4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
71 #include <asm/irq_regs.h>
74 * Scheduler clock - returns current time in nanosec units.
75 * This is default implementation.
76 * Architectures and sub-architectures can override this.
78 unsigned long long __attribute__((weak
)) sched_clock(void)
80 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
124 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
133 sg
->__cpu_power
+= val
;
134 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
138 static inline int rt_policy(int policy
)
140 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
145 static inline int task_has_rt_policy(struct task_struct
*p
)
147 return rt_policy(p
->policy
);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array
{
154 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
155 struct list_head queue
[MAX_RT_PRIO
];
158 #ifdef CONFIG_GROUP_SCHED
160 #include <linux/cgroup.h>
164 static LIST_HEAD(task_groups
);
166 /* task group related information */
168 #ifdef CONFIG_CGROUP_SCHED
169 struct cgroup_subsys_state css
;
172 #ifdef CONFIG_FAIR_GROUP_SCHED
173 /* schedulable entities of this group on each cpu */
174 struct sched_entity
**se
;
175 /* runqueue "owned" by this group on each cpu */
176 struct cfs_rq
**cfs_rq
;
177 unsigned long shares
;
180 #ifdef CONFIG_RT_GROUP_SCHED
181 struct sched_rt_entity
**rt_se
;
182 struct rt_rq
**rt_rq
;
188 struct list_head list
;
191 #ifdef CONFIG_FAIR_GROUP_SCHED
192 /* Default task group's sched entity on each cpu */
193 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
194 /* Default task group's cfs_rq on each cpu */
195 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
197 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
198 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
201 #ifdef CONFIG_RT_GROUP_SCHED
202 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
203 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
205 static struct sched_rt_entity
*init_sched_rt_entity_p
[NR_CPUS
];
206 static struct rt_rq
*init_rt_rq_p
[NR_CPUS
];
209 /* task_group_lock serializes add/remove of task groups and also changes to
210 * a task group's cpu shares.
212 static DEFINE_SPINLOCK(task_group_lock
);
214 /* doms_cur_mutex serializes access to doms_cur[] array */
215 static DEFINE_MUTEX(doms_cur_mutex
);
217 #ifdef CONFIG_FAIR_GROUP_SCHED
218 #ifdef CONFIG_USER_SCHED
219 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
221 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
224 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
227 /* Default task group.
228 * Every task in system belong to this group at bootup.
230 struct task_group init_task_group
= {
231 #ifdef CONFIG_FAIR_GROUP_SCHED
232 .se
= init_sched_entity_p
,
233 .cfs_rq
= init_cfs_rq_p
,
236 #ifdef CONFIG_RT_GROUP_SCHED
237 .rt_se
= init_sched_rt_entity_p
,
238 .rt_rq
= init_rt_rq_p
,
242 /* return group to which a task belongs */
243 static inline struct task_group
*task_group(struct task_struct
*p
)
245 struct task_group
*tg
;
247 #ifdef CONFIG_USER_SCHED
249 #elif defined(CONFIG_CGROUP_SCHED)
250 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
251 struct task_group
, css
);
253 tg
= &init_task_group
;
258 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
259 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
261 #ifdef CONFIG_FAIR_GROUP_SCHED
262 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
263 p
->se
.parent
= task_group(p
)->se
[cpu
];
266 #ifdef CONFIG_RT_GROUP_SCHED
267 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
268 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
272 static inline void lock_doms_cur(void)
274 mutex_lock(&doms_cur_mutex
);
277 static inline void unlock_doms_cur(void)
279 mutex_unlock(&doms_cur_mutex
);
284 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
285 static inline void lock_doms_cur(void) { }
286 static inline void unlock_doms_cur(void) { }
288 #endif /* CONFIG_GROUP_SCHED */
290 /* CFS-related fields in a runqueue */
292 struct load_weight load
;
293 unsigned long nr_running
;
298 struct rb_root tasks_timeline
;
299 struct rb_node
*rb_leftmost
;
300 struct rb_node
*rb_load_balance_curr
;
301 /* 'curr' points to currently running entity on this cfs_rq.
302 * It is set to NULL otherwise (i.e when none are currently running).
304 struct sched_entity
*curr
;
306 unsigned long nr_spread_over
;
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
312 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
313 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
314 * (like users, containers etc.)
316 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
317 * list is used during load balance.
319 struct list_head leaf_cfs_rq_list
;
320 struct task_group
*tg
; /* group that "owns" this runqueue */
324 /* Real-Time classes' related field in a runqueue: */
326 struct rt_prio_array active
;
327 unsigned long rt_nr_running
;
328 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
329 int highest_prio
; /* highest queued rt task prio */
332 unsigned long rt_nr_migratory
;
338 #ifdef CONFIG_RT_GROUP_SCHED
339 unsigned long rt_nr_boosted
;
342 struct list_head leaf_rt_rq_list
;
343 struct task_group
*tg
;
344 struct sched_rt_entity
*rt_se
;
351 * We add the notion of a root-domain which will be used to define per-domain
352 * variables. Each exclusive cpuset essentially defines an island domain by
353 * fully partitioning the member cpus from any other cpuset. Whenever a new
354 * exclusive cpuset is created, we also create and attach a new root-domain
364 * The "RT overload" flag: it gets set if a CPU has more than
365 * one runnable RT task.
372 * By default the system creates a single root-domain with all cpus as
373 * members (mimicking the global state we have today).
375 static struct root_domain def_root_domain
;
380 * This is the main, per-CPU runqueue data structure.
382 * Locking rule: those places that want to lock multiple runqueues
383 * (such as the load balancing or the thread migration code), lock
384 * acquire operations must be ordered by ascending &runqueue.
391 * nr_running and cpu_load should be in the same cacheline because
392 * remote CPUs use both these fields when doing load calculation.
394 unsigned long nr_running
;
395 #define CPU_LOAD_IDX_MAX 5
396 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
397 unsigned char idle_at_tick
;
399 unsigned char in_nohz_recently
;
401 /* capture load from *all* tasks on this cpu: */
402 struct load_weight load
;
403 unsigned long nr_load_updates
;
408 u64 rt_period_expire
;
411 #ifdef CONFIG_FAIR_GROUP_SCHED
412 /* list of leaf cfs_rq on this cpu: */
413 struct list_head leaf_cfs_rq_list
;
415 #ifdef CONFIG_RT_GROUP_SCHED
416 struct list_head leaf_rt_rq_list
;
420 * This is part of a global counter where only the total sum
421 * over all CPUs matters. A task can increase this counter on
422 * one CPU and if it got migrated afterwards it may decrease
423 * it on another CPU. Always updated under the runqueue lock:
425 unsigned long nr_uninterruptible
;
427 struct task_struct
*curr
, *idle
;
428 unsigned long next_balance
;
429 struct mm_struct
*prev_mm
;
431 u64 clock
, prev_clock_raw
;
434 unsigned int clock_warps
, clock_overflows
, clock_underflows
;
436 unsigned int clock_deep_idle_events
;
442 struct root_domain
*rd
;
443 struct sched_domain
*sd
;
445 /* For active balancing */
448 /* cpu of this runqueue: */
451 struct task_struct
*migration_thread
;
452 struct list_head migration_queue
;
455 #ifdef CONFIG_SCHED_HRTICK
456 unsigned long hrtick_flags
;
457 ktime_t hrtick_expire
;
458 struct hrtimer hrtick_timer
;
461 #ifdef CONFIG_SCHEDSTATS
463 struct sched_info rq_sched_info
;
465 /* sys_sched_yield() stats */
466 unsigned int yld_exp_empty
;
467 unsigned int yld_act_empty
;
468 unsigned int yld_both_empty
;
469 unsigned int yld_count
;
471 /* schedule() stats */
472 unsigned int sched_switch
;
473 unsigned int sched_count
;
474 unsigned int sched_goidle
;
476 /* try_to_wake_up() stats */
477 unsigned int ttwu_count
;
478 unsigned int ttwu_local
;
481 unsigned int bkl_count
;
483 struct lock_class_key rq_lock_key
;
486 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
488 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
490 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
493 static inline int cpu_of(struct rq
*rq
)
503 * Update the per-runqueue clock, as finegrained as the platform can give
504 * us, but without assuming monotonicity, etc.:
506 static void __update_rq_clock(struct rq
*rq
)
508 u64 prev_raw
= rq
->prev_clock_raw
;
509 u64 now
= sched_clock();
510 s64 delta
= now
- prev_raw
;
511 u64 clock
= rq
->clock
;
513 #ifdef CONFIG_SCHED_DEBUG
514 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
517 * Protect against sched_clock() occasionally going backwards:
519 if (unlikely(delta
< 0)) {
524 * Catch too large forward jumps too:
526 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
527 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
528 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
531 rq
->clock_overflows
++;
533 if (unlikely(delta
> rq
->clock_max_delta
))
534 rq
->clock_max_delta
= delta
;
539 rq
->prev_clock_raw
= now
;
543 static void update_rq_clock(struct rq
*rq
)
545 if (likely(smp_processor_id() == cpu_of(rq
)))
546 __update_rq_clock(rq
);
550 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
551 * See detach_destroy_domains: synchronize_sched for details.
553 * The domain tree of any CPU may only be accessed from within
554 * preempt-disabled sections.
556 #define for_each_domain(cpu, __sd) \
557 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
559 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
560 #define this_rq() (&__get_cpu_var(runqueues))
561 #define task_rq(p) cpu_rq(task_cpu(p))
562 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
564 unsigned long rt_needs_cpu(int cpu
)
566 struct rq
*rq
= cpu_rq(cpu
);
569 if (!rq
->rt_throttled
)
572 if (rq
->clock
> rq
->rt_period_expire
)
575 delta
= rq
->rt_period_expire
- rq
->clock
;
576 do_div(delta
, NSEC_PER_SEC
/ HZ
);
578 return (unsigned long)delta
;
582 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
584 #ifdef CONFIG_SCHED_DEBUG
585 # define const_debug __read_mostly
587 # define const_debug static const
591 * Debugging: various feature bits
594 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
595 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
596 SCHED_FEAT_START_DEBIT
= 4,
597 SCHED_FEAT_TREE_AVG
= 8,
598 SCHED_FEAT_APPROX_AVG
= 16,
599 SCHED_FEAT_HRTICK
= 32,
600 SCHED_FEAT_DOUBLE_TICK
= 64,
603 const_debug
unsigned int sysctl_sched_features
=
604 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
605 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
606 SCHED_FEAT_START_DEBIT
* 1 |
607 SCHED_FEAT_TREE_AVG
* 0 |
608 SCHED_FEAT_APPROX_AVG
* 0 |
609 SCHED_FEAT_HRTICK
* 1 |
610 SCHED_FEAT_DOUBLE_TICK
* 0;
612 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
615 * Number of tasks to iterate in a single balance run.
616 * Limited because this is done with IRQs disabled.
618 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
621 * period over which we measure -rt task cpu usage in us.
624 unsigned int sysctl_sched_rt_period
= 1000000;
626 static __read_mostly
int scheduler_running
;
629 * part of the period that we allow rt tasks to run in us.
632 int sysctl_sched_rt_runtime
= 950000;
635 * single value that denotes runtime == period, ie unlimited time.
637 #define RUNTIME_INF ((u64)~0ULL)
640 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
641 * clock constructed from sched_clock():
643 unsigned long long cpu_clock(int cpu
)
645 unsigned long long now
;
650 * Only call sched_clock() if the scheduler has already been
651 * initialized (some code might call cpu_clock() very early):
653 if (unlikely(!scheduler_running
))
656 local_irq_save(flags
);
660 local_irq_restore(flags
);
664 EXPORT_SYMBOL_GPL(cpu_clock
);
666 #ifndef prepare_arch_switch
667 # define prepare_arch_switch(next) do { } while (0)
669 #ifndef finish_arch_switch
670 # define finish_arch_switch(prev) do { } while (0)
673 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
675 return rq
->curr
== p
;
678 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
679 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
681 return task_current(rq
, p
);
684 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
688 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
690 #ifdef CONFIG_DEBUG_SPINLOCK
691 /* this is a valid case when another task releases the spinlock */
692 rq
->lock
.owner
= current
;
695 * If we are tracking spinlock dependencies then we have to
696 * fix up the runqueue lock - which gets 'carried over' from
699 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
701 spin_unlock_irq(&rq
->lock
);
704 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
705 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
710 return task_current(rq
, p
);
714 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
718 * We can optimise this out completely for !SMP, because the
719 * SMP rebalancing from interrupt is the only thing that cares
724 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
725 spin_unlock_irq(&rq
->lock
);
727 spin_unlock(&rq
->lock
);
731 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
735 * After ->oncpu is cleared, the task can be moved to a different CPU.
736 * We must ensure this doesn't happen until the switch is completely
742 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
746 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
749 * __task_rq_lock - lock the runqueue a given task resides on.
750 * Must be called interrupts disabled.
752 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
756 struct rq
*rq
= task_rq(p
);
757 spin_lock(&rq
->lock
);
758 if (likely(rq
== task_rq(p
)))
760 spin_unlock(&rq
->lock
);
765 * task_rq_lock - lock the runqueue a given task resides on and disable
766 * interrupts. Note the ordering: we can safely lookup the task_rq without
767 * explicitly disabling preemption.
769 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
775 local_irq_save(*flags
);
777 spin_lock(&rq
->lock
);
778 if (likely(rq
== task_rq(p
)))
780 spin_unlock_irqrestore(&rq
->lock
, *flags
);
784 static void __task_rq_unlock(struct rq
*rq
)
787 spin_unlock(&rq
->lock
);
790 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
793 spin_unlock_irqrestore(&rq
->lock
, *flags
);
797 * this_rq_lock - lock this runqueue and disable interrupts.
799 static struct rq
*this_rq_lock(void)
806 spin_lock(&rq
->lock
);
812 * We are going deep-idle (irqs are disabled):
814 void sched_clock_idle_sleep_event(void)
816 struct rq
*rq
= cpu_rq(smp_processor_id());
818 spin_lock(&rq
->lock
);
819 __update_rq_clock(rq
);
820 spin_unlock(&rq
->lock
);
821 rq
->clock_deep_idle_events
++;
823 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
826 * We just idled delta nanoseconds (called with irqs disabled):
828 void sched_clock_idle_wakeup_event(u64 delta_ns
)
830 struct rq
*rq
= cpu_rq(smp_processor_id());
831 u64 now
= sched_clock();
833 rq
->idle_clock
+= delta_ns
;
835 * Override the previous timestamp and ignore all
836 * sched_clock() deltas that occured while we idled,
837 * and use the PM-provided delta_ns to advance the
840 spin_lock(&rq
->lock
);
841 rq
->prev_clock_raw
= now
;
842 rq
->clock
+= delta_ns
;
843 spin_unlock(&rq
->lock
);
844 touch_softlockup_watchdog();
846 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
848 static void __resched_task(struct task_struct
*p
, int tif_bit
);
850 static inline void resched_task(struct task_struct
*p
)
852 __resched_task(p
, TIF_NEED_RESCHED
);
855 #ifdef CONFIG_SCHED_HRTICK
857 * Use HR-timers to deliver accurate preemption points.
859 * Its all a bit involved since we cannot program an hrt while holding the
860 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
863 * When we get rescheduled we reprogram the hrtick_timer outside of the
866 static inline void resched_hrt(struct task_struct
*p
)
868 __resched_task(p
, TIF_HRTICK_RESCHED
);
871 static inline void resched_rq(struct rq
*rq
)
875 spin_lock_irqsave(&rq
->lock
, flags
);
876 resched_task(rq
->curr
);
877 spin_unlock_irqrestore(&rq
->lock
, flags
);
881 HRTICK_SET
, /* re-programm hrtick_timer */
882 HRTICK_RESET
, /* not a new slice */
887 * - enabled by features
888 * - hrtimer is actually high res
890 static inline int hrtick_enabled(struct rq
*rq
)
892 if (!sched_feat(HRTICK
))
894 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
898 * Called to set the hrtick timer state.
900 * called with rq->lock held and irqs disabled
902 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
904 assert_spin_locked(&rq
->lock
);
907 * preempt at: now + delay
910 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
912 * indicate we need to program the timer
914 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
916 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
919 * New slices are called from the schedule path and don't need a
923 resched_hrt(rq
->curr
);
926 static void hrtick_clear(struct rq
*rq
)
928 if (hrtimer_active(&rq
->hrtick_timer
))
929 hrtimer_cancel(&rq
->hrtick_timer
);
933 * Update the timer from the possible pending state.
935 static void hrtick_set(struct rq
*rq
)
941 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
943 spin_lock_irqsave(&rq
->lock
, flags
);
944 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
945 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
946 time
= rq
->hrtick_expire
;
947 clear_thread_flag(TIF_HRTICK_RESCHED
);
948 spin_unlock_irqrestore(&rq
->lock
, flags
);
951 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
952 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
959 * High-resolution timer tick.
960 * Runs from hardirq context with interrupts disabled.
962 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
964 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
966 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
968 spin_lock(&rq
->lock
);
969 __update_rq_clock(rq
);
970 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
971 spin_unlock(&rq
->lock
);
973 return HRTIMER_NORESTART
;
976 static inline void init_rq_hrtick(struct rq
*rq
)
978 rq
->hrtick_flags
= 0;
979 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
980 rq
->hrtick_timer
.function
= hrtick
;
981 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
984 void hrtick_resched(void)
989 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
992 local_irq_save(flags
);
993 rq
= cpu_rq(smp_processor_id());
995 local_irq_restore(flags
);
998 static inline void hrtick_clear(struct rq
*rq
)
1002 static inline void hrtick_set(struct rq
*rq
)
1006 static inline void init_rq_hrtick(struct rq
*rq
)
1010 void hrtick_resched(void)
1016 * resched_task - mark a task 'to be rescheduled now'.
1018 * On UP this means the setting of the need_resched flag, on SMP it
1019 * might also involve a cross-CPU call to trigger the scheduler on
1024 #ifndef tsk_is_polling
1025 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1028 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1032 assert_spin_locked(&task_rq(p
)->lock
);
1034 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1037 set_tsk_thread_flag(p
, tif_bit
);
1040 if (cpu
== smp_processor_id())
1043 /* NEED_RESCHED must be visible before we test polling */
1045 if (!tsk_is_polling(p
))
1046 smp_send_reschedule(cpu
);
1049 static void resched_cpu(int cpu
)
1051 struct rq
*rq
= cpu_rq(cpu
);
1052 unsigned long flags
;
1054 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1056 resched_task(cpu_curr(cpu
));
1057 spin_unlock_irqrestore(&rq
->lock
, flags
);
1060 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1062 assert_spin_locked(&task_rq(p
)->lock
);
1063 set_tsk_thread_flag(p
, tif_bit
);
1067 #if BITS_PER_LONG == 32
1068 # define WMULT_CONST (~0UL)
1070 # define WMULT_CONST (1UL << 32)
1073 #define WMULT_SHIFT 32
1076 * Shift right and round:
1078 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1080 static unsigned long
1081 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1082 struct load_weight
*lw
)
1086 if (unlikely(!lw
->inv_weight
))
1087 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
1089 tmp
= (u64
)delta_exec
* weight
;
1091 * Check whether we'd overflow the 64-bit multiplication:
1093 if (unlikely(tmp
> WMULT_CONST
))
1094 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1097 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1099 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1102 static inline unsigned long
1103 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1105 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1108 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1114 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1121 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1122 * of tasks with abnormal "nice" values across CPUs the contribution that
1123 * each task makes to its run queue's load is weighted according to its
1124 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1125 * scaled version of the new time slice allocation that they receive on time
1129 #define WEIGHT_IDLEPRIO 2
1130 #define WMULT_IDLEPRIO (1 << 31)
1133 * Nice levels are multiplicative, with a gentle 10% change for every
1134 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1135 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1136 * that remained on nice 0.
1138 * The "10% effect" is relative and cumulative: from _any_ nice level,
1139 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1140 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1141 * If a task goes up by ~10% and another task goes down by ~10% then
1142 * the relative distance between them is ~25%.)
1144 static const int prio_to_weight
[40] = {
1145 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1146 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1147 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1148 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1149 /* 0 */ 1024, 820, 655, 526, 423,
1150 /* 5 */ 335, 272, 215, 172, 137,
1151 /* 10 */ 110, 87, 70, 56, 45,
1152 /* 15 */ 36, 29, 23, 18, 15,
1156 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1158 * In cases where the weight does not change often, we can use the
1159 * precalculated inverse to speed up arithmetics by turning divisions
1160 * into multiplications:
1162 static const u32 prio_to_wmult
[40] = {
1163 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1164 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1165 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1166 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1167 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1168 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1169 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1170 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1173 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1176 * runqueue iterator, to support SMP load-balancing between different
1177 * scheduling classes, without having to expose their internal data
1178 * structures to the load-balancing proper:
1180 struct rq_iterator
{
1182 struct task_struct
*(*start
)(void *);
1183 struct task_struct
*(*next
)(void *);
1187 static unsigned long
1188 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1189 unsigned long max_load_move
, struct sched_domain
*sd
,
1190 enum cpu_idle_type idle
, int *all_pinned
,
1191 int *this_best_prio
, struct rq_iterator
*iterator
);
1194 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1195 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1196 struct rq_iterator
*iterator
);
1199 #ifdef CONFIG_CGROUP_CPUACCT
1200 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1202 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1206 static unsigned long source_load(int cpu
, int type
);
1207 static unsigned long target_load(int cpu
, int type
);
1208 static unsigned long cpu_avg_load_per_task(int cpu
);
1209 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1210 #endif /* CONFIG_SMP */
1212 #include "sched_stats.h"
1213 #include "sched_idletask.c"
1214 #include "sched_fair.c"
1215 #include "sched_rt.c"
1216 #ifdef CONFIG_SCHED_DEBUG
1217 # include "sched_debug.c"
1220 #define sched_class_highest (&rt_sched_class)
1222 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
1224 update_load_add(&rq
->load
, p
->se
.load
.weight
);
1227 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
1229 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
1232 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1238 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1244 static void set_load_weight(struct task_struct
*p
)
1246 if (task_has_rt_policy(p
)) {
1247 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1248 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1253 * SCHED_IDLE tasks get minimal weight:
1255 if (p
->policy
== SCHED_IDLE
) {
1256 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1257 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1261 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1262 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1265 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1267 sched_info_queued(p
);
1268 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1272 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1274 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1279 * __normal_prio - return the priority that is based on the static prio
1281 static inline int __normal_prio(struct task_struct
*p
)
1283 return p
->static_prio
;
1287 * Calculate the expected normal priority: i.e. priority
1288 * without taking RT-inheritance into account. Might be
1289 * boosted by interactivity modifiers. Changes upon fork,
1290 * setprio syscalls, and whenever the interactivity
1291 * estimator recalculates.
