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 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
116 * Since cpu_power is a 'constant', we can use a reciprocal divide.
118 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
120 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
124 * Each time a sched group cpu_power is changed,
125 * we must compute its reciprocal value
127 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
129 sg
->__cpu_power
+= val
;
130 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
134 static inline int rt_policy(int policy
)
136 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
141 static inline int task_has_rt_policy(struct task_struct
*p
)
143 return rt_policy(p
->policy
);
147 * This is the priority-queue data structure of the RT scheduling class:
149 struct rt_prio_array
{
150 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
151 struct list_head queue
[MAX_RT_PRIO
];
154 #ifdef CONFIG_FAIR_GROUP_SCHED
158 /* task group related information */
160 /* schedulable entities of this group on each cpu */
161 struct sched_entity
**se
;
162 /* runqueue "owned" by this group on each cpu */
163 struct cfs_rq
**cfs_rq
;
164 unsigned long shares
;
165 /* spinlock to serialize modification to shares */
169 /* Default task group's sched entity on each cpu */
170 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
171 /* Default task group's cfs_rq on each cpu */
172 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
174 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
175 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
177 /* Default task group.
178 * Every task in system belong to this group at bootup.
180 struct task_group init_task_group
= {
181 .se
= init_sched_entity_p
,
182 .cfs_rq
= init_cfs_rq_p
,
185 #ifdef CONFIG_FAIR_USER_SCHED
186 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
188 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
191 static int init_task_group_load
= INIT_TASK_GRP_LOAD
;
193 /* return group to which a task belongs */
194 static inline struct task_group
*task_group(struct task_struct
*p
)
196 struct task_group
*tg
;
198 #ifdef CONFIG_FAIR_USER_SCHED
201 tg
= &init_task_group
;
207 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
208 static inline void set_task_cfs_rq(struct task_struct
*p
)
210 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[task_cpu(p
)];
211 p
->se
.parent
= task_group(p
)->se
[task_cpu(p
)];
216 static inline void set_task_cfs_rq(struct task_struct
*p
) { }
218 #endif /* CONFIG_FAIR_GROUP_SCHED */
220 /* CFS-related fields in a runqueue */
222 struct load_weight load
;
223 unsigned long nr_running
;
228 struct rb_root tasks_timeline
;
229 struct rb_node
*rb_leftmost
;
230 struct rb_node
*rb_load_balance_curr
;
231 /* 'curr' points to currently running entity on this cfs_rq.
232 * It is set to NULL otherwise (i.e when none are currently running).
234 struct sched_entity
*curr
;
236 unsigned long nr_spread_over
;
238 #ifdef CONFIG_FAIR_GROUP_SCHED
239 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
241 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
242 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
243 * (like users, containers etc.)
245 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
246 * list is used during load balance.
248 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
249 struct task_group
*tg
; /* group that "owns" this runqueue */
254 /* Real-Time classes' related field in a runqueue: */
256 struct rt_prio_array active
;
257 int rt_load_balance_idx
;
258 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
262 * This is the main, per-CPU runqueue data structure.
264 * Locking rule: those places that want to lock multiple runqueues
265 * (such as the load balancing or the thread migration code), lock
266 * acquire operations must be ordered by ascending &runqueue.
269 spinlock_t lock
; /* runqueue lock */
272 * nr_running and cpu_load should be in the same cacheline because
273 * remote CPUs use both these fields when doing load calculation.
275 unsigned long nr_running
;
276 #define CPU_LOAD_IDX_MAX 5
277 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
278 unsigned char idle_at_tick
;
280 unsigned char in_nohz_recently
;
282 struct load_weight load
; /* capture load from *all* tasks on this cpu */
283 unsigned long nr_load_updates
;
287 #ifdef CONFIG_FAIR_GROUP_SCHED
288 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
293 * This is part of a global counter where only the total sum
294 * over all CPUs matters. A task can increase this counter on
295 * one CPU and if it got migrated afterwards it may decrease
296 * it on another CPU. Always updated under the runqueue lock:
298 unsigned long nr_uninterruptible
;
300 struct task_struct
*curr
, *idle
;
301 unsigned long next_balance
;
302 struct mm_struct
*prev_mm
;
304 u64 clock
, prev_clock_raw
;
307 unsigned int clock_warps
, clock_overflows
;
309 unsigned int clock_deep_idle_events
;
315 struct sched_domain
*sd
;
317 /* For active balancing */
320 int cpu
; /* cpu of this runqueue */
322 struct task_struct
*migration_thread
;
323 struct list_head migration_queue
;
326 #ifdef CONFIG_SCHEDSTATS
328 struct sched_info rq_sched_info
;
330 /* sys_sched_yield() stats */
331 unsigned long yld_exp_empty
;
332 unsigned long yld_act_empty
;
333 unsigned long yld_both_empty
;
334 unsigned long yld_count
;
336 /* schedule() stats */
337 unsigned long sched_switch
;
338 unsigned long sched_count
;
339 unsigned long sched_goidle
;
341 /* try_to_wake_up() stats */
342 unsigned long ttwu_count
;
343 unsigned long ttwu_local
;
346 unsigned long bkl_count
;
348 struct lock_class_key rq_lock_key
;
351 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
352 static DEFINE_MUTEX(sched_hotcpu_mutex
);
354 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
356 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
359 static inline int cpu_of(struct rq
*rq
)
369 * Update the per-runqueue clock, as finegrained as the platform can give
370 * us, but without assuming monotonicity, etc.:
372 static void __update_rq_clock(struct rq
*rq
)
374 u64 prev_raw
= rq
->prev_clock_raw
;
375 u64 now
= sched_clock();
376 s64 delta
= now
- prev_raw
;
377 u64 clock
= rq
->clock
;
379 #ifdef CONFIG_SCHED_DEBUG
380 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
383 * Protect against sched_clock() occasionally going backwards:
385 if (unlikely(delta
< 0)) {
390 * Catch too large forward jumps too:
392 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
393 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
394 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
397 rq
->clock_overflows
++;
399 if (unlikely(delta
> rq
->clock_max_delta
))
400 rq
->clock_max_delta
= delta
;
405 rq
->prev_clock_raw
= now
;
409 static void update_rq_clock(struct rq
*rq
)
411 if (likely(smp_processor_id() == cpu_of(rq
)))
412 __update_rq_clock(rq
);
416 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
417 * See detach_destroy_domains: synchronize_sched for details.
419 * The domain tree of any CPU may only be accessed from within
420 * preempt-disabled sections.
422 #define for_each_domain(cpu, __sd) \
423 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
425 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
426 #define this_rq() (&__get_cpu_var(runqueues))
427 #define task_rq(p) cpu_rq(task_cpu(p))
428 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
431 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
433 #ifdef CONFIG_SCHED_DEBUG
434 # define const_debug __read_mostly
436 # define const_debug static const
440 * Debugging: various feature bits
443 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
444 SCHED_FEAT_START_DEBIT
= 2,
445 SCHED_FEAT_TREE_AVG
= 4,
446 SCHED_FEAT_APPROX_AVG
= 8,
447 SCHED_FEAT_WAKEUP_PREEMPT
= 16,
448 SCHED_FEAT_PREEMPT_RESTRICT
= 32,
451 const_debug
unsigned int sysctl_sched_features
=
452 SCHED_FEAT_NEW_FAIR_SLEEPERS
*1 |
453 SCHED_FEAT_START_DEBIT
*1 |
454 SCHED_FEAT_TREE_AVG
*0 |
455 SCHED_FEAT_APPROX_AVG
*0 |
456 SCHED_FEAT_WAKEUP_PREEMPT
*1 |
457 SCHED_FEAT_PREEMPT_RESTRICT
*1;
459 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
462 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
463 * clock constructed from sched_clock():
465 unsigned long long cpu_clock(int cpu
)
467 unsigned long long now
;
471 local_irq_save(flags
);
475 local_irq_restore(flags
);
479 EXPORT_SYMBOL_GPL(cpu_clock
);
481 #ifndef prepare_arch_switch
482 # define prepare_arch_switch(next) do { } while (0)
484 #ifndef finish_arch_switch
485 # define finish_arch_switch(prev) do { } while (0)
488 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
489 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
491 return rq
->curr
== p
;
494 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
498 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
500 #ifdef CONFIG_DEBUG_SPINLOCK
501 /* this is a valid case when another task releases the spinlock */
502 rq
->lock
.owner
= current
;
505 * If we are tracking spinlock dependencies then we have to
506 * fix up the runqueue lock - which gets 'carried over' from
509 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
511 spin_unlock_irq(&rq
->lock
);
514 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
515 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
520 return rq
->curr
== p
;
524 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
528 * We can optimise this out completely for !SMP, because the
529 * SMP rebalancing from interrupt is the only thing that cares
534 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
535 spin_unlock_irq(&rq
->lock
);
537 spin_unlock(&rq
->lock
);
541 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
545 * After ->oncpu is cleared, the task can be moved to a different CPU.
546 * We must ensure this doesn't happen until the switch is completely
552 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
556 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
559 * __task_rq_lock - lock the runqueue a given task resides on.
560 * Must be called interrupts disabled.
562 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
566 struct rq
*rq
= task_rq(p
);
567 spin_lock(&rq
->lock
);
568 if (likely(rq
== task_rq(p
)))
570 spin_unlock(&rq
->lock
);
575 * task_rq_lock - lock the runqueue a given task resides on and disable
576 * interrupts. Note the ordering: we can safely lookup the task_rq without
577 * explicitly disabling preemption.
579 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
585 local_irq_save(*flags
);
587 spin_lock(&rq
->lock
);
588 if (likely(rq
== task_rq(p
)))
590 spin_unlock_irqrestore(&rq
->lock
, *flags
);
594 static void __task_rq_unlock(struct rq
*rq
)
597 spin_unlock(&rq
->lock
);
600 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
603 spin_unlock_irqrestore(&rq
->lock
, *flags
);
607 * this_rq_lock - lock this runqueue and disable interrupts.
609 static struct rq
*this_rq_lock(void)
616 spin_lock(&rq
->lock
);
622 * We are going deep-idle (irqs are disabled):
624 void sched_clock_idle_sleep_event(void)
626 struct rq
*rq
= cpu_rq(smp_processor_id());
628 spin_lock(&rq
->lock
);
629 __update_rq_clock(rq
);
630 spin_unlock(&rq
->lock
);
631 rq
->clock_deep_idle_events
++;
633 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
636 * We just idled delta nanoseconds (called with irqs disabled):
638 void sched_clock_idle_wakeup_event(u64 delta_ns
)
640 struct rq
*rq
= cpu_rq(smp_processor_id());
641 u64 now
= sched_clock();
643 rq
->idle_clock
+= delta_ns
;
645 * Override the previous timestamp and ignore all
646 * sched_clock() deltas that occured while we idled,
647 * and use the PM-provided delta_ns to advance the
650 spin_lock(&rq
->lock
);
651 rq
->prev_clock_raw
= now
;
652 rq
->clock
+= delta_ns
;
653 spin_unlock(&rq
->lock
);
655 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
658 * resched_task - mark a task 'to be rescheduled now'.
660 * On UP this means the setting of the need_resched flag, on SMP it
661 * might also involve a cross-CPU call to trigger the scheduler on
666 #ifndef tsk_is_polling
667 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
670 static void resched_task(struct task_struct
*p
)
674 assert_spin_locked(&task_rq(p
)->lock
);
676 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
679 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
682 if (cpu
== smp_processor_id())
685 /* NEED_RESCHED must be visible before we test polling */
687 if (!tsk_is_polling(p
))
688 smp_send_reschedule(cpu
);
691 static void resched_cpu(int cpu
)
693 struct rq
*rq
= cpu_rq(cpu
);
696 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
698 resched_task(cpu_curr(cpu
));
699 spin_unlock_irqrestore(&rq
->lock
, flags
);
702 static inline void resched_task(struct task_struct
*p
)
704 assert_spin_locked(&task_rq(p
)->lock
);
705 set_tsk_need_resched(p
);
709 #if BITS_PER_LONG == 32
710 # define WMULT_CONST (~0UL)
712 # define WMULT_CONST (1UL << 32)
715 #define WMULT_SHIFT 32
718 * Shift right and round:
720 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
723 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
724 struct load_weight
*lw
)
728 if (unlikely(!lw
->inv_weight
))
729 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
731 tmp
= (u64
)delta_exec
* weight
;
733 * Check whether we'd overflow the 64-bit multiplication:
735 if (unlikely(tmp
> WMULT_CONST
))
736 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
739 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
741 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
744 static inline unsigned long
745 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
747 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
750 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
755 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
761 * To aid in avoiding the subversion of "niceness" due to uneven distribution
762 * of tasks with abnormal "nice" values across CPUs the contribution that
763 * each task makes to its run queue's load is weighted according to its
764 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
765 * scaled version of the new time slice allocation that they receive on time
769 #define WEIGHT_IDLEPRIO 2
770 #define WMULT_IDLEPRIO (1 << 31)
773 * Nice levels are multiplicative, with a gentle 10% change for every
774 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
775 * nice 1, it will get ~10% less CPU time than another CPU-bound task
776 * that remained on nice 0.
778 * The "10% effect" is relative and cumulative: from _any_ nice level,
779 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
780 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
781 * If a task goes up by ~10% and another task goes down by ~10% then
782 * the relative distance between them is ~25%.)
784 static const int prio_to_weight
[40] = {
785 /* -20 */ 88761, 71755, 56483, 46273, 36291,
786 /* -15 */ 29154, 23254, 18705, 14949, 11916,
787 /* -10 */ 9548, 7620, 6100, 4904, 3906,
788 /* -5 */ 3121, 2501, 1991, 1586, 1277,
789 /* 0 */ 1024, 820, 655, 526, 423,
790 /* 5 */ 335, 272, 215, 172, 137,
791 /* 10 */ 110, 87, 70, 56, 45,
792 /* 15 */ 36, 29, 23, 18, 15,
796 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
798 * In cases where the weight does not change often, we can use the
799 * precalculated inverse to speed up arithmetics by turning divisions
800 * into multiplications:
802 static const u32 prio_to_wmult
[40] = {
803 /* -20 */ 48388, 59856, 76040, 92818, 118348,
804 /* -15 */ 147320, 184698, 229616, 287308, 360437,
805 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
806 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
807 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
808 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
809 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
810 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
813 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
816 * runqueue iterator, to support SMP load-balancing between different
817 * scheduling classes, without having to expose their internal data
818 * structures to the load-balancing proper:
822 struct task_struct
*(*start
)(void *);
823 struct task_struct
*(*next
)(void *);
826 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
827 unsigned long max_nr_move
, unsigned long max_load_move
,
828 struct sched_domain
*sd
, enum cpu_idle_type idle
,
829 int *all_pinned
, unsigned long *load_moved
,
830 int *this_best_prio
, struct rq_iterator
*iterator
);
832 #include "sched_stats.h"
833 #include "sched_idletask.c"
834 #include "sched_fair.c"
835 #include "sched_rt.c"
836 #ifdef CONFIG_SCHED_DEBUG
837 # include "sched_debug.c"
840 #define sched_class_highest (&rt_sched_class)
843 * Update delta_exec, delta_fair fields for rq.
