4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak
)) sched_clock(void)
75 return (unsigned long long)jiffies
* (1000000000 / HZ
);
79 #define is_migration_thread(p, rq) ((p) == (rq)->migration_thread)
81 #define is_migration_thread(p, rq) 0
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Some helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (1000000000 / HZ))
106 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
122 * Since cpu_power is a 'constant', we can use a reciprocal divide.
124 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
126 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
130 * Each time a sched group cpu_power is changed,
131 * we must compute its reciprocal value
133 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
135 sg
->__cpu_power
+= val
;
136 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
140 static inline int rt_policy(int policy
)
142 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
147 static inline int task_has_rt_policy(struct task_struct
*p
)
149 return rt_policy(p
->policy
);
153 * This is the priority-queue data structure of the RT scheduling class:
155 struct rt_prio_array
{
156 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
157 struct list_head queue
[MAX_RT_PRIO
];
160 #ifdef CONFIG_FAIR_GROUP_SCHED
164 /* task group related information */
166 /* schedulable entities of this group on each cpu */
167 struct sched_entity
**se
;
168 /* runqueue "owned" by this group on each cpu */
169 struct cfs_rq
**cfs_rq
;
170 unsigned long shares
;
171 /* spinlock to serialize modification to shares */
175 /* Default task group's sched entity on each cpu */
176 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
177 /* Default task group's cfs_rq on each cpu */
178 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
180 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
181 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
183 /* Default task group.
184 * Every task in system belong to this group at bootup.
186 struct task_group init_task_group
= {
187 .se
= init_sched_entity_p
,
188 .cfs_rq
= init_cfs_rq_p
,
191 #ifdef CONFIG_FAIR_USER_SCHED
192 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
194 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
197 static int init_task_group_load
= INIT_TASK_GRP_LOAD
;
199 /* return group to which a task belongs */
200 static inline struct task_group
*task_group(struct task_struct
*p
)
202 struct task_group
*tg
;
204 #ifdef CONFIG_FAIR_USER_SCHED
207 tg
= &init_task_group
;
213 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
214 static inline void set_task_cfs_rq(struct task_struct
*p
)
216 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[task_cpu(p
)];
217 p
->se
.parent
= task_group(p
)->se
[task_cpu(p
)];
222 static inline void set_task_cfs_rq(struct task_struct
*p
) { }
224 #endif /* CONFIG_FAIR_GROUP_SCHED */
226 /* CFS-related fields in a runqueue */
228 struct load_weight load
;
229 unsigned long nr_running
;
234 struct rb_root tasks_timeline
;
235 struct rb_node
*rb_leftmost
;
236 struct rb_node
*rb_load_balance_curr
;
237 /* 'curr' points to currently running entity on this cfs_rq.
238 * It is set to NULL otherwise (i.e when none are currently running).
240 struct sched_entity
*curr
;
242 unsigned long nr_spread_over
;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
247 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
248 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
249 * (like users, containers etc.)
251 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
252 * list is used during load balance.
254 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
255 struct task_group
*tg
; /* group that "owns" this runqueue */
260 /* Real-Time classes' related field in a runqueue: */
262 struct rt_prio_array active
;
263 int rt_load_balance_idx
;
264 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
268 * This is the main, per-CPU runqueue data structure.
270 * Locking rule: those places that want to lock multiple runqueues
271 * (such as the load balancing or the thread migration code), lock
272 * acquire operations must be ordered by ascending &runqueue.
275 spinlock_t lock
; /* runqueue lock */
278 * nr_running and cpu_load should be in the same cacheline because
279 * remote CPUs use both these fields when doing load calculation.
281 unsigned long nr_running
;
282 #define CPU_LOAD_IDX_MAX 5
283 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
284 unsigned char idle_at_tick
;
286 unsigned char in_nohz_recently
;
288 struct load_weight load
; /* capture load from *all* tasks on this cpu */
289 unsigned long nr_load_updates
;
293 #ifdef CONFIG_FAIR_GROUP_SCHED
294 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
299 * This is part of a global counter where only the total sum
300 * over all CPUs matters. A task can increase this counter on
301 * one CPU and if it got migrated afterwards it may decrease
302 * it on another CPU. Always updated under the runqueue lock:
304 unsigned long nr_uninterruptible
;
306 struct task_struct
*curr
, *idle
;
307 unsigned long next_balance
;
308 struct mm_struct
*prev_mm
;
310 u64 clock
, prev_clock_raw
;
313 unsigned int clock_warps
, clock_overflows
;
315 unsigned int clock_deep_idle_events
;
321 struct sched_domain
*sd
;
323 /* For active balancing */
326 int cpu
; /* cpu of this runqueue */
328 struct task_struct
*migration_thread
;
329 struct list_head migration_queue
;
332 #ifdef CONFIG_SCHEDSTATS
334 struct sched_info rq_sched_info
;
336 /* sys_sched_yield() stats */
337 unsigned long yld_exp_empty
;
338 unsigned long yld_act_empty
;
339 unsigned long yld_both_empty
;
340 unsigned long yld_count
;
342 /* schedule() stats */
343 unsigned long sched_switch
;
344 unsigned long sched_count
;
345 unsigned long sched_goidle
;
347 /* try_to_wake_up() stats */
348 unsigned long ttwu_count
;
349 unsigned long ttwu_local
;
352 unsigned long bkl_count
;
354 struct lock_class_key rq_lock_key
;
357 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
358 static DEFINE_MUTEX(sched_hotcpu_mutex
);
360 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
362 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
365 static inline int cpu_of(struct rq
*rq
)
375 * Update the per-runqueue clock, as finegrained as the platform can give
376 * us, but without assuming monotonicity, etc.:
378 static void __update_rq_clock(struct rq
*rq
)
380 u64 prev_raw
= rq
->prev_clock_raw
;
381 u64 now
= sched_clock();
382 s64 delta
= now
- prev_raw
;
383 u64 clock
= rq
->clock
;
385 #ifdef CONFIG_SCHED_DEBUG
386 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
389 * Protect against sched_clock() occasionally going backwards:
391 if (unlikely(delta
< 0)) {
396 * Catch too large forward jumps too:
398 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
399 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
400 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
403 rq
->clock_overflows
++;
405 if (unlikely(delta
> rq
->clock_max_delta
))
406 rq
->clock_max_delta
= delta
;
411 rq
->prev_clock_raw
= now
;
415 static void update_rq_clock(struct rq
*rq
)
417 if (likely(smp_processor_id() == cpu_of(rq
)))
418 __update_rq_clock(rq
);
422 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
423 * See detach_destroy_domains: synchronize_sched for details.
425 * The domain tree of any CPU may only be accessed from within
426 * preempt-disabled sections.
428 #define for_each_domain(cpu, __sd) \
429 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
431 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
432 #define this_rq() (&__get_cpu_var(runqueues))
433 #define task_rq(p) cpu_rq(task_cpu(p))
434 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
437 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
439 #ifdef CONFIG_SCHED_DEBUG
440 # define const_debug __read_mostly
442 # define const_debug static const
446 * Debugging: various feature bits
449 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
450 SCHED_FEAT_START_DEBIT
= 2,
451 SCHED_FEAT_TREE_AVG
= 4,
452 SCHED_FEAT_APPROX_AVG
= 8,
453 SCHED_FEAT_WAKEUP_PREEMPT
= 16,
454 SCHED_FEAT_PREEMPT_RESTRICT
= 32,
457 const_debug
unsigned int sysctl_sched_features
=
458 SCHED_FEAT_NEW_FAIR_SLEEPERS
*1 |
459 SCHED_FEAT_START_DEBIT
*1 |
460 SCHED_FEAT_TREE_AVG
*0 |
461 SCHED_FEAT_APPROX_AVG
*0 |
462 SCHED_FEAT_WAKEUP_PREEMPT
*1 |
463 SCHED_FEAT_PREEMPT_RESTRICT
*1;
465 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
468 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
469 * clock constructed from sched_clock():
471 unsigned long long cpu_clock(int cpu
)
473 unsigned long long now
;
477 local_irq_save(flags
);
481 local_irq_restore(flags
);
485 EXPORT_SYMBOL_GPL(cpu_clock
);
487 #ifndef prepare_arch_switch
488 # define prepare_arch_switch(next) do { } while (0)
490 #ifndef finish_arch_switch
491 # define finish_arch_switch(prev) do { } while (0)
494 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
495 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
497 return rq
->curr
== p
;
500 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
504 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
506 #ifdef CONFIG_DEBUG_SPINLOCK
507 /* this is a valid case when another task releases the spinlock */
508 rq
->lock
.owner
= current
;
511 * If we are tracking spinlock dependencies then we have to
512 * fix up the runqueue lock - which gets 'carried over' from
515 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
517 spin_unlock_irq(&rq
->lock
);
520 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
521 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
526 return rq
->curr
== p
;
530 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
534 * We can optimise this out completely for !SMP, because the
535 * SMP rebalancing from interrupt is the only thing that cares
540 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
541 spin_unlock_irq(&rq
->lock
);
543 spin_unlock(&rq
->lock
);
547 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
551 * After ->oncpu is cleared, the task can be moved to a different CPU.
552 * We must ensure this doesn't happen until the switch is completely
558 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
562 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
565 * __task_rq_lock - lock the runqueue a given task resides on.
566 * Must be called interrupts disabled.
568 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
572 struct rq
*rq
= task_rq(p
);
573 spin_lock(&rq
->lock
);
574 if (likely(rq
== task_rq(p
)))
576 spin_unlock(&rq
->lock
);
581 * task_rq_lock - lock the runqueue a given task resides on and disable
582 * interrupts. Note the ordering: we can safely lookup the task_rq without
583 * explicitly disabling preemption.
585 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
591 local_irq_save(*flags
);
593 spin_lock(&rq
->lock
);
594 if (likely(rq
== task_rq(p
)))
596 spin_unlock_irqrestore(&rq
->lock
, *flags
);
600 static void __task_rq_unlock(struct rq
*rq
)
603 spin_unlock(&rq
->lock
);
606 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
609 spin_unlock_irqrestore(&rq
->lock
, *flags
);
613 * this_rq_lock - lock this runqueue and disable interrupts.
615 static struct rq
*this_rq_lock(void)
622 spin_lock(&rq
->lock
);
628 * We are going deep-idle (irqs are disabled):
630 void sched_clock_idle_sleep_event(void)
632 struct rq
*rq
= cpu_rq(smp_processor_id());
634 spin_lock(&rq
->lock
);
635 __update_rq_clock(rq
);
636 spin_unlock(&rq
->lock
);
637 rq
->clock_deep_idle_events
++;
639 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
642 * We just idled delta nanoseconds (called with irqs disabled):
644 void sched_clock_idle_wakeup_event(u64 delta_ns
)
646 struct rq
*rq
= cpu_rq(smp_processor_id());
647 u64 now
= sched_clock();
649 rq
->idle_clock
+= delta_ns
;
651 * Override the previous timestamp and ignore all
652 * sched_clock() deltas that occured while we idled,
653 * and use the PM-provided delta_ns to advance the
656 spin_lock(&rq
->lock
);
657 rq
->prev_clock_raw
= now
;
658 rq
->clock
+= delta_ns
;
659 spin_unlock(&rq
->lock
);
661 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
664 * resched_task - mark a task 'to be rescheduled now'.
666 * On UP this means the setting of the need_resched flag, on SMP it
667 * might also involve a cross-CPU call to trigger the scheduler on
672 #ifndef tsk_is_polling
673 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
676 static void resched_task(struct task_struct
*p
)
680 assert_spin_locked(&task_rq(p
)->lock
);
682 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
685 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
688 if (cpu
== smp_processor_id())
691 /* NEED_RESCHED must be visible before we test polling */
693 if (!tsk_is_polling(p
))
694 smp_send_reschedule(cpu
);
697 static void resched_cpu(int cpu
)
699 struct rq
*rq
= cpu_rq(cpu
);
702 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
704 resched_task(cpu_curr(cpu
));
705 spin_unlock_irqrestore(&rq
->lock
, flags
);
708 static inline void resched_task(struct task_struct
*p
)
710 assert_spin_locked(&task_rq(p
)->lock
);
711 set_tsk_need_resched(p
);
715 #if BITS_PER_LONG == 32
716 # define WMULT_CONST (~0UL)
718 # define WMULT_CONST (1UL << 32)
721 #define WMULT_SHIFT 32
724 * Shift right and round:
726 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
729 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
730 struct load_weight
*lw
)
734 if (unlikely(!lw
->inv_weight
))
735 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
737 tmp
= (u64
)delta_exec
* weight
;
739 * Check whether we'd overflow the 64-bit multiplication:
741 if (unlikely(tmp
> WMULT_CONST
))
742 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
745 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
747 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
750 static inline unsigned long
751 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
753 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
756 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
761 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
767 * To aid in avoiding the subversion of "niceness" due to uneven distribution
768 * of tasks with abnormal "nice" values across CPUs the contribution that
769 * each task makes to its run queue's load is weighted according to its
770 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
771 * scaled version of the new time slice allocation that they receive on time
775 #define WEIGHT_IDLEPRIO 2
776 #define WMULT_IDLEPRIO (1 << 31)
779 * Nice levels are multiplicative, with a gentle 10% change for every
780 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
781 * nice 1, it will get ~10% less CPU time than another CPU-bound task
782 * that remained on nice 0.
784 * The "10% effect" is relative and cumulative: from _any_ nice level,
785 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
786 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
787 * If a task goes up by ~10% and another task goes down by ~10% then
788 * the relative distance between them is ~25%.)
790 static const int prio_to_weight
[40] = {
791 /* -20 */ 88761, 71755, 56483, 46273, 36291,
792 /* -15 */ 29154, 23254, 18705, 14949, 11916,
793 /* -10 */ 9548, 7620, 6100, 4904, 3906,
794 /* -5 */ 3121, 2501, 1991, 1586, 1277,
795 /* 0 */ 1024, 820, 655, 526, 423,
796 /* 5 */ 335, 272, 215, 172, 137,
797 /* 10 */ 110, 87, 70, 56, 45,
798 /* 15 */ 36, 29, 23, 18, 15,
802 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
804 * In cases where the weight does not change often, we can use the
805 * precalculated inverse to speed up arithmetics by turning divisions
806 * into multiplications:
808 static const u32 prio_to_wmult
[40] = {
809 /* -20 */ 48388, 59856, 76040, 92818, 118348,
810 /* -15 */ 147320, 184698, 229616, 287308, 360437,
811 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
812 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
813 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
814 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
815 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
816 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
819 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
822 * runqueue iterator, to support SMP load-balancing between different
823 * scheduling classes, without having to expose their internal data
824 * structures to the load-balancing proper:
828 struct task_struct
*(*start
)(void *);
829 struct task_struct
*(*next
)(void *);
832 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
833 unsigned long max_nr_move
, unsigned long max_load_move
,
834 struct sched_domain
*sd
, enum cpu_idle_type idle
,
835 int *all_pinned
, unsigned long *load_moved
,
836 int *this_best_prio
, struct rq_iterator
*iterator
);
838 #include "sched_stats.h"
839 #include "sched_idletask.c"
840 #include "sched_fair.c"
841 #include "sched_rt.c"
842 #ifdef CONFIG_SCHED_DEBUG
843 # include "sched_debug.c"
846 #define sched_class_highest (&rt_sched_class)
849 * Update delta_exec, delta_fair fields for rq.
851 * delta_fair clock advances at a rate inversely proportional to
852 * total load (rq->load.weight) on the runqueue, while
853 * delta_exec advances at the same rate as wall-clock (provided
856 * delta_exec / delta_fair is a measure of the (smoothened) load on this
857 * runqueue over any given interval. This (smoothened) load is used
858 * during load balance.
860 * This function is called /before/ updating rq->load
861 * and when switching tasks.
863 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
865 update_load_add(&rq
->load
, p
->se
.load
.weight
);
868 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
870 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
873 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
879 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
885 static void set_load_weight(struct task_struct
*p
)
887 if (task_has_rt_policy(p
)) {
888 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
889 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
894 * SCHED_IDLE tasks get minimal weight:
896 if (p
->policy
== SCHED_IDLE
) {
897 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
898 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
902 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
903 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
906 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
908 sched_info_queued(p
);
909 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
913 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
915 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
920 * __normal_prio - return the priority that is based on the static prio
922 static inline int __normal_prio(struct task_struct
*p
)
924 return p
->static_prio
;
928 * Calculate the expected normal priority: i.e. priority
929 * without taking RT-inheritance into account. Might be
930 * boosted by interactivity modifiers. Changes upon fork,
931 * setprio syscalls, and whenever the interactivity
932 * estimator recalculates.
934 static inline int normal_prio(struct task_struct
*p
)
938 if (task_has_rt_policy(p
))
939 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
941 prio
= __normal_prio(p
);
946 * Calculate the current priority, i.e. the priority
947 * taken into account by the scheduler. This value might
948 * be boosted by RT tasks, or might be boosted by
949 * interactivity modifiers. Will be RT if the task got
950 * RT-boosted. If not then it returns p->normal_prio.
952 static int effective_prio(struct task_struct
*p
)
954 p
->normal_prio
= normal_prio(p
);
956 * If we are RT tasks or we were boosted to RT priority,
957 * keep the priority unchanged. Otherwise, update priority
958 * to the normal priority:
960 if (!rt_prio(p
->prio
))
961 return p
->normal_prio
;
966 * activate_task - move a task to the runqueue.
968 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
970 if (p
->state
== TASK_UNINTERRUPTIBLE
)
971 rq
->nr_uninterruptible
--;
973 enqueue_task(rq
, p
, wakeup
);
974 inc_nr_running(p
, rq
);
978 * deactivate_task - remove a task from the runqueue.
980 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
982 if (p
->state
== TASK_UNINTERRUPTIBLE
)
983 rq
->nr_uninterruptible
++;
985 dequeue_task(rq
, p
, sleep
);
986 dec_nr_running(p
, rq
);
990 * task_curr - is this task currently executing on a CPU?
991 * @p: the task in question.
993 inline int task_curr(const struct task_struct
*p
)
995 return cpu_curr(task_cpu(p
)) == p
;
998 /* Used instead of source_load when we know the type == 0 */
999 unsigned long weighted_cpuload(const int cpu
)
1001 return cpu_rq(cpu
)->load
.weight
;
1004 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1007 task_thread_info(p
)->cpu
= cpu
;
1014 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1016 int old_cpu
= task_cpu(p
);
1017 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1018 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1019 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1022 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1024 #ifdef CONFIG_SCHEDSTATS
1025 if (p
->se
.wait_start
)
1026 p
->se
.wait_start
-= clock_offset
;
1027 if (p
->se
.sleep_start
)
1028 p
->se
.sleep_start
-= clock_offset
;
1029 if (p
->se
.block_start
)
1030 p
->se
.block_start
-= clock_offset
;
1032 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1033 new_cfsrq
->min_vruntime
;
1035 __set_task_cpu(p
, new_cpu
);
1038 struct migration_req
{
1039 struct list_head list
;
1041 struct task_struct
*task
;
1044 struct completion done
;
1048 * The task's runqueue lock must be held.
