 * ownership of it and will kfree it when done with it. If the caller
6977 * failed the kmalloc call, then it can pass in doms_new == NULL,
6978 * and partition_sched_domains() will fallback to the single partition
6981 * Call with hotplug lock held
6983 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6989 /* always unregister in case we don't destroy any domains */
6990 unregister_sched_domain_sysctl();
6992 if (doms_new
== NULL
) {
6994 doms_new
= &fallback_doms
;
6995 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6998 /* Destroy deleted domains */
6999 for (i
= 0; i
< ndoms_cur
; i
++) {
7000 for (j
= 0; j
< ndoms_new
; j
++) {
7001 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
7004 /* no match - a current sched domain not in new doms_new[] */
7005 detach_destroy_domains(doms_cur
+ i
);
7010 /* Build new domains */
7011 for (i
= 0; i
< ndoms_new
; i
++) {
7012 for (j
= 0; j
< ndoms_cur
; j
++) {
7013 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
7016 /* no match - add a new doms_new */
7017 build_sched_domains(doms_new
+ i
);
7022 /* Remember the new sched domains */
7023 if (doms_cur
!= &fallback_doms
)
7025 doms_cur
= doms_new
;
7026 ndoms_cur
= ndoms_new
;
7028 register_sched_domain_sysctl();
7033 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7034 int arch_reinit_sched_domains(void)
7039 detach_destroy_domains(&cpu_online_map
);
7040 err
= arch_init_sched_domains(&cpu_online_map
);
7046 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7050 if (buf
[0] != '0' && buf
[0] != '1')
7054 sched_smt_power_savings
= (buf
[0] == '1');
7056 sched_mc_power_savings
= (buf
[0] == '1');
7058 ret
= arch_reinit_sched_domains();
7060 return ret
? ret
: count
;
7063 #ifdef CONFIG_SCHED_MC
7064 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7066 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7068 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7069 const char *buf
, size_t count
)
7071 return sched_power_savings_store(buf
, count
, 0);
7073 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7074 sched_mc_power_savings_store
);
7077 #ifdef CONFIG_SCHED_SMT
7078 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7080 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7082 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7083 const char *buf
, size_t count
)
7085 return sched_power_savings_store(buf
, count
, 1);
7087 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7088 sched_smt_power_savings_store
);
7091 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7095 #ifdef CONFIG_SCHED_SMT
7097 err
= sysfs_create_file(&cls
->kset
.kobj
,
7098 &attr_sched_smt_power_savings
.attr
);
7100 #ifdef CONFIG_SCHED_MC
7101 if (!err
&& mc_capable())
7102 err
= sysfs_create_file(&cls
->kset
.kobj
,
7103 &attr_sched_mc_power_savings
.attr
);
7110 * Force a reinitialization of the sched domains hierarchy. The domains
7111 * and groups cannot be updated in place without racing with the balancing
7112 * code, so we temporarily attach all running cpus to the NULL domain
7113 * which will prevent rebalancing while the sched domains are recalculated.
7115 static int update_sched_domains(struct notifier_block
*nfb
,
7116 unsigned long action
, void *hcpu
)
7119 case CPU_UP_PREPARE
:
7120 case CPU_UP_PREPARE_FROZEN
:
7121 case CPU_DOWN_PREPARE
:
7122 case CPU_DOWN_PREPARE_FROZEN
:
7123 detach_destroy_domains(&cpu_online_map
);
7126 case CPU_UP_CANCELED
:
7127 case CPU_UP_CANCELED_FROZEN
:
7128 case CPU_DOWN_FAILED
:
7129 case CPU_DOWN_FAILED_FROZEN
:
7131 case CPU_ONLINE_FROZEN
:
7133 case CPU_DEAD_FROZEN
:
7135 * Fall through and re-initialise the domains.