845 * delta_fair clock advances at a rate inversely proportional to
846 * total load (rq->load.weight) on the runqueue, while
847 * delta_exec advances at the same rate as wall-clock (provided
850 * delta_exec / delta_fair is a measure of the (smoothened) load on this
851 * runqueue over any given interval. This (smoothened) load is used
852 * during load balance.
854 * This function is called /before/ updating rq->load
855 * and when switching tasks.
857 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
859 update_load_add(&rq
->load
, p
->se
.load
.weight
);
862 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
864 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
867 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
873 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
879 static void set_load_weight(struct task_struct
*p
)
881 if (task_has_rt_policy(p
)) {
882 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
883 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
888 * SCHED_IDLE tasks get minimal weight:
890 if (p
->policy
== SCHED_IDLE
) {
891 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
892 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
896 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
897 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
900 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
902 sched_info_queued(p
);
903 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
907 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
909 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
914 * __normal_prio - return the priority that is based on the static prio
916 static inline int __normal_prio(struct task_struct
*p
)
918 return p
->static_prio
;
922 * Calculate the expected normal priority: i.e. priority
923 * without taking RT-inheritance into account. Might be
924 * boosted by interactivity modifiers. Changes upon fork,
925 * setprio syscalls, and whenever the interactivity
926 * estimator recalculates.
928 static inline int normal_prio(struct task_struct
*p
)
932 if (task_has_rt_policy(p
))
933 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
935 prio
= __normal_prio(p
);
940 * Calculate the current priority, i.e. the priority
941 * taken into account by the scheduler. This value might
942 * be boosted by RT tasks, or might be boosted by
943 * interactivity modifiers. Will be RT if the task got
944 * RT-boosted. If not then it returns p->normal_prio.
946 static int effective_prio(struct task_struct
*p
)
948 p
->normal_prio
= normal_prio(p
);
950 * If we are RT tasks or we were boosted to RT priority,
951 * keep the priority unchanged. Otherwise, update priority
952 * to the normal priority:
954 if (!rt_prio(p
->prio
))
955 return p
->normal_prio
;
960 * activate_task - move a task to the runqueue.
962 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
964 if (p
->state
== TASK_UNINTERRUPTIBLE
)
965 rq
->nr_uninterruptible
--;
967 enqueue_task(rq
, p
, wakeup
);
968 inc_nr_running(p
, rq
);
972 * deactivate_task - remove a task from the runqueue.
974 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
976 if (p
->state
== TASK_UNINTERRUPTIBLE
)
977 rq
->nr_uninterruptible
++;
979 dequeue_task(rq
, p
, sleep
);
980 dec_nr_running(p
, rq
);
984 * task_curr - is this task currently executing on a CPU?
985 * @p: the task in question.
987 inline int task_curr(const struct task_struct
*p
)
989 return cpu_curr(task_cpu(p
)) == p
;
992 /* Used instead of source_load when we know the type == 0 */
993 unsigned long weighted_cpuload(const int cpu
)
995 return cpu_rq(cpu
)->load
.weight
;
998 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1001 task_thread_info(p
)->cpu
= cpu
;
1008 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1010 int old_cpu
= task_cpu(p
);
1011 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1012 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1013 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1016 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1018 #ifdef CONFIG_SCHEDSTATS
1019 if (p
->se
.wait_start
)
1020 p
->se
.wait_start
-= clock_offset
;
1021 if (p
->se
.sleep_start
)
1022 p
->se
.sleep_start
-= clock_offset
;
1023 if (p
->se
.block_start
)
1024 p
->se
.block_start
-= clock_offset
;
1026 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1027 new_cfsrq
->min_vruntime
;
1029 __set_task_cpu(p
, new_cpu
);
1032 struct migration_req
{
1033 struct list_head list
;
1035 struct task_struct
*task
;
1038 struct completion done
;
1042 * The task's runqueue lock must be held.
1043 * Returns true if you have to wait for migration thread.
1046 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1048 struct rq
*rq
= task_rq(p
);
1051 * If the task is not on a runqueue (and not running), then
1052 * it is sufficient to simply update the task's cpu field.
1054 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1055 set_task_cpu(p
, dest_cpu
);
1059 init_completion(&req
->done
);
1061 req
->dest_cpu
= dest_cpu
;
1062 list_add(&req
->list
, &rq
->migration_queue
);
1068 * wait_task_inactive - wait for a thread to unschedule.
1070 * The caller must ensure that the task *will* unschedule sometime soon,
1071 * else this function might spin for a *long* time. This function can't
1072 * be called with interrupts off, or it may introduce deadlock with
1073 * smp_call_function() if an IPI is sent by the same process we are
1074 * waiting to become inactive.
1076 void wait_task_inactive(struct task_struct
*p
)
1078 unsigned long flags
;
1084 * We do the initial early heuristics without holding
1085 * any task-queue locks at all. We'll only try to get
1086 * the runqueue lock when things look like they will
1092 * If the task is actively running on another CPU
1093 * still, just relax and busy-wait without holding
1096 * NOTE! Since we don't hold any locks, it's not
1097 * even sure that "rq" stays as the right runqueue!
1098 * But we don't care, since "task_running()" will
1099 * return false if the runqueue has changed and p
1100 * is actually now running somewhere else!
1102 while (task_running(rq
, p
))
1106 * Ok, time to look more closely! We need the rq
1107 * lock now, to be *sure*. If we're wrong, we'll
1108 * just go back and repeat.
1110 rq
= task_rq_lock(p
, &flags
);
1111 running
= task_running(rq
, p
);
1112 on_rq
= p
->se
.on_rq
;
1113 task_rq_unlock(rq
, &flags
);
1116 * Was it really running after all now that we
1117 * checked with the proper locks actually held?
1119 * Oops. Go back and try again..
1121 if (unlikely(running
)) {
1127 * It's not enough that it's not actively running,
1128 * it must be off the runqueue _entirely_, and not
1131 * So if it wa still runnable (but just not actively
1132 * running right now), it's preempted, and we should
1133 * yield - it could be a while.
1135 if (unlikely(on_rq
)) {
1136 schedule_timeout_uninterruptible(1);
1141 * Ahh, all good. It wasn't running, and it wasn't
1142 * runnable, which means that it will never become
1143 * running in the future either. We're all done!
1150 * kick_process - kick a running thread to enter/exit the kernel
1151 * @p: the to-be-kicked thread
1153 * Cause a process which is running on another CPU to enter
1154 * kernel-mode, without any delay. (to get signals handled.)
1156 * NOTE: this function doesnt have to take the runqueue lock,
1157 * because all it wants to ensure is that the remote task enters
1158 * the kernel. If the IPI races and the task has been migrated
1159 * to another CPU then no harm is done and the purpose has been
1162 void kick_process(struct task_struct
*p
)
1168 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1169 smp_send_reschedule(cpu
);
1174 * Return a low guess at the load of a migration-source cpu weighted
1175 * according to the scheduling class and "nice" value.
1177 * We want to under-estimate the load of migration sources, to
1178 * balance conservatively.
1180 static unsigned long source_load(int cpu
, int type
)
1182 struct rq
*rq
= cpu_rq(cpu
);
1183 unsigned long total
= weighted_cpuload(cpu
);
1188 return min(rq
->cpu_load
[type
-1], total
);
1192 * Return a high guess at the load of a migration-target cpu weighted
1193 * according to the scheduling class and "nice" value.
1195 static unsigned long target_load(int cpu
, int type
)
1197 struct rq
*rq
= cpu_rq(cpu
);
1198 unsigned long total
= weighted_cpuload(cpu
);
1203 return max(rq
->cpu_load
[type
-1], total
);
1207 * Return the average load per task on the cpu's run queue
1209 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1211 struct rq
*rq
= cpu_rq(cpu
);
1212 unsigned long total
= weighted_cpuload(cpu
);
1213 unsigned long n
= rq
->nr_running
;
1215 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1219 * find_idlest_group finds and returns the least busy CPU group within the
1222 static struct sched_group
*
1223 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1225 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1226 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1227 int load_idx
= sd
->forkexec_idx
;
1228 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1231 unsigned long load
, avg_load
;
1235 /* Skip over this group if it has no CPUs allowed */
1236 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1239 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1241 /* Tally up the load of all CPUs in the group */
1244 for_each_cpu_mask(i
, group
->cpumask
) {
1245 /* Bias balancing toward cpus of our domain */
1247 load
= source_load(i
, load_idx
);
1249 load
= target_load(i
, load_idx
);
1254 /* Adjust by relative CPU power of the group */
1255 avg_load
= sg_div_cpu_power(group
,
1256 avg_load
* SCHED_LOAD_SCALE
);
1259 this_load
= avg_load
;
1261 } else if (avg_load
< min_load
) {
1262 min_load
= avg_load
;
1265 } while (group
= group
->next
, group
!= sd
->groups
);
1267 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1273 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1276 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1279 unsigned long load
, min_load
= ULONG_MAX
;
1283 /* Traverse only the allowed CPUs */
1284 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1286 for_each_cpu_mask(i
, tmp
) {
1287 load
= weighted_cpuload(i
);
1289 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1299 * sched_balance_self: balance the current task (running on cpu) in domains
1300 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1303 * Balance, ie. select the least loaded group.
1305 * Returns the target CPU number, or the same CPU if no balancing is needed.
1307 * preempt must be disabled.
1309 static int sched_balance_self(int cpu
, int flag
)
1311 struct task_struct
*t
= current
;
1312 struct sched_domain
*tmp
, *sd
= NULL
;
1314 for_each_domain(cpu
, tmp
) {
1316 * If power savings logic is enabled for a domain, stop there.
1318 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1320 if (tmp
->flags
& flag
)
1326 struct sched_group
*group
;
1327 int new_cpu
, weight
;
1329 if (!(sd
->flags
& flag
)) {
1335 group
= find_idlest_group(sd
, t
, cpu
);
1341 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1342 if (new_cpu
== -1 || new_cpu
== cpu
) {
1343 /* Now try balancing at a lower domain level of cpu */
1348 /* Now try balancing at a lower domain level of new_cpu */
1351 weight
= cpus_weight(span
);
1352 for_each_domain(cpu
, tmp
) {
1353 if (weight
<= cpus_weight(tmp
->span
))
1355 if (tmp
->flags
& flag
)
1358 /* while loop will break here if sd == NULL */
1364 #endif /* CONFIG_SMP */
1367 * wake_idle() will wake a task on an idle cpu if task->cpu is
1368 * not idle and an idle cpu is available. The span of cpus to
1369 * search starts with cpus closest then further out as needed,
1370 * so we always favor a closer, idle cpu.
1372 * Returns the CPU we should wake onto.
1374 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1375 static int wake_idle(int cpu
, struct task_struct
*p
)
1378 struct sched_domain
*sd
;
1382 * If it is idle, then it is the best cpu to run this task.
1384 * This cpu is also the best, if it has more than one task already.
1385 * Siblings must be also busy(in most cases) as they didn't already
1386 * pickup the extra load from this cpu and hence we need not check
1387 * sibling runqueue info. This will avoid the checks and cache miss
1388 * penalities associated with that.
1390 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1393 for_each_domain(cpu
, sd
) {
1394 if (sd
->flags
& SD_WAKE_IDLE
) {
1395 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1396 for_each_cpu_mask(i
, tmp
) {
1407 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1414 * try_to_wake_up - wake up a thread
1415 * @p: the to-be-woken-up thread
1416 * @state: the mask of task states that can be woken
1417 * @sync: do a synchronous wakeup?
1419 * Put it on the run-queue if it's not already there. The "current"
1420 * thread is always on the run-queue (except when the actual
1421 * re-schedule is in progress), and as such you're allowed to do
1422 * the simpler "current->state = TASK_RUNNING" to mark yourself
1423 * runnable without the overhead of this.
1425 * returns failure only if the task is already active.
1427 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1429 int cpu
, this_cpu
, success
= 0;
1430 unsigned long flags
;
1434 struct sched_domain
*sd
, *this_sd
= NULL
;
1435 unsigned long load
, this_load
;
1439 rq
= task_rq_lock(p
, &flags
);
1440 old_state
= p
->state
;
1441 if (!(old_state
& state
))
1448 this_cpu
= smp_processor_id();
1451 if (unlikely(task_running(rq
, p
)))
1456 schedstat_inc(rq
, ttwu_count
);
1457 if (cpu
== this_cpu
) {
1458 schedstat_inc(rq
, ttwu_local
);
1462 for_each_domain(this_cpu
, sd
) {
1463 if (cpu_isset(cpu
, sd
->span
)) {
1464 schedstat_inc(sd
, ttwu_wake_remote
);
1470 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1474 * Check for affine wakeup and passive balancing possibilities.
1477 int idx
= this_sd
->wake_idx
;
1478 unsigned int imbalance
;
1480 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1482 load
= source_load(cpu
, idx
);
1483 this_load
= target_load(this_cpu
, idx
);
1485 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1487 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1488 unsigned long tl
= this_load
;
1489 unsigned long tl_per_task
;
1491 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1494 * If sync wakeup then subtract the (maximum possible)
1495 * effect of the currently running task from the load
1496 * of the current CPU:
1499 tl
-= current
->se
.load
.weight
;
1502 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1503 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1505 * This domain has SD_WAKE_AFFINE and
1506 * p is cache cold in this domain, and
1507 * there is no bad imbalance.
1509 schedstat_inc(this_sd
, ttwu_move_affine
);
1515 * Start passive balancing when half the imbalance_pct
1518 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1519 if (imbalance
*this_load
<= 100*load
) {
1520 schedstat_inc(this_sd
, ttwu_move_balance
);
1526 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1528 new_cpu
= wake_idle(new_cpu
, p
);
1529 if (new_cpu
!= cpu
) {
1530 set_task_cpu(p
, new_cpu
);
1531 task_rq_unlock(rq
, &flags
);
1532 /* might preempt at this point */
1533 rq
= task_rq_lock(p
, &flags
);
1534 old_state
= p
->state
;
1535 if (!(old_state
& state
))
1540 this_cpu
= smp_processor_id();
1545 #endif /* CONFIG_SMP */
1546 update_rq_clock(rq
);
1547 activate_task(rq
, p
, 1);
1549 * Sync wakeups (i.e. those types of wakeups where the waker
1550 * has indicated that it will leave the CPU in short order)
1551 * don't trigger a preemption, if the woken up task will run on
1552 * this cpu. (in this case the 'I will reschedule' promise of
1553 * the waker guarantees that the freshly woken up task is going
1554 * to be considered on this CPU.)
1556 if (!sync
|| cpu
!= this_cpu
)
1557 check_preempt_curr(rq
, p
);
1561 p
->state
= TASK_RUNNING
;
1563 task_rq_unlock(rq
, &flags
);
1568 int fastcall
wake_up_process(struct task_struct
*p
)
1570 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1571 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1573 EXPORT_SYMBOL(wake_up_process
);
1575 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1577 return try_to_wake_up(p
, state
, 0);
1581 * Perform scheduler related setup for a newly forked process p.