1049 * Returns true if you have to wait for migration thread.
1052 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1054 struct rq
*rq
= task_rq(p
);
1057 * If the task is not on a runqueue (and not running), then
1058 * it is sufficient to simply update the task's cpu field.
1060 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1061 set_task_cpu(p
, dest_cpu
);
1065 init_completion(&req
->done
);
1067 req
->dest_cpu
= dest_cpu
;
1068 list_add(&req
->list
, &rq
->migration_queue
);
1074 * wait_task_inactive - wait for a thread to unschedule.
1076 * The caller must ensure that the task *will* unschedule sometime soon,
1077 * else this function might spin for a *long* time. This function can't
1078 * be called with interrupts off, or it may introduce deadlock with
1079 * smp_call_function() if an IPI is sent by the same process we are
1080 * waiting to become inactive.
1082 void wait_task_inactive(struct task_struct
*p
)
1084 unsigned long flags
;
1090 * We do the initial early heuristics without holding
1091 * any task-queue locks at all. We'll only try to get
1092 * the runqueue lock when things look like they will
1098 * If the task is actively running on another CPU
1099 * still, just relax and busy-wait without holding
1102 * NOTE! Since we don't hold any locks, it's not
1103 * even sure that "rq" stays as the right runqueue!
1104 * But we don't care, since "task_running()" will
1105 * return false if the runqueue has changed and p
1106 * is actually now running somewhere else!
1108 while (task_running(rq
, p
))
1112 * Ok, time to look more closely! We need the rq
1113 * lock now, to be *sure*. If we're wrong, we'll
1114 * just go back and repeat.
1116 rq
= task_rq_lock(p
, &flags
);
1117 running
= task_running(rq
, p
);
1118 on_rq
= p
->se
.on_rq
;
1119 task_rq_unlock(rq
, &flags
);
1122 * Was it really running after all now that we
1123 * checked with the proper locks actually held?
1125 * Oops. Go back and try again..
1127 if (unlikely(running
)) {
1133 * It's not enough that it's not actively running,
1134 * it must be off the runqueue _entirely_, and not
1137 * So if it wa still runnable (but just not actively
1138 * running right now), it's preempted, and we should
1139 * yield - it could be a while.
1141 if (unlikely(on_rq
)) {
1142 schedule_timeout_uninterruptible(1);
1147 * Ahh, all good. It wasn't running, and it wasn't
1148 * runnable, which means that it will never become
1149 * running in the future either. We're all done!
1156 * kick_process - kick a running thread to enter/exit the kernel
1157 * @p: the to-be-kicked thread
1159 * Cause a process which is running on another CPU to enter
1160 * kernel-mode, without any delay. (to get signals handled.)
1162 * NOTE: this function doesnt have to take the runqueue lock,
1163 * because all it wants to ensure is that the remote task enters
1164 * the kernel. If the IPI races and the task has been migrated
1165 * to another CPU then no harm is done and the purpose has been
1168 void kick_process(struct task_struct
*p
)
1174 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1175 smp_send_reschedule(cpu
);
1180 * Return a low guess at the load of a migration-source cpu weighted
1181 * according to the scheduling class and "nice" value.
1183 * We want to under-estimate the load of migration sources, to
1184 * balance conservatively.
1186 static unsigned long source_load(int cpu
, int type
)
1188 struct rq
*rq
= cpu_rq(cpu
);
1189 unsigned long total
= weighted_cpuload(cpu
);
1194 return min(rq
->cpu_load
[type
-1], total
);
1198 * Return a high guess at the load of a migration-target cpu weighted
1199 * according to the scheduling class and "nice" value.
1201 static unsigned long target_load(int cpu
, int type
)
1203 struct rq
*rq
= cpu_rq(cpu
);
1204 unsigned long total
= weighted_cpuload(cpu
);
1209 return max(rq
->cpu_load
[type
-1], total
);
1213 * Return the average load per task on the cpu's run queue
1215 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1217 struct rq
*rq
= cpu_rq(cpu
);
1218 unsigned long total
= weighted_cpuload(cpu
);
1219 unsigned long n
= rq
->nr_running
;
1221 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1225 * find_idlest_group finds and returns the least busy CPU group within the
1228 static struct sched_group
*
1229 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1231 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1232 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1233 int load_idx
= sd
->forkexec_idx
;
1234 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1237 unsigned long load
, avg_load
;
1241 /* Skip over this group if it has no CPUs allowed */
1242 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1245 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1247 /* Tally up the load of all CPUs in the group */
1250 for_each_cpu_mask(i
, group
->cpumask
) {
1251 /* Bias balancing toward cpus of our domain */
1253 load
= source_load(i
, load_idx
);
1255 load
= target_load(i
, load_idx
);
1260 /* Adjust by relative CPU power of the group */
1261 avg_load
= sg_div_cpu_power(group
,
1262 avg_load
* SCHED_LOAD_SCALE
);
1265 this_load
= avg_load
;
1267 } else if (avg_load
< min_load
) {
1268 min_load
= avg_load
;
1271 } while (group
= group
->next
, group
!= sd
->groups
);
1273 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1279 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1282 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1285 unsigned long load
, min_load
= ULONG_MAX
;
1289 /* Traverse only the allowed CPUs */
1290 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1292 for_each_cpu_mask(i
, tmp
) {
1293 load
= weighted_cpuload(i
);
1295 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1305 * sched_balance_self: balance the current task (running on cpu) in domains
1306 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1309 * Balance, ie. select the least loaded group.
1311 * Returns the target CPU number, or the same CPU if no balancing is needed.
1313 * preempt must be disabled.
1315 static int sched_balance_self(int cpu
, int flag
)
1317 struct task_struct
*t
= current
;
1318 struct sched_domain
*tmp
, *sd
= NULL
;
1320 for_each_domain(cpu
, tmp
) {
1322 * If power savings logic is enabled for a domain, stop there.
1324 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1326 if (tmp
->flags
& flag
)
1332 struct sched_group
*group
;
1333 int new_cpu
, weight
;
1335 if (!(sd
->flags
& flag
)) {
1341 group
= find_idlest_group(sd
, t
, cpu
);
1347 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1348 if (new_cpu
== -1 || new_cpu
== cpu
) {
1349 /* Now try balancing at a lower domain level of cpu */
1354 /* Now try balancing at a lower domain level of new_cpu */
1357 weight
= cpus_weight(span
);
1358 for_each_domain(cpu
, tmp
) {
1359 if (weight
<= cpus_weight(tmp
->span
))
1361 if (tmp
->flags
& flag
)
1364 /* while loop will break here if sd == NULL */
1370 #endif /* CONFIG_SMP */
1373 * wake_idle() will wake a task on an idle cpu if task->cpu is
1374 * not idle and an idle cpu is available. The span of cpus to
1375 * search starts with cpus closest then further out as needed,
1376 * so we always favor a closer, idle cpu.
1378 * Returns the CPU we should wake onto.
1380 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1381 static int wake_idle(int cpu
, struct task_struct
*p
)
1384 struct sched_domain
*sd
;
1388 * If it is idle, then it is the best cpu to run this task.
1390 * This cpu is also the best, if it has more than one task already.
1391 * Siblings must be also busy(in most cases) as they didn't already
1392 * pickup the extra load from this cpu and hence we need not check
1393 * sibling runqueue info. This will avoid the checks and cache miss
1394 * penalities associated with that.
1396 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1399 for_each_domain(cpu
, sd
) {
1400 if (sd
->flags
& SD_WAKE_IDLE
) {
1401 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1402 for_each_cpu_mask(i
, tmp
) {
1413 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1420 * try_to_wake_up - wake up a thread
1421 * @p: the to-be-woken-up thread
1422 * @state: the mask of task states that can be woken
1423 * @sync: do a synchronous wakeup?
1425 * Put it on the run-queue if it's not already there. The "current"
1426 * thread is always on the run-queue (except when the actual
1427 * re-schedule is in progress), and as such you're allowed to do
1428 * the simpler "current->state = TASK_RUNNING" to mark yourself
1429 * runnable without the overhead of this.
1431 * returns failure only if the task is already active.
1433 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1435 int cpu
, this_cpu
, success
= 0;
1436 unsigned long flags
;
1440 struct sched_domain
*sd
, *this_sd
= NULL
;
1441 unsigned long load
, this_load
;
1445 rq
= task_rq_lock(p
, &flags
);
1446 old_state
= p
->state
;
1447 if (!(old_state
& state
))
1454 this_cpu
= smp_processor_id();
1457 if (unlikely(task_running(rq
, p
)))
1462 schedstat_inc(rq
, ttwu_count
);
1463 if (cpu
== this_cpu
) {
1464 schedstat_inc(rq
, ttwu_local
);
1468 for_each_domain(this_cpu
, sd
) {
1469 if (cpu_isset(cpu
, sd
->span
)) {
1470 schedstat_inc(sd
, ttwu_wake_remote
);
1476 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1480 * Check for affine wakeup and passive balancing possibilities.
1483 int idx
= this_sd
->wake_idx
;
1484 unsigned int imbalance
;
1486 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1488 load
= source_load(cpu
, idx
);
1489 this_load
= target_load(this_cpu
, idx
);
1491 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1493 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1494 unsigned long tl
= this_load
;
1495 unsigned long tl_per_task
;
1497 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1500 * If sync wakeup then subtract the (maximum possible)
1501 * effect of the currently running task from the load
1502 * of the current CPU:
1505 tl
-= current
->se
.load
.weight
;
1508 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1509 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1511 * This domain has SD_WAKE_AFFINE and
1512 * p is cache cold in this domain, and
1513 * there is no bad imbalance.
1515 schedstat_inc(this_sd
, ttwu_move_affine
);
1521 * Start passive balancing when half the imbalance_pct
1524 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1525 if (imbalance
*this_load
<= 100*load
) {
1526 schedstat_inc(this_sd
, ttwu_move_balance
);
1532 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1534 new_cpu
= wake_idle(new_cpu
, p
);
1535 if (new_cpu
!= cpu
) {
1536 set_task_cpu(p
, new_cpu
);
1537 task_rq_unlock(rq
, &flags
);
1538 /* might preempt at this point */
1539 rq
= task_rq_lock(p
, &flags
);
1540 old_state
= p
->state
;
1541 if (!(old_state
& state
))
1546 this_cpu
= smp_processor_id();
1551 #endif /* CONFIG_SMP */
1552 update_rq_clock(rq
);
1553 activate_task(rq
, p
, 1);
1555 * Sync wakeups (i.e. those types of wakeups where the waker
1556 * has indicated that it will leave the CPU in short order)
1557 * don't trigger a preemption, if the woken up task will run on
1558 * this cpu. (in this case the 'I will reschedule' promise of
1559 * the waker guarantees that the freshly woken up task is going
1560 * to be considered on this CPU.)
1562 if (!sync
|| cpu
!= this_cpu
)
1563 check_preempt_curr(rq
, p
);
1567 p
->state
= TASK_RUNNING
;
1569 task_rq_unlock(rq
, &flags
);
1574 int fastcall
wake_up_process(struct task_struct
*p
)
1576 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1577 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1579 EXPORT_SYMBOL(wake_up_process
);
1581 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1583 return try_to_wake_up(p
, state
, 0);
1587 * Perform scheduler related setup for a newly forked process p.
1588 * p is forked by current.
1590 * __sched_fork() is basic setup used by init_idle() too:
1592 static void __sched_fork(struct task_struct
*p
)
1594 p
->se
.exec_start
= 0;
1595 p
->se
.sum_exec_runtime
= 0;
1596 p
->se
.prev_sum_exec_runtime
= 0;
1598 #ifdef CONFIG_SCHEDSTATS
1599 p
->se
.wait_start
= 0;
1600 p
->se
.sum_sleep_runtime
= 0;
1601 p
->se
.sleep_start
= 0;
1602 p
->se
.block_start
= 0;
1603 p
->se
.sleep_max
= 0;
1604 p
->se
.block_max
= 0;
1606 p
->se
.slice_max
= 0;
1610 INIT_LIST_HEAD(&p
->run_list
);
1613 #ifdef CONFIG_PREEMPT_NOTIFIERS
1614 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1618 * We mark the process as running here, but have not actually
1619 * inserted it onto the runqueue yet. This guarantees that
1620 * nobody will actually run it, and a signal or other external
1621 * event cannot wake it up and insert it on the runqueue either.
1623 p
->state
= TASK_RUNNING
;
1627 * fork()/clone()-time setup:
1629 void sched_fork(struct task_struct
*p
, int clone_flags
)
1631 int cpu
= get_cpu();
1636 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1638 set_task_cpu(p
, cpu
);
1641 * Make sure we do not leak PI boosting priority to the child:
1643 p
->prio
= current
->normal_prio
;
1644 if (!rt_prio(p
->prio
))
1645 p
->sched_class
= &fair_sched_class
;
1647 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1648 if (likely(sched_info_on()))
1649 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1651 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1654 #ifdef CONFIG_PREEMPT
1655 /* Want to start with kernel preemption disabled. */
1656 task_thread_info(p
)->preempt_count
= 1;
1662 * wake_up_new_task - wake up a newly created task for the first time.
1664 * This function will do some initial scheduler statistics housekeeping
1665 * that must be done for every newly created context, then puts the task
1666 * on the runqueue and wakes it.
1668 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1670 unsigned long flags
;
1673 rq
= task_rq_lock(p
, &flags
);
1674 BUG_ON(p
->state
!= TASK_RUNNING
);
1675 update_rq_clock(rq
);
1677 p
->prio
= effective_prio(p
);
1679 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
|| !rq
->cfs
.curr
) {
1680 activate_task(rq
, p
, 0);
1683 * Let the scheduling class do new task startup
1684 * management (if any):
1686 p
->sched_class
->task_new(rq
, p
);
1687 inc_nr_running(p
, rq
);
1689 check_preempt_curr(rq
, p
);
1690 task_rq_unlock(rq
, &flags
);
1693 #ifdef CONFIG_PREEMPT_NOTIFIERS
1696 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1697 * @notifier: notifier struct to register
1699 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1701 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1703 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1706 * preempt_notifier_unregister - no longer interested in preemption notifications
1707 * @notifier: notifier struct to unregister
1709 * This is safe to call from within a preemption notifier.
1711 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1713 hlist_del(¬ifier
->link
);
1715 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1717 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1719 struct preempt_notifier
*notifier
;
1720 struct hlist_node
*node
;
1722 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1723 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1727 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1728 struct task_struct
*next
)
1730 struct preempt_notifier
*notifier
;
1731 struct hlist_node
*node
;
1733 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1734 notifier
->ops
->sched_out(notifier
, next
);
1739 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1744 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1745 struct task_struct
*next
)
1752 * prepare_task_switch - prepare to switch tasks
1753 * @rq: the runqueue preparing to switch
1754 * @prev: the current task that is being switched out
1755 * @next: the task we are going to switch to.
1757 * This is called with the rq lock held and interrupts off. It must
1758 * be paired with a subsequent finish_task_switch after the context
1761 * prepare_task_switch sets up locking and calls architecture specific
1765 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1766 struct task_struct
*next
)
1768 fire_sched_out_preempt_notifiers(prev
, next
);
1769 prepare_lock_switch(rq
, next
);
1770 prepare_arch_switch(next
);
1774 * finish_task_switch - clean up after a task-switch
1775 * @rq: runqueue associated with task-switch
1776 * @prev: the thread we just switched away from.
1778 * finish_task_switch must be called after the context switch, paired
1779 * with a prepare_task_switch call before the context switch.
1780 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1781 * and do any other architecture-specific cleanup actions.
1783 * Note that we may have delayed dropping an mm in context_switch(). If
1784 * so, we finish that here outside of the runqueue lock. (Doing it
1785 * with the lock held can cause deadlocks; see schedule() for
1788 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1789 __releases(rq
->lock
)
1791 struct mm_struct
*mm
= rq
->prev_mm
;
1797 * A task struct has one reference for the use as "current".
1798 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1799 * schedule one last time. The schedule call will never return, and
1800 * the scheduled task must drop that reference.
1801 * The test for TASK_DEAD must occur while the runqueue locks are
1802 * still held, otherwise prev could be scheduled on another cpu, die
1803 * there before we look at prev->state, and then the reference would
1805 * Manfred Spraul <manfred@colorfullife.com>
1807 prev_state
= prev
->state
;
1808 finish_arch_switch(prev
);
1809 finish_lock_switch(rq
, prev
);
1810 fire_sched_in_preempt_notifiers(current
);
1813 if (unlikely(prev_state
== TASK_DEAD
)) {
1815 * Remove function-return probe instances associated with this
1816 * task and put them back on the free list.
1818 kprobe_flush_task(prev
);
1819 put_task_struct(prev
);
1824 * schedule_tail - first thing a freshly forked thread must call.
1825 * @prev: the thread we just switched away from.
1827 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1828 __releases(rq
->lock
)
1830 struct rq
*rq
= this_rq();
1832 finish_task_switch(rq
, prev
);
1833 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1834 /* In this case, finish_task_switch does not reenable preemption */
1837 if (current
->set_child_tid
)
1838 put_user(current
->pid
, current
->set_child_tid
);
1842 * context_switch - switch to the new MM and the new
1843 * thread's register state.
1846 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1847 struct task_struct
*next
)
1849 struct mm_struct
*mm
, *oldmm
;
1851 prepare_task_switch(rq
, prev
, next
);
1853 oldmm
= prev
->active_mm
;
1855 * For paravirt, this is coupled with an exit in switch_to to
1856 * combine the page table reload and the switch backend into
1859 arch_enter_lazy_cpu_mode();
1861 if (unlikely(!mm
)) {
1862 next
->active_mm
= oldmm
;
1863 atomic_inc(&oldmm
->mm_count
);
1864 enter_lazy_tlb(oldmm
, next
);
1866 switch_mm(oldmm
, mm
, next
);
1868 if (unlikely(!prev
->mm
)) {
1869 prev
->active_mm
= NULL
;
1870 rq
->prev_mm
= oldmm
;
1873 * Since the runqueue lock will be released by the next
1874 * task (which is an invalid locking op but in the case
1875 * of the scheduler it's an obvious special-case), so we
1876 * do an early lockdep release here:
1878 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1879 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1882 /* Here we just switch the register state and the stack. */
1883 switch_to(prev
, next
, prev
);
1887 * this_rq must be evaluated again because prev may have moved
1888 * CPUs since it called schedule(), thus the 'rq' on its stack
1889 * frame will be invalid.
