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_FAIR_GROUP_SCHED
160 #include <linux/cgroup.h>
164 static LIST_HEAD(task_groups
);
166 /* task group related information */
168 #ifdef CONFIG_FAIR_CGROUP_SCHED
169 struct cgroup_subsys_state css
;
171 /* schedulable entities of this group on each cpu */
172 struct sched_entity
**se
;
173 /* runqueue "owned" by this group on each cpu */
174 struct cfs_rq
**cfs_rq
;
176 struct sched_rt_entity
**rt_se
;
177 struct rt_rq
**rt_rq
;
179 unsigned int rt_ratio
;
182 * shares assigned to a task group governs how much of cpu bandwidth
183 * is allocated to the group. The more shares a group has, the more is
184 * the cpu bandwidth allocated to it.
186 * For ex, lets say that there are three task groups, A, B and C which
187 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
188 * cpu bandwidth allocated by the scheduler to task groups A, B and C
191 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
192 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
193 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
195 * The weight assigned to a task group's schedulable entities on every
196 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
197 * group's shares. For ex: lets say that task group A has been
198 * assigned shares of 1000 and there are two CPUs in a system. Then,
200 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
202 * Note: It's not necessary that each of a task's group schedulable
203 * entity have the same weight on all CPUs. If the group
204 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
205 * better distribution of weight could be:
207 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
208 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
210 * rebalance_shares() is responsible for distributing the shares of a
211 * task groups like this among the group's schedulable entities across
215 unsigned long shares
;
218 struct list_head list
;
221 /* Default task group's sched entity on each cpu */
222 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
223 /* Default task group's cfs_rq on each cpu */
224 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
226 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
227 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
229 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
230 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
232 static struct sched_rt_entity
*init_sched_rt_entity_p
[NR_CPUS
];
233 static struct rt_rq
*init_rt_rq_p
[NR_CPUS
];
235 /* task_group_mutex serializes add/remove of task groups and also changes to
236 * a task group's cpu shares.
238 static DEFINE_MUTEX(task_group_mutex
);
240 /* doms_cur_mutex serializes access to doms_cur[] array */
241 static DEFINE_MUTEX(doms_cur_mutex
);
244 /* kernel thread that runs rebalance_shares() periodically */
245 static struct task_struct
*lb_monitor_task
;
246 static int load_balance_monitor(void *unused
);
249 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
);
251 /* Default task group.
252 * Every task in system belong to this group at bootup.
254 struct task_group init_task_group
= {
255 .se
= init_sched_entity_p
,
256 .cfs_rq
= init_cfs_rq_p
,
258 .rt_se
= init_sched_rt_entity_p
,
259 .rt_rq
= init_rt_rq_p
,
262 #ifdef CONFIG_FAIR_USER_SCHED
263 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
265 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
268 #define MIN_GROUP_SHARES 2
270 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
272 /* return group to which a task belongs */
273 static inline struct task_group
*task_group(struct task_struct
*p
)
275 struct task_group
*tg
;
277 #ifdef CONFIG_FAIR_USER_SCHED
279 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
280 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
281 struct task_group
, css
);
283 tg
= &init_task_group
;
288 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
289 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
291 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
292 p
->se
.parent
= task_group(p
)->se
[cpu
];
294 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
295 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
298 static inline void lock_task_group_list(void)
300 mutex_lock(&task_group_mutex
);
303 static inline void unlock_task_group_list(void)
305 mutex_unlock(&task_group_mutex
);
308 static inline void lock_doms_cur(void)
310 mutex_lock(&doms_cur_mutex
);
313 static inline void unlock_doms_cur(void)
315 mutex_unlock(&doms_cur_mutex
);
320 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
321 static inline void lock_task_group_list(void) { }
322 static inline void unlock_task_group_list(void) { }
323 static inline void lock_doms_cur(void) { }
324 static inline void unlock_doms_cur(void) { }
326 #endif /* CONFIG_FAIR_GROUP_SCHED */
328 /* CFS-related fields in a runqueue */
330 struct load_weight load
;
331 unsigned long nr_running
;
336 struct rb_root tasks_timeline
;
337 struct rb_node
*rb_leftmost
;
338 struct rb_node
*rb_load_balance_curr
;
339 /* 'curr' points to currently running entity on this cfs_rq.
340 * It is set to NULL otherwise (i.e when none are currently running).
342 struct sched_entity
*curr
;
344 unsigned long nr_spread_over
;
346 #ifdef CONFIG_FAIR_GROUP_SCHED
347 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
350 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
351 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
352 * (like users, containers etc.)
354 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
355 * list is used during load balance.
357 struct list_head leaf_cfs_rq_list
;
358 struct task_group
*tg
; /* group that "owns" this runqueue */
362 /* Real-Time classes' related field in a runqueue: */
364 struct rt_prio_array active
;
365 unsigned long rt_nr_running
;
366 #if defined CONFIG_SMP || defined CONFIG_FAIR_GROUP_SCHED
367 int highest_prio
; /* highest queued rt task prio */
370 unsigned long rt_nr_migratory
;
376 #ifdef CONFIG_FAIR_GROUP_SCHED
378 struct list_head leaf_rt_rq_list
;
379 struct task_group
*tg
;
380 struct sched_rt_entity
*rt_se
;
387 * We add the notion of a root-domain which will be used to define per-domain
388 * variables. Each exclusive cpuset essentially defines an island domain by
389 * fully partitioning the member cpus from any other cpuset. Whenever a new
390 * exclusive cpuset is created, we also create and attach a new root-domain
400 * The "RT overload" flag: it gets set if a CPU has more than
401 * one runnable RT task.
408 * By default the system creates a single root-domain with all cpus as
409 * members (mimicking the global state we have today).
411 static struct root_domain def_root_domain
;
416 * This is the main, per-CPU runqueue data structure.
418 * Locking rule: those places that want to lock multiple runqueues
419 * (such as the load balancing or the thread migration code), lock
420 * acquire operations must be ordered by ascending &runqueue.
427 * nr_running and cpu_load should be in the same cacheline because
428 * remote CPUs use both these fields when doing load calculation.
430 unsigned long nr_running
;
431 #define CPU_LOAD_IDX_MAX 5
432 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
433 unsigned char idle_at_tick
;
435 unsigned char in_nohz_recently
;
437 /* capture load from *all* tasks on this cpu: */
438 struct load_weight load
;
439 unsigned long nr_load_updates
;
444 u64 rt_period_expire
;
447 #ifdef CONFIG_FAIR_GROUP_SCHED
448 /* list of leaf cfs_rq on this cpu: */
449 struct list_head leaf_cfs_rq_list
;
450 struct list_head leaf_rt_rq_list
;
454 * This is part of a global counter where only the total sum
455 * over all CPUs matters. A task can increase this counter on
456 * one CPU and if it got migrated afterwards it may decrease
457 * it on another CPU. Always updated under the runqueue lock:
459 unsigned long nr_uninterruptible
;
461 struct task_struct
*curr
, *idle
;
462 unsigned long next_balance
;
463 struct mm_struct
*prev_mm
;
465 u64 clock
, prev_clock_raw
;
468 unsigned int clock_warps
, clock_overflows
;
470 unsigned int clock_deep_idle_events
;
476 struct root_domain
*rd
;
477 struct sched_domain
*sd
;
479 /* For active balancing */
482 /* cpu of this runqueue: */
485 struct task_struct
*migration_thread
;
486 struct list_head migration_queue
;
489 #ifdef CONFIG_SCHED_HRTICK
490 unsigned long hrtick_flags
;
491 ktime_t hrtick_expire
;
492 struct hrtimer hrtick_timer
;
495 #ifdef CONFIG_SCHEDSTATS
497 struct sched_info rq_sched_info
;
499 /* sys_sched_yield() stats */
500 unsigned int yld_exp_empty
;
501 unsigned int yld_act_empty
;
502 unsigned int yld_both_empty
;
503 unsigned int yld_count
;
505 /* schedule() stats */
506 unsigned int sched_switch
;
507 unsigned int sched_count
;
508 unsigned int sched_goidle
;
510 /* try_to_wake_up() stats */
511 unsigned int ttwu_count
;
512 unsigned int ttwu_local
;
515 unsigned int bkl_count
;
517 struct lock_class_key rq_lock_key
;
520 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
522 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
524 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
527 static inline int cpu_of(struct rq
*rq
)
537 * Update the per-runqueue clock, as finegrained as the platform can give
538 * us, but without assuming monotonicity, etc.:
540 static void __update_rq_clock(struct rq
*rq
)
542 u64 prev_raw
= rq
->prev_clock_raw
;
543 u64 now
= sched_clock();
544 s64 delta
= now
- prev_raw
;
545 u64 clock
= rq
->clock
;
547 #ifdef CONFIG_SCHED_DEBUG
548 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
551 * Protect against sched_clock() occasionally going backwards:
553 if (unlikely(delta
< 0)) {
558 * Catch too large forward jumps too:
560 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
561 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
562 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
565 rq
->clock_overflows
++;
567 if (unlikely(delta
> rq
->clock_max_delta
))
568 rq
->clock_max_delta
= delta
;
573 rq
->prev_clock_raw
= now
;
577 static void update_rq_clock(struct rq
*rq
)
579 if (likely(smp_processor_id() == cpu_of(rq
)))
580 __update_rq_clock(rq
);
584 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
585 * See detach_destroy_domains: synchronize_sched for details.
587 * The domain tree of any CPU may only be accessed from within
588 * preempt-disabled sections.
590 #define for_each_domain(cpu, __sd) \
591 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
593 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
594 #define this_rq() (&__get_cpu_var(runqueues))
595 #define task_rq(p) cpu_rq(task_cpu(p))
596 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
598 unsigned long rt_needs_cpu(int cpu
)
600 struct rq
*rq
= cpu_rq(cpu
);
603 if (!rq
->rt_throttled
)
606 if (rq
->clock
> rq
->rt_period_expire
)
609 delta
= rq
->rt_period_expire
- rq
->clock
;
610 do_div(delta
, NSEC_PER_SEC
/ HZ
);
612 return (unsigned long)delta
;
616 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
618 #ifdef CONFIG_SCHED_DEBUG
619 # define const_debug __read_mostly
621 # define const_debug static const
625 * Debugging: various feature bits
628 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
629 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
630 SCHED_FEAT_START_DEBIT
= 4,
631 SCHED_FEAT_TREE_AVG
= 8,
632 SCHED_FEAT_APPROX_AVG
= 16,
633 SCHED_FEAT_HRTICK
= 32,
634 SCHED_FEAT_DOUBLE_TICK
= 64,
637 const_debug
unsigned int sysctl_sched_features
=
638 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
639 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
640 SCHED_FEAT_START_DEBIT
* 1 |
641 SCHED_FEAT_TREE_AVG
* 0 |
642 SCHED_FEAT_APPROX_AVG
* 0 |
643 SCHED_FEAT_HRTICK
* 1 |
644 SCHED_FEAT_DOUBLE_TICK
* 0;
646 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
649 * Number of tasks to iterate in a single balance run.
650 * Limited because this is done with IRQs disabled.
652 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
655 * period over which we measure -rt task cpu usage in ms.
658 const_debug
unsigned int sysctl_sched_rt_period
= 1000;
660 #define SCHED_RT_FRAC_SHIFT 16
661 #define SCHED_RT_FRAC (1UL << SCHED_RT_FRAC_SHIFT)
664 * ratio of time -rt tasks may consume.
667 const_debug
unsigned int sysctl_sched_rt_ratio
= 62259;
670 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
671 * clock constructed from sched_clock():
673 unsigned long long cpu_clock(int cpu
)
675 unsigned long long now
;
679 local_irq_save(flags
);
682 * Only call sched_clock() if the scheduler has already been
683 * initialized (some code might call cpu_clock() very early):
688 local_irq_restore(flags
);
692 EXPORT_SYMBOL_GPL(cpu_clock
);
694 #ifndef prepare_arch_switch
695 # define prepare_arch_switch(next) do { } while (0)
697 #ifndef finish_arch_switch
698 # define finish_arch_switch(prev) do { } while (0)
701 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
703 return rq
->curr
== p
;
706 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
707 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
709 return task_current(rq
, p
);
712 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
716 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
718 #ifdef CONFIG_DEBUG_SPINLOCK
719 /* this is a valid case when another task releases the spinlock */
720 rq
->lock
.owner
= current
;
723 * If we are tracking spinlock dependencies then we have to
724 * fix up the runqueue lock - which gets 'carried over' from
727 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
729 spin_unlock_irq(&rq
->lock
);
732 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
733 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
738 return task_current(rq
, p
);
742 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
746 * We can optimise this out completely for !SMP, because the
747 * SMP rebalancing from interrupt is the only thing that cares
752 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
753 spin_unlock_irq(&rq
->lock
);
755 spin_unlock(&rq
->lock
);
759 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
763 * After ->oncpu is cleared, the task can be moved to a different CPU.
764 * We must ensure this doesn't happen until the switch is completely
770 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
774 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
777 * __task_rq_lock - lock the runqueue a given task resides on.
778 * Must be called interrupts disabled.
780 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
784 struct rq
*rq
= task_rq(p
);
785 spin_lock(&rq
->lock
);
786 if (likely(rq
== task_rq(p
)))
788 spin_unlock(&rq
->lock
);
793 * task_rq_lock - lock the runqueue a given task resides on and disable
794 * interrupts. Note the ordering: we can safely lookup the task_rq without
795 * explicitly disabling preemption.
797 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
803 local_irq_save(*flags
);
805 spin_lock(&rq
->lock
);
806 if (likely(rq
== task_rq(p
)))
808 spin_unlock_irqrestore(&rq
->lock
, *flags
);
812 static void __task_rq_unlock(struct rq
*rq
)
815 spin_unlock(&rq
->lock
);
818 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
821 spin_unlock_irqrestore(&rq
->lock
, *flags
);
825 * this_rq_lock - lock this runqueue and disable interrupts.
827 static struct rq
*this_rq_lock(void)
834 spin_lock(&rq
->lock
);
840 * We are going deep-idle (irqs are disabled):
842 void sched_clock_idle_sleep_event(void)
844 struct rq
*rq
= cpu_rq(smp_processor_id());
846 spin_lock(&rq
->lock
);
847 __update_rq_clock(rq
);
848 spin_unlock(&rq
->lock
);
849 rq
->clock_deep_idle_events
++;
851 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
854 * We just idled delta nanoseconds (called with irqs disabled):
856 void sched_clock_idle_wakeup_event(u64 delta_ns
)
858 struct rq
*rq
= cpu_rq(smp_processor_id());
859 u64 now
= sched_clock();
861 touch_softlockup_watchdog();
862 rq
->idle_clock
+= delta_ns
;
864 * Override the previous timestamp and ignore all
865 * sched_clock() deltas that occured while we idled,
866 * and use the PM-provided delta_ns to advance the
869 spin_lock(&rq
->lock
);
870 rq
->prev_clock_raw
= now
;
871 rq
->clock
+= delta_ns
;
872 spin_unlock(&rq
->lock
);
874 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
876 static void __resched_task(struct task_struct
*p
, int tif_bit
);
878 static inline void resched_task(struct task_struct
*p
)
880 __resched_task(p
, TIF_NEED_RESCHED
);
883 #ifdef CONFIG_SCHED_HRTICK
885 * Use HR-timers to deliver accurate preemption points.
887 * Its all a bit involved since we cannot program an hrt while holding the
888 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
891 * When we get rescheduled we reprogram the hrtick_timer outside of the
894 static inline void resched_hrt(struct task_struct
*p
)
896 __resched_task(p
, TIF_HRTICK_RESCHED
);
899 static inline void resched_rq(struct rq
*rq
)
903 spin_lock_irqsave(&rq
->lock
, flags
);
904 resched_task(rq
->curr
);
905 spin_unlock_irqrestore(&rq
->lock
, flags
);
909 HRTICK_SET
, /* re-programm hrtick_timer */
910 HRTICK_RESET
, /* not a new slice */
915 * - enabled by features
916 * - hrtimer is actually high res
918 static inline int hrtick_enabled(struct rq
*rq
)
920 if (!sched_feat(HRTICK
))
922 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
926 * Called to set the hrtick timer state.
928 * called with rq->lock held and irqs disabled
930 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
932 assert_spin_locked(&rq
->lock
);
935 * preempt at: now + delay
938 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
940 * indicate we need to program the timer
942 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
944 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
947 * New slices are called from the schedule path and don't need a
951 resched_hrt(rq
->curr
);
954 static void hrtick_clear(struct rq
*rq
)
956 if (hrtimer_active(&rq
->hrtick_timer
))
957 hrtimer_cancel(&rq
->hrtick_timer
);
961 * Update the timer from the possible pending state.
963 static void hrtick_set(struct rq
*rq
)
969 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
971 spin_lock_irqsave(&rq
->lock
, flags
);
972 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
973 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
974 time
= rq
->hrtick_expire
;
975 clear_thread_flag(TIF_HRTICK_RESCHED
);
976 spin_unlock_irqrestore(&rq
->lock
, flags
);
979 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
980 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
987 * High-resolution timer tick.
988 * Runs from hardirq context with interrupts disabled.
990 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
992 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
994 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
996 spin_lock(&rq
->lock
);
997 __update_rq_clock(rq
);
998 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
999 spin_unlock(&rq
->lock
);
1001 return HRTIMER_NORESTART
;
1004 static inline void init_rq_hrtick(struct rq
*rq
)
1006 rq
->hrtick_flags
= 0;
1007 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1008 rq
->hrtick_timer
.function
= hrtick
;
1009 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1012 void hrtick_resched(void)
1015 unsigned long flags
;
1017 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1020 local_irq_save(flags
);
1021 rq
= cpu_rq(smp_processor_id());
1023 local_irq_restore(flags
);
1026 static inline void hrtick_clear(struct rq
*rq
)
1030 static inline void hrtick_set(struct rq
*rq
)
1034 static inline void init_rq_hrtick(struct rq
*rq
)
1038 void hrtick_resched(void)
1044 * resched_task - mark a task 'to be rescheduled now'.
1046 * On UP this means the setting of the need_resched flag, on SMP it
1047 * might also involve a cross-CPU call to trigger the scheduler on
1052 #ifndef tsk_is_polling
1053 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1056 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1060 assert_spin_locked(&task_rq(p
)->lock
);
1062 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1065 set_tsk_thread_flag(p
, tif_bit
);
1068 if (cpu
== smp_processor_id())
1071 /* NEED_RESCHED must be visible before we test polling */
1073 if (!tsk_is_polling(p
))
1074 smp_send_reschedule(cpu
);
1077 static void resched_cpu(int cpu
)
1079 struct rq
*rq
= cpu_rq(cpu
);
1080 unsigned long flags
;
1082 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1084 resched_task(cpu_curr(cpu
));
1085 spin_unlock_irqrestore(&rq
->lock
, flags
);
1088 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1090 assert_spin_locked(&task_rq(p
)->lock
);
1091 set_tsk_thread_flag(p
, tif_bit
);
1095 #if BITS_PER_LONG == 32
1096 # define WMULT_CONST (~0UL)
1098 # define WMULT_CONST (1UL << 32)
1101 #define WMULT_SHIFT 32
1104 * Shift right and round:
1106 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1108 static unsigned long
1109 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1110 struct load_weight
*lw
)
1114 if (unlikely(!lw
->inv_weight
))
1115 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
1117 tmp
= (u64
)delta_exec
* weight
;
1119 * Check whether we'd overflow the 64-bit multiplication:
1121 if (unlikely(tmp
> WMULT_CONST
))
1122 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1125 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1127 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1130 static inline unsigned long
1131 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1133 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1136 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1141 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1147 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1148 * of tasks with abnormal "nice" values across CPUs the contribution that
1149 * each task makes to its run queue's load is weighted according to its
1150 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1151 * scaled version of the new time slice allocation that they receive on time
1155 #define WEIGHT_IDLEPRIO 2
1156 #define WMULT_IDLEPRIO (1 << 31)
1159 * Nice levels are multiplicative, with a gentle 10% change for every
1160 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1161 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1162 * that remained on nice 0.
1164 * The "10% effect" is relative and cumulative: from _any_ nice level,
1165 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1166 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1167 * If a task goes up by ~10% and another task goes down by ~10% then
1168 * the relative distance between them is ~25%.)