1293 static inline int normal_prio(struct task_struct
*p
)
1297 if (task_has_rt_policy(p
))
1298 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1300 prio
= __normal_prio(p
);
1305 * Calculate the current priority, i.e. the priority
1306 * taken into account by the scheduler. This value might
1307 * be boosted by RT tasks, or might be boosted by
1308 * interactivity modifiers. Will be RT if the task got
1309 * RT-boosted. If not then it returns p->normal_prio.
1311 static int effective_prio(struct task_struct
*p
)
1313 p
->normal_prio
= normal_prio(p
);
1315 * If we are RT tasks or we were boosted to RT priority,
1316 * keep the priority unchanged. Otherwise, update priority
1317 * to the normal priority:
1319 if (!rt_prio(p
->prio
))
1320 return p
->normal_prio
;
1325 * activate_task - move a task to the runqueue.
1327 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1329 if (task_contributes_to_load(p
))
1330 rq
->nr_uninterruptible
--;
1332 enqueue_task(rq
, p
, wakeup
);
1333 inc_nr_running(p
, rq
);
1337 * deactivate_task - remove a task from the runqueue.
1339 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1341 if (task_contributes_to_load(p
))
1342 rq
->nr_uninterruptible
++;
1344 dequeue_task(rq
, p
, sleep
);
1345 dec_nr_running(p
, rq
);
1349 * task_curr - is this task currently executing on a CPU?
1350 * @p: the task in question.
1352 inline int task_curr(const struct task_struct
*p
)
1354 return cpu_curr(task_cpu(p
)) == p
;
1357 /* Used instead of source_load when we know the type == 0 */
1358 unsigned long weighted_cpuload(const int cpu
)
1360 return cpu_rq(cpu
)->load
.weight
;
1363 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1365 set_task_rq(p
, cpu
);
1368 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1369 * successfuly executed on another CPU. We must ensure that updates of
1370 * per-task data have been completed by this moment.
1373 task_thread_info(p
)->cpu
= cpu
;
1377 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1378 const struct sched_class
*prev_class
,
1379 int oldprio
, int running
)
1381 if (prev_class
!= p
->sched_class
) {
1382 if (prev_class
->switched_from
)
1383 prev_class
->switched_from(rq
, p
, running
);
1384 p
->sched_class
->switched_to(rq
, p
, running
);
1386 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1392 * Is this task likely cache-hot:
1395 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1399 if (p
->sched_class
!= &fair_sched_class
)
1402 if (sysctl_sched_migration_cost
== -1)
1404 if (sysctl_sched_migration_cost
== 0)
1407 delta
= now
- p
->se
.exec_start
;
1409 return delta
< (s64
)sysctl_sched_migration_cost
;
1413 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1415 int old_cpu
= task_cpu(p
);
1416 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1417 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1418 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1421 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1423 #ifdef CONFIG_SCHEDSTATS
1424 if (p
->se
.wait_start
)
1425 p
->se
.wait_start
-= clock_offset
;
1426 if (p
->se
.sleep_start
)
1427 p
->se
.sleep_start
-= clock_offset
;
1428 if (p
->se
.block_start
)
1429 p
->se
.block_start
-= clock_offset
;
1430 if (old_cpu
!= new_cpu
) {
1431 schedstat_inc(p
, se
.nr_migrations
);
1432 if (task_hot(p
, old_rq
->clock
, NULL
))
1433 schedstat_inc(p
, se
.nr_forced2_migrations
);
1436 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1437 new_cfsrq
->min_vruntime
;
1439 __set_task_cpu(p
, new_cpu
);
1442 struct migration_req
{
1443 struct list_head list
;
1445 struct task_struct
*task
;
1448 struct completion done
;
1452 * The task's runqueue lock must be held.
1453 * Returns true if you have to wait for migration thread.
1456 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1458 struct rq
*rq
= task_rq(p
);
1461 * If the task is not on a runqueue (and not running), then
1462 * it is sufficient to simply update the task's cpu field.
1464 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1465 set_task_cpu(p
, dest_cpu
);
1469 init_completion(&req
->done
);
1471 req
->dest_cpu
= dest_cpu
;
1472 list_add(&req
->list
, &rq
->migration_queue
);
1478 * wait_task_inactive - wait for a thread to unschedule.
1480 * The caller must ensure that the task *will* unschedule sometime soon,
1481 * else this function might spin for a *long* time. This function can't
1482 * be called with interrupts off, or it may introduce deadlock with
1483 * smp_call_function() if an IPI is sent by the same process we are
1484 * waiting to become inactive.
1486 void wait_task_inactive(struct task_struct
*p
)
1488 unsigned long flags
;
1494 * We do the initial early heuristics without holding
1495 * any task-queue locks at all. We'll only try to get
1496 * the runqueue lock when things look like they will
1502 * If the task is actively running on another CPU
1503 * still, just relax and busy-wait without holding
1506 * NOTE! Since we don't hold any locks, it's not
1507 * even sure that "rq" stays as the right runqueue!
1508 * But we don't care, since "task_running()" will
1509 * return false if the runqueue has changed and p
1510 * is actually now running somewhere else!
1512 while (task_running(rq
, p
))
1516 * Ok, time to look more closely! We need the rq
1517 * lock now, to be *sure*. If we're wrong, we'll
1518 * just go back and repeat.
1520 rq
= task_rq_lock(p
, &flags
);
1521 running
= task_running(rq
, p
);
1522 on_rq
= p
->se
.on_rq
;
1523 task_rq_unlock(rq
, &flags
);
1526 * Was it really running after all now that we
1527 * checked with the proper locks actually held?
1529 * Oops. Go back and try again..
1531 if (unlikely(running
)) {
1537 * It's not enough that it's not actively running,
1538 * it must be off the runqueue _entirely_, and not
1541 * So if it wa still runnable (but just not actively
1542 * running right now), it's preempted, and we should
1543 * yield - it could be a while.
1545 if (unlikely(on_rq
)) {
1546 schedule_timeout_uninterruptible(1);
1551 * Ahh, all good. It wasn't running, and it wasn't
1552 * runnable, which means that it will never become
1553 * running in the future either. We're all done!
1560 * kick_process - kick a running thread to enter/exit the kernel
1561 * @p: the to-be-kicked thread
1563 * Cause a process which is running on another CPU to enter
1564 * kernel-mode, without any delay. (to get signals handled.)
1566 * NOTE: this function doesnt have to take the runqueue lock,
1567 * because all it wants to ensure is that the remote task enters
1568 * the kernel. If the IPI races and the task has been migrated
1569 * to another CPU then no harm is done and the purpose has been
1572 void kick_process(struct task_struct
*p
)
1578 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1579 smp_send_reschedule(cpu
);
1584 * Return a low guess at the load of a migration-source cpu weighted
1585 * according to the scheduling class and "nice" value.
1587 * We want to under-estimate the load of migration sources, to
1588 * balance conservatively.
1590 static unsigned long source_load(int cpu
, int type
)
1592 struct rq
*rq
= cpu_rq(cpu
);
1593 unsigned long total
= weighted_cpuload(cpu
);
1598 return min(rq
->cpu_load
[type
-1], total
);
1602 * Return a high guess at the load of a migration-target cpu weighted
1603 * according to the scheduling class and "nice" value.
1605 static unsigned long target_load(int cpu
, int type
)
1607 struct rq
*rq
= cpu_rq(cpu
);
1608 unsigned long total
= weighted_cpuload(cpu
);
1613 return max(rq
->cpu_load
[type
-1], total
);
1617 * Return the average load per task on the cpu's run queue
1619 static unsigned long cpu_avg_load_per_task(int cpu
)
1621 struct rq
*rq
= cpu_rq(cpu
);
1622 unsigned long total
= weighted_cpuload(cpu
);
1623 unsigned long n
= rq
->nr_running
;
1625 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1629 * find_idlest_group finds and returns the least busy CPU group within the
1632 static struct sched_group
*
1633 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1635 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1636 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1637 int load_idx
= sd
->forkexec_idx
;
1638 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1641 unsigned long load
, avg_load
;
1645 /* Skip over this group if it has no CPUs allowed */
1646 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1649 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1651 /* Tally up the load of all CPUs in the group */
1654 for_each_cpu_mask(i
, group
->cpumask
) {
1655 /* Bias balancing toward cpus of our domain */
1657 load
= source_load(i
, load_idx
);
1659 load
= target_load(i
, load_idx
);
1664 /* Adjust by relative CPU power of the group */
1665 avg_load
= sg_div_cpu_power(group
,
1666 avg_load
* SCHED_LOAD_SCALE
);
1669 this_load
= avg_load
;
1671 } else if (avg_load
< min_load
) {
1672 min_load
= avg_load
;
1675 } while (group
= group
->next
, group
!= sd
->groups
);
1677 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1683 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1686 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1689 unsigned long load
, min_load
= ULONG_MAX
;
1693 /* Traverse only the allowed CPUs */
1694 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1696 for_each_cpu_mask(i
, tmp
) {
1697 load
= weighted_cpuload(i
);
1699 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1709 * sched_balance_self: balance the current task (running on cpu) in domains
1710 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1713 * Balance, ie. select the least loaded group.
1715 * Returns the target CPU number, or the same CPU if no balancing is needed.
1717 * preempt must be disabled.
1719 static int sched_balance_self(int cpu
, int flag
)
1721 struct task_struct
*t
= current
;
1722 struct sched_domain
*tmp
, *sd
= NULL
;
1724 for_each_domain(cpu
, tmp
) {
1726 * If power savings logic is enabled for a domain, stop there.
1728 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1730 if (tmp
->flags
& flag
)
1736 struct sched_group
*group
;
1737 int new_cpu
, weight
;
1739 if (!(sd
->flags
& flag
)) {
1745 group
= find_idlest_group(sd
, t
, cpu
);
1751 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1752 if (new_cpu
== -1 || new_cpu
== cpu
) {
1753 /* Now try balancing at a lower domain level of cpu */
1758 /* Now try balancing at a lower domain level of new_cpu */
1761 weight
= cpus_weight(span
);
1762 for_each_domain(cpu
, tmp
) {
1763 if (weight
<= cpus_weight(tmp
->span
))
1765 if (tmp
->flags
& flag
)
1768 /* while loop will break here if sd == NULL */
1774 #endif /* CONFIG_SMP */
1777 * try_to_wake_up - wake up a thread
1778 * @p: the to-be-woken-up thread
1779 * @state: the mask of task states that can be woken
1780 * @sync: do a synchronous wakeup?
1782 * Put it on the run-queue if it's not already there. The "current"
1783 * thread is always on the run-queue (except when the actual
1784 * re-schedule is in progress), and as such you're allowed to do
1785 * the simpler "current->state = TASK_RUNNING" to mark yourself
1786 * runnable without the overhead of this.
1788 * returns failure only if the task is already active.
1790 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1792 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1793 unsigned long flags
;
1798 rq
= task_rq_lock(p
, &flags
);
1799 old_state
= p
->state
;
1800 if (!(old_state
& state
))
1808 this_cpu
= smp_processor_id();
1811 if (unlikely(task_running(rq
, p
)))
1814 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1815 if (cpu
!= orig_cpu
) {
1816 set_task_cpu(p
, cpu
);
1817 task_rq_unlock(rq
, &flags
);
1818 /* might preempt at this point */
1819 rq
= task_rq_lock(p
, &flags
);
1820 old_state
= p
->state
;
1821 if (!(old_state
& state
))
1826 this_cpu
= smp_processor_id();
1830 #ifdef CONFIG_SCHEDSTATS
1831 schedstat_inc(rq
, ttwu_count
);
1832 if (cpu
== this_cpu
)
1833 schedstat_inc(rq
, ttwu_local
);
1835 struct sched_domain
*sd
;
1836 for_each_domain(this_cpu
, sd
) {
1837 if (cpu_isset(cpu
, sd
->span
)) {
1838 schedstat_inc(sd
, ttwu_wake_remote
);
1846 #endif /* CONFIG_SMP */
1847 schedstat_inc(p
, se
.nr_wakeups
);
1849 schedstat_inc(p
, se
.nr_wakeups_sync
);
1850 if (orig_cpu
!= cpu
)
1851 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1852 if (cpu
== this_cpu
)
1853 schedstat_inc(p
, se
.nr_wakeups_local
);
1855 schedstat_inc(p
, se
.nr_wakeups_remote
);
1856 update_rq_clock(rq
);
1857 activate_task(rq
, p
, 1);
1858 check_preempt_curr(rq
, p
);
1862 p
->state
= TASK_RUNNING
;
1864 if (p
->sched_class
->task_wake_up
)
1865 p
->sched_class
->task_wake_up(rq
, p
);
1868 task_rq_unlock(rq
, &flags
);
1873 int wake_up_process(struct task_struct
*p
)
1875 return try_to_wake_up(p
, TASK_ALL
, 0);
1877 EXPORT_SYMBOL(wake_up_process
);
1879 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1881 return try_to_wake_up(p
, state
, 0);
1885 * Perform scheduler related setup for a newly forked process p.
1886 * p is forked by current.
1888 * __sched_fork() is basic setup used by init_idle() too:
1890 static void __sched_fork(struct task_struct
*p
)
1892 p
->se
.exec_start
= 0;
1893 p
->se
.sum_exec_runtime
= 0;
1894 p
->se
.prev_sum_exec_runtime
= 0;
1896 #ifdef CONFIG_SCHEDSTATS
1897 p
->se
.wait_start
= 0;
1898 p
->se
.sum_sleep_runtime
= 0;
1899 p
->se
.sleep_start
= 0;
1900 p
->se
.block_start
= 0;
1901 p
->se
.sleep_max
= 0;
1902 p
->se
.block_max
= 0;
1904 p
->se
.slice_max
= 0;
1908 INIT_LIST_HEAD(&p
->rt
.run_list
);
1911 #ifdef CONFIG_PREEMPT_NOTIFIERS
1912 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1916 * We mark the process as running here, but have not actually
1917 * inserted it onto the runqueue yet. This guarantees that
1918 * nobody will actually run it, and a signal or other external
1919 * event cannot wake it up and insert it on the runqueue either.
1921 p
->state
= TASK_RUNNING
;
1925 * fork()/clone()-time setup:
1927 void sched_fork(struct task_struct
*p
, int clone_flags
)
1929 int cpu
= get_cpu();
1934 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1936 set_task_cpu(p
, cpu
);
1939 * Make sure we do not leak PI boosting priority to the child:
1941 p
->prio
= current
->normal_prio
;
1942 if (!rt_prio(p
->prio
))
1943 p
->sched_class
= &fair_sched_class
;
1945 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1946 if (likely(sched_info_on()))
1947 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1949 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1952 #ifdef CONFIG_PREEMPT
1953 /* Want to start with kernel preemption disabled. */
1954 task_thread_info(p
)->preempt_count
= 1;
1960 * wake_up_new_task - wake up a newly created task for the first time.
1962 * This function will do some initial scheduler statistics housekeeping
1963 * that must be done for every newly created context, then puts the task
1964 * on the runqueue and wakes it.
1966 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1968 unsigned long flags
;
1971 rq
= task_rq_lock(p
, &flags
);
1972 BUG_ON(p
->state
!= TASK_RUNNING
);
1973 update_rq_clock(rq
);
1975 p
->prio
= effective_prio(p
);
1977 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1978 activate_task(rq
, p
, 0);
1981 * Let the scheduling class do new task startup
1982 * management (if any):
1984 p
->sched_class
->task_new(rq
, p
);
1985 inc_nr_running(p
, rq
);
1987 check_preempt_curr(rq
, p
);
1989 if (p
->sched_class
->task_wake_up
)
1990 p
->sched_class
->task_wake_up(rq
, p
);
1992 task_rq_unlock(rq
, &flags
);
1995 #ifdef CONFIG_PREEMPT_NOTIFIERS
1998 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1999 * @notifier: notifier struct to register
2001 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2003 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2005 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2008 * preempt_notifier_unregister - no longer interested in preemption notifications
2009 * @notifier: notifier struct to unregister
2011 * This is safe to call from within a preemption notifier.
2013 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2015 hlist_del(¬ifier
->link
);
2017 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2019 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2021 struct preempt_notifier
*notifier
;
2022 struct hlist_node
*node
;
2024 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2025 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2029 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2030 struct task_struct
*next
)
2032 struct preempt_notifier
*notifier
;
2033 struct hlist_node
*node
;
2035 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2036 notifier
->ops
->sched_out(notifier
, next
);
2041 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2046 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2047 struct task_struct
*next
)
2054 * prepare_task_switch - prepare to switch tasks
2055 * @rq: the runqueue preparing to switch
2056 * @prev: the current task that is being switched out
2057 * @next: the task we are going to switch to.
2059 * This is called with the rq lock held and interrupts off. It must
2060 * be paired with a subsequent finish_task_switch after the context
2063 * prepare_task_switch sets up locking and calls architecture specific
2067 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2068 struct task_struct
*next
)
2070 fire_sched_out_preempt_notifiers(prev
, next
);
2071 prepare_lock_switch(rq
, next
);
2072 prepare_arch_switch(next
);
2076 * finish_task_switch - clean up after a task-switch
2077 * @rq: runqueue associated with task-switch
2078 * @prev: the thread we just switched away from.
2080 * finish_task_switch must be called after the context switch, paired
2081 * with a prepare_task_switch call before the context switch.
2082 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2083 * and do any other architecture-specific cleanup actions.
2085 * Note that we may have delayed dropping an mm in context_switch(). If
2086 * so, we finish that here outside of the runqueue lock. (Doing it
2087 * with the lock held can cause deadlocks; see schedule() for
2090 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2091 __releases(rq
->lock
)
2093 struct mm_struct
*mm
= rq
->prev_mm
;
2099 * A task struct has one reference for the use as "current".
2100 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2101 * schedule one last time. The schedule call will never return, and
2102 * the scheduled task must drop that reference.
2103 * The test for TASK_DEAD must occur while the runqueue locks are
2104 * still held, otherwise prev could be scheduled on another cpu, die
2105 * there before we look at prev->state, and then the reference would
2107 * Manfred Spraul <manfred@colorfullife.com>
2109 prev_state
= prev
->state
;
2110 finish_arch_switch(prev
);
2111 finish_lock_switch(rq
, prev
);
2113 if (current
->sched_class
->post_schedule
)
2114 current
->sched_class
->post_schedule(rq
);
2117 fire_sched_in_preempt_notifiers(current
);
2120 if (unlikely(prev_state
== TASK_DEAD
)) {
2122 * Remove function-return probe instances associated with this
2123 * task and put them back on the free list.
2125 kprobe_flush_task(prev
);
2126 put_task_struct(prev
);
2131 * schedule_tail - first thing a freshly forked thread must call.
2132 * @prev: the thread we just switched away from.
2134 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2135 __releases(rq
->lock
)
2137 struct rq
*rq
= this_rq();
2139 finish_task_switch(rq
, prev
);
2140 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2141 /* In this case, finish_task_switch does not reenable preemption */
2144 if (current
->set_child_tid
)
2145 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2149 * context_switch - switch to the new MM and the new
2150 * thread's register state.
2153 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2154 struct task_struct
*next
)
2156 struct mm_struct
*mm
, *oldmm
;
2158 prepare_task_switch(rq
, prev
, next
);
2160 oldmm
= prev
->active_mm
;
2162 * For paravirt, this is coupled with an exit in switch_to to
2163 * combine the page table reload and the switch backend into
2166 arch_enter_lazy_cpu_mode();
2168 if (unlikely(!mm
)) {
2169 next
->active_mm
= oldmm
;
2170 atomic_inc(&oldmm
->mm_count
);
2171 enter_lazy_tlb(oldmm
, next
);
2173 switch_mm(oldmm
, mm
, next
);
2175 if (unlikely(!prev
->mm
)) {
2176 prev
->active_mm
= NULL
;
2177 rq
->prev_mm
= oldmm
;
2180 * Since the runqueue lock will be released by the next
2181 * task (which is an invalid locking op but in the case
2182 * of the scheduler it's an obvious special-case), so we
2183 * do an early lockdep release here:
2185 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2186 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2189 /* Here we just switch the register state and the stack. */
2190 switch_to(prev
, next
, prev
);
2194 * this_rq must be evaluated again because prev may have moved
2195 * CPUs since it called schedule(), thus the 'rq' on its stack
2196 * frame will be invalid.