1891 finish_task_switch(this_rq(), prev
);
1895 * nr_running, nr_uninterruptible and nr_context_switches:
1897 * externally visible scheduler statistics: current number of runnable
1898 * threads, current number of uninterruptible-sleeping threads, total
1899 * number of context switches performed since bootup.
1901 unsigned long nr_running(void)
1903 unsigned long i
, sum
= 0;
1905 for_each_online_cpu(i
)
1906 sum
+= cpu_rq(i
)->nr_running
;
1911 unsigned long nr_uninterruptible(void)
1913 unsigned long i
, sum
= 0;
1915 for_each_possible_cpu(i
)
1916 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1919 * Since we read the counters lockless, it might be slightly
1920 * inaccurate. Do not allow it to go below zero though:
1922 if (unlikely((long)sum
< 0))
1928 unsigned long long nr_context_switches(void)
1931 unsigned long long sum
= 0;
1933 for_each_possible_cpu(i
)
1934 sum
+= cpu_rq(i
)->nr_switches
;
1939 unsigned long nr_iowait(void)
1941 unsigned long i
, sum
= 0;
1943 for_each_possible_cpu(i
)
1944 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1949 unsigned long nr_active(void)
1951 unsigned long i
, running
= 0, uninterruptible
= 0;
1953 for_each_online_cpu(i
) {
1954 running
+= cpu_rq(i
)->nr_running
;
1955 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1958 if (unlikely((long)uninterruptible
< 0))
1959 uninterruptible
= 0;
1961 return running
+ uninterruptible
;
1965 * Update rq->cpu_load[] statistics. This function is usually called every
1966 * scheduler tick (TICK_NSEC).
1968 static void update_cpu_load(struct rq
*this_rq
)
1970 unsigned long this_load
= this_rq
->load
.weight
;
1973 this_rq
->nr_load_updates
++;
1975 /* Update our load: */
1976 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1977 unsigned long old_load
, new_load
;
1979 /* scale is effectively 1 << i now, and >> i divides by scale */
1981 old_load
= this_rq
->cpu_load
[i
];
1982 new_load
= this_load
;
1984 * Round up the averaging division if load is increasing. This
1985 * prevents us from getting stuck on 9 if the load is 10, for
1988 if (new_load
> old_load
)
1989 new_load
+= scale
-1;
1990 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
1997 * double_rq_lock - safely lock two runqueues
1999 * Note this does not disable interrupts like task_rq_lock,
2000 * you need to do so manually before calling.
2002 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2003 __acquires(rq1
->lock
)
2004 __acquires(rq2
->lock
)
2006 BUG_ON(!irqs_disabled());
2008 spin_lock(&rq1
->lock
);
2009 __acquire(rq2
->lock
); /* Fake it out ;) */
2012 spin_lock(&rq1
->lock
);
2013 spin_lock(&rq2
->lock
);
2015 spin_lock(&rq2
->lock
);
2016 spin_lock(&rq1
->lock
);
2019 update_rq_clock(rq1
);
2020 update_rq_clock(rq2
);
2024 * double_rq_unlock - safely unlock two runqueues
2026 * Note this does not restore interrupts like task_rq_unlock,
2027 * you need to do so manually after calling.
2029 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2030 __releases(rq1
->lock
)
2031 __releases(rq2
->lock
)
2033 spin_unlock(&rq1
->lock
);
2035 spin_unlock(&rq2
->lock
);
2037 __release(rq2
->lock
);
2041 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2043 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2044 __releases(this_rq
->lock
)
2045 __acquires(busiest
->lock
)
2046 __acquires(this_rq
->lock
)
2048 if (unlikely(!irqs_disabled())) {
2049 /* printk() doesn't work good under rq->lock */
2050 spin_unlock(&this_rq
->lock
);
2053 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2054 if (busiest
< this_rq
) {
2055 spin_unlock(&this_rq
->lock
);
2056 spin_lock(&busiest
->lock
);
2057 spin_lock(&this_rq
->lock
);
2059 spin_lock(&busiest
->lock
);
2064 * If dest_cpu is allowed for this process, migrate the task to it.
2065 * This is accomplished by forcing the cpu_allowed mask to only
2066 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2067 * the cpu_allowed mask is restored.
2069 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2071 struct migration_req req
;
2072 unsigned long flags
;
2075 rq
= task_rq_lock(p
, &flags
);
2076 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2077 || unlikely(cpu_is_offline(dest_cpu
)))
2080 /* force the process onto the specified CPU */
2081 if (migrate_task(p
, dest_cpu
, &req
)) {
2082 /* Need to wait for migration thread (might exit: take ref). */
2083 struct task_struct
*mt
= rq
->migration_thread
;
2085 get_task_struct(mt
);
2086 task_rq_unlock(rq
, &flags
);
2087 wake_up_process(mt
);
2088 put_task_struct(mt
);
2089 wait_for_completion(&req
.done
);
2094 task_rq_unlock(rq
, &flags
);
2098 * sched_exec - execve() is a valuable balancing opportunity, because at
2099 * this point the task has the smallest effective memory and cache footprint.
2101 void sched_exec(void)
2103 int new_cpu
, this_cpu
= get_cpu();
2104 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2106 if (new_cpu
!= this_cpu
)
2107 sched_migrate_task(current
, new_cpu
);
2111 * pull_task - move a task from a remote runqueue to the local runqueue.
2112 * Both runqueues must be locked.
2114 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2115 struct rq
*this_rq
, int this_cpu
)
2117 deactivate_task(src_rq
, p
, 0);
2118 set_task_cpu(p
, this_cpu
);
2119 activate_task(this_rq
, p
, 0);
2121 * Note that idle threads have a prio of MAX_PRIO, for this test
2122 * to be always true for them.
2124 check_preempt_curr(this_rq
, p
);
2128 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2131 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2132 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2136 * We do not migrate tasks that are:
2137 * 1) running (obviously), or
2138 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2139 * 3) are cache-hot on their current CPU.
2141 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2145 if (task_running(rq
, p
))
2151 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2152 unsigned long max_nr_move
, unsigned long max_load_move
,
2153 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2154 int *all_pinned
, unsigned long *load_moved
,
2155 int *this_best_prio
, struct rq_iterator
*iterator
)
2157 int pulled
= 0, pinned
= 0, skip_for_load
;
2158 struct task_struct
*p
;
2159 long rem_load_move
= max_load_move
;
2161 if (max_nr_move
== 0 || max_load_move
== 0)
2167 * Start the load-balancing iterator:
2169 p
= iterator
->start(iterator
->arg
);
2174 * To help distribute high priority tasks accross CPUs we don't
2175 * skip a task if it will be the highest priority task (i.e. smallest
2176 * prio value) on its new queue regardless of its load weight
2178 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2179 SCHED_LOAD_SCALE_FUZZ
;
2180 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2181 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2182 p
= iterator
->next(iterator
->arg
);
2186 pull_task(busiest
, p
, this_rq
, this_cpu
);
2188 rem_load_move
-= p
->se
.load
.weight
;
2191 * We only want to steal up to the prescribed number of tasks
2192 * and the prescribed amount of weighted load.
2194 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2195 if (p
->prio
< *this_best_prio
)
2196 *this_best_prio
= p
->prio
;
2197 p
= iterator
->next(iterator
->arg
);
2202 * Right now, this is the only place pull_task() is called,
2203 * so we can safely collect pull_task() stats here rather than
2204 * inside pull_task().
2206 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2209 *all_pinned
= pinned
;
2210 *load_moved
= max_load_move
- rem_load_move
;
2215 * move_tasks tries to move up to max_load_move weighted load from busiest to
2216 * this_rq, as part of a balancing operation within domain "sd".
2217 * Returns 1 if successful and 0 otherwise.
2219 * Called with both runqueues locked.
2221 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2222 unsigned long max_load_move
,
2223 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2226 const struct sched_class
*class = sched_class_highest
;
2227 unsigned long total_load_moved
= 0;
2228 int this_best_prio
= this_rq
->curr
->prio
;
2232 class->load_balance(this_rq
, this_cpu
, busiest
,
2233 ULONG_MAX
, max_load_move
- total_load_moved
,
2234 sd
, idle
, all_pinned
, &this_best_prio
);
2235 class = class->next
;
2236 } while (class && max_load_move
> total_load_moved
);
2238 return total_load_moved
> 0;
2242 * move_one_task tries to move exactly one task from busiest to this_rq, as
2243 * part of active balancing operations within "domain".
2244 * Returns 1 if successful and 0 otherwise.
2246 * Called with both runqueues locked.
2248 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2249 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2251 const struct sched_class
*class;
2252 int this_best_prio
= MAX_PRIO
;
2254 for (class = sched_class_highest
; class; class = class->next
)
2255 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2256 1, ULONG_MAX
, sd
, idle
, NULL
,
2264 * find_busiest_group finds and returns the busiest CPU group within the
2265 * domain. It calculates and returns the amount of weighted load which
2266 * should be moved to restore balance via the imbalance parameter.
2268 static struct sched_group
*
2269 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2270 unsigned long *imbalance
, enum cpu_idle_type idle
,
2271 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2273 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2274 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2275 unsigned long max_pull
;
2276 unsigned long busiest_load_per_task
, busiest_nr_running
;
2277 unsigned long this_load_per_task
, this_nr_running
;
2279 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2280 int power_savings_balance
= 1;
2281 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2282 unsigned long min_nr_running
= ULONG_MAX
;
2283 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2286 max_load
= this_load
= total_load
= total_pwr
= 0;
2287 busiest_load_per_task
= busiest_nr_running
= 0;
2288 this_load_per_task
= this_nr_running
= 0;
2289 if (idle
== CPU_NOT_IDLE
)
2290 load_idx
= sd
->busy_idx
;
2291 else if (idle
== CPU_NEWLY_IDLE
)
2292 load_idx
= sd
->newidle_idx
;
2294 load_idx
= sd
->idle_idx
;
2297 unsigned long load
, group_capacity
;
2300 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2301 unsigned long sum_nr_running
, sum_weighted_load
;
2303 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2306 balance_cpu
= first_cpu(group
->cpumask
);
2308 /* Tally up the load of all CPUs in the group */
2309 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2311 for_each_cpu_mask(i
, group
->cpumask
) {
2314 if (!cpu_isset(i
, *cpus
))
2319 if (*sd_idle
&& rq
->nr_running
)
2322 /* Bias balancing toward cpus of our domain */
2324 if (idle_cpu(i
) && !first_idle_cpu
) {
2329 load
= target_load(i
, load_idx
);
2331 load
= source_load(i
, load_idx
);
2334 sum_nr_running
+= rq
->nr_running
;
2335 sum_weighted_load
+= weighted_cpuload(i
);
2339 * First idle cpu or the first cpu(busiest) in this sched group
2340 * is eligible for doing load balancing at this and above
2341 * domains. In the newly idle case, we will allow all the cpu's
2342 * to do the newly idle load balance.
2344 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2345 balance_cpu
!= this_cpu
&& balance
) {
2350 total_load
+= avg_load
;
2351 total_pwr
+= group
->__cpu_power
;
2353 /* Adjust by relative CPU power of the group */
2354 avg_load
= sg_div_cpu_power(group
,
2355 avg_load
* SCHED_LOAD_SCALE
);
2357 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2360 this_load
= avg_load
;
2362 this_nr_running
= sum_nr_running
;
2363 this_load_per_task
= sum_weighted_load
;
2364 } else if (avg_load
> max_load
&&
2365 sum_nr_running
> group_capacity
) {
2366 max_load
= avg_load
;
2368 busiest_nr_running
= sum_nr_running
;
2369 busiest_load_per_task
= sum_weighted_load
;
2372 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2374 * Busy processors will not participate in power savings
2377 if (idle
== CPU_NOT_IDLE
||
2378 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2382 * If the local group is idle or completely loaded
2383 * no need to do power savings balance at this domain
2385 if (local_group
&& (this_nr_running
>= group_capacity
||
2387 power_savings_balance
= 0;
2390 * If a group is already running at full capacity or idle,
2391 * don't include that group in power savings calculations
2393 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2398 * Calculate the group which has the least non-idle load.
2399 * This is the group from where we need to pick up the load
2402 if ((sum_nr_running
< min_nr_running
) ||
2403 (sum_nr_running
== min_nr_running
&&
2404 first_cpu(group
->cpumask
) <
2405 first_cpu(group_min
->cpumask
))) {
2407 min_nr_running
= sum_nr_running
;
2408 min_load_per_task
= sum_weighted_load
/
2413 * Calculate the group which is almost near its
2414 * capacity but still has some space to pick up some load
2415 * from other group and save more power
2417 if (sum_nr_running
<= group_capacity
- 1) {
2418 if (sum_nr_running
> leader_nr_running
||
2419 (sum_nr_running
== leader_nr_running
&&
2420 first_cpu(group
->cpumask
) >
2421 first_cpu(group_leader
->cpumask
))) {
2422 group_leader
= group
;
2423 leader_nr_running
= sum_nr_running
;
2428 group
= group
->next
;
2429 } while (group
!= sd
->groups
);
2431 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2434 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2436 if (this_load
>= avg_load
||
2437 100*max_load
<= sd
->imbalance_pct
*this_load
)
2440 busiest_load_per_task
/= busiest_nr_running
;
2442 * We're trying to get all the cpus to the average_load, so we don't
2443 * want to push ourselves above the average load, nor do we wish to
2444 * reduce the max loaded cpu below the average load, as either of these
2445 * actions would just result in more rebalancing later, and ping-pong
2446 * tasks around. Thus we look for the minimum possible imbalance.
2447 * Negative imbalances (*we* are more loaded than anyone else) will
2448 * be counted as no imbalance for these purposes -- we can't fix that
2449 * by pulling tasks to us. Be careful of negative numbers as they'll
2450 * appear as very large values with unsigned longs.
2452 if (max_load
<= busiest_load_per_task
)
2456 * In the presence of smp nice balancing, certain scenarios can have
2457 * max load less than avg load(as we skip the groups at or below
2458 * its cpu_power, while calculating max_load..)
2460 if (max_load
< avg_load
) {
2462 goto small_imbalance
;
2465 /* Don't want to pull so many tasks that a group would go idle */
2466 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2468 /* How much load to actually move to equalise the imbalance */
2469 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2470 (avg_load
- this_load
) * this->__cpu_power
)
2474 * if *imbalance is less than the average load per runnable task
2475 * there is no gaurantee that any tasks will be moved so we'll have
2476 * a think about bumping its value to force at least one task to be
2479 if (*imbalance
< busiest_load_per_task
) {
2480 unsigned long tmp
, pwr_now
, pwr_move
;
2484 pwr_move
= pwr_now
= 0;
2486 if (this_nr_running
) {
2487 this_load_per_task
/= this_nr_running
;
2488 if (busiest_load_per_task
> this_load_per_task
)
2491 this_load_per_task
= SCHED_LOAD_SCALE
;
2493 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2494 busiest_load_per_task
* imbn
) {
2495 *imbalance
= busiest_load_per_task
;
2500 * OK, we don't have enough imbalance to justify moving tasks,
2501 * however we may be able to increase total CPU power used by
2505 pwr_now
+= busiest
->__cpu_power
*
2506 min(busiest_load_per_task
, max_load
);
2507 pwr_now
+= this->__cpu_power
*
2508 min(this_load_per_task
, this_load
);
2509 pwr_now
/= SCHED_LOAD_SCALE
;
2511 /* Amount of load we'd subtract */
2512 tmp
= sg_div_cpu_power(busiest
,
2513 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2515 pwr_move
+= busiest
->__cpu_power
*
2516 min(busiest_load_per_task
, max_load
- tmp
);
2518 /* Amount of load we'd add */
2519 if (max_load
* busiest
->__cpu_power
<
2520 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2521 tmp
= sg_div_cpu_power(this,
2522 max_load
* busiest
->__cpu_power
);
2524 tmp
= sg_div_cpu_power(this,
2525 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2526 pwr_move
+= this->__cpu_power
*
2527 min(this_load_per_task
, this_load
+ tmp
);
2528 pwr_move
/= SCHED_LOAD_SCALE
;
2530 /* Move if we gain throughput */
2531 if (pwr_move
> pwr_now
)
2532 *imbalance
= busiest_load_per_task
;
2538 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2539 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2542 if (this == group_leader
&& group_leader
!= group_min
) {
2543 *imbalance
= min_load_per_task
;
2553 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2556 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2557 unsigned long imbalance
, cpumask_t
*cpus
)
2559 struct rq
*busiest
= NULL
, *rq
;
2560 unsigned long max_load
= 0;
2563 for_each_cpu_mask(i
, group
->cpumask
) {
2566 if (!cpu_isset(i
, *cpus
))
2570 wl
= weighted_cpuload(i
);
2572 if (rq
->nr_running
== 1 && wl
> imbalance
)
2575 if (wl
> max_load
) {
2585 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2586 * so long as it is large enough.
2588 #define MAX_PINNED_INTERVAL 512
2591 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2592 * tasks if there is an imbalance.
2594 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2595 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2598 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2599 struct sched_group
*group
;
2600 unsigned long imbalance
;
2602 cpumask_t cpus
= CPU_MASK_ALL
;
2603 unsigned long flags
;
2606 * When power savings policy is enabled for the parent domain, idle
2607 * sibling can pick up load irrespective of busy siblings. In this case,
2608 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2609 * portraying it as CPU_NOT_IDLE.
2611 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2612 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2615 schedstat_inc(sd
, lb_count
[idle
]);
2618 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2625 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2629 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2631 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2635 BUG_ON(busiest
== this_rq
);
2637 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2640 if (busiest
->nr_running
> 1) {
2642 * Attempt to move tasks. If find_busiest_group has found
2643 * an imbalance but busiest->nr_running <= 1, the group is
2644 * still unbalanced. ld_moved simply stays zero, so it is
2645 * correctly treated as an imbalance.
2647 local_irq_save(flags
);
2648 double_rq_lock(this_rq
, busiest
);
2649 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2650 imbalance
, sd
, idle
, &all_pinned
);
2651 double_rq_unlock(this_rq
, busiest
);
2652 local_irq_restore(flags
);
2655 * some other cpu did the load balance for us.