1582 * p is forked by current.
1584 * __sched_fork() is basic setup used by init_idle() too:
1586 static void __sched_fork(struct task_struct
*p
)
1588 p
->se
.exec_start
= 0;
1589 p
->se
.sum_exec_runtime
= 0;
1590 p
->se
.prev_sum_exec_runtime
= 0;
1592 #ifdef CONFIG_SCHEDSTATS
1593 p
->se
.wait_start
= 0;
1594 p
->se
.sum_sleep_runtime
= 0;
1595 p
->se
.sleep_start
= 0;
1596 p
->se
.block_start
= 0;
1597 p
->se
.sleep_max
= 0;
1598 p
->se
.block_max
= 0;
1600 p
->se
.slice_max
= 0;
1604 INIT_LIST_HEAD(&p
->run_list
);
1607 #ifdef CONFIG_PREEMPT_NOTIFIERS
1608 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1612 * We mark the process as running here, but have not actually
1613 * inserted it onto the runqueue yet. This guarantees that
1614 * nobody will actually run it, and a signal or other external
1615 * event cannot wake it up and insert it on the runqueue either.
1617 p
->state
= TASK_RUNNING
;
1621 * fork()/clone()-time setup:
1623 void sched_fork(struct task_struct
*p
, int clone_flags
)
1625 int cpu
= get_cpu();
1630 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1632 set_task_cpu(p
, cpu
);
1635 * Make sure we do not leak PI boosting priority to the child:
1637 p
->prio
= current
->normal_prio
;
1638 if (!rt_prio(p
->prio
))
1639 p
->sched_class
= &fair_sched_class
;
1641 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1642 if (likely(sched_info_on()))
1643 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1645 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1648 #ifdef CONFIG_PREEMPT
1649 /* Want to start with kernel preemption disabled. */
1650 task_thread_info(p
)->preempt_count
= 1;
1656 * wake_up_new_task - wake up a newly created task for the first time.
1658 * This function will do some initial scheduler statistics housekeeping
1659 * that must be done for every newly created context, then puts the task
1660 * on the runqueue and wakes it.
1662 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1664 unsigned long flags
;
1667 rq
= task_rq_lock(p
, &flags
);
1668 BUG_ON(p
->state
!= TASK_RUNNING
);
1669 update_rq_clock(rq
);
1671 p
->prio
= effective_prio(p
);
1673 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
|| !rq
->cfs
.curr
) {
1674 activate_task(rq
, p
, 0);
1677 * Let the scheduling class do new task startup
1678 * management (if any):
1680 p
->sched_class
->task_new(rq
, p
);
1681 inc_nr_running(p
, rq
);
1683 check_preempt_curr(rq
, p
);
1684 task_rq_unlock(rq
, &flags
);
1687 #ifdef CONFIG_PREEMPT_NOTIFIERS
1690 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1691 * @notifier: notifier struct to register
1693 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1695 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1697 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1700 * preempt_notifier_unregister - no longer interested in preemption notifications
1701 * @notifier: notifier struct to unregister
1703 * This is safe to call from within a preemption notifier.
1705 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1707 hlist_del(¬ifier
->link
);
1709 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1711 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1713 struct preempt_notifier
*notifier
;
1714 struct hlist_node
*node
;
1716 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1717 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1721 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1722 struct task_struct
*next
)
1724 struct preempt_notifier
*notifier
;
1725 struct hlist_node
*node
;
1727 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1728 notifier
->ops
->sched_out(notifier
, next
);
1733 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1738 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1739 struct task_struct
*next
)
1746 * prepare_task_switch - prepare to switch tasks
1747 * @rq: the runqueue preparing to switch
1748 * @prev: the current task that is being switched out
1749 * @next: the task we are going to switch to.
1751 * This is called with the rq lock held and interrupts off. It must
1752 * be paired with a subsequent finish_task_switch after the context
1755 * prepare_task_switch sets up locking and calls architecture specific
1759 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1760 struct task_struct
*next
)
1762 fire_sched_out_preempt_notifiers(prev
, next
);
1763 prepare_lock_switch(rq
, next
);
1764 prepare_arch_switch(next
);
1768 * finish_task_switch - clean up after a task-switch
1769 * @rq: runqueue associated with task-switch
1770 * @prev: the thread we just switched away from.
1772 * finish_task_switch must be called after the context switch, paired
1773 * with a prepare_task_switch call before the context switch.
1774 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1775 * and do any other architecture-specific cleanup actions.
1777 * Note that we may have delayed dropping an mm in context_switch(). If
1778 * so, we finish that here outside of the runqueue lock. (Doing it
1779 * with the lock held can cause deadlocks; see schedule() for
1782 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1783 __releases(rq
->lock
)
1785 struct mm_struct
*mm
= rq
->prev_mm
;
1791 * A task struct has one reference for the use as "current".
1792 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1793 * schedule one last time. The schedule call will never return, and
1794 * the scheduled task must drop that reference.
1795 * The test for TASK_DEAD must occur while the runqueue locks are
1796 * still held, otherwise prev could be scheduled on another cpu, die
1797 * there before we look at prev->state, and then the reference would
1799 * Manfred Spraul <manfred@colorfullife.com>
1801 prev_state
= prev
->state
;
1802 finish_arch_switch(prev
);
1803 finish_lock_switch(rq
, prev
);
1804 fire_sched_in_preempt_notifiers(current
);
1807 if (unlikely(prev_state
== TASK_DEAD
)) {
1809 * Remove function-return probe instances associated with this
1810 * task and put them back on the free list.
1812 kprobe_flush_task(prev
);
1813 put_task_struct(prev
);
1818 * schedule_tail - first thing a freshly forked thread must call.
1819 * @prev: the thread we just switched away from.
1821 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1822 __releases(rq
->lock
)
1824 struct rq
*rq
= this_rq();
1826 finish_task_switch(rq
, prev
);
1827 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1828 /* In this case, finish_task_switch does not reenable preemption */
1831 if (current
->set_child_tid
)
1832 put_user(current
->pid
, current
->set_child_tid
);
1836 * context_switch - switch to the new MM and the new
1837 * thread's register state.
1840 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1841 struct task_struct
*next
)
1843 struct mm_struct
*mm
, *oldmm
;
1845 prepare_task_switch(rq
, prev
, next
);
1847 oldmm
= prev
->active_mm
;
1849 * For paravirt, this is coupled with an exit in switch_to to
1850 * combine the page table reload and the switch backend into
1853 arch_enter_lazy_cpu_mode();
1855 if (unlikely(!mm
)) {
1856 next
->active_mm
= oldmm
;
1857 atomic_inc(&oldmm
->mm_count
);
1858 enter_lazy_tlb(oldmm
, next
);
1860 switch_mm(oldmm
, mm
, next
);
1862 if (unlikely(!prev
->mm
)) {
1863 prev
->active_mm
= NULL
;
1864 rq
->prev_mm
= oldmm
;
1867 * Since the runqueue lock will be released by the next
1868 * task (which is an invalid locking op but in the case
1869 * of the scheduler it's an obvious special-case), so we
1870 * do an early lockdep release here:
1872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1873 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1876 /* Here we just switch the register state and the stack. */
1877 switch_to(prev
, next
, prev
);
1881 * this_rq must be evaluated again because prev may have moved
1882 * CPUs since it called schedule(), thus the 'rq' on its stack
1883 * frame will be invalid.
1885 finish_task_switch(this_rq(), prev
);
1889 * nr_running, nr_uninterruptible and nr_context_switches:
1891 * externally visible scheduler statistics: current number of runnable
1892 * threads, current number of uninterruptible-sleeping threads, total
1893 * number of context switches performed since bootup.
1895 unsigned long nr_running(void)
1897 unsigned long i
, sum
= 0;
1899 for_each_online_cpu(i
)
1900 sum
+= cpu_rq(i
)->nr_running
;
1905 unsigned long nr_uninterruptible(void)
1907 unsigned long i
, sum
= 0;
1909 for_each_possible_cpu(i
)
1910 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1913 * Since we read the counters lockless, it might be slightly
1914 * inaccurate. Do not allow it to go below zero though:
1916 if (unlikely((long)sum
< 0))
1922 unsigned long long nr_context_switches(void)
1925 unsigned long long sum
= 0;
1927 for_each_possible_cpu(i
)
1928 sum
+= cpu_rq(i
)->nr_switches
;
1933 unsigned long nr_iowait(void)
1935 unsigned long i
, sum
= 0;
1937 for_each_possible_cpu(i
)
1938 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1943 unsigned long nr_active(void)
1945 unsigned long i
, running
= 0, uninterruptible
= 0;
1947 for_each_online_cpu(i
) {
1948 running
+= cpu_rq(i
)->nr_running
;
1949 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1952 if (unlikely((long)uninterruptible
< 0))
1953 uninterruptible
= 0;
1955 return running
+ uninterruptible
;
1959 * Update rq->cpu_load[] statistics. This function is usually called every
1960 * scheduler tick (TICK_NSEC).
1962 static void update_cpu_load(struct rq
*this_rq
)
1964 unsigned long this_load
= this_rq
->load
.weight
;
1967 this_rq
->nr_load_updates
++;
1969 /* Update our load: */
1970 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1971 unsigned long old_load
, new_load
;
1973 /* scale is effectively 1 << i now, and >> i divides by scale */
1975 old_load
= this_rq
->cpu_load
[i
];
1976 new_load
= this_load
;
1978 * Round up the averaging division if load is increasing. This
1979 * prevents us from getting stuck on 9 if the load is 10, for
1982 if (new_load
> old_load
)
1983 new_load
+= scale
-1;
1984 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
1991 * double_rq_lock - safely lock two runqueues
1993 * Note this does not disable interrupts like task_rq_lock,
1994 * you need to do so manually before calling.
1996 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1997 __acquires(rq1
->lock
)
1998 __acquires(rq2
->lock
)
2000 BUG_ON(!irqs_disabled());
2002 spin_lock(&rq1
->lock
);
2003 __acquire(rq2
->lock
); /* Fake it out ;) */
2006 spin_lock(&rq1
->lock
);
2007 spin_lock(&rq2
->lock
);
2009 spin_lock(&rq2
->lock
);
2010 spin_lock(&rq1
->lock
);
2013 update_rq_clock(rq1
);
2014 update_rq_clock(rq2
);
2018 * double_rq_unlock - safely unlock two runqueues
2020 * Note this does not restore interrupts like task_rq_unlock,
2021 * you need to do so manually after calling.
2023 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2024 __releases(rq1
->lock
)
2025 __releases(rq2
->lock
)
2027 spin_unlock(&rq1
->lock
);
2029 spin_unlock(&rq2
->lock
);
2031 __release(rq2
->lock
);
2035 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2037 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2038 __releases(this_rq
->lock
)
2039 __acquires(busiest
->lock
)
2040 __acquires(this_rq
->lock
)
2042 if (unlikely(!irqs_disabled())) {
2043 /* printk() doesn't work good under rq->lock */
2044 spin_unlock(&this_rq
->lock
);
2047 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2048 if (busiest
< this_rq
) {
2049 spin_unlock(&this_rq
->lock
);
2050 spin_lock(&busiest
->lock
);
2051 spin_lock(&this_rq
->lock
);
2053 spin_lock(&busiest
->lock
);
2058 * If dest_cpu is allowed for this process, migrate the task to it.
2059 * This is accomplished by forcing the cpu_allowed mask to only
2060 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2061 * the cpu_allowed mask is restored.
2063 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2065 struct migration_req req
;
2066 unsigned long flags
;
2069 rq
= task_rq_lock(p
, &flags
);
2070 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2071 || unlikely(cpu_is_offline(dest_cpu
)))
2074 /* force the process onto the specified CPU */
2075 if (migrate_task(p
, dest_cpu
, &req
)) {
2076 /* Need to wait for migration thread (might exit: take ref). */
2077 struct task_struct
*mt
= rq
->migration_thread
;
2079 get_task_struct(mt
);
2080 task_rq_unlock(rq
, &flags
);
2081 wake_up_process(mt
);
2082 put_task_struct(mt
);
2083 wait_for_completion(&req
.done
);
2088 task_rq_unlock(rq
, &flags
);
2092 * sched_exec - execve() is a valuable balancing opportunity, because at
2093 * this point the task has the smallest effective memory and cache footprint.
2095 void sched_exec(void)
2097 int new_cpu
, this_cpu
= get_cpu();
2098 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2100 if (new_cpu
!= this_cpu
)
2101 sched_migrate_task(current
, new_cpu
);
2105 * pull_task - move a task from a remote runqueue to the local runqueue.
2106 * Both runqueues must be locked.
2108 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2109 struct rq
*this_rq
, int this_cpu
)
2111 deactivate_task(src_rq
, p
, 0);
2112 set_task_cpu(p
, this_cpu
);
2113 activate_task(this_rq
, p
, 0);
2115 * Note that idle threads have a prio of MAX_PRIO, for this test
2116 * to be always true for them.
2118 check_preempt_curr(this_rq
, p
);
2122 * Is this task likely cache-hot:
2125 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
2127 s64 delta
= now
- p
->se
.exec_start
;
2129 return delta
< (long long)sysctl_sched_migration_cost
;
2133 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2136 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2137 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2141 * We do not migrate tasks that are:
2142 * 1) running (obviously), or
2143 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2144 * 3) are cache-hot on their current CPU.
2146 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2150 if (task_running(rq
, p
))
2154 * Aggressive migration if:
2155 * 1) task is cache cold, or
2156 * 2) too many balance attempts have failed.
2159 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2160 #ifdef CONFIG_SCHEDSTATS
2161 if (task_hot(p
, rq
->clock
, sd
))
2162 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2167 if (task_hot(p
, rq
->clock
, sd
))
2172 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2173 unsigned long max_nr_move
, unsigned long max_load_move
,
2174 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2175 int *all_pinned
, unsigned long *load_moved
,
2176 int *this_best_prio
, struct rq_iterator
*iterator
)
2178 int pulled
= 0, pinned
= 0, skip_for_load
;
2179 struct task_struct
*p
;
2180 long rem_load_move
= max_load_move
;
2182 if (max_nr_move
== 0 || max_load_move
== 0)
2188 * Start the load-balancing iterator:
2190 p
= iterator
->start(iterator
->arg
);
2195 * To help distribute high priority tasks accross CPUs we don't
2196 * skip a task if it will be the highest priority task (i.e. smallest
2197 * prio value) on its new queue regardless of its load weight
2199 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2200 SCHED_LOAD_SCALE_FUZZ
;
2201 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2202 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2203 p
= iterator
->next(iterator
->arg
);
2207 pull_task(busiest
, p
, this_rq
, this_cpu
);
2209 rem_load_move
-= p
->se
.load
.weight
;
2212 * We only want to steal up to the prescribed number of tasks
2213 * and the prescribed amount of weighted load.