7142 /* The hotplug lock is already held by cpu_up/cpu_down */
7143 arch_init_sched_domains(&cpu_online_map
);
7148 void __init
sched_init_smp(void)
7150 cpumask_t non_isolated_cpus
;
7153 arch_init_sched_domains(&cpu_online_map
);
7154 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7155 if (cpus_empty(non_isolated_cpus
))
7156 cpu_set(smp_processor_id(), non_isolated_cpus
);
7158 /* XXX: Theoretical race here - CPU may be hotplugged now */
7159 hotcpu_notifier(update_sched_domains
, 0);
7162 /* Move init over to a non-isolated CPU */
7163 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
7165 sched_init_granularity();
7168 void __init
sched_init_smp(void)
7170 sched_init_granularity();
7172 #endif /* CONFIG_SMP */
7174 int in_sched_functions(unsigned long addr
)
7176 return in_lock_functions(addr
) ||
7177 (addr
>= (unsigned long)__sched_text_start
7178 && addr
< (unsigned long)__sched_text_end
);
7181 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7183 cfs_rq
->tasks_timeline
= RB_ROOT
;
7184 #ifdef CONFIG_FAIR_GROUP_SCHED
7187 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7190 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7192 struct rt_prio_array
*array
;
7195 array
= &rt_rq
->active
;
7196 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7197 INIT_LIST_HEAD(array
->queue
+ i
);
7198 __clear_bit(i
, array
->bitmap
);
7200 /* delimiter for bitsearch: */
7201 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7203 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7204 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7207 rt_rq
->rt_nr_migratory
= 0;
7208 rt_rq
->overloaded
= 0;
7212 rt_rq
->rt_throttled
= 0;
7214 #ifdef CONFIG_RT_GROUP_SCHED
7215 rt_rq
->rt_nr_boosted
= 0;
7220 #ifdef CONFIG_FAIR_GROUP_SCHED
7221 static void init_tg_cfs_entry(struct rq
*rq
, struct task_group
*tg
,
7222 struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
7225 tg
->cfs_rq
[cpu
] = cfs_rq
;
7226 init_cfs_rq(cfs_rq
, rq
);
7229 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7232 se
->cfs_rq
= &rq
->cfs
;
7234 se
->load
.weight
= tg
->shares
;
7235 se
->load
.inv_weight
= div64_64(1ULL<<32, se
->load
.weight
);
7240 #ifdef CONFIG_RT_GROUP_SCHED
7241 static void init_tg_rt_entry(struct rq
*rq
, struct task_group
*tg
,
7242 struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
,
7245 tg
->rt_rq
[cpu
] = rt_rq
;
7246 init_rt_rq(rt_rq
, rq
);
7248 rt_rq
->rt_se
= rt_se
;
7250 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7252 tg
->rt_se
[cpu
] = rt_se
;
7253 rt_se
->rt_rq
= &rq
->rt
;
7254 rt_se
->my_q
= rt_rq
;
7255 rt_se
->parent
= NULL
;
7256 INIT_LIST_HEAD(&rt_se
->run_list
);
7260 void __init
sched_init(void)
7262 int highest_cpu
= 0;
7266 init_defrootdomain();
7269 #ifdef CONFIG_GROUP_SCHED
7270 list_add(&init_task_group
.list
, &task_groups
);
7273 for_each_possible_cpu(i
) {
7277 spin_lock_init(&rq
->lock
);
7278 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7281 init_cfs_rq(&rq
->cfs
, rq
);
7282 init_rt_rq(&rq
->rt
, rq
);
7283 #ifdef CONFIG_FAIR_GROUP_SCHED
7284 init_task_group
.shares
= init_task_group_load
;
7285 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7286 init_tg_cfs_entry(rq
, &init_task_group
,
7287 &per_cpu(init_cfs_rq
, i
),
7288 &per_cpu(init_sched_entity
, i
), i
, 1);
7291 #ifdef CONFIG_RT_GROUP_SCHED
7292 init_task_group
.rt_runtime
=
7293 sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
7294 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7295 init_tg_rt_entry(rq