2198 finish_task_switch(this_rq(), prev
);
2202 * nr_running, nr_uninterruptible and nr_context_switches:
2204 * externally visible scheduler statistics: current number of runnable
2205 * threads, current number of uninterruptible-sleeping threads, total
2206 * number of context switches performed since bootup.
2208 unsigned long nr_running(void)
2210 unsigned long i
, sum
= 0;
2212 for_each_online_cpu(i
)
2213 sum
+= cpu_rq(i
)->nr_running
;
2218 unsigned long nr_uninterruptible(void)
2220 unsigned long i
, sum
= 0;
2222 for_each_possible_cpu(i
)
2223 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2226 * Since we read the counters lockless, it might be slightly
2227 * inaccurate. Do not allow it to go below zero though:
2229 if (unlikely((long)sum
< 0))
2235 unsigned long long nr_context_switches(void)
2238 unsigned long long sum
= 0;
2240 for_each_possible_cpu(i
)
2241 sum
+= cpu_rq(i
)->nr_switches
;
2246 unsigned long nr_iowait(void)
2248 unsigned long i
, sum
= 0;
2250 for_each_possible_cpu(i
)
2251 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2256 unsigned long nr_active(void)
2258 unsigned long i
, running
= 0, uninterruptible
= 0;
2260 for_each_online_cpu(i
) {
2261 running
+= cpu_rq(i
)->nr_running
;
2262 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2265 if (unlikely((long)uninterruptible
< 0))
2266 uninterruptible
= 0;
2268 return running
+ uninterruptible
;
2272 * Update rq->cpu_load[] statistics. This function is usually called every
2273 * scheduler tick (TICK_NSEC).
2275 static void update_cpu_load(struct rq
*this_rq
)
2277 unsigned long this_load
= this_rq
->load
.weight
;
2280 this_rq
->nr_load_updates
++;
2282 /* Update our load: */
2283 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2284 unsigned long old_load
, new_load
;
2286 /* scale is effectively 1 << i now, and >> i divides by scale */
2288 old_load
= this_rq
->cpu_load
[i
];
2289 new_load
= this_load
;
2291 * Round up the averaging division if load is increasing. This
2292 * prevents us from getting stuck on 9 if the load is 10, for
2295 if (new_load
> old_load
)
2296 new_load
+= scale
-1;
2297 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2304 * double_rq_lock - safely lock two runqueues
2306 * Note this does not disable interrupts like task_rq_lock,
2307 * you need to do so manually before calling.
2309 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2310 __acquires(rq1
->lock
)
2311 __acquires(rq2
->lock
)
2313 BUG_ON(!irqs_disabled());
2315 spin_lock(&rq1
->lock
);
2316 __acquire(rq2
->lock
); /* Fake it out ;) */
2319 spin_lock(&rq1
->lock
);
2320 spin_lock(&rq2
->lock
);
2322 spin_lock(&rq2
->lock
);
2323 spin_lock(&rq1
->lock
);
2326 update_rq_clock(rq1
);
2327 update_rq_clock(rq2
);
2331 * double_rq_unlock - safely unlock two runqueues
2333 * Note this does not restore interrupts like task_rq_unlock,
2334 * you need to do so manually after calling.
2336 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2337 __releases(rq1
->lock
)
2338 __releases(rq2
->lock
)
2340 spin_unlock(&rq1
->lock
);
2342 spin_unlock(&rq2
->lock
);
2344 __release(rq2
->lock
);
2348 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2350 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2351 __releases(this_rq
->lock
)
2352 __acquires(busiest
->lock
)
2353 __acquires(this_rq
->lock
)
2357 if (unlikely(!irqs_disabled())) {
2358 /* printk() doesn't work good under rq->lock */
2359 spin_unlock(&this_rq
->lock
);
2362 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2363 if (busiest
< this_rq
) {
2364 spin_unlock(&this_rq
->lock
);
2365 spin_lock(&busiest
->lock
);
2366 spin_lock(&this_rq
->lock
);
2369 spin_lock(&busiest
->lock
);
2375 * If dest_cpu is allowed for this process, migrate the task to it.
2376 * This is accomplished by forcing the cpu_allowed mask to only
2377 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2378 * the cpu_allowed mask is restored.
2380 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2382 struct migration_req req
;
2383 unsigned long flags
;
2386 rq
= task_rq_lock(p
, &flags
);
2387 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2388 || unlikely(cpu_is_offline(dest_cpu
)))
2391 /* force the process onto the specified CPU */
2392 if (migrate_task(p
, dest_cpu
, &req
)) {
2393 /* Need to wait for migration thread (might exit: take ref). */
2394 struct task_struct
*mt
= rq
->migration_thread
;
2396 get_task_struct(mt
);
2397 task_rq_unlock(rq
, &flags
);
2398 wake_up_process(mt
);
2399 put_task_struct(mt
);
2400 wait_for_completion(&req
.done
);
2405 task_rq_unlock(rq
, &flags
);
2409 * sched_exec - execve() is a valuable balancing opportunity, because at
2410 * this point the task has the smallest effective memory and cache footprint.
2412 void sched_exec(void)
2414 int new_cpu
, this_cpu
= get_cpu();
2415 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2417 if (new_cpu
!= this_cpu
)
2418 sched_migrate_task(current
, new_cpu
);
2422 * pull_task - move a task from a remote runqueue to the local runqueue.
2423 * Both runqueues must be locked.
2425 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2426 struct rq
*this_rq
, int this_cpu
)
2428 deactivate_task(src_rq
, p
, 0);
2429 set_task_cpu(p
, this_cpu
);
2430 activate_task(this_rq
, p
, 0);
2432 * Note that idle threads have a prio of MAX_PRIO, for this test
2433 * to be always true for them.
2435 check_preempt_curr(this_rq
, p
);
2439 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2442 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2443 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2447 * We do not migrate tasks that are:
2448 * 1) running (obviously), or
2449 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2450 * 3) are cache-hot on their current CPU.
2452 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2453 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2458 if (task_running(rq
, p
)) {
2459 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2464 * Aggressive migration if:
2465 * 1) task is cache cold, or
2466 * 2) too many balance attempts have failed.
2469 if (!task_hot(p
, rq
->clock
, sd
) ||
2470 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2471 #ifdef CONFIG_SCHEDSTATS
2472 if (task_hot(p
, rq
->clock
, sd
)) {
2473 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2474 schedstat_inc(p
, se
.nr_forced_migrations
);
2480 if (task_hot(p
, rq
->clock
, sd
)) {
2481 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2487 static unsigned long
2488 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2489 unsigned long max_load_move
, struct sched_domain
*sd
,
2490 enum cpu_idle_type idle
, int *all_pinned
,
2491 int *this_best_prio
, struct rq_iterator
*iterator
)
2493 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2494 struct task_struct
*p
;
2495 long rem_load_move
= max_load_move
;
2497 if (max_load_move
== 0)
2503 * Start the load-balancing iterator:
2505 p
= iterator
->start(iterator
->arg
);
2507 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2510 * To help distribute high priority tasks across CPUs we don't
2511 * skip a task if it will be the highest priority task (i.e. smallest
2512 * prio value) on its new queue regardless of its load weight
2514 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2515 SCHED_LOAD_SCALE_FUZZ
;
2516 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2517 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2518 p
= iterator
->next(iterator
->arg
);
2522 pull_task(busiest
, p
, this_rq
, this_cpu
);
2524 rem_load_move
-= p
->se
.load
.weight
;
2527 * We only want to steal up to the prescribed amount of weighted load.
2529 if (rem_load_move
> 0) {
2530 if (p
->prio
< *this_best_prio
)
2531 *this_best_prio
= p
->prio
;
2532 p
= iterator
->next(iterator
->arg
);
2537 * Right now, this is one of only two places pull_task() is called,
2538 * so we can safely collect pull_task() stats here rather than
2539 * inside pull_task().
2541 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2544 *all_pinned
= pinned
;
2546 return max_load_move
- rem_load_move
;
2550 * move_tasks tries to move up to max_load_move weighted load from busiest to
2551 * this_rq, as part of a balancing operation within domain "sd".
2552 * Returns 1 if successful and 0 otherwise.
2554 * Called with both runqueues locked.
2556 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2557 unsigned long max_load_move
,
2558 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2561 const struct sched_class
*class = sched_class_highest
;
2562 unsigned long total_load_moved
= 0;
2563 int this_best_prio
= this_rq
->curr
->prio
;
2567 class->load_balance(this_rq
, this_cpu
, busiest
,
2568 max_load_move
- total_load_moved
,
2569 sd
, idle
, all_pinned
, &this_best_prio
);
2570 class = class->next
;
2571 } while (class && max_load_move
> total_load_moved
);
2573 return total_load_moved
> 0;
2577 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2578 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2579 struct rq_iterator
*iterator
)
2581 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2585 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2586 pull_task(busiest
, p
, this_rq
, this_cpu
);
2588 * Right now, this is only the second place pull_task()
2589 * is called, so we can safely collect pull_task()
2590 * stats here rather than inside pull_task().
2592 schedstat_inc(sd
, lb_gained
[idle
]);
2596 p
= iterator
->next(iterator
->arg
);
2603 * move_one_task tries to move exactly one task from busiest to this_rq, as
2604 * part of active balancing operations within "domain".
2605 * Returns 1 if successful and 0 otherwise.
2607 * Called with both runqueues locked.
2609 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2610 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2612 const struct sched_class
*class;
2614 for (class = sched_class_highest
; class; class = class->next
)
2615 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2622 * find_busiest_group finds and returns the busiest CPU group within the
2623 * domain. It calculates and returns the amount of weighted load which
2624 * should be moved to restore balance via the imbalance parameter.
2626 static struct sched_group
*
2627 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2628 unsigned long *imbalance
, enum cpu_idle_type idle
,
2629 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2631 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2632 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2633 unsigned long max_pull
;
2634 unsigned long busiest_load_per_task
, busiest_nr_running
;
2635 unsigned long this_load_per_task
, this_nr_running
;
2636 int load_idx
, group_imb
= 0;
2637 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2638 int power_savings_balance
= 1;
2639 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2640 unsigned long min_nr_running
= ULONG_MAX
;
2641 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2644 max_load
= this_load
= total_load
= total_pwr
= 0;
2645 busiest_load_per_task
= busiest_nr_running
= 0;
2646 this_load_per_task
= this_nr_running
= 0;
2647 if (idle
== CPU_NOT_IDLE
)
2648 load_idx
= sd
->busy_idx
;
2649 else if (idle
== CPU_NEWLY_IDLE
)
2650 load_idx
= sd
->newidle_idx
;
2652 load_idx
= sd
->idle_idx
;
2655 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2658 int __group_imb
= 0;
2659 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2660 unsigned long sum_nr_running
, sum_weighted_load
;
2662 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2665 balance_cpu
= first_cpu(group
->cpumask
);
2667 /* Tally up the load of all CPUs in the group */
2668 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2670 min_cpu_load
= ~0UL;
2672 for_each_cpu_mask(i
, group
->cpumask
) {
2675 if (!cpu_isset(i
, *cpus
))
2680 if (*sd_idle
&& rq
->nr_running
)
2683 /* Bias balancing toward cpus of our domain */
2685 if (idle_cpu(i
) && !first_idle_cpu
) {
2690 load
= target_load(i
, load_idx
);
2692 load
= source_load(i
, load_idx
);
2693 if (load
> max_cpu_load
)
2694 max_cpu_load
= load
;
2695 if (min_cpu_load
> load
)
2696 min_cpu_load
= load
;
2700 sum_nr_running
+= rq
->nr_running
;
2701 sum_weighted_load
+= weighted_cpuload(i
);
2705 * First idle cpu or the first cpu(busiest) in this sched group
2706 * is eligible for doing load balancing at this and above
2707 * domains. In the newly idle case, we will allow all the cpu's
2708 * to do the newly idle load balance.
2710 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2711 balance_cpu
!= this_cpu
&& balance
) {
2716 total_load
+= avg_load
;
2717 total_pwr
+= group
->__cpu_power
;
2719 /* Adjust by relative CPU power of the group */
2720 avg_load
= sg_div_cpu_power(group
,
2721 avg_load
* SCHED_LOAD_SCALE
);
2723 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2726 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2729 this_load
= avg_load
;
2731 this_nr_running
= sum_nr_running
;
2732 this_load_per_task
= sum_weighted_load
;
2733 } else if (avg_load
> max_load
&&
2734 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2735 max_load
= avg_load
;
2737 busiest_nr_running
= sum_nr_running
;
2738 busiest_load_per_task
= sum_weighted_load
;
2739 group_imb
= __group_imb
;
2742 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2744 * Busy processors will not participate in power savings
2747 if (idle
== CPU_NOT_IDLE
||
2748 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2752 * If the local group is idle or completely loaded
2753 * no need to do power savings balance at this domain
2755 if (local_group
&& (this_nr_running
>= group_capacity
||
2757 power_savings_balance
= 0;
2760 * If a group is already running at full capacity or idle,
2761 * don't include that group in power savings calculations
2763 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2768 * Calculate the group which has the least non-idle load.
2769 * This is the group from where we need to pick up the load
2772 if ((sum_nr_running
< min_nr_running
) ||
2773 (sum_nr_running
== min_nr_running
&&
2774 first_cpu(group
->cpumask
) <
2775 first_cpu(group_min
->cpumask
))) {
2777 min_nr_running
= sum_nr_running
;
2778 min_load_per_task
= sum_weighted_load
/
2783 * Calculate the group which is almost near its
2784 * capacity but still has some space to pick up some load
2785 * from other group and save more power
2787 if (sum_nr_running
<= group_capacity
- 1) {
2788 if (sum_nr_running
> leader_nr_running
||
2789 (sum_nr_running
== leader_nr_running
&&
2790 first_cpu(group
->cpumask
) >
2791 first_cpu(group_leader
->cpumask
))) {
2792 group_leader
= group
;
2793 leader_nr_running
= sum_nr_running
;
2798 group
= group
->next
;
2799 } while (group
!= sd
->groups
);
2801 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2804 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2806 if (this_load
>= avg_load
||
2807 100*max_load
<= sd
->imbalance_pct
*this_load
)
2810 busiest_load_per_task
/= busiest_nr_running
;
2812 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2815 * We're trying to get all the cpus to the average_load, so we don't
2816 * want to push ourselves above the average load, nor do we wish to
2817 * reduce the max loaded cpu below the average load, as either of these
2818 * actions would just result in more rebalancing later, and ping-pong
2819 * tasks around. Thus we look for the minimum possible imbalance.
2820 * Negative imbalances (*we* are more loaded than anyone else) will
2821 * be counted as no imbalance for these purposes -- we can't fix that
2822 * by pulling tasks to us. Be careful of negative numbers as they'll
2823 * appear as very large values with unsigned longs.
2825 if (max_load
<= busiest_load_per_task
)
2829 * In the presence of smp nice balancing, certain scenarios can have
2830 * max load less than avg load(as we skip the groups at or below
2831 * its cpu_power, while calculating max_load..)
2833 if (max_load
< avg_load
) {
2835 goto small_imbalance
;
2838 /* Don't want to pull so many tasks that a group would go idle */
2839 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2841 /* How much load to actually move to equalise the imbalance */
2842 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2843 (avg_load
- this_load
) * this->__cpu_power
)
2847 * if *imbalance is less than the average load per runnable task
2848 * there is no gaurantee that any tasks will be moved so we'll have
2849 * a think about bumping its value to force at least one task to be
2852 if (*imbalance
< busiest_load_per_task
) {
2853 unsigned long tmp
, pwr_now
, pwr_move
;
2857 pwr_move
= pwr_now
= 0;
2859 if (this_nr_running
) {
2860 this_load_per_task
/= this_nr_running
;
2861 if (busiest_load_per_task
> this_load_per_task
)
2864 this_load_per_task
= SCHED_LOAD_SCALE
;
2866 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2867 busiest_load_per_task
* imbn
) {
2868 *imbalance
= busiest_load_per_task
;
2873 * OK, we don't have enough imbalance to justify moving tasks,
2874 * however we may be able to increase total CPU power used by
2878 pwr_now
+= busiest
->__cpu_power
*
2879 min(busiest_load_per_task
, max_load
);
2880 pwr_now
+= this->__cpu_power
*
2881 min(this_load_per_task
, this_load
);
2882 pwr_now
/= SCHED_LOAD_SCALE
;
2884 /* Amount of load we'd subtract */
2885 tmp
= sg_div_cpu_power(busiest
,
2886 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2888 pwr_move
+= busiest
->__cpu_power
*
2889 min(busiest_load_per_task
, max_load
- tmp
);
2891 /* Amount of load we'd add */
2892 if (max_load
* busiest
->__cpu_power
<
2893 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2894 tmp
= sg_div_cpu_power(this,
2895 max_load
* busiest
->__cpu_power
);
2897 tmp
= sg_div_cpu_power(this,
2898 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2899 pwr_move
+= this->__cpu_power
*
2900 min(this_load_per_task
, this_load
+ tmp
);
2901 pwr_move
/= SCHED_LOAD_SCALE
;
2903 /* Move if we gain throughput */
2904 if (pwr_move
> pwr_now
)
2905 *imbalance
= busiest_load_per_task
;
2911 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2912 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2915 if (this == group_leader
&& group_leader
!= group_min
) {
2916 *imbalance
= min_load_per_task
;
2926 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2929 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2930 unsigned long imbalance
, cpumask_t
*cpus
)
2932 struct rq
*busiest
= NULL
, *rq
;
2933 unsigned long max_load
= 0;
2936 for_each_cpu_mask(i
, group
->cpumask
) {
2939 if (!cpu_isset(i
, *cpus
))
2943 wl
= weighted_cpuload(i
);
2945 if (rq
->nr_running
== 1 && wl
> imbalance
)
2948 if (wl
> max_load
) {
2958 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2959 * so long as it is large enough.
2961 #define MAX_PINNED_INTERVAL 512
2964 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2965 * tasks if there is an imbalance.
2967 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2968 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2971 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2972 struct sched_group
*group
;
2973 unsigned long imbalance
;
2975 cpumask_t cpus
= CPU_MASK_ALL
;
2976 unsigned long flags
;
2979 * When power savings policy is enabled for the parent domain, idle
2980 * sibling can pick up load irrespective of busy siblings. In this case,
2981 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2982 * portraying it as CPU_NOT_IDLE.
2984 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2985 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2988 schedstat_inc(sd
, lb_count
[idle
]);
2991 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2998 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3002 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
3004 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3008 BUG_ON(busiest
== this_rq
);
3010 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3013 if (busiest
->nr_running
> 1) {
3015 * Attempt to move tasks. If find_busiest_group has found
3016 * an imbalance but busiest->nr_running <= 1, the group is
3017 * still unbalanced. ld_moved simply stays zero, so it is
3018 * correctly treated as an imbalance.
3020 local_irq_save(flags
);
3021 double_rq_lock(this_rq
, busiest
);
3022 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3023 imbalance
, sd
, idle
, &all_pinned
);
3024 double_rq_unlock(this_rq
, busiest
);
3025 local_irq_restore(flags
);
3028 * some other cpu did the load balance for us.
3030 if (ld_moved
&& this_cpu
!= smp_processor_id())
3031 resched_cpu(this_cpu
);
3033 /* All tasks on this runqueue were pinned by CPU affinity */
3034 if (unlikely(all_pinned
)) {
3035 cpu_clear(cpu_of(busiest
), cpus
);
3036 if (!cpus_empty(cpus
))
3043 schedstat_inc(sd
, lb_failed
[idle
]);
3044 sd
->nr_balance_failed
++;
3046 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3048 spin_lock_irqsave(&busiest
->lock
, flags
);
3050 /* don't kick the migration_thread, if the curr
3051 * task on busiest cpu can't be moved to this_cpu
3053 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3054 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3056 goto out_one_pinned
;
3059 if (!busiest
->active_balance
) {
3060 busiest
->active_balance
= 1;
3061 busiest
->push_cpu
= this_cpu
;
3064 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3066 wake_up_process(busiest
->migration_thread
);
3069 * We've kicked active balancing, reset the failure
3072 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3075 sd
->nr_balance_failed
= 0;
3077 if (likely(!active_balance
)) {
3078 /* We were unbalanced, so reset the balancing interval */
3079 sd
->balance_interval
= sd
->min_interval
;
3082 * If we've begun active balancing, start to back off. This
3083 * case may not be covered by the all_pinned logic if there
3084 * is only 1 task on the busy runqueue (because we don't call
3087 if (sd
->balance_interval
< sd
->max_interval
)
3088 sd
->balance_interval
*= 2;
3091 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3092 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3097 schedstat_inc(sd
, lb_balanced
[idle
]);
3099 sd
->nr_balance_failed
= 0;
3102 /* tune up the balancing interval */
3103 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3104 (sd
->balance_interval
< sd
->max_interval
))
3105 sd
->balance_interval
*= 2;
3107 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3108 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3114 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3115 * tasks if there is an imbalance.
3117 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3118 * this_rq is locked.
3121 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
3123 struct sched_group
*group
;
3124 struct rq
*busiest
= NULL
;
3125 unsigned long imbalance
;
3129 cpumask_t cpus
= CPU_MASK_ALL
;
3132 * When power savings policy is enabled for the parent domain, idle
3133 * sibling can pick up load irrespective of busy siblings. In this case,
3134 * let the state of idle sibling percolate up as IDLE, instead of
3135 * portraying it as CPU_NOT_IDLE.