2215 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2216 if (p
->prio
< *this_best_prio
)
2217 *this_best_prio
= p
->prio
;
2218 p
= iterator
->next(iterator
->arg
);
2223 * Right now, this is the only place pull_task() is called,
2224 * so we can safely collect pull_task() stats here rather than
2225 * inside pull_task().
2227 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2230 *all_pinned
= pinned
;
2231 *load_moved
= max_load_move
- rem_load_move
;
2236 * move_tasks tries to move up to max_load_move weighted load from busiest to
2237 * this_rq, as part of a balancing operation within domain "sd".
2238 * Returns 1 if successful and 0 otherwise.
2240 * Called with both runqueues locked.
2242 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2243 unsigned long max_load_move
,
2244 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2247 const struct sched_class
*class = sched_class_highest
;
2248 unsigned long total_load_moved
= 0;
2249 int this_best_prio
= this_rq
->curr
->prio
;
2253 class->load_balance(this_rq
, this_cpu
, busiest
,
2254 ULONG_MAX
, max_load_move
- total_load_moved
,
2255 sd
, idle
, all_pinned
, &this_best_prio
);
2256 class = class->next
;
2257 } while (class && max_load_move
> total_load_moved
);
2259 return total_load_moved
> 0;
2263 * move_one_task tries to move exactly one task from busiest to this_rq, as
2264 * part of active balancing operations within "domain".
2265 * Returns 1 if successful and 0 otherwise.
2267 * Called with both runqueues locked.
2269 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2270 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2272 const struct sched_class
*class;
2273 int this_best_prio
= MAX_PRIO
;
2275 for (class = sched_class_highest
; class; class = class->next
)
2276 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2277 1, ULONG_MAX
, sd
, idle
, NULL
,
2285 * find_busiest_group finds and returns the busiest CPU group within the
2286 * domain. It calculates and returns the amount of weighted load which
2287 * should be moved to restore balance via the imbalance parameter.
2289 static struct sched_group
*
2290 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2291 unsigned long *imbalance
, enum cpu_idle_type idle
,
2292 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2294 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2295 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2296 unsigned long max_pull
;
2297 unsigned long busiest_load_per_task
, busiest_nr_running
;
2298 unsigned long this_load_per_task
, this_nr_running
;
2300 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2301 int power_savings_balance
= 1;
2302 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2303 unsigned long min_nr_running
= ULONG_MAX
;
2304 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2307 max_load
= this_load
= total_load
= total_pwr
= 0;
2308 busiest_load_per_task
= busiest_nr_running
= 0;
2309 this_load_per_task
= this_nr_running
= 0;
2310 if (idle
== CPU_NOT_IDLE
)
2311 load_idx
= sd
->busy_idx
;
2312 else if (idle
== CPU_NEWLY_IDLE
)
2313 load_idx
= sd
->newidle_idx
;
2315 load_idx
= sd
->idle_idx
;
2318 unsigned long load
, group_capacity
;
2321 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2322 unsigned long sum_nr_running
, sum_weighted_load
;
2324 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2327 balance_cpu
= first_cpu(group
->cpumask
);
2329 /* Tally up the load of all CPUs in the group */
2330 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2332 for_each_cpu_mask(i
, group
->cpumask
) {
2335 if (!cpu_isset(i
, *cpus
))
2340 if (*sd_idle
&& rq
->nr_running
)
2343 /* Bias balancing toward cpus of our domain */
2345 if (idle_cpu(i
) && !first_idle_cpu
) {
2350 load
= target_load(i
, load_idx
);
2352 load
= source_load(i
, load_idx
);
2355 sum_nr_running
+= rq
->nr_running
;
2356 sum_weighted_load
+= weighted_cpuload(i
);
2360 * First idle cpu or the first cpu(busiest) in this sched group
2361 * is eligible for doing load balancing at this and above
2362 * domains. In the newly idle case, we will allow all the cpu's
2363 * to do the newly idle load balance.
2365 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2366 balance_cpu
!= this_cpu
&& balance
) {
2371 total_load
+= avg_load
;
2372 total_pwr
+= group
->__cpu_power
;
2374 /* Adjust by relative CPU power of the group */
2375 avg_load
= sg_div_cpu_power(group
,
2376 avg_load
* SCHED_LOAD_SCALE
);
2378 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2381 this_load
= avg_load
;
2383 this_nr_running
= sum_nr_running
;
2384 this_load_per_task
= sum_weighted_load
;
2385 } else if (avg_load
> max_load
&&
2386 sum_nr_running
> group_capacity
) {
2387 max_load
= avg_load
;
2389 busiest_nr_running
= sum_nr_running
;
2390 busiest_load_per_task
= sum_weighted_load
;
2393 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2395 * Busy processors will not participate in power savings
2398 if (idle
== CPU_NOT_IDLE
||
2399 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2403 * If the local group is idle or completely loaded
2404 * no need to do power savings balance at this domain
2406 if (local_group
&& (this_nr_running
>= group_capacity
||
2408 power_savings_balance
= 0;
2411 * If a group is already running at full capacity or idle,
2412 * don't include that group in power savings calculations
2414 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2419 * Calculate the group which has the least non-idle load.
2420 * This is the group from where we need to pick up the load
2423 if ((sum_nr_running
< min_nr_running
) ||
2424 (sum_nr_running
== min_nr_running
&&
2425 first_cpu(group
->cpumask
) <
2426 first_cpu(group_min
->cpumask
))) {
2428 min_nr_running
= sum_nr_running
;
2429 min_load_per_task
= sum_weighted_load
/
2434 * Calculate the group which is almost near its
2435 * capacity but still has some space to pick up some load
2436 * from other group and save more power
2438 if (sum_nr_running
<= group_capacity
- 1) {
2439 if (sum_nr_running
> leader_nr_running
||
2440 (sum_nr_running
== leader_nr_running
&&
2441 first_cpu(group
->cpumask
) >
2442 first_cpu(group_leader
->cpumask
))) {
2443 group_leader
= group
;
2444 leader_nr_running
= sum_nr_running
;
2449 group
= group
->next
;
2450 } while (group
!= sd
->groups
);
2452 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2455 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2457 if (this_load
>= avg_load
||
2458 100*max_load
<= sd
->imbalance_pct
*this_load
)
2461 busiest_load_per_task
/= busiest_nr_running
;
2463 * We're trying to get all the cpus to the average_load, so we don't
2464 * want to push ourselves above the average load, nor do we wish to
2465 * reduce the max loaded cpu below the average load, as either of these
2466 * actions would just result in more rebalancing later, and ping-pong
2467 * tasks around. Thus we look for the minimum possible imbalance.
2468 * Negative imbalances (*we* are more loaded than anyone else) will
2469 * be counted as no imbalance for these purposes -- we can't fix that
2470 * by pulling tasks to us. Be careful of negative numbers as they'll
2471 * appear as very large values with unsigned longs.
2473 if (max_load
<= busiest_load_per_task
)
2477 * In the presence of smp nice balancing, certain scenarios can have
2478 * max load less than avg load(as we skip the groups at or below
2479 * its cpu_power, while calculating max_load..)
2481 if (max_load
< avg_load
) {
2483 goto small_imbalance
;
2486 /* Don't want to pull so many tasks that a group would go idle */
2487 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2489 /* How much load to actually move to equalise the imbalance */
2490 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2491 (avg_load
- this_load
) * this->__cpu_power
)
2495 * if *imbalance is less than the average load per runnable task
2496 * there is no gaurantee that any tasks will be moved so we'll have
2497 * a think about bumping its value to force at least one task to be
2500 if (*imbalance
< busiest_load_per_task
) {
2501 unsigned long tmp
, pwr_now
, pwr_move
;
2505 pwr_move
= pwr_now
= 0;
2507 if (this_nr_running
) {
2508 this_load_per_task
/= this_nr_running
;
2509 if (busiest_load_per_task
> this_load_per_task
)
2512 this_load_per_task
= SCHED_LOAD_SCALE
;
2514 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2515 busiest_load_per_task
* imbn
) {
2516 *imbalance
= busiest_load_per_task
;
2521 * OK, we don't have enough imbalance to justify moving tasks,
2522 * however we may be able to increase total CPU power used by
2526 pwr_now
+= busiest
->__cpu_power
*
2527 min(busiest_load_per_task
, max_load
);
2528 pwr_now
+= this->__cpu_power
*
2529 min(this_load_per_task
, this_load
);
2530 pwr_now
/= SCHED_LOAD_SCALE
;
2532 /* Amount of load we'd subtract */
2533 tmp
= sg_div_cpu_power(busiest
,
2534 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2536 pwr_move
+= busiest
->__cpu_power
*
2537 min(busiest_load_per_task
, max_load
- tmp
);
2539 /* Amount of load we'd add */
2540 if (max_load
* busiest
->__cpu_power
<
2541 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2542 tmp
= sg_div_cpu_power(this,
2543 max_load
* busiest
->__cpu_power
);
2545 tmp
= sg_div_cpu_power(this,
2546 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2547 pwr_move
+= this->__cpu_power
*
2548 min(this_load_per_task
, this_load
+ tmp
);
2549 pwr_move
/= SCHED_LOAD_SCALE
;
2551 /* Move if we gain throughput */
2552 if (pwr_move
> pwr_now
)
2553 *imbalance
= busiest_load_per_task
;
2559 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2560 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2563 if (this == group_leader
&& group_leader
!= group_min
) {
2564 *imbalance
= min_load_per_task
;
2574 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2577 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2578 unsigned long imbalance
, cpumask_t
*cpus
)
2580 struct rq
*busiest
= NULL
, *rq
;
2581 unsigned long max_load
= 0;
2584 for_each_cpu_mask(i
, group
->cpumask
) {
2587 if (!cpu_isset(i
, *cpus
))
2591 wl
= weighted_cpuload(i
);
2593 if (rq
->nr_running
== 1 && wl
> imbalance
)
2596 if (wl
> max_load
) {
2606 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2607 * so long as it is large enough.
2609 #define MAX_PINNED_INTERVAL 512
2612 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2613 * tasks if there is an imbalance.
2615 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2616 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2619 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2620 struct sched_group
*group
;
2621 unsigned long imbalance
;
2623 cpumask_t cpus
= CPU_MASK_ALL
;
2624 unsigned long flags
;
2627 * When power savings policy is enabled for the parent domain, idle
2628 * sibling can pick up load irrespective of busy siblings. In this case,
2629 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2630 * portraying it as CPU_NOT_IDLE.
2632 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2633 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2636 schedstat_inc(sd
, lb_count
[idle
]);
2639 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2646 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2650 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2652 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2656 BUG_ON(busiest
== this_rq
);
2658 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2661 if (busiest
->nr_running
> 1) {
2663 * Attempt to move tasks. If find_busiest_group has found
2664 * an imbalance but busiest->nr_running <= 1, the group is
2665 * still unbalanced. ld_moved simply stays zero, so it is
2666 * correctly treated as an imbalance.
2668 local_irq_save(flags
);
2669 double_rq_lock(this_rq
, busiest
);
2670 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2671 imbalance
, sd
, idle
, &all_pinned
);
2672 double_rq_unlock(this_rq
, busiest
);
2673 local_irq_restore(flags
);
2676 * some other cpu did the load balance for us.
2678 if (ld_moved
&& this_cpu
!= smp_processor_id())
2679 resched_cpu(this_cpu
);
2681 /* All tasks on this runqueue were pinned by CPU affinity */
2682 if (unlikely(all_pinned
)) {
2683 cpu_clear(cpu_of(busiest
), cpus
);
2684 if (!cpus_empty(cpus
))
2691 schedstat_inc(sd
, lb_failed
[idle
]);
2692 sd
->nr_balance_failed
++;
2694 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2696 spin_lock_irqsave(&busiest
->lock
, flags
);
2698 /* don't kick the migration_thread, if the curr
2699 * task on busiest cpu can't be moved to this_cpu
2701 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2702 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2704 goto out_one_pinned
;
2707 if (!busiest
->active_balance
) {
2708 busiest
->active_balance
= 1;
2709 busiest
->push_cpu
= this_cpu
;
2712 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2714 wake_up_process(busiest
->migration_thread
);
2717 * We've kicked active balancing, reset the failure
2720 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2723 sd
->nr_balance_failed
= 0;
2725 if (likely(!active_balance
)) {
2726 /* We were unbalanced, so reset the balancing interval */
2727 sd
->balance_interval
= sd
->min_interval
;
2730 * If we've begun active balancing, start to back off. This
2731 * case may not be covered by the all_pinned logic if there
2732 * is only 1 task on the busy runqueue (because we don't call
2735 if (sd
->balance_interval
< sd
->max_interval
)
2736 sd
->balance_interval
*= 2;
2739 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2740 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2745 schedstat_inc(sd
, lb_balanced
[idle
]);
2747 sd
->nr_balance_failed
= 0;
2750 /* tune up the balancing interval */
2751 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2752 (sd
->balance_interval
< sd
->max_interval
))
2753 sd
->balance_interval
*= 2;
2755 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2756 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2762 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2763 * tasks if there is an imbalance.
2765 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2766 * this_rq is locked.
2769 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2771 struct sched_group
*group
;
2772 struct rq
*busiest
= NULL
;
2773 unsigned long imbalance
;
2777 cpumask_t cpus
= CPU_MASK_ALL
;
2780 * When power savings policy is enabled for the parent domain, idle
2781 * sibling can pick up load irrespective of busy siblings. In this case,
2782 * let the state of idle sibling percolate up as IDLE, instead of
2783 * portraying it as CPU_NOT_IDLE.
2785 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2786 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2789 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2791 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2792 &sd_idle
, &cpus
, NULL
);
2794 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2798 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2801 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2805 BUG_ON(busiest
== this_rq
);
2807 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2810 if (busiest
->nr_running
> 1) {
2811 /* Attempt to move tasks */
2812 double_lock_balance(this_rq
, busiest
);
2813 /* this_rq->clock is already updated */
2814 update_rq_clock(busiest
);
2815 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2816 imbalance
, sd
, CPU_NEWLY_IDLE
,
2818 spin_unlock(&busiest
->lock
);
2820 if (unlikely(all_pinned
)) {
2821 cpu_clear(cpu_of(busiest
), cpus
);
2822 if (!cpus_empty(cpus
))
2828 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2829 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2830 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2833 sd
->nr_balance_failed
= 0;
2838 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2839 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2840 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2842 sd
->nr_balance_failed
= 0;
2848 * idle_balance is called by schedule() if this_cpu is about to become
2849 * idle. Attempts to pull tasks from other CPUs.
2851 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2853 struct sched_domain
*sd
;
2854 int pulled_task
= -1;
2855 unsigned long next_balance
= jiffies
+ HZ
;
2857 for_each_domain(this_cpu
, sd
) {
2858 unsigned long interval
;
2860 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2863 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2864 /* If we've pulled tasks over stop searching: */
2865 pulled_task
= load_balance_newidle(this_cpu
,
2868 interval
= msecs_to_jiffies(sd
->balance_interval
);
2869 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2870 next_balance
= sd
->last_balance
+ interval
;
2874 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2876 * We are going idle. next_balance may be set based on
2877 * a busy processor. So reset next_balance.