, &init_task_group
,
7296 &per_cpu(init_rt_rq
, i
),
7297 &per_cpu(init_sched_rt_entity
, i
), i
, 1);
7299 rq
->rt_period_expire
= 0;
7300 rq
->rt_throttled
= 0;
7302 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7303 rq
->cpu_load
[j
] = 0;
7307 rq
->active_balance
= 0;
7308 rq
->next_balance
= jiffies
;
7311 rq
->migration_thread
= NULL
;
7312 INIT_LIST_HEAD(&rq
->migration_queue
);
7313 rq_attach_root(rq
, &def_root_domain
);
7316 atomic_set(&rq
->nr_iowait
, 0);
7320 set_load_weight(&init_task
);
7322 #ifdef CONFIG_PREEMPT_NOTIFIERS
7323 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7327 nr_cpu_ids
= highest_cpu
+ 1;
7328 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7331 #ifdef CONFIG_RT_MUTEXES
7332 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7336 * The boot idle thread does lazy MMU switching as well:
7338 atomic_inc(&init_mm
.mm_count
);
7339 enter_lazy_tlb(&init_mm
, current
);
7342 * Make us the idle thread. Technically, schedule() should not be
7343 * called from this thread, however somewhere below it might be,
7344 * but because we are the idle thread, we just pick up running again
7345 * when this runqueue becomes "idle".
7347 init_idle(current
, smp_processor_id());
7349 * During early bootup we pretend to be a normal task:
7351 current
->sched_class
= &fair_sched_class
;
7353 scheduler_running
= 1;
7356 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7357 void __might_sleep(char *file
, int line
)
7360 static unsigned long prev_jiffy
; /* ratelimiting */
7362 if ((in_atomic() || irqs_disabled()) &&
7363 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7364 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7366 prev_jiffy
= jiffies
;
7367 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7368 " context at %s:%d\n", file
, line
);
7369 printk("in_atomic():%d, irqs_disabled():%d\n",
7370 in_atomic(), irqs_disabled());
7371 debug_show_held_locks(current
);
7372 if (irqs_disabled())
7373 print_irqtrace_events(current
);
7378 EXPORT_SYMBOL(__might_sleep
);
7381 #ifdef CONFIG_MAGIC_SYSRQ
7382 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7385 update_rq_clock(rq
);
7386 on_rq
= p
->se
.on_rq
;
7388 deactivate_task(rq
, p
, 0);
7389 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7391 activate_task(rq
, p
, 0);
7392 resched_task(rq
->curr
);
7396 void normalize_rt_tasks(void)
7398 struct task_struct
*g
, *p
;
7399 unsigned long flags
;
7402 read_lock_irqsave(&tasklist_lock
, flags
);
7403 do_each_thread(g
, p
) {
7405 * Only normalize user tasks:
7410 p
->se
.exec_start
= 0;
7411 #ifdef CONFIG_SCHEDSTATS
7412 p
->se
.wait_start
= 0;
7413 p
->se
.sleep_start
= 0;
7414 p
->se
.block_start
= 0;
7416 task_rq(p
)->clock
= 0;
7420 * Renice negative nice level userspace
7423 if (TASK_NICE(p
) < 0 && p
->mm
)
7424 set_user_nice(p
, 0);
7428 spin_lock(&p
->pi_lock
);
7429 rq
= __task_rq_lock(p
);
7431 normalize_task(rq
, p
);
7433 __task_rq_unlock(rq
);
7434 spin_unlock(&p
->pi_lock
);
7435 } while_each_thread(g
, p
);
7437 read_unlock_irqrestore(&tasklist_lock
, flags
);
7440 #endif /* CONFIG_MAGIC_SYSRQ */
7444 * These functions are only useful for the IA64 MCA handling.
7446 * They can only be called when the whole system has been
7447 * stopped - every CPU needs to be quiescent, and no scheduling
7448 * activity can take place. Using them for anything else would
7449 * be a serious bug, and as a result, they aren't even visible
7450 * under any other configuration.