1170 static const int prio_to_weight
[40] = {
1171 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1172 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1173 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1174 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1175 /* 0 */ 1024, 820, 655, 526, 423,
1176 /* 5 */ 335, 272, 215, 172, 137,
1177 /* 10 */ 110, 87, 70, 56, 45,
1178 /* 15 */ 36, 29, 23, 18, 15,
1182 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1184 * In cases where the weight does not change often, we can use the
1185 * precalculated inverse to speed up arithmetics by turning divisions
1186 * into multiplications:
1188 static const u32 prio_to_wmult
[40] = {
1189 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1190 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1191 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1192 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1193 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1194 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1195 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1196 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1199 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1202 * runqueue iterator, to support SMP load-balancing between different
1203 * scheduling classes, without having to expose their internal data
1204 * structures to the load-balancing proper:
1206 struct rq_iterator
{
1208 struct task_struct
*(*start
)(void *);
1209 struct task_struct
*(*next
)(void *);
1213 static unsigned long
1214 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1215 unsigned long max_load_move
, struct sched_domain
*sd
,
1216 enum cpu_idle_type idle
, int *all_pinned
,
1217 int *this_best_prio
, struct rq_iterator
*iterator
);
1220 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1221 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1222 struct rq_iterator
*iterator
);
1225 #ifdef CONFIG_CGROUP_CPUACCT
1226 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1228 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1231 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1233 update_load_add(&rq
->load
, load
);
1236 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1238 update_load_sub(&rq
->load
, load
);
1242 static unsigned long source_load(int cpu
, int type
);
1243 static unsigned long target_load(int cpu
, int type
);
1244 static unsigned long cpu_avg_load_per_task(int cpu
);
1245 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1246 #endif /* CONFIG_SMP */
1248 #include "sched_stats.h"
1249 #include "sched_idletask.c"
1250 #include "sched_fair.c"
1251 #include "sched_rt.c"
1252 #ifdef CONFIG_SCHED_DEBUG
1253 # include "sched_debug.c"
1256 #define sched_class_highest (&rt_sched_class)
1258 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1263 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1268 static void set_load_weight(struct task_struct
*p
)
1270 if (task_has_rt_policy(p
)) {
1271 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1272 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1277 * SCHED_IDLE tasks get minimal weight:
1279 if (p
->policy
== SCHED_IDLE
) {
1280 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1281 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1285 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1286 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1289 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1291 sched_info_queued(p
);
1292 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1296 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1298 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1303 * __normal_prio - return the priority that is based on the static prio
1305 static inline int __normal_prio(struct task_struct
*p
)
1307 return p
->static_prio
;
1311 * Calculate the expected normal priority: i.e. priority
1312 * without taking RT-inheritance into account. Might be
1313 * boosted by interactivity modifiers. Changes upon fork,
1314 * setprio syscalls, and whenever the interactivity
1315 * estimator recalculates.
1317 static inline int normal_prio(struct task_struct
*p
)
1321 if (task_has_rt_policy(p
))
1322 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1324 prio
= __normal_prio(p
);
1329 * Calculate the current priority, i.e. the priority
1330 * taken into account by the scheduler. This value might
1331 * be boosted by RT tasks, or might be boosted by
1332 * interactivity modifiers. Will be RT if the task got
1333 * RT-boosted. If not then it returns p->normal_prio.
1335 static int effective_prio(struct task_struct
*p
)
1337 p
->normal_prio
= normal_prio(p
);
1339 * If we are RT tasks or we were boosted to RT priority,
1340 * keep the priority unchanged. Otherwise, update priority
1341 * to the normal priority:
1343 if (!rt_prio(p
->prio
))
1344 return p
->normal_prio
;
1349 * activate_task - move a task to the runqueue.
1351 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1353 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1354 rq
->nr_uninterruptible
--;
1356 enqueue_task(rq
, p
, wakeup
);
1357 inc_nr_running(p
, rq
);
1361 * deactivate_task - remove a task from the runqueue.
1363 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1365 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1366 rq
->nr_uninterruptible
++;
1368 dequeue_task(rq
, p
, sleep
);
1369 dec_nr_running(p
, rq
);
1373 * task_curr - is this task currently executing on a CPU?
1374 * @p: the task in question.
1376 inline int task_curr(const struct task_struct
*p
)
1378 return cpu_curr(task_cpu(p
)) == p
;
1381 /* Used instead of source_load when we know the type == 0 */
1382 unsigned long weighted_cpuload(const int cpu
)
1384 return cpu_rq(cpu
)->load
.weight
;
1387 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1389 set_task_rq(p
, cpu
);
1392 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1393 * successfuly executed on another CPU. We must ensure that updates of
1394 * per-task data have been completed by this moment.
1397 task_thread_info(p
)->cpu
= cpu
;
1401 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1402 const struct sched_class
*prev_class
,
1403 int oldprio
, int running
)
1405 if (prev_class
!= p
->sched_class
) {
1406 if (prev_class
->switched_from
)
1407 prev_class
->switched_from(rq
, p
, running
);
1408 p
->sched_class
->switched_to(rq
, p
, running
);
1410 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1416 * Is this task likely cache-hot:
1419 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1423 if (p
->sched_class
!= &fair_sched_class
)
1426 if (sysctl_sched_migration_cost
== -1)
1428 if (sysctl_sched_migration_cost
== 0)
1431 delta
= now
- p
->se
.exec_start
;
1433 return delta
< (s64
)sysctl_sched_migration_cost
;
1437 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1439 int old_cpu
= task_cpu(p
);
1440 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1441 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1442 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1445 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1447 #ifdef CONFIG_SCHEDSTATS
1448 if (p
->se
.wait_start
)
1449 p
->se
.wait_start
-= clock_offset
;
1450 if (p
->se
.sleep_start
)
1451 p
->se
.sleep_start
-= clock_offset
;
1452 if (p
->se
.block_start
)
1453 p
->se
.block_start
-= clock_offset
;
1454 if (old_cpu
!= new_cpu
) {
1455 schedstat_inc(p
, se
.nr_migrations
);
1456 if (task_hot(p
, old_rq
->clock
, NULL
))
1457 schedstat_inc(p
, se
.nr_forced2_migrations
);
1460 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1461 new_cfsrq
->min_vruntime
;
1463 __set_task_cpu(p
, new_cpu
);
1466 struct migration_req
{
1467 struct list_head list
;
1469 struct task_struct
*task
;
1472 struct completion done
;
1476 * The task's runqueue lock must be held.
1477 * Returns true if you have to wait for migration thread.
1480 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1482 struct rq
*rq
= task_rq(p
);
1485 * If the task is not on a runqueue (and not running), then
1486 * it is sufficient to simply update the task's cpu field.
1488 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1489 set_task_cpu(p
, dest_cpu
);
1493 init_completion(&req
->done
);
1495 req
->dest_cpu
= dest_cpu
;
1496 list_add(&req
->list
, &rq
->migration_queue
);
1502 * wait_task_inactive - wait for a thread to unschedule.
1504 * The caller must ensure that the task *will* unschedule sometime soon,
1505 * else this function might spin for a *long* time. This function can't
1506 * be called with interrupts off, or it may introduce deadlock with
1507 * smp_call_function() if an IPI is sent by the same process we are
1508 * waiting to become inactive.
1510 void wait_task_inactive(struct task_struct
*p
)
1512 unsigned long flags
;
1518 * We do the initial early heuristics without holding
1519 * any task-queue locks at all. We'll only try to get
1520 * the runqueue lock when things look like they will
1526 * If the task is actively running on another CPU
1527 * still, just relax and busy-wait without holding
1530 * NOTE! Since we don't hold any locks, it's not
1531 * even sure that "rq" stays as the right runqueue!
1532 * But we don't care, since "task_running()" will
1533 * return false if the runqueue has changed and p
1534 * is actually now running somewhere else!
1536 while (task_running(rq
, p
))
1540 * Ok, time to look more closely! We need the rq
1541 * lock now, to be *sure*. If we're wrong, we'll
1542 * just go back and repeat.
1544 rq
= task_rq_lock(p
, &flags
);
1545 running
= task_running(rq
, p
);
1546 on_rq
= p
->se
.on_rq
;
1547 task_rq_unlock(rq
, &flags
);
1550 * Was it really running after all now that we
1551 * checked with the proper locks actually held?
1553 * Oops. Go back and try again..
1555 if (unlikely(running
)) {
1561 * It's not enough that it's not actively running,
1562 * it must be off the runqueue _entirely_, and not
1565 * So if it wa still runnable (but just not actively
1566 * running right now), it's preempted, and we should
1567 * yield - it could be a while.
1569 if (unlikely(on_rq
)) {
1570 schedule_timeout_uninterruptible(1);
1575 * Ahh, all good. It wasn't running, and it wasn't
1576 * runnable, which means that it will never become
1577 * running in the future either. We're all done!
1584 * kick_process - kick a running thread to enter/exit the kernel
1585 * @p: the to-be-kicked thread
1587 * Cause a process which is running on another CPU to enter
1588 * kernel-mode, without any delay. (to get signals handled.)
1590 * NOTE: this function doesnt have to take the runqueue lock,
1591 * because all it wants to ensure is that the remote task enters
1592 * the kernel. If the IPI races and the task has been migrated
1593 * to another CPU then no harm is done and the purpose has been
1596 void kick_process(struct task_struct
*p
)
1602 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1603 smp_send_reschedule(cpu
);
1608 * Return a low guess at the load of a migration-source cpu weighted
1609 * according to the scheduling class and "nice" value.
1611 * We want to under-estimate the load of migration sources, to
1612 * balance conservatively.
1614 static unsigned long source_load(int cpu
, int type
)
1616 struct rq
*rq
= cpu_rq(cpu
);
1617 unsigned long total
= weighted_cpuload(cpu
);
1622 return min(rq
->cpu_load
[type
-1], total
);
1626 * Return a high guess at the load of a migration-target cpu weighted
1627 * according to the scheduling class and "nice" value.
1629 static unsigned long target_load(int cpu
, int type
)
1631 struct rq
*rq
= cpu_rq(cpu
);
1632 unsigned long total
= weighted_cpuload(cpu
);
1637 return max(rq
->cpu_load
[type
-1], total
);
1641 * Return the average load per task on the cpu's run queue
1643 static unsigned long cpu_avg_load_per_task(int cpu
)
1645 struct rq
*rq
= cpu_rq(cpu
);
1646 unsigned long total
= weighted_cpuload(cpu
);
1647 unsigned long n
= rq
->nr_running
;
1649 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1653 * find_idlest_group finds and returns the least busy CPU group within the
1656 static struct sched_group
*
1657 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1659 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1660 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1661 int load_idx
= sd
->forkexec_idx
;
1662 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1665 unsigned long load
, avg_load
;
1669 /* Skip over this group if it has no CPUs allowed */
1670 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1673 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1675 /* Tally up the load of all CPUs in the group */
1678 for_each_cpu_mask(i
, group
->cpumask
) {
1679 /* Bias balancing toward cpus of our domain */
1681 load
= source_load(i
, load_idx
);
1683 load
= target_load(i
, load_idx
);
1688 /* Adjust by relative CPU power of the group */
1689 avg_load
= sg_div_cpu_power(group
,
1690 avg_load
* SCHED_LOAD_SCALE
);
1693 this_load
= avg_load
;
1695 } else if (avg_load
< min_load
) {
1696 min_load
= avg_load
;
1699 } while (group
= group
->next
, group
!= sd
->groups
);
1701 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1707 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1710 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1713 unsigned long load
, min_load
= ULONG_MAX
;
1717 /* Traverse only the allowed CPUs */
1718 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1720 for_each_cpu_mask(i
, tmp
) {
1721 load
= weighted_cpuload(i
);
1723 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1733 * sched_balance_self: balance the current task (running on cpu) in domains
1734 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1737 * Balance, ie. select the least loaded group.
1739 * Returns the target CPU number, or the same CPU if no balancing is needed.
1741 * preempt must be disabled.
1743 static int sched_balance_self(int cpu
, int flag
)
1745 struct task_struct
*t
= current
;
1746 struct sched_domain
*tmp
, *sd
= NULL
;
1748 for_each_domain(cpu
, tmp
) {
1750 * If power savings logic is enabled for a domain, stop there.
1752 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1754 if (tmp
->flags
& flag
)
1760 struct sched_group
*group
;
1761 int new_cpu
, weight
;
1763 if (!(sd
->flags
& flag
)) {
1769 group
= find_idlest_group(sd
, t
, cpu
);
1775 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1776 if (new_cpu
== -1 || new_cpu
== cpu
) {
1777 /* Now try balancing at a lower domain level of cpu */
1782 /* Now try balancing at a lower domain level of new_cpu */
1785 weight
= cpus_weight(span
);
1786 for_each_domain(cpu
, tmp
) {
1787 if (weight
<= cpus_weight(tmp
->span
))
1789 if (tmp
->flags
& flag
)
1792 /* while loop will break here if sd == NULL */
1798 #endif /* CONFIG_SMP */
1801 * try_to_wake_up - wake up a thread
1802 * @p: the to-be-woken-up thread
1803 * @state: the mask of task states that can be woken
1804 * @sync: do a synchronous wakeup?
1806 * Put it on the run-queue if it's not already there. The "current"
1807 * thread is always on the run-queue (except when the actual
1808 * re-schedule is in progress), and as such you're allowed to do
1809 * the simpler "current->state = TASK_RUNNING" to mark yourself
1810 * runnable without the overhead of this.
1812 * returns failure only if the task is already active.
1814 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1816 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1817 unsigned long flags
;
1821 rq
= task_rq_lock(p
, &flags
);
1822 old_state
= p
->state
;
1823 if (!(old_state
& state
))
1831 this_cpu
= smp_processor_id();
1834 if (unlikely(task_running(rq
, p
)))
1837 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1838 if (cpu
!= orig_cpu
) {
1839 set_task_cpu(p
, cpu
);
1840 task_rq_unlock(rq
, &flags
);
1841 /* might preempt at this point */
1842 rq
= task_rq_lock(p
, &flags
);
1843 old_state
= p
->state
;
1844 if (!(old_state
& state
))
1849 this_cpu
= smp_processor_id();
1853 #ifdef CONFIG_SCHEDSTATS
1854 schedstat_inc(rq
, ttwu_count
);
1855 if (cpu
== this_cpu
)
1856 schedstat_inc(rq
, ttwu_local
);
1858 struct sched_domain
*sd
;
1859 for_each_domain(this_cpu
, sd
) {
1860 if (cpu_isset(cpu
, sd
->span
)) {
1861 schedstat_inc(sd
, ttwu_wake_remote
);
1869 #endif /* CONFIG_SMP */
1870 schedstat_inc(p
, se
.nr_wakeups
);
1872 schedstat_inc(p
, se
.nr_wakeups_sync
);
1873 if (orig_cpu
!= cpu
)
1874 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1875 if (cpu
== this_cpu
)
1876 schedstat_inc(p
, se
.nr_wakeups_local
);
1878 schedstat_inc(p
, se
.nr_wakeups_remote
);
1879 update_rq_clock(rq
);
1880 activate_task(rq
, p
, 1);
1881 check_preempt_curr(rq
, p
);
1885 p
->state
= TASK_RUNNING
;
1887 if (p
->sched_class
->task_wake_up
)
1888 p
->sched_class
->task_wake_up(rq
, p
);
1891 task_rq_unlock(rq
, &flags
);
1896 int fastcall
wake_up_process(struct task_struct
*p
)
1898 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1899 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1901 EXPORT_SYMBOL(wake_up_process
);
1903 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1905 return try_to_wake_up(p
, state
, 0);
1909 * Perform scheduler related setup for a newly forked process p.
1910 * p is forked by current.
1912 * __sched_fork() is basic setup used by init_idle() too:
1914 static void __sched_fork(struct task_struct
*p
)
1916 p
->se
.exec_start
= 0;
1917 p
->se
.sum_exec_runtime
= 0;
1918 p
->se
.prev_sum_exec_runtime
= 0;
1920 #ifdef CONFIG_SCHEDSTATS
1921 p
->se
.wait_start
= 0;
1922 p
->se
.sum_sleep_runtime
= 0;
1923 p
->se
.sleep_start
= 0;
1924 p
->se
.block_start
= 0;
1925 p
->se
.sleep_max
= 0;
1926 p
->se
.block_max
= 0;
1928 p
->se
.slice_max
= 0;
1932 INIT_LIST_HEAD(&p
->rt
.run_list
);
1935 #ifdef CONFIG_PREEMPT_NOTIFIERS
1936 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1940 * We mark the process as running here, but have not actually
1941 * inserted it onto the runqueue yet. This guarantees that
1942 * nobody will actually run it, and a signal or other external
1943 * event cannot wake it up and insert it on the runqueue either.
1945 p
->state
= TASK_RUNNING
;
1949 * fork()/clone()-time setup:
1951 void sched_fork(struct task_struct
*p
, int clone_flags
)
1953 int cpu
= get_cpu();
1958 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1960 set_task_cpu(p
, cpu
);
1963 * Make sure we do not leak PI boosting priority to the child:
1965 p
->prio
= current
->normal_prio
;
1966 if (!rt_prio(p
->prio
))
1967 p
->sched_class
= &fair_sched_class
;
1969 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1970 if (likely(sched_info_on()))
1971 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1973 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1976 #ifdef CONFIG_PREEMPT
1977 /* Want to start with kernel preemption disabled. */
1978 task_thread_info(p
)->preempt_count
= 1;
1984 * wake_up_new_task - wake up a newly created task for the first time.
1986 * This function will do some initial scheduler statistics housekeeping
1987 * that must be done for every newly created context, then puts the task
1988 * on the runqueue and wakes it.
1990 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1992 unsigned long flags
;
1995 rq
= task_rq_lock(p
, &flags
);
1996 BUG_ON(p
->state
!= TASK_RUNNING
);
1997 update_rq_clock(rq
);
1999 p
->prio
= effective_prio(p
);
2001 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2002 activate_task(rq
, p
, 0);
2005 * Let the scheduling class do new task startup
2006 * management (if any):
2008 p
->sched_class
->task_new(rq
, p
);
2009 inc_nr_running(p
, rq
);
2011 check_preempt_curr(rq
, p
);
2013 if (p
->sched_class
->task_wake_up
)
2014 p
->sched_class
->task_wake_up(rq
, p
);
2016 task_rq_unlock(rq
, &flags
);
2019 #ifdef CONFIG_PREEMPT_NOTIFIERS
2022 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2023 * @notifier: notifier struct to register
2025 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2027 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2029 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2032 * preempt_notifier_unregister - no longer interested in preemption notifications
2033 * @notifier: notifier struct to unregister
2035 * This is safe to call from within a preemption notifier.
2037 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2039 hlist_del(¬ifier
->link
);
2041 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2043 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2045 struct preempt_notifier
*notifier
;
2046 struct hlist_node
*node
;
2048 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2049 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2053 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2054 struct task_struct
*next
)
2056 struct preempt_notifier
*notifier
;
2057 struct hlist_node
*node
;
2059 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2060 notifier
->ops
->sched_out(notifier
, next
);
2065 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2070 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2071 struct task_struct
*next
)
2078 * prepare_task_switch - prepare to switch tasks
2079 * @rq: the runqueue preparing to switch
2080 * @prev: the current task that is being switched out
2081 * @next: the task we are going to switch to.
2083 * This is called with the rq lock held and interrupts off. It must
2084 * be paired with a subsequent finish_task_switch after the context
2087 * prepare_task_switch sets up locking and calls architecture specific
2091 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2092 struct task_struct
*next
)
2094 fire_sched_out_preempt_notifiers(prev
, next
);
2095 prepare_lock_switch(rq
, next
);
2096 prepare_arch_switch(next
);
2100 * finish_task_switch - clean up after a task-switch
2101 * @rq: runqueue associated with task-switch
2102 * @prev: the thread we just switched away from.
2104 * finish_task_switch must be called after the context switch, paired
2105 * with a prepare_task_switch call before the context switch.
2106 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2107 * and do any other architecture-specific cleanup actions.
2109 * Note that we may have delayed dropping an mm in context_switch(). If
2110 * so, we finish that here outside of the runqueue lock. (Doing it
2111 * with the lock held can cause deadlocks; see schedule() for
2114 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2115 __releases(rq
->lock
)
2117 struct mm_struct
*mm
= rq
->prev_mm
;
2123 * A task struct has one reference for the use as "current".
2124 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2125 * schedule one last time. The schedule call will never return, and
2126 * the scheduled task must drop that reference.
2127 * The test for TASK_DEAD must occur while the runqueue locks are
2128 * still held, otherwise prev could be scheduled on another cpu, die
2129 * there before we look at prev->state, and then the reference would
2131 * Manfred Spraul <manfred@colorfullife.com>
2133 prev_state
= prev
->state
;
2134 finish_arch_switch(prev
);
2135 finish_lock_switch(rq
, prev
);
2137 if (current
->sched_class
->post_schedule
)
2138 current
->sched_class
->post_schedule(rq
);
2141 fire_sched_in_preempt_notifiers(current
);
2144 if (unlikely(prev_state
== TASK_DEAD
)) {
2146 * Remove function-return probe instances associated with this
2147 * task and put them back on the free list.
2149 kprobe_flush_task(prev
);
2150 put_task_struct(prev
);
2155 * schedule_tail - first thing a freshly forked thread must call.
2156 * @prev: the thread we just switched away from.
2158 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2159 __releases(rq
->lock
)
2161 struct rq
*rq
= this_rq();
2163 finish_task_switch(rq
, prev
);
2164 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2165 /* In this case, finish_task_switch does not reenable preemption */
2168 if (current
->set_child_tid
)
2169 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2173 * context_switch - switch to the new MM and the new
2174 * thread's register state.
2177 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2178 struct task_struct
*next
)
2180 struct mm_struct
*mm
, *oldmm
;
2182 prepare_task_switch(rq
, prev
, next
);
2184 oldmm
= prev
->active_mm
;
2186 * For paravirt, this is coupled with an exit in switch_to to
2187 * combine the page table reload and the switch backend into
2190 arch_enter_lazy_cpu_mode();
2192 if (unlikely(!mm
)) {
2193 next
->active_mm
= oldmm
;
2194 atomic_inc(&oldmm
->mm_count
);
2195 enter_lazy_tlb(oldmm
, next
);
2197 switch_mm(oldmm
, mm
, next
);
2199 if (unlikely(!prev
->mm
)) {
2200 prev
->active_mm
= NULL
;
2201 rq
->prev_mm
= oldmm
;
2204 * Since the runqueue lock will be released by the next
2205 * task (which is an invalid locking op but in the case
2206 * of the scheduler it's an obvious special-case), so we
2207 * do an early lockdep release here:
2209 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2210 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2213 /* Here we just switch the register state and the stack. */
2214 switch_to(prev
, next
, prev
);
2218 * this_rq must be evaluated again because prev may have moved
2219 * CPUs since it called schedule(), thus the 'rq' on its stack
2220 * frame will be invalid.