2879 this_rq
->next_balance
= next_balance
;
2884 * active_load_balance is run by migration threads. It pushes running tasks
2885 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2886 * running on each physical CPU where possible, and avoids physical /
2887 * logical imbalances.
2889 * Called with busiest_rq locked.
2891 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2893 int target_cpu
= busiest_rq
->push_cpu
;
2894 struct sched_domain
*sd
;
2895 struct rq
*target_rq
;
2897 /* Is there any task to move? */
2898 if (busiest_rq
->nr_running
<= 1)
2901 target_rq
= cpu_rq(target_cpu
);
2904 * This condition is "impossible", if it occurs
2905 * we need to fix it. Originally reported by
2906 * Bjorn Helgaas on a 128-cpu setup.
2908 BUG_ON(busiest_rq
== target_rq
);
2910 /* move a task from busiest_rq to target_rq */
2911 double_lock_balance(busiest_rq
, target_rq
);
2912 update_rq_clock(busiest_rq
);
2913 update_rq_clock(target_rq
);
2915 /* Search for an sd spanning us and the target CPU. */
2916 for_each_domain(target_cpu
, sd
) {
2917 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2918 cpu_isset(busiest_cpu
, sd
->span
))
2923 schedstat_inc(sd
, alb_count
);
2925 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2927 schedstat_inc(sd
, alb_pushed
);
2929 schedstat_inc(sd
, alb_failed
);
2931 spin_unlock(&target_rq
->lock
);
2936 atomic_t load_balancer
;
2938 } nohz ____cacheline_aligned
= {
2939 .load_balancer
= ATOMIC_INIT(-1),
2940 .cpu_mask
= CPU_MASK_NONE
,
2944 * This routine will try to nominate the ilb (idle load balancing)
2945 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2946 * load balancing on behalf of all those cpus. If all the cpus in the system
2947 * go into this tickless mode, then there will be no ilb owner (as there is
2948 * no need for one) and all the cpus will sleep till the next wakeup event
2951 * For the ilb owner, tick is not stopped. And this tick will be used
2952 * for idle load balancing. ilb owner will still be part of
2955 * While stopping the tick, this cpu will become the ilb owner if there
2956 * is no other owner. And will be the owner till that cpu becomes busy
2957 * or if all cpus in the system stop their ticks at which point
2958 * there is no need for ilb owner.
2960 * When the ilb owner becomes busy, it nominates another owner, during the
2961 * next busy scheduler_tick()
2963 int select_nohz_load_balancer(int stop_tick
)
2965 int cpu
= smp_processor_id();
2968 cpu_set(cpu
, nohz
.cpu_mask
);
2969 cpu_rq(cpu
)->in_nohz_recently
= 1;
2972 * If we are going offline and still the leader, give up!
2974 if (cpu_is_offline(cpu
) &&
2975 atomic_read(&nohz
.load_balancer
) == cpu
) {
2976 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2981 /* time for ilb owner also to sleep */
2982 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2983 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2984 atomic_set(&nohz
.load_balancer
, -1);
2988 if (atomic_read(&nohz
.load_balancer
) == -1) {
2989 /* make me the ilb owner */
2990 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2992 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2995 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2998 cpu_clear(cpu
, nohz
.cpu_mask
);
3000 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3001 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3008 static DEFINE_SPINLOCK(balancing
);
3011 * It checks each scheduling domain to see if it is due to be balanced,
3012 * and initiates a balancing operation if so.
3014 * Balancing parameters are set up in arch_init_sched_domains.
3016 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3019 struct rq
*rq
= cpu_rq(cpu
);
3020 unsigned long interval
;
3021 struct sched_domain
*sd
;
3022 /* Earliest time when we have to do rebalance again */
3023 unsigned long next_balance
= jiffies
+ 60*HZ
;
3024 int update_next_balance
= 0;
3026 for_each_domain(cpu
, sd
) {
3027 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3030 interval
= sd
->balance_interval
;
3031 if (idle
!= CPU_IDLE
)
3032 interval
*= sd
->busy_factor
;
3034 /* scale ms to jiffies */
3035 interval
= msecs_to_jiffies(interval
);
3036 if (unlikely(!interval
))
3038 if (interval
> HZ
*NR_CPUS
/10)
3039 interval
= HZ
*NR_CPUS
/10;
3042 if (sd
->flags
& SD_SERIALIZE
) {
3043 if (!spin_trylock(&balancing
))
3047 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3048 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3050 * We've pulled tasks over so either we're no
3051 * longer idle, or one of our SMT siblings is
3054 idle
= CPU_NOT_IDLE
;
3056 sd
->last_balance
= jiffies
;
3058 if (sd
->flags
& SD_SERIALIZE
)
3059 spin_unlock(&balancing
);
3061 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3062 next_balance
= sd
->last_balance
+ interval
;
3063 update_next_balance
= 1;
3067 * Stop the load balance at this level. There is another
3068 * CPU in our sched group which is doing load balancing more
3076 * next_balance will be updated only when there is a need.
3077 * When the cpu is attached to null domain for ex, it will not be
3080 if (likely(update_next_balance
))
3081 rq
->next_balance
= next_balance
;
3085 * run_rebalance_domains is triggered when needed from the scheduler tick.
3086 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3087 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3089 static void run_rebalance_domains(struct softirq_action
*h
)
3091 int this_cpu
= smp_processor_id();
3092 struct rq
*this_rq
= cpu_rq(this_cpu
);
3093 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3094 CPU_IDLE
: CPU_NOT_IDLE
;
3096 rebalance_domains(this_cpu
, idle
);
3100 * If this cpu is the owner for idle load balancing, then do the
3101 * balancing on behalf of the other idle cpus whose ticks are
3104 if (this_rq
->idle_at_tick
&&
3105 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3106 cpumask_t cpus
= nohz
.cpu_mask
;
3110 cpu_clear(this_cpu
, cpus
);
3111 for_each_cpu_mask(balance_cpu
, cpus
) {
3113 * If this cpu gets work to do, stop the load balancing
3114 * work being done for other cpus. Next load
3115 * balancing owner will pick it up.
3120 rebalance_domains(balance_cpu
, CPU_IDLE
);
3122 rq
= cpu_rq(balance_cpu
);
3123 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3124 this_rq
->next_balance
= rq
->next_balance
;
3131 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3133 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3134 * idle load balancing owner or decide to stop the periodic load balancing,
3135 * if the whole system is idle.
3137 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3141 * If we were in the nohz mode recently and busy at the current
3142 * scheduler tick, then check if we need to nominate new idle
3145 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3146 rq
->in_nohz_recently
= 0;
3148 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3149 cpu_clear(cpu
, nohz
.cpu_mask
);
3150 atomic_set(&nohz
.load_balancer
, -1);
3153 if (atomic_read(&nohz
.load_balancer
) == -1) {
3155 * simple selection for now: Nominate the
3156 * first cpu in the nohz list to be the next
3159 * TBD: Traverse the sched domains and nominate
3160 * the nearest cpu in the nohz.cpu_mask.
3162 int ilb
= first_cpu(nohz
.cpu_mask
);
3170 * If this cpu is idle and doing idle load balancing for all the
3171 * cpus with ticks stopped, is it time for that to stop?
3173 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3174 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3180 * If this cpu is idle and the idle load balancing is done by
3181 * someone else, then no need raise the SCHED_SOFTIRQ
3183 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3184 cpu_isset(cpu
, nohz
.cpu_mask
))
3187 if (time_after_eq(jiffies
, rq
->next_balance
))
3188 raise_softirq(SCHED_SOFTIRQ
);
3191 #else /* CONFIG_SMP */
3194 * on UP we do not need to balance between CPUs:
3196 static inline void idle_balance(int cpu
, struct rq
*rq
)
3200 /* Avoid "used but not defined" warning on UP */
3201 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3202 unsigned long max_nr_move
, unsigned long max_load_move
,
3203 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3204 int *all_pinned
, unsigned long *load_moved
,
3205 int *this_best_prio
, struct rq_iterator
*iterator
)
3214 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3216 EXPORT_PER_CPU_SYMBOL(kstat
);
3219 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3220 * that have not yet been banked in case the task is currently running.
3222 unsigned long long task_sched_runtime(struct task_struct
*p
)
3224 unsigned long flags
;
3228 rq
= task_rq_lock(p
, &flags
);
3229 ns
= p
->se
.sum_exec_runtime
;
3230 if (rq
->curr
== p
) {
3231 update_rq_clock(rq
);
3232 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3233 if ((s64
)delta_exec
> 0)
3236 task_rq_unlock(rq
, &flags
);
3242 * Account user 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 user space since the last update
3247 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3249 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3252 p
->utime
= cputime_add(p
->utime
, cputime
);
3254 /* Add user time to cpustat. */
3255 tmp
= cputime_to_cputime64(cputime
);
3256 if (TASK_NICE(p
) > 0)
3257 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3259 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3263 * Account system cpu time to a process.
3264 * @p: the process that the cpu time gets accounted to
3265 * @hardirq_offset: the offset to subtract from hardirq_count()
3266 * @cputime: the cpu time spent in kernel space since the last update
3268 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3271 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3272 struct rq
*rq
= this_rq();
3275 p
->stime
= cputime_add(p
->stime
, cputime
);
3277 /* Add system time to cpustat. */
3278 tmp
= cputime_to_cputime64(cputime
);
3279 if (hardirq_count() - hardirq_offset
)
3280 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3281 else if (softirq_count())
3282 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3283 else if (p
!= rq
->idle
)
3284 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3285 else if (atomic_read(&rq
->nr_iowait
) > 0)
3286 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3288 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3289 /* Account for system time used */
3290 acct_update_integrals(p
);
3294 * Account for involuntary wait time.
3295 * @p: the process from which the cpu time has been stolen
3296 * @steal: the cpu time spent in involuntary wait
3298 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3300 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3301 cputime64_t tmp
= cputime_to_cputime64(steal
);
3302 struct rq
*rq
= this_rq();
3304 if (p
== rq
->idle
) {
3305 p
->stime
= cputime_add(p
->stime
, steal
);
3306 if (atomic_read(&rq
->nr_iowait
) > 0)
3307 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3309 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3311 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3315 * This function gets called by the timer code, with HZ frequency.
3316 * We call it with interrupts disabled.
3318 * It also gets called by the fork code, when changing the parent's
3321 void scheduler_tick(void)
3323 int cpu
= smp_processor_id();
3324 struct rq
*rq
= cpu_rq(cpu
);
3325 struct task_struct
*curr
= rq
->curr
;
3326 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3328 spin_lock(&rq
->lock
);
3329 __update_rq_clock(rq
);
3331 * Let rq->clock advance by at least TICK_NSEC:
3333 if (unlikely(rq
->clock
< next_tick
))
3334 rq
->clock
= next_tick
;
3335 rq
->tick_timestamp
= rq
->clock
;
3336 update_cpu_load(rq
);
3337 if (curr
!= rq
->idle
) /* FIXME: needed? */
3338 curr
->sched_class
->task_tick(rq
, curr
);
3339 spin_unlock(&rq
->lock
);
3342 rq
->idle_at_tick
= idle_cpu(cpu
);
3343 trigger_load_balance(rq
, cpu
);
3347 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3349 void fastcall
add_preempt_count(int val
)
3354 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3356 preempt_count() += val
;
3358 * Spinlock count overflowing soon?
3360 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3363 EXPORT_SYMBOL(add_preempt_count
);
3365 void fastcall
sub_preempt_count(int val
)
3370 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3373 * Is the spinlock portion underflowing?
3375 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3376 !(preempt_count() & PREEMPT_MASK
)))
3379 preempt_count() -= val
;
3381 EXPORT_SYMBOL(sub_preempt_count
);
3386 * Print scheduling while atomic bug:
3388 static noinline
void __schedule_bug(struct task_struct
*prev
)
3390 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3391 prev
->comm
, preempt_count(), prev
->pid
);
3392 debug_show_held_locks(prev
);
3393 if (irqs_disabled())
3394 print_irqtrace_events(prev
);
3399 * Various schedule()-time debugging checks and statistics:
3401 static inline void schedule_debug(struct task_struct
*prev
)
3404 * Test if we are atomic. Since do_exit() needs to call into
3405 * schedule() atomically, we ignore that path for now.
3406 * Otherwise, whine if we are scheduling when we should not be.
3408 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3409 __schedule_bug(prev
);
3411 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3413 schedstat_inc(this_rq(), sched_count
);
3414 #ifdef CONFIG_SCHEDSTATS
3415 if (unlikely(prev
->lock_depth
>= 0)) {
3416 schedstat_inc(this_rq(), bkl_count
);
3417 schedstat_inc(prev
, sched_info
.bkl_count
);
3423 * Pick up the highest-prio task:
3425 static inline struct task_struct
*
3426 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3428 const struct sched_class
*class;
3429 struct task_struct
*p
;
3432 * Optimization: we know that if all tasks are in
3433 * the fair class we can call that function directly:
3435 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3436 p
= fair_sched_class
.pick_next_task(rq
);
3441 class = sched_class_highest
;
3443 p
= class->pick_next_task(rq
);
3447 * Will never be NULL as the idle class always
3448 * returns a non-NULL p:
3450 class = class->next
;
3455 * schedule() is the main scheduler function.
3457 asmlinkage
void __sched
schedule(void)
3459 struct task_struct
*prev
, *next
;
3466 cpu
= smp_processor_id();
3470 switch_count
= &prev
->nivcsw
;
3472 release_kernel_lock(prev
);
3473 need_resched_nonpreemptible
:
3475 schedule_debug(prev
);
3478 * Do the rq-clock update outside the rq lock:
3480 local_irq_disable();
3481 __update_rq_clock(rq
);
3482 spin_lock(&rq
->lock
);
3483 clear_tsk_need_resched(prev
);
3485 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3486 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3487 unlikely(signal_pending(prev
)))) {
3488 prev
->state
= TASK_RUNNING
;
3490 deactivate_task(rq
, prev
, 1);
3492 switch_count
= &prev
->nvcsw
;
3495 if (unlikely(!rq
->nr_running
))
3496 idle_balance(cpu
, rq
);
3498 prev
->sched_class
->put_prev_task(rq
, prev
);
3499 next
= pick_next_task(rq
, prev
);
3501 sched_info_switch(prev
, next
);
3503 if (likely(prev
!= next
)) {
3508 context_switch(rq
, prev
, next
); /* unlocks the rq */
3510 spin_unlock_irq(&rq
->lock
);
3512 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3513 cpu
= smp_processor_id();
3515 goto need_resched_nonpreemptible
;
3517 preempt_enable_no_resched();
3518 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3521 EXPORT_SYMBOL(schedule
);
3523 #ifdef CONFIG_PREEMPT
3525 * this is the entry point to schedule() from in-kernel preemption
3526 * off of preempt_enable. Kernel preemptions off return from interrupt
3527 * occur there and call schedule directly.