7454 * curr_task - return the current task for a given cpu.
7455 * @cpu: the processor in question.
7457 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7459 struct task_struct
*curr_task(int cpu
)
7461 return cpu_curr(cpu
);
7465 * set_curr_task - set the current task for a given cpu.
7466 * @cpu: the processor in question.
7467 * @p: the task pointer to set.
7469 * Description: This function must only be used when non-maskable interrupts
7470 * are serviced on a separate stack. It allows the architecture to switch the
7471 * notion of the current task on a cpu in a non-blocking manner. This function
7472 * must be called with all CPU's synchronized, and interrupts disabled, the
7473 * and caller must save the original value of the current task (see
7474 * curr_task() above) and restore that value before reenabling interrupts and
7475 * re-starting the system.
7477 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7479 void set_curr_task(int cpu
, struct task_struct
*p
)
7486 #ifdef CONFIG_GROUP_SCHED
7488 #ifdef CONFIG_FAIR_GROUP_SCHED
7489 static void free_fair_sched_group(struct task_group
*tg
)
7493 for_each_possible_cpu(i
) {
7495 kfree(tg
->cfs_rq
[i
]);
7504 static int alloc_fair_sched_group(struct task_group
*tg
)
7506 struct cfs_rq
*cfs_rq
;
7507 struct sched_entity
*se
;
7511 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7514 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7518 tg
->shares
= NICE_0_LOAD
;
7520 for_each_possible_cpu(i
) {
7523 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
7524 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7528 se
= kmalloc_node(sizeof(struct sched_entity
),
7529 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7533 init_tg_cfs_entry(rq
, tg
, cfs_rq
, se
, i
, 0);
7542 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7544 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
7545 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
7548 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7550 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
7553 static inline void free_fair_sched_group(struct task_group
*tg
)
7557 static inline int alloc_fair_sched_group(struct task_group
*tg
)
7562 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7566 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7571 #ifdef CONFIG_RT_GROUP_SCHED
7572 static void free_rt_sched_group(struct task_group
*tg
)
7576 for_each_possible_cpu(i
) {
7578 kfree(tg
->rt_rq
[i
]);
7580 kfree(tg
->rt_se
[i
]);
7587 static int alloc_rt_sched_group(struct task_group
*tg
)
7589 struct rt_rq
*rt_rq
;
7590 struct sched_rt_entity
*rt_se
;
7594 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * NR_CPUS
, GFP_KERNEL
);
7597 tg
->rt_se
= kzalloc(sizeof(rt_se
) * NR_CPUS
, GFP_KERNEL
);
7603 for_each_possible_cpu(i
) {
7606 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
7607 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7611 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
7612 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7616 init_tg_rt_entry(rq
, tg
, rt_rq
, rt_se
, i
, 0);
7625 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7627 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
7628 &cpu_rq(cpu
)->leaf_rt_rq_list
);
7631 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7633 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
7636 static inline void free_rt_sched_group(struct task_group
*tg
)
7640 static inline int alloc_rt_sched_group(struct task_group
*tg
)
7645 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7649 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7654 static void free_sched_group(struct task_group
*tg
)
7656 free_fair_sched_group(tg
);
7657 free_rt_sched_group(tg
);
7661 /* allocate runqueue etc for a new task group */
7662 struct task_group
*sched_create_group(void)
7664 struct task_group
*tg
;
7665 unsigned long flags
;
7668 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7670 return ERR_PTR(-ENOMEM
);
7672 if (!alloc_fair_sched_group(tg
))
7675 if (!alloc_rt_sched_group(tg
))
7678 spin_lock_irqsave(&task_group_lock
, flags
);
7679 for_each_possible_cpu(i
) {
7680 register_fair_sched_group(tg
, i
);
7681 register_rt_sched_group(tg
, i
);
7683 list_add_rcu(&tg
->list
, &task_groups
);
7684 spin_unlock_irqrestore(&task_group_lock
, flags
);
7689 free_sched_group(tg
);
7690 return ERR_PTR(-ENOMEM
);
7693 /* rcu callback to free various structures associated with a task group */
7694 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7696 /* now it should be safe to free those cfs_rqs */
7697 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7700 /* Destroy runqueue etc associated with a task group */
7701 void sched_destroy_group(struct task_group
*tg
)
7703 unsigned long flags
;
7706 spin_lock_irqsave(&task_group_lock
, flags
);
7707 for_each_possible_cpu(i
) {
7708 unregister_fair_sched_group(tg
, i
);
7709 unregister_rt_sched_group(tg
, i
);
7711 list_del_rcu(&tg
->list
);
7712 spin_unlock_irqrestore(&task_group_lock
, flags
);
7714 /* wait for possible concurrent references to cfs_rqs complete */
7715 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7718 /* change task's runqueue when it moves between groups.