2222 finish_task_switch(this_rq(), prev
);
2226 * nr_running, nr_uninterruptible and nr_context_switches:
2228 * externally visible scheduler statistics: current number of runnable
2229 * threads, current number of uninterruptible-sleeping threads, total
2230 * number of context switches performed since bootup.
2232 unsigned long nr_running(void)
2234 unsigned long i
, sum
= 0;
2236 for_each_online_cpu(i
)
2237 sum
+= cpu_rq(i
)->nr_running
;
2242 unsigned long nr_uninterruptible(void)
2244 unsigned long i
, sum
= 0;
2246 for_each_possible_cpu(i
)
2247 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2250 * Since we read the counters lockless, it might be slightly
2251 * inaccurate. Do not allow it to go below zero though:
2253 if (unlikely((long)sum
< 0))
2259 unsigned long long nr_context_switches(void)
2262 unsigned long long sum
= 0;
2264 for_each_possible_cpu(i
)
2265 sum
+= cpu_rq(i
)->nr_switches
;
2270 unsigned long nr_iowait(void)
2272 unsigned long i
, sum
= 0;
2274 for_each_possible_cpu(i
)
2275 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2280 unsigned long nr_active(void)
2282 unsigned long i
, running
= 0, uninterruptible
= 0;
2284 for_each_online_cpu(i
) {
2285 running
+= cpu_rq(i
)->nr_running
;
2286 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2289 if (unlikely((long)uninterruptible
< 0))
2290 uninterruptible
= 0;
2292 return running
+ uninterruptible
;
2296 * Update rq->cpu_load[] statistics. This function is usually called every
2297 * scheduler tick (TICK_NSEC).
2299 static void update_cpu_load(struct rq
*this_rq
)
2301 unsigned long this_load
= this_rq
->load
.weight
;
2304 this_rq
->nr_load_updates
++;
2306 /* Update our load: */
2307 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2308 unsigned long old_load
, new_load
;
2310 /* scale is effectively 1 << i now, and >> i divides by scale */
2312 old_load
= this_rq
->cpu_load
[i
];
2313 new_load
= this_load
;
2315 * Round up the averaging division if load is increasing. This
2316 * prevents us from getting stuck on 9 if the load is 10, for
2319 if (new_load
> old_load
)
2320 new_load
+= scale
-1;
2321 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2328 * double_rq_lock - safely lock two runqueues
2330 * Note this does not disable interrupts like task_rq_lock,
2331 * you need to do so manually before calling.
2333 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2334 __acquires(rq1
->lock
)
2335 __acquires(rq2
->lock
)
2337 BUG_ON(!irqs_disabled());
2339 spin_lock(&rq1
->lock
);
2340 __acquire(rq2
->lock
); /* Fake it out ;) */
2343 spin_lock(&rq1
->lock
);
2344 spin_lock(&rq2
->lock
);
2346 spin_lock(&rq2
->lock
);
2347 spin_lock(&rq1
->lock
);
2350 update_rq_clock(rq1
);
2351 update_rq_clock(rq2
);
2355 * double_rq_unlock - safely unlock two runqueues
2357 * Note this does not restore interrupts like task_rq_unlock,
2358 * you need to do so manually after calling.
2360 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2361 __releases(rq1
->lock
)
2362 __releases(rq2
->lock
)
2364 spin_unlock(&rq1
->lock
);
2366 spin_unlock(&rq2
->lock
);
2368 __release(rq2
->lock
);
2372 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2374 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2375 __releases(this_rq
->lock
)
2376 __acquires(busiest
->lock
)
2377 __acquires(this_rq
->lock
)
2381 if (unlikely(!irqs_disabled())) {
2382 /* printk() doesn't work good under rq->lock */
2383 spin_unlock(&this_rq
->lock
);
2386 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2387 if (busiest
< this_rq
) {
2388 spin_unlock(&this_rq
->lock
);
2389 spin_lock(&busiest
->lock
);
2390 spin_lock(&this_rq
->lock
);
2393 spin_lock(&busiest
->lock
);
2399 * If dest_cpu is allowed for this process, migrate the task to it.
2400 * This is accomplished by forcing the cpu_allowed mask to only
2401 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2402 * the cpu_allowed mask is restored.
2404 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2406 struct migration_req req
;
2407 unsigned long flags
;
2410 rq
= task_rq_lock(p
, &flags
);
2411 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2412 || unlikely(cpu_is_offline(dest_cpu
)))
2415 /* force the process onto the specified CPU */
2416 if (migrate_task(p
, dest_cpu
, &req
)) {
2417 /* Need to wait for migration thread (might exit: take ref). */
2418 struct task_struct
*mt
= rq
->migration_thread
;
2420 get_task_struct(mt
);
2421 task_rq_unlock(rq
, &flags
);
2422 wake_up_process(mt
);
2423 put_task_struct(mt
);
2424 wait_for_completion(&req
.done
);
2429 task_rq_unlock(rq
, &flags
);
2433 * sched_exec - execve() is a valuable balancing opportunity, because at
2434 * this point the task has the smallest effective memory and cache footprint.
2436 void sched_exec(void)
2438 int new_cpu
, this_cpu
= get_cpu();
2439 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2441 if (new_cpu
!= this_cpu
)
2442 sched_migrate_task(current
, new_cpu
);
2446 * pull_task - move a task from a remote runqueue to the local runqueue.
2447 * Both runqueues must be locked.
2449 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2450 struct rq
*this_rq
, int this_cpu
)
2452 deactivate_task(src_rq
, p
, 0);
2453 set_task_cpu(p
, this_cpu
);
2454 activate_task(this_rq
, p
, 0);
2456 * Note that idle threads have a prio of MAX_PRIO, for this test
2457 * to be always true for them.
2459 check_preempt_curr(this_rq
, p
);
2463 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2466 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2467 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2471 * We do not migrate tasks that are:
2472 * 1) running (obviously), or
2473 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2474 * 3) are cache-hot on their current CPU.
2476 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2477 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2482 if (task_running(rq
, p
)) {
2483 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2488 * Aggressive migration if:
2489 * 1) task is cache cold, or
2490 * 2) too many balance attempts have failed.
2493 if (!task_hot(p
, rq
->clock
, sd
) ||
2494 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2495 #ifdef CONFIG_SCHEDSTATS
2496 if (task_hot(p
, rq
->clock
, sd
)) {
2497 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2498 schedstat_inc(p
, se
.nr_forced_migrations
);
2504 if (task_hot(p
, rq
->clock
, sd
)) {
2505 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2511 static unsigned long
2512 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2513 unsigned long max_load_move
, struct sched_domain
*sd
,
2514 enum cpu_idle_type idle
, int *all_pinned
,
2515 int *this_best_prio
, struct rq_iterator
*iterator
)
2517 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2518 struct task_struct
*p
;
2519 long rem_load_move
= max_load_move
;
2521 if (max_load_move
== 0)
2527 * Start the load-balancing iterator:
2529 p
= iterator
->start(iterator
->arg
);
2531 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2534 * To help distribute high priority tasks across CPUs we don't
2535 * skip a task if it will be the highest priority task (i.e. smallest
2536 * prio value) on its new queue regardless of its load weight
2538 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2539 SCHED_LOAD_SCALE_FUZZ
;
2540 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2541 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2542 p
= iterator
->next(iterator
->arg
);
2546 pull_task(busiest
, p
, this_rq
, this_cpu
);
2548 rem_load_move
-= p
->se
.load
.weight
;
2551 * We only want to steal up to the prescribed amount of weighted load.
2553 if (rem_load_move
> 0) {
2554 if (p
->prio
< *this_best_prio
)
2555 *this_best_prio
= p
->prio
;
2556 p
= iterator
->next(iterator
->arg
);
2561 * Right now, this is one of only two places pull_task() is called,
2562 * so we can safely collect pull_task() stats here rather than
2563 * inside pull_task().
2565 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2568 *all_pinned
= pinned
;
2570 return max_load_move
- rem_load_move
;
2574 * move_tasks tries to move up to max_load_move weighted load from busiest to
2575 * this_rq, as part of a balancing operation within domain "sd".
2576 * Returns 1 if successful and 0 otherwise.
2578 * Called with both runqueues locked.
2580 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2581 unsigned long max_load_move
,
2582 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2585 const struct sched_class
*class = sched_class_highest
;
2586 unsigned long total_load_moved
= 0;
2587 int this_best_prio
= this_rq
->curr
->prio
;
2591 class->load_balance(this_rq
, this_cpu
, busiest
,
2592 max_load_move
- total_load_moved
,
2593 sd
, idle
, all_pinned
, &this_best_prio
);
2594 class = class->next
;
2595 } while (class && max_load_move
> total_load_moved
);
2597 return total_load_moved
> 0;
2601 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2602 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2603 struct rq_iterator
*iterator
)
2605 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2609 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2610 pull_task(busiest
, p
, this_rq
, this_cpu
);
2612 * Right now, this is only the second place pull_task()
2613 * is called, so we can safely collect pull_task()
2614 * stats here rather than inside pull_task().
2616 schedstat_inc(sd
, lb_gained
[idle
]);
2620 p
= iterator
->next(iterator
->arg
);
2627 * move_one_task tries to move exactly one task from busiest to this_rq, as
2628 * part of active balancing operations within "domain".
2629 * Returns 1 if successful and 0 otherwise.
2631 * Called with both runqueues locked.
2633 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2634 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2636 const struct sched_class
*class;
2638 for (class = sched_class_highest
; class; class = class->next
)
2639 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2646 * find_busiest_group finds and returns the busiest CPU group within the
2647 * domain. It calculates and returns the amount of weighted load which
2648 * should be moved to restore balance via the imbalance parameter.
2650 static struct sched_group
*
2651 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2652 unsigned long *imbalance
, enum cpu_idle_type idle
,
2653 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2655 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2656 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2657 unsigned long max_pull
;
2658 unsigned long busiest_load_per_task
, busiest_nr_running
;
2659 unsigned long this_load_per_task
, this_nr_running
;
2660 int load_idx
, group_imb
= 0;
2661 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2662 int power_savings_balance
= 1;
2663 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2664 unsigned long min_nr_running
= ULONG_MAX
;
2665 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2668 max_load
= this_load
= total_load
= total_pwr
= 0;
2669 busiest_load_per_task
= busiest_nr_running
= 0;
2670 this_load_per_task
= this_nr_running
= 0;
2671 if (idle
== CPU_NOT_IDLE
)
2672 load_idx
= sd
->busy_idx
;
2673 else if (idle
== CPU_NEWLY_IDLE
)
2674 load_idx
= sd
->newidle_idx
;
2676 load_idx
= sd
->idle_idx
;
2679 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2682 int __group_imb
= 0;
2683 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2684 unsigned long sum_nr_running
, sum_weighted_load
;
2686 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2689 balance_cpu
= first_cpu(group
->cpumask
);
2691 /* Tally up the load of all CPUs in the group */
2692 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2694 min_cpu_load
= ~0UL;
2696 for_each_cpu_mask(i
, group
->cpumask
) {
2699 if (!cpu_isset(i
, *cpus
))
2704 if (*sd_idle
&& rq
->nr_running
)
2707 /* Bias balancing toward cpus of our domain */
2709 if (idle_cpu(i
) && !first_idle_cpu
) {
2714 load
= target_load(i
, load_idx
);
2716 load
= source_load(i
, load_idx
);
2717 if (load
> max_cpu_load
)
2718 max_cpu_load
= load
;
2719 if (min_cpu_load
> load
)
2720 min_cpu_load
= load
;
2724 sum_nr_running
+= rq
->nr_running
;
2725 sum_weighted_load
+= weighted_cpuload(i
);
2729 * First idle cpu or the first cpu(busiest) in this sched group
2730 * is eligible for doing load balancing at this and above
2731 * domains. In the newly idle case, we will allow all the cpu's
2732 * to do the newly idle load balance.
2734 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2735 balance_cpu
!= this_cpu
&& balance
) {
2740 total_load
+= avg_load
;
2741 total_pwr
+= group
->__cpu_power
;
2743 /* Adjust by relative CPU power of the group */
2744 avg_load
= sg_div_cpu_power(group
,
2745 avg_load
* SCHED_LOAD_SCALE
);
2747 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2750 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2753 this_load
= avg_load
;
2755 this_nr_running
= sum_nr_running
;
2756 this_load_per_task
= sum_weighted_load
;
2757 } else if (avg_load
> max_load
&&
2758 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2759 max_load
= avg_load
;
2761 busiest_nr_running
= sum_nr_running
;
2762 busiest_load_per_task
= sum_weighted_load
;
2763 group_imb
= __group_imb
;
2766 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2768 * Busy processors will not participate in power savings
2771 if (idle
== CPU_NOT_IDLE
||
2772 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2776 * If the local group is idle or completely loaded
2777 * no need to do power savings balance at this domain
2779 if (local_group
&& (this_nr_running
>= group_capacity
||
2781 power_savings_balance
= 0;
2784 * If a group is already running at full capacity or idle,
2785 * don't include that group in power savings calculations
2787 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2792 * Calculate the group which has the least non-idle load.
2793 * This is the group from where we need to pick up the load
2796 if ((sum_nr_running
< min_nr_running
) ||
2797 (sum_nr_running
== min_nr_running
&&
2798 first_cpu(group
->cpumask
) <
2799 first_cpu(group_min
->cpumask
))) {
2801 min_nr_running
= sum_nr_running
;
2802 min_load_per_task
= sum_weighted_load
/
2807 * Calculate the group which is almost near its
2808 * capacity but still has some space to pick up some load
2809 * from other group and save more power
2811 if (sum_nr_running
<= group_capacity
- 1) {
2812 if (sum_nr_running
> leader_nr_running
||
2813 (sum_nr_running
== leader_nr_running
&&
2814 first_cpu(group
->cpumask
) >
2815 first_cpu(group_leader
->cpumask
))) {
2816 group_leader
= group
;
2817 leader_nr_running
= sum_nr_running
;
2822 group
= group
->next
;
2823 } while (group
!= sd
->groups
);
2825 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2828 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2830 if (this_load
>= avg_load
||
2831 100*max_load
<= sd
->imbalance_pct
*this_load
)
2834 busiest_load_per_task
/= busiest_nr_running
;
2836 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2839 * We're trying to get all the cpus to the average_load, so we don't
2840 * want to push ourselves above the average load, nor do we wish to
2841 * reduce the max loaded cpu below the average load, as either of these
2842 * actions would just result in more rebalancing later, and ping-pong
2843 * tasks around. Thus we look for the minimum possible imbalance.
2844 * Negative imbalances (*we* are more loaded than anyone else) will
2845 * be counted as no imbalance for these purposes -- we can't fix that
2846 * by pulling tasks to us. Be careful of negative numbers as they'll
2847 * appear as very large values with unsigned longs.
2849 if (max_load
<= busiest_load_per_task
)
2853 * In the presence of smp nice balancing, certain scenarios can have
2854 * max load less than avg load(as we skip the groups at or below
2855 * its cpu_power, while calculating max_load..)
2857 if (max_load
< avg_load
) {
2859 goto small_imbalance
;
2862 /* Don't want to pull so many tasks that a group would go idle */
2863 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2865 /* How much load to actually move to equalise the imbalance */
2866 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2867 (avg_load
- this_load
) * this->__cpu_power
)
2871 * if *imbalance is less than the average load per runnable task
2872 * there is no gaurantee that any tasks will be moved so we'll have
2873 * a think about bumping its value to force at least one task to be
2876 if (*imbalance
< busiest_load_per_task
) {
2877 unsigned long tmp
, pwr_now
, pwr_move
;
2881 pwr_move
= pwr_now
= 0;
2883 if (this_nr_running
) {
2884 this_load_per_task
/= this_nr_running
;
2885 if (busiest_load_per_task
> this_load_per_task
)
2888 this_load_per_task
= SCHED_LOAD_SCALE
;
2890 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2891 busiest_load_per_task
* imbn
) {
2892 *imbalance
= busiest_load_per_task
;
2897 * OK, we don't have enough imbalance to justify moving tasks,
2898 * however we may be able to increase total CPU power used by
2902 pwr_now
+= busiest
->__cpu_power
*
2903 min(busiest_load_per_task
, max_load
);
2904 pwr_now
+= this->__cpu_power
*
2905 min(this_load_per_task
, this_load
);
2906 pwr_now
/= SCHED_LOAD_SCALE
;
2908 /* Amount of load we'd subtract */
2909 tmp
= sg_div_cpu_power(busiest
,
2910 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2912 pwr_move
+= busiest
->__cpu_power
*
2913 min(busiest_load_per_task
, max_load
- tmp
);
2915 /* Amount of load we'd add */
2916 if (max_load
* busiest
->__cpu_power
<
2917 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2918 tmp
= sg_div_cpu_power(this,
2919 max_load
* busiest
->__cpu_power
);
2921 tmp
= sg_div_cpu_power(this,
2922 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2923 pwr_move
+= this->__cpu_power
*
2924 min(this_load_per_task
, this_load
+ tmp
);
2925 pwr_move
/= SCHED_LOAD_SCALE
;
2927 /* Move if we gain throughput */
2928 if (pwr_move
> pwr_now
)
2929 *imbalance
= busiest_load_per_task
;
2935 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2936 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2939 if (this == group_leader
&& group_leader
!= group_min
) {
2940 *imbalance
= min_load_per_task
;
2950 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2953 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2954 unsigned long imbalance
, cpumask_t
*cpus
)
2956 struct rq
*busiest
= NULL
, *rq
;
2957 unsigned long max_load
= 0;
2960 for_each_cpu_mask(i
, group
->cpumask
) {
2963 if (!cpu_isset(i
, *cpus
))
2967 wl
= weighted_cpuload(i
);
2969 if (rq
->nr_running
== 1 && wl
> imbalance
)
2972 if (wl
> max_load
) {
2982 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2983 * so long as it is large enough.
2985 #define MAX_PINNED_INTERVAL 512
2988 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2989 * tasks if there is an imbalance.
2991 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2992 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2995 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2996 struct sched_group
*group
;
2997 unsigned long imbalance
;
2999 cpumask_t cpus
= CPU_MASK_ALL
;
3000 unsigned long flags
;
3003 * When power savings policy is enabled for the parent domain, idle
3004 * sibling can pick up load irrespective of busy siblings. In this case,
3005 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3006 * portraying it as CPU_NOT_IDLE.
3008 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3009 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3012 schedstat_inc(sd
, lb_count
[idle
]);
3015 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3022 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3026 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
3028 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3032 BUG_ON(busiest
== this_rq
);
3034 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3037 if (busiest
->nr_running
> 1) {
3039 * Attempt to move tasks. If find_busiest_group has found
3040 * an imbalance but busiest->nr_running <= 1, the group is
3041 * still unbalanced. ld_moved simply stays zero, so it is
3042 * correctly treated as an imbalance.
3044 local_irq_save(flags
);
3045 double_rq_lock(this_rq
, busiest
);
3046 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3047 imbalance
, sd
, idle
, &all_pinned
);
3048 double_rq_unlock(this_rq
, busiest
);
3049 local_irq_restore(flags
);
3052 * some other cpu did the load balance for us.
3054 if (ld_moved
&& this_cpu
!= smp_processor_id())
3055 resched_cpu(this_cpu
);
3057 /* All tasks on this runqueue were pinned by CPU affinity */
3058 if (unlikely(all_pinned
)) {
3059 cpu_clear(cpu_of(busiest
), cpus
);
3060 if (!cpus_empty(cpus
))
3067 schedstat_inc(sd
, lb_failed
[idle
]);
3068 sd
->nr_balance_failed
++;
3070 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3072 spin_lock_irqsave(&busiest
->lock
, flags
);
3074 /* don't kick the migration_thread, if the curr
3075 * task on busiest cpu can't be moved to this_cpu
3077 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3078 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3080 goto out_one_pinned
;
3083 if (!busiest
->active_balance
) {
3084 busiest
->active_balance
= 1;
3085 busiest
->push_cpu
= this_cpu
;
3088 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3090 wake_up_process(busiest
->migration_thread
);
3093 * We've kicked active balancing, reset the failure
3096 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3099 sd
->nr_balance_failed
= 0;
3101 if (likely(!active_balance
)) {
3102 /* We were unbalanced, so reset the balancing interval */
3103 sd
->balance_interval
= sd
->min_interval
;
3106 * If we've begun active balancing, start to back off. This
3107 * case may not be covered by the all_pinned logic if there
3108 * is only 1 task on the busy runqueue (because we don't call
3111 if (sd
->balance_interval
< sd
->max_interval
)
3112 sd
->balance_interval
*= 2;
3115 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3116 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3121 schedstat_inc(sd
, lb_balanced
[idle
]);
3123 sd
->nr_balance_failed
= 0;
3126 /* tune up the balancing interval */
3127 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3128 (sd
->balance_interval
< sd
->max_interval
))
3129 sd
->balance_interval
*= 2;
3131 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3132 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3138 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3139 * tasks if there is an imbalance.
3141 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3142 * this_rq is locked.
3145 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
3147 struct sched_group
*group
;
3148 struct rq
*busiest
= NULL
;
3149 unsigned long imbalance
;
3153 cpumask_t cpus
= CPU_MASK_ALL
;
3156 * When power savings policy is enabled for the parent domain, idle
3157 * sibling can pick up load irrespective of busy siblings. In this case,
3158 * let the state of idle sibling percolate up as IDLE, instead of
3159 * portraying it as CPU_NOT_IDLE.