3529 asmlinkage
void __sched
preempt_schedule(void)
3531 struct thread_info
*ti
= current_thread_info();
3532 #ifdef CONFIG_PREEMPT_BKL
3533 struct task_struct
*task
= current
;
3534 int saved_lock_depth
;
3537 * If there is a non-zero preempt_count or interrupts are disabled,
3538 * we do not want to preempt the current task. Just return..
3540 if (likely(ti
->preempt_count
|| irqs_disabled()))
3544 add_preempt_count(PREEMPT_ACTIVE
);
3547 * We keep the big kernel semaphore locked, but we
3548 * clear ->lock_depth so that schedule() doesnt
3549 * auto-release the semaphore:
3551 #ifdef CONFIG_PREEMPT_BKL
3552 saved_lock_depth
= task
->lock_depth
;
3553 task
->lock_depth
= -1;
3556 #ifdef CONFIG_PREEMPT_BKL
3557 task
->lock_depth
= saved_lock_depth
;
3559 sub_preempt_count(PREEMPT_ACTIVE
);
3562 * Check again in case we missed a preemption opportunity
3563 * between schedule and now.
3566 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3568 EXPORT_SYMBOL(preempt_schedule
);
3571 * this is the entry point to schedule() from kernel preemption
3572 * off of irq context.
3573 * Note, that this is called and return with irqs disabled. This will
3574 * protect us against recursive calling from irq.
3576 asmlinkage
void __sched
preempt_schedule_irq(void)
3578 struct thread_info
*ti
= current_thread_info();
3579 #ifdef CONFIG_PREEMPT_BKL
3580 struct task_struct
*task
= current
;
3581 int saved_lock_depth
;
3583 /* Catch callers which need to be fixed */
3584 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3587 add_preempt_count(PREEMPT_ACTIVE
);
3590 * We keep the big kernel semaphore locked, but we
3591 * clear ->lock_depth so that schedule() doesnt
3592 * auto-release the semaphore:
3594 #ifdef CONFIG_PREEMPT_BKL
3595 saved_lock_depth
= task
->lock_depth
;
3596 task
->lock_depth
= -1;
3600 local_irq_disable();
3601 #ifdef CONFIG_PREEMPT_BKL
3602 task
->lock_depth
= saved_lock_depth
;
3604 sub_preempt_count(PREEMPT_ACTIVE
);
3607 * Check again in case we missed a preemption opportunity
3608 * between schedule and now.
3611 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3614 #endif /* CONFIG_PREEMPT */
3616 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3619 return try_to_wake_up(curr
->private, mode
, sync
);
3621 EXPORT_SYMBOL(default_wake_function
);
3624 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3625 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3626 * number) then we wake all the non-exclusive tasks and one exclusive task.
3628 * There are circumstances in which we can try to wake a task which has already
3629 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3630 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3632 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3633 int nr_exclusive
, int sync
, void *key
)
3635 wait_queue_t
*curr
, *next
;
3637 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3638 unsigned flags
= curr
->flags
;
3640 if (curr
->func(curr
, mode
, sync
, key
) &&
3641 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3647 * __wake_up - wake up threads blocked on a waitqueue.
3649 * @mode: which threads
3650 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3651 * @key: is directly passed to the wakeup function
3653 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3654 int nr_exclusive
, void *key
)
3656 unsigned long flags
;
3658 spin_lock_irqsave(&q
->lock
, flags
);
3659 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3660 spin_unlock_irqrestore(&q
->lock
, flags
);
3662 EXPORT_SYMBOL(__wake_up
);
3665 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3667 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3669 __wake_up_common(q
, mode
, 1, 0, NULL
);
3673 * __wake_up_sync - wake up threads blocked on a waitqueue.
3675 * @mode: which threads
3676 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3678 * The sync wakeup differs that the waker knows that it will schedule
3679 * away soon, so while the target thread will be woken up, it will not
3680 * be migrated to another CPU - ie. the two threads are 'synchronized'
3681 * with each other. This can prevent needless bouncing between CPUs.
3683 * On UP it can prevent extra preemption.
3686 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3688 unsigned long flags
;
3694 if (unlikely(!nr_exclusive
))
3697 spin_lock_irqsave(&q
->lock
, flags
);
3698 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3699 spin_unlock_irqrestore(&q
->lock
, flags
);
3701 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3703 void fastcall
complete(struct completion
*x
)
3705 unsigned long flags
;
3707 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3709 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3711 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3713 EXPORT_SYMBOL(complete
);
3715 void fastcall
complete_all(struct completion
*x
)
3717 unsigned long flags
;
3719 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3720 x
->done
+= UINT_MAX
/2;
3721 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3723 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3725 EXPORT_SYMBOL(complete_all
);
3727 static inline long __sched
3728 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3731 DECLARE_WAITQUEUE(wait
, current
);
3733 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3734 __add_wait_queue_tail(&x
->wait
, &wait
);
3736 if (state
== TASK_INTERRUPTIBLE
&&
3737 signal_pending(current
)) {
3738 __remove_wait_queue(&x
->wait
, &wait
);
3739 return -ERESTARTSYS
;
3741 __set_current_state(state
);
3742 spin_unlock_irq(&x
->wait
.lock
);
3743 timeout
= schedule_timeout(timeout
);
3744 spin_lock_irq(&x
->wait
.lock
);
3746 __remove_wait_queue(&x
->wait
, &wait
);
3750 __remove_wait_queue(&x
->wait
, &wait
);
3757 wait_for_common(struct completion
*x
, long timeout
, int state
)
3761 spin_lock_irq(&x
->wait
.lock
);
3762 timeout
= do_wait_for_common(x
, timeout
, state
);
3763 spin_unlock_irq(&x
->wait
.lock
);
3767 void fastcall __sched
wait_for_completion(struct completion
*x
)
3769 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3771 EXPORT_SYMBOL(wait_for_completion
);
3773 unsigned long fastcall __sched
3774 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3776 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3778 EXPORT_SYMBOL(wait_for_completion_timeout
);
3780 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3782 return wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3784 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3786 unsigned long fastcall __sched
3787 wait_for_completion_interruptible_timeout(struct completion
*x
,
3788 unsigned long timeout
)
3790 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3792 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3795 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3797 unsigned long flags
;
3800 init_waitqueue_entry(&wait
, current
);
3802 __set_current_state(state
);
3804 spin_lock_irqsave(&q
->lock
, flags
);
3805 __add_wait_queue(q
, &wait
);
3806 spin_unlock(&q
->lock
);
3807 timeout
= schedule_timeout(timeout
);
3808 spin_lock_irq(&q
->lock
);
3809 __remove_wait_queue(q
, &wait
);
3810 spin_unlock_irqrestore(&q
->lock
, flags
);
3815 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3817 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3819 EXPORT_SYMBOL(interruptible_sleep_on
);
3822 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3824 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3826 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3828 void __sched
sleep_on(wait_queue_head_t
*q
)
3830 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3832 EXPORT_SYMBOL(sleep_on
);
3834 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3836 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3838 EXPORT_SYMBOL(sleep_on_timeout
);
3840 #ifdef CONFIG_RT_MUTEXES
3843 * rt_mutex_setprio - set the current priority of a task
3845 * @prio: prio value (kernel-internal form)
3847 * This function changes the 'effective' priority of a task. It does
3848 * not touch ->normal_prio like __setscheduler().
3850 * Used by the rt_mutex code to implement priority inheritance logic.
3852 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3854 unsigned long flags
;
3855 int oldprio
, on_rq
, running
;
3858 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3860 rq
= task_rq_lock(p
, &flags
);
3861 update_rq_clock(rq
);
3864 on_rq
= p
->se
.on_rq
;
3865 running
= task_running(rq
, p
);
3867 dequeue_task(rq
, p
, 0);
3869 p
->sched_class
->put_prev_task(rq
, p
);
3873 p
->sched_class
= &rt_sched_class
;
3875 p
->sched_class
= &fair_sched_class
;
3881 p
->sched_class
->set_curr_task(rq
);
3882 enqueue_task(rq
, p
, 0);
3884 * Reschedule if we are currently running on this runqueue and
3885 * our priority decreased, or if we are not currently running on
3886 * this runqueue and our priority is higher than the current's
3889 if (p
->prio
> oldprio
)
3890 resched_task(rq
->curr
);
3892 check_preempt_curr(rq
, p
);
3895 task_rq_unlock(rq
, &flags
);
3900 void set_user_nice(struct task_struct
*p
, long nice
)
3902 int old_prio
, delta
, on_rq
;
3903 unsigned long flags
;
3906 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3909 * We have to be careful, if called from sys_setpriority(),
3910 * the task might be in the middle of scheduling on another CPU.
3912 rq
= task_rq_lock(p
, &flags
);
3913 update_rq_clock(rq
);
3915 * The RT priorities are set via sched_setscheduler(), but we still
3916 * allow the 'normal' nice value to be set - but as expected
3917 * it wont have any effect on scheduling until the task is
3918 * SCHED_FIFO/SCHED_RR:
3920 if (task_has_rt_policy(p
)) {
3921 p
->static_prio
= NICE_TO_PRIO(nice
);
3924 on_rq
= p
->se
.on_rq
;
3926 dequeue_task(rq
, p
, 0);
3930 p
->static_prio
= NICE_TO_PRIO(nice
);
3933 p
->prio
= effective_prio(p
);
3934 delta
= p
->prio
- old_prio
;
3937 enqueue_task(rq
, p
, 0);
3940 * If the task increased its priority or is running and
3941 * lowered its priority, then reschedule its CPU:
3943 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3944 resched_task(rq
->curr
);
3947 task_rq_unlock(rq
, &flags
);
3949 EXPORT_SYMBOL(set_user_nice
);
3952 * can_nice - check if a task can reduce its nice value
3956 int can_nice(const struct task_struct
*p
, const int nice
)
3958 /* convert nice value [19,-20] to rlimit style value [1,40] */
3959 int nice_rlim
= 20 - nice
;
3961 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3962 capable(CAP_SYS_NICE
));
3965 #ifdef __ARCH_WANT_SYS_NICE
3968 * sys_nice - change the priority of the current process.
3969 * @increment: priority increment
3971 * sys_setpriority is a more generic, but much slower function that
3972 * does similar things.
3974 asmlinkage
long sys_nice(int increment
)
3979 * Setpriority might change our priority at the same moment.
3980 * We don't have to worry. Conceptually one call occurs first
3981 * and we have a single winner.
3983 if (increment
< -40)
3988 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3994 if (increment
< 0 && !can_nice(current
, nice
))
3997 retval
= security_task_setnice(current
, nice
);
4001 set_user_nice(current
, nice
);
4008 * task_prio - return the priority value of a given task.
4009 * @p: the task in question.
4011 * This is the priority value as seen by users in /proc.
4012 * RT tasks are offset by -200. Normal tasks are centered
4013 * around 0, value goes from -16 to +15.
4015 int task_prio(const struct task_struct
*p
)
4017 return p
->prio
- MAX_RT_PRIO
;
4021 * task_nice - return the nice value of a given task.
4022 * @p: the task in question.
4024 int task_nice(const struct task_struct
*p
)
4026 return TASK_NICE(p
);
4028 EXPORT_SYMBOL_GPL(task_nice
);
4031 * idle_cpu - is a given cpu idle currently?
4032 * @cpu: the processor in question.
4034 int idle_cpu(int cpu
)
4036 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4040 * idle_task - return the idle task for a given cpu.
4041 * @cpu: the processor in question.
4043 struct task_struct
*idle_task(int cpu
)
4045 return cpu_rq(cpu
)->idle
;
4049 * find_process_by_pid - find a process with a matching PID value.
4050 * @pid: the pid in question.
4052 static struct task_struct
*find_process_by_pid(pid_t pid
)
4054 return pid
? find_task_by_pid(pid
) : current
;
4057 /* Actually do priority change: must hold rq lock. */
4059 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4061 BUG_ON(p
->se
.on_rq
);
4064 switch (p
->policy
) {
4068 p
->sched_class
= &fair_sched_class
;
4072 p
->sched_class
= &rt_sched_class
;
4076 p
->rt_priority
= prio
;
4077 p
->normal_prio
= normal_prio(p
);
4078 /* we are holding p->pi_lock already */
4079 p
->prio
= rt_mutex_getprio(p
);
4084 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4085 * @p: the task in question.
4086 * @policy: new policy.
4087 * @param: structure containing the new RT priority.
4089 * NOTE that the task may be already dead.
4091 int sched_setscheduler(struct task_struct
*p
, int policy
,
4092 struct sched_param
*param
)
4094 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4095 unsigned long flags
;
4098 /* may grab non-irq protected spin_locks */
4099 BUG_ON(in_interrupt());
4101 /* double check policy once rq lock held */
4103 policy
= oldpolicy
= p
->policy
;
4104 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4105 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4106 policy
!= SCHED_IDLE
)
4109 * Valid priorities for SCHED_FIFO and SCHED_RR are
4110 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4111 * SCHED_BATCH and SCHED_IDLE is 0.
4113 if (param
->sched_priority
< 0 ||
4114 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4115 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4117 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4121 * Allow unprivileged RT tasks to decrease priority:
4123 if (!capable(CAP_SYS_NICE
)) {
4124 if (rt_policy(policy
)) {
4125 unsigned long rlim_rtprio
;
4127 if (!lock_task_sighand(p
, &flags
))
4129 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4130 unlock_task_sighand(p
, &flags
);
4132 /* can't set/change the rt policy */
4133 if (policy
!= p
->policy
&& !rlim_rtprio
)
4136 /* can't increase priority */
4137 if (param
->sched_priority
> p
->rt_priority
&&
4138 param
->sched_priority
> rlim_rtprio
)
4142 * Like positive nice levels, dont allow tasks to
4143 * move out of SCHED_IDLE either:
4145 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4148 /* can't change other user's priorities */
4149 if ((current
->euid
!= p
->euid
) &&
4150 (current
->euid
!= p
->uid
))
4154 retval
= security_task_setscheduler(p
, policy
, param
);
4158 * make sure no PI-waiters arrive (or leave) while we are
4159 * changing the priority of the task:
4161 spin_lock_irqsave(&p
->pi_lock
, flags
);
4163 * To be able to change p->policy safely, the apropriate
4164 * runqueue lock must be held.