7719 * The caller of this function should have put the task in its new group
7720 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7721 * reflect its new group.
7723 void sched_move_task(struct task_struct
*tsk
)
7726 unsigned long flags
;
7729 rq
= task_rq_lock(tsk
, &flags
);
7731 update_rq_clock(rq
);
7733 running
= task_current(rq
, tsk
);
7734 on_rq
= tsk
->se
.on_rq
;
7737 dequeue_task(rq
, tsk
, 0);
7738 if (unlikely(running
))
7739 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7741 set_task_rq(tsk
, task_cpu(tsk
));
7743 #ifdef CONFIG_FAIR_GROUP_SCHED
7744 if (tsk
->sched_class
->moved_group
)
7745 tsk
->sched_class
->moved_group(tsk
);
7748 if (unlikely(running
))
7749 tsk
->sched_class
->set_curr_task(rq
);
7751 enqueue_task(rq
, tsk
, 0);
7753 task_rq_unlock(rq
, &flags
);
7756 #ifdef CONFIG_FAIR_GROUP_SCHED
7757 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7759 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7760 struct rq
*rq
= cfs_rq
->rq
;
7763 spin_lock_irq(&rq
->lock
);
7767 dequeue_entity(cfs_rq
, se
, 0);
7769 se
->load
.weight
= shares
;
7770 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7773 enqueue_entity(cfs_rq
, se
, 0);
7775 spin_unlock_irq(&rq
->lock
);
7778 static DEFINE_MUTEX(shares_mutex
);
7780 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7783 unsigned long flags
;
7786 * A weight of 0 or 1 can cause arithmetics problems.
7787 * (The default weight is 1024 - so there's no practical
7788 * limitation from this.)
7793 mutex_lock(&shares_mutex
);
7794 if (tg
->shares
== shares
)
7797 spin_lock_irqsave(&task_group_lock
, flags
);
7798 for_each_possible_cpu(i
)
7799 unregister_fair_sched_group(tg
, i
);
7800 spin_unlock_irqrestore(&task_group_lock
, flags
);
7802 /* wait for any ongoing reference to this group to finish */
7803 synchronize_sched();
7806 * Now we are free to modify the group's share on each cpu
7807 * w/o tripping rebalance_share or load_balance_fair.
7809 tg
->shares
= shares
;
7810 for_each_possible_cpu(i
)
7811 set_se_shares(tg
->se
[i
], shares
);
7814 * Enable load balance activity on this group, by inserting it back on
7815 * each cpu's rq->leaf_cfs_rq_list.
7817 spin_lock_irqsave(&task_group_lock
, flags
);
7818 for_each_possible_cpu(i
)
7819 register_fair_sched_group(tg
, i
);
7820 spin_unlock_irqrestore(&task_group_lock
, flags
);
7822 mutex_unlock(&shares_mutex
);
7826 unsigned long sched_group_shares(struct task_group
*tg
)
7832 #ifdef CONFIG_RT_GROUP_SCHED
7834 * Ensure that the real time constraints are schedulable.