3137 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3138 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3141 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3143 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3144 &sd_idle
, &cpus
, NULL
);
3146 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3150 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
3153 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3157 BUG_ON(busiest
== this_rq
);
3159 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3162 if (busiest
->nr_running
> 1) {
3163 /* Attempt to move tasks */
3164 double_lock_balance(this_rq
, busiest
);
3165 /* this_rq->clock is already updated */
3166 update_rq_clock(busiest
);
3167 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3168 imbalance
, sd
, CPU_NEWLY_IDLE
,
3170 spin_unlock(&busiest
->lock
);
3172 if (unlikely(all_pinned
)) {
3173 cpu_clear(cpu_of(busiest
), cpus
);
3174 if (!cpus_empty(cpus
))
3180 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3181 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3182 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3185 sd
->nr_balance_failed
= 0;
3190 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3191 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3192 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3194 sd
->nr_balance_failed
= 0;
3200 * idle_balance is called by schedule() if this_cpu is about to become
3201 * idle. Attempts to pull tasks from other CPUs.
3203 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3205 struct sched_domain
*sd
;
3206 int pulled_task
= -1;
3207 unsigned long next_balance
= jiffies
+ HZ
;
3209 for_each_domain(this_cpu
, sd
) {
3210 unsigned long interval
;
3212 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3215 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3216 /* If we've pulled tasks over stop searching: */
3217 pulled_task
= load_balance_newidle(this_cpu
,
3220 interval
= msecs_to_jiffies(sd
->balance_interval
);
3221 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3222 next_balance
= sd
->last_balance
+ interval
;
3226 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3228 * We are going idle. next_balance may be set based on
3229 * a busy processor. So reset next_balance.
3231 this_rq
->next_balance
= next_balance
;
3236 * active_load_balance is run by migration threads. It pushes running tasks
3237 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3238 * running on each physical CPU where possible, and avoids physical /
3239 * logical imbalances.
3241 * Called with busiest_rq locked.
3243 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3245 int target_cpu
= busiest_rq
->push_cpu
;
3246 struct sched_domain
*sd
;
3247 struct rq
*target_rq
;
3249 /* Is there any task to move? */
3250 if (busiest_rq
->nr_running
<= 1)
3253 target_rq
= cpu_rq(target_cpu
);
3256 * This condition is "impossible", if it occurs
3257 * we need to fix it. Originally reported by
3258 * Bjorn Helgaas on a 128-cpu setup.
3260 BUG_ON(busiest_rq
== target_rq
);
3262 /* move a task from busiest_rq to target_rq */
3263 double_lock_balance(busiest_rq
, target_rq
);
3264 update_rq_clock(busiest_rq
);
3265 update_rq_clock(target_rq
);
3267 /* Search for an sd spanning us and the target CPU. */
3268 for_each_domain(target_cpu
, sd
) {
3269 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3270 cpu_isset(busiest_cpu
, sd
->span
))
3275 schedstat_inc(sd
, alb_count
);
3277 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3279 schedstat_inc(sd
, alb_pushed
);
3281 schedstat_inc(sd
, alb_failed
);
3283 spin_unlock(&target_rq
->lock
);
3288 atomic_t load_balancer
;
3290 } nohz ____cacheline_aligned
= {
3291 .load_balancer
= ATOMIC_INIT(-1),
3292 .cpu_mask
= CPU_MASK_NONE
,
3296 * This routine will try to nominate the ilb (idle load balancing)
3297 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3298 * load balancing on behalf of all those cpus. If all the cpus in the system
3299 * go into this tickless mode, then there will be no ilb owner (as there is
3300 * no need for one) and all the cpus will sleep till the next wakeup event
3303 * For the ilb owner, tick is not stopped. And this tick will be used
3304 * for idle load balancing. ilb owner will still be part of
3307 * While stopping the tick, this cpu will become the ilb owner if there
3308 * is no other owner. And will be the owner till that cpu becomes busy
3309 * or if all cpus in the system stop their ticks at which point
3310 * there is no need for ilb owner.
3312 * When the ilb owner becomes busy, it nominates another owner, during the
3313 * next busy scheduler_tick()
3315 int select_nohz_load_balancer(int stop_tick
)
3317 int cpu
= smp_processor_id();
3320 cpu_set(cpu
, nohz
.cpu_mask
);
3321 cpu_rq(cpu
)->in_nohz_recently
= 1;
3324 * If we are going offline and still the leader, give up!
3326 if (cpu_is_offline(cpu
) &&
3327 atomic_read(&nohz
.load_balancer
) == cpu
) {
3328 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3333 /* time for ilb owner also to sleep */
3334 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3335 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3336 atomic_set(&nohz
.load_balancer
, -1);
3340 if (atomic_read(&nohz
.load_balancer
) == -1) {
3341 /* make me the ilb owner */
3342 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3344 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3347 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3350 cpu_clear(cpu
, nohz
.cpu_mask
);
3352 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3353 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3360 static DEFINE_SPINLOCK(balancing
);
3363 * It checks each scheduling domain to see if it is due to be balanced,
3364 * and initiates a balancing operation if so.
3366 * Balancing parameters are set up in arch_init_sched_domains.
3368 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3371 struct rq
*rq
= cpu_rq(cpu
);
3372 unsigned long interval
;
3373 struct sched_domain
*sd
;
3374 /* Earliest time when we have to do rebalance again */
3375 unsigned long next_balance
= jiffies
+ 60*HZ
;
3376 int update_next_balance
= 0;
3378 for_each_domain(cpu
, sd
) {
3379 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3382 interval
= sd
->balance_interval
;
3383 if (idle
!= CPU_IDLE
)
3384 interval
*= sd
->busy_factor
;
3386 /* scale ms to jiffies */
3387 interval
= msecs_to_jiffies(interval
);
3388 if (unlikely(!interval
))
3390 if (interval
> HZ
*NR_CPUS
/10)
3391 interval
= HZ
*NR_CPUS
/10;
3394 if (sd
->flags
& SD_SERIALIZE
) {
3395 if (!spin_trylock(&balancing
))
3399 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3400 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3402 * We've pulled tasks over so either we're no
3403 * longer idle, or one of our SMT siblings is
3406 idle
= CPU_NOT_IDLE
;
3408 sd
->last_balance
= jiffies
;
3410 if (sd
->flags
& SD_SERIALIZE
)
3411 spin_unlock(&balancing
);
3413 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3414 next_balance
= sd
->last_balance
+ interval
;
3415 update_next_balance
= 1;
3419 * Stop the load balance at this level. There is another
3420 * CPU in our sched group which is doing load balancing more
3428 * next_balance will be updated only when there is a need.
3429 * When the cpu is attached to null domain for ex, it will not be
3432 if (likely(update_next_balance
))
3433 rq
->next_balance
= next_balance
;
3437 * run_rebalance_domains is triggered when needed from the scheduler tick.
3438 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3439 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3441 static void run_rebalance_domains(struct softirq_action
*h
)
3443 int this_cpu
= smp_processor_id();
3444 struct rq
*this_rq
= cpu_rq(this_cpu
);
3445 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3446 CPU_IDLE
: CPU_NOT_IDLE
;
3448 rebalance_domains(this_cpu
, idle
);
3452 * If this cpu is the owner for idle load balancing, then do the
3453 * balancing on behalf of the other idle cpus whose ticks are
3456 if (this_rq
->idle_at_tick
&&
3457 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3458 cpumask_t cpus
= nohz
.cpu_mask
;
3462 cpu_clear(this_cpu
, cpus
);
3463 for_each_cpu_mask(balance_cpu
, cpus
) {
3465 * If this cpu gets work to do, stop the load balancing
3466 * work being done for other cpus. Next load
3467 * balancing owner will pick it up.
3472 rebalance_domains(balance_cpu
, CPU_IDLE
);
3474 rq
= cpu_rq(balance_cpu
);
3475 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3476 this_rq
->next_balance
= rq
->next_balance
;
3483 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3485 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3486 * idle load balancing owner or decide to stop the periodic load balancing,
3487 * if the whole system is idle.
3489 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3493 * If we were in the nohz mode recently and busy at the current
3494 * scheduler tick, then check if we need to nominate new idle
3497 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3498 rq
->in_nohz_recently
= 0;
3500 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3501 cpu_clear(cpu
, nohz
.cpu_mask
);
3502 atomic_set(&nohz
.load_balancer
, -1);
3505 if (atomic_read(&nohz
.load_balancer
) == -1) {
3507 * simple selection for now: Nominate the
3508 * first cpu in the nohz list to be the next
3511 * TBD: Traverse the sched domains and nominate
3512 * the nearest cpu in the nohz.cpu_mask.
3514 int ilb
= first_cpu(nohz
.cpu_mask
);
3522 * If this cpu is idle and doing idle load balancing for all the
3523 * cpus with ticks stopped, is it time for that to stop?
3525 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3526 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3532 * If this cpu is idle and the idle load balancing is done by
3533 * someone else, then no need raise the SCHED_SOFTIRQ
3535 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3536 cpu_isset(cpu
, nohz
.cpu_mask
))
3539 if (time_after_eq(jiffies
, rq
->next_balance
))
3540 raise_softirq(SCHED_SOFTIRQ
);
3543 #else /* CONFIG_SMP */
3546 * on UP we do not need to balance between CPUs:
3548 static inline void idle_balance(int cpu
, struct rq
*rq
)
3554 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3556 EXPORT_PER_CPU_SYMBOL(kstat
);
3559 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3560 * that have not yet been banked in case the task is currently running.
3562 unsigned long long task_sched_runtime(struct task_struct
*p
)
3564 unsigned long flags
;
3568 rq
= task_rq_lock(p
, &flags
);
3569 ns
= p
->se
.sum_exec_runtime
;
3570 if (task_current(rq
, p
)) {
3571 update_rq_clock(rq
);
3572 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3573 if ((s64
)delta_exec
> 0)
3576 task_rq_unlock(rq
, &flags
);
3582 * Account user cpu time to a process.
3583 * @p: the process that the cpu time gets accounted to
3584 * @cputime: the cpu time spent in user space since the last update
3586 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3588 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3591 p
->utime
= cputime_add(p
->utime
, cputime
);
3593 /* Add user time to cpustat. */
3594 tmp
= cputime_to_cputime64(cputime
);
3595 if (TASK_NICE(p
) > 0)
3596 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3598 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3602 * Account guest cpu time to a process.
3603 * @p: the process that the cpu time gets accounted to
3604 * @cputime: the cpu time spent in virtual machine since the last update
3606 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3609 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3611 tmp
= cputime_to_cputime64(cputime
);
3613 p
->utime
= cputime_add(p
->utime
, cputime
);
3614 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3616 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3617 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3621 * Account scaled user cpu time to a process.
3622 * @p: the process that the cpu time gets accounted to
3623 * @cputime: the cpu time spent in user space since the last update
3625 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3627 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3631 * Account system cpu time to a process.
3632 * @p: the process that the cpu time gets accounted to
3633 * @hardirq_offset: the offset to subtract from hardirq_count()
3634 * @cputime: the cpu time spent in kernel space since the last update
3636 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3639 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3640 struct rq
*rq
= this_rq();
3643 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3644 return account_guest_time(p
, cputime
);
3646 p
->stime
= cputime_add(p
->stime
, cputime
);
3648 /* Add system time to cpustat. */
3649 tmp
= cputime_to_cputime64(cputime
);
3650 if (hardirq_count() - hardirq_offset
)
3651 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3652 else if (softirq_count())
3653 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3654 else if (p
!= rq
->idle
)
3655 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3656 else if (atomic_read(&rq
->nr_iowait
) > 0)
3657 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3659 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3660 /* Account for system time used */
3661 acct_update_integrals(p
);
3665 * Account scaled system cpu time to a process.
3666 * @p: the process that the cpu time gets accounted to
3667 * @hardirq_offset: the offset to subtract from hardirq_count()
3668 * @cputime: the cpu time spent in kernel space since the last update
3670 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3672 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3676 * Account for involuntary wait time.
3677 * @p: the process from which the cpu time has been stolen
3678 * @steal: the cpu time spent in involuntary wait
3680 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3682 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3683 cputime64_t tmp
= cputime_to_cputime64(steal
);
3684 struct rq
*rq
= this_rq();
3686 if (p
== rq
->idle
) {
3687 p
->stime
= cputime_add(p
->stime
, steal
);
3688 if (atomic_read(&rq
->nr_iowait
) > 0)
3689 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3691 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3693 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3697 * This function gets called by the timer code, with HZ frequency.
3698 * We call it with interrupts disabled.
3700 * It also gets called by the fork code, when changing the parent's
3703 void scheduler_tick(void)
3705 int cpu
= smp_processor_id();
3706 struct rq
*rq
= cpu_rq(cpu
);
3707 struct task_struct
*curr
= rq
->curr
;
3708 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3710 spin_lock(&rq
->lock
);
3711 __update_rq_clock(rq
);
3713 * Let rq->clock advance by at least TICK_NSEC:
3715 if (unlikely(rq
->clock
< next_tick
)) {
3716 rq
->clock
= next_tick
;
3717 rq
->clock_underflows
++;
3719 rq
->tick_timestamp
= rq
->clock
;
3720 update_cpu_load(rq
);
3721 curr
->sched_class
->task_tick(rq
, curr
, 0);
3722 update_sched_rt_period(rq
);
3723 spin_unlock(&rq
->lock
);
3726 rq
->idle_at_tick
= idle_cpu(cpu
);
3727 trigger_load_balance(rq
, cpu
);
3731 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3733 void __kprobes
add_preempt_count(int val
)
3738 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3740 preempt_count() += val
;
3742 * Spinlock count overflowing soon?
3744 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3747 EXPORT_SYMBOL(add_preempt_count
);
3749 void __kprobes
sub_preempt_count(int val
)
3754 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3757 * Is the spinlock portion underflowing?
3759 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3760 !(preempt_count() & PREEMPT_MASK
)))
3763 preempt_count() -= val
;
3765 EXPORT_SYMBOL(sub_preempt_count
);
3770 * Print scheduling while atomic bug:
3772 static noinline
void __schedule_bug(struct task_struct
*prev
)
3774 struct pt_regs
*regs
= get_irq_regs();
3776 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3777 prev
->comm
, prev
->pid
, preempt_count());
3779 debug_show_held_locks(prev
);
3780 if (irqs_disabled())
3781 print_irqtrace_events(prev
);
3790 * Various schedule()-time debugging checks and statistics:
3792 static inline void schedule_debug(struct task_struct
*prev
)
3795 * Test if we are atomic. Since do_exit() needs to call into
3796 * schedule() atomically, we ignore that path for now.
3797 * Otherwise, whine if we are scheduling when we should not be.
3799 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3800 __schedule_bug(prev
);
3802 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3804 schedstat_inc(this_rq(), sched_count
);
3805 #ifdef CONFIG_SCHEDSTATS
3806 if (unlikely(prev
->lock_depth
>= 0)) {
3807 schedstat_inc(this_rq(), bkl_count
);
3808 schedstat_inc(prev
, sched_info
.bkl_count
);
3814 * Pick up the highest-prio task:
3816 static inline struct task_struct
*
3817 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3819 const struct sched_class
*class;
3820 struct task_struct
*p
;
3823 * Optimization: we know that if all tasks are in
3824 * the fair class we can call that function directly:
3826 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3827 p
= fair_sched_class
.pick_next_task(rq
);
3832 class = sched_class_highest
;
3834 p
= class->pick_next_task(rq
);
3838 * Will never be NULL as the idle class always
3839 * returns a non-NULL p:
3841 class = class->next
;
3846 * schedule() is the main scheduler function.
3848 asmlinkage
void __sched
schedule(void)
3850 struct task_struct
*prev
, *next
;
3851 unsigned long *switch_count
;
3857 cpu
= smp_processor_id();
3861 switch_count
= &prev
->nivcsw
;
3863 release_kernel_lock(prev
);
3864 need_resched_nonpreemptible
:
3866 schedule_debug(prev
);
3871 * Do the rq-clock update outside the rq lock:
3873 local_irq_disable();
3874 __update_rq_clock(rq
);
3875 spin_lock(&rq
->lock
);
3876 clear_tsk_need_resched(prev
);
3878 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3879 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3880 unlikely(signal_pending(prev
)))) {
3881 prev
->state
= TASK_RUNNING
;
3883 deactivate_task(rq
, prev
, 1);
3885 switch_count
= &prev
->nvcsw
;
3889 if (prev
->sched_class
->pre_schedule
)
3890 prev
->sched_class
->pre_schedule(rq
, prev
);
3893 if (unlikely(!rq
->nr_running
))
3894 idle_balance(cpu
, rq
);
3896 prev
->sched_class
->put_prev_task(rq
, prev
);
3897 next
= pick_next_task(rq
, prev
);
3899 sched_info_switch(prev
, next
);
3901 if (likely(prev
!= next
)) {
3906 context_switch(rq
, prev
, next
); /* unlocks the rq */
3908 * the context switch might have flipped the stack from under
3909 * us, hence refresh the local variables.
3911 cpu
= smp_processor_id();
3914 spin_unlock_irq(&rq
->lock
);
3918 if (unlikely(reacquire_kernel_lock(current
) < 0))
3919 goto need_resched_nonpreemptible
;
3921 preempt_enable_no_resched();
3922 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3925 EXPORT_SYMBOL(schedule
);
3927 #ifdef CONFIG_PREEMPT
3929 * this is the entry point to schedule() from in-kernel preemption
3930 * off of preempt_enable. Kernel preemptions off return from interrupt
3931 * occur there and call schedule directly.
3933 asmlinkage
void __sched
preempt_schedule(void)
3935 struct thread_info
*ti
= current_thread_info();
3936 struct task_struct
*task
= current
;
3937 int saved_lock_depth
;
3940 * If there is a non-zero preempt_count or interrupts are disabled,
3941 * we do not want to preempt the current task. Just return..
3943 if (likely(ti
->preempt_count
|| irqs_disabled()))
3947 add_preempt_count(PREEMPT_ACTIVE
);
3950 * We keep the big kernel semaphore locked, but we
3951 * clear ->lock_depth so that schedule() doesnt
3952 * auto-release the semaphore:
3954 saved_lock_depth
= task
->lock_depth
;
3955 task
->lock_depth
= -1;
3957 task
->lock_depth
= saved_lock_depth
;
3958 sub_preempt_count(PREEMPT_ACTIVE
);
3961 * Check again in case we missed a preemption opportunity
3962 * between schedule and now.
3965 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3967 EXPORT_SYMBOL(preempt_schedule
);
3970 * this is the entry point to schedule() from kernel preemption
3971 * off of irq context.
3972 * Note, that this is called and return with irqs disabled. This will
3973 * protect us against recursive calling from irq.
3975 asmlinkage
void __sched
preempt_schedule_irq(void)
3977 struct thread_info
*ti
= current_thread_info();
3978 struct task_struct
*task
= current
;
3979 int saved_lock_depth
;
3981 /* Catch callers which need to be fixed */
3982 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3985 add_preempt_count(PREEMPT_ACTIVE
);
3988 * We keep the big kernel semaphore locked, but we
3989 * clear ->lock_depth so that schedule() doesnt
3990 * auto-release the semaphore:
3992 saved_lock_depth
= task
->lock_depth
;
3993 task
->lock_depth
= -1;
3996 local_irq_disable();
3997 task
->lock_depth
= saved_lock_depth
;
3998 sub_preempt_count(PREEMPT_ACTIVE
);
4001 * Check again in case we missed a preemption opportunity
4002 * between schedule and now.
4005 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4008 #endif /* CONFIG_PREEMPT */
4010 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4013 return try_to_wake_up(curr
->private, mode
, sync
);
4015 EXPORT_SYMBOL(default_wake_function
);
4018 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4019 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4020 * number) then we wake all the non-exclusive tasks and one exclusive task.
4022 * There are circumstances in which we can try to wake a task which has already
4023 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4024 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4026 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4027 int nr_exclusive
, int sync
, void *key
)
4029 wait_queue_t
*curr
, *next
;
4031 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4032 unsigned flags
= curr
->flags
;
4034 if (curr
->func(curr
, mode
, sync
, key
) &&
4035 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4041 * __wake_up - wake up threads blocked on a waitqueue.
4043 * @mode: which threads
4044 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4045 * @key: is directly passed to the wakeup function
4047 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4048 int nr_exclusive
, void *key
)
4050 unsigned long flags
;
4052 spin_lock_irqsave(&q
->lock
, flags
);
4053 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4054 spin_unlock_irqrestore(&q
->lock
, flags
);
4056 EXPORT_SYMBOL(__wake_up
);
4059 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4061 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4063 __wake_up_common(q
, mode
, 1, 0, NULL
);
4067 * __wake_up_sync - wake up threads blocked on a waitqueue.
4069 * @mode: which threads
4070 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4072 * The sync wakeup differs that the waker knows that it will schedule
4073 * away soon, so while the target thread will be woken up, it will not
4074 * be migrated to another CPU - ie. the two threads are 'synchronized'
4075 * with each other. This can prevent needless bouncing between CPUs.
4077 * On UP it can prevent extra preemption.