2657 if (ld_moved
&& this_cpu
!= smp_processor_id())
2658 resched_cpu(this_cpu
);
2660 /* All tasks on this runqueue were pinned by CPU affinity */
2661 if (unlikely(all_pinned
)) {
2662 cpu_clear(cpu_of(busiest
), cpus
);
2663 if (!cpus_empty(cpus
))
2670 schedstat_inc(sd
, lb_failed
[idle
]);
2671 sd
->nr_balance_failed
++;
2673 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2675 spin_lock_irqsave(&busiest
->lock
, flags
);
2677 /* don't kick the migration_thread, if the curr
2678 * task on busiest cpu can't be moved to this_cpu
2680 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2681 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2683 goto out_one_pinned
;
2686 if (!busiest
->active_balance
) {
2687 busiest
->active_balance
= 1;
2688 busiest
->push_cpu
= this_cpu
;
2691 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2693 wake_up_process(busiest
->migration_thread
);
2696 * We've kicked active balancing, reset the failure
2699 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2702 sd
->nr_balance_failed
= 0;
2704 if (likely(!active_balance
)) {
2705 /* We were unbalanced, so reset the balancing interval */
2706 sd
->balance_interval
= sd
->min_interval
;
2709 * If we've begun active balancing, start to back off. This
2710 * case may not be covered by the all_pinned logic if there
2711 * is only 1 task on the busy runqueue (because we don't call
2714 if (sd
->balance_interval
< sd
->max_interval
)
2715 sd
->balance_interval
*= 2;
2718 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2719 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2724 schedstat_inc(sd
, lb_balanced
[idle
]);
2726 sd
->nr_balance_failed
= 0;
2729 /* tune up the balancing interval */
2730 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2731 (sd
->balance_interval
< sd
->max_interval
))
2732 sd
->balance_interval
*= 2;
2734 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2735 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2741 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2742 * tasks if there is an imbalance.
2744 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2745 * this_rq is locked.
2748 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2750 struct sched_group
*group
;
2751 struct rq
*busiest
= NULL
;
2752 unsigned long imbalance
;
2756 cpumask_t cpus
= CPU_MASK_ALL
;
2759 * When power savings policy is enabled for the parent domain, idle
2760 * sibling can pick up load irrespective of busy siblings. In this case,
2761 * let the state of idle sibling percolate up as IDLE, instead of
2762 * portraying it as CPU_NOT_IDLE.
2764 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2765 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2768 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2770 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2771 &sd_idle
, &cpus
, NULL
);
2773 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2777 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2780 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2784 BUG_ON(busiest
== this_rq
);
2786 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2789 if (busiest
->nr_running
> 1) {
2790 /* Attempt to move tasks */
2791 double_lock_balance(this_rq
, busiest
);
2792 /* this_rq->clock is already updated */
2793 update_rq_clock(busiest
);
2794 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2795 imbalance
, sd
, CPU_NEWLY_IDLE
,
2797 spin_unlock(&busiest
->lock
);
2799 if (unlikely(all_pinned
)) {
2800 cpu_clear(cpu_of(busiest
), cpus
);
2801 if (!cpus_empty(cpus
))
2807 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2808 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2809 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2812 sd
->nr_balance_failed
= 0;
2817 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2818 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2819 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2821 sd
->nr_balance_failed
= 0;
2827 * idle_balance is called by schedule() if this_cpu is about to become
2828 * idle. Attempts to pull tasks from other CPUs.
2830 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2832 struct sched_domain
*sd
;
2833 int pulled_task
= -1;
2834 unsigned long next_balance
= jiffies
+ HZ
;
2836 for_each_domain(this_cpu
, sd
) {
2837 unsigned long interval
;
2839 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2842 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2843 /* If we've pulled tasks over stop searching: */
2844 pulled_task
= load_balance_newidle(this_cpu
,
2847 interval
= msecs_to_jiffies(sd
->balance_interval
);
2848 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2849 next_balance
= sd
->last_balance
+ interval
;
2853 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2855 * We are going idle. next_balance may be set based on
2856 * a busy processor. So reset next_balance.
2858 this_rq
->next_balance
= next_balance
;
2863 * active_load_balance is run by migration threads. It pushes running tasks
2864 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2865 * running on each physical CPU where possible, and avoids physical /
2866 * logical imbalances.
2868 * Called with busiest_rq locked.
2870 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2872 int target_cpu
= busiest_rq
->push_cpu
;
2873 struct sched_domain
*sd
;
2874 struct rq
*target_rq
;
2876 /* Is there any task to move? */
2877 if (busiest_rq
->nr_running
<= 1)
2880 target_rq
= cpu_rq(target_cpu
);
2883 * This condition is "impossible", if it occurs
2884 * we need to fix it. Originally reported by
2885 * Bjorn Helgaas on a 128-cpu setup.
2887 BUG_ON(busiest_rq
== target_rq
);
2889 /* move a task from busiest_rq to target_rq */
2890 double_lock_balance(busiest_rq
, target_rq
);
2891 update_rq_clock(busiest_rq
);
2892 update_rq_clock(target_rq
);
2894 /* Search for an sd spanning us and the target CPU. */
2895 for_each_domain(target_cpu
, sd
) {
2896 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2897 cpu_isset(busiest_cpu
, sd
->span
))
2902 schedstat_inc(sd
, alb_count
);
2904 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2906 schedstat_inc(sd
, alb_pushed
);
2908 schedstat_inc(sd
, alb_failed
);
2910 spin_unlock(&target_rq
->lock
);
2915 atomic_t load_balancer
;
2917 } nohz ____cacheline_aligned
= {
2918 .load_balancer
= ATOMIC_INIT(-1),
2919 .cpu_mask
= CPU_MASK_NONE
,
2923 * This routine will try to nominate the ilb (idle load balancing)
2924 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2925 * load balancing on behalf of all those cpus. If all the cpus in the system
2926 * go into this tickless mode, then there will be no ilb owner (as there is
2927 * no need for one) and all the cpus will sleep till the next wakeup event
2930 * For the ilb owner, tick is not stopped. And this tick will be used
2931 * for idle load balancing. ilb owner will still be part of
2934 * While stopping the tick, this cpu will become the ilb owner if there
2935 * is no other owner. And will be the owner till that cpu becomes busy
2936 * or if all cpus in the system stop their ticks at which point
2937 * there is no need for ilb owner.
2939 * When the ilb owner becomes busy, it nominates another owner, during the
2940 * next busy scheduler_tick()
2942 int select_nohz_load_balancer(int stop_tick
)
2944 int cpu
= smp_processor_id();
2947 cpu_set(cpu
, nohz
.cpu_mask
);
2948 cpu_rq(cpu
)->in_nohz_recently
= 1;
2951 * If we are going offline and still the leader, give up!
2953 if (cpu_is_offline(cpu
) &&
2954 atomic_read(&nohz
.load_balancer
) == cpu
) {
2955 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2960 /* time for ilb owner also to sleep */
2961 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2962 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2963 atomic_set(&nohz
.load_balancer
, -1);
2967 if (atomic_read(&nohz
.load_balancer
) == -1) {
2968 /* make me the ilb owner */
2969 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2971 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2974 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2977 cpu_clear(cpu
, nohz
.cpu_mask
);
2979 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2980 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2987 static DEFINE_SPINLOCK(balancing
);
2990 * It checks each scheduling domain to see if it is due to be balanced,
2991 * and initiates a balancing operation if so.
2993 * Balancing parameters are set up in arch_init_sched_domains.
2995 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
2998 struct rq
*rq
= cpu_rq(cpu
);
2999 unsigned long interval
;
3000 struct sched_domain
*sd
;
3001 /* Earliest time when we have to do rebalance again */
3002 unsigned long next_balance
= jiffies
+ 60*HZ
;
3003 int update_next_balance
= 0;
3005 for_each_domain(cpu
, sd
) {
3006 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3009 interval
= sd
->balance_interval
;
3010 if (idle
!= CPU_IDLE
)
3011 interval
*= sd
->busy_factor
;
3013 /* scale ms to jiffies */
3014 interval
= msecs_to_jiffies(interval
);
3015 if (unlikely(!interval
))
3017 if (interval
> HZ
*NR_CPUS
/10)
3018 interval
= HZ
*NR_CPUS
/10;
3021 if (sd
->flags
& SD_SERIALIZE
) {
3022 if (!spin_trylock(&balancing
))
3026 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3027 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3029 * We've pulled tasks over so either we're no
3030 * longer idle, or one of our SMT siblings is
3033 idle
= CPU_NOT_IDLE
;
3035 sd
->last_balance
= jiffies
;
3037 if (sd
->flags
& SD_SERIALIZE
)
3038 spin_unlock(&balancing
);
3040 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3041 next_balance
= sd
->last_balance
+ interval
;
3042 update_next_balance
= 1;
3046 * Stop the load balance at this level. There is another
3047 * CPU in our sched group which is doing load balancing more
3055 * next_balance will be updated only when there is a need.
3056 * When the cpu is attached to null domain for ex, it will not be
3059 if (likely(update_next_balance
))
3060 rq
->next_balance
= next_balance
;
3064 * run_rebalance_domains is triggered when needed from the scheduler tick.
3065 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3066 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3068 static void run_rebalance_domains(struct softirq_action
*h
)
3070 int this_cpu
= smp_processor_id();
3071 struct rq
*this_rq
= cpu_rq(this_cpu
);
3072 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3073 CPU_IDLE
: CPU_NOT_IDLE
;
3075 rebalance_domains(this_cpu
, idle
);
3079 * If this cpu is the owner for idle load balancing, then do the
3080 * balancing on behalf of the other idle cpus whose ticks are
3083 if (this_rq
->idle_at_tick
&&
3084 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3085 cpumask_t cpus
= nohz
.cpu_mask
;
3089 cpu_clear(this_cpu
, cpus
);
3090 for_each_cpu_mask(balance_cpu
, cpus
) {
3092 * If this cpu gets work to do, stop the load balancing
3093 * work being done for other cpus. Next load
3094 * balancing owner will pick it up.
3099 rebalance_domains(balance_cpu
, CPU_IDLE
);
3101 rq
= cpu_rq(balance_cpu
);
3102 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3103 this_rq
->next_balance
= rq
->next_balance
;
3110 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3112 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3113 * idle load balancing owner or decide to stop the periodic load balancing,
3114 * if the whole system is idle.
3116 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3120 * If we were in the nohz mode recently and busy at the current
3121 * scheduler tick, then check if we need to nominate new idle
3124 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3125 rq
->in_nohz_recently
= 0;
3127 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3128 cpu_clear(cpu
, nohz
.cpu_mask
);
3129 atomic_set(&nohz
.load_balancer
, -1);
3132 if (atomic_read(&nohz
.load_balancer
) == -1) {
3134 * simple selection for now: Nominate the
3135 * first cpu in the nohz list to be the next
3138 * TBD: Traverse the sched domains and nominate
3139 * the nearest cpu in the nohz.cpu_mask.
3141 int ilb
= first_cpu(nohz
.cpu_mask
);
3149 * If this cpu is idle and doing idle load balancing for all the
3150 * cpus with ticks stopped, is it time for that to stop?
3152 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3153 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3159 * If this cpu is idle and the idle load balancing is done by
3160 * someone else, then no need raise the SCHED_SOFTIRQ
3162 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3163 cpu_isset(cpu
, nohz
.cpu_mask
))
3166 if (time_after_eq(jiffies
, rq
->next_balance
))
3167 raise_softirq(SCHED_SOFTIRQ
);
3170 #else /* CONFIG_SMP */
3173 * on UP we do not need to balance between CPUs:
3175 static inline void idle_balance(int cpu
, struct rq
*rq
)
3179 /* Avoid "used but not defined" warning on UP */
3180 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3181 unsigned long max_nr_move
, unsigned long max_load_move
,
3182 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3183 int *all_pinned
, unsigned long *load_moved
,
3184 int *this_best_prio
, struct rq_iterator
*iterator
)
3193 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3195 EXPORT_PER_CPU_SYMBOL(kstat
);
3198 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3199 * that have not yet been banked in case the task is currently running.
3201 unsigned long long task_sched_runtime(struct task_struct
*p
)
3203 unsigned long flags
;
3207 rq
= task_rq_lock(p
, &flags
);
3208 ns
= p
->se
.sum_exec_runtime
;
3209 if (rq
->curr
== p
) {
3210 update_rq_clock(rq
);
3211 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3212 if ((s64
)delta_exec
> 0)
3215 task_rq_unlock(rq
, &flags
);
3221 * Account user cpu time to a process.
3222 * @p: the process that the cpu time gets accounted to
3223 * @hardirq_offset: the offset to subtract from hardirq_count()
3224 * @cputime: the cpu time spent in user space since the last update
3226 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3228 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3231 p
->utime
= cputime_add(p
->utime
, cputime
);
3233 /* Add user time to cpustat. */
3234 tmp
= cputime_to_cputime64(cputime
);
3235 if (TASK_NICE(p
) > 0)
3236 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3238 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3242 * Account system cpu time to a process.
3243 * @p: the process that the cpu time gets accounted to
3244 * @hardirq_offset: the offset to subtract from hardirq_count()
3245 * @cputime: the cpu time spent in kernel space since the last update
3247 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3250 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3251 struct rq
*rq
= this_rq();
3254 p
->stime
= cputime_add(p
->stime
, cputime
);
3256 /* Add system time to cpustat. */
3257 tmp
= cputime_to_cputime64(cputime
);
3258 if (hardirq_count() - hardirq_offset
)
3259 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3260 else if (softirq_count())
3261 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3262 else if (p
!= rq
->idle
)
3263 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3264 else if (atomic_read(&rq
->nr_iowait
) > 0)
3265 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3267 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3268 /* Account for system time used */
3269 acct_update_integrals(p
);
3273 * Account for involuntary wait time.
3274 * @p: the process from which the cpu time has been stolen
3275 * @steal: the cpu time spent in involuntary wait
3277 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3279 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3280 cputime64_t tmp
= cputime_to_cputime64(steal
);
3281 struct rq
*rq
= this_rq();
3283 if (p
== rq
->idle
) {
3284 p
->stime
= cputime_add(p
->stime
, steal
);
3285 if (atomic_read(&rq
->nr_iowait
) > 0)
3286 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3288 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3290 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3294 * This function gets called by the timer code, with HZ frequency.
3295 * We call it with interrupts disabled.
3297 * It also gets called by the fork code, when changing the parent's
3300 void scheduler_tick(void)
3302 int cpu
= smp_processor_id();
3303 struct rq
*rq
= cpu_rq(cpu
);
3304 struct task_struct
*curr
= rq
->curr
;
3305 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3307 spin_lock(&rq
->lock
);
3308 __update_rq_clock(rq
);
3310 * Let rq->clock advance by at least TICK_NSEC:
3312 if (unlikely(rq
->clock
< next_tick
))
3313 rq
->clock
= next_tick
;
3314 rq
->tick_timestamp
= rq
->clock
;
3315 update_cpu_load(rq
);
3316 if (curr
!= rq
->idle
) /* FIXME: needed? */
3317 curr
->sched_class
->task_tick(rq
, curr
);
3318 spin_unlock(&rq
->lock
);
3321 rq
->idle_at_tick
= idle_cpu(cpu
);
3322 trigger_load_balance(rq
, cpu
);
3326 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3328 void fastcall
add_preempt_count(int val
)
3333 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3335 preempt_count() += val
;
3337 * Spinlock count overflowing soon?
3339 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3342 EXPORT_SYMBOL(add_preempt_count
);
3344 void fastcall
sub_preempt_count(int val
)
3349 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3352 * Is the spinlock portion underflowing?
3354 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3355 !(preempt_count() & PREEMPT_MASK
)))
3358 preempt_count() -= val
;
3360 EXPORT_SYMBOL(sub_preempt_count
);
3365 * Print scheduling while atomic bug:
3367 static noinline
void __schedule_bug(struct task_struct
*prev
)
3369 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3370 prev
->comm
, preempt_count(), prev
->pid
);
3371 debug_show_held_locks(prev
);
3372 if (irqs_disabled())
3373 print_irqtrace_events(prev
);
3378 * Various schedule()-time debugging checks and statistics:
3380 static inline void schedule_debug(struct task_struct
*prev
)
3383 * Test if we are atomic. Since do_exit() needs to call into
3384 * schedule() atomically, we ignore that path for now.
3385 * Otherwise, whine if we are scheduling when we should not be.
3387 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3388 __schedule_bug(prev
);
3390 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3392 schedstat_inc(this_rq(), sched_count
);
3393 #ifdef CONFIG_SCHEDSTATS
3394 if (unlikely(prev
->lock_depth
>= 0)) {
3395 schedstat_inc(this_rq(), bkl_count
);
3396 schedstat_inc(prev
, sched_info
.bkl_count
);
3402 * Pick up the highest-prio task:
3404 static inline struct task_struct
*
3405 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3407 const struct sched_class
*class;
3408 struct task_struct
*p
;
3411 * Optimization: we know that if all tasks are in
3412 * the fair class we can call that function directly:
3414 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3415 p
= fair_sched_class
.pick_next_task(rq
);
3420 class = sched_class_highest
;
3422 p
= class->pick_next_task(rq
);
3426 * Will never be NULL as the idle class always
3427 * returns a non-NULL p:
3429 class = class->next
;
3434 * schedule() is the main scheduler function.
3436 asmlinkage
void __sched
schedule(void)
3438 struct task_struct
*prev
, *next
;
3445 cpu
= smp_processor_id();
3449 switch_count
= &prev
->nivcsw
;
3451 release_kernel_lock(prev
);
3452 need_resched_nonpreemptible
:
3454 schedule_debug(prev
);
3457 * Do the rq-clock update outside the rq lock:
3459 local_irq_disable();
3460 __update_rq_clock(rq
);
3461 spin_lock(&rq
->lock
);
3462 clear_tsk_need_resched(prev
);
3464 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3465 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3466 unlikely(signal_pending(prev
)))) {
3467 prev
->state
= TASK_RUNNING
;
3469 deactivate_task(rq
, prev
, 1);
3471 switch_count
= &prev
->nvcsw
;
3474 if (unlikely(!rq
->nr_running
))
3475 idle_balance(cpu
, rq
);
3477 prev
->sched_class
->put_prev_task(rq
, prev
);
3478 next
= pick_next_task(rq
, prev
);
3480 sched_info_switch(prev
, next
);
3482 if (likely(prev
!= next
)) {
3487 context_switch(rq
, prev
, next
); /* unlocks the rq */
3489 spin_unlock_irq(&rq
->lock
);
3491 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3492 cpu
= smp_processor_id();
3494 goto need_resched_nonpreemptible
;
3496 preempt_enable_no_resched();
3497 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3500 EXPORT_SYMBOL(schedule
);
3502 #ifdef CONFIG_PREEMPT
3504 * this is the entry point to schedule() from in-kernel preemption
3505 * off of preempt_enable. Kernel preemptions off return from interrupt
3506 * occur there and call schedule directly.