7836 static DEFINE_MUTEX(rt_constraints_mutex
);
7838 static unsigned long to_ratio(u64 period
, u64 runtime
)
7840 if (runtime
== RUNTIME_INF
)
7843 return div64_64(runtime
<< 16, period
);
7846 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7848 struct task_group
*tgi
;
7849 unsigned long total
= 0;
7850 unsigned long global_ratio
=
7851 to_ratio(sysctl_sched_rt_period
,
7852 sysctl_sched_rt_runtime
< 0 ?
7853 RUNTIME_INF
: sysctl_sched_rt_runtime
);
7856 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
7860 total
+= to_ratio(period
, tgi
->rt_runtime
);
7864 return total
+ to_ratio(period
, runtime
) < global_ratio
;
7867 /* Must be called with tasklist_lock held */
7868 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7870 struct task_struct
*g
, *p
;
7871 do_each_thread(g
, p
) {
7872 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
7874 } while_each_thread(g
, p
);
7878 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7880 u64 rt_runtime
, rt_period
;
7883 rt_period
= (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
7884 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7885 if (rt_runtime_us
== -1)
7886 rt_runtime
= RUNTIME_INF
;
7888 mutex_lock(&rt_constraints_mutex
);
7889 read_lock(&tasklist_lock
);
7890 if (rt_runtime_us
== 0 && tg_has_rt_tasks(tg
)) {
7894 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
7898 tg
->rt_runtime
= rt_runtime
;
7900 read_unlock(&tasklist_lock
);
7901 mutex_unlock(&rt_constraints_mutex
);
7906 long sched_group_rt_runtime(struct task_group
*tg
)
7910 if (tg
->rt_runtime
== RUNTIME_INF
)
7913 rt_runtime_us
= tg
->rt_runtime
;
7914 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7915 return rt_runtime_us
;
7918 #endif /* CONFIG_GROUP_SCHED */
7920 #ifdef CONFIG_CGROUP_SCHED
7922 /* return corresponding task_group object of a cgroup */
7923 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7925 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7926 struct task_group
, css
);
7929 static struct cgroup_subsys_state
*
7930 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7932 struct task_group
*tg
;
7934 if (!cgrp
->parent
) {
7935 /* This is early initialization for the top cgroup */
7936 init_task_group
.css
.cgroup
= cgrp
;
7937 return &init_task_group
.css
;
7940 /* we support only 1-level deep hierarchical scheduler atm */
7941 if (cgrp
->parent
->parent
)
7942 return ERR_PTR(-EINVAL
);
7944 tg
= sched_create_group();
7946 return ERR_PTR(-ENOMEM
);
7948 /* Bind the cgroup to task_group object we just created */
7949 tg
->css
.cgroup
= cgrp
;
7955 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7957 struct task_group
*tg
= cgroup_tg(cgrp
);
7959 sched_destroy_group(tg
);
7963 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7964 struct task_struct
*tsk
)
7966 #ifdef CONFIG_RT_GROUP_SCHED
7967 /* Don't accept realtime tasks when there is no way for them to run */
7968 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_runtime
== 0)
7971 /* We don't support RT-tasks being in separate groups */
7972 if (tsk
->sched_class
!= &fair_sched_class
)
7980 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7981 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7983 sched_move_task(tsk
);
7986 #ifdef CONFIG_FAIR_GROUP_SCHED
7987 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7990 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7993 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7995 struct task_group
*tg
= cgroup_tg(cgrp
);
7997 return (u64
) tg
->shares
;
8001 #ifdef CONFIG_RT_GROUP_SCHED
8002 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8004 const char __user
*userbuf
,
8005 size_t nbytes
, loff_t
*unused_ppos
)
8014 if (nbytes
>= sizeof(buffer
))
8016 if (copy_from_user(buffer
, userbuf
, nbytes
))
8019 buffer
[nbytes