3161 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3162 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3165 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3167 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3168 &sd_idle
, &cpus
, NULL
);
3170 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3174 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
3177 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3181 BUG_ON(busiest
== this_rq
);
3183 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3186 if (busiest
->nr_running
> 1) {
3187 /* Attempt to move tasks */
3188 double_lock_balance(this_rq
, busiest
);
3189 /* this_rq->clock is already updated */
3190 update_rq_clock(busiest
);
3191 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3192 imbalance
, sd
, CPU_NEWLY_IDLE
,
3194 spin_unlock(&busiest
->lock
);
3196 if (unlikely(all_pinned
)) {
3197 cpu_clear(cpu_of(busiest
), cpus
);
3198 if (!cpus_empty(cpus
))
3204 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3205 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3206 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3209 sd
->nr_balance_failed
= 0;
3214 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3215 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3216 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3218 sd
->nr_balance_failed
= 0;
3224 * idle_balance is called by schedule() if this_cpu is about to become
3225 * idle. Attempts to pull tasks from other CPUs.
3227 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3229 struct sched_domain
*sd
;
3230 int pulled_task
= -1;
3231 unsigned long next_balance
= jiffies
+ HZ
;
3233 for_each_domain(this_cpu
, sd
) {
3234 unsigned long interval
;
3236 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3239 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3240 /* If we've pulled tasks over stop searching: */
3241 pulled_task
= load_balance_newidle(this_cpu
,
3244 interval
= msecs_to_jiffies(sd
->balance_interval
);
3245 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3246 next_balance
= sd
->last_balance
+ interval
;
3250 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3252 * We are going idle. next_balance may be set based on
3253 * a busy processor. So reset next_balance.
3255 this_rq
->next_balance
= next_balance
;
3260 * active_load_balance is run by migration threads. It pushes running tasks
3261 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3262 * running on each physical CPU where possible, and avoids physical /
3263 * logical imbalances.
3265 * Called with busiest_rq locked.
3267 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3269 int target_cpu
= busiest_rq
->push_cpu
;
3270 struct sched_domain
*sd
;
3271 struct rq
*target_rq
;
3273 /* Is there any task to move? */
3274 if (busiest_rq
->nr_running
<= 1)
3277 target_rq
= cpu_rq(target_cpu
);
3280 * This condition is "impossible", if it occurs
3281 * we need to fix it. Originally reported by
3282 * Bjorn Helgaas on a 128-cpu setup.
3284 BUG_ON(busiest_rq
== target_rq
);
3286 /* move a task from busiest_rq to target_rq */
3287 double_lock_balance(busiest_rq
, target_rq
);
3288 update_rq_clock(busiest_rq
);
3289 update_rq_clock(target_rq
);
3291 /* Search for an sd spanning us and the target CPU. */
3292 for_each_domain(target_cpu
, sd
) {
3293 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3294 cpu_isset(busiest_cpu
, sd
->span
))
3299 schedstat_inc(sd
, alb_count
);
3301 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3303 schedstat_inc(sd
, alb_pushed
);
3305 schedstat_inc(sd
, alb_failed
);
3307 spin_unlock(&target_rq
->lock
);
3312 atomic_t load_balancer
;
3314 } nohz ____cacheline_aligned
= {
3315 .load_balancer
= ATOMIC_INIT(-1),
3316 .cpu_mask
= CPU_MASK_NONE
,
3320 * This routine will try to nominate the ilb (idle load balancing)
3321 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3322 * load balancing on behalf of all those cpus. If all the cpus in the system
3323 * go into this tickless mode, then there will be no ilb owner (as there is
3324 * no need for one) and all the cpus will sleep till the next wakeup event
3327 * For the ilb owner, tick is not stopped. And this tick will be used
3328 * for idle load balancing. ilb owner will still be part of
3331 * While stopping the tick, this cpu will become the ilb owner if there
3332 * is no other owner. And will be the owner till that cpu becomes busy
3333 * or if all cpus in the system stop their ticks at which point
3334 * there is no need for ilb owner.
3336 * When the ilb owner becomes busy, it nominates another owner, during the
3337 * next busy scheduler_tick()
3339 int select_nohz_load_balancer(int stop_tick
)
3341 int cpu
= smp_processor_id();
3344 cpu_set(cpu
, nohz
.cpu_mask
);
3345 cpu_rq(cpu
)->in_nohz_recently
= 1;
3348 * If we are going offline and still the leader, give up!
3350 if (cpu_is_offline(cpu
) &&
3351 atomic_read(&nohz
.load_balancer
) == cpu
) {
3352 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3357 /* time for ilb owner also to sleep */
3358 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3359 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3360 atomic_set(&nohz
.load_balancer
, -1);
3364 if (atomic_read(&nohz
.load_balancer
) == -1) {
3365 /* make me the ilb owner */
3366 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3368 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3371 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3374 cpu_clear(cpu
, nohz
.cpu_mask
);
3376 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3377 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3384 static DEFINE_SPINLOCK(balancing
);
3387 * It checks each scheduling domain to see if it is due to be balanced,
3388 * and initiates a balancing operation if so.
3390 * Balancing parameters are set up in arch_init_sched_domains.
3392 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3395 struct rq
*rq
= cpu_rq(cpu
);
3396 unsigned long interval
;
3397 struct sched_domain
*sd
;
3398 /* Earliest time when we have to do rebalance again */
3399 unsigned long next_balance
= jiffies
+ 60*HZ
;
3400 int update_next_balance
= 0;
3402 for_each_domain(cpu
, sd
) {
3403 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3406 interval
= sd
->balance_interval
;
3407 if (idle
!= CPU_IDLE
)
3408 interval
*= sd
->busy_factor
;
3410 /* scale ms to jiffies */
3411 interval
= msecs_to_jiffies(interval
);
3412 if (unlikely(!interval
))
3414 if (interval
> HZ
*NR_CPUS
/10)
3415 interval
= HZ
*NR_CPUS
/10;
3418 if (sd
->flags
& SD_SERIALIZE
) {
3419 if (!spin_trylock(&balancing
))
3423 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3424 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3426 * We've pulled tasks over so either we're no
3427 * longer idle, or one of our SMT siblings is
3430 idle
= CPU_NOT_IDLE
;
3432 sd
->last_balance
= jiffies
;
3434 if (sd
->flags
& SD_SERIALIZE
)
3435 spin_unlock(&balancing
);
3437 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3438 next_balance
= sd
->last_balance
+ interval
;
3439 update_next_balance
= 1;
3443 * Stop the load balance at this level. There is another
3444 * CPU in our sched group which is doing load balancing more
3452 * next_balance will be updated only when there is a need.
3453 * When the cpu is attached to null domain for ex, it will not be
3456 if (likely(update_next_balance
))
3457 rq
->next_balance
= next_balance
;
3461 * run_rebalance_domains is triggered when needed from the scheduler tick.
3462 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3463 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3465 static void run_rebalance_domains(struct softirq_action
*h
)
3467 int this_cpu
= smp_processor_id();
3468 struct rq
*this_rq
= cpu_rq(this_cpu
);
3469 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3470 CPU_IDLE
: CPU_NOT_IDLE
;
3472 rebalance_domains(this_cpu
, idle
);
3476 * If this cpu is the owner for idle load balancing, then do the
3477 * balancing on behalf of the other idle cpus whose ticks are
3480 if (this_rq
->idle_at_tick
&&
3481 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3482 cpumask_t cpus
= nohz
.cpu_mask
;
3486 cpu_clear(this_cpu
, cpus
);
3487 for_each_cpu_mask(balance_cpu
, cpus
) {
3489 * If this cpu gets work to do, stop the load balancing
3490 * work being done for other cpus. Next load
3491 * balancing owner will pick it up.
3496 rebalance_domains(balance_cpu
, CPU_IDLE
);
3498 rq
= cpu_rq(balance_cpu
);
3499 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3500 this_rq
->next_balance
= rq
->next_balance
;
3507 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3509 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3510 * idle load balancing owner or decide to stop the periodic load balancing,
3511 * if the whole system is idle.
3513 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3517 * If we were in the nohz mode recently and busy at the current
3518 * scheduler tick, then check if we need to nominate new idle
3521 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3522 rq
->in_nohz_recently
= 0;
3524 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3525 cpu_clear(cpu
, nohz
.cpu_mask
);
3526 atomic_set(&nohz
.load_balancer
, -1);
3529 if (atomic_read(&nohz
.load_balancer
) == -1) {
3531 * simple selection for now: Nominate the
3532 * first cpu in the nohz list to be the next
3535 * TBD: Traverse the sched domains and nominate
3536 * the nearest cpu in the nohz.cpu_mask.
3538 int ilb
= first_cpu(nohz
.cpu_mask
);
3546 * If this cpu is idle and doing idle load balancing for all the
3547 * cpus with ticks stopped, is it time for that to stop?
3549 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3550 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3556 * If this cpu is idle and the idle load balancing is done by
3557 * someone else, then no need raise the SCHED_SOFTIRQ
3559 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3560 cpu_isset(cpu
, nohz
.cpu_mask
))
3563 if (time_after_eq(jiffies
, rq
->next_balance
))
3564 raise_softirq(SCHED_SOFTIRQ
);
3567 #else /* CONFIG_SMP */
3570 * on UP we do not need to balance between CPUs:
3572 static inline void idle_balance(int cpu
, struct rq
*rq
)
3578 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3580 EXPORT_PER_CPU_SYMBOL(kstat
);
3583 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3584 * that have not yet been banked in case the task is currently running.
3586 unsigned long long task_sched_runtime(struct task_struct
*p
)
3588 unsigned long flags
;
3592 rq
= task_rq_lock(p
, &flags
);
3593 ns
= p
->se
.sum_exec_runtime
;
3594 if (task_current(rq
, p
)) {
3595 update_rq_clock(rq
);
3596 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3597 if ((s64
)delta_exec
> 0)
3600 task_rq_unlock(rq
, &flags
);
3606 * Account user cpu time to a process.
3607 * @p: the process that the cpu time gets accounted to
3608 * @cputime: the cpu time spent in user space since the last update
3610 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3612 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3615 p
->utime
= cputime_add(p
->utime
, cputime
);
3617 /* Add user time to cpustat. */
3618 tmp
= cputime_to_cputime64(cputime
);
3619 if (TASK_NICE(p
) > 0)
3620 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3622 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3626 * Account guest cpu time to a process.
3627 * @p: the process that the cpu time gets accounted to
3628 * @cputime: the cpu time spent in virtual machine since the last update
3630 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3633 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3635 tmp
= cputime_to_cputime64(cputime
);
3637 p
->utime
= cputime_add(p
->utime
, cputime
);
3638 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3640 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3641 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3645 * Account scaled user cpu time to a process.
3646 * @p: the process that the cpu time gets accounted to
3647 * @cputime: the cpu time spent in user space since the last update
3649 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3651 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3655 * Account system cpu time to a process.
3656 * @p: the process that the cpu time gets accounted to
3657 * @hardirq_offset: the offset to subtract from hardirq_count()
3658 * @cputime: the cpu time spent in kernel space since the last update
3660 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3663 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3664 struct rq
*rq
= this_rq();
3667 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3668 return account_guest_time(p
, cputime
);
3670 p
->stime
= cputime_add(p
->stime
, cputime
);
3672 /* Add system time to cpustat. */
3673 tmp
= cputime_to_cputime64(cputime
);
3674 if (hardirq_count() - hardirq_offset
)
3675 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3676 else if (softirq_count())
3677 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3678 else if (p
!= rq
->idle
)
3679 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3680 else if (atomic_read(&rq
->nr_iowait
) > 0)
3681 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3683 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3684 /* Account for system time used */
3685 acct_update_integrals(p
);
3689 * Account scaled system cpu time to a process.
3690 * @p: the process that the cpu time gets accounted to
3691 * @hardirq_offset: the offset to subtract from hardirq_count()
3692 * @cputime: the cpu time spent in kernel space since the last update
3694 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3696 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3700 * Account for involuntary wait time.
3701 * @p: the process from which the cpu time has been stolen
3702 * @steal: the cpu time spent in involuntary wait
3704 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3706 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3707 cputime64_t tmp
= cputime_to_cputime64(steal
);
3708 struct rq
*rq
= this_rq();
3710 if (p
== rq
->idle
) {
3711 p
->stime
= cputime_add(p
->stime
, steal
);
3712 if (atomic_read(&rq
->nr_iowait
) > 0)
3713 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3715 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3717 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3721 * This function gets called by the timer code, with HZ frequency.
3722 * We call it with interrupts disabled.
3724 * It also gets called by the fork code, when changing the parent's
3727 void scheduler_tick(void)
3729 int cpu
= smp_processor_id();
3730 struct rq
*rq
= cpu_rq(cpu
);
3731 struct task_struct
*curr
= rq
->curr
;
3732 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3734 spin_lock(&rq
->lock
);
3735 __update_rq_clock(rq
);
3737 * Let rq->clock advance by at least TICK_NSEC:
3739 if (unlikely(rq
->clock
< next_tick
))
3740 rq
->clock
= next_tick
;
3741 rq
->tick_timestamp
= rq
->clock
;
3742 update_cpu_load(rq
);
3743 curr
->sched_class
->task_tick(rq
, curr
, 0);
3744 update_sched_rt_period(rq
);
3745 spin_unlock(&rq
->lock
);
3748 rq
->idle_at_tick
= idle_cpu(cpu
);
3749 trigger_load_balance(rq
, cpu
);
3753 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3755 void fastcall
add_preempt_count(int val
)
3760 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3762 preempt_count() += val
;
3764 * Spinlock count overflowing soon?
3766 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3769 EXPORT_SYMBOL(add_preempt_count
);
3771 void fastcall
sub_preempt_count(int val
)
3776 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3779 * Is the spinlock portion underflowing?
3781 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3782 !(preempt_count() & PREEMPT_MASK
)))
3785 preempt_count() -= val
;
3787 EXPORT_SYMBOL(sub_preempt_count
);
3792 * Print scheduling while atomic bug:
3794 static noinline
void __schedule_bug(struct task_struct
*prev
)
3796 struct pt_regs
*regs
= get_irq_regs();
3798 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3799 prev
->comm
, prev
->pid
, preempt_count());
3801 debug_show_held_locks(prev
);
3802 if (irqs_disabled())
3803 print_irqtrace_events(prev
);
3812 * Various schedule()-time debugging checks and statistics:
3814 static inline void schedule_debug(struct task_struct
*prev
)
3817 * Test if we are atomic. Since do_exit() needs to call into
3818 * schedule() atomically, we ignore that path for now.
3819 * Otherwise, whine if we are scheduling when we should not be.
3821 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3822 __schedule_bug(prev
);
3824 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3826 schedstat_inc(this_rq(), sched_count
);
3827 #ifdef CONFIG_SCHEDSTATS
3828 if (unlikely(prev
->lock_depth
>= 0)) {
3829 schedstat_inc(this_rq(), bkl_count
);
3830 schedstat_inc(prev
, sched_info
.bkl_count
);
3836 * Pick up the highest-prio task:
3838 static inline struct task_struct
*
3839 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3841 const struct sched_class
*class;
3842 struct task_struct
*p
;
3845 * Optimization: we know that if all tasks are in
3846 * the fair class we can call that function directly:
3848 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3849 p
= fair_sched_class
.pick_next_task(rq
);
3854 class = sched_class_highest
;
3856 p
= class->pick_next_task(rq
);
3860 * Will never be NULL as the idle class always
3861 * returns a non-NULL p:
3863 class = class->next
;
3868 * schedule() is the main scheduler function.
3870 asmlinkage
void __sched
schedule(void)
3872 struct task_struct
*prev
, *next
;
3879 cpu
= smp_processor_id();
3883 switch_count
= &prev
->nivcsw
;
3885 release_kernel_lock(prev
);
3886 need_resched_nonpreemptible
:
3888 schedule_debug(prev
);
3893 * Do the rq-clock update outside the rq lock:
3895 local_irq_disable();
3896 __update_rq_clock(rq
);
3897 spin_lock(&rq
->lock
);
3898 clear_tsk_need_resched(prev
);
3900 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3901 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3902 unlikely(signal_pending(prev
)))) {
3903 prev
->state
= TASK_RUNNING
;
3905 deactivate_task(rq
, prev
, 1);
3907 switch_count
= &prev
->nvcsw
;
3911 if (prev
->sched_class
->pre_schedule
)
3912 prev
->sched_class
->pre_schedule(rq
, prev
);
3915 if (unlikely(!rq
->nr_running
))
3916 idle_balance(cpu
, rq
);
3918 prev
->sched_class
->put_prev_task(rq
, prev
);
3919 next
= pick_next_task(rq
, prev
);
3921 sched_info_switch(prev
, next
);
3923 if (likely(prev
!= next
)) {
3928 context_switch(rq
, prev
, next
); /* unlocks the rq */
3930 * the context switch might have flipped the stack from under
3931 * us, hence refresh the local variables.
3933 cpu
= smp_processor_id();
3936 spin_unlock_irq(&rq
->lock
);
3940 if (unlikely(reacquire_kernel_lock(current
) < 0))
3941 goto need_resched_nonpreemptible
;
3943 preempt_enable_no_resched();
3944 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3947 EXPORT_SYMBOL(schedule
);
3949 #ifdef CONFIG_PREEMPT
3951 * this is the entry point to schedule() from in-kernel preemption
3952 * off of preempt_enable. Kernel preemptions off return from interrupt
3953 * occur there and call schedule directly.
3955 asmlinkage
void __sched
preempt_schedule(void)
3957 struct thread_info
*ti
= current_thread_info();
3958 struct task_struct
*task
= current
;
3959 int saved_lock_depth
;
3962 * If there is a non-zero preempt_count or interrupts are disabled,
3963 * we do not want to preempt the current task. Just return..
3965 if (likely(ti
->preempt_count
|| irqs_disabled()))
3969 add_preempt_count(PREEMPT_ACTIVE
);
3972 * We keep the big kernel semaphore locked, but we
3973 * clear ->lock_depth so that schedule() doesnt
3974 * auto-release the semaphore:
3976 saved_lock_depth
= task
->lock_depth
;
3977 task
->lock_depth
= -1;
3979 task
->lock_depth
= saved_lock_depth
;
3980 sub_preempt_count(PREEMPT_ACTIVE
);
3983 * Check again in case we missed a preemption opportunity
3984 * between schedule and now.
3987 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3989 EXPORT_SYMBOL(preempt_schedule
);
3992 * this is the entry point to schedule() from kernel preemption
3993 * off of irq context.
3994 * Note, that this is called and return with irqs disabled. This will
3995 * protect us against recursive calling from irq.
3997 asmlinkage
void __sched
preempt_schedule_irq(void)
3999 struct thread_info
*ti
= current_thread_info();
4000 struct task_struct
*task
= current
;
4001 int saved_lock_depth
;
4003 /* Catch callers which need to be fixed */
4004 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4007 add_preempt_count(PREEMPT_ACTIVE
);
4010 * We keep the big kernel semaphore locked, but we
4011 * clear ->lock_depth so that schedule() doesnt
4012 * auto-release the semaphore:
4014 saved_lock_depth
= task
->lock_depth
;
4015 task
->lock_depth
= -1;
4018 local_irq_disable();
4019 task
->lock_depth
= saved_lock_depth
;
4020 sub_preempt_count(PREEMPT_ACTIVE
);
4023 * Check again in case we missed a preemption opportunity
4024 * between schedule and now.
4027 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4030 #endif /* CONFIG_PREEMPT */
4032 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4035 return try_to_wake_up(curr
->private, mode
, sync
);
4037 EXPORT_SYMBOL(default_wake_function
);
4040 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4041 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4042 * number) then we wake all the non-exclusive tasks and one exclusive task.
4044 * There are circumstances in which we can try to wake a task which has already
4045 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4046 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4048 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4049 int nr_exclusive
, int sync
, void *key
)
4051 wait_queue_t
*curr
, *next
;
4053 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4054 unsigned flags
= curr
->flags
;
4056 if (curr
->func(curr
, mode
, sync
, key
) &&
4057 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4063 * __wake_up - wake up threads blocked on a waitqueue.
4065 * @mode: which threads
4066 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4067 * @key: is directly passed to the wakeup function
4069 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4070 int nr_exclusive
, void *key
)
4072 unsigned long flags
;
4074 spin_lock_irqsave(&q
->lock
, flags
);
4075 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4076 spin_unlock_irqrestore(&q
->lock
, flags
);
4078 EXPORT_SYMBOL(__wake_up
);
4081 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4083 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4085 __wake_up_common(q
, mode
, 1, 0, NULL
);
4089 * __wake_up_sync - wake up threads blocked on a waitqueue.
4091 * @mode: which threads
4092 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4094 * The sync wakeup differs that the waker knows that it will schedule
4095 * away soon, so while the target thread will be woken up, it will not
4096 * be migrated to another CPU - ie. the two threads are 'synchronized'
4097 * with each other. This can prevent needless bouncing between CPUs.
4099 * On UP it can prevent extra preemption.