4080 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4082 unsigned long flags
;
4088 if (unlikely(!nr_exclusive
))
4091 spin_lock_irqsave(&q
->lock
, flags
);
4092 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4093 spin_unlock_irqrestore(&q
->lock
, flags
);
4095 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4097 void complete(struct completion
*x
)
4099 unsigned long flags
;
4101 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4103 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4104 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4106 EXPORT_SYMBOL(complete
);
4108 void complete_all(struct completion
*x
)
4110 unsigned long flags
;
4112 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4113 x
->done
+= UINT_MAX
/2;
4114 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4115 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4117 EXPORT_SYMBOL(complete_all
);
4119 static inline long __sched
4120 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4123 DECLARE_WAITQUEUE(wait
, current
);
4125 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4126 __add_wait_queue_tail(&x
->wait
, &wait
);
4128 if ((state
== TASK_INTERRUPTIBLE
&&
4129 signal_pending(current
)) ||
4130 (state
== TASK_KILLABLE
&&
4131 fatal_signal_pending(current
))) {
4132 __remove_wait_queue(&x
->wait
, &wait
);
4133 return -ERESTARTSYS
;
4135 __set_current_state(state
);
4136 spin_unlock_irq(&x
->wait
.lock
);
4137 timeout
= schedule_timeout(timeout
);
4138 spin_lock_irq(&x
->wait
.lock
);
4140 __remove_wait_queue(&x
->wait
, &wait
);
4144 __remove_wait_queue(&x
->wait
, &wait
);
4151 wait_for_common(struct completion
*x
, long timeout
, int state
)
4155 spin_lock_irq(&x
->wait
.lock
);
4156 timeout
= do_wait_for_common(x
, timeout
, state
);
4157 spin_unlock_irq(&x
->wait
.lock
);
4161 void __sched
wait_for_completion(struct completion
*x
)
4163 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4165 EXPORT_SYMBOL(wait_for_completion
);
4167 unsigned long __sched
4168 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4170 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4172 EXPORT_SYMBOL(wait_for_completion_timeout
);
4174 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4176 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4177 if (t
== -ERESTARTSYS
)
4181 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4183 unsigned long __sched
4184 wait_for_completion_interruptible_timeout(struct completion
*x
,
4185 unsigned long timeout
)
4187 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4189 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4191 int __sched
wait_for_completion_killable(struct completion
*x
)
4193 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4194 if (t
== -ERESTARTSYS
)
4198 EXPORT_SYMBOL(wait_for_completion_killable
);
4201 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4203 unsigned long flags
;
4206 init_waitqueue_entry(&wait
, current
);
4208 __set_current_state(state
);
4210 spin_lock_irqsave(&q
->lock
, flags
);
4211 __add_wait_queue(q
, &wait
);
4212 spin_unlock(&q
->lock
);
4213 timeout
= schedule_timeout(timeout
);
4214 spin_lock_irq(&q
->lock
);
4215 __remove_wait_queue(q
, &wait
);
4216 spin_unlock_irqrestore(&q
->lock
, flags
);
4221 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4223 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4225 EXPORT_SYMBOL(interruptible_sleep_on
);
4228 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4230 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4232 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4234 void __sched
sleep_on(wait_queue_head_t
*q
)
4236 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4238 EXPORT_SYMBOL(sleep_on
);
4240 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4242 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4244 EXPORT_SYMBOL(sleep_on_timeout
);
4246 #ifdef CONFIG_RT_MUTEXES
4249 * rt_mutex_setprio - set the current priority of a task
4251 * @prio: prio value (kernel-internal form)
4253 * This function changes the 'effective' priority of a task. It does
4254 * not touch ->normal_prio like __setscheduler().
4256 * Used by the rt_mutex code to implement priority inheritance logic.
4258 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4260 unsigned long flags
;
4261 int oldprio
, on_rq
, running
;
4263 const struct sched_class
*prev_class
= p
->sched_class
;
4265 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4267 rq
= task_rq_lock(p
, &flags
);
4268 update_rq_clock(rq
);
4271 on_rq
= p
->se
.on_rq
;
4272 running
= task_current(rq
, p
);
4274 dequeue_task(rq
, p
, 0);
4276 p
->sched_class
->put_prev_task(rq
, p
);
4279 p
->sched_class
= &rt_sched_class
;
4281 p
->sched_class
= &fair_sched_class
;
4286 p
->sched_class
->set_curr_task(rq
);
4288 enqueue_task(rq
, p
, 0);
4290 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4292 task_rq_unlock(rq
, &flags
);
4297 void set_user_nice(struct task_struct
*p
, long nice
)
4299 int old_prio
, delta
, on_rq
;
4300 unsigned long flags
;
4303 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4306 * We have to be careful, if called from sys_setpriority(),
4307 * the task might be in the middle of scheduling on another CPU.
4309 rq
= task_rq_lock(p
, &flags
);
4310 update_rq_clock(rq
);
4312 * The RT priorities are set via sched_setscheduler(), but we still
4313 * allow the 'normal' nice value to be set - but as expected
4314 * it wont have any effect on scheduling until the task is
4315 * SCHED_FIFO/SCHED_RR:
4317 if (task_has_rt_policy(p
)) {
4318 p
->static_prio
= NICE_TO_PRIO(nice
);
4321 on_rq
= p
->se
.on_rq
;
4323 dequeue_task(rq
, p
, 0);
4327 p
->static_prio
= NICE_TO_PRIO(nice
);
4330 p
->prio
= effective_prio(p
);
4331 delta
= p
->prio
- old_prio
;
4334 enqueue_task(rq
, p
, 0);
4337 * If the task increased its priority or is running and
4338 * lowered its priority, then reschedule its CPU:
4340 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4341 resched_task(rq
->curr
);
4344 task_rq_unlock(rq
, &flags
);
4346 EXPORT_SYMBOL(set_user_nice
);
4349 * can_nice - check if a task can reduce its nice value
4353 int can_nice(const struct task_struct
*p
, const int nice
)
4355 /* convert nice value [19,-20] to rlimit style value [1,40] */
4356 int nice_rlim
= 20 - nice
;
4358 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4359 capable(CAP_SYS_NICE
));
4362 #ifdef __ARCH_WANT_SYS_NICE
4365 * sys_nice - change the priority of the current process.
4366 * @increment: priority increment
4368 * sys_setpriority is a more generic, but much slower function that
4369 * does similar things.
4371 asmlinkage
long sys_nice(int increment
)
4376 * Setpriority might change our priority at the same moment.
4377 * We don't have to worry. Conceptually one call occurs first
4378 * and we have a single winner.
4380 if (increment
< -40)
4385 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4391 if (increment
< 0 && !can_nice(current
, nice
))
4394 retval
= security_task_setnice(current
, nice
);
4398 set_user_nice(current
, nice
);
4405 * task_prio - return the priority value of a given task.
4406 * @p: the task in question.
4408 * This is the priority value as seen by users in /proc.
4409 * RT tasks are offset by -200. Normal tasks are centered
4410 * around 0, value goes from -16 to +15.
4412 int task_prio(const struct task_struct
*p
)
4414 return p
->prio
- MAX_RT_PRIO
;
4418 * task_nice - return the nice value of a given task.
4419 * @p: the task in question.
4421 int task_nice(const struct task_struct
*p
)
4423 return TASK_NICE(p
);
4425 EXPORT_SYMBOL(task_nice
);
4428 * idle_cpu - is a given cpu idle currently?
4429 * @cpu: the processor in question.
4431 int idle_cpu(int cpu
)
4433 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4437 * idle_task - return the idle task for a given cpu.
4438 * @cpu: the processor in question.
4440 struct task_struct
*idle_task(int cpu
)
4442 return cpu_rq(cpu
)->idle
;
4446 * find_process_by_pid - find a process with a matching PID value.
4447 * @pid: the pid in question.
4449 static struct task_struct
*find_process_by_pid(pid_t pid
)
4451 return pid
? find_task_by_vpid(pid
) : current
;
4454 /* Actually do priority change: must hold rq lock. */
4456 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4458 BUG_ON(p
->se
.on_rq
);
4461 switch (p
->policy
) {
4465 p
->sched_class
= &fair_sched_class
;
4469 p
->sched_class
= &rt_sched_class
;
4473 p
->rt_priority
= prio
;
4474 p
->normal_prio
= normal_prio(p
);
4475 /* we are holding p->pi_lock already */
4476 p
->prio
= rt_mutex_getprio(p
);
4481 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4482 * @p: the task in question.
4483 * @policy: new policy.
4484 * @param: structure containing the new RT priority.
4486 * NOTE that the task may be already dead.
4488 int sched_setscheduler(struct task_struct
*p
, int policy
,
4489 struct sched_param
*param
)
4491 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4492 unsigned long flags
;
4493 const struct sched_class
*prev_class
= p
->sched_class
;
4496 /* may grab non-irq protected spin_locks */
4497 BUG_ON(in_interrupt());
4499 /* double check policy once rq lock held */
4501 policy
= oldpolicy
= p
->policy
;
4502 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4503 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4504 policy
!= SCHED_IDLE
)
4507 * Valid priorities for SCHED_FIFO and SCHED_RR are
4508 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4509 * SCHED_BATCH and SCHED_IDLE is 0.
4511 if (param
->sched_priority
< 0 ||
4512 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4513 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4515 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4519 * Allow unprivileged RT tasks to decrease priority:
4521 if (!capable(CAP_SYS_NICE
)) {
4522 if (rt_policy(policy
)) {
4523 unsigned long rlim_rtprio
;
4525 if (!lock_task_sighand(p
, &flags
))
4527 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4528 unlock_task_sighand(p
, &flags
);
4530 /* can't set/change the rt policy */
4531 if (policy
!= p
->policy
&& !rlim_rtprio
)
4534 /* can't increase priority */
4535 if (param
->sched_priority
> p
->rt_priority
&&
4536 param
->sched_priority
> rlim_rtprio
)
4540 * Like positive nice levels, dont allow tasks to
4541 * move out of SCHED_IDLE either:
4543 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4546 /* can't change other user's priorities */
4547 if ((current
->euid
!= p
->euid
) &&
4548 (current
->euid
!= p
->uid
))
4552 #ifdef CONFIG_RT_GROUP_SCHED
4554 * Do not allow realtime tasks into groups that have no runtime
4557 if (rt_policy(policy
) && task_group(p
)->rt_runtime
== 0)
4561 retval
= security_task_setscheduler(p
, policy
, param
);
4565 * make sure no PI-waiters arrive (or leave) while we are
4566 * changing the priority of the task:
4568 spin_lock_irqsave(&p
->pi_lock
, flags
);
4570 * To be able to change p->policy safely, the apropriate
4571 * runqueue lock must be held.
4573 rq
= __task_rq_lock(p
);
4574 /* recheck policy now with rq lock held */
4575 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4576 policy
= oldpolicy
= -1;
4577 __task_rq_unlock(rq
);
4578 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4581 update_rq_clock(rq
);
4582 on_rq
= p
->se
.on_rq
;
4583 running
= task_current(rq
, p
);
4585 deactivate_task(rq
, p
, 0);
4587 p
->sched_class
->put_prev_task(rq
, p
);
4590 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4593 p
->sched_class
->set_curr_task(rq
);
4595 activate_task(rq
, p
, 0);
4597 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4599 __task_rq_unlock(rq
);
4600 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4602 rt_mutex_adjust_pi(p
);
4606 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4609 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4611 struct sched_param lparam
;
4612 struct task_struct
*p
;
4615 if (!param
|| pid
< 0)
4617 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4622 p
= find_process_by_pid(pid
);
4624 retval
= sched_setscheduler(p
, policy
, &lparam
);
4631 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4632 * @pid: the pid in question.
4633 * @policy: new policy.
4634 * @param: structure containing the new RT priority.
4637 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4639 /* negative values for policy are not valid */
4643 return do_sched_setscheduler(pid
, policy
, param
);
4647 * sys_sched_setparam - set/change the RT priority of a thread
4648 * @pid: the pid in question.
4649 * @param: structure containing the new RT priority.
4651 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4653 return do_sched_setscheduler(pid
, -1, param
);
4657 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4658 * @pid: the pid in question.
4660 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4662 struct task_struct
*p
;
4669 read_lock(&tasklist_lock
);
4670 p
= find_process_by_pid(pid
);
4672 retval
= security_task_getscheduler(p
);
4676 read_unlock(&tasklist_lock
);
4681 * sys_sched_getscheduler - get the RT priority of a thread
4682 * @pid: the pid in question.
4683 * @param: structure containing the RT priority.
4685 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4687 struct sched_param lp
;
4688 struct task_struct
*p
;
4691 if (!param
|| pid
< 0)
4694 read_lock(&tasklist_lock
);
4695 p
= find_process_by_pid(pid
);
4700 retval
= security_task_getscheduler(p
);
4704 lp
.sched_priority
= p
->rt_priority
;
4705 read_unlock(&tasklist_lock
);
4708 * This one might sleep, we cannot do it with a spinlock held ...
4710 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4715 read_unlock(&tasklist_lock
);
4719 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4721 cpumask_t cpus_allowed
;
4722 struct task_struct
*p
;
4726 read_lock(&tasklist_lock
);
4728 p
= find_process_by_pid(pid
);
4730 read_unlock(&tasklist_lock
);
4736 * It is not safe to call set_cpus_allowed with the
4737 * tasklist_lock held. We will bump the task_struct's
4738 * usage count and then drop tasklist_lock.
4741 read_unlock(&tasklist_lock
);
4744 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4745 !capable(CAP_SYS_NICE
))
4748 retval
= security_task_setscheduler(p
, 0, NULL
);
4752 cpus_allowed
= cpuset_cpus_allowed(p
);
4753 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4755 retval
= set_cpus_allowed(p
, new_mask
);
4758 cpus_allowed
= cpuset_cpus_allowed(p
);
4759 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4761 * We must have raced with a concurrent cpuset
4762 * update. Just reset the cpus_allowed to the
4763 * cpuset's cpus_allowed
4765 new_mask
= cpus_allowed
;
4775 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4776 cpumask_t
*new_mask
)
4778 if (len
< sizeof(cpumask_t
)) {
4779 memset(new_mask
, 0, sizeof(cpumask_t
));
4780 } else if (len
> sizeof(cpumask_t
)) {
4781 len
= sizeof(cpumask_t
);
4783 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4787 * sys_sched_setaffinity - set the cpu affinity of a process
4788 * @pid: pid of the process
4789 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4790 * @user_mask_ptr: user-space pointer to the new cpu mask
4792 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4793 unsigned long __user
*user_mask_ptr
)
4798 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4802 return sched_setaffinity(pid
, new_mask
);
4806 * Represents all cpu's present in the system
4807 * In systems capable of hotplug, this map could dynamically grow
4808 * as new cpu's are detected in the system via any platform specific
4809 * method, such as ACPI for e.g.
4812 cpumask_t cpu_present_map __read_mostly
;
4813 EXPORT_SYMBOL(cpu_present_map
);
4816 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4817 EXPORT_SYMBOL(cpu_online_map
);
4819 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4820 EXPORT_SYMBOL(cpu_possible_map
);
4823 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4825 struct task_struct
*p
;
4829 read_lock(&tasklist_lock
);
4832 p
= find_process_by_pid(pid
);
4836 retval
= security_task_getscheduler(p
);
4840 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4843 read_unlock(&tasklist_lock
);
4850 * sys_sched_getaffinity - get the cpu affinity of a process
4851 * @pid: pid of the process
4852 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4853 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4855 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4856 unsigned long __user
*user_mask_ptr
)
4861 if (len
< sizeof(cpumask_t
))
4864 ret
= sched_getaffinity(pid
, &mask
);
4868 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4871 return sizeof(cpumask_t
);
4875 * sys_sched_yield - yield the current processor to other threads.
4877 * This function yields the current CPU to other tasks. If there are no
4878 * other threads running on this CPU then this function will return.
4880 asmlinkage
long sys_sched_yield(void)
4882 struct rq
*rq
= this_rq_lock();
4884 schedstat_inc(rq
, yld_count
);
4885 current
->sched_class
->yield_task(rq
);
4888 * Since we are going to call schedule() anyway, there's
4889 * no need to preempt or enable interrupts:
4891 __release(rq
->lock
);
4892 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4893 _raw_spin_unlock(&rq
->lock
);
4894 preempt_enable_no_resched();
4901 static void __cond_resched(void)
4903 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4904 __might_sleep(__FILE__
, __LINE__
);
4907 * The BKS might be reacquired before we have dropped
4908 * PREEMPT_ACTIVE, which could trigger a second
4909 * cond_resched() call.
4912 add_preempt_count(PREEMPT_ACTIVE
);
4914 sub_preempt_count(PREEMPT_ACTIVE
);
4915 } while (need_resched());
4918 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4919 int __sched
_cond_resched(void)
4921 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4922 system_state
== SYSTEM_RUNNING
) {
4928 EXPORT_SYMBOL(_cond_resched
);
4932 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4933 * call schedule, and on return reacquire the lock.
4935 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4936 * operations here to prevent schedule() from being called twice (once via
4937 * spin_unlock(), once by hand).
4939 int cond_resched_lock(spinlock_t
*lock
)
4941 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
4944 if (spin_needbreak(lock
) || resched
) {
4946 if (resched
&& need_resched())
4955 EXPORT_SYMBOL(cond_resched_lock
);
4957 int __sched
cond_resched_softirq(void)
4959 BUG_ON(!in_softirq());
4961 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4969 EXPORT_SYMBOL(cond_resched_softirq
);
4972 * yield - yield the current processor to other threads.
4974 * This is a shortcut for kernel-space yielding - it marks the
4975 * thread runnable and calls sys_sched_yield().
4977 void __sched
yield(void)
4979 set_current_state(TASK_RUNNING
);
4982 EXPORT_SYMBOL(yield
);
4985 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4986 * that process accounting knows that this is a task in IO wait state.
4988 * But don't do that if it is a deliberate, throttling IO wait (this task
4989 * has set its backing_dev_info: the queue against which it should throttle)
4991 void __sched
io_schedule(void)
4993 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4995 delayacct_blkio_start();
4996 atomic_inc(&rq
->nr_iowait
);
4998 atomic_dec(&rq
->nr_iowait
);
4999 delayacct_blkio_end();
5001 EXPORT_SYMBOL(io_schedule
);
5003 long __sched
io_schedule_timeout(long timeout
)
5005 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5008 delayacct_blkio_start();
5009 atomic_inc(&rq
->nr_iowait
);
5010 ret
= schedule_timeout(timeout
);
5011 atomic_dec(&rq
->nr_iowait
);
5012 delayacct_blkio_end();
5017 * sys_sched_get_priority_max - return maximum RT priority.
5018 * @policy: scheduling class.
5020 * this syscall returns the maximum rt_priority that can be used
5021 * by a given scheduling class.
5023 asmlinkage
long sys_sched_get_priority_max(int policy
)
5030 ret
= MAX_USER_RT_PRIO
-1;
5042 * sys_sched_get_priority_min - return minimum RT priority.
5043 * @policy: scheduling class.
5045 * this syscall returns the minimum rt_priority that can be used
5046 * by a given scheduling class.
5048 asmlinkage
long sys_sched_get_priority_min(int policy
)
5066 * sys_sched_rr_get_interval - return the default timeslice of a process.
5067 * @pid: pid of the process.
5068 * @interval: userspace pointer to the timeslice value.
5070 * this syscall writes the default timeslice value of a given process
5071 * into the user-space timespec buffer. A value of '0' means infinity.
5074 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5076 struct task_struct
*p
;
5077 unsigned int time_slice
;
5085 read_lock(&tasklist_lock
);
5086 p
= find_process_by_pid(pid
);
5090 retval
= security_task_getscheduler(p
);
5095 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5096 * tasks that are on an otherwise idle runqueue:
5099 if (p
->policy
== SCHED_RR
) {
5100 time_slice
= DEF_TIMESLICE
;
5101 } else if (p
->policy
!= SCHED_FIFO
) {
5102 struct sched_entity
*se
= &p
->se
;
5103 unsigned long flags
;
5106 rq
= task_rq_lock(p
, &flags
);
5107 if (rq
->cfs
.load
.weight
)
5108 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5109 task_rq_unlock(rq
, &flags
);
5111 read_unlock(&tasklist_lock
);
5112 jiffies_to_timespec(time_slice
, &t
);
5113 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5117 read_unlock(&tasklist_lock
);
5121 static const char stat_nam
[] = "RSDTtZX";
5123 void sched_show_task(struct task_struct
*p
)
5125 unsigned long free
= 0;
5128 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5129 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5130 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5131 #if BITS_PER_LONG == 32
5132 if (state
== TASK_RUNNING
)
5133 printk(KERN_CONT
" running ");
5135 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5137 if (state
== TASK_RUNNING
)
5138 printk(KERN_CONT
" running task ");
5140 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5142 #ifdef CONFIG_DEBUG_STACK_USAGE
5144 unsigned long *n
= end_of_stack(p
);
5147 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5150 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5151 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5153 show_stack(p
, NULL
);
5156 void show_state_filter(unsigned long state_filter
)
5158 struct task_struct
*g
, *p
;
5160 #if BITS_PER_LONG == 32
5162 " task PC stack pid father\n");
5165 " task PC stack pid father\n");
5167 read_lock(&tasklist_lock
);
5168 do_each_thread(g
, p
) {
5170 * reset the NMI-timeout, listing all files on a slow
5171 * console might take alot of time:
5173 touch_nmi_watchdog();
5174 if (!state_filter
|| (p
->state
& state_filter
))
5176 } while_each_thread(g
, p
);
5178 touch_all_softlockup_watchdogs();
5180 #ifdef CONFIG_SCHED_DEBUG
5181 sysrq_sched_debug_show();
5183 read_unlock(&tasklist_lock
);
5185 * Only show locks if all tasks are dumped:
5187 if (state_filter
== -1)
5188 debug_show_all_locks();
5191 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5193 idle
->sched_class
= &idle_sched_class
;
5197 * init_idle - set up an idle thread for a given CPU
5198 * @idle: task in question
5199 * @cpu: cpu the idle task belongs to
5201 * NOTE: this function does not set the idle thread's NEED_RESCHED
5202 * flag, to make booting more robust.