] = 0; /* nul-terminate */
8021 /* strip newline if necessary */
8022 if (nbytes
&& (buffer
[nbytes
-1] == '\n'))
8023 buffer
[nbytes
-1] = 0;
8024 val
= simple_strtoll(buffer
, &end
, 0);
8028 /* Pass to subsystem */
8029 retval
= sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8035 static ssize_t
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
,
8037 char __user
*buf
, size_t nbytes
,
8041 long val
= sched_group_rt_runtime(cgroup_tg(cgrp
));
8042 int len
= sprintf(tmp
, "%ld\n", val
);
8044 return simple_read_from_buffer(buf
, nbytes
, ppos
, tmp
, len
);
8048 static struct cftype cpu_files
[] = {
8049 #ifdef CONFIG_FAIR_GROUP_SCHED
8052 .read_uint
= cpu_shares_read_uint
,
8053 .write_uint
= cpu_shares_write_uint
,
8056 #ifdef CONFIG_RT_GROUP_SCHED
8058 .name
= "rt_runtime_us",
8059 .read
= cpu_rt_runtime_read
,
8060 .write
= cpu_rt_runtime_write
,
8065 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8067 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8070 struct cgroup_subsys cpu_cgroup_subsys
= {
8072 .create
= cpu_cgroup_create
,
8073 .destroy
= cpu_cgroup_destroy
,
8074 .can_attach
= cpu_cgroup_can_attach
,
8075 .attach
= cpu_cgroup_attach
,
8076 .populate
= cpu_cgroup_populate
,
8077 .subsys_id
= cpu_cgroup_subsys_id
,
8081 #endif /* CONFIG_CGROUP_SCHED */
8083 #ifdef CONFIG_CGROUP_CPUACCT
8086 * CPU accounting code for task groups.
8088 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8089 * (balbir@in.ibm.com).
8092 /* track cpu usage of a group of tasks */
8094 struct cgroup_subsys_state css
;
8095 /* cpuusage holds pointer to a u64-type object on every cpu */
8099 struct cgroup_subsys cpuacct_subsys
;
8101 /* return cpu accounting group corresponding to this container */
8102 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
8104 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
8105 struct cpuacct
, css
);
8108 /* return cpu accounting group to which this task belongs */
8109 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8111 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8112 struct cpuacct
, css
);
8115 /* create a new cpu accounting group */
8116 static struct cgroup_subsys_state
*cpuacct_create(
8117 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8119 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8122 return ERR_PTR(-ENOMEM
);
8124 ca
->cpuusage
= alloc_percpu(u64
);
8125 if (!ca
->cpuusage
) {
8127 return ERR_PTR(-ENOMEM
);
8133 /* destroy an existing cpu accounting group */
8135 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8137 struct cpuacct
*ca
= cgroup_ca(cont
);
8139 free_percpu(ca
->cpuusage
);
8143 /* return total cpu usage (in nanoseconds) of a group */
8144 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
8146 struct cpuacct
*ca
= cgroup_ca(cont
);
8147 u64 totalcpuusage
= 0;
8150 for_each_possible_cpu(i
) {
8151 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8154 * Take rq->lock to make 64-bit addition safe on 32-bit
8157 spin_lock_irq(&cpu_rq(i
)->lock
);
8158 totalcpuusage
+= *cpuusage
;
8159 spin_unlock_irq(&cpu_rq(i
)->lock
);
8162 return totalcpuusage
;
8165 static struct cftype files
[] = {
8168 .read_uint
= cpuusage_read
,
8172 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8174 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
8178 * charge this task's execution time to its accounting group.
8180 * called with rq->lock held.
8182 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8186 if (!cpuacct_subsys
.active
)
8191 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
8193 *cpuusage
+= cputime
;
8197 struct cgroup_subsys cpuacct_subsys
= {
8199 .create
= cpuacct_create
,
8200 .destroy
= cpuacct_destroy
,
8201 .populate
= cpuacct_populate
,
8202 .subsys_id
= cpuacct_subsys_id
,
8204 #endif /* CONFIG_CGROUP_CPUACCT */