4102 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4104 unsigned long flags
;
4110 if (unlikely(!nr_exclusive
))
4113 spin_lock_irqsave(&q
->lock
, flags
);
4114 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4115 spin_unlock_irqrestore(&q
->lock
, flags
);
4117 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4119 void complete(struct completion
*x
)
4121 unsigned long flags
;
4123 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4125 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
4127 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4129 EXPORT_SYMBOL(complete
);
4131 void complete_all(struct completion
*x
)
4133 unsigned long flags
;
4135 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4136 x
->done
+= UINT_MAX
/2;
4137 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
4139 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4141 EXPORT_SYMBOL(complete_all
);
4143 static inline long __sched
4144 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4147 DECLARE_WAITQUEUE(wait
, current
);
4149 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4150 __add_wait_queue_tail(&x
->wait
, &wait
);
4152 if (state
== TASK_INTERRUPTIBLE
&&
4153 signal_pending(current
)) {
4154 __remove_wait_queue(&x
->wait
, &wait
);
4155 return -ERESTARTSYS
;
4157 __set_current_state(state
);
4158 spin_unlock_irq(&x
->wait
.lock
);
4159 timeout
= schedule_timeout(timeout
);
4160 spin_lock_irq(&x
->wait
.lock
);
4162 __remove_wait_queue(&x
->wait
, &wait
);
4166 __remove_wait_queue(&x
->wait
, &wait
);
4173 wait_for_common(struct completion
*x
, long timeout
, int state
)
4177 spin_lock_irq(&x
->wait
.lock
);
4178 timeout
= do_wait_for_common(x
, timeout
, state
);
4179 spin_unlock_irq(&x
->wait
.lock
);
4183 void __sched
wait_for_completion(struct completion
*x
)
4185 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4187 EXPORT_SYMBOL(wait_for_completion
);
4189 unsigned long __sched
4190 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4192 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4194 EXPORT_SYMBOL(wait_for_completion_timeout
);
4196 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4198 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4199 if (t
== -ERESTARTSYS
)
4203 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4205 unsigned long __sched
4206 wait_for_completion_interruptible_timeout(struct completion
*x
,
4207 unsigned long timeout
)
4209 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4211 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4214 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4216 unsigned long flags
;
4219 init_waitqueue_entry(&wait
, current
);
4221 __set_current_state(state
);
4223 spin_lock_irqsave(&q
->lock
, flags
);
4224 __add_wait_queue(q
, &wait
);
4225 spin_unlock(&q
->lock
);
4226 timeout
= schedule_timeout(timeout
);
4227 spin_lock_irq(&q
->lock
);
4228 __remove_wait_queue(q
, &wait
);
4229 spin_unlock_irqrestore(&q
->lock
, flags
);
4234 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4236 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4238 EXPORT_SYMBOL(interruptible_sleep_on
);
4241 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4243 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4245 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4247 void __sched
sleep_on(wait_queue_head_t
*q
)
4249 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4251 EXPORT_SYMBOL(sleep_on
);
4253 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4255 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4257 EXPORT_SYMBOL(sleep_on_timeout
);
4259 #ifdef CONFIG_RT_MUTEXES
4262 * rt_mutex_setprio - set the current priority of a task
4264 * @prio: prio value (kernel-internal form)
4266 * This function changes the 'effective' priority of a task. It does
4267 * not touch ->normal_prio like __setscheduler().
4269 * Used by the rt_mutex code to implement priority inheritance logic.
4271 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4273 unsigned long flags
;
4274 int oldprio
, on_rq
, running
;
4276 const struct sched_class
*prev_class
= p
->sched_class
;
4278 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4280 rq
= task_rq_lock(p
, &flags
);
4281 update_rq_clock(rq
);
4284 on_rq
= p
->se
.on_rq
;
4285 running
= task_current(rq
, p
);
4287 dequeue_task(rq
, p
, 0);
4289 p
->sched_class
->put_prev_task(rq
, p
);
4293 p
->sched_class
= &rt_sched_class
;
4295 p
->sched_class
= &fair_sched_class
;
4301 p
->sched_class
->set_curr_task(rq
);
4303 enqueue_task(rq
, p
, 0);
4305 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4307 task_rq_unlock(rq
, &flags
);
4312 void set_user_nice(struct task_struct
*p
, long nice
)
4314 int old_prio
, delta
, on_rq
;
4315 unsigned long flags
;
4318 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4321 * We have to be careful, if called from sys_setpriority(),
4322 * the task might be in the middle of scheduling on another CPU.
4324 rq
= task_rq_lock(p
, &flags
);
4325 update_rq_clock(rq
);
4327 * The RT priorities are set via sched_setscheduler(), but we still
4328 * allow the 'normal' nice value to be set - but as expected
4329 * it wont have any effect on scheduling until the task is
4330 * SCHED_FIFO/SCHED_RR:
4332 if (task_has_rt_policy(p
)) {
4333 p
->static_prio
= NICE_TO_PRIO(nice
);
4336 on_rq
= p
->se
.on_rq
;
4338 dequeue_task(rq
, p
, 0);
4340 p
->static_prio
= NICE_TO_PRIO(nice
);
4343 p
->prio
= effective_prio(p
);
4344 delta
= p
->prio
- old_prio
;
4347 enqueue_task(rq
, p
, 0);
4349 * If the task increased its priority or is running and
4350 * lowered its priority, then reschedule its CPU:
4352 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4353 resched_task(rq
->curr
);
4356 task_rq_unlock(rq
, &flags
);
4358 EXPORT_SYMBOL(set_user_nice
);
4361 * can_nice - check if a task can reduce its nice value
4365 int can_nice(const struct task_struct
*p
, const int nice
)
4367 /* convert nice value [19,-20] to rlimit style value [1,40] */
4368 int nice_rlim
= 20 - nice
;
4370 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4371 capable(CAP_SYS_NICE
));
4374 #ifdef __ARCH_WANT_SYS_NICE
4377 * sys_nice - change the priority of the current process.
4378 * @increment: priority increment
4380 * sys_setpriority is a more generic, but much slower function that
4381 * does similar things.
4383 asmlinkage
long sys_nice(int increment
)
4388 * Setpriority might change our priority at the same moment.
4389 * We don't have to worry. Conceptually one call occurs first
4390 * and we have a single winner.
4392 if (increment
< -40)
4397 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4403 if (increment
< 0 && !can_nice(current
, nice
))
4406 retval
= security_task_setnice(current
, nice
);
4410 set_user_nice(current
, nice
);
4417 * task_prio - return the priority value of a given task.
4418 * @p: the task in question.
4420 * This is the priority value as seen by users in /proc.
4421 * RT tasks are offset by -200. Normal tasks are centered
4422 * around 0, value goes from -16 to +15.
4424 int task_prio(const struct task_struct
*p
)
4426 return p
->prio
- MAX_RT_PRIO
;
4430 * task_nice - return the nice value of a given task.
4431 * @p: the task in question.
4433 int task_nice(const struct task_struct
*p
)
4435 return TASK_NICE(p
);
4437 EXPORT_SYMBOL_GPL(task_nice
);
4440 * idle_cpu - is a given cpu idle currently?
4441 * @cpu: the processor in question.
4443 int idle_cpu(int cpu
)
4445 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4449 * idle_task - return the idle task for a given cpu.
4450 * @cpu: the processor in question.
4452 struct task_struct
*idle_task(int cpu
)
4454 return cpu_rq(cpu
)->idle
;
4458 * find_process_by_pid - find a process with a matching PID value.
4459 * @pid: the pid in question.
4461 static struct task_struct
*find_process_by_pid(pid_t pid
)
4463 return pid
? find_task_by_vpid(pid
) : current
;
4466 /* Actually do priority change: must hold rq lock. */
4468 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4470 BUG_ON(p
->se
.on_rq
);
4473 switch (p
->policy
) {
4477 p
->sched_class
= &fair_sched_class
;
4481 p
->sched_class
= &rt_sched_class
;
4485 p
->rt_priority
= prio
;
4486 p
->normal_prio
= normal_prio(p
);
4487 /* we are holding p->pi_lock already */
4488 p
->prio
= rt_mutex_getprio(p
);
4493 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4494 * @p: the task in question.
4495 * @policy: new policy.
4496 * @param: structure containing the new RT priority.
4498 * NOTE that the task may be already dead.
4500 int sched_setscheduler(struct task_struct
*p
, int policy
,
4501 struct sched_param
*param
)
4503 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4504 unsigned long flags
;
4505 const struct sched_class
*prev_class
= p
->sched_class
;
4508 /* may grab non-irq protected spin_locks */
4509 BUG_ON(in_interrupt());
4511 /* double check policy once rq lock held */
4513 policy
= oldpolicy
= p
->policy
;
4514 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4515 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4516 policy
!= SCHED_IDLE
)
4519 * Valid priorities for SCHED_FIFO and SCHED_RR are
4520 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4521 * SCHED_BATCH and SCHED_IDLE is 0.
4523 if (param
->sched_priority
< 0 ||
4524 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4525 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4527 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4531 * Allow unprivileged RT tasks to decrease priority:
4533 if (!capable(CAP_SYS_NICE
)) {
4534 if (rt_policy(policy
)) {
4535 unsigned long rlim_rtprio
;
4537 if (!lock_task_sighand(p
, &flags
))
4539 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4540 unlock_task_sighand(p
, &flags
);
4542 /* can't set/change the rt policy */
4543 if (policy
!= p
->policy
&& !rlim_rtprio
)
4546 /* can't increase priority */
4547 if (param
->sched_priority
> p
->rt_priority
&&
4548 param
->sched_priority
> rlim_rtprio
)
4552 * Like positive nice levels, dont allow tasks to
4553 * move out of SCHED_IDLE either:
4555 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4558 /* can't change other user's priorities */
4559 if ((current
->euid
!= p
->euid
) &&
4560 (current
->euid
!= p
->uid
))
4564 retval
= security_task_setscheduler(p
, policy
, param
);
4568 * make sure no PI-waiters arrive (or leave) while we are
4569 * changing the priority of the task:
4571 spin_lock_irqsave(&p
->pi_lock
, flags
);
4573 * To be able to change p->policy safely, the apropriate
4574 * runqueue lock must be held.
4576 rq
= __task_rq_lock(p
);
4577 /* recheck policy now with rq lock held */
4578 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4579 policy
= oldpolicy
= -1;
4580 __task_rq_unlock(rq
);
4581 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4584 update_rq_clock(rq
);
4585 on_rq
= p
->se
.on_rq
;
4586 running
= task_current(rq
, p
);
4588 deactivate_task(rq
, p
, 0);
4590 p
->sched_class
->put_prev_task(rq
, p
);
4594 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4598 p
->sched_class
->set_curr_task(rq
);
4600 activate_task(rq
, p
, 0);
4602 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4604 __task_rq_unlock(rq
);
4605 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4607 rt_mutex_adjust_pi(p
);
4611 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4614 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4616 struct sched_param lparam
;
4617 struct task_struct
*p
;
4620 if (!param
|| pid
< 0)
4622 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4627 p
= find_process_by_pid(pid
);
4629 retval
= sched_setscheduler(p
, policy
, &lparam
);
4636 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4637 * @pid: the pid in question.
4638 * @policy: new policy.
4639 * @param: structure containing the new RT priority.
4642 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4644 /* negative values for policy are not valid */
4648 return do_sched_setscheduler(pid
, policy
, param
);
4652 * sys_sched_setparam - set/change the RT priority of a thread
4653 * @pid: the pid in question.
4654 * @param: structure containing the new RT priority.
4656 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4658 return do_sched_setscheduler(pid
, -1, param
);
4662 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4663 * @pid: the pid in question.
4665 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4667 struct task_struct
*p
;
4674 read_lock(&tasklist_lock
);
4675 p
= find_process_by_pid(pid
);
4677 retval
= security_task_getscheduler(p
);
4681 read_unlock(&tasklist_lock
);
4686 * sys_sched_getscheduler - get the RT priority of a thread
4687 * @pid: the pid in question.
4688 * @param: structure containing the RT priority.
4690 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4692 struct sched_param lp
;
4693 struct task_struct
*p
;
4696 if (!param
|| pid
< 0)
4699 read_lock(&tasklist_lock
);
4700 p
= find_process_by_pid(pid
);
4705 retval
= security_task_getscheduler(p
);
4709 lp
.sched_priority
= p
->rt_priority
;
4710 read_unlock(&tasklist_lock
);
4713 * This one might sleep, we cannot do it with a spinlock held ...
4715 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4720 read_unlock(&tasklist_lock
);
4724 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4726 cpumask_t cpus_allowed
;
4727 struct task_struct
*p
;
4731 read_lock(&tasklist_lock
);
4733 p
= find_process_by_pid(pid
);
4735 read_unlock(&tasklist_lock
);
4741 * It is not safe to call set_cpus_allowed with the
4742 * tasklist_lock held. We will bump the task_struct's
4743 * usage count and then drop tasklist_lock.
4746 read_unlock(&tasklist_lock
);
4749 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4750 !capable(CAP_SYS_NICE
))
4753 retval
= security_task_setscheduler(p
, 0, NULL
);
4757 cpus_allowed
= cpuset_cpus_allowed(p
);
4758 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4760 retval
= set_cpus_allowed(p
, new_mask
);
4763 cpus_allowed
= cpuset_cpus_allowed(p
);
4764 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4766 * We must have raced with a concurrent cpuset
4767 * update. Just reset the cpus_allowed to the
4768 * cpuset's cpus_allowed
4770 new_mask
= cpus_allowed
;
4780 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4781 cpumask_t
*new_mask
)
4783 if (len
< sizeof(cpumask_t
)) {
4784 memset(new_mask
, 0, sizeof(cpumask_t
));
4785 } else if (len
> sizeof(cpumask_t
)) {
4786 len
= sizeof(cpumask_t
);
4788 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4792 * sys_sched_setaffinity - set the cpu affinity of a process
4793 * @pid: pid of the process
4794 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4795 * @user_mask_ptr: user-space pointer to the new cpu mask
4797 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4798 unsigned long __user
*user_mask_ptr
)
4803 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4807 return sched_setaffinity(pid
, new_mask
);
4811 * Represents all cpu's present in the system
4812 * In systems capable of hotplug, this map could dynamically grow
4813 * as new cpu's are detected in the system via any platform specific
4814 * method, such as ACPI for e.g.
4817 cpumask_t cpu_present_map __read_mostly
;
4818 EXPORT_SYMBOL(cpu_present_map
);
4821 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4822 EXPORT_SYMBOL(cpu_online_map
);
4824 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4825 EXPORT_SYMBOL(cpu_possible_map
);
4828 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4830 struct task_struct
*p
;
4834 read_lock(&tasklist_lock
);
4837 p
= find_process_by_pid(pid
);
4841 retval
= security_task_getscheduler(p
);
4845 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4848 read_unlock(&tasklist_lock
);
4855 * sys_sched_getaffinity - get the cpu affinity of a process
4856 * @pid: pid of the process
4857 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4858 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4860 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4861 unsigned long __user
*user_mask_ptr
)
4866 if (len
< sizeof(cpumask_t
))
4869 ret
= sched_getaffinity(pid
, &mask
);
4873 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4876 return sizeof(cpumask_t
);
4880 * sys_sched_yield - yield the current processor to other threads.
4882 * This function yields the current CPU to other tasks. If there are no
4883 * other threads running on this CPU then this function will return.
4885 asmlinkage
long sys_sched_yield(void)
4887 struct rq
*rq
= this_rq_lock();
4889 schedstat_inc(rq
, yld_count
);
4890 current
->sched_class
->yield_task(rq
);
4893 * Since we are going to call schedule() anyway, there's
4894 * no need to preempt or enable interrupts:
4896 __release(rq
->lock
);
4897 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4898 _raw_spin_unlock(&rq
->lock
);
4899 preempt_enable_no_resched();
4906 static void __cond_resched(void)
4908 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4909 __might_sleep(__FILE__
, __LINE__
);
4912 * The BKS might be reacquired before we have dropped
4913 * PREEMPT_ACTIVE, which could trigger a second
4914 * cond_resched() call.
4917 add_preempt_count(PREEMPT_ACTIVE
);
4919 sub_preempt_count(PREEMPT_ACTIVE
);
4920 } while (need_resched());
4923 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4924 int __sched
_cond_resched(void)
4926 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4927 system_state
== SYSTEM_RUNNING
) {
4933 EXPORT_SYMBOL(_cond_resched
);
4937 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4938 * call schedule, and on return reacquire the lock.
4940 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4941 * operations here to prevent schedule() from being called twice (once via
4942 * spin_unlock(), once by hand).
4944 int cond_resched_lock(spinlock_t
*lock
)
4948 if (need_lockbreak(lock
)) {
4954 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4955 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4956 _raw_spin_unlock(lock
);
4957 preempt_enable_no_resched();
4964 EXPORT_SYMBOL(cond_resched_lock
);
4966 int __sched
cond_resched_softirq(void)
4968 BUG_ON(!in_softirq());
4970 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4978 EXPORT_SYMBOL(cond_resched_softirq
);
4981 * yield - yield the current processor to other threads.
4983 * This is a shortcut for kernel-space yielding - it marks the
4984 * thread runnable and calls sys_sched_yield().
4986 void __sched
yield(void)
4988 set_current_state(TASK_RUNNING
);
4991 EXPORT_SYMBOL(yield
);
4994 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4995 * that process accounting knows that this is a task in IO wait state.
4997 * But don't do that if it is a deliberate, throttling IO wait (this task
4998 * has set its backing_dev_info: the queue against which it should throttle)
5000 void __sched
io_schedule(void)
5002 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5004 delayacct_blkio_start();
5005 atomic_inc(&rq
->nr_iowait
);
5007 atomic_dec(&rq
->nr_iowait
);
5008 delayacct_blkio_end();
5010 EXPORT_SYMBOL(io_schedule
);
5012 long __sched
io_schedule_timeout(long timeout
)
5014 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5017 delayacct_blkio_start();
5018 atomic_inc(&rq
->nr_iowait
);
5019 ret
= schedule_timeout(timeout
);
5020 atomic_dec(&rq
->nr_iowait
);
5021 delayacct_blkio_end();
5026 * sys_sched_get_priority_max - return maximum RT priority.
5027 * @policy: scheduling class.
5029 * this syscall returns the maximum rt_priority that can be used
5030 * by a given scheduling class.
5032 asmlinkage
long sys_sched_get_priority_max(int policy
)
5039 ret
= MAX_USER_RT_PRIO
-1;
5051 * sys_sched_get_priority_min - return minimum RT priority.
5052 * @policy: scheduling class.
5054 * this syscall returns the minimum rt_priority that can be used
5055 * by a given scheduling class.
5057 asmlinkage
long sys_sched_get_priority_min(int policy
)
5075 * sys_sched_rr_get_interval - return the default timeslice of a process.
5076 * @pid: pid of the process.
5077 * @interval: userspace pointer to the timeslice value.
5079 * this syscall writes the default timeslice value of a given process
5080 * into the user-space timespec buffer. A value of '0' means infinity.
5083 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5085 struct task_struct
*p
;
5086 unsigned int time_slice
;
5094 read_lock(&tasklist_lock
);
5095 p
= find_process_by_pid(pid
);
5099 retval
= security_task_getscheduler(p
);
5104 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5105 * tasks that are on an otherwise idle runqueue:
5108 if (p
->policy
== SCHED_RR
) {
5109 time_slice
= DEF_TIMESLICE
;
5111 struct sched_entity
*se
= &p
->se
;
5112 unsigned long flags
;
5115 rq
= task_rq_lock(p
, &flags
);
5116 if (rq
->cfs
.load
.weight
)
5117 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5118 task_rq_unlock(rq
, &flags
);
5120 read_unlock(&tasklist_lock
);
5121 jiffies_to_timespec(time_slice
, &t
);
5122 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5126 read_unlock(&tasklist_lock
);
5130 static const char stat_nam
[] = "RSDTtZX";
5132 void sched_show_task(struct task_struct
*p
)
5134 unsigned long free
= 0;
5137 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5138 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5139 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5140 #if BITS_PER_LONG == 32
5141 if (state
== TASK_RUNNING
)
5142 printk(KERN_CONT
" running ");
5144 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5146 if (state
== TASK_RUNNING
)
5147 printk(KERN_CONT
" running task ");
5149 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5151 #ifdef CONFIG_DEBUG_STACK_USAGE
5153 unsigned long *n
= end_of_stack(p
);
5156 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5159 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5160 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5162 if (state
!= TASK_RUNNING
)
5163 show_stack(p
, NULL
);
5166 void show_state_filter(unsigned long state_filter
)
5168 struct task_struct
*g
, *p
;
5170 #if BITS_PER_LONG == 32
5172 " task PC stack pid father\n");
5175 " task PC stack pid father\n");
5177 read_lock(&tasklist_lock
);
5178 do_each_thread(g
, p
) {
5180 * reset the NMI-timeout, listing all files on a slow
5181 * console might take alot of time:
5183 touch_nmi_watchdog();
5184 if (!state_filter
|| (p
->state
& state_filter
))
5186 } while_each_thread(g
, p
);
5188 touch_all_softlockup_watchdogs();
5190 #ifdef CONFIG_SCHED_DEBUG
5191 sysrq_sched_debug_show();
5193 read_unlock(&tasklist_lock
);
5195 * Only show locks if all tasks are dumped:
5197 if (state_filter
== -1)
5198 debug_show_all_locks();
5201 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5203 idle
->sched_class
= &idle_sched_class
;
5207 * init_idle - set up an idle thread for a given CPU
5208 * @idle: task in question
5209 * @cpu: cpu the idle task belongs to
5211 * NOTE: this function does not set the idle thread's NEED_RESCHED
5212 * flag, to make booting more robust.
5214 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5216 struct rq
*rq
= cpu_rq(cpu
);
5217 unsigned long flags
;
5220 idle
->se
.exec_start
= sched_clock();
5222 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5223 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5224 __set_task_cpu(idle
, cpu
);
5226 spin_lock_irqsave(&rq
->lock
, flags
);
5227 rq
->curr
= rq
->idle
= idle
;
5228 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5231 spin_unlock_irqrestore(&rq
->lock
, flags
);
5233 /* Set the preempt count _outside_ the spinlocks! */
5234 task_thread_info(idle
)->preempt_count
= 0;
5237 * The idle tasks have their own, simple scheduling class:
5239 idle
->sched_class
= &idle_sched_class
;
5243 * In a system that switches off the HZ timer nohz_cpu_mask
5244 * indicates which cpus entered this state. This is used
5245 * in the rcu update to wait only for active cpus. For system
5246 * which do not switch off the HZ timer nohz_cpu_mask should
5247 * always be CPU_MASK_NONE.