4166 rq
= __task_rq_lock(p
);
4167 /* recheck policy now with rq lock held */
4168 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4169 policy
= oldpolicy
= -1;
4170 __task_rq_unlock(rq
);
4171 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4174 update_rq_clock(rq
);
4175 on_rq
= p
->se
.on_rq
;
4176 running
= task_running(rq
, p
);
4178 deactivate_task(rq
, p
, 0);
4180 p
->sched_class
->put_prev_task(rq
, p
);
4184 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4188 p
->sched_class
->set_curr_task(rq
);
4189 activate_task(rq
, p
, 0);
4191 * Reschedule if we are currently running on this runqueue and
4192 * our priority decreased, or if we are not currently running on
4193 * this runqueue and our priority is higher than the current's
4196 if (p
->prio
> oldprio
)
4197 resched_task(rq
->curr
);
4199 check_preempt_curr(rq
, p
);
4202 __task_rq_unlock(rq
);
4203 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4205 rt_mutex_adjust_pi(p
);
4209 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4212 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4214 struct sched_param lparam
;
4215 struct task_struct
*p
;
4218 if (!param
|| pid
< 0)
4220 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4225 p
= find_process_by_pid(pid
);
4227 retval
= sched_setscheduler(p
, policy
, &lparam
);
4234 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4235 * @pid: the pid in question.
4236 * @policy: new policy.
4237 * @param: structure containing the new RT priority.
4239 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4240 struct sched_param __user
*param
)
4242 /* negative values for policy are not valid */
4246 return do_sched_setscheduler(pid
, policy
, param
);
4250 * sys_sched_setparam - set/change the RT priority of a thread
4251 * @pid: the pid in question.
4252 * @param: structure containing the new RT priority.
4254 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4256 return do_sched_setscheduler(pid
, -1, param
);
4260 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4261 * @pid: the pid in question.
4263 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4265 struct task_struct
*p
;
4272 read_lock(&tasklist_lock
);
4273 p
= find_process_by_pid(pid
);
4275 retval
= security_task_getscheduler(p
);
4279 read_unlock(&tasklist_lock
);
4284 * sys_sched_getscheduler - get the RT priority of a thread
4285 * @pid: the pid in question.
4286 * @param: structure containing the RT priority.
4288 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4290 struct sched_param lp
;
4291 struct task_struct
*p
;
4294 if (!param
|| pid
< 0)
4297 read_lock(&tasklist_lock
);
4298 p
= find_process_by_pid(pid
);
4303 retval
= security_task_getscheduler(p
);
4307 lp
.sched_priority
= p
->rt_priority
;
4308 read_unlock(&tasklist_lock
);
4311 * This one might sleep, we cannot do it with a spinlock held ...
4313 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4318 read_unlock(&tasklist_lock
);
4322 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4324 cpumask_t cpus_allowed
;
4325 struct task_struct
*p
;
4328 mutex_lock(&sched_hotcpu_mutex
);
4329 read_lock(&tasklist_lock
);
4331 p
= find_process_by_pid(pid
);
4333 read_unlock(&tasklist_lock
);
4334 mutex_unlock(&sched_hotcpu_mutex
);
4339 * It is not safe to call set_cpus_allowed with the
4340 * tasklist_lock held. We will bump the task_struct's
4341 * usage count and then drop tasklist_lock.
4344 read_unlock(&tasklist_lock
);
4347 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4348 !capable(CAP_SYS_NICE
))
4351 retval
= security_task_setscheduler(p
, 0, NULL
);
4355 cpus_allowed
= cpuset_cpus_allowed(p
);
4356 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4357 retval
= set_cpus_allowed(p
, new_mask
);
4361 mutex_unlock(&sched_hotcpu_mutex
);
4365 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4366 cpumask_t
*new_mask
)
4368 if (len
< sizeof(cpumask_t
)) {
4369 memset(new_mask
, 0, sizeof(cpumask_t
));
4370 } else if (len
> sizeof(cpumask_t
)) {
4371 len
= sizeof(cpumask_t
);
4373 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4377 * sys_sched_setaffinity - set the cpu affinity of a process
4378 * @pid: pid of the process
4379 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4380 * @user_mask_ptr: user-space pointer to the new cpu mask
4382 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4383 unsigned long __user
*user_mask_ptr
)
4388 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4392 return sched_setaffinity(pid
, new_mask
);
4396 * Represents all cpu's present in the system
4397 * In systems capable of hotplug, this map could dynamically grow
4398 * as new cpu's are detected in the system via any platform specific
4399 * method, such as ACPI for e.g.
4402 cpumask_t cpu_present_map __read_mostly
;
4403 EXPORT_SYMBOL(cpu_present_map
);
4406 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4407 EXPORT_SYMBOL(cpu_online_map
);
4409 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4410 EXPORT_SYMBOL(cpu_possible_map
);
4413 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4415 struct task_struct
*p
;
4418 mutex_lock(&sched_hotcpu_mutex
);
4419 read_lock(&tasklist_lock
);
4422 p
= find_process_by_pid(pid
);
4426 retval
= security_task_getscheduler(p
);
4430 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4433 read_unlock(&tasklist_lock
);
4434 mutex_unlock(&sched_hotcpu_mutex
);
4440 * sys_sched_getaffinity - get the cpu affinity of a process
4441 * @pid: pid of the process
4442 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4443 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4445 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4446 unsigned long __user
*user_mask_ptr
)
4451 if (len
< sizeof(cpumask_t
))
4454 ret
= sched_getaffinity(pid
, &mask
);
4458 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4461 return sizeof(cpumask_t
);
4465 * sys_sched_yield - yield the current processor to other threads.
4467 * This function yields the current CPU to other tasks. If there are no
4468 * other threads running on this CPU then this function will return.
4470 asmlinkage
long sys_sched_yield(void)
4472 struct rq
*rq
= this_rq_lock();
4474 schedstat_inc(rq
, yld_count
);
4475 current
->sched_class
->yield_task(rq
);
4478 * Since we are going to call schedule() anyway, there's
4479 * no need to preempt or enable interrupts:
4481 __release(rq
->lock
);
4482 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4483 _raw_spin_unlock(&rq
->lock
);
4484 preempt_enable_no_resched();
4491 static void __cond_resched(void)
4493 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4494 __might_sleep(__FILE__
, __LINE__
);
4497 * The BKS might be reacquired before we have dropped
4498 * PREEMPT_ACTIVE, which could trigger a second
4499 * cond_resched() call.
4502 add_preempt_count(PREEMPT_ACTIVE
);
4504 sub_preempt_count(PREEMPT_ACTIVE
);
4505 } while (need_resched());
4508 int __sched
cond_resched(void)
4510 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4511 system_state
== SYSTEM_RUNNING
) {
4517 EXPORT_SYMBOL(cond_resched
);
4520 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4521 * call schedule, and on return reacquire the lock.
4523 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4524 * operations here to prevent schedule() from being called twice (once via
4525 * spin_unlock(), once by hand).
4527 int cond_resched_lock(spinlock_t
*lock
)
4531 if (need_lockbreak(lock
)) {
4537 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4538 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4539 _raw_spin_unlock(lock
);
4540 preempt_enable_no_resched();
4547 EXPORT_SYMBOL(cond_resched_lock
);
4549 int __sched
cond_resched_softirq(void)
4551 BUG_ON(!in_softirq());
4553 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4561 EXPORT_SYMBOL(cond_resched_softirq
);
4564 * yield - yield the current processor to other threads.
4566 * This is a shortcut for kernel-space yielding - it marks the
4567 * thread runnable and calls sys_sched_yield().
4569 void __sched
yield(void)
4571 set_current_state(TASK_RUNNING
);
4574 EXPORT_SYMBOL(yield
);
4577 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4578 * that process accounting knows that this is a task in IO wait state.
4580 * But don't do that if it is a deliberate, throttling IO wait (this task
4581 * has set its backing_dev_info: the queue against which it should throttle)
4583 void __sched
io_schedule(void)
4585 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4587 delayacct_blkio_start();
4588 atomic_inc(&rq
->nr_iowait
);
4590 atomic_dec(&rq
->nr_iowait
);
4591 delayacct_blkio_end();
4593 EXPORT_SYMBOL(io_schedule
);
4595 long __sched
io_schedule_timeout(long timeout
)
4597 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4600 delayacct_blkio_start();
4601 atomic_inc(&rq
->nr_iowait
);
4602 ret
= schedule_timeout(timeout
);
4603 atomic_dec(&rq
->nr_iowait
);
4604 delayacct_blkio_end();
4609 * sys_sched_get_priority_max - return maximum RT priority.
4610 * @policy: scheduling class.
4612 * this syscall returns the maximum rt_priority that can be used
4613 * by a given scheduling class.
4615 asmlinkage
long sys_sched_get_priority_max(int policy
)
4622 ret
= MAX_USER_RT_PRIO
-1;
4634 * sys_sched_get_priority_min - return minimum RT priority.
4635 * @policy: scheduling class.
4637 * this syscall returns the minimum rt_priority that can be used
4638 * by a given scheduling class.
4640 asmlinkage
long sys_sched_get_priority_min(int policy
)
4658 * sys_sched_rr_get_interval - return the default timeslice of a process.
4659 * @pid: pid of the process.
4660 * @interval: userspace pointer to the timeslice value.
4662 * this syscall writes the default timeslice value of a given process
4663 * into the user-space timespec buffer. A value of '0' means infinity.
4666 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4668 struct task_struct
*p
;
4669 unsigned int time_slice
;
4677 read_lock(&tasklist_lock
);
4678 p
= find_process_by_pid(pid
);
4682 retval
= security_task_getscheduler(p
);
4686 if (p
->policy
== SCHED_FIFO
)
4688 else if (p
->policy
== SCHED_RR
)
4689 time_slice
= DEF_TIMESLICE
;
4691 struct sched_entity
*se
= &p
->se
;
4692 unsigned long flags
;
4695 rq
= task_rq_lock(p
, &flags
);
4696 time_slice
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
4697 task_rq_unlock(rq
, &flags
);
4699 read_unlock(&tasklist_lock
);
4700 jiffies_to_timespec(time_slice
, &t
);
4701 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4705 read_unlock(&tasklist_lock
);
4709 static const char stat_nam
[] = "RSDTtZX";
4711 static void show_task(struct task_struct
*p
)
4713 unsigned long free
= 0;
4716 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4717 printk("%-13.13s %c", p
->comm
,
4718 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4719 #if BITS_PER_LONG == 32
4720 if (state
== TASK_RUNNING
)
4721 printk(" running ");
4723 printk(" %08lx ", thread_saved_pc(p
));
4725 if (state
== TASK_RUNNING
)
4726 printk(" running task ");
4728 printk(" %016lx ", thread_saved_pc(p
));
4730 #ifdef CONFIG_DEBUG_STACK_USAGE
4732 unsigned long *n
= end_of_stack(p
);
4735 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4738 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4740 if (state
!= TASK_RUNNING
)
4741 show_stack(p
, NULL
);
4744 void show_state_filter(unsigned long state_filter
)
4746 struct task_struct
*g
, *p
;
4748 #if BITS_PER_LONG == 32
4750 " task PC stack pid father\n");
4753 " task PC stack pid father\n");
4755 read_lock(&tasklist_lock
);
4756 do_each_thread(g
, p
) {
4758 * reset the NMI-timeout, listing all files on a slow
4759 * console might take alot of time:
4761 touch_nmi_watchdog();
4762 if (!state_filter
|| (p
->state
& state_filter
))
4764 } while_each_thread(g
, p
);
4766 touch_all_softlockup_watchdogs();
4768 #ifdef CONFIG_SCHED_DEBUG
4769 sysrq_sched_debug_show();
4771 read_unlock(&tasklist_lock
);
4773 * Only show locks if all tasks are dumped:
4775 if (state_filter
== -1)
4776 debug_show_all_locks();
4779 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4781 idle
->sched_class
= &idle_sched_class
;
4785 * init_idle - set up an idle thread for a given CPU
4786 * @idle: task in question
4787 * @cpu: cpu the idle task belongs to
4789 * NOTE: this function does not set the idle thread's NEED_RESCHED
4790 * flag, to make booting more robust.
4792 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4794 struct rq
*rq
= cpu_rq(cpu
);
4795 unsigned long flags
;
4798 idle
->se
.exec_start
= sched_clock();
4800 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4801 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4802 __set_task_cpu(idle
, cpu
);
4804 spin_lock_irqsave(&rq
->lock
, flags
);
4805 rq
->curr
= rq
->idle
= idle
;
4806 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4809 spin_unlock_irqrestore(&rq
->lock
, flags
);
4811 /* Set the preempt count _outside_ the spinlocks! */
4812 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4813 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4815 task_thread_info(idle
)->preempt_count
= 0;
4818 * The idle tasks have their own, simple scheduling class:
4820 idle
->sched_class
= &idle_sched_class
;
4824 * In a system that switches off the HZ timer nohz_cpu_mask
4825 * indicates which cpus entered this state. This is used
4826 * in the rcu update to wait only for active cpus. For system
4827 * which do not switch off the HZ timer nohz_cpu_mask should
4828 * always be CPU_MASK_NONE.
4830 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4834 * This is how migration works:
4836 * 1) we queue a struct migration_req structure in the source CPU's
4837 * runqueue and wake up that CPU's migration thread.
4838 * 2) we down() the locked semaphore => thread blocks.
4839 * 3) migration thread wakes up (implicitly it forces the migrated
4840 * thread off the CPU)
4841 * 4) it gets the migration request and checks whether the migrated
4842 * task is still in the wrong runqueue.
4843 * 5) if it's in the wrong runqueue then the migration thread removes
4844 * it and puts it into the right queue.
4845 * 6) migration thread up()s the semaphore.
4846 * 7) we wake up and the migration is done.
4850 * Change a given task's CPU affinity. Migrate the thread to a
4851 * proper CPU and schedule it away if the CPU it's executing on
4852 * is removed from the allowed bitmask.
4854 * NOTE: the caller must have a valid reference to the task, the
4855 * task must not exit() & deallocate itself prematurely. The
4856 * call is not atomic; no spinlocks may be held.
4858 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4860 struct migration_req req
;
4861 unsigned long flags
;
4865 rq
= task_rq_lock(p
, &flags
);
4866 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4871 p
->cpus_allowed
= new_mask
;
4872 /* Can the task run on the task's current CPU? If so, we're done */
4873 if (cpu_isset(task_cpu(p
), new_mask
))
4876 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4877 /* Need help from migration thread: drop lock and wait. */
4878 task_rq_unlock(rq
, &flags
);
4879 wake_up_process(rq
->migration_thread
);
4880 wait_for_completion(&req
.done
);
4881 tlb_migrate_finish(p
->mm
);
4885 task_rq_unlock(rq
, &flags
);
4889 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4892 * Move (not current) task off this cpu, onto dest cpu. We're doing
4893 * this because either it can't run here any more (set_cpus_allowed()
4894 * away from this CPU, or CPU going down), or because we're
4895 * attempting to rebalance this task on exec (sched_exec).
4897 * So we race with normal scheduler movements, but that's OK, as long
4898 * as the task is no longer on this CPU.
4900 * Returns non-zero if task was successfully migrated.