3508 asmlinkage
void __sched
preempt_schedule(void)
3510 struct thread_info
*ti
= current_thread_info();
3511 #ifdef CONFIG_PREEMPT_BKL
3512 struct task_struct
*task
= current
;
3513 int saved_lock_depth
;
3516 * If there is a non-zero preempt_count or interrupts are disabled,
3517 * we do not want to preempt the current task. Just return..
3519 if (likely(ti
->preempt_count
|| irqs_disabled()))
3523 add_preempt_count(PREEMPT_ACTIVE
);
3526 * We keep the big kernel semaphore locked, but we
3527 * clear ->lock_depth so that schedule() doesnt
3528 * auto-release the semaphore:
3530 #ifdef CONFIG_PREEMPT_BKL
3531 saved_lock_depth
= task
->lock_depth
;
3532 task
->lock_depth
= -1;
3535 #ifdef CONFIG_PREEMPT_BKL
3536 task
->lock_depth
= saved_lock_depth
;
3538 sub_preempt_count(PREEMPT_ACTIVE
);
3541 * Check again in case we missed a preemption opportunity
3542 * between schedule and now.
3545 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3547 EXPORT_SYMBOL(preempt_schedule
);
3550 * this is the entry point to schedule() from kernel preemption
3551 * off of irq context.
3552 * Note, that this is called and return with irqs disabled. This will
3553 * protect us against recursive calling from irq.
3555 asmlinkage
void __sched
preempt_schedule_irq(void)
3557 struct thread_info
*ti
= current_thread_info();
3558 #ifdef CONFIG_PREEMPT_BKL
3559 struct task_struct
*task
= current
;
3560 int saved_lock_depth
;
3562 /* Catch callers which need to be fixed */
3563 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3566 add_preempt_count(PREEMPT_ACTIVE
);
3569 * We keep the big kernel semaphore locked, but we
3570 * clear ->lock_depth so that schedule() doesnt
3571 * auto-release the semaphore:
3573 #ifdef CONFIG_PREEMPT_BKL
3574 saved_lock_depth
= task
->lock_depth
;
3575 task
->lock_depth
= -1;
3579 local_irq_disable();
3580 #ifdef CONFIG_PREEMPT_BKL
3581 task
->lock_depth
= saved_lock_depth
;
3583 sub_preempt_count(PREEMPT_ACTIVE
);
3586 * Check again in case we missed a preemption opportunity
3587 * between schedule and now.
3590 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3593 #endif /* CONFIG_PREEMPT */
3595 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3598 return try_to_wake_up(curr
->private, mode
, sync
);
3600 EXPORT_SYMBOL(default_wake_function
);
3603 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3604 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3605 * number) then we wake all the non-exclusive tasks and one exclusive task.
3607 * There are circumstances in which we can try to wake a task which has already
3608 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3609 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3611 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3612 int nr_exclusive
, int sync
, void *key
)
3614 wait_queue_t
*curr
, *next
;
3616 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3617 unsigned flags
= curr
->flags
;
3619 if (curr
->func(curr
, mode
, sync
, key
) &&
3620 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3626 * __wake_up - wake up threads blocked on a waitqueue.
3628 * @mode: which threads
3629 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3630 * @key: is directly passed to the wakeup function
3632 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3633 int nr_exclusive
, void *key
)
3635 unsigned long flags
;
3637 spin_lock_irqsave(&q
->lock
, flags
);
3638 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3639 spin_unlock_irqrestore(&q
->lock
, flags
);
3641 EXPORT_SYMBOL(__wake_up
);
3644 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3646 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3648 __wake_up_common(q
, mode
, 1, 0, NULL
);
3652 * __wake_up_sync - wake up threads blocked on a waitqueue.
3654 * @mode: which threads
3655 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3657 * The sync wakeup differs that the waker knows that it will schedule
3658 * away soon, so while the target thread will be woken up, it will not
3659 * be migrated to another CPU - ie. the two threads are 'synchronized'
3660 * with each other. This can prevent needless bouncing between CPUs.
3662 * On UP it can prevent extra preemption.
3665 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3667 unsigned long flags
;
3673 if (unlikely(!nr_exclusive
))
3676 spin_lock_irqsave(&q
->lock
, flags
);
3677 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3678 spin_unlock_irqrestore(&q
->lock
, flags
);
3680 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3682 void fastcall
complete(struct completion
*x
)
3684 unsigned long flags
;
3686 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3688 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3690 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3692 EXPORT_SYMBOL(complete
);
3694 void fastcall
complete_all(struct completion
*x
)
3696 unsigned long flags
;
3698 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3699 x
->done
+= UINT_MAX
/2;
3700 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3702 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3704 EXPORT_SYMBOL(complete_all
);
3706 static inline long __sched
3707 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3710 DECLARE_WAITQUEUE(wait
, current
);
3712 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3713 __add_wait_queue_tail(&x
->wait
, &wait
);
3715 if (state
== TASK_INTERRUPTIBLE
&&
3716 signal_pending(current
)) {
3717 __remove_wait_queue(&x
->wait
, &wait
);
3718 return -ERESTARTSYS
;
3720 __set_current_state(state
);
3721 spin_unlock_irq(&x
->wait
.lock
);
3722 timeout
= schedule_timeout(timeout
);
3723 spin_lock_irq(&x
->wait
.lock
);
3725 __remove_wait_queue(&x
->wait
, &wait
);
3729 __remove_wait_queue(&x
->wait
, &wait
);
3736 wait_for_common(struct completion
*x
, long timeout
, int state
)
3740 spin_lock_irq(&x
->wait
.lock
);
3741 timeout
= do_wait_for_common(x
, timeout
, state
);
3742 spin_unlock_irq(&x
->wait
.lock
);
3746 void fastcall __sched
wait_for_completion(struct completion
*x
)
3748 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3750 EXPORT_SYMBOL(wait_for_completion
);
3752 unsigned long fastcall __sched
3753 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3755 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3757 EXPORT_SYMBOL(wait_for_completion_timeout
);
3759 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3761 return wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3763 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3765 unsigned long fastcall __sched
3766 wait_for_completion_interruptible_timeout(struct completion
*x
,
3767 unsigned long timeout
)
3769 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3771 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3774 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3776 unsigned long flags
;
3779 init_waitqueue_entry(&wait
, current
);
3781 __set_current_state(state
);
3783 spin_lock_irqsave(&q
->lock
, flags
);
3784 __add_wait_queue(q
, &wait
);
3785 spin_unlock(&q
->lock
);
3786 timeout
= schedule_timeout(timeout
);
3787 spin_lock_irq(&q
->lock
);
3788 __remove_wait_queue(q
, &wait
);
3789 spin_unlock_irqrestore(&q
->lock
, flags
);
3794 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3796 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3798 EXPORT_SYMBOL(interruptible_sleep_on
);
3801 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3803 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3805 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3807 void __sched
sleep_on(wait_queue_head_t
*q
)
3809 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3811 EXPORT_SYMBOL(sleep_on
);
3813 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3815 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3817 EXPORT_SYMBOL(sleep_on_timeout
);
3819 #ifdef CONFIG_RT_MUTEXES
3822 * rt_mutex_setprio - set the current priority of a task
3824 * @prio: prio value (kernel-internal form)
3826 * This function changes the 'effective' priority of a task. It does
3827 * not touch ->normal_prio like __setscheduler().
3829 * Used by the rt_mutex code to implement priority inheritance logic.
3831 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3833 unsigned long flags
;
3834 int oldprio
, on_rq
, running
;
3837 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3839 rq
= task_rq_lock(p
, &flags
);
3840 update_rq_clock(rq
);
3843 on_rq
= p
->se
.on_rq
;
3844 running
= task_running(rq
, p
);
3846 dequeue_task(rq
, p
, 0);
3848 p
->sched_class
->put_prev_task(rq
, p
);
3852 p
->sched_class
= &rt_sched_class
;
3854 p
->sched_class
= &fair_sched_class
;
3860 p
->sched_class
->set_curr_task(rq
);
3861 enqueue_task(rq
, p
, 0);
3863 * Reschedule if we are currently running on this runqueue and
3864 * our priority decreased, or if we are not currently running on
3865 * this runqueue and our priority is higher than the current's
3868 if (p
->prio
> oldprio
)
3869 resched_task(rq
->curr
);
3871 check_preempt_curr(rq
, p
);
3874 task_rq_unlock(rq
, &flags
);
3879 void set_user_nice(struct task_struct
*p
, long nice
)
3881 int old_prio
, delta
, on_rq
;
3882 unsigned long flags
;
3885 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3888 * We have to be careful, if called from sys_setpriority(),
3889 * the task might be in the middle of scheduling on another CPU.
3891 rq
= task_rq_lock(p
, &flags
);
3892 update_rq_clock(rq
);
3894 * The RT priorities are set via sched_setscheduler(), but we still
3895 * allow the 'normal' nice value to be set - but as expected
3896 * it wont have any effect on scheduling until the task is
3897 * SCHED_FIFO/SCHED_RR:
3899 if (task_has_rt_policy(p
)) {
3900 p
->static_prio
= NICE_TO_PRIO(nice
);
3903 on_rq
= p
->se
.on_rq
;
3905 dequeue_task(rq
, p
, 0);
3909 p
->static_prio
= NICE_TO_PRIO(nice
);
3912 p
->prio
= effective_prio(p
);
3913 delta
= p
->prio
- old_prio
;
3916 enqueue_task(rq
, p
, 0);
3919 * If the task increased its priority or is running and
3920 * lowered its priority, then reschedule its CPU:
3922 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3923 resched_task(rq
->curr
);
3926 task_rq_unlock(rq
, &flags
);
3928 EXPORT_SYMBOL(set_user_nice
);
3931 * can_nice - check if a task can reduce its nice value
3935 int can_nice(const struct task_struct
*p
, const int nice
)
3937 /* convert nice value [19,-20] to rlimit style value [1,40] */
3938 int nice_rlim
= 20 - nice
;
3940 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3941 capable(CAP_SYS_NICE
));
3944 #ifdef __ARCH_WANT_SYS_NICE
3947 * sys_nice - change the priority of the current process.
3948 * @increment: priority increment
3950 * sys_setpriority is a more generic, but much slower function that
3951 * does similar things.
3953 asmlinkage
long sys_nice(int increment
)
3958 * Setpriority might change our priority at the same moment.
3959 * We don't have to worry. Conceptually one call occurs first
3960 * and we have a single winner.
3962 if (increment
< -40)
3967 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3973 if (increment
< 0 && !can_nice(current
, nice
))
3976 retval
= security_task_setnice(current
, nice
);
3980 set_user_nice(current
, nice
);
3987 * task_prio - return the priority value of a given task.
3988 * @p: the task in question.
3990 * This is the priority value as seen by users in /proc.
3991 * RT tasks are offset by -200. Normal tasks are centered
3992 * around 0, value goes from -16 to +15.
3994 int task_prio(const struct task_struct
*p
)
3996 return p
->prio
- MAX_RT_PRIO
;
4000 * task_nice - return the nice value of a given task.
4001 * @p: the task in question.
4003 int task_nice(const struct task_struct
*p
)
4005 return TASK_NICE(p
);
4007 EXPORT_SYMBOL_GPL(task_nice
);
4010 * idle_cpu - is a given cpu idle currently?
4011 * @cpu: the processor in question.
4013 int idle_cpu(int cpu
)
4015 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4019 * idle_task - return the idle task for a given cpu.
4020 * @cpu: the processor in question.
4022 struct task_struct
*idle_task(int cpu
)
4024 return cpu_rq(cpu
)->idle
;
4028 * find_process_by_pid - find a process with a matching PID value.
4029 * @pid: the pid in question.
4031 static struct task_struct
*find_process_by_pid(pid_t pid
)
4033 return pid
? find_task_by_pid(pid
) : current
;
4036 /* Actually do priority change: must hold rq lock. */
4038 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4040 BUG_ON(p
->se
.on_rq
);
4043 switch (p
->policy
) {
4047 p
->sched_class
= &fair_sched_class
;
4051 p
->sched_class
= &rt_sched_class
;
4055 p
->rt_priority
= prio
;
4056 p
->normal_prio
= normal_prio(p
);
4057 /* we are holding p->pi_lock already */
4058 p
->prio
= rt_mutex_getprio(p
);
4063 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4064 * @p: the task in question.
4065 * @policy: new policy.
4066 * @param: structure containing the new RT priority.
4068 * NOTE that the task may be already dead.
4070 int sched_setscheduler(struct task_struct
*p
, int policy
,
4071 struct sched_param
*param
)
4073 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4074 unsigned long flags
;
4077 /* may grab non-irq protected spin_locks */
4078 BUG_ON(in_interrupt());
4080 /* double check policy once rq lock held */
4082 policy
= oldpolicy
= p
->policy
;
4083 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4084 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4085 policy
!= SCHED_IDLE
)
4088 * Valid priorities for SCHED_FIFO and SCHED_RR are
4089 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4090 * SCHED_BATCH and SCHED_IDLE is 0.
4092 if (param
->sched_priority
< 0 ||
4093 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4094 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4096 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4100 * Allow unprivileged RT tasks to decrease priority:
4102 if (!capable(CAP_SYS_NICE
)) {
4103 if (rt_policy(policy
)) {
4104 unsigned long rlim_rtprio
;
4106 if (!lock_task_sighand(p
, &flags
))
4108 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4109 unlock_task_sighand(p
, &flags
);
4111 /* can't set/change the rt policy */
4112 if (policy
!= p
->policy
&& !rlim_rtprio
)
4115 /* can't increase priority */
4116 if (param
->sched_priority
> p
->rt_priority
&&
4117 param
->sched_priority
> rlim_rtprio
)
4121 * Like positive nice levels, dont allow tasks to
4122 * move out of SCHED_IDLE either:
4124 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4127 /* can't change other user's priorities */
4128 if ((current
->euid
!= p
->euid
) &&
4129 (current
->euid
!= p
->uid
))
4133 retval
= security_task_setscheduler(p
, policy
, param
);
4137 * make sure no PI-waiters arrive (or leave) while we are
4138 * changing the priority of the task:
4140 spin_lock_irqsave(&p
->pi_lock
, flags
);
4142 * To be able to change p->policy safely, the apropriate
4143 * runqueue lock must be held.
4145 rq
= __task_rq_lock(p
);
4146 /* recheck policy now with rq lock held */
4147 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4148 policy
= oldpolicy
= -1;
4149 __task_rq_unlock(rq
);
4150 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4153 update_rq_clock(rq
);
4154 on_rq
= p
->se
.on_rq
;
4155 running
= task_running(rq
, p
);
4157 deactivate_task(rq
, p
, 0);
4159 p
->sched_class
->put_prev_task(rq
, p
);
4163 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4167 p
->sched_class
->set_curr_task(rq
);
4168 activate_task(rq
, p
, 0);
4170 * Reschedule if we are currently running on this runqueue and
4171 * our priority decreased, or if we are not currently running on
4172 * this runqueue and our priority is higher than the current's
4175 if (p
->prio
> oldprio
)
4176 resched_task(rq
->curr
);
4178 check_preempt_curr(rq
, p
);
4181 __task_rq_unlock(rq
);
4182 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4184 rt_mutex_adjust_pi(p
);
4188 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4191 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4193 struct sched_param lparam
;
4194 struct task_struct
*p
;
4197 if (!param
|| pid
< 0)
4199 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4204 p
= find_process_by_pid(pid
);
4206 retval
= sched_setscheduler(p
, policy
, &lparam
);
4213 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4214 * @pid: the pid in question.
4215 * @policy: new policy.
4216 * @param: structure containing the new RT priority.
4218 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4219 struct sched_param __user
*param
)
4221 /* negative values for policy are not valid */
4225 return do_sched_setscheduler(pid
, policy
, param
);
4229 * sys_sched_setparam - set/change the RT priority of a thread
4230 * @pid: the pid in question.
4231 * @param: structure containing the new RT priority.
4233 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4235 return do_sched_setscheduler(pid
, -1, param
);
4239 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4240 * @pid: the pid in question.
4242 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4244 struct task_struct
*p
;
4251 read_lock(&tasklist_lock
);
4252 p
= find_process_by_pid(pid
);
4254 retval
= security_task_getscheduler(p
);
4258 read_unlock(&tasklist_lock
);
4263 * sys_sched_getscheduler - get the RT priority of a thread
4264 * @pid: the pid in question.
4265 * @param: structure containing the RT priority.
4267 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4269 struct sched_param lp
;
4270 struct task_struct
*p
;
4273 if (!param
|| pid
< 0)
4276 read_lock(&tasklist_lock
);
4277 p
= find_process_by_pid(pid
);
4282 retval
= security_task_getscheduler(p
);
4286 lp
.sched_priority
= p
->rt_priority
;
4287 read_unlock(&tasklist_lock
);
4290 * This one might sleep, we cannot do it with a spinlock held ...
4292 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4297 read_unlock(&tasklist_lock
);
4301 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4303 cpumask_t cpus_allowed
;
4304 struct task_struct
*p
;
4307 mutex_lock(&sched_hotcpu_mutex
);
4308 read_lock(&tasklist_lock
);
4310 p
= find_process_by_pid(pid
);
4312 read_unlock(&tasklist_lock
);
4313 mutex_unlock(&sched_hotcpu_mutex
);
4318 * It is not safe to call set_cpus_allowed with the
4319 * tasklist_lock held. We will bump the task_struct's
4320 * usage count and then drop tasklist_lock.
4323 read_unlock(&tasklist_lock
);
4326 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4327 !capable(CAP_SYS_NICE
))
4330 retval
= security_task_setscheduler(p
, 0, NULL
);
4334 cpus_allowed
= cpuset_cpus_allowed(p
);
4335 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4336 retval
= set_cpus_allowed(p
, new_mask
);
4340 mutex_unlock(&sched_hotcpu_mutex
);
4344 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4345 cpumask_t
*new_mask
)
4347 if (len
< sizeof(cpumask_t
)) {
4348 memset(new_mask
, 0, sizeof(cpumask_t
));
4349 } else if (len
> sizeof(cpumask_t
)) {
4350 len
= sizeof(cpumask_t
);
4352 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4356 * sys_sched_setaffinity - set the cpu affinity of a process
4357 * @pid: pid of the process
4358 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4359 * @user_mask_ptr: user-space pointer to the new cpu mask
4361 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4362 unsigned long __user
*user_mask_ptr
)
4367 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4371 return sched_setaffinity(pid
, new_mask
);
4375 * Represents all cpu's present in the system
4376 * In systems capable of hotplug, this map could dynamically grow
4377 * as new cpu's are detected in the system via any platform specific
4378 * method, such as ACPI for e.g.