5249 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5252 * Increase the granularity value when there are more CPUs,
5253 * because with more CPUs the 'effective latency' as visible
5254 * to users decreases. But the relationship is not linear,
5255 * so pick a second-best guess by going with the log2 of the
5258 * This idea comes from the SD scheduler of Con Kolivas:
5260 static inline void sched_init_granularity(void)
5262 unsigned int factor
= 1 + ilog2(num_online_cpus());
5263 const unsigned long limit
= 200000000;
5265 sysctl_sched_min_granularity
*= factor
;
5266 if (sysctl_sched_min_granularity
> limit
)
5267 sysctl_sched_min_granularity
= limit
;
5269 sysctl_sched_latency
*= factor
;
5270 if (sysctl_sched_latency
> limit
)
5271 sysctl_sched_latency
= limit
;
5273 sysctl_sched_wakeup_granularity
*= factor
;
5274 sysctl_sched_batch_wakeup_granularity
*= factor
;
5279 * This is how migration works:
5281 * 1) we queue a struct migration_req structure in the source CPU's
5282 * runqueue and wake up that CPU's migration thread.
5283 * 2) we down() the locked semaphore => thread blocks.
5284 * 3) migration thread wakes up (implicitly it forces the migrated
5285 * thread off the CPU)
5286 * 4) it gets the migration request and checks whether the migrated
5287 * task is still in the wrong runqueue.
5288 * 5) if it's in the wrong runqueue then the migration thread removes
5289 * it and puts it into the right queue.
5290 * 6) migration thread up()s the semaphore.
5291 * 7) we wake up and the migration is done.
5295 * Change a given task's CPU affinity. Migrate the thread to a
5296 * proper CPU and schedule it away if the CPU it's executing on
5297 * is removed from the allowed bitmask.
5299 * NOTE: the caller must have a valid reference to the task, the
5300 * task must not exit() & deallocate itself prematurely. The
5301 * call is not atomic; no spinlocks may be held.
5303 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5305 struct migration_req req
;
5306 unsigned long flags
;
5310 rq
= task_rq_lock(p
, &flags
);
5311 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5316 if (p
->sched_class
->set_cpus_allowed
)
5317 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5319 p
->cpus_allowed
= new_mask
;
5320 p
->rt
.nr_cpus_allowed
= cpus_weight(new_mask
);
5323 /* Can the task run on the task's current CPU? If so, we're done */
5324 if (cpu_isset(task_cpu(p
), new_mask
))
5327 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5328 /* Need help from migration thread: drop lock and wait. */
5329 task_rq_unlock(rq
, &flags
);
5330 wake_up_process(rq
->migration_thread
);
5331 wait_for_completion(&req
.done
);
5332 tlb_migrate_finish(p
->mm
);
5336 task_rq_unlock(rq
, &flags
);
5340 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5343 * Move (not current) task off this cpu, onto dest cpu. We're doing
5344 * this because either it can't run here any more (set_cpus_allowed()
5345 * away from this CPU, or CPU going down), or because we're
5346 * attempting to rebalance this task on exec (sched_exec).
5348 * So we race with normal scheduler movements, but that's OK, as long
5349 * as the task is no longer on this CPU.
5351 * Returns non-zero if task was successfully migrated.
5353 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5355 struct rq
*rq_dest
, *rq_src
;
5358 if (unlikely(cpu_is_offline(dest_cpu
)))
5361 rq_src
= cpu_rq(src_cpu
);
5362 rq_dest
= cpu_rq(dest_cpu
);
5364 double_rq_lock(rq_src
, rq_dest
);
5365 /* Already moved. */
5366 if (task_cpu(p
) != src_cpu
)
5368 /* Affinity changed (again). */
5369 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5372 on_rq
= p
->se
.on_rq
;
5374 deactivate_task(rq_src
, p
, 0);
5376 set_task_cpu(p
, dest_cpu
);
5378 activate_task(rq_dest
, p
, 0);
5379 check_preempt_curr(rq_dest
, p
);
5383 double_rq_unlock(rq_src
, rq_dest
);
5388 * migration_thread - this is a highprio system thread that performs
5389 * thread migration by bumping thread off CPU then 'pushing' onto
5392 static int migration_thread(void *data
)
5394 int cpu
= (long)data
;
5398 BUG_ON(rq
->migration_thread
!= current
);
5400 set_current_state(TASK_INTERRUPTIBLE
);
5401 while (!kthread_should_stop()) {
5402 struct migration_req
*req
;
5403 struct list_head
*head
;
5405 spin_lock_irq(&rq
->lock
);
5407 if (cpu_is_offline(cpu
)) {
5408 spin_unlock_irq(&rq
->lock
);
5412 if (rq
->active_balance
) {
5413 active_load_balance(rq
, cpu
);
5414 rq
->active_balance
= 0;
5417 head
= &rq
->migration_queue
;
5419 if (list_empty(head
)) {
5420 spin_unlock_irq(&rq
->lock
);
5422 set_current_state(TASK_INTERRUPTIBLE
);
5425 req
= list_entry(head
->next
, struct migration_req
, list
);
5426 list_del_init(head
->next
);
5428 spin_unlock(&rq
->lock
);
5429 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5432 complete(&req
->done
);
5434 __set_current_state(TASK_RUNNING
);
5438 /* Wait for kthread_stop */
5439 set_current_state(TASK_INTERRUPTIBLE
);
5440 while (!kthread_should_stop()) {
5442 set_current_state(TASK_INTERRUPTIBLE
);
5444 __set_current_state(TASK_RUNNING
);
5448 #ifdef CONFIG_HOTPLUG_CPU
5450 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5454 local_irq_disable();
5455 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5461 * Figure out where task on dead CPU should go, use force if necessary.
5462 * NOTE: interrupts should be disabled by the caller
5464 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5466 unsigned long flags
;
5473 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5474 cpus_and(mask
, mask
, p
->cpus_allowed
);
5475 dest_cpu
= any_online_cpu(mask
);
5477 /* On any allowed CPU? */
5478 if (dest_cpu
== NR_CPUS
)
5479 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5481 /* No more Mr. Nice Guy. */
5482 if (dest_cpu
== NR_CPUS
) {
5483 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5485 * Try to stay on the same cpuset, where the
5486 * current cpuset may be a subset of all cpus.
5487 * The cpuset_cpus_allowed_locked() variant of
5488 * cpuset_cpus_allowed() will not block. It must be
5489 * called within calls to cpuset_lock/cpuset_unlock.
5491 rq
= task_rq_lock(p
, &flags
);
5492 p
->cpus_allowed
= cpus_allowed
;
5493 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5494 task_rq_unlock(rq
, &flags
);
5497 * Don't tell them about moving exiting tasks or
5498 * kernel threads (both mm NULL), since they never
5501 if (p
->mm
&& printk_ratelimit()) {
5502 printk(KERN_INFO
"process %d (%s) no "
5503 "longer affine to cpu%d\n",
5504 task_pid_nr(p
), p
->comm
, dead_cpu
);
5507 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5511 * While a dead CPU has no uninterruptible tasks queued at this point,
5512 * it might still have a nonzero ->nr_uninterruptible counter, because
5513 * for performance reasons the counter is not stricly tracking tasks to
5514 * their home CPUs. So we just add the counter to another CPU's counter,
5515 * to keep the global sum constant after CPU-down:
5517 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5519 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5520 unsigned long flags
;
5522 local_irq_save(flags
);
5523 double_rq_lock(rq_src
, rq_dest
);
5524 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5525 rq_src
->nr_uninterruptible
= 0;
5526 double_rq_unlock(rq_src
, rq_dest
);
5527 local_irq_restore(flags
);
5530 /* Run through task list and migrate tasks from the dead cpu. */
5531 static void migrate_live_tasks(int src_cpu
)
5533 struct task_struct
*p
, *t
;
5535 read_lock(&tasklist_lock
);
5537 do_each_thread(t
, p
) {
5541 if (task_cpu(p
) == src_cpu
)
5542 move_task_off_dead_cpu(src_cpu
, p
);
5543 } while_each_thread(t
, p
);
5545 read_unlock(&tasklist_lock
);
5549 * Schedules idle task to be the next runnable task on current CPU.
5550 * It does so by boosting its priority to highest possible.
5551 * Used by CPU offline code.
5553 void sched_idle_next(void)
5555 int this_cpu
= smp_processor_id();
5556 struct rq
*rq
= cpu_rq(this_cpu
);
5557 struct task_struct
*p
= rq
->idle
;
5558 unsigned long flags
;
5560 /* cpu has to be offline */
5561 BUG_ON(cpu_online(this_cpu
));
5564 * Strictly not necessary since rest of the CPUs are stopped by now
5565 * and interrupts disabled on the current cpu.
5567 spin_lock_irqsave(&rq
->lock
, flags
);
5569 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5571 update_rq_clock(rq
);
5572 activate_task(rq
, p
, 0);
5574 spin_unlock_irqrestore(&rq
->lock
, flags
);
5578 * Ensures that the idle task is using init_mm right before its cpu goes
5581 void idle_task_exit(void)
5583 struct mm_struct
*mm
= current
->active_mm
;
5585 BUG_ON(cpu_online(smp_processor_id()));
5588 switch_mm(mm
, &init_mm
, current
);
5592 /* called under rq->lock with disabled interrupts */
5593 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5595 struct rq
*rq
= cpu_rq(dead_cpu
);
5597 /* Must be exiting, otherwise would be on tasklist. */
5598 BUG_ON(!p
->exit_state
);
5600 /* Cannot have done final schedule yet: would have vanished. */
5601 BUG_ON(p
->state
== TASK_DEAD
);
5606 * Drop lock around migration; if someone else moves it,
5607 * that's OK. No task can be added to this CPU, so iteration is
5610 spin_unlock_irq(&rq
->lock
);
5611 move_task_off_dead_cpu(dead_cpu
, p
);
5612 spin_lock_irq(&rq
->lock
);
5617 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5618 static void migrate_dead_tasks(unsigned int dead_cpu
)
5620 struct rq
*rq
= cpu_rq(dead_cpu
);
5621 struct task_struct
*next
;
5624 if (!rq
->nr_running
)
5626 update_rq_clock(rq
);
5627 next
= pick_next_task(rq
, rq
->curr
);
5630 migrate_dead(dead_cpu
, next
);
5634 #endif /* CONFIG_HOTPLUG_CPU */
5636 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5638 static struct ctl_table sd_ctl_dir
[] = {
5640 .procname
= "sched_domain",
5646 static struct ctl_table sd_ctl_root
[] = {
5648 .ctl_name
= CTL_KERN
,
5649 .procname
= "kernel",
5651 .child
= sd_ctl_dir
,
5656 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5658 struct ctl_table
*entry
=
5659 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5664 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5666 struct ctl_table
*entry
;
5669 * In the intermediate directories, both the child directory and
5670 * procname are dynamically allocated and could fail but the mode
5671 * will always be set. In the lowest directory the names are
5672 * static strings and all have proc handlers.
5674 for (entry
= *tablep
; entry
->mode
; entry
++) {
5676 sd_free_ctl_entry(&entry
->child
);
5677 if (entry
->proc_handler
== NULL
)
5678 kfree(entry
->procname
);
5686 set_table_entry(struct ctl_table
*entry
,
5687 const char *procname
, void *data
, int maxlen
,
5688 mode_t mode
, proc_handler
*proc_handler
)
5690 entry
->procname
= procname
;
5692 entry
->maxlen
= maxlen
;
5694 entry
->proc_handler
= proc_handler
;
5697 static struct ctl_table
*
5698 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5700 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5705 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5706 sizeof(long), 0644, proc_doulongvec_minmax
);
5707 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5708 sizeof(long), 0644, proc_doulongvec_minmax
);
5709 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5710 sizeof(int), 0644, proc_dointvec_minmax
);
5711 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5712 sizeof(int), 0644, proc_dointvec_minmax
);
5713 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5714 sizeof(int), 0644, proc_dointvec_minmax
);
5715 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5716 sizeof(int), 0644, proc_dointvec_minmax
);
5717 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5718 sizeof(int), 0644, proc_dointvec_minmax
);
5719 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5720 sizeof(int), 0644, proc_dointvec_minmax
);
5721 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5722 sizeof(int), 0644, proc_dointvec_minmax
);
5723 set_table_entry(&table
[9], "cache_nice_tries",
5724 &sd
->cache_nice_tries
,
5725 sizeof(int), 0644, proc_dointvec_minmax
);
5726 set_table_entry(&table
[10], "flags", &sd
->flags
,
5727 sizeof(int), 0644, proc_dointvec_minmax
);
5728 /* &table[11] is terminator */
5733 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5735 struct ctl_table
*entry
, *table
;
5736 struct sched_domain
*sd
;
5737 int domain_num
= 0, i
;
5740 for_each_domain(cpu
, sd
)
5742 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5747 for_each_domain(cpu
, sd
) {
5748 snprintf(buf
, 32, "domain%d", i
);
5749 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5751 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5758 static struct ctl_table_header
*sd_sysctl_header
;
5759 static void register_sched_domain_sysctl(void)
5761 int i
, cpu_num
= num_online_cpus();
5762 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5765 WARN_ON(sd_ctl_dir
[0].child
);
5766 sd_ctl_dir
[0].child
= entry
;
5771 for_each_online_cpu(i
) {
5772 snprintf(buf
, 32, "cpu%d", i
);
5773 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5775 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5779 WARN_ON(sd_sysctl_header
);
5780 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5783 /* may be called multiple times per register */
5784 static void unregister_sched_domain_sysctl(void)
5786 if (sd_sysctl_header
)
5787 unregister_sysctl_table(sd_sysctl_header
);
5788 sd_sysctl_header
= NULL
;
5789 if (sd_ctl_dir
[0].child
)
5790 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5793 static void register_sched_domain_sysctl(void)
5796 static void unregister_sched_domain_sysctl(void)
5802 * migration_call - callback that gets triggered when a CPU is added.
5803 * Here we can start up the necessary migration thread for the new CPU.
5805 static int __cpuinit
5806 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5808 struct task_struct
*p
;
5809 int cpu
= (long)hcpu
;
5810 unsigned long flags
;
5815 case CPU_UP_PREPARE
:
5816 case CPU_UP_PREPARE_FROZEN
:
5817 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5820 kthread_bind(p
, cpu
);
5821 /* Must be high prio: stop_machine expects to yield to it. */
5822 rq
= task_rq_lock(p
, &flags
);
5823 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5824 task_rq_unlock(rq
, &flags
);
5825 cpu_rq(cpu
)->migration_thread
= p
;
5829 case CPU_ONLINE_FROZEN
:
5830 /* Strictly unnecessary, as first user will wake it. */
5831 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5833 /* Update our root-domain */
5835 spin_lock_irqsave(&rq
->lock
, flags
);
5837 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5838 cpu_set(cpu
, rq
->rd
->online
);
5840 spin_unlock_irqrestore(&rq
->lock
, flags
);
5843 #ifdef CONFIG_HOTPLUG_CPU
5844 case CPU_UP_CANCELED
:
5845 case CPU_UP_CANCELED_FROZEN
:
5846 if (!cpu_rq(cpu
)->migration_thread
)
5848 /* Unbind it from offline cpu so it can run. Fall thru. */
5849 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5850 any_online_cpu(cpu_online_map
));
5851 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5852 cpu_rq(cpu
)->migration_thread
= NULL
;
5856 case CPU_DEAD_FROZEN
:
5857 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5858 migrate_live_tasks(cpu
);
5860 kthread_stop(rq
->migration_thread
);
5861 rq
->migration_thread
= NULL
;
5862 /* Idle task back to normal (off runqueue, low prio) */
5863 spin_lock_irq(&rq
->lock
);
5864 update_rq_clock(rq
);
5865 deactivate_task(rq
, rq
->idle
, 0);
5866 rq
->idle
->static_prio
= MAX_PRIO
;
5867 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5868 rq
->idle
->sched_class
= &idle_sched_class
;
5869 migrate_dead_tasks(cpu
);
5870 spin_unlock_irq(&rq
->lock
);
5872 migrate_nr_uninterruptible(rq
);
5873 BUG_ON(rq
->nr_running
!= 0);
5876 * No need to migrate the tasks: it was best-effort if
5877 * they didn't take sched_hotcpu_mutex. Just wake up
5880 spin_lock_irq(&rq
->lock
);
5881 while (!list_empty(&rq
->migration_queue
)) {
5882 struct migration_req
*req
;
5884 req
= list_entry(rq
->migration_queue
.next
,
5885 struct migration_req
, list
);
5886 list_del_init(&req
->list
);
5887 complete(&req
->done
);
5889 spin_unlock_irq(&rq
->lock
);
5892 case CPU_DOWN_PREPARE
:
5893 /* Update our root-domain */
5895 spin_lock_irqsave(&rq
->lock
, flags
);
5897 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5898 cpu_clear(cpu
, rq
->rd
->online
);
5900 spin_unlock_irqrestore(&rq
->lock
, flags
);
5907 /* Register at highest priority so that task migration (migrate_all_tasks)
5908 * happens before everything else.
5910 static struct notifier_block __cpuinitdata migration_notifier
= {
5911 .notifier_call
= migration_call
,
5915 void __init
migration_init(void)
5917 void *cpu
= (void *)(long)smp_processor_id();
5920 /* Start one for the boot CPU: */
5921 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5922 BUG_ON(err
== NOTIFY_BAD
);
5923 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5924 register_cpu_notifier(&migration_notifier
);
5930 /* Number of possible processor ids */
5931 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5932 EXPORT_SYMBOL(nr_cpu_ids
);
5934 #ifdef CONFIG_SCHED_DEBUG
5936 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5938 struct sched_group
*group
= sd
->groups
;
5939 cpumask_t groupmask
;
5942 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5943 cpus_clear(groupmask
);
5945 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5947 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5948 printk("does not load-balance\n");
5950 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5955 printk(KERN_CONT
"span %s\n", str
);
5957 if (!cpu_isset(cpu
, sd
->span
)) {
5958 printk(KERN_ERR
"ERROR: domain->span does not contain "
5961 if (!cpu_isset(cpu
, group
->cpumask
)) {
5962 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5966 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5970 printk(KERN_ERR
"ERROR: group is NULL\n");
5974 if (!group
->__cpu_power
) {
5975 printk(KERN_CONT
"\n");
5976 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5981 if (!cpus_weight(group
->cpumask
)) {
5982 printk(KERN_CONT
"\n");
5983 printk(KERN_ERR
"ERROR: empty group\n");
5987 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5988 printk(KERN_CONT
"\n");
5989 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5993 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5995 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5996 printk(KERN_CONT
" %s", str
);
5998 group
= group
->next
;
5999 } while (group
!= sd
->groups
);
6000 printk(KERN_CONT
"\n");
6002 if (!cpus_equal(sd
->span
, groupmask
))
6003 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6005 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
6006 printk(KERN_ERR
"ERROR: parent span is not a superset "
6007 "of domain->span\n");
6011 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6016 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6020 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6023 if (sched_domain_debug_one(sd
, cpu
, level
))
6032 # define sched_domain_debug(sd, cpu) do { } while (0)
6035 static int sd_degenerate(struct sched_domain
*sd
)
6037 if (cpus_weight(sd
->span
) == 1)
6040 /* Following flags need at least 2 groups */
6041 if (sd
->flags
& (SD_LOAD_BALANCE
|
6042 SD_BALANCE_NEWIDLE
|
6046 SD_SHARE_PKG_RESOURCES
)) {
6047 if (sd
->groups
!= sd
->groups
->next
)
6051 /* Following flags don't use groups */
6052 if (sd
->flags
& (SD_WAKE_IDLE
|
6061 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6063 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6065 if (sd_degenerate(parent
))
6068 if (!cpus_equal(sd
->span
, parent
->span
))
6071 /* Does parent contain flags not in child? */
6072 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6073 if (cflags
& SD_WAKE_AFFINE
)
6074 pflags
&= ~SD_WAKE_BALANCE
;
6075 /* Flags needing groups don't count if only 1 group in parent */
6076 if (parent
->groups
== parent
->groups
->next
) {
6077 pflags
&= ~(SD_LOAD_BALANCE
|
6078 SD_BALANCE_NEWIDLE
|
6082 SD_SHARE_PKG_RESOURCES
);
6084 if (~cflags
& pflags
)
6090 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6092 unsigned long flags
;
6093 const struct sched_class
*class;
6095 spin_lock_irqsave(&rq
->lock
, flags
);
6098 struct root_domain
*old_rd
= rq
->rd
;
6100 for (class = sched_class_highest
; class; class = class->next
) {
6101 if (class->leave_domain
)
6102 class->leave_domain(rq
);
6105 cpu_clear(rq
->cpu
, old_rd
->span
);
6106 cpu_clear(rq
->cpu
, old_rd
->online
);
6108 if (atomic_dec_and_test(&old_rd
->refcount
))
6112 atomic_inc(&rd
->refcount
);
6115 cpu_set(rq
->cpu
, rd
->span
);
6116 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6117 cpu_set(rq
->cpu
, rd
->online
);
6119 for (class = sched_class_highest
; class; class = class->next
) {
6120 if (class->join_domain
)
6121 class->join_domain(rq
);
6124 spin_unlock_irqrestore(&rq
->lock
, flags
);
6127 static void init_rootdomain(struct root_domain
*rd
)
6129 memset(rd
, 0, sizeof(*rd
));
6131 cpus_clear(rd
->span
);
6132 cpus_clear(rd
->online
);
6135 static void init_defrootdomain(void)
6137 init_rootdomain(&def_root_domain
);
6138 atomic_set(&def_root_domain
.refcount
, 1);
6141 static struct root_domain
*alloc_rootdomain(void)
6143 struct root_domain
*rd
;
6145 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6149 init_rootdomain(rd
);
6155 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6156 * hold the hotplug lock.