5204 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5206 struct rq
*rq
= cpu_rq(cpu
);
5207 unsigned long flags
;
5210 idle
->se
.exec_start
= sched_clock();
5212 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5213 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5214 __set_task_cpu(idle
, cpu
);
5216 spin_lock_irqsave(&rq
->lock
, flags
);
5217 rq
->curr
= rq
->idle
= idle
;
5218 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5221 spin_unlock_irqrestore(&rq
->lock
, flags
);
5223 /* Set the preempt count _outside_ the spinlocks! */
5224 task_thread_info(idle
)->preempt_count
= 0;
5227 * The idle tasks have their own, simple scheduling class:
5229 idle
->sched_class
= &idle_sched_class
;
5233 * In a system that switches off the HZ timer nohz_cpu_mask
5234 * indicates which cpus entered this state. This is used
5235 * in the rcu update to wait only for active cpus. For system
5236 * which do not switch off the HZ timer nohz_cpu_mask should
5237 * always be CPU_MASK_NONE.
5239 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5242 * Increase the granularity value when there are more CPUs,
5243 * because with more CPUs the 'effective latency' as visible
5244 * to users decreases. But the relationship is not linear,
5245 * so pick a second-best guess by going with the log2 of the
5248 * This idea comes from the SD scheduler of Con Kolivas:
5250 static inline void sched_init_granularity(void)
5252 unsigned int factor
= 1 + ilog2(num_online_cpus());
5253 const unsigned long limit
= 200000000;
5255 sysctl_sched_min_granularity
*= factor
;
5256 if (sysctl_sched_min_granularity
> limit
)
5257 sysctl_sched_min_granularity
= limit
;
5259 sysctl_sched_latency
*= factor
;
5260 if (sysctl_sched_latency
> limit
)
5261 sysctl_sched_latency
= limit
;
5263 sysctl_sched_wakeup_granularity
*= factor
;
5264 sysctl_sched_batch_wakeup_granularity
*= factor
;
5269 * This is how migration works:
5271 * 1) we queue a struct migration_req structure in the source CPU's
5272 * runqueue and wake up that CPU's migration thread.
5273 * 2) we down() the locked semaphore => thread blocks.
5274 * 3) migration thread wakes up (implicitly it forces the migrated
5275 * thread off the CPU)
5276 * 4) it gets the migration request and checks whether the migrated
5277 * task is still in the wrong runqueue.
5278 * 5) if it's in the wrong runqueue then the migration thread removes
5279 * it and puts it into the right queue.
5280 * 6) migration thread up()s the semaphore.
5281 * 7) we wake up and the migration is done.
5285 * Change a given task's CPU affinity. Migrate the thread to a
5286 * proper CPU and schedule it away if the CPU it's executing on
5287 * is removed from the allowed bitmask.
5289 * NOTE: the caller must have a valid reference to the task, the
5290 * task must not exit() & deallocate itself prematurely. The
5291 * call is not atomic; no spinlocks may be held.
5293 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5295 struct migration_req req
;
5296 unsigned long flags
;
5300 rq
= task_rq_lock(p
, &flags
);
5301 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5306 if (p
->sched_class
->set_cpus_allowed
)
5307 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5309 p
->cpus_allowed
= new_mask
;
5310 p
->rt
.nr_cpus_allowed
= cpus_weight(new_mask
);
5313 /* Can the task run on the task's current CPU? If so, we're done */
5314 if (cpu_isset(task_cpu(p
), new_mask
))
5317 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5318 /* Need help from migration thread: drop lock and wait. */
5319 task_rq_unlock(rq
, &flags
);
5320 wake_up_process(rq
->migration_thread
);
5321 wait_for_completion(&req
.done
);
5322 tlb_migrate_finish(p
->mm
);
5326 task_rq_unlock(rq
, &flags
);
5330 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5333 * Move (not current) task off this cpu, onto dest cpu. We're doing
5334 * this because either it can't run here any more (set_cpus_allowed()
5335 * away from this CPU, or CPU going down), or because we're
5336 * attempting to rebalance this task on exec (sched_exec).
5338 * So we race with normal scheduler movements, but that's OK, as long
5339 * as the task is no longer on this CPU.
5341 * Returns non-zero if task was successfully migrated.
5343 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5345 struct rq
*rq_dest
, *rq_src
;
5348 if (unlikely(cpu_is_offline(dest_cpu
)))
5351 rq_src
= cpu_rq(src_cpu
);
5352 rq_dest
= cpu_rq(dest_cpu
);
5354 double_rq_lock(rq_src
, rq_dest
);
5355 /* Already moved. */
5356 if (task_cpu(p
) != src_cpu
)
5358 /* Affinity changed (again). */
5359 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5362 on_rq
= p
->se
.on_rq
;
5364 deactivate_task(rq_src
, p
, 0);
5366 set_task_cpu(p
, dest_cpu
);
5368 activate_task(rq_dest
, p
, 0);
5369 check_preempt_curr(rq_dest
, p
);
5373 double_rq_unlock(rq_src
, rq_dest
);
5378 * migration_thread - this is a highprio system thread that performs
5379 * thread migration by bumping thread off CPU then 'pushing' onto
5382 static int migration_thread(void *data
)
5384 int cpu
= (long)data
;
5388 BUG_ON(rq
->migration_thread
!= current
);
5390 set_current_state(TASK_INTERRUPTIBLE
);
5391 while (!kthread_should_stop()) {
5392 struct migration_req
*req
;
5393 struct list_head
*head
;
5395 spin_lock_irq(&rq
->lock
);
5397 if (cpu_is_offline(cpu
)) {
5398 spin_unlock_irq(&rq
->lock
);
5402 if (rq
->active_balance
) {
5403 active_load_balance(rq
, cpu
);
5404 rq
->active_balance
= 0;
5407 head
= &rq
->migration_queue
;
5409 if (list_empty(head
)) {
5410 spin_unlock_irq(&rq
->lock
);
5412 set_current_state(TASK_INTERRUPTIBLE
);
5415 req
= list_entry(head
->next
, struct migration_req
, list
);
5416 list_del_init(head
->next
);
5418 spin_unlock(&rq
->lock
);
5419 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5422 complete(&req
->done
);
5424 __set_current_state(TASK_RUNNING
);
5428 /* Wait for kthread_stop */
5429 set_current_state(TASK_INTERRUPTIBLE
);
5430 while (!kthread_should_stop()) {
5432 set_current_state(TASK_INTERRUPTIBLE
);
5434 __set_current_state(TASK_RUNNING
);
5438 #ifdef CONFIG_HOTPLUG_CPU
5440 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5444 local_irq_disable();
5445 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5451 * Figure out where task on dead CPU should go, use force if necessary.
5452 * NOTE: interrupts should be disabled by the caller
5454 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5456 unsigned long flags
;
5463 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5464 cpus_and(mask
, mask
, p
->cpus_allowed
);
5465 dest_cpu
= any_online_cpu(mask
);
5467 /* On any allowed CPU? */
5468 if (dest_cpu
== NR_CPUS
)
5469 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5471 /* No more Mr. Nice Guy. */
5472 if (dest_cpu
== NR_CPUS
) {
5473 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5475 * Try to stay on the same cpuset, where the
5476 * current cpuset may be a subset of all cpus.
5477 * The cpuset_cpus_allowed_locked() variant of
5478 * cpuset_cpus_allowed() will not block. It must be
5479 * called within calls to cpuset_lock/cpuset_unlock.
5481 rq
= task_rq_lock(p
, &flags
);
5482 p
->cpus_allowed
= cpus_allowed
;
5483 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5484 task_rq_unlock(rq
, &flags
);
5487 * Don't tell them about moving exiting tasks or
5488 * kernel threads (both mm NULL), since they never
5491 if (p
->mm
&& printk_ratelimit()) {
5492 printk(KERN_INFO
"process %d (%s) no "
5493 "longer affine to cpu%d\n",
5494 task_pid_nr(p
), p
->comm
, dead_cpu
);
5497 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5501 * While a dead CPU has no uninterruptible tasks queued at this point,
5502 * it might still have a nonzero ->nr_uninterruptible counter, because
5503 * for performance reasons the counter is not stricly tracking tasks to
5504 * their home CPUs. So we just add the counter to another CPU's counter,
5505 * to keep the global sum constant after CPU-down:
5507 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5509 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5510 unsigned long flags
;
5512 local_irq_save(flags
);
5513 double_rq_lock(rq_src
, rq_dest
);
5514 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5515 rq_src
->nr_uninterruptible
= 0;
5516 double_rq_unlock(rq_src
, rq_dest
);
5517 local_irq_restore(flags
);
5520 /* Run through task list and migrate tasks from the dead cpu. */
5521 static void migrate_live_tasks(int src_cpu
)
5523 struct task_struct
*p
, *t
;
5525 read_lock(&tasklist_lock
);
5527 do_each_thread(t
, p
) {
5531 if (task_cpu(p
) == src_cpu
)
5532 move_task_off_dead_cpu(src_cpu
, p
);
5533 } while_each_thread(t
, p
);
5535 read_unlock(&tasklist_lock
);
5539 * Schedules idle task to be the next runnable task on current CPU.
5540 * It does so by boosting its priority to highest possible.
5541 * Used by CPU offline code.
5543 void sched_idle_next(void)
5545 int this_cpu
= smp_processor_id();
5546 struct rq
*rq
= cpu_rq(this_cpu
);
5547 struct task_struct
*p
= rq
->idle
;
5548 unsigned long flags
;
5550 /* cpu has to be offline */
5551 BUG_ON(cpu_online(this_cpu
));
5554 * Strictly not necessary since rest of the CPUs are stopped by now
5555 * and interrupts disabled on the current cpu.
5557 spin_lock_irqsave(&rq
->lock
, flags
);
5559 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5561 update_rq_clock(rq
);
5562 activate_task(rq
, p
, 0);
5564 spin_unlock_irqrestore(&rq
->lock
, flags
);
5568 * Ensures that the idle task is using init_mm right before its cpu goes
5571 void idle_task_exit(void)
5573 struct mm_struct
*mm
= current
->active_mm
;
5575 BUG_ON(cpu_online(smp_processor_id()));
5578 switch_mm(mm
, &init_mm
, current
);
5582 /* called under rq->lock with disabled interrupts */
5583 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5585 struct rq
*rq
= cpu_rq(dead_cpu
);
5587 /* Must be exiting, otherwise would be on tasklist. */
5588 BUG_ON(!p
->exit_state
);
5590 /* Cannot have done final schedule yet: would have vanished. */
5591 BUG_ON(p
->state
== TASK_DEAD
);
5596 * Drop lock around migration; if someone else moves it,
5597 * that's OK. No task can be added to this CPU, so iteration is
5600 spin_unlock_irq(&rq
->lock
);
5601 move_task_off_dead_cpu(dead_cpu
, p
);
5602 spin_lock_irq(&rq
->lock
);
5607 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5608 static void migrate_dead_tasks(unsigned int dead_cpu
)
5610 struct rq
*rq
= cpu_rq(dead_cpu
);
5611 struct task_struct
*next
;
5614 if (!rq
->nr_running
)
5616 update_rq_clock(rq
);
5617 next
= pick_next_task(rq
, rq
->curr
);
5620 migrate_dead(dead_cpu
, next
);
5624 #endif /* CONFIG_HOTPLUG_CPU */
5626 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5628 static struct ctl_table sd_ctl_dir
[] = {
5630 .procname
= "sched_domain",
5636 static struct ctl_table sd_ctl_root
[] = {
5638 .ctl_name
= CTL_KERN
,
5639 .procname
= "kernel",
5641 .child
= sd_ctl_dir
,
5646 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5648 struct ctl_table
*entry
=
5649 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5654 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5656 struct ctl_table
*entry
;
5659 * In the intermediate directories, both the child directory and
5660 * procname are dynamically allocated and could fail but the mode
5661 * will always be set. In the lowest directory the names are
5662 * static strings and all have proc handlers.
5664 for (entry
= *tablep
; entry
->mode
; entry
++) {
5666 sd_free_ctl_entry(&entry
->child
);
5667 if (entry
->proc_handler
== NULL
)
5668 kfree(entry
->procname
);
5676 set_table_entry(struct ctl_table
*entry
,
5677 const char *procname
, void *data
, int maxlen
,
5678 mode_t mode
, proc_handler
*proc_handler
)
5680 entry
->procname
= procname
;
5682 entry
->maxlen
= maxlen
;
5684 entry
->proc_handler
= proc_handler
;
5687 static struct ctl_table
*
5688 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5690 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5695 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5696 sizeof(long), 0644, proc_doulongvec_minmax
);
5697 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5698 sizeof(long), 0644, proc_doulongvec_minmax
);
5699 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5700 sizeof(int), 0644, proc_dointvec_minmax
);
5701 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5702 sizeof(int), 0644, proc_dointvec_minmax
);
5703 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5704 sizeof(int), 0644, proc_dointvec_minmax
);
5705 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5706 sizeof(int), 0644, proc_dointvec_minmax
);
5707 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5708 sizeof(int), 0644, proc_dointvec_minmax
);
5709 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5710 sizeof(int), 0644, proc_dointvec_minmax
);
5711 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5712 sizeof(int), 0644, proc_dointvec_minmax
);
5713 set_table_entry(&table
[9], "cache_nice_tries",
5714 &sd
->cache_nice_tries
,
5715 sizeof(int), 0644, proc_dointvec_minmax
);
5716 set_table_entry(&table
[10], "flags", &sd
->flags
,
5717 sizeof(int), 0644, proc_dointvec_minmax
);
5718 /* &table[11] is terminator */
5723 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5725 struct ctl_table
*entry
, *table
;
5726 struct sched_domain
*sd
;
5727 int domain_num
= 0, i
;
5730 for_each_domain(cpu
, sd
)
5732 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5737 for_each_domain(cpu
, sd
) {
5738 snprintf(buf
, 32, "domain%d", i
);
5739 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5741 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5748 static struct ctl_table_header
*sd_sysctl_header
;
5749 static void register_sched_domain_sysctl(void)
5751 int i
, cpu_num
= num_online_cpus();
5752 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5755 WARN_ON(sd_ctl_dir
[0].child
);
5756 sd_ctl_dir
[0].child
= entry
;
5761 for_each_online_cpu(i
) {
5762 snprintf(buf
, 32, "cpu%d", i
);
5763 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5765 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5769 WARN_ON(sd_sysctl_header
);
5770 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5773 /* may be called multiple times per register */
5774 static void unregister_sched_domain_sysctl(void)
5776 if (sd_sysctl_header
)
5777 unregister_sysctl_table(sd_sysctl_header
);
5778 sd_sysctl_header
= NULL
;
5779 if (sd_ctl_dir
[0].child
)
5780 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5783 static void register_sched_domain_sysctl(void)
5786 static void unregister_sched_domain_sysctl(void)
5792 * migration_call - callback that gets triggered when a CPU is added.
5793 * Here we can start up the necessary migration thread for the new CPU.
5795 static int __cpuinit
5796 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5798 struct task_struct
*p
;
5799 int cpu
= (long)hcpu
;
5800 unsigned long flags
;
5805 case CPU_UP_PREPARE
:
5806 case CPU_UP_PREPARE_FROZEN
:
5807 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5810 kthread_bind(p
, cpu
);
5811 /* Must be high prio: stop_machine expects to yield to it. */
5812 rq
= task_rq_lock(p
, &flags
);
5813 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5814 task_rq_unlock(rq
, &flags
);
5815 cpu_rq(cpu
)->migration_thread
= p
;
5819 case CPU_ONLINE_FROZEN
:
5820 /* Strictly unnecessary, as first user will wake it. */
5821 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5823 /* Update our root-domain */
5825 spin_lock_irqsave(&rq
->lock
, flags
);
5827 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5828 cpu_set(cpu
, rq
->rd
->online
);
5830 spin_unlock_irqrestore(&rq
->lock
, flags
);
5833 #ifdef CONFIG_HOTPLUG_CPU
5834 case CPU_UP_CANCELED
:
5835 case CPU_UP_CANCELED_FROZEN
:
5836 if (!cpu_rq(cpu
)->migration_thread
)
5838 /* Unbind it from offline cpu so it can run. Fall thru. */
5839 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5840 any_online_cpu(cpu_online_map
));
5841 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5842 cpu_rq(cpu
)->migration_thread
= NULL
;
5846 case CPU_DEAD_FROZEN
:
5847 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5848 migrate_live_tasks(cpu
);
5850 kthread_stop(rq
->migration_thread
);
5851 rq
->migration_thread
= NULL
;
5852 /* Idle task back to normal (off runqueue, low prio) */
5853 spin_lock_irq(&rq
->lock
);
5854 update_rq_clock(rq
);
5855 deactivate_task(rq
, rq
->idle
, 0);
5856 rq
->idle
->static_prio
= MAX_PRIO
;
5857 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5858 rq
->idle
->sched_class
= &idle_sched_class
;
5859 migrate_dead_tasks(cpu
);
5860 spin_unlock_irq(&rq
->lock
);
5862 migrate_nr_uninterruptible(rq
);
5863 BUG_ON(rq
->nr_running
!= 0);
5866 * No need to migrate the tasks: it was best-effort if
5867 * they didn't take sched_hotcpu_mutex. Just wake up
5870 spin_lock_irq(&rq
->lock
);
5871 while (!list_empty(&rq
->migration_queue
)) {
5872 struct migration_req
*req
;
5874 req
= list_entry(rq
->migration_queue
.next
,
5875 struct migration_req
, list
);
5876 list_del_init(&req
->list
);
5877 complete(&req
->done
);
5879 spin_unlock_irq(&rq
->lock
);
5883 case CPU_DYING_FROZEN
:
5884 /* Update our root-domain */
5886 spin_lock_irqsave(&rq
->lock
, flags
);
5888 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5889 cpu_clear(cpu
, rq
->rd
->online
);
5891 spin_unlock_irqrestore(&rq
->lock
, flags
);
5898 /* Register at highest priority so that task migration (migrate_all_tasks)
5899 * happens before everything else.
5901 static struct notifier_block __cpuinitdata migration_notifier
= {
5902 .notifier_call
= migration_call
,
5906 void __init
migration_init(void)
5908 void *cpu
= (void *)(long)smp_processor_id();
5911 /* Start one for the boot CPU: */
5912 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5913 BUG_ON(err
== NOTIFY_BAD
);
5914 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5915 register_cpu_notifier(&migration_notifier
);
5921 /* Number of possible processor ids */
5922 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5923 EXPORT_SYMBOL(nr_cpu_ids
);
5925 #ifdef CONFIG_SCHED_DEBUG
5927 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5929 struct sched_group
*group
= sd
->groups
;
5930 cpumask_t groupmask
;
5933 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5934 cpus_clear(groupmask
);
5936 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5938 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5939 printk("does not load-balance\n");
5941 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5946 printk(KERN_CONT
"span %s\n", str
);
5948 if (!cpu_isset(cpu
, sd
->span
)) {
5949 printk(KERN_ERR
"ERROR: domain->span does not contain "
5952 if (!cpu_isset(cpu
, group
->cpumask
)) {
5953 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5957 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5961 printk(KERN_ERR
"ERROR: group is NULL\n");
5965 if (!group
->__cpu_power
) {
5966 printk(KERN_CONT
"\n");
5967 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5972 if (!cpus_weight(group
->cpumask
)) {
5973 printk(KERN_CONT
"\n");
5974 printk(KERN_ERR
"ERROR: empty group\n");
5978 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5979 printk(KERN_CONT
"\n");
5980 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5984 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5986 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5987 printk(KERN_CONT
" %s", str
);
5989 group
= group
->next
;
5990 } while (group
!= sd
->groups
);
5991 printk(KERN_CONT
"\n");
5993 if (!cpus_equal(sd
->span
, groupmask
))
5994 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5996 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
5997 printk(KERN_ERR
"ERROR: parent span is not a superset "
5998 "of domain->span\n");
6002 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6007 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6011 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6014 if (sched_domain_debug_one(sd
, cpu
, level
))
6023 # define sched_domain_debug(sd, cpu) do { } while (0)
6026 static int sd_degenerate(struct sched_domain
*sd
)
6028 if (cpus_weight(sd
->span
) == 1)
6031 /* Following flags need at least 2 groups */
6032 if (sd
->flags
& (SD_LOAD_BALANCE
|
6033 SD_BALANCE_NEWIDLE
|
6037 SD_SHARE_PKG_RESOURCES
)) {
6038 if (sd
->groups
!= sd
->groups
->next
)
6042 /* Following flags don't use groups */
6043 if (sd
->flags
& (SD_WAKE_IDLE
|
6052 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6054 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6056 if (sd_degenerate(parent
))
6059 if (!cpus_equal(sd
->span
, parent
->span
))
6062 /* Does parent contain flags not in child? */
6063 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6064 if (cflags
& SD_WAKE_AFFINE
)
6065 pflags
&= ~SD_WAKE_BALANCE
;
6066 /* Flags needing groups don't count if only 1 group in parent */
6067 if (parent
->groups
== parent
->groups
->next
) {
6068 pflags
&= ~(SD_LOAD_BALANCE
|
6069 SD_BALANCE_NEWIDLE
|
6073 SD_SHARE_PKG_RESOURCES
);
6075 if (~cflags
& pflags
)
6081 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6083 unsigned long flags
;
6084 const struct sched_class
*class;
6086 spin_lock_irqsave(&rq
->lock
, flags
);
6089 struct root_domain
*old_rd
= rq
->rd
;
6091 for (class = sched_class_highest
; class; class = class->next
) {
6092 if (class->leave_domain
)
6093 class->leave_domain(rq
);
6096 cpu_clear(rq
->cpu
, old_rd
->span
);
6097 cpu_clear(rq
->cpu
, old_rd
->online
);
6099 if (atomic_dec_and_test(&old_rd
->refcount
))
6103 atomic_inc(&rd
->refcount
);
6106 cpu_set(rq
->cpu
, rd
->span
);
6107 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6108 cpu_set(rq
->cpu
, rd
->online
);
6110 for (class = sched_class_highest
; class; class = class->next
) {
6111 if (class->join_domain
)
6112 class->join_domain(rq
);
6115 spin_unlock_irqrestore(&rq
->lock
, flags
);
6118 static void init_rootdomain(struct root_domain
*rd
)
6120 memset(rd
, 0, sizeof(*rd
));
6122 cpus_clear(rd
->span
);
6123 cpus_clear(rd
->online
);
6126 static void init_defrootdomain(void)
6128 init_rootdomain(&def_root_domain
);
6129 atomic_set(&def_root_domain
.refcount
, 1);
6132 static struct root_domain
*alloc_rootdomain(void)
6134 struct root_domain
*rd
;
6136 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6140 init_rootdomain(rd
);
6146 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6147 * hold the hotplug lock.