4381 cpumask_t cpu_present_map __read_mostly
;
4382 EXPORT_SYMBOL(cpu_present_map
);
4385 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4386 EXPORT_SYMBOL(cpu_online_map
);
4388 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4389 EXPORT_SYMBOL(cpu_possible_map
);
4392 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4394 struct task_struct
*p
;
4397 mutex_lock(&sched_hotcpu_mutex
);
4398 read_lock(&tasklist_lock
);
4401 p
= find_process_by_pid(pid
);
4405 retval
= security_task_getscheduler(p
);
4409 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4412 read_unlock(&tasklist_lock
);
4413 mutex_unlock(&sched_hotcpu_mutex
);
4419 * sys_sched_getaffinity - get the cpu affinity of a process
4420 * @pid: pid of the process
4421 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4422 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4424 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4425 unsigned long __user
*user_mask_ptr
)
4430 if (len
< sizeof(cpumask_t
))
4433 ret
= sched_getaffinity(pid
, &mask
);
4437 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4440 return sizeof(cpumask_t
);
4444 * sys_sched_yield - yield the current processor to other threads.
4446 * This function yields the current CPU to other tasks. If there are no
4447 * other threads running on this CPU then this function will return.
4449 asmlinkage
long sys_sched_yield(void)
4451 struct rq
*rq
= this_rq_lock();
4453 schedstat_inc(rq
, yld_count
);
4454 current
->sched_class
->yield_task(rq
);
4457 * Since we are going to call schedule() anyway, there's
4458 * no need to preempt or enable interrupts:
4460 __release(rq
->lock
);
4461 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4462 _raw_spin_unlock(&rq
->lock
);
4463 preempt_enable_no_resched();
4470 static void __cond_resched(void)
4472 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4473 __might_sleep(__FILE__
, __LINE__
);
4476 * The BKS might be reacquired before we have dropped
4477 * PREEMPT_ACTIVE, which could trigger a second
4478 * cond_resched() call.
4481 add_preempt_count(PREEMPT_ACTIVE
);
4483 sub_preempt_count(PREEMPT_ACTIVE
);
4484 } while (need_resched());
4487 int __sched
cond_resched(void)
4489 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4490 system_state
== SYSTEM_RUNNING
) {
4496 EXPORT_SYMBOL(cond_resched
);
4499 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4500 * call schedule, and on return reacquire the lock.
4502 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4503 * operations here to prevent schedule() from being called twice (once via
4504 * spin_unlock(), once by hand).
4506 int cond_resched_lock(spinlock_t
*lock
)
4510 if (need_lockbreak(lock
)) {
4516 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4517 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4518 _raw_spin_unlock(lock
);
4519 preempt_enable_no_resched();
4526 EXPORT_SYMBOL(cond_resched_lock
);
4528 int __sched
cond_resched_softirq(void)
4530 BUG_ON(!in_softirq());
4532 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4540 EXPORT_SYMBOL(cond_resched_softirq
);
4543 * yield - yield the current processor to other threads.
4545 * This is a shortcut for kernel-space yielding - it marks the
4546 * thread runnable and calls sys_sched_yield().
4548 void __sched
yield(void)
4550 set_current_state(TASK_RUNNING
);
4553 EXPORT_SYMBOL(yield
);
4556 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4557 * that process accounting knows that this is a task in IO wait state.
4559 * But don't do that if it is a deliberate, throttling IO wait (this task
4560 * has set its backing_dev_info: the queue against which it should throttle)
4562 void __sched
io_schedule(void)
4564 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4566 delayacct_blkio_start();
4567 atomic_inc(&rq
->nr_iowait
);
4569 atomic_dec(&rq
->nr_iowait
);
4570 delayacct_blkio_end();
4572 EXPORT_SYMBOL(io_schedule
);
4574 long __sched
io_schedule_timeout(long timeout
)
4576 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4579 delayacct_blkio_start();
4580 atomic_inc(&rq
->nr_iowait
);
4581 ret
= schedule_timeout(timeout
);
4582 atomic_dec(&rq
->nr_iowait
);
4583 delayacct_blkio_end();
4588 * sys_sched_get_priority_max - return maximum RT priority.
4589 * @policy: scheduling class.
4591 * this syscall returns the maximum rt_priority that can be used
4592 * by a given scheduling class.
4594 asmlinkage
long sys_sched_get_priority_max(int policy
)
4601 ret
= MAX_USER_RT_PRIO
-1;
4613 * sys_sched_get_priority_min - return minimum RT priority.
4614 * @policy: scheduling class.
4616 * this syscall returns the minimum rt_priority that can be used
4617 * by a given scheduling class.
4619 asmlinkage
long sys_sched_get_priority_min(int policy
)
4637 * sys_sched_rr_get_interval - return the default timeslice of a process.
4638 * @pid: pid of the process.
4639 * @interval: userspace pointer to the timeslice value.
4641 * this syscall writes the default timeslice value of a given process
4642 * into the user-space timespec buffer. A value of '0' means infinity.
4645 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4647 struct task_struct
*p
;
4648 unsigned int time_slice
;
4656 read_lock(&tasklist_lock
);
4657 p
= find_process_by_pid(pid
);
4661 retval
= security_task_getscheduler(p
);
4665 if (p
->policy
== SCHED_FIFO
)
4667 else if (p
->policy
== SCHED_RR
)
4668 time_slice
= DEF_TIMESLICE
;
4670 struct sched_entity
*se
= &p
->se
;
4671 unsigned long flags
;
4674 rq
= task_rq_lock(p
, &flags
);
4675 time_slice
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
4676 task_rq_unlock(rq
, &flags
);
4678 read_unlock(&tasklist_lock
);
4679 jiffies_to_timespec(time_slice
, &t
);
4680 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4684 read_unlock(&tasklist_lock
);
4688 static const char stat_nam
[] = "RSDTtZX";
4690 static void show_task(struct task_struct
*p
)
4692 unsigned long free
= 0;
4695 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4696 printk("%-13.13s %c", p
->comm
,
4697 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4698 #if BITS_PER_LONG == 32
4699 if (state
== TASK_RUNNING
)
4700 printk(" running ");
4702 printk(" %08lx ", thread_saved_pc(p
));
4704 if (state
== TASK_RUNNING
)
4705 printk(" running task ");
4707 printk(" %016lx ", thread_saved_pc(p
));
4709 #ifdef CONFIG_DEBUG_STACK_USAGE
4711 unsigned long *n
= end_of_stack(p
);
4714 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4717 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4719 if (state
!= TASK_RUNNING
)
4720 show_stack(p
, NULL
);
4723 void show_state_filter(unsigned long state_filter
)
4725 struct task_struct
*g
, *p
;
4727 #if BITS_PER_LONG == 32
4729 " task PC stack pid father\n");
4732 " task PC stack pid father\n");
4734 read_lock(&tasklist_lock
);
4735 do_each_thread(g
, p
) {
4737 * reset the NMI-timeout, listing all files on a slow
4738 * console might take alot of time:
4740 touch_nmi_watchdog();
4741 if (!state_filter
|| (p
->state
& state_filter
))
4743 } while_each_thread(g
, p
);
4745 touch_all_softlockup_watchdogs();
4747 #ifdef CONFIG_SCHED_DEBUG
4748 sysrq_sched_debug_show();
4750 read_unlock(&tasklist_lock
);
4752 * Only show locks if all tasks are dumped:
4754 if (state_filter
== -1)
4755 debug_show_all_locks();
4758 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4760 idle
->sched_class
= &idle_sched_class
;
4764 * init_idle - set up an idle thread for a given CPU
4765 * @idle: task in question
4766 * @cpu: cpu the idle task belongs to
4768 * NOTE: this function does not set the idle thread's NEED_RESCHED
4769 * flag, to make booting more robust.
4771 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4773 struct rq
*rq
= cpu_rq(cpu
);
4774 unsigned long flags
;
4777 idle
->se
.exec_start
= sched_clock();
4779 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4780 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4781 __set_task_cpu(idle
, cpu
);
4783 spin_lock_irqsave(&rq
->lock
, flags
);
4784 rq
->curr
= rq
->idle
= idle
;
4785 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4788 spin_unlock_irqrestore(&rq
->lock
, flags
);
4790 /* Set the preempt count _outside_ the spinlocks! */
4791 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4792 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4794 task_thread_info(idle
)->preempt_count
= 0;
4797 * The idle tasks have their own, simple scheduling class:
4799 idle
->sched_class
= &idle_sched_class
;
4803 * In a system that switches off the HZ timer nohz_cpu_mask
4804 * indicates which cpus entered this state. This is used
4805 * in the rcu update to wait only for active cpus. For system
4806 * which do not switch off the HZ timer nohz_cpu_mask should
4807 * always be CPU_MASK_NONE.
4809 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4813 * This is how migration works:
4815 * 1) we queue a struct migration_req structure in the source CPU's
4816 * runqueue and wake up that CPU's migration thread.
4817 * 2) we down() the locked semaphore => thread blocks.
4818 * 3) migration thread wakes up (implicitly it forces the migrated
4819 * thread off the CPU)
4820 * 4) it gets the migration request and checks whether the migrated
4821 * task is still in the wrong runqueue.
4822 * 5) if it's in the wrong runqueue then the migration thread removes
4823 * it and puts it into the right queue.
4824 * 6) migration thread up()s the semaphore.
4825 * 7) we wake up and the migration is done.
4829 * Change a given task's CPU affinity. Migrate the thread to a
4830 * proper CPU and schedule it away if the CPU it's executing on
4831 * is removed from the allowed bitmask.
4833 * NOTE: the caller must have a valid reference to the task, the
4834 * task must not exit() & deallocate itself prematurely. The
4835 * call is not atomic; no spinlocks may be held.
4837 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4839 struct migration_req req
;
4840 unsigned long flags
;
4844 rq
= task_rq_lock(p
, &flags
);
4845 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4850 p
->cpus_allowed
= new_mask
;
4851 /* Can the task run on the task's current CPU? If so, we're done */
4852 if (cpu_isset(task_cpu(p
), new_mask
))
4855 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4856 /* Need help from migration thread: drop lock and wait. */
4857 task_rq_unlock(rq
, &flags
);
4858 wake_up_process(rq
->migration_thread
);
4859 wait_for_completion(&req
.done
);
4860 tlb_migrate_finish(p
->mm
);
4864 task_rq_unlock(rq
, &flags
);
4868 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4871 * Move (not current) task off this cpu, onto dest cpu. We're doing
4872 * this because either it can't run here any more (set_cpus_allowed()
4873 * away from this CPU, or CPU going down), or because we're
4874 * attempting to rebalance this task on exec (sched_exec).
4876 * So we race with normal scheduler movements, but that's OK, as long
4877 * as the task is no longer on this CPU.
4879 * Returns non-zero if task was successfully migrated.
4881 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4883 struct rq
*rq_dest
, *rq_src
;
4886 if (unlikely(cpu_is_offline(dest_cpu
)))
4889 rq_src
= cpu_rq(src_cpu
);
4890 rq_dest
= cpu_rq(dest_cpu
);
4892 double_rq_lock(rq_src
, rq_dest
);
4893 /* Already moved. */
4894 if (task_cpu(p
) != src_cpu
)
4896 /* Affinity changed (again). */
4897 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4900 on_rq
= p
->se
.on_rq
;
4902 deactivate_task(rq_src
, p
, 0);
4904 set_task_cpu(p
, dest_cpu
);
4906 activate_task(rq_dest
, p
, 0);
4907 check_preempt_curr(rq_dest
, p
);
4911 double_rq_unlock(rq_src
, rq_dest
);
4916 * migration_thread - this is a highprio system thread that performs
4917 * thread migration by bumping thread off CPU then 'pushing' onto
4920 static int migration_thread(void *data
)
4922 int cpu
= (long)data
;
4926 BUG_ON(rq
->migration_thread
!= current
);
4928 set_current_state(TASK_INTERRUPTIBLE
);
4929 while (!kthread_should_stop()) {
4930 struct migration_req
*req
;
4931 struct list_head
*head
;
4933 spin_lock_irq(&rq
->lock
);
4935 if (cpu_is_offline(cpu
)) {
4936 spin_unlock_irq(&rq
->lock
);
4940 if (rq
->active_balance
) {
4941 active_load_balance(rq
, cpu
);
4942 rq
->active_balance
= 0;
4945 head
= &rq
->migration_queue
;
4947 if (list_empty(head
)) {
4948 spin_unlock_irq(&rq
->lock
);
4950 set_current_state(TASK_INTERRUPTIBLE
);
4953 req
= list_entry(head
->next
, struct migration_req
, list
);
4954 list_del_init(head
->next
);
4956 spin_unlock(&rq
->lock
);
4957 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4960 complete(&req
->done
);
4962 __set_current_state(TASK_RUNNING
);
4966 /* Wait for kthread_stop */
4967 set_current_state(TASK_INTERRUPTIBLE
);
4968 while (!kthread_should_stop()) {
4970 set_current_state(TASK_INTERRUPTIBLE
);
4972 __set_current_state(TASK_RUNNING
);
4976 #ifdef CONFIG_HOTPLUG_CPU
4978 * Figure out where task on dead CPU should go, use force if neccessary.
4979 * NOTE: interrupts should be disabled by the caller
4981 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
4983 unsigned long flags
;
4990 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4991 cpus_and(mask
, mask
, p
->cpus_allowed
);
4992 dest_cpu
= any_online_cpu(mask
);
4994 /* On any allowed CPU? */
4995 if (dest_cpu
== NR_CPUS
)
4996 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
4998 /* No more Mr. Nice Guy. */
4999 if (dest_cpu
== NR_CPUS
) {
5000 rq
= task_rq_lock(p
, &flags
);
5001 cpus_setall(p
->cpus_allowed
);
5002 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5003 task_rq_unlock(rq
, &flags
);
5006 * Don't tell them about moving exiting tasks or
5007 * kernel threads (both mm NULL), since they never
5010 if (p
->mm
&& printk_ratelimit())
5011 printk(KERN_INFO
"process %d (%s) no "
5012 "longer affine to cpu%d\n",
5013 p
->pid
, p
->comm
, dead_cpu
);
5015 } while (!__migrate_task(p
, dead_cpu
, dest_cpu
));
5019 * While a dead CPU has no uninterruptible tasks queued at this point,
5020 * it might still have a nonzero ->nr_uninterruptible counter, because
5021 * for performance reasons the counter is not stricly tracking tasks to
5022 * their home CPUs. So we just add the counter to another CPU's counter,
5023 * to keep the global sum constant after CPU-down:
5025 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5027 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5028 unsigned long flags
;
5030 local_irq_save(flags
);
5031 double_rq_lock(rq_src
, rq_dest
);
5032 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5033 rq_src
->nr_uninterruptible
= 0;
5034 double_rq_unlock(rq_src
, rq_dest
);
5035 local_irq_restore(flags
);
5038 /* Run through task list and migrate tasks from the dead cpu. */
5039 static void migrate_live_tasks(int src_cpu
)
5041 struct task_struct
*p
, *t
;
5043 write_lock_irq(&tasklist_lock
);
5045 do_each_thread(t
, p
) {
5049 if (task_cpu(p
) == src_cpu
)
5050 move_task_off_dead_cpu(src_cpu
, p
);
5051 } while_each_thread(t
, p
);
5053 write_unlock_irq(&tasklist_lock
);
5057 * activate_idle_task - move idle task to the _front_ of runqueue.
5059 static void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
5061 update_rq_clock(rq
);
5063 if (p
->state
== TASK_UNINTERRUPTIBLE
)
5064 rq
->nr_uninterruptible
--;
5066 enqueue_task(rq
, p
, 0);
5067 inc_nr_running(p
, rq
);
5071 * Schedules idle task to be the next runnable task on current CPU.
5072 * It does so by boosting its priority to highest possible and adding it to
5073 * the _front_ of the runqueue. Used by CPU offline code.
5075 void sched_idle_next(void)
5077 int this_cpu
= smp_processor_id();
5078 struct rq
*rq
= cpu_rq(this_cpu
);
5079 struct task_struct
*p
= rq
->idle
;
5080 unsigned long flags
;
5082 /* cpu has to be offline */
5083 BUG_ON(cpu_online(this_cpu
));
5086 * Strictly not necessary since rest of the CPUs are stopped by now
5087 * and interrupts disabled on the current cpu.