6159 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6161 struct rq
*rq
= cpu_rq(cpu
);
6162 struct sched_domain
*tmp
;
6164 /* Remove the sched domains which do not contribute to scheduling. */
6165 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6166 struct sched_domain
*parent
= tmp
->parent
;
6169 if (sd_parent_degenerate(tmp
, parent
)) {
6170 tmp
->parent
= parent
->parent
;
6172 parent
->parent
->child
= tmp
;
6176 if (sd
&& sd_degenerate(sd
)) {
6182 sched_domain_debug(sd
, cpu
);
6184 rq_attach_root(rq
, rd
);
6185 rcu_assign_pointer(rq
->sd
, sd
);
6188 /* cpus with isolated domains */
6189 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6191 /* Setup the mask of cpus configured for isolated domains */
6192 static int __init
isolated_cpu_setup(char *str
)
6194 int ints
[NR_CPUS
], i
;
6196 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6197 cpus_clear(cpu_isolated_map
);
6198 for (i
= 1; i
<= ints
[0]; i
++)
6199 if (ints
[i
] < NR_CPUS
)
6200 cpu_set(ints
[i
], cpu_isolated_map
);
6204 __setup("isolcpus=", isolated_cpu_setup
);
6207 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6208 * to a function which identifies what group(along with sched group) a CPU
6209 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6210 * (due to the fact that we keep track of groups covered with a cpumask_t).
6212 * init_sched_build_groups will build a circular linked list of the groups
6213 * covered by the given span, and will set each group's ->cpumask correctly,
6214 * and ->cpu_power to 0.
6217 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
6218 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6219 struct sched_group
**sg
))
6221 struct sched_group
*first
= NULL
, *last
= NULL
;
6222 cpumask_t covered
= CPU_MASK_NONE
;
6225 for_each_cpu_mask(i
, span
) {
6226 struct sched_group
*sg
;
6227 int group
= group_fn(i
, cpu_map
, &sg
);
6230 if (cpu_isset(i
, covered
))
6233 sg
->cpumask
= CPU_MASK_NONE
;
6234 sg
->__cpu_power
= 0;
6236 for_each_cpu_mask(j
, span
) {
6237 if (group_fn(j
, cpu_map
, NULL
) != group
)
6240 cpu_set(j
, covered
);
6241 cpu_set(j
, sg
->cpumask
);
6252 #define SD_NODES_PER_DOMAIN 16
6257 * find_next_best_node - find the next node to include in a sched_domain
6258 * @node: node whose sched_domain we're building
6259 * @used_nodes: nodes already in the sched_domain
6261 * Find the next node to include in a given scheduling domain. Simply
6262 * finds the closest node not already in the @used_nodes map.
6264 * Should use nodemask_t.
6266 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6268 int i
, n
, val
, min_val
, best_node
= 0;
6272 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6273 /* Start at @node */
6274 n
= (node
+ i
) % MAX_NUMNODES
;
6276 if (!nr_cpus_node(n
))
6279 /* Skip already used nodes */
6280 if (test_bit(n
, used_nodes
))
6283 /* Simple min distance search */
6284 val
= node_distance(node
, n
);
6286 if (val
< min_val
) {
6292 set_bit(best_node
, used_nodes
);
6297 * sched_domain_node_span - get a cpumask for a node's sched_domain
6298 * @node: node whose cpumask we're constructing
6299 * @size: number of nodes to include in this span
6301 * Given a node, construct a good cpumask for its sched_domain to span. It
6302 * should be one that prevents unnecessary balancing, but also spreads tasks
6305 static cpumask_t
sched_domain_node_span(int node
)
6307 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6308 cpumask_t span
, nodemask
;
6312 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6314 nodemask
= node_to_cpumask(node
);
6315 cpus_or(span
, span
, nodemask
);
6316 set_bit(node
, used_nodes
);
6318 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6319 int next_node
= find_next_best_node(node
, used_nodes
);
6321 nodemask
= node_to_cpumask(next_node
);
6322 cpus_or(span
, span
, nodemask
);
6329 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6332 * SMT sched-domains:
6334 #ifdef CONFIG_SCHED_SMT
6335 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6336 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6339 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6342 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6348 * multi-core sched-domains:
6350 #ifdef CONFIG_SCHED_MC
6351 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6352 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6355 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6357 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6360 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6361 cpus_and(mask
, mask
, *cpu_map
);
6362 group
= first_cpu(mask
);
6364 *sg
= &per_cpu(sched_group_core
, group
);
6367 #elif defined(CONFIG_SCHED_MC)
6369 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6372 *sg
= &per_cpu(sched_group_core
, cpu
);
6377 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6378 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6381 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6384 #ifdef CONFIG_SCHED_MC
6385 cpumask_t mask
= cpu_coregroup_map(cpu
);
6386 cpus_and(mask
, mask
, *cpu_map
);
6387 group
= first_cpu(mask
);
6388 #elif defined(CONFIG_SCHED_SMT)
6389 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6390 cpus_and(mask
, mask
, *cpu_map
);
6391 group
= first_cpu(mask
);
6396 *sg
= &per_cpu(sched_group_phys
, group
);
6402 * The init_sched_build_groups can't handle what we want to do with node
6403 * groups, so roll our own. Now each node has its own list of groups which
6404 * gets dynamically allocated.
6406 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6407 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6409 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6410 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6412 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6413 struct sched_group
**sg
)
6415 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6418 cpus_and(nodemask
, nodemask
, *cpu_map
);
6419 group
= first_cpu(nodemask
);
6422 *sg
= &per_cpu(sched_group_allnodes
, group
);
6426 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6428 struct sched_group
*sg
= group_head
;
6434 for_each_cpu_mask(j
, sg
->cpumask
) {
6435 struct sched_domain
*sd
;
6437 sd
= &per_cpu(phys_domains
, j
);
6438 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6440 * Only add "power" once for each
6446 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6449 } while (sg
!= group_head
);
6454 /* Free memory allocated for various sched_group structures */
6455 static void free_sched_groups(const cpumask_t
*cpu_map
)
6459 for_each_cpu_mask(cpu
, *cpu_map
) {
6460 struct sched_group
**sched_group_nodes
6461 = sched_group_nodes_bycpu
[cpu
];
6463 if (!sched_group_nodes
)
6466 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6467 cpumask_t nodemask
= node_to_cpumask(i
);
6468 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6470 cpus_and(nodemask
, nodemask
, *cpu_map
);
6471 if (cpus_empty(nodemask
))
6481 if (oldsg
!= sched_group_nodes
[i
])
6484 kfree(sched_group_nodes
);
6485 sched_group_nodes_bycpu
[cpu
] = NULL
;
6489 static void free_sched_groups(const cpumask_t
*cpu_map
)
6495 * Initialize sched groups cpu_power.
6497 * cpu_power indicates the capacity of sched group, which is used while
6498 * distributing the load between different sched groups in a sched domain.
6499 * Typically cpu_power for all the groups in a sched domain will be same unless
6500 * there are asymmetries in the topology. If there are asymmetries, group
6501 * having more cpu_power will pickup more load compared to the group having
6504 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6505 * the maximum number of tasks a group can handle in the presence of other idle
6506 * or lightly loaded groups in the same sched domain.
6508 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6510 struct sched_domain
*child
;
6511 struct sched_group
*group
;
6513 WARN_ON(!sd
|| !sd
->groups
);
6515 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6520 sd
->groups
->__cpu_power
= 0;
6523 * For perf policy, if the groups in child domain share resources
6524 * (for example cores sharing some portions of the cache hierarchy
6525 * or SMT), then set this domain groups cpu_power such that each group
6526 * can handle only one task, when there are other idle groups in the
6527 * same sched domain.
6529 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6531 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6532 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6537 * add cpu_power of each child group to this groups cpu_power
6539 group
= child
->groups
;
6541 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6542 group
= group
->next
;
6543 } while (group
!= child
->groups
);
6547 * Build sched domains for a given set of cpus and attach the sched domains
6548 * to the individual cpus
6550 static int build_sched_domains(const cpumask_t
*cpu_map
)
6553 struct root_domain
*rd
;
6555 struct sched_group
**sched_group_nodes
= NULL
;
6556 int sd_allnodes
= 0;
6559 * Allocate the per-node list of sched groups
6561 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6563 if (!sched_group_nodes
) {
6564 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6567 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6570 rd
= alloc_rootdomain();
6572 printk(KERN_WARNING
"Cannot alloc root domain\n");
6577 * Set up domains for cpus specified by the cpu_map.
6579 for_each_cpu_mask(i
, *cpu_map
) {
6580 struct sched_domain
*sd
= NULL
, *p
;
6581 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6583 cpus_and(nodemask
, nodemask
, *cpu_map
);
6586 if (cpus_weight(*cpu_map
) >
6587 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6588 sd
= &per_cpu(allnodes_domains
, i
);
6589 *sd
= SD_ALLNODES_INIT
;
6590 sd
->span
= *cpu_map
;
6591 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6597 sd
= &per_cpu(node_domains
, i
);
6599 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6603 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6607 sd
= &per_cpu(phys_domains
, i
);
6609 sd
->span
= nodemask
;
6613 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6615 #ifdef CONFIG_SCHED_MC
6617 sd
= &per_cpu(core_domains
, i
);
6619 sd
->span
= cpu_coregroup_map(i
);
6620 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6623 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6626 #ifdef CONFIG_SCHED_SMT
6628 sd
= &per_cpu(cpu_domains
, i
);
6629 *sd
= SD_SIBLING_INIT
;
6630 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6631 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6634 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6638 #ifdef CONFIG_SCHED_SMT
6639 /* Set up CPU (sibling) groups */
6640 for_each_cpu_mask(i
, *cpu_map
) {
6641 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6642 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6643 if (i
!= first_cpu(this_sibling_map
))
6646 init_sched_build_groups(this_sibling_map
, cpu_map
,
6651 #ifdef CONFIG_SCHED_MC
6652 /* Set up multi-core groups */
6653 for_each_cpu_mask(i
, *cpu_map
) {
6654 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6655 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6656 if (i
!= first_cpu(this_core_map
))
6658 init_sched_build_groups(this_core_map
, cpu_map
,
6659 &cpu_to_core_group
);
6663 /* Set up physical groups */
6664 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6665 cpumask_t nodemask
= node_to_cpumask(i
);
6667 cpus_and(nodemask
, nodemask
, *cpu_map
);
6668 if (cpus_empty(nodemask
))
6671 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6675 /* Set up node groups */
6677 init_sched_build_groups(*cpu_map
, cpu_map
,
6678 &cpu_to_allnodes_group
);
6680 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6681 /* Set up node groups */
6682 struct sched_group
*sg
, *prev
;
6683 cpumask_t nodemask
= node_to_cpumask(i
);
6684 cpumask_t domainspan
;
6685 cpumask_t covered
= CPU_MASK_NONE
;
6688 cpus_and(nodemask
, nodemask
, *cpu_map
);
6689 if (cpus_empty(nodemask
)) {
6690 sched_group_nodes
[i
] = NULL
;
6694 domainspan
= sched_domain_node_span(i
);
6695 cpus_and(domainspan
, domainspan
, *cpu_map
);
6697 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6699 printk(KERN_WARNING
"Can not alloc domain group for "
6703 sched_group_nodes
[i
] = sg
;
6704 for_each_cpu_mask(j
, nodemask
) {
6705 struct sched_domain
*sd
;
6707 sd
= &per_cpu(node_domains
, j
);
6710 sg
->__cpu_power
= 0;
6711 sg
->cpumask
= nodemask
;
6713 cpus_or(covered
, covered
, nodemask
);
6716 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6717 cpumask_t tmp
, notcovered
;
6718 int n
= (i
+ j
) % MAX_NUMNODES
;
6720 cpus_complement(notcovered
, covered
);
6721 cpus_and(tmp
, notcovered
, *cpu_map
);
6722 cpus_and(tmp
, tmp
, domainspan
);
6723 if (cpus_empty(tmp
))
6726 nodemask
= node_to_cpumask(n
);
6727 cpus_and(tmp
, tmp
, nodemask
);
6728 if (cpus_empty(tmp
))
6731 sg
= kmalloc_node(sizeof(struct sched_group
),
6735 "Can not alloc domain group for node %d\n", j
);
6738 sg
->__cpu_power
= 0;
6740 sg
->next
= prev
->next
;
6741 cpus_or(covered
, covered
, tmp
);
6748 /* Calculate CPU power for physical packages and nodes */
6749 #ifdef CONFIG_SCHED_SMT
6750 for_each_cpu_mask(i
, *cpu_map
) {
6751 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6753 init_sched_groups_power(i
, sd
);
6756 #ifdef CONFIG_SCHED_MC
6757 for_each_cpu_mask(i
, *cpu_map
) {
6758 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6760 init_sched_groups_power(i
, sd
);
6764 for_each_cpu_mask(i
, *cpu_map
) {
6765 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6767 init_sched_groups_power(i
, sd
);
6771 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6772 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6775 struct sched_group
*sg
;
6777 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6778 init_numa_sched_groups_power(sg
);
6782 /* Attach the domains */
6783 for_each_cpu_mask(i
, *cpu_map
) {
6784 struct sched_domain
*sd
;
6785 #ifdef CONFIG_SCHED_SMT
6786 sd
= &per_cpu(cpu_domains
, i
);
6787 #elif defined(CONFIG_SCHED_MC)
6788 sd
= &per_cpu(core_domains
, i
);
6790 sd
= &per_cpu(phys_domains
, i
);
6792 cpu_attach_domain(sd
, rd
, i
);
6799 free_sched_groups(cpu_map
);
6804 static cpumask_t
*doms_cur
; /* current sched domains */
6805 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6808 * Special case: If a kmalloc of a doms_cur partition (array of
6809 * cpumask_t) fails, then fallback to a single sched domain,
6810 * as determined by the single cpumask_t fallback_doms.
6812 static cpumask_t fallback_doms
;
6815 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6816 * For now this just excludes isolated cpus, but could be used to
6817 * exclude other special cases in the future.
6819 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6824 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6826 doms_cur
= &fallback_doms
;
6827 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6828 err
= build_sched_domains(doms_cur
);
6829 register_sched_domain_sysctl();
6834 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6836 free_sched_groups(cpu_map
);
6840 * Detach sched domains from a group of cpus specified in cpu_map
6841 * These cpus will now be attached to the NULL domain
6843 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6847 unregister_sched_domain_sysctl();
6849 for_each_cpu_mask(i
, *cpu_map
)
6850 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6851 synchronize_sched();
6852 arch_destroy_sched_domains(cpu_map
);
6856 * Partition sched domains as specified by the 'ndoms_new'
6857 * cpumasks in the array doms_new[] of cpumasks. This compares
6858 * doms_new[] to the current sched domain partitioning, doms_cur[].
6859 * It destroys each deleted domain and builds each new domain.
6861 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6862 * The masks don't intersect (don't overlap.) We should setup one
6863 * sched domain for each mask. CPUs not in any of the cpumasks will
6864 * not be load balanced. If the same cpumask appears both in the
6865 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6868 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6869 * ownership of it and will kfree it when done with it. If the caller
6870 * failed the kmalloc call, then it can pass in doms_new == NULL,
6871 * and partition_sched_domains() will fallback to the single partition
6874 * Call with hotplug lock held
6876 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6882 /* always unregister in case we don't destroy any domains */
6883 unregister_sched_domain_sysctl();
6885 if (doms_new
== NULL
) {
6887 doms_new
= &fallback_doms
;
6888 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6891 /* Destroy deleted domains */
6892 for (i
= 0; i
< ndoms_cur
; i
++) {
6893 for (j
= 0; j
< ndoms_new
; j
++) {
6894 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6897 /* no match - a current sched domain not in new doms_new[] */
6898 detach_destroy_domains(doms_cur
+ i
);
6903 /* Build new domains */
6904 for (i
= 0; i
< ndoms_new
; i
++) {
6905 for (j
= 0; j
< ndoms_cur
; j
++) {
6906 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6909 /* no match - add a new doms_new */
6910 build_sched_domains(doms_new
+ i
);
6915 /* Remember the new sched domains */
6916 if (doms_cur
!= &fallback_doms
)
6918 doms_cur
= doms_new
;
6919 ndoms_cur
= ndoms_new
;
6921 register_sched_domain_sysctl();
6926 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6927 static int arch_reinit_sched_domains(void)
6932 detach_destroy_domains(&cpu_online_map
);
6933 err
= arch_init_sched_domains(&cpu_online_map
);
6939 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6943 if (buf
[0] != '0' && buf
[0] != '1')
6947 sched_smt_power_savings
= (buf
[0] == '1');
6949 sched_mc_power_savings
= (buf
[0] == '1');
6951 ret
= arch_reinit_sched_domains();
6953 return ret
? ret
: count
;
6956 #ifdef CONFIG_SCHED_MC
6957 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6959 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6961 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6962 const char *buf
, size_t count
)
6964 return sched_power_savings_store(buf
, count
, 0);
6966 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6967 sched_mc_power_savings_store
);
6970 #ifdef CONFIG_SCHED_SMT
6971 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6973 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6975 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6976 const char *buf
, size_t count
)
6978 return sched_power_savings_store(buf
, count
, 1);
6980 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6981 sched_smt_power_savings_store
);
6984 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6988 #ifdef CONFIG_SCHED_SMT
6990 err
= sysfs_create_file(&cls
->kset
.kobj
,
6991 &attr_sched_smt_power_savings
.attr
);
6993 #ifdef CONFIG_SCHED_MC
6994 if (!err
&& mc_capable())
6995 err
= sysfs_create_file(&cls
->kset
.kobj
,
6996 &attr_sched_mc_power_savings
.attr
);
7003 * Force a reinitialization of the sched domains hierarchy. The domains
7004 * and groups cannot be updated in place without racing with the balancing
7005 * code, so we temporarily attach all running cpus to the NULL domain
7006 * which will prevent rebalancing while the sched domains are recalculated.
7008 static int update_sched_domains(struct notifier_block
*nfb
,
7009 unsigned long action
, void *hcpu
)
7012 case CPU_UP_PREPARE
:
7013 case CPU_UP_PREPARE_FROZEN
:
7014 case CPU_DOWN_PREPARE
:
7015 case CPU_DOWN_PREPARE_FROZEN
:
7016 detach_destroy_domains(&cpu_online_map
);
7019 case CPU_UP_CANCELED
:
7020 case CPU_UP_CANCELED_FROZEN
:
7021 case CPU_DOWN_FAILED
:
7022 case CPU_DOWN_FAILED_FROZEN
:
7024 case CPU_ONLINE_FROZEN
:
7026 case CPU_DEAD_FROZEN
:
7028 * Fall through and re-initialise the domains.