6150 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6152 struct rq
*rq
= cpu_rq(cpu
);
6153 struct sched_domain
*tmp
;
6155 /* Remove the sched domains which do not contribute to scheduling. */
6156 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6157 struct sched_domain
*parent
= tmp
->parent
;
6160 if (sd_parent_degenerate(tmp
, parent
)) {
6161 tmp
->parent
= parent
->parent
;
6163 parent
->parent
->child
= tmp
;
6167 if (sd
&& sd_degenerate(sd
)) {
6173 sched_domain_debug(sd
, cpu
);
6175 rq_attach_root(rq
, rd
);
6176 rcu_assign_pointer(rq
->sd
, sd
);
6179 /* cpus with isolated domains */
6180 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6182 /* Setup the mask of cpus configured for isolated domains */
6183 static int __init
isolated_cpu_setup(char *str
)
6185 int ints
[NR_CPUS
], i
;
6187 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6188 cpus_clear(cpu_isolated_map
);
6189 for (i
= 1; i
<= ints
[0]; i
++)
6190 if (ints
[i
] < NR_CPUS
)
6191 cpu_set(ints
[i
], cpu_isolated_map
);
6195 __setup("isolcpus=", isolated_cpu_setup
);
6198 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6199 * to a function which identifies what group(along with sched group) a CPU
6200 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6201 * (due to the fact that we keep track of groups covered with a cpumask_t).
6203 * init_sched_build_groups will build a circular linked list of the groups
6204 * covered by the given span, and will set each group's ->cpumask correctly,
6205 * and ->cpu_power to 0.
6208 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
6209 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6210 struct sched_group
**sg
))
6212 struct sched_group
*first
= NULL
, *last
= NULL
;
6213 cpumask_t covered
= CPU_MASK_NONE
;
6216 for_each_cpu_mask(i
, span
) {
6217 struct sched_group
*sg
;
6218 int group
= group_fn(i
, cpu_map
, &sg
);
6221 if (cpu_isset(i
, covered
))
6224 sg
->cpumask
= CPU_MASK_NONE
;
6225 sg
->__cpu_power
= 0;
6227 for_each_cpu_mask(j
, span
) {
6228 if (group_fn(j
, cpu_map
, NULL
) != group
)
6231 cpu_set(j
, covered
);
6232 cpu_set(j
, sg
->cpumask
);
6243 #define SD_NODES_PER_DOMAIN 16
6248 * find_next_best_node - find the next node to include in a sched_domain
6249 * @node: node whose sched_domain we're building
6250 * @used_nodes: nodes already in the sched_domain
6252 * Find the next node to include in a given scheduling domain. Simply
6253 * finds the closest node not already in the @used_nodes map.
6255 * Should use nodemask_t.
6257 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6259 int i
, n
, val
, min_val
, best_node
= 0;
6263 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6264 /* Start at @node */
6265 n
= (node
+ i
) % MAX_NUMNODES
;
6267 if (!nr_cpus_node(n
))
6270 /* Skip already used nodes */
6271 if (test_bit(n
, used_nodes
))
6274 /* Simple min distance search */
6275 val
= node_distance(node
, n
);
6277 if (val
< min_val
) {
6283 set_bit(best_node
, used_nodes
);
6288 * sched_domain_node_span - get a cpumask for a node's sched_domain
6289 * @node: node whose cpumask we're constructing
6290 * @size: number of nodes to include in this span
6292 * Given a node, construct a good cpumask for its sched_domain to span. It
6293 * should be one that prevents unnecessary balancing, but also spreads tasks
6296 static cpumask_t
sched_domain_node_span(int node
)
6298 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6299 cpumask_t span
, nodemask
;
6303 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6305 nodemask
= node_to_cpumask(node
);
6306 cpus_or(span
, span
, nodemask
);
6307 set_bit(node
, used_nodes
);
6309 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6310 int next_node
= find_next_best_node(node
, used_nodes
);
6312 nodemask
= node_to_cpumask(next_node
);
6313 cpus_or(span
, span
, nodemask
);
6320 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6323 * SMT sched-domains:
6325 #ifdef CONFIG_SCHED_SMT
6326 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6327 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6330 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6333 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6339 * multi-core sched-domains:
6341 #ifdef CONFIG_SCHED_MC
6342 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6343 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6346 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6348 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6351 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6352 cpus_and(mask
, mask
, *cpu_map
);
6353 group
= first_cpu(mask
);
6355 *sg
= &per_cpu(sched_group_core
, group
);
6358 #elif defined(CONFIG_SCHED_MC)
6360 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6363 *sg
= &per_cpu(sched_group_core
, cpu
);
6368 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6369 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6372 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6375 #ifdef CONFIG_SCHED_MC
6376 cpumask_t mask
= cpu_coregroup_map(cpu
);
6377 cpus_and(mask
, mask
, *cpu_map
);
6378 group
= first_cpu(mask
);
6379 #elif defined(CONFIG_SCHED_SMT)
6380 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6381 cpus_and(mask
, mask
, *cpu_map
);
6382 group
= first_cpu(mask
);
6387 *sg
= &per_cpu(sched_group_phys
, group
);
6393 * The init_sched_build_groups can't handle what we want to do with node
6394 * groups, so roll our own. Now each node has its own list of groups which
6395 * gets dynamically allocated.
6397 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6398 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6400 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6401 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6403 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6404 struct sched_group
**sg
)
6406 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6409 cpus_and(nodemask
, nodemask
, *cpu_map
);
6410 group
= first_cpu(nodemask
);
6413 *sg
= &per_cpu(sched_group_allnodes
, group
);
6417 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6419 struct sched_group
*sg
= group_head
;
6425 for_each_cpu_mask(j
, sg
->cpumask
) {
6426 struct sched_domain
*sd
;
6428 sd
= &per_cpu(phys_domains
, j
);
6429 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6431 * Only add "power" once for each
6437 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6440 } while (sg
!= group_head
);
6445 /* Free memory allocated for various sched_group structures */
6446 static void free_sched_groups(const cpumask_t
*cpu_map
)
6450 for_each_cpu_mask(cpu
, *cpu_map
) {
6451 struct sched_group
**sched_group_nodes
6452 = sched_group_nodes_bycpu
[cpu
];
6454 if (!sched_group_nodes
)
6457 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6458 cpumask_t nodemask
= node_to_cpumask(i
);
6459 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6461 cpus_and(nodemask
, nodemask
, *cpu_map
);
6462 if (cpus_empty(nodemask
))
6472 if (oldsg
!= sched_group_nodes
[i
])
6475 kfree(sched_group_nodes
);
6476 sched_group_nodes_bycpu
[cpu
] = NULL
;
6480 static void free_sched_groups(const cpumask_t
*cpu_map
)
6486 * Initialize sched groups cpu_power.
6488 * cpu_power indicates the capacity of sched group, which is used while
6489 * distributing the load between different sched groups in a sched domain.
6490 * Typically cpu_power for all the groups in a sched domain will be same unless
6491 * there are asymmetries in the topology. If there are asymmetries, group
6492 * having more cpu_power will pickup more load compared to the group having
6495 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6496 * the maximum number of tasks a group can handle in the presence of other idle
6497 * or lightly loaded groups in the same sched domain.
6499 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6501 struct sched_domain
*child
;
6502 struct sched_group
*group
;
6504 WARN_ON(!sd
|| !sd
->groups
);
6506 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6511 sd
->groups
->__cpu_power
= 0;
6514 * For perf policy, if the groups in child domain share resources
6515 * (for example cores sharing some portions of the cache hierarchy
6516 * or SMT), then set this domain groups cpu_power such that each group
6517 * can handle only one task, when there are other idle groups in the
6518 * same sched domain.
6520 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6522 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6523 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6528 * add cpu_power of each child group to this groups cpu_power
6530 group
= child
->groups
;
6532 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6533 group
= group
->next
;
6534 } while (group
!= child
->groups
);
6538 * Build sched domains for a given set of cpus and attach the sched domains
6539 * to the individual cpus
6541 static int build_sched_domains(const cpumask_t
*cpu_map
)
6544 struct root_domain
*rd
;
6546 struct sched_group
**sched_group_nodes
= NULL
;
6547 int sd_allnodes
= 0;
6550 * Allocate the per-node list of sched groups
6552 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6554 if (!sched_group_nodes
) {
6555 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6558 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6561 rd
= alloc_rootdomain();
6563 printk(KERN_WARNING
"Cannot alloc root domain\n");
6568 * Set up domains for cpus specified by the cpu_map.
6570 for_each_cpu_mask(i
, *cpu_map
) {
6571 struct sched_domain
*sd
= NULL
, *p
;
6572 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6574 cpus_and(nodemask
, nodemask
, *cpu_map
);
6577 if (cpus_weight(*cpu_map
) >
6578 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6579 sd
= &per_cpu(allnodes_domains
, i
);
6580 *sd
= SD_ALLNODES_INIT
;
6581 sd
->span
= *cpu_map
;
6582 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6588 sd
= &per_cpu(node_domains
, i
);
6590 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6594 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6598 sd
= &per_cpu(phys_domains
, i
);
6600 sd
->span
= nodemask
;
6604 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6606 #ifdef CONFIG_SCHED_MC
6608 sd
= &per_cpu(core_domains
, i
);
6610 sd
->span
= cpu_coregroup_map(i
);
6611 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6614 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6617 #ifdef CONFIG_SCHED_SMT
6619 sd
= &per_cpu(cpu_domains
, i
);
6620 *sd
= SD_SIBLING_INIT
;
6621 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6622 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6625 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6629 #ifdef CONFIG_SCHED_SMT
6630 /* Set up CPU (sibling) groups */
6631 for_each_cpu_mask(i
, *cpu_map
) {
6632 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6633 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6634 if (i
!= first_cpu(this_sibling_map
))
6637 init_sched_build_groups(this_sibling_map
, cpu_map
,
6642 #ifdef CONFIG_SCHED_MC
6643 /* Set up multi-core groups */
6644 for_each_cpu_mask(i
, *cpu_map
) {
6645 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6646 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6647 if (i
!= first_cpu(this_core_map
))
6649 init_sched_build_groups(this_core_map
, cpu_map
,
6650 &cpu_to_core_group
);
6654 /* Set up physical groups */
6655 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6656 cpumask_t nodemask
= node_to_cpumask(i
);
6658 cpus_and(nodemask
, nodemask
, *cpu_map
);
6659 if (cpus_empty(nodemask
))
6662 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6666 /* Set up node groups */
6668 init_sched_build_groups(*cpu_map
, cpu_map
,
6669 &cpu_to_allnodes_group
);
6671 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6672 /* Set up node groups */
6673 struct sched_group
*sg
, *prev
;
6674 cpumask_t nodemask
= node_to_cpumask(i
);
6675 cpumask_t domainspan
;
6676 cpumask_t covered
= CPU_MASK_NONE
;
6679 cpus_and(nodemask
, nodemask
, *cpu_map
);
6680 if (cpus_empty(nodemask
)) {
6681 sched_group_nodes
[i
] = NULL
;
6685 domainspan
= sched_domain_node_span(i
);
6686 cpus_and(domainspan
, domainspan
, *cpu_map
);
6688 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6690 printk(KERN_WARNING
"Can not alloc domain group for "
6694 sched_group_nodes
[i
] = sg
;
6695 for_each_cpu_mask(j
, nodemask
) {
6696 struct sched_domain
*sd
;
6698 sd
= &per_cpu(node_domains
, j
);
6701 sg
->__cpu_power
= 0;
6702 sg
->cpumask
= nodemask
;
6704 cpus_or(covered
, covered
, nodemask
);
6707 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6708 cpumask_t tmp
, notcovered
;
6709 int n
= (i
+ j
) % MAX_NUMNODES
;
6711 cpus_complement(notcovered
, covered
);
6712 cpus_and(tmp
, notcovered
, *cpu_map
);
6713 cpus_and(tmp
, tmp
, domainspan
);
6714 if (cpus_empty(tmp
))
6717 nodemask
= node_to_cpumask(n
);
6718 cpus_and(tmp
, tmp
, nodemask
);
6719 if (cpus_empty(tmp
))
6722 sg
= kmalloc_node(sizeof(struct sched_group
),
6726 "Can not alloc domain group for node %d\n", j
);
6729 sg
->__cpu_power
= 0;
6731 sg
->next
= prev
->next
;
6732 cpus_or(covered
, covered
, tmp
);
6739 /* Calculate CPU power for physical packages and nodes */
6740 #ifdef CONFIG_SCHED_SMT
6741 for_each_cpu_mask(i
, *cpu_map
) {
6742 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6744 init_sched_groups_power(i
, sd
);
6747 #ifdef CONFIG_SCHED_MC
6748 for_each_cpu_mask(i
, *cpu_map
) {
6749 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6751 init_sched_groups_power(i
, sd
);
6755 for_each_cpu_mask(i
, *cpu_map
) {
6756 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6758 init_sched_groups_power(i
, sd
);
6762 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6763 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6766 struct sched_group
*sg
;
6768 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6769 init_numa_sched_groups_power(sg
);
6773 /* Attach the domains */
6774 for_each_cpu_mask(i
, *cpu_map
) {
6775 struct sched_domain
*sd
;
6776 #ifdef CONFIG_SCHED_SMT
6777 sd
= &per_cpu(cpu_domains
, i
);
6778 #elif defined(CONFIG_SCHED_MC)
6779 sd
= &per_cpu(core_domains
, i
);
6781 sd
= &per_cpu(phys_domains
, i
);
6783 cpu_attach_domain(sd
, rd
, i
);
6790 free_sched_groups(cpu_map
);
6795 static cpumask_t
*doms_cur
; /* current sched domains */
6796 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6799 * Special case: If a kmalloc of a doms_cur partition (array of
6800 * cpumask_t) fails, then fallback to a single sched domain,
6801 * as determined by the single cpumask_t fallback_doms.
6803 static cpumask_t fallback_doms
;
6806 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6807 * For now this just excludes isolated cpus, but could be used to
6808 * exclude other special cases in the future.
6810 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6815 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6817 doms_cur
= &fallback_doms
;
6818 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6819 err
= build_sched_domains(doms_cur
);
6820 register_sched_domain_sysctl();
6825 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6827 free_sched_groups(cpu_map
);
6831 * Detach sched domains from a group of cpus specified in cpu_map
6832 * These cpus will now be attached to the NULL domain
6834 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6838 unregister_sched_domain_sysctl();
6840 for_each_cpu_mask(i
, *cpu_map
)
6841 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6842 synchronize_sched();
6843 arch_destroy_sched_domains(cpu_map
);
6847 * Partition sched domains as specified by the 'ndoms_new'
6848 * cpumasks in the array doms_new[] of cpumasks. This compares
6849 * doms_new[] to the current sched domain partitioning, doms_cur[].
6850 * It destroys each deleted domain and builds each new domain.
6852 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6853 * The masks don't intersect (don't overlap.) We should setup one
6854 * sched domain for each mask. CPUs not in any of the cpumasks will
6855 * not be load balanced. If the same cpumask appears both in the
6856 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6859 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6860 * ownership of it and will kfree it when done with it. If the caller
6861 * failed the kmalloc call, then it can pass in doms_new == NULL,
6862 * and partition_sched_domains() will fallback to the single partition
6865 * Call with hotplug lock held
6867 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6873 /* always unregister in case we don't destroy any domains */
6874 unregister_sched_domain_sysctl();
6876 if (doms_new
== NULL
) {
6878 doms_new
= &fallback_doms
;
6879 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6882 /* Destroy deleted domains */
6883 for (i
= 0; i
< ndoms_cur
; i
++) {
6884 for (j
= 0; j
< ndoms_new
; j
++) {
6885 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6888 /* no match - a current sched domain not in new doms_new[] */
6889 detach_destroy_domains(doms_cur
+ i
);
6894 /* Build new domains */
6895 for (i
= 0; i
< ndoms_new
; i
++) {
6896 for (j
= 0; j
< ndoms_cur
; j
++) {
6897 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6900 /* no match - add a new doms_new */
6901 build_sched_domains(doms_new
+ i
);
6906 /* Remember the new sched domains */
6907 if (doms_cur
!= &fallback_doms
)
6909 doms_cur
= doms_new
;
6910 ndoms_cur
= ndoms_new
;
6912 register_sched_domain_sysctl();
6917 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6918 static int arch_reinit_sched_domains(void)
6923 detach_destroy_domains(&cpu_online_map
);
6924 err
= arch_init_sched_domains(&cpu_online_map
);
6930 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6934 if (buf
[0] != '0' && buf
[0] != '1')
6938 sched_smt_power_savings
= (buf
[0] == '1');
6940 sched_mc_power_savings
= (buf
[0] == '1');
6942 ret
= arch_reinit_sched_domains();
6944 return ret
? ret
: count
;
6947 #ifdef CONFIG_SCHED_MC
6948 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6950 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6952 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6953 const char *buf
, size_t count
)
6955 return sched_power_savings_store(buf
, count
, 0);
6957 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6958 sched_mc_power_savings_store
);
6961 #ifdef CONFIG_SCHED_SMT
6962 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6964 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6966 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6967 const char *buf
, size_t count
)
6969 return sched_power_savings_store(buf
, count
, 1);
6971 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6972 sched_smt_power_savings_store
);
6975 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6979 #ifdef CONFIG_SCHED_SMT
6981 err
= sysfs_create_file(&cls
->kset
.kobj
,
6982 &attr_sched_smt_power_savings
.attr
);
6984 #ifdef CONFIG_SCHED_MC
6985 if (!err
&& mc_capable())
6986 err
= sysfs_create_file(&cls
->kset
.kobj
,
6987 &attr_sched_mc_power_savings
.attr
);
6994 * Force a reinitialization of the sched domains hierarchy. The domains
6995 * and groups cannot be updated in place without racing with the balancing
6996 * code, so we temporarily attach all running cpus to the NULL domain
6997 * which will prevent rebalancing while the sched domains are recalculated.
6999 static int update_sched_domains(struct notifier_block
*nfb
,
7000 unsigned long action
, void *hcpu
)
7003 case CPU_UP_PREPARE
:
7004 case CPU_UP_PREPARE_FROZEN
:
7005 case CPU_DOWN_PREPARE
:
7006 case CPU_DOWN_PREPARE_FROZEN
:
7007 detach_destroy_domains(&cpu_online_map
);
7010 case CPU_UP_CANCELED
:
7011 case CPU_UP_CANCELED_FROZEN
:
7012 case CPU_DOWN_FAILED
:
7013 case CPU_DOWN_FAILED_FROZEN
:
7015 case CPU_ONLINE_FROZEN
:
7017 case CPU_DEAD_FROZEN
:
7019 * Fall through and re-initialise the domains.