5089 spin_lock_irqsave(&rq
->lock
, flags
);
5091 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5093 /* Add idle task to the _front_ of its priority queue: */
5094 activate_idle_task(p
, rq
);
5096 spin_unlock_irqrestore(&rq
->lock
, flags
);
5100 * Ensures that the idle task is using init_mm right before its cpu goes
5103 void idle_task_exit(void)
5105 struct mm_struct
*mm
= current
->active_mm
;
5107 BUG_ON(cpu_online(smp_processor_id()));
5110 switch_mm(mm
, &init_mm
, current
);
5114 /* called under rq->lock with disabled interrupts */
5115 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5117 struct rq
*rq
= cpu_rq(dead_cpu
);
5119 /* Must be exiting, otherwise would be on tasklist. */
5120 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5122 /* Cannot have done final schedule yet: would have vanished. */
5123 BUG_ON(p
->state
== TASK_DEAD
);
5128 * Drop lock around migration; if someone else moves it,
5129 * that's OK. No task can be added to this CPU, so iteration is
5131 * NOTE: interrupts should be left disabled --dev@
5133 spin_unlock(&rq
->lock
);
5134 move_task_off_dead_cpu(dead_cpu
, p
);
5135 spin_lock(&rq
->lock
);
5140 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5141 static void migrate_dead_tasks(unsigned int dead_cpu
)
5143 struct rq
*rq
= cpu_rq(dead_cpu
);
5144 struct task_struct
*next
;
5147 if (!rq
->nr_running
)
5149 update_rq_clock(rq
);
5150 next
= pick_next_task(rq
, rq
->curr
);
5153 migrate_dead(dead_cpu
, next
);
5157 #endif /* CONFIG_HOTPLUG_CPU */
5159 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5161 static struct ctl_table sd_ctl_dir
[] = {
5163 .procname
= "sched_domain",
5169 static struct ctl_table sd_ctl_root
[] = {
5171 .ctl_name
= CTL_KERN
,
5172 .procname
= "kernel",
5174 .child
= sd_ctl_dir
,
5179 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5181 struct ctl_table
*entry
=
5182 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5185 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5191 set_table_entry(struct ctl_table
*entry
,
5192 const char *procname
, void *data
, int maxlen
,
5193 mode_t mode
, proc_handler
*proc_handler
)
5195 entry
->procname
= procname
;
5197 entry
->maxlen
= maxlen
;
5199 entry
->proc_handler
= proc_handler
;
5202 static struct ctl_table
*
5203 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5205 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5207 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5208 sizeof(long), 0644, proc_doulongvec_minmax
);
5209 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5210 sizeof(long), 0644, proc_doulongvec_minmax
);
5211 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5212 sizeof(int), 0644, proc_dointvec_minmax
);
5213 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5214 sizeof(int), 0644, proc_dointvec_minmax
);
5215 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5216 sizeof(int), 0644, proc_dointvec_minmax
);
5217 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5218 sizeof(int), 0644, proc_dointvec_minmax
);
5219 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5220 sizeof(int), 0644, proc_dointvec_minmax
);
5221 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5222 sizeof(int), 0644, proc_dointvec_minmax
);
5223 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5224 sizeof(int), 0644, proc_dointvec_minmax
);
5225 set_table_entry(&table
[9], "cache_nice_tries",
5226 &sd
->cache_nice_tries
,
5227 sizeof(int), 0644, proc_dointvec_minmax
);
5228 set_table_entry(&table
[10], "flags", &sd
->flags
,
5229 sizeof(int), 0644, proc_dointvec_minmax
);
5234 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5236 struct ctl_table
*entry
, *table
;
5237 struct sched_domain
*sd
;
5238 int domain_num
= 0, i
;
5241 for_each_domain(cpu
, sd
)
5243 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5246 for_each_domain(cpu
, sd
) {
5247 snprintf(buf
, 32, "domain%d", i
);
5248 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5250 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5257 static struct ctl_table_header
*sd_sysctl_header
;
5258 static void init_sched_domain_sysctl(void)
5260 int i
, cpu_num
= num_online_cpus();
5261 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5264 sd_ctl_dir
[0].child
= entry
;
5266 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5267 snprintf(buf
, 32, "cpu%d", i
);
5268 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5270 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5272 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5275 static void init_sched_domain_sysctl(void)
5281 * migration_call - callback that gets triggered when a CPU is added.
5282 * Here we can start up the necessary migration thread for the new CPU.
5284 static int __cpuinit
5285 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5287 struct task_struct
*p
;
5288 int cpu
= (long)hcpu
;
5289 unsigned long flags
;
5293 case CPU_LOCK_ACQUIRE
:
5294 mutex_lock(&sched_hotcpu_mutex
);
5297 case CPU_UP_PREPARE
:
5298 case CPU_UP_PREPARE_FROZEN
:
5299 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5302 kthread_bind(p
, cpu
);
5303 /* Must be high prio: stop_machine expects to yield to it. */
5304 rq
= task_rq_lock(p
, &flags
);
5305 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5306 task_rq_unlock(rq
, &flags
);
5307 cpu_rq(cpu
)->migration_thread
= p
;
5311 case CPU_ONLINE_FROZEN
:
5312 /* Strictly unneccessary, as first user will wake it. */
5313 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5316 #ifdef CONFIG_HOTPLUG_CPU
5317 case CPU_UP_CANCELED
:
5318 case CPU_UP_CANCELED_FROZEN
:
5319 if (!cpu_rq(cpu
)->migration_thread
)
5321 /* Unbind it from offline cpu so it can run. Fall thru. */
5322 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5323 any_online_cpu(cpu_online_map
));
5324 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5325 cpu_rq(cpu
)->migration_thread
= NULL
;
5329 case CPU_DEAD_FROZEN
:
5330 migrate_live_tasks(cpu
);
5332 kthread_stop(rq
->migration_thread
);
5333 rq
->migration_thread
= NULL
;
5334 /* Idle task back to normal (off runqueue, low prio) */
5335 rq
= task_rq_lock(rq
->idle
, &flags
);
5336 update_rq_clock(rq
);
5337 deactivate_task(rq
, rq
->idle
, 0);
5338 rq
->idle
->static_prio
= MAX_PRIO
;
5339 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5340 rq
->idle
->sched_class
= &idle_sched_class
;
5341 migrate_dead_tasks(cpu
);
5342 task_rq_unlock(rq
, &flags
);
5343 migrate_nr_uninterruptible(rq
);
5344 BUG_ON(rq
->nr_running
!= 0);
5346 /* No need to migrate the tasks: it was best-effort if
5347 * they didn't take sched_hotcpu_mutex. Just wake up
5348 * the requestors. */
5349 spin_lock_irq(&rq
->lock
);
5350 while (!list_empty(&rq
->migration_queue
)) {
5351 struct migration_req
*req
;
5353 req
= list_entry(rq
->migration_queue
.next
,
5354 struct migration_req
, list
);
5355 list_del_init(&req
->list
);
5356 complete(&req
->done
);
5358 spin_unlock_irq(&rq
->lock
);
5361 case CPU_LOCK_RELEASE
:
5362 mutex_unlock(&sched_hotcpu_mutex
);
5368 /* Register at highest priority so that task migration (migrate_all_tasks)
5369 * happens before everything else.
5371 static struct notifier_block __cpuinitdata migration_notifier
= {
5372 .notifier_call
= migration_call
,
5376 int __init
migration_init(void)
5378 void *cpu
= (void *)(long)smp_processor_id();
5381 /* Start one for the boot CPU: */
5382 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5383 BUG_ON(err
== NOTIFY_BAD
);
5384 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5385 register_cpu_notifier(&migration_notifier
);
5393 /* Number of possible processor ids */
5394 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5395 EXPORT_SYMBOL(nr_cpu_ids
);
5397 #ifdef CONFIG_SCHED_DEBUG
5398 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5403 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5407 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5412 struct sched_group
*group
= sd
->groups
;
5413 cpumask_t groupmask
;
5415 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5416 cpus_clear(groupmask
);
5419 for (i
= 0; i
< level
+ 1; i
++)
5421 printk("domain %d: ", level
);
5423 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5424 printk("does not load-balance\n");
5426 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5431 printk("span %s\n", str
);
5433 if (!cpu_isset(cpu
, sd
->span
))
5434 printk(KERN_ERR
"ERROR: domain->span does not contain "
5436 if (!cpu_isset(cpu
, group
->cpumask
))
5437 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5441 for (i
= 0; i
< level
+ 2; i
++)
5447 printk(KERN_ERR
"ERROR: group is NULL\n");
5451 if (!group
->__cpu_power
) {
5453 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5458 if (!cpus_weight(group
->cpumask
)) {
5460 printk(KERN_ERR
"ERROR: empty group\n");
5464 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5466 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5470 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5472 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5475 group
= group
->next
;
5476 } while (group
!= sd
->groups
);
5479 if (!cpus_equal(sd
->span
, groupmask
))
5480 printk(KERN_ERR
"ERROR: groups don't span "
5488 if (!cpus_subset(groupmask
, sd
->span
))
5489 printk(KERN_ERR
"ERROR: parent span is not a superset "
5490 "of domain->span\n");
5495 # define sched_domain_debug(sd, cpu) do { } while (0)
5498 static int sd_degenerate(struct sched_domain
*sd
)
5500 if (cpus_weight(sd
->span
) == 1)
5503 /* Following flags need at least 2 groups */
5504 if (sd
->flags
& (SD_LOAD_BALANCE
|
5505 SD_BALANCE_NEWIDLE
|
5509 SD_SHARE_PKG_RESOURCES
)) {
5510 if (sd
->groups
!= sd
->groups
->next
)
5514 /* Following flags don't use groups */
5515 if (sd
->flags
& (SD_WAKE_IDLE
|
5524 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5526 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5528 if (sd_degenerate(parent
))
5531 if (!cpus_equal(sd
->span
, parent
->span
))
5534 /* Does parent contain flags not in child? */
5535 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5536 if (cflags
& SD_WAKE_AFFINE
)
5537 pflags
&= ~SD_WAKE_BALANCE
;
5538 /* Flags needing groups don't count if only 1 group in parent */
5539 if (parent
->groups
== parent
->groups
->next
) {
5540 pflags
&= ~(SD_LOAD_BALANCE
|
5541 SD_BALANCE_NEWIDLE
|
5545 SD_SHARE_PKG_RESOURCES
);
5547 if (~cflags
& pflags
)
5554 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5555 * hold the hotplug lock.
5557 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5559 struct rq
*rq
= cpu_rq(cpu
);
5560 struct sched_domain
*tmp
;
5562 /* Remove the sched domains which do not contribute to scheduling. */
5563 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5564 struct sched_domain
*parent
= tmp
->parent
;
5567 if (sd_parent_degenerate(tmp
, parent
)) {
5568 tmp
->parent
= parent
->parent
;
5570 parent
->parent
->child
= tmp
;
5574 if (sd
&& sd_degenerate(sd
)) {
5580 sched_domain_debug(sd
, cpu
);
5582 rcu_assign_pointer(rq
->sd
, sd
);
5585 /* cpus with isolated domains */
5586 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5588 /* Setup the mask of cpus configured for isolated domains */
5589 static int __init
isolated_cpu_setup(char *str
)
5591 int ints
[NR_CPUS
], i
;
5593 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5594 cpus_clear(cpu_isolated_map
);
5595 for (i
= 1; i
<= ints
[0]; i
++)
5596 if (ints
[i
] < NR_CPUS
)
5597 cpu_set(ints
[i
], cpu_isolated_map
);
5601 __setup("isolcpus=", isolated_cpu_setup
);
5604 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5605 * to a function which identifies what group(along with sched group) a CPU
5606 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5607 * (due to the fact that we keep track of groups covered with a cpumask_t).
5609 * init_sched_build_groups will build a circular linked list of the groups
5610 * covered by the given span, and will set each group's ->cpumask correctly,
5611 * and ->cpu_power to 0.
5614 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5615 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5616 struct sched_group
**sg
))
5618 struct sched_group
*first
= NULL
, *last
= NULL
;
5619 cpumask_t covered
= CPU_MASK_NONE
;
5622 for_each_cpu_mask(i
, span
) {
5623 struct sched_group
*sg
;
5624 int group
= group_fn(i
, cpu_map
, &sg
);
5627 if (cpu_isset(i
, covered
))
5630 sg
->cpumask
= CPU_MASK_NONE
;
5631 sg
->__cpu_power
= 0;
5633 for_each_cpu_mask(j
, span
) {
5634 if (group_fn(j
, cpu_map
, NULL
) != group
)
5637 cpu_set(j
, covered
);
5638 cpu_set(j
, sg
->cpumask
);
5649 #define SD_NODES_PER_DOMAIN 16
5654 * find_next_best_node - find the next node to include in a sched_domain
5655 * @node: node whose sched_domain we're building
5656 * @used_nodes: nodes already in the sched_domain
5658 * Find the next node to include in a given scheduling domain. Simply
5659 * finds the closest node not already in the @used_nodes map.
5661 * Should use nodemask_t.
5663 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5665 int i
, n
, val
, min_val
, best_node
= 0;
5669 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5670 /* Start at @node */
5671 n
= (node
+ i
) % MAX_NUMNODES
;
5673 if (!nr_cpus_node(n
))
5676 /* Skip already used nodes */
5677 if (test_bit(n
, used_nodes
))
5680 /* Simple min distance search */
5681 val
= node_distance(node
, n
);
5683 if (val
< min_val
) {
5689 set_bit(best_node
, used_nodes
);
5694 * sched_domain_node_span - get a cpumask for a node's sched_domain
5695 * @node: node whose cpumask we're constructing
5696 * @size: number of nodes to include in this span
5698 * Given a node, construct a good cpumask for its sched_domain to span. It
5699 * should be one that prevents unnecessary balancing, but also spreads tasks
5702 static cpumask_t
sched_domain_node_span(int node
)
5704 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5705 cpumask_t span
, nodemask
;
5709 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5711 nodemask
= node_to_cpumask(node
);
5712 cpus_or(span
, span
, nodemask
);
5713 set_bit(node
, used_nodes
);
5715 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5716 int next_node
= find_next_best_node(node
, used_nodes
);
5718 nodemask
= node_to_cpumask(next_node
);
5719 cpus_or(span
, span
, nodemask
);
5726 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5729 * SMT sched-domains:
5731 #ifdef CONFIG_SCHED_SMT
5732 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5733 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5735 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5736 struct sched_group
**sg
)
5739 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5745 * multi-core sched-domains:
5747 #ifdef CONFIG_SCHED_MC
5748 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5749 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5752 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5753 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5754 struct sched_group
**sg
)
5757 cpumask_t mask
= cpu_sibling_map
[cpu
];
5758 cpus_and(mask
, mask
, *cpu_map
);
5759 group
= first_cpu(mask
);
5761 *sg
= &per_cpu(sched_group_core
, group
);
5764 #elif defined(CONFIG_SCHED_MC)
5765 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5766 struct sched_group
**sg
)
5769 *sg
= &per_cpu(sched_group_core
, cpu
);
5774 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5775 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5777 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5778 struct sched_group
**sg
)
5781 #ifdef CONFIG_SCHED_MC
5782 cpumask_t mask
= cpu_coregroup_map(cpu
);
5783 cpus_and(mask
, mask
, *cpu_map
);
5784 group
= first_cpu(mask
);
5785 #elif defined(CONFIG_SCHED_SMT)
5786 cpumask_t mask
= cpu_sibling_map
[cpu
];
5787 cpus_and(mask
, mask
, *cpu_map
);
5788 group
= first_cpu(mask
);
5793 *sg
= &per_cpu(sched_group_phys
, group
);
5799 * The init_sched_build_groups can't handle what we want to do with node
5800 * groups, so roll our own. Now each node has its own list of groups which
5801 * gets dynamically allocated.
5803 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5804 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5806 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5807 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5809 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5810 struct sched_group
**sg
)
5812 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5815 cpus_and(nodemask
, nodemask
, *cpu_map
);
5816 group
= first_cpu(nodemask
);
5819 *sg
= &per_cpu(sched_group_allnodes
, group
);
5823 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5825 struct sched_group
*sg
= group_head
;
5831 for_each_cpu_mask(j
, sg
->cpumask
) {
5832 struct sched_domain
*sd
;
5834 sd
= &per_cpu(phys_domains
, j
);
5835 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5837 * Only add "power" once for each
5843 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5846 } while (sg
!= group_head
);
5851 /* Free memory allocated for various sched_group structures */
5852 static void free_sched_groups(const cpumask_t
*cpu_map
)
5856 for_each_cpu_mask(cpu
, *cpu_map
) {
5857 struct sched_group
**sched_group_nodes
5858 = sched_group_nodes_bycpu
[cpu
];
5860 if (!sched_group_nodes
)
5863 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5864 cpumask_t nodemask
= node_to_cpumask(i
);
5865 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5867 cpus_and(nodemask
, nodemask
, *cpu_map
);
5868 if (cpus_empty(nodemask
))
5878 if (oldsg
!= sched_group_nodes
[i
])
5881 kfree(sched_group_nodes
);
5882 sched_group_nodes_bycpu
[cpu
] = NULL
;
5886 static void free_sched_groups(const cpumask_t
*cpu_map
)
5892 * Initialize sched groups cpu_power.
5894 * cpu_power indicates the capacity of sched group, which is used while
5895 * distributing the load between different sched groups in a sched domain.
5896 * Typically cpu_power for all the groups in a sched domain will be same unless
5897 * there are asymmetries in the topology. If there are asymmetries, group
5898 * having more cpu_power will pickup more load compared to the group having
5901 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5902 * the maximum number of tasks a group can handle in the presence of other idle
5903 * or lightly loaded groups in the same sched domain.
5905 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5907 struct sched_domain
*child
;
5908 struct sched_group
*group
;
5910 WARN_ON(!sd
|| !sd
->groups
);
5912 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5917 sd
->groups
->__cpu_power
= 0;
5920 * For perf policy, if the groups in child domain share resources
5921 * (for example cores sharing some portions of the cache hierarchy
5922 * or SMT), then set this domain groups cpu_power such that each group
5923 * can handle only one task, when there are other idle groups in the
5924 * same sched domain.
5926 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5928 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5929 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5934 * add cpu_power of each child group to this groups cpu_power
5936 group
= child
->groups
;
5938 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5939 group
= group
->next
;
5940 } while (group
!= child
->groups
);
5944 * Build sched domains for a given set of cpus and attach the sched domains
5945 * to the individual cpus
5947 static int build_sched_domains(const cpumask_t
*cpu_map
)
5951 struct sched_group
**sched_group_nodes
= NULL
;
5952 int sd_allnodes
= 0;
5955 * Allocate the per-node list of sched groups
5957 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5959 if (!sched_group_nodes
) {
5960 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5963 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5967 * Set up domains for cpus specified by the cpu_map.