7035 /* The hotplug lock is already held by cpu_up/cpu_down */
7036 arch_init_sched_domains(&cpu_online_map
);
7041 void __init
sched_init_smp(void)
7043 cpumask_t non_isolated_cpus
;
7046 arch_init_sched_domains(&cpu_online_map
);
7047 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7048 if (cpus_empty(non_isolated_cpus
))
7049 cpu_set(smp_processor_id(), non_isolated_cpus
);
7051 /* XXX: Theoretical race here - CPU may be hotplugged now */
7052 hotcpu_notifier(update_sched_domains
, 0);
7054 /* Move init over to a non-isolated CPU */
7055 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
7057 sched_init_granularity();
7059 #ifdef CONFIG_FAIR_GROUP_SCHED
7060 if (nr_cpu_ids
== 1)
7063 lb_monitor_task
= kthread_create(load_balance_monitor
, NULL
,
7065 if (!IS_ERR(lb_monitor_task
)) {
7066 lb_monitor_task
->flags
|= PF_NOFREEZE
;
7067 wake_up_process(lb_monitor_task
);
7069 printk(KERN_ERR
"Could not create load balance monitor thread"
7070 "(error = %ld) \n", PTR_ERR(lb_monitor_task
));
7075 void __init
sched_init_smp(void)
7077 sched_init_granularity();
7079 #endif /* CONFIG_SMP */
7081 int in_sched_functions(unsigned long addr
)
7083 return in_lock_functions(addr
) ||
7084 (addr
>= (unsigned long)__sched_text_start
7085 && addr
< (unsigned long)__sched_text_end
);
7088 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7090 cfs_rq
->tasks_timeline
= RB_ROOT
;
7091 #ifdef CONFIG_FAIR_GROUP_SCHED
7094 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7097 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7099 struct rt_prio_array
*array
;
7102 array
= &rt_rq
->active
;
7103 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7104 INIT_LIST_HEAD(array
->queue
+ i
);
7105 __clear_bit(i
, array
->bitmap
);
7107 /* delimiter for bitsearch: */
7108 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7110 #if defined CONFIG_SMP || defined CONFIG_FAIR_GROUP_SCHED
7111 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7114 rt_rq
->rt_nr_migratory
= 0;
7115 rt_rq
->overloaded
= 0;
7119 rt_rq
->rt_throttled
= 0;
7121 #ifdef CONFIG_FAIR_GROUP_SCHED
7126 #ifdef CONFIG_FAIR_GROUP_SCHED
7127 static void init_tg_cfs_entry(struct rq
*rq
, struct task_group
*tg
,
7128 struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
7131 tg
->cfs_rq
[cpu
] = cfs_rq
;
7132 init_cfs_rq(cfs_rq
, rq
);
7135 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7138 se
->cfs_rq
= &rq
->cfs
;
7140 se
->load
.weight
= tg
->shares
;
7141 se
->load
.inv_weight
= div64_64(1ULL<<32, se
->load
.weight
);
7145 static void init_tg_rt_entry(struct rq
*rq
, struct task_group
*tg
,
7146 struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
,
7149 tg
->rt_rq
[cpu
] = rt_rq
;
7150 init_rt_rq(rt_rq
, rq
);
7152 rt_rq
->rt_se
= rt_se
;
7154 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7156 tg
->rt_se
[cpu
] = rt_se
;
7157 rt_se
->rt_rq
= &rq
->rt
;
7158 rt_se
->my_q
= rt_rq
;
7159 rt_se
->parent
= NULL
;
7160 INIT_LIST_HEAD(&rt_se
->run_list
);
7164 void __init
sched_init(void)
7166 int highest_cpu
= 0;
7170 init_defrootdomain();
7173 #ifdef CONFIG_FAIR_GROUP_SCHED
7174 list_add(&init_task_group
.list
, &task_groups
);
7177 for_each_possible_cpu(i
) {
7181 spin_lock_init(&rq
->lock
);
7182 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7185 init_cfs_rq(&rq
->cfs
, rq
);
7186 init_rt_rq(&rq
->rt
, rq
);
7187 #ifdef CONFIG_FAIR_GROUP_SCHED
7188 init_task_group
.shares
= init_task_group_load
;
7189 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7190 init_tg_cfs_entry(rq
, &init_task_group
,
7191 &per_cpu(init_cfs_rq
, i
),
7192 &per_cpu(init_sched_entity
, i
), i
, 1);
7194 init_task_group
.rt_ratio
= sysctl_sched_rt_ratio
; /* XXX */
7195 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7196 init_tg_rt_entry(rq
, &init_task_group
,
7197 &per_cpu(init_rt_rq
, i
),
7198 &per_cpu(init_sched_rt_entity
, i
), i
, 1);
7200 rq
->rt_period_expire
= 0;
7201 rq
->rt_throttled
= 0;
7203 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7204 rq
->cpu_load
[j
] = 0;
7208 rq
->active_balance
= 0;
7209 rq
->next_balance
= jiffies
;
7212 rq
->migration_thread
= NULL
;
7213 INIT_LIST_HEAD(&rq
->migration_queue
);
7214 rq_attach_root(rq
, &def_root_domain
);
7217 atomic_set(&rq
->nr_iowait
, 0);
7221 set_load_weight(&init_task
);
7223 #ifdef CONFIG_PREEMPT_NOTIFIERS
7224 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7228 nr_cpu_ids
= highest_cpu
+ 1;
7229 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7232 #ifdef CONFIG_RT_MUTEXES
7233 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7237 * The boot idle thread does lazy MMU switching as well:
7239 atomic_inc(&init_mm
.mm_count
);
7240 enter_lazy_tlb(&init_mm
, current
);
7243 * Make us the idle thread. Technically, schedule() should not be
7244 * called from this thread, however somewhere below it might be,
7245 * but because we are the idle thread, we just pick up running again
7246 * when this runqueue becomes "idle".
7248 init_idle(current
, smp_processor_id());
7250 * During early bootup we pretend to be a normal task:
7252 current
->sched_class
= &fair_sched_class
;
7255 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7256 void __might_sleep(char *file
, int line
)
7259 static unsigned long prev_jiffy
; /* ratelimiting */
7261 if ((in_atomic() || irqs_disabled()) &&
7262 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7263 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7265 prev_jiffy
= jiffies
;
7266 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7267 " context at %s:%d\n", file
, line
);
7268 printk("in_atomic():%d, irqs_disabled():%d\n",
7269 in_atomic(), irqs_disabled());
7270 debug_show_held_locks(current
);
7271 if (irqs_disabled())
7272 print_irqtrace_events(current
);
7277 EXPORT_SYMBOL(__might_sleep
);
7280 #ifdef CONFIG_MAGIC_SYSRQ
7281 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7284 update_rq_clock(rq
);
7285 on_rq
= p
->se
.on_rq
;
7287 deactivate_task(rq
, p
, 0);
7288 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7290 activate_task(rq
, p
, 0);
7291 resched_task(rq
->curr
);
7295 void normalize_rt_tasks(void)
7297 struct task_struct
*g
, *p
;
7298 unsigned long flags
;
7301 read_lock_irq(&tasklist_lock
);
7302 do_each_thread(g
, p
) {
7304 * Only normalize user tasks:
7309 p
->se
.exec_start
= 0;
7310 #ifdef CONFIG_SCHEDSTATS
7311 p
->se
.wait_start
= 0;
7312 p
->se
.sleep_start
= 0;
7313 p
->se
.block_start
= 0;
7315 task_rq(p
)->clock
= 0;
7319 * Renice negative nice level userspace
7322 if (TASK_NICE(p
) < 0 && p
->mm
)
7323 set_user_nice(p
, 0);
7327 spin_lock_irqsave(&p
->pi_lock
, flags
);
7328 rq
= __task_rq_lock(p
);
7330 normalize_task(rq
, p
);
7332 __task_rq_unlock(rq
);
7333 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7334 } while_each_thread(g
, p
);
7336 read_unlock_irq(&tasklist_lock
);
7339 #endif /* CONFIG_MAGIC_SYSRQ */
7343 * These functions are only useful for the IA64 MCA handling.
7345 * They can only be called when the whole system has been
7346 * stopped - every CPU needs to be quiescent, and no scheduling
7347 * activity can take place. Using them for anything else would
7348 * be a serious bug, and as a result, they aren't even visible
7349 * under any other configuration.
7353 * curr_task - return the current task for a given cpu.
7354 * @cpu: the processor in question.
7356 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7358 struct task_struct
*curr_task(int cpu
)
7360 return cpu_curr(cpu
);
7364 * set_curr_task - set the current task for a given cpu.
7365 * @cpu: the processor in question.
7366 * @p: the task pointer to set.
7368 * Description: This function must only be used when non-maskable interrupts
7369 * are serviced on a separate stack. It allows the architecture to switch the
7370 * notion of the current task on a cpu in a non-blocking manner. This function
7371 * must be called with all CPU's synchronized, and interrupts disabled, the
7372 * and caller must save the original value of the current task (see
7373 * curr_task() above) and restore that value before reenabling interrupts and
7374 * re-starting the system.
7376 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7378 void set_curr_task(int cpu
, struct task_struct
*p
)
7385 #ifdef CONFIG_FAIR_GROUP_SCHED
7389 * distribute shares of all task groups among their schedulable entities,
7390 * to reflect load distribution across cpus.
7392 static int rebalance_shares(struct sched_domain
*sd
, int this_cpu
)
7394 struct cfs_rq
*cfs_rq
;
7395 struct rq
*rq
= cpu_rq(this_cpu
);
7396 cpumask_t sdspan
= sd
->span
;
7399 /* Walk thr' all the task groups that we have */
7400 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
7402 unsigned long total_load
= 0, total_shares
;
7403 struct task_group
*tg
= cfs_rq
->tg
;
7405 /* Gather total task load of this group across cpus */
7406 for_each_cpu_mask(i
, sdspan
)
7407 total_load
+= tg
->cfs_rq
[i
]->load
.weight
;
7409 /* Nothing to do if this group has no load */
7414 * tg->shares represents the number of cpu shares the task group
7415 * is eligible to hold on a single cpu. On N cpus, it is
7416 * eligible to hold (N * tg->shares) number of cpu shares.
7418 total_shares
= tg
->shares
* cpus_weight(sdspan
);
7421 * redistribute total_shares across cpus as per the task load
7424 for_each_cpu_mask(i
, sdspan
) {
7425 unsigned long local_load
, local_shares
;
7427 local_load
= tg
->cfs_rq
[i
]->load
.weight
;
7428 local_shares
= (local_load
* total_shares
) / total_load
;
7430 local_shares
= MIN_GROUP_SHARES
;
7431 if (local_shares
== tg
->se
[i
]->load
.weight
)
7434 spin_lock_irq(&cpu_rq(i
)->lock
);
7435 set_se_shares(tg
->se
[i
], local_shares
);
7436 spin_unlock_irq(&cpu_rq(i
)->lock
);
7445 * How frequently should we rebalance_shares() across cpus?
7447 * The more frequently we rebalance shares, the more accurate is the fairness
7448 * of cpu bandwidth distribution between task groups. However higher frequency
7449 * also implies increased scheduling overhead.
7451 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7452 * consecutive calls to rebalance_shares() in the same sched domain.
7454 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7455 * consecutive calls to rebalance_shares() in the same sched domain.
7457 * These settings allows for the appropriate trade-off between accuracy of
7458 * fairness and the associated overhead.
7462 /* default: 8ms, units: milliseconds */
7463 const_debug
unsigned int sysctl_sched_min_bal_int_shares
= 8;
7465 /* default: 128ms, units: milliseconds */
7466 const_debug
unsigned int sysctl_sched_max_bal_int_shares
= 128;
7468 /* kernel thread that runs rebalance_shares() periodically */
7469 static int load_balance_monitor(void *unused
)
7471 unsigned int timeout
= sysctl_sched_min_bal_int_shares
;
7472 struct sched_param schedparm
;
7476 * We don't want this thread's execution to be limited by the shares
7477 * assigned to default group (init_task_group). Hence make it run
7478 * as a SCHED_RR RT task at the lowest priority.
7480 schedparm
.sched_priority
= 1;
7481 ret
= sched_setscheduler(current
, SCHED_RR
, &schedparm
);
7483 printk(KERN_ERR
"Couldn't set SCHED_RR policy for load balance"
7484 " monitor thread (error = %d) \n", ret
);
7486 while (!kthread_should_stop()) {
7487 int i
, cpu
, balanced
= 1;
7489 /* Prevent cpus going down or coming up */
7491 /* lockout changes to doms_cur[] array */
7494 * Enter a rcu read-side critical section to safely walk rq->sd
7495 * chain on various cpus and to walk task group list
7496 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7500 for (i
= 0; i
< ndoms_cur
; i
++) {
7501 cpumask_t cpumap
= doms_cur
[i
];
7502 struct sched_domain
*sd
= NULL
, *sd_prev
= NULL
;
7504 cpu
= first_cpu(cpumap
);
7506 /* Find the highest domain at which to balance shares */
7507 for_each_domain(cpu
, sd
) {
7508 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7514 /* sd == NULL? No load balance reqd in this domain */
7518 balanced
&= rebalance_shares(sd
, cpu
);
7527 timeout
= sysctl_sched_min_bal_int_shares
;
7528 else if (timeout
< sysctl_sched_max_bal_int_shares
)
7531 msleep_interruptible(timeout
);
7536 #endif /* CONFIG_SMP */
7538 static void free_sched_group(struct task_group
*tg
)
7542 for_each_possible_cpu(i
) {
7544 kfree(tg
->cfs_rq
[i
]);
7548 kfree(tg
->rt_rq
[i
]);
7550 kfree(tg
->rt_se
[i
]);
7560 /* allocate runqueue etc for a new task group */
7561 struct task_group
*sched_create_group(void)
7563 struct task_group
*tg
;
7564 struct cfs_rq
*cfs_rq
;
7565 struct sched_entity
*se
;
7566 struct rt_rq
*rt_rq
;
7567 struct sched_rt_entity
*rt_se
;
7571 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7573 return ERR_PTR(-ENOMEM
);
7575 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7578 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7581 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * NR_CPUS
, GFP_KERNEL
);
7584 tg
->rt_se
= kzalloc(sizeof(rt_se
) * NR_CPUS
, GFP_KERNEL
);
7588 tg
->shares
= NICE_0_LOAD
;
7589 tg
->rt_ratio
= 0; /* XXX */
7591 for_each_possible_cpu(i
) {
7594 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
7595 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7599 se
= kmalloc_node(sizeof(struct sched_entity
),
7600 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7604 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
7605 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7609 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
7610 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7614 init_tg_cfs_entry(rq
, tg
, cfs_rq
, se
, i
, 0);
7615 init_tg_rt_entry(rq
, tg
, rt_rq
, rt_se
, i
, 0);
7618 lock_task_group_list();
7619 for_each_possible_cpu(i
) {
7621 cfs_rq
= tg
->cfs_rq
[i
];
7622 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7623 rt_rq
= tg
->rt_rq
[i
];
7624 list_add_rcu(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7626 list_add_rcu(&tg
->list
, &task_groups
);
7627 unlock_task_group_list();
7632 free_sched_group(tg
);
7633 return ERR_PTR(-ENOMEM
);
7636 /* rcu callback to free various structures associated with a task group */
7637 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7639 /* now it should be safe to free those cfs_rqs */
7640 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7643 /* Destroy runqueue etc associated with a task group */
7644 void sched_destroy_group(struct task_group
*tg
)
7646 struct cfs_rq
*cfs_rq
= NULL
;
7647 struct rt_rq
*rt_rq
= NULL
;
7650 lock_task_group_list();
7651 for_each_possible_cpu(i
) {
7652 cfs_rq
= tg
->cfs_rq
[i
];
7653 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7654 rt_rq
= tg
->rt_rq
[i
];
7655 list_del_rcu(&rt_rq
->leaf_rt_rq_list
);
7657 list_del_rcu(&tg
->list
);
7658 unlock_task_group_list();
7662 /* wait for possible concurrent references to cfs_rqs complete */
7663 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7666 /* change task's runqueue when it moves between groups.
7667 * The caller of this function should have put the task in its new group
7668 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7669 * reflect its new group.
7671 void sched_move_task(struct task_struct
*tsk
)
7674 unsigned long flags
;
7677 rq
= task_rq_lock(tsk
, &flags
);
7679 update_rq_clock(rq
);
7681 running
= task_current(rq
, tsk
);
7682 on_rq
= tsk
->se
.on_rq
;
7685 dequeue_task(rq
, tsk
, 0);
7686 if (unlikely(running
))
7687 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7690 set_task_rq(tsk
, task_cpu(tsk
));
7693 if (unlikely(running
))
7694 tsk
->sched_class
->set_curr_task(rq
);
7695 enqueue_task(rq
, tsk
, 0);
7698 task_rq_unlock(rq
, &flags
);
7701 /* rq->lock to be locked by caller */
7702 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7704 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7705 struct rq
*rq
= cfs_rq
->rq
;
7709 shares
= MIN_GROUP_SHARES
;
7713 dequeue_entity(cfs_rq
, se
, 0);
7714 dec_cpu_load(rq
, se
->load
.weight
);
7717 se
->load
.weight
= shares
;
7718 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7721 enqueue_entity(cfs_rq
, se
, 0);
7722 inc_cpu_load(rq
, se
->load
.weight
);
7726 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7729 struct cfs_rq
*cfs_rq
;
7732 lock_task_group_list();
7733 if (tg
->shares
== shares
)
7736 if (shares
< MIN_GROUP_SHARES
)
7737 shares
= MIN_GROUP_SHARES
;
7740 * Prevent any load balance activity (rebalance_shares,
7741 * load_balance_fair) from referring to this group first,
7742 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7744 for_each_possible_cpu(i
) {
7745 cfs_rq
= tg
->cfs_rq
[i
];
7746 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7749 /* wait for any ongoing reference to this group to finish */
7750 synchronize_sched();
7753 * Now we are free to modify the group's share on each cpu
7754 * w/o tripping rebalance_share or load_balance_fair.
7756 tg
->shares
= shares
;
7757 for_each_possible_cpu(i
) {
7758 spin_lock_irq(&cpu_rq(i
)->lock
);
7759 set_se_shares(tg
->se
[i
], shares
);
7760 spin_unlock_irq(&cpu_rq(i
)->lock
);
7764 * Enable load balance activity on this group, by inserting it back on
7765 * each cpu's rq->leaf_cfs_rq_list.
7767 for_each_possible_cpu(i
) {
7769 cfs_rq
= tg
->cfs_rq
[i
];
7770 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7773 unlock_task_group_list();
7777 unsigned long sched_group_shares(struct task_group
*tg
)
7783 * Ensure the total rt_ratio <= sysctl_sched_rt_ratio
7785 int sched_group_set_rt_ratio(struct task_group
*tg
, unsigned long rt_ratio
)
7787 struct task_group
*tgi
;
7788 unsigned long total
= 0;
7791 list_for_each_entry_rcu(tgi
, &task_groups
, list
)
7792 total
+= tgi
->rt_ratio
;
7795 if (total
+ rt_ratio
- tg
->rt_ratio
> sysctl_sched_rt_ratio
)
7798 tg
->rt_ratio
= rt_ratio
;
7802 unsigned long sched_group_rt_ratio(struct task_group
*tg
)
7804 return tg
->rt_ratio
;
7807 #endif /* CONFIG_FAIR_GROUP_SCHED */
7809 #ifdef CONFIG_FAIR_CGROUP_SCHED
7811 /* return corresponding task_group object of a cgroup */
7812 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7814 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7815 struct task_group
, css
);
7818 static struct cgroup_subsys_state
*
7819 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7821 struct task_group
*tg
;
7823 if (!cgrp
->parent
) {
7824 /* This is early initialization for the top cgroup */
7825 init_task_group
.css
.cgroup
= cgrp
;
7826 return &init_task_group
.css
;
7829 /* we support only 1-level deep hierarchical scheduler atm */
7830 if (cgrp
->parent
->parent
)
7831 return ERR_PTR(-EINVAL
);
7833 tg
= sched_create_group();
7835 return ERR_PTR(-ENOMEM
);
7837 /* Bind the cgroup to task_group object we just created */
7838 tg
->css
.cgroup
= cgrp
;
7844 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7846 struct task_group
*tg
= cgroup_tg(cgrp
);
7848 sched_destroy_group(tg
);
7852 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7853 struct task_struct
*tsk
)
7855 /* We don't support RT-tasks being in separate groups */
7856 if (tsk
->sched_class
!= &fair_sched_class
)
7863 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7864 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7866 sched_move_task(tsk
);
7869 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7872 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7875 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7877 struct task_group
*tg
= cgroup_tg(cgrp
);
7879 return (u64
) tg
->shares
;
7882 static int cpu_rt_ratio_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7885 return sched_group_set_rt_ratio(cgroup_tg(cgrp
), rt_ratio_val
);
7888 static u64
cpu_rt_ratio_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7890 struct task_group
*tg
= cgroup_tg(cgrp
);
7892 return (u64
) tg
->rt_ratio
;
7895 static struct cftype cpu_files
[] = {
7898 .read_uint
= cpu_shares_read_uint
,
7899 .write_uint
= cpu_shares_write_uint
,
7903 .read_uint
= cpu_rt_ratio_read_uint
,
7904 .write_uint
= cpu_rt_ratio_write_uint
,
7908 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7910 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7913 struct cgroup_subsys cpu_cgroup_subsys
= {
7915 .create
= cpu_cgroup_create
,
7916 .destroy
= cpu_cgroup_destroy
,
7917 .can_attach
= cpu_cgroup_can_attach
,
7918 .attach
= cpu_cgroup_attach
,
7919 .populate
= cpu_cgroup_populate
,
7920 .subsys_id
= cpu_cgroup_subsys_id
,
7924 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7926 #ifdef CONFIG_CGROUP_CPUACCT
7929 * CPU accounting code for task groups.
7931 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7932 * (balbir@in.ibm.com).
7935 /* track cpu usage of a group of tasks */
7937 struct cgroup_subsys_state css
;
7938 /* cpuusage holds pointer to a u64-type object on every cpu */
7942 struct cgroup_subsys cpuacct_subsys
;
7944 /* return cpu accounting group corresponding to this container */
7945 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7947 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7948 struct cpuacct
, css
);
7951 /* return cpu accounting group to which this task belongs */
7952 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7954 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7955 struct cpuacct
, css
);
7958 /* create a new cpu accounting group */
7959 static struct cgroup_subsys_state
*cpuacct_create(
7960 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7962 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7965 return ERR_PTR(-ENOMEM
);
7967 ca
->cpuusage
= alloc_percpu(u64
);
7968 if (!ca
->cpuusage
) {
7970 return ERR_PTR(-ENOMEM
);
7976 /* destroy an existing cpu accounting group */
7978 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7980 struct cpuacct
*ca
= cgroup_ca(cont
);
7982 free_percpu(ca
->cpuusage
);
7986 /* return total cpu usage (in nanoseconds) of a group */
7987 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
7989 struct cpuacct
*ca
= cgroup_ca(cont
);
7990 u64 totalcpuusage
= 0;
7993 for_each_possible_cpu(i
) {
7994 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
7997 * Take rq->lock to make 64-bit addition safe on 32-bit
8000 spin_lock_irq(&cpu_rq(i
)->lock
);
8001 totalcpuusage
+= *cpuusage
;
8002 spin_unlock_irq(&cpu_rq(i
)->lock
);
8005 return totalcpuusage
;
8008 static struct cftype files
[] = {
8011 .read_uint
= cpuusage_read
,
8015 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8017 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
8021 * charge this task's execution time to its accounting group.
8023 * called with rq->lock held.
8025 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8029 if (!cpuacct_subsys
.active
)
8034 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
8036 *cpuusage
+= cputime
;
8040 struct cgroup_subsys cpuacct_subsys
= {
8042 .create
= cpuacct_create
,
8043 .destroy
= cpuacct_destroy
,
8044 .populate
= cpuacct_populate
,
8045 .subsys_id
= cpuacct_subsys_id
,
8047 #endif /* CONFIG_CGROUP_CPUACCT */