7026 /* The hotplug lock is already held by cpu_up/cpu_down */
7027 arch_init_sched_domains(&cpu_online_map
);
7032 void __init
sched_init_smp(void)
7034 cpumask_t non_isolated_cpus
;
7037 arch_init_sched_domains(&cpu_online_map
);
7038 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7039 if (cpus_empty(non_isolated_cpus
))
7040 cpu_set(smp_processor_id(), non_isolated_cpus
);
7042 /* XXX: Theoretical race here - CPU may be hotplugged now */
7043 hotcpu_notifier(update_sched_domains
, 0);
7045 /* Move init over to a non-isolated CPU */
7046 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
7048 sched_init_granularity();
7051 void __init
sched_init_smp(void)
7053 sched_init_granularity();
7055 #endif /* CONFIG_SMP */
7057 int in_sched_functions(unsigned long addr
)
7059 return in_lock_functions(addr
) ||
7060 (addr
>= (unsigned long)__sched_text_start
7061 && addr
< (unsigned long)__sched_text_end
);
7064 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7066 cfs_rq
->tasks_timeline
= RB_ROOT
;
7067 #ifdef CONFIG_FAIR_GROUP_SCHED
7070 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7073 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7075 struct rt_prio_array
*array
;
7078 array
= &rt_rq
->active
;
7079 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7080 INIT_LIST_HEAD(array
->queue
+ i
);
7081 __clear_bit(i
, array
->bitmap
);
7083 /* delimiter for bitsearch: */
7084 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7086 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7087 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7090 rt_rq
->rt_nr_migratory
= 0;
7091 rt_rq
->overloaded
= 0;
7095 rt_rq
->rt_throttled
= 0;
7097 #ifdef CONFIG_RT_GROUP_SCHED
7098 rt_rq
->rt_nr_boosted
= 0;
7103 #ifdef CONFIG_FAIR_GROUP_SCHED
7104 static void init_tg_cfs_entry(struct rq
*rq
, struct task_group
*tg
,
7105 struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
7108 tg
->cfs_rq
[cpu
] = cfs_rq
;
7109 init_cfs_rq(cfs_rq
, rq
);
7112 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7115 se
->cfs_rq
= &rq
->cfs
;
7117 se
->load
.weight
= tg
->shares
;
7118 se
->load
.inv_weight
= div64_64(1ULL<<32, se
->load
.weight
);
7123 #ifdef CONFIG_RT_GROUP_SCHED
7124 static void init_tg_rt_entry(struct rq
*rq
, struct task_group
*tg
,
7125 struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
,
7128 tg
->rt_rq
[cpu
] = rt_rq
;
7129 init_rt_rq(rt_rq
, rq
);
7131 rt_rq
->rt_se
= rt_se
;
7133 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7135 tg
->rt_se
[cpu
] = rt_se
;
7136 rt_se
->rt_rq
= &rq
->rt
;
7137 rt_se
->my_q
= rt_rq
;
7138 rt_se
->parent
= NULL
;
7139 INIT_LIST_HEAD(&rt_se
->run_list
);
7143 void __init
sched_init(void)
7145 int highest_cpu
= 0;
7149 init_defrootdomain();
7152 #ifdef CONFIG_GROUP_SCHED
7153 list_add(&init_task_group
.list
, &task_groups
);
7156 for_each_possible_cpu(i
) {
7160 spin_lock_init(&rq
->lock
);
7161 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7164 init_cfs_rq(&rq
->cfs
, rq
);
7165 init_rt_rq(&rq
->rt
, rq
);
7166 #ifdef CONFIG_FAIR_GROUP_SCHED
7167 init_task_group
.shares
= init_task_group_load
;
7168 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7169 init_tg_cfs_entry(rq
, &init_task_group
,
7170 &per_cpu(init_cfs_rq
, i
),
7171 &per_cpu(init_sched_entity
, i
), i
, 1);
7174 #ifdef CONFIG_RT_GROUP_SCHED
7175 init_task_group
.rt_runtime
=
7176 sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
7177 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7178 init_tg_rt_entry(rq
, &init_task_group
,
7179 &per_cpu(init_rt_rq
, i
),
7180 &per_cpu(init_sched_rt_entity
, i
), i
, 1);
7182 rq
->rt_period_expire
= 0;
7183 rq
->rt_throttled
= 0;
7185 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7186 rq
->cpu_load
[j
] = 0;
7190 rq
->active_balance
= 0;
7191 rq
->next_balance
= jiffies
;
7194 rq
->migration_thread
= NULL
;
7195 INIT_LIST_HEAD(&rq
->migration_queue
);
7196 rq_attach_root(rq
, &def_root_domain
);
7199 atomic_set(&rq
->nr_iowait
, 0);
7203 set_load_weight(&init_task
);
7205 #ifdef CONFIG_PREEMPT_NOTIFIERS
7206 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7210 nr_cpu_ids
= highest_cpu
+ 1;
7211 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7214 #ifdef CONFIG_RT_MUTEXES
7215 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7219 * The boot idle thread does lazy MMU switching as well:
7221 atomic_inc(&init_mm
.mm_count
);
7222 enter_lazy_tlb(&init_mm
, current
);
7225 * Make us the idle thread. Technically, schedule() should not be
7226 * called from this thread, however somewhere below it might be,
7227 * but because we are the idle thread, we just pick up running again
7228 * when this runqueue becomes "idle".
7230 init_idle(current
, smp_processor_id());
7232 * During early bootup we pretend to be a normal task:
7234 current
->sched_class
= &fair_sched_class
;
7236 scheduler_running
= 1;
7239 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7240 void __might_sleep(char *file
, int line
)
7243 static unsigned long prev_jiffy
; /* ratelimiting */
7245 if ((in_atomic() || irqs_disabled()) &&
7246 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7247 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7249 prev_jiffy
= jiffies
;
7250 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7251 " context at %s:%d\n", file
, line
);
7252 printk("in_atomic():%d, irqs_disabled():%d\n",
7253 in_atomic(), irqs_disabled());
7254 debug_show_held_locks(current
);
7255 if (irqs_disabled())
7256 print_irqtrace_events(current
);
7261 EXPORT_SYMBOL(__might_sleep
);
7264 #ifdef CONFIG_MAGIC_SYSRQ
7265 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7268 update_rq_clock(rq
);
7269 on_rq
= p
->se
.on_rq
;
7271 deactivate_task(rq
, p
, 0);
7272 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7274 activate_task(rq
, p
, 0);
7275 resched_task(rq
->curr
);
7279 void normalize_rt_tasks(void)
7281 struct task_struct
*g
, *p
;
7282 unsigned long flags
;
7285 read_lock_irqsave(&tasklist_lock
, flags
);
7286 do_each_thread(g
, p
) {
7288 * Only normalize user tasks:
7293 p
->se
.exec_start
= 0;
7294 #ifdef CONFIG_SCHEDSTATS
7295 p
->se
.wait_start
= 0;
7296 p
->se
.sleep_start
= 0;
7297 p
->se
.block_start
= 0;
7299 task_rq(p
)->clock
= 0;
7303 * Renice negative nice level userspace
7306 if (TASK_NICE(p
) < 0 && p
->mm
)
7307 set_user_nice(p
, 0);
7311 spin_lock(&p
->pi_lock
);
7312 rq
= __task_rq_lock(p
);
7314 normalize_task(rq
, p
);
7316 __task_rq_unlock(rq
);
7317 spin_unlock(&p
->pi_lock
);
7318 } while_each_thread(g
, p
);
7320 read_unlock_irqrestore(&tasklist_lock
, flags
);
7323 #endif /* CONFIG_MAGIC_SYSRQ */
7327 * These functions are only useful for the IA64 MCA handling.
7329 * They can only be called when the whole system has been
7330 * stopped - every CPU needs to be quiescent, and no scheduling
7331 * activity can take place. Using them for anything else would
7332 * be a serious bug, and as a result, they aren't even visible
7333 * under any other configuration.
7337 * curr_task - return the current task for a given cpu.
7338 * @cpu: the processor in question.
7340 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7342 struct task_struct
*curr_task(int cpu
)
7344 return cpu_curr(cpu
);
7348 * set_curr_task - set the current task for a given cpu.
7349 * @cpu: the processor in question.
7350 * @p: the task pointer to set.
7352 * Description: This function must only be used when non-maskable interrupts
7353 * are serviced on a separate stack. It allows the architecture to switch the
7354 * notion of the current task on a cpu in a non-blocking manner. This function
7355 * must be called with all CPU's synchronized, and interrupts disabled, the
7356 * and caller must save the original value of the current task (see
7357 * curr_task() above) and restore that value before reenabling interrupts and
7358 * re-starting the system.
7360 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7362 void set_curr_task(int cpu
, struct task_struct
*p
)
7369 #ifdef CONFIG_GROUP_SCHED
7371 #ifdef CONFIG_FAIR_GROUP_SCHED
7372 static void free_fair_sched_group(struct task_group
*tg
)
7376 for_each_possible_cpu(i
) {
7378 kfree(tg
->cfs_rq
[i
]);
7387 static int alloc_fair_sched_group(struct task_group
*tg
)
7389 struct cfs_rq
*cfs_rq
;
7390 struct sched_entity
*se
;
7394 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7397 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7401 tg
->shares
= NICE_0_LOAD
;
7403 for_each_possible_cpu(i
) {
7406 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
7407 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7411 se
= kmalloc_node(sizeof(struct sched_entity
),
7412 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7416 init_tg_cfs_entry(rq
, tg
, cfs_rq
, se
, i
, 0);
7425 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7427 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
7428 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
7431 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7433 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
7436 static inline void free_fair_sched_group(struct task_group
*tg
)
7440 static inline int alloc_fair_sched_group(struct task_group
*tg
)
7445 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7449 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7454 #ifdef CONFIG_RT_GROUP_SCHED
7455 static void free_rt_sched_group(struct task_group
*tg
)
7459 for_each_possible_cpu(i
) {
7461 kfree(tg
->rt_rq
[i
]);
7463 kfree(tg
->rt_se
[i
]);
7470 static int alloc_rt_sched_group(struct task_group
*tg
)
7472 struct rt_rq
*rt_rq
;
7473 struct sched_rt_entity
*rt_se
;
7477 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * NR_CPUS
, GFP_KERNEL
);
7480 tg
->rt_se
= kzalloc(sizeof(rt_se
) * NR_CPUS
, GFP_KERNEL
);
7486 for_each_possible_cpu(i
) {
7489 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
7490 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7494 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
7495 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7499 init_tg_rt_entry(rq
, tg
, rt_rq
, rt_se
, i
, 0);
7508 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7510 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
7511 &cpu_rq(cpu
)->leaf_rt_rq_list
);
7514 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7516 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
7519 static inline void free_rt_sched_group(struct task_group
*tg
)
7523 static inline int alloc_rt_sched_group(struct task_group
*tg
)
7528 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7532 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7537 static void free_sched_group(struct task_group
*tg
)
7539 free_fair_sched_group(tg
);
7540 free_rt_sched_group(tg
);
7544 /* allocate runqueue etc for a new task group */
7545 struct task_group
*sched_create_group(void)
7547 struct task_group
*tg
;
7548 unsigned long flags
;
7551 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7553 return ERR_PTR(-ENOMEM
);
7555 if (!alloc_fair_sched_group(tg
))
7558 if (!alloc_rt_sched_group(tg
))
7561 spin_lock_irqsave(&task_group_lock
, flags
);
7562 for_each_possible_cpu(i
) {
7563 register_fair_sched_group(tg
, i
);
7564 register_rt_sched_group(tg
, i
);
7566 list_add_rcu(&tg
->list
, &task_groups
);
7567 spin_unlock_irqrestore(&task_group_lock
, flags
);
7572 free_sched_group(tg
);
7573 return ERR_PTR(-ENOMEM
);
7576 /* rcu callback to free various structures associated with a task group */
7577 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7579 /* now it should be safe to free those cfs_rqs */
7580 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7583 /* Destroy runqueue etc associated with a task group */
7584 void sched_destroy_group(struct task_group
*tg
)
7586 unsigned long flags
;
7589 spin_lock_irqsave(&task_group_lock
, flags
);
7590 for_each_possible_cpu(i
) {
7591 unregister_fair_sched_group(tg
, i
);
7592 unregister_rt_sched_group(tg
, i
);
7594 list_del_rcu(&tg
->list
);
7595 spin_unlock_irqrestore(&task_group_lock
, flags
);
7597 /* wait for possible concurrent references to cfs_rqs complete */
7598 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7601 /* change task's runqueue when it moves between groups.
7602 * The caller of this function should have put the task in its new group
7603 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7604 * reflect its new group.
7606 void sched_move_task(struct task_struct
*tsk
)
7609 unsigned long flags
;
7612 rq
= task_rq_lock(tsk
, &flags
);
7614 update_rq_clock(rq
);
7616 running
= task_current(rq
, tsk
);
7617 on_rq
= tsk
->se
.on_rq
;
7620 dequeue_task(rq
, tsk
, 0);
7621 if (unlikely(running
))
7622 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7624 set_task_rq(tsk
, task_cpu(tsk
));
7626 #ifdef CONFIG_FAIR_GROUP_SCHED
7627 if (tsk
->sched_class
->moved_group
)
7628 tsk
->sched_class
->moved_group(tsk
);
7631 if (unlikely(running
))
7632 tsk
->sched_class
->set_curr_task(rq
);
7634 enqueue_task(rq
, tsk
, 0);
7636 task_rq_unlock(rq
, &flags
);
7639 #ifdef CONFIG_FAIR_GROUP_SCHED
7640 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7642 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7643 struct rq
*rq
= cfs_rq
->rq
;
7646 spin_lock_irq(&rq
->lock
);
7650 dequeue_entity(cfs_rq
, se
, 0);
7652 se
->load
.weight
= shares
;
7653 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7656 enqueue_entity(cfs_rq
, se
, 0);
7658 spin_unlock_irq(&rq
->lock
);
7661 static DEFINE_MUTEX(shares_mutex
);
7663 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7666 unsigned long flags
;
7669 * A weight of 0 or 1 can cause arithmetics problems.
7670 * (The default weight is 1024 - so there's no practical
7671 * limitation from this.)
7676 mutex_lock(&shares_mutex
);
7677 if (tg
->shares
== shares
)
7680 spin_lock_irqsave(&task_group_lock
, flags
);
7681 for_each_possible_cpu(i
)
7682 unregister_fair_sched_group(tg
, i
);
7683 spin_unlock_irqrestore(&task_group_lock
, flags
);
7685 /* wait for any ongoing reference to this group to finish */
7686 synchronize_sched();
7689 * Now we are free to modify the group's share on each cpu
7690 * w/o tripping rebalance_share or load_balance_fair.
7692 tg
->shares
= shares
;
7693 for_each_possible_cpu(i
)
7694 set_se_shares(tg
->se
[i
], shares
);
7697 * Enable load balance activity on this group, by inserting it back on
7698 * each cpu's rq->leaf_cfs_rq_list.
7700 spin_lock_irqsave(&task_group_lock
, flags
);
7701 for_each_possible_cpu(i
)
7702 register_fair_sched_group(tg
, i
);
7703 spin_unlock_irqrestore(&task_group_lock
, flags
);
7705 mutex_unlock(&shares_mutex
);
7709 unsigned long sched_group_shares(struct task_group
*tg
)
7715 #ifdef CONFIG_RT_GROUP_SCHED
7717 * Ensure that the real time constraints are schedulable.
7719 static DEFINE_MUTEX(rt_constraints_mutex
);
7721 static unsigned long to_ratio(u64 period
, u64 runtime
)
7723 if (runtime
== RUNTIME_INF
)
7726 return div64_64(runtime
<< 16, period
);
7729 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7731 struct task_group
*tgi
;
7732 unsigned long total
= 0;
7733 unsigned long global_ratio
=
7734 to_ratio(sysctl_sched_rt_period
,
7735 sysctl_sched_rt_runtime
< 0 ?
7736 RUNTIME_INF
: sysctl_sched_rt_runtime
);
7739 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
7743 total
+= to_ratio(period
, tgi
->rt_runtime
);
7747 return total
+ to_ratio(period
, runtime
) < global_ratio
;
7750 /* Must be called with tasklist_lock held */
7751 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7753 struct task_struct
*g
, *p
;
7754 do_each_thread(g
, p
) {
7755 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
7757 } while_each_thread(g
, p
);
7761 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7763 u64 rt_runtime
, rt_period
;
7766 rt_period
= (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
7767 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7768 if (rt_runtime_us
== -1)
7769 rt_runtime
= RUNTIME_INF
;
7771 mutex_lock(&rt_constraints_mutex
);
7772 read_lock(&tasklist_lock
);
7773 if (rt_runtime_us
== 0 && tg_has_rt_tasks(tg
)) {
7777 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
7781 tg
->rt_runtime
= rt_runtime
;
7783 read_unlock(&tasklist_lock
);
7784 mutex_unlock(&rt_constraints_mutex
);
7789 long sched_group_rt_runtime(struct task_group
*tg
)
7793 if (tg
->rt_runtime
== RUNTIME_INF
)
7796 rt_runtime_us
= tg
->rt_runtime
;
7797 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7798 return rt_runtime_us
;
7801 #endif /* CONFIG_GROUP_SCHED */
7803 #ifdef CONFIG_CGROUP_SCHED
7805 /* return corresponding task_group object of a cgroup */
7806 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7808 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7809 struct task_group
, css
);
7812 static struct cgroup_subsys_state
*
7813 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7815 struct task_group
*tg
;
7817 if (!cgrp
->parent
) {
7818 /* This is early initialization for the top cgroup */
7819 init_task_group
.css
.cgroup
= cgrp
;
7820 return &init_task_group
.css
;
7823 /* we support only 1-level deep hierarchical scheduler atm */
7824 if (cgrp
->parent
->parent
)
7825 return ERR_PTR(-EINVAL
);
7827 tg
= sched_create_group();
7829 return ERR_PTR(-ENOMEM
);
7831 /* Bind the cgroup to task_group object we just created */
7832 tg
->css
.cgroup
= cgrp
;
7838 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7840 struct task_group
*tg
= cgroup_tg(cgrp
);
7842 sched_destroy_group(tg
);
7846 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7847 struct task_struct
*tsk
)
7849 #ifdef CONFIG_RT_GROUP_SCHED
7850 /* Don't accept realtime tasks when there is no way for them to run */
7851 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_runtime
== 0)
7854 /* We don't support RT-tasks being in separate groups */
7855 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 #ifdef CONFIG_FAIR_GROUP_SCHED
7870 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7873 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7876 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7878 struct task_group
*tg
= cgroup_tg(cgrp
);
7880 return (u64
) tg
->shares
;
7884 #ifdef CONFIG_RT_GROUP_SCHED
7885 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7887 const char __user
*userbuf
,
7888 size_t nbytes
, loff_t
*unused_ppos
)
7897 if (nbytes
>= sizeof(buffer
))
7899 if (copy_from_user(buffer
, userbuf
, nbytes
))
7902 buffer
[nbytes
] = 0; /* nul-terminate */
7904 /* strip newline if necessary */
7905 if (nbytes
&& (buffer
[nbytes
-1] == '\n'))
7906 buffer
[nbytes
-1] = 0;
7907 val
= simple_strtoll(buffer
, &end
, 0);
7911 /* Pass to subsystem */
7912 retval
= sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7918 static ssize_t
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
,
7920 char __user
*buf
, size_t nbytes
,
7924 long val
= sched_group_rt_runtime(cgroup_tg(cgrp
));
7925 int len
= sprintf(tmp
, "%ld\n", val
);
7927 return simple_read_from_buffer(buf
, nbytes
, ppos
, tmp
, len
);
7931 static struct cftype cpu_files
[] = {
7932 #ifdef CONFIG_FAIR_GROUP_SCHED
7935 .read_uint
= cpu_shares_read_uint
,
7936 .write_uint
= cpu_shares_write_uint
,
7939 #ifdef CONFIG_RT_GROUP_SCHED
7941 .name
= "rt_runtime_us",
7942 .read
= cpu_rt_runtime_read
,
7943 .write
= cpu_rt_runtime_write
,
7948 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7950 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7953 struct cgroup_subsys cpu_cgroup_subsys
= {
7955 .create
= cpu_cgroup_create
,
7956 .destroy
= cpu_cgroup_destroy
,
7957 .can_attach
= cpu_cgroup_can_attach
,
7958 .attach
= cpu_cgroup_attach
,
7959 .populate
= cpu_cgroup_populate
,
7960 .subsys_id
= cpu_cgroup_subsys_id
,
7964 #endif /* CONFIG_CGROUP_SCHED */
7966 #ifdef CONFIG_CGROUP_CPUACCT
7969 * CPU accounting code for task groups.
7971 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7972 * (balbir@in.ibm.com).
7975 /* track cpu usage of a group of tasks */
7977 struct cgroup_subsys_state css
;
7978 /* cpuusage holds pointer to a u64-type object on every cpu */
7982 struct cgroup_subsys cpuacct_subsys
;
7984 /* return cpu accounting group corresponding to this container */
7985 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7987 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7988 struct cpuacct
, css
);
7991 /* return cpu accounting group to which this task belongs */
7992 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7994 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7995 struct cpuacct
, css
);
7998 /* create a new cpu accounting group */
7999 static struct cgroup_subsys_state
*cpuacct_create(
8000 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8002 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8005 return ERR_PTR(-ENOMEM
);
8007 ca
->cpuusage
= alloc_percpu(u64
);
8008 if (!ca
->cpuusage
) {
8010 return ERR_PTR(-ENOMEM
);
8016 /* destroy an existing cpu accounting group */
8018 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8020 struct cpuacct
*ca
= cgroup_ca(cont
);
8022 free_percpu(ca
->cpuusage
);
8026 /* return total cpu usage (in nanoseconds) of a group */
8027 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
8029 struct cpuacct
*ca
= cgroup_ca(cont
);
8030 u64 totalcpuusage
= 0;
8033 for_each_possible_cpu(i
) {
8034 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8037 * Take rq->lock to make 64-bit addition safe on 32-bit
8040 spin_lock_irq(&cpu_rq(i
)->lock
);
8041 totalcpuusage
+= *cpuusage
;
8042 spin_unlock_irq(&cpu_rq(i
)->lock
);
8045 return totalcpuusage
;
8048 static struct cftype files
[] = {
8051 .read_uint
= cpuusage_read
,
8055 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8057 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
8061 * charge this task's execution time to its accounting group.
8063 * called with rq->lock held.
8065 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8069 if (!cpuacct_subsys
.active
)
8074 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
8076 *cpuusage
+= cputime
;
8080 struct cgroup_subsys cpuacct_subsys
= {
8082 .create
= cpuacct_create
,
8083 .destroy
= cpuacct_destroy
,
8084 .populate
= cpuacct_populate
,
8085 .subsys_id
= cpuacct_subsys_id
,
8087 #endif /* CONFIG_CGROUP_CPUACCT */