5969 for_each_cpu_mask(i
, *cpu_map
) {
5970 struct sched_domain
*sd
= NULL
, *p
;
5971 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5973 cpus_and(nodemask
, nodemask
, *cpu_map
);
5976 if (cpus_weight(*cpu_map
) >
5977 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5978 sd
= &per_cpu(allnodes_domains
, i
);
5979 *sd
= SD_ALLNODES_INIT
;
5980 sd
->span
= *cpu_map
;
5981 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
5987 sd
= &per_cpu(node_domains
, i
);
5989 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5993 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
5997 sd
= &per_cpu(phys_domains
, i
);
5999 sd
->span
= nodemask
;
6003 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6005 #ifdef CONFIG_SCHED_MC
6007 sd
= &per_cpu(core_domains
, i
);
6009 sd
->span
= cpu_coregroup_map(i
);
6010 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6013 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6016 #ifdef CONFIG_SCHED_SMT
6018 sd
= &per_cpu(cpu_domains
, i
);
6019 *sd
= SD_SIBLING_INIT
;
6020 sd
->span
= cpu_sibling_map
[i
];
6021 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6024 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6028 #ifdef CONFIG_SCHED_SMT
6029 /* Set up CPU (sibling) groups */
6030 for_each_cpu_mask(i
, *cpu_map
) {
6031 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6032 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6033 if (i
!= first_cpu(this_sibling_map
))
6036 init_sched_build_groups(this_sibling_map
, cpu_map
,
6041 #ifdef CONFIG_SCHED_MC
6042 /* Set up multi-core groups */
6043 for_each_cpu_mask(i
, *cpu_map
) {
6044 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6045 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6046 if (i
!= first_cpu(this_core_map
))
6048 init_sched_build_groups(this_core_map
, cpu_map
,
6049 &cpu_to_core_group
);
6053 /* Set up physical groups */
6054 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6055 cpumask_t nodemask
= node_to_cpumask(i
);
6057 cpus_and(nodemask
, nodemask
, *cpu_map
);
6058 if (cpus_empty(nodemask
))
6061 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6065 /* Set up node groups */
6067 init_sched_build_groups(*cpu_map
, cpu_map
,
6068 &cpu_to_allnodes_group
);
6070 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6071 /* Set up node groups */
6072 struct sched_group
*sg
, *prev
;
6073 cpumask_t nodemask
= node_to_cpumask(i
);
6074 cpumask_t domainspan
;
6075 cpumask_t covered
= CPU_MASK_NONE
;
6078 cpus_and(nodemask
, nodemask
, *cpu_map
);
6079 if (cpus_empty(nodemask
)) {
6080 sched_group_nodes
[i
] = NULL
;
6084 domainspan
= sched_domain_node_span(i
);
6085 cpus_and(domainspan
, domainspan
, *cpu_map
);
6087 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6089 printk(KERN_WARNING
"Can not alloc domain group for "
6093 sched_group_nodes
[i
] = sg
;
6094 for_each_cpu_mask(j
, nodemask
) {
6095 struct sched_domain
*sd
;
6097 sd
= &per_cpu(node_domains
, j
);
6100 sg
->__cpu_power
= 0;
6101 sg
->cpumask
= nodemask
;
6103 cpus_or(covered
, covered
, nodemask
);
6106 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6107 cpumask_t tmp
, notcovered
;
6108 int n
= (i
+ j
) % MAX_NUMNODES
;
6110 cpus_complement(notcovered
, covered
);
6111 cpus_and(tmp
, notcovered
, *cpu_map
);
6112 cpus_and(tmp
, tmp
, domainspan
);
6113 if (cpus_empty(tmp
))
6116 nodemask
= node_to_cpumask(n
);
6117 cpus_and(tmp
, tmp
, nodemask
);
6118 if (cpus_empty(tmp
))
6121 sg
= kmalloc_node(sizeof(struct sched_group
),
6125 "Can not alloc domain group for node %d\n", j
);
6128 sg
->__cpu_power
= 0;
6130 sg
->next
= prev
->next
;
6131 cpus_or(covered
, covered
, tmp
);
6138 /* Calculate CPU power for physical packages and nodes */
6139 #ifdef CONFIG_SCHED_SMT
6140 for_each_cpu_mask(i
, *cpu_map
) {
6141 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6143 init_sched_groups_power(i
, sd
);
6146 #ifdef CONFIG_SCHED_MC
6147 for_each_cpu_mask(i
, *cpu_map
) {
6148 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6150 init_sched_groups_power(i
, sd
);
6154 for_each_cpu_mask(i
, *cpu_map
) {
6155 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6157 init_sched_groups_power(i
, sd
);
6161 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6162 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6165 struct sched_group
*sg
;
6167 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6168 init_numa_sched_groups_power(sg
);
6172 /* Attach the domains */
6173 for_each_cpu_mask(i
, *cpu_map
) {
6174 struct sched_domain
*sd
;
6175 #ifdef CONFIG_SCHED_SMT
6176 sd
= &per_cpu(cpu_domains
, i
);
6177 #elif defined(CONFIG_SCHED_MC)
6178 sd
= &per_cpu(core_domains
, i
);
6180 sd
= &per_cpu(phys_domains
, i
);
6182 cpu_attach_domain(sd
, i
);
6189 free_sched_groups(cpu_map
);
6194 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6196 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6198 cpumask_t cpu_default_map
;
6202 * Setup mask for cpus without special case scheduling requirements.
6203 * For now this just excludes isolated cpus, but could be used to
6204 * exclude other special cases in the future.
6206 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6208 err
= build_sched_domains(&cpu_default_map
);
6213 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6215 free_sched_groups(cpu_map
);
6219 * Detach sched domains from a group of cpus specified in cpu_map
6220 * These cpus will now be attached to the NULL domain
6222 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6226 for_each_cpu_mask(i
, *cpu_map
)
6227 cpu_attach_domain(NULL
, i
);
6228 synchronize_sched();
6229 arch_destroy_sched_domains(cpu_map
);
6233 * Partition sched domains as specified by the cpumasks below.
6234 * This attaches all cpus from the cpumasks to the NULL domain,
6235 * waits for a RCU quiescent period, recalculates sched
6236 * domain information and then attaches them back to the
6237 * correct sched domains
6238 * Call with hotplug lock held
6240 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6242 cpumask_t change_map
;
6245 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6246 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6247 cpus_or(change_map
, *partition1
, *partition2
);
6249 /* Detach sched domains from all of the affected cpus */
6250 detach_destroy_domains(&change_map
);
6251 if (!cpus_empty(*partition1
))
6252 err
= build_sched_domains(partition1
);
6253 if (!err
&& !cpus_empty(*partition2
))
6254 err
= build_sched_domains(partition2
);
6259 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6260 static int arch_reinit_sched_domains(void)
6264 mutex_lock(&sched_hotcpu_mutex
);
6265 detach_destroy_domains(&cpu_online_map
);
6266 err
= arch_init_sched_domains(&cpu_online_map
);
6267 mutex_unlock(&sched_hotcpu_mutex
);
6272 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6276 if (buf
[0] != '0' && buf
[0] != '1')
6280 sched_smt_power_savings
= (buf
[0] == '1');
6282 sched_mc_power_savings
= (buf
[0] == '1');
6284 ret
= arch_reinit_sched_domains();
6286 return ret
? ret
: count
;
6289 #ifdef CONFIG_SCHED_MC
6290 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6292 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6294 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6295 const char *buf
, size_t count
)
6297 return sched_power_savings_store(buf
, count
, 0);
6299 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6300 sched_mc_power_savings_store
);
6303 #ifdef CONFIG_SCHED_SMT
6304 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6306 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6308 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6309 const char *buf
, size_t count
)
6311 return sched_power_savings_store(buf
, count
, 1);
6313 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6314 sched_smt_power_savings_store
);
6317 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6321 #ifdef CONFIG_SCHED_SMT
6323 err
= sysfs_create_file(&cls
->kset
.kobj
,
6324 &attr_sched_smt_power_savings
.attr
);
6326 #ifdef CONFIG_SCHED_MC
6327 if (!err
&& mc_capable())
6328 err
= sysfs_create_file(&cls
->kset
.kobj
,
6329 &attr_sched_mc_power_savings
.attr
);
6336 * Force a reinitialization of the sched domains hierarchy. The domains
6337 * and groups cannot be updated in place without racing with the balancing
6338 * code, so we temporarily attach all running cpus to the NULL domain
6339 * which will prevent rebalancing while the sched domains are recalculated.
6341 static int update_sched_domains(struct notifier_block
*nfb
,
6342 unsigned long action
, void *hcpu
)
6345 case CPU_UP_PREPARE
:
6346 case CPU_UP_PREPARE_FROZEN
:
6347 case CPU_DOWN_PREPARE
:
6348 case CPU_DOWN_PREPARE_FROZEN
:
6349 detach_destroy_domains(&cpu_online_map
);
6352 case CPU_UP_CANCELED
:
6353 case CPU_UP_CANCELED_FROZEN
:
6354 case CPU_DOWN_FAILED
:
6355 case CPU_DOWN_FAILED_FROZEN
:
6357 case CPU_ONLINE_FROZEN
:
6359 case CPU_DEAD_FROZEN
:
6361 * Fall through and re-initialise the domains.
6368 /* The hotplug lock is already held by cpu_up/cpu_down */
6369 arch_init_sched_domains(&cpu_online_map
);
6374 void __init
sched_init_smp(void)
6376 cpumask_t non_isolated_cpus
;
6378 mutex_lock(&sched_hotcpu_mutex
);
6379 arch_init_sched_domains(&cpu_online_map
);
6380 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6381 if (cpus_empty(non_isolated_cpus
))
6382 cpu_set(smp_processor_id(), non_isolated_cpus
);
6383 mutex_unlock(&sched_hotcpu_mutex
);
6384 /* XXX: Theoretical race here - CPU may be hotplugged now */
6385 hotcpu_notifier(update_sched_domains
, 0);
6387 init_sched_domain_sysctl();
6389 /* Move init over to a non-isolated CPU */
6390 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6394 void __init
sched_init_smp(void)
6397 #endif /* CONFIG_SMP */
6399 int in_sched_functions(unsigned long addr
)
6401 /* Linker adds these: start and end of __sched functions */
6402 extern char __sched_text_start
[], __sched_text_end
[];
6404 return in_lock_functions(addr
) ||
6405 (addr
>= (unsigned long)__sched_text_start
6406 && addr
< (unsigned long)__sched_text_end
);
6409 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6411 cfs_rq
->tasks_timeline
= RB_ROOT
;
6412 #ifdef CONFIG_FAIR_GROUP_SCHED
6415 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6418 void __init
sched_init(void)
6420 int highest_cpu
= 0;
6423 for_each_possible_cpu(i
) {
6424 struct rt_prio_array
*array
;
6428 spin_lock_init(&rq
->lock
);
6429 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6432 init_cfs_rq(&rq
->cfs
, rq
);
6433 #ifdef CONFIG_FAIR_GROUP_SCHED
6434 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6436 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6437 struct sched_entity
*se
=
6438 &per_cpu(init_sched_entity
, i
);
6440 init_cfs_rq_p
[i
] = cfs_rq
;
6441 init_cfs_rq(cfs_rq
, rq
);
6442 cfs_rq
->tg
= &init_task_group
;
6443 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6444 &rq
->leaf_cfs_rq_list
);
6446 init_sched_entity_p
[i
] = se
;
6447 se
->cfs_rq
= &rq
->cfs
;
6449 se
->load
.weight
= init_task_group_load
;
6450 se
->load
.inv_weight
=
6451 div64_64(1ULL<<32, init_task_group_load
);
6454 init_task_group
.shares
= init_task_group_load
;
6455 spin_lock_init(&init_task_group
.lock
);
6458 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6459 rq
->cpu_load
[j
] = 0;
6462 rq
->active_balance
= 0;
6463 rq
->next_balance
= jiffies
;
6466 rq
->migration_thread
= NULL
;
6467 INIT_LIST_HEAD(&rq
->migration_queue
);
6469 atomic_set(&rq
->nr_iowait
, 0);
6471 array
= &rq
->rt
.active
;
6472 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6473 INIT_LIST_HEAD(array
->queue
+ j
);
6474 __clear_bit(j
, array
->bitmap
);
6477 /* delimiter for bitsearch: */
6478 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6481 set_load_weight(&init_task
);
6483 #ifdef CONFIG_PREEMPT_NOTIFIERS
6484 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6488 nr_cpu_ids
= highest_cpu
+ 1;
6489 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6492 #ifdef CONFIG_RT_MUTEXES
6493 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6497 * The boot idle thread does lazy MMU switching as well:
6499 atomic_inc(&init_mm
.mm_count
);
6500 enter_lazy_tlb(&init_mm
, current
);
6503 * Make us the idle thread. Technically, schedule() should not be
6504 * called from this thread, however somewhere below it might be,
6505 * but because we are the idle thread, we just pick up running again
6506 * when this runqueue becomes "idle".
6508 init_idle(current
, smp_processor_id());
6510 * During early bootup we pretend to be a normal task:
6512 current
->sched_class
= &fair_sched_class
;
6515 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6516 void __might_sleep(char *file
, int line
)
6519 static unsigned long prev_jiffy
; /* ratelimiting */
6521 if ((in_atomic() || irqs_disabled()) &&
6522 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6523 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6525 prev_jiffy
= jiffies
;
6526 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6527 " context at %s:%d\n", file
, line
);
6528 printk("in_atomic():%d, irqs_disabled():%d\n",
6529 in_atomic(), irqs_disabled());
6530 debug_show_held_locks(current
);
6531 if (irqs_disabled())
6532 print_irqtrace_events(current
);
6537 EXPORT_SYMBOL(__might_sleep
);
6540 #ifdef CONFIG_MAGIC_SYSRQ
6541 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6544 update_rq_clock(rq
);
6545 on_rq
= p
->se
.on_rq
;
6547 deactivate_task(rq
, p
, 0);
6548 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6550 activate_task(rq
, p
, 0);
6551 resched_task(rq
->curr
);
6555 void normalize_rt_tasks(void)
6557 struct task_struct
*g
, *p
;
6558 unsigned long flags
;
6561 read_lock_irq(&tasklist_lock
);
6562 do_each_thread(g
, p
) {
6563 p
->se
.exec_start
= 0;
6564 #ifdef CONFIG_SCHEDSTATS
6565 p
->se
.wait_start
= 0;
6566 p
->se
.sleep_start
= 0;
6567 p
->se
.block_start
= 0;
6569 task_rq(p
)->clock
= 0;
6573 * Renice negative nice level userspace
6576 if (TASK_NICE(p
) < 0 && p
->mm
)
6577 set_user_nice(p
, 0);
6581 spin_lock_irqsave(&p
->pi_lock
, flags
);
6582 rq
= __task_rq_lock(p
);
6584 if (!is_migration_thread(p
, rq
))
6585 normalize_task(rq
, p
);
6587 __task_rq_unlock(rq
);
6588 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6589 } while_each_thread(g
, p
);
6591 read_unlock_irq(&tasklist_lock
);
6594 #endif /* CONFIG_MAGIC_SYSRQ */
6598 * These functions are only useful for the IA64 MCA handling.
6600 * They can only be called when the whole system has been
6601 * stopped - every CPU needs to be quiescent, and no scheduling
6602 * activity can take place. Using them for anything else would
6603 * be a serious bug, and as a result, they aren't even visible
6604 * under any other configuration.
6608 * curr_task - return the current task for a given cpu.
6609 * @cpu: the processor in question.
6611 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6613 struct task_struct
*curr_task(int cpu
)
6615 return cpu_curr(cpu
);
6619 * set_curr_task - set the current task for a given cpu.
6620 * @cpu: the processor in question.
6621 * @p: the task pointer to set.
6623 * Description: This function must only be used when non-maskable interrupts
6624 * are serviced on a separate stack. It allows the architecture to switch the
6625 * notion of the current task on a cpu in a non-blocking manner. This function
6626 * must be called with all CPU's synchronized, and interrupts disabled, the
6627 * and caller must save the original value of the current task (see
6628 * curr_task() above) and restore that value before reenabling interrupts and
6629 * re-starting the system.
6631 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6633 void set_curr_task(int cpu
, struct task_struct
*p
)
6640 #ifdef CONFIG_FAIR_GROUP_SCHED
6642 /* allocate runqueue etc for a new task group */
6643 struct task_group
*sched_create_group(void)
6645 struct task_group
*tg
;
6646 struct cfs_rq
*cfs_rq
;
6647 struct sched_entity
*se
;
6651 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6653 return ERR_PTR(-ENOMEM
);
6655 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
6658 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
6662 for_each_possible_cpu(i
) {
6665 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
6670 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
6675 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
6676 memset(se
, 0, sizeof(struct sched_entity
));
6678 tg
->cfs_rq
[i
] = cfs_rq
;
6679 init_cfs_rq(cfs_rq
, rq
);
6683 se
->cfs_rq
= &rq
->cfs
;
6685 se
->load
.weight
= NICE_0_LOAD
;
6686 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6690 for_each_possible_cpu(i
) {
6692 cfs_rq
= tg
->cfs_rq
[i
];
6693 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6696 tg
->shares
= NICE_0_LOAD
;
6697 spin_lock_init(&tg
->lock
);
6702 for_each_possible_cpu(i
) {
6704 kfree(tg
->cfs_rq
[i
]);
6712 return ERR_PTR(-ENOMEM
);
6715 /* rcu callback to free various structures associated with a task group */
6716 static void free_sched_group(struct rcu_head
*rhp
)
6718 struct cfs_rq
*cfs_rq
= container_of(rhp
, struct cfs_rq
, rcu
);
6719 struct task_group
*tg
= cfs_rq
->tg
;
6720 struct sched_entity
*se
;
6723 /* now it should be safe to free those cfs_rqs */
6724 for_each_possible_cpu(i
) {
6725 cfs_rq
= tg
->cfs_rq
[i
];
6737 /* Destroy runqueue etc associated with a task group */
6738 void sched_destroy_group(struct task_group
*tg
)
6740 struct cfs_rq
*cfs_rq
;
6743 for_each_possible_cpu(i
) {
6744 cfs_rq
= tg
->cfs_rq
[i
];
6745 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
6748 cfs_rq
= tg
->cfs_rq
[0];
6750 /* wait for possible concurrent references to cfs_rqs complete */
6751 call_rcu(&cfs_rq
->rcu
, free_sched_group
);
6754 /* change task's runqueue when it moves between groups.
6755 * The caller of this function should have put the task in its new group
6756 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6757 * reflect its new group.
6759 void sched_move_task(struct task_struct
*tsk
)
6762 unsigned long flags
;
6765 rq
= task_rq_lock(tsk
, &flags
);
6767 if (tsk
->sched_class
!= &fair_sched_class
)
6770 update_rq_clock(rq
);
6772 running
= task_running(rq
, tsk
);
6773 on_rq
= tsk
->se
.on_rq
;
6776 dequeue_task(rq
, tsk
, 0);
6777 if (unlikely(running
))
6778 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6781 set_task_cfs_rq(tsk
);
6784 if (unlikely(running
))
6785 tsk
->sched_class
->set_curr_task(rq
);
6786 enqueue_task(rq
, tsk
, 0);
6790 task_rq_unlock(rq
, &flags
);
6793 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
6795 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
6796 struct rq
*rq
= cfs_rq
->rq
;
6799 spin_lock_irq(&rq
->lock
);
6803 dequeue_entity(cfs_rq
, se
, 0);
6805 se
->load
.weight
= shares
;
6806 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
6809 enqueue_entity(cfs_rq
, se
, 0);
6811 spin_unlock_irq(&rq
->lock
);
6814 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6818 spin_lock(&tg
->lock
);
6819 if (tg
->shares
== shares
)
6822 /* return -EINVAL if the new value is not sane */
6824 tg
->shares
= shares
;
6825 for_each_possible_cpu(i
)
6826 set_se_shares(tg
->se
[i
], shares
);
6829 spin_unlock(&tg
->lock
);
6833 unsigned long sched_group_shares(struct task_group
*tg
)
6838 #endif /* CONFIG_FAIR_GROUP_SCHED */