4902 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4904 struct rq
*rq_dest
, *rq_src
;
4907 if (unlikely(cpu_is_offline(dest_cpu
)))
4910 rq_src
= cpu_rq(src_cpu
);
4911 rq_dest
= cpu_rq(dest_cpu
);
4913 double_rq_lock(rq_src
, rq_dest
);
4914 /* Already moved. */
4915 if (task_cpu(p
) != src_cpu
)
4917 /* Affinity changed (again). */
4918 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4921 on_rq
= p
->se
.on_rq
;
4923 deactivate_task(rq_src
, p
, 0);
4925 set_task_cpu(p
, dest_cpu
);
4927 activate_task(rq_dest
, p
, 0);
4928 check_preempt_curr(rq_dest
, p
);
4932 double_rq_unlock(rq_src
, rq_dest
);
4937 * migration_thread - this is a highprio system thread that performs
4938 * thread migration by bumping thread off CPU then 'pushing' onto
4941 static int migration_thread(void *data
)
4943 int cpu
= (long)data
;
4947 BUG_ON(rq
->migration_thread
!= current
);
4949 set_current_state(TASK_INTERRUPTIBLE
);
4950 while (!kthread_should_stop()) {
4951 struct migration_req
*req
;
4952 struct list_head
*head
;
4954 spin_lock_irq(&rq
->lock
);
4956 if (cpu_is_offline(cpu
)) {
4957 spin_unlock_irq(&rq
->lock
);
4961 if (rq
->active_balance
) {
4962 active_load_balance(rq
, cpu
);
4963 rq
->active_balance
= 0;
4966 head
= &rq
->migration_queue
;
4968 if (list_empty(head
)) {
4969 spin_unlock_irq(&rq
->lock
);
4971 set_current_state(TASK_INTERRUPTIBLE
);
4974 req
= list_entry(head
->next
, struct migration_req
, list
);
4975 list_del_init(head
->next
);
4977 spin_unlock(&rq
->lock
);
4978 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4981 complete(&req
->done
);
4983 __set_current_state(TASK_RUNNING
);
4987 /* Wait for kthread_stop */
4988 set_current_state(TASK_INTERRUPTIBLE
);
4989 while (!kthread_should_stop()) {
4991 set_current_state(TASK_INTERRUPTIBLE
);
4993 __set_current_state(TASK_RUNNING
);
4997 #ifdef CONFIG_HOTPLUG_CPU
4999 * Figure out where task on dead CPU should go, use force if neccessary.
5000 * NOTE: interrupts should be disabled by the caller
5002 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5004 unsigned long flags
;
5011 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5012 cpus_and(mask
, mask
, p
->cpus_allowed
);
5013 dest_cpu
= any_online_cpu(mask
);
5015 /* On any allowed CPU? */
5016 if (dest_cpu
== NR_CPUS
)
5017 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5019 /* No more Mr. Nice Guy. */
5020 if (dest_cpu
== NR_CPUS
) {
5021 rq
= task_rq_lock(p
, &flags
);
5022 cpus_setall(p
->cpus_allowed
);
5023 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5024 task_rq_unlock(rq
, &flags
);
5027 * Don't tell them about moving exiting tasks or
5028 * kernel threads (both mm NULL), since they never
5031 if (p
->mm
&& printk_ratelimit())
5032 printk(KERN_INFO
"process %d (%s) no "
5033 "longer affine to cpu%d\n",
5034 p
->pid
, p
->comm
, dead_cpu
);
5036 } while (!__migrate_task(p
, dead_cpu
, dest_cpu
));
5040 * While a dead CPU has no uninterruptible tasks queued at this point,
5041 * it might still have a nonzero ->nr_uninterruptible counter, because
5042 * for performance reasons the counter is not stricly tracking tasks to
5043 * their home CPUs. So we just add the counter to another CPU's counter,
5044 * to keep the global sum constant after CPU-down:
5046 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5048 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5049 unsigned long flags
;
5051 local_irq_save(flags
);
5052 double_rq_lock(rq_src
, rq_dest
);
5053 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5054 rq_src
->nr_uninterruptible
= 0;
5055 double_rq_unlock(rq_src
, rq_dest
);
5056 local_irq_restore(flags
);
5059 /* Run through task list and migrate tasks from the dead cpu. */
5060 static void migrate_live_tasks(int src_cpu
)
5062 struct task_struct
*p
, *t
;
5064 write_lock_irq(&tasklist_lock
);
5066 do_each_thread(t
, p
) {
5070 if (task_cpu(p
) == src_cpu
)
5071 move_task_off_dead_cpu(src_cpu
, p
);
5072 } while_each_thread(t
, p
);
5074 write_unlock_irq(&tasklist_lock
);
5078 * activate_idle_task - move idle task to the _front_ of runqueue.
5080 static void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
5082 update_rq_clock(rq
);
5084 if (p
->state
== TASK_UNINTERRUPTIBLE
)
5085 rq
->nr_uninterruptible
--;
5087 enqueue_task(rq
, p
, 0);
5088 inc_nr_running(p
, rq
);
5092 * Schedules idle task to be the next runnable task on current CPU.
5093 * It does so by boosting its priority to highest possible and adding it to
5094 * the _front_ of the runqueue. Used by CPU offline code.
5096 void sched_idle_next(void)
5098 int this_cpu
= smp_processor_id();
5099 struct rq
*rq
= cpu_rq(this_cpu
);
5100 struct task_struct
*p
= rq
->idle
;
5101 unsigned long flags
;
5103 /* cpu has to be offline */
5104 BUG_ON(cpu_online(this_cpu
));
5107 * Strictly not necessary since rest of the CPUs are stopped by now
5108 * and interrupts disabled on the current cpu.
5110 spin_lock_irqsave(&rq
->lock
, flags
);
5112 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5114 /* Add idle task to the _front_ of its priority queue: */
5115 activate_idle_task(p
, rq
);
5117 spin_unlock_irqrestore(&rq
->lock
, flags
);
5121 * Ensures that the idle task is using init_mm right before its cpu goes
5124 void idle_task_exit(void)
5126 struct mm_struct
*mm
= current
->active_mm
;
5128 BUG_ON(cpu_online(smp_processor_id()));
5131 switch_mm(mm
, &init_mm
, current
);
5135 /* called under rq->lock with disabled interrupts */
5136 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5138 struct rq
*rq
= cpu_rq(dead_cpu
);
5140 /* Must be exiting, otherwise would be on tasklist. */
5141 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5143 /* Cannot have done final schedule yet: would have vanished. */
5144 BUG_ON(p
->state
== TASK_DEAD
);
5149 * Drop lock around migration; if someone else moves it,
5150 * that's OK. No task can be added to this CPU, so iteration is
5152 * NOTE: interrupts should be left disabled --dev@
5154 spin_unlock(&rq
->lock
);
5155 move_task_off_dead_cpu(dead_cpu
, p
);
5156 spin_lock(&rq
->lock
);
5161 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5162 static void migrate_dead_tasks(unsigned int dead_cpu
)
5164 struct rq
*rq
= cpu_rq(dead_cpu
);
5165 struct task_struct
*next
;
5168 if (!rq
->nr_running
)
5170 update_rq_clock(rq
);
5171 next
= pick_next_task(rq
, rq
->curr
);
5174 migrate_dead(dead_cpu
, next
);
5178 #endif /* CONFIG_HOTPLUG_CPU */
5180 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5182 static struct ctl_table sd_ctl_dir
[] = {
5184 .procname
= "sched_domain",
5190 static struct ctl_table sd_ctl_root
[] = {
5192 .ctl_name
= CTL_KERN
,
5193 .procname
= "kernel",
5195 .child
= sd_ctl_dir
,
5200 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5202 struct ctl_table
*entry
=
5203 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5206 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5212 set_table_entry(struct ctl_table
*entry
,
5213 const char *procname
, void *data
, int maxlen
,
5214 mode_t mode
, proc_handler
*proc_handler
)
5216 entry
->procname
= procname
;
5218 entry
->maxlen
= maxlen
;
5220 entry
->proc_handler
= proc_handler
;
5223 static struct ctl_table
*
5224 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5226 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5228 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5229 sizeof(long), 0644, proc_doulongvec_minmax
);
5230 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5231 sizeof(long), 0644, proc_doulongvec_minmax
);
5232 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5233 sizeof(int), 0644, proc_dointvec_minmax
);
5234 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5235 sizeof(int), 0644, proc_dointvec_minmax
);
5236 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5237 sizeof(int), 0644, proc_dointvec_minmax
);
5238 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5239 sizeof(int), 0644, proc_dointvec_minmax
);
5240 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5241 sizeof(int), 0644, proc_dointvec_minmax
);
5242 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5243 sizeof(int), 0644, proc_dointvec_minmax
);
5244 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5245 sizeof(int), 0644, proc_dointvec_minmax
);
5246 set_table_entry(&table
[9], "cache_nice_tries",
5247 &sd
->cache_nice_tries
,
5248 sizeof(int), 0644, proc_dointvec_minmax
);
5249 set_table_entry(&table
[10], "flags", &sd
->flags
,
5250 sizeof(int), 0644, proc_dointvec_minmax
);
5255 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5257 struct ctl_table
*entry
, *table
;
5258 struct sched_domain
*sd
;
5259 int domain_num
= 0, i
;
5262 for_each_domain(cpu
, sd
)
5264 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5267 for_each_domain(cpu
, sd
) {
5268 snprintf(buf
, 32, "domain%d", i
);
5269 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5271 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5278 static struct ctl_table_header
*sd_sysctl_header
;
5279 static void init_sched_domain_sysctl(void)
5281 int i
, cpu_num
= num_online_cpus();
5282 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5285 sd_ctl_dir
[0].child
= entry
;
5287 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5288 snprintf(buf
, 32, "cpu%d", i
);
5289 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5291 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5293 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5296 static void init_sched_domain_sysctl(void)
5302 * migration_call - callback that gets triggered when a CPU is added.
5303 * Here we can start up the necessary migration thread for the new CPU.
5305 static int __cpuinit
5306 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5308 struct task_struct
*p
;
5309 int cpu
= (long)hcpu
;
5310 unsigned long flags
;
5314 case CPU_LOCK_ACQUIRE
:
5315 mutex_lock(&sched_hotcpu_mutex
);
5318 case CPU_UP_PREPARE
:
5319 case CPU_UP_PREPARE_FROZEN
:
5320 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5323 kthread_bind(p
, cpu
);
5324 /* Must be high prio: stop_machine expects to yield to it. */
5325 rq
= task_rq_lock(p
, &flags
);
5326 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5327 task_rq_unlock(rq
, &flags
);
5328 cpu_rq(cpu
)->migration_thread
= p
;
5332 case CPU_ONLINE_FROZEN
:
5333 /* Strictly unneccessary, as first user will wake it. */
5334 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5337 #ifdef CONFIG_HOTPLUG_CPU
5338 case CPU_UP_CANCELED
:
5339 case CPU_UP_CANCELED_FROZEN
:
5340 if (!cpu_rq(cpu
)->migration_thread
)
5342 /* Unbind it from offline cpu so it can run. Fall thru. */
5343 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5344 any_online_cpu(cpu_online_map
));
5345 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5346 cpu_rq(cpu
)->migration_thread
= NULL
;
5350 case CPU_DEAD_FROZEN
:
5351 migrate_live_tasks(cpu
);
5353 kthread_stop(rq
->migration_thread
);
5354 rq
->migration_thread
= NULL
;
5355 /* Idle task back to normal (off runqueue, low prio) */
5356 rq
= task_rq_lock(rq
->idle
, &flags
);
5357 update_rq_clock(rq
);
5358 deactivate_task(rq
, rq
->idle
, 0);
5359 rq
->idle
->static_prio
= MAX_PRIO
;
5360 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5361 rq
->idle
->sched_class
= &idle_sched_class
;
5362 migrate_dead_tasks(cpu
);
5363 task_rq_unlock(rq
, &flags
);
5364 migrate_nr_uninterruptible(rq
);
5365 BUG_ON(rq
->nr_running
!= 0);
5367 /* No need to migrate the tasks: it was best-effort if
5368 * they didn't take sched_hotcpu_mutex. Just wake up
5369 * the requestors. */
5370 spin_lock_irq(&rq
->lock
);
5371 while (!list_empty(&rq
->migration_queue
)) {
5372 struct migration_req
*req
;
5374 req
= list_entry(rq
->migration_queue
.next
,
5375 struct migration_req
, list
);
5376 list_del_init(&req
->list
);
5377 complete(&req
->done
);
5379 spin_unlock_irq(&rq
->lock
);
5382 case CPU_LOCK_RELEASE
:
5383 mutex_unlock(&sched_hotcpu_mutex
);
5389 /* Register at highest priority so that task migration (migrate_all_tasks)
5390 * happens before everything else.
5392 static struct notifier_block __cpuinitdata migration_notifier
= {
5393 .notifier_call
= migration_call
,
5397 int __init
migration_init(void)
5399 void *cpu
= (void *)(long)smp_processor_id();
5402 /* Start one for the boot CPU: */
5403 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5404 BUG_ON(err
== NOTIFY_BAD
);
5405 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5406 register_cpu_notifier(&migration_notifier
);
5414 /* Number of possible processor ids */
5415 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5416 EXPORT_SYMBOL(nr_cpu_ids
);
5418 #ifdef CONFIG_SCHED_DEBUG
5419 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5424 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5428 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5433 struct sched_group
*group
= sd
->groups
;
5434 cpumask_t groupmask
;
5436 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5437 cpus_clear(groupmask
);
5440 for (i
= 0; i
< level
+ 1; i
++)
5442 printk("domain %d: ", level
);
5444 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5445 printk("does not load-balance\n");
5447 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5452 printk("span %s\n", str
);
5454 if (!cpu_isset(cpu
, sd
->span
))
5455 printk(KERN_ERR
"ERROR: domain->span does not contain "
5457 if (!cpu_isset(cpu
, group
->cpumask
))
5458 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5462 for (i
= 0; i
< level
+ 2; i
++)
5468 printk(KERN_ERR
"ERROR: group is NULL\n");
5472 if (!group
->__cpu_power
) {
5474 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5479 if (!cpus_weight(group
->cpumask
)) {
5481 printk(KERN_ERR
"ERROR: empty group\n");
5485 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5487 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5491 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5493 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5496 group
= group
->next
;
5497 } while (group
!= sd
->groups
);
5500 if (!cpus_equal(sd
->span
, groupmask
))
5501 printk(KERN_ERR
"ERROR: groups don't span "
5509 if (!cpus_subset(groupmask
, sd
->span
))
5510 printk(KERN_ERR
"ERROR: parent span is not a superset "
5511 "of domain->span\n");
5516 # define sched_domain_debug(sd, cpu) do { } while (0)
5519 static int sd_degenerate(struct sched_domain
*sd
)
5521 if (cpus_weight(sd
->span
) == 1)
5524 /* Following flags need at least 2 groups */
5525 if (sd
->flags
& (SD_LOAD_BALANCE
|
5526 SD_BALANCE_NEWIDLE
|
5530 SD_SHARE_PKG_RESOURCES
)) {
5531 if (sd
->groups
!= sd
->groups
->next
)
5535 /* Following flags don't use groups */
5536 if (sd
->flags
& (SD_WAKE_IDLE
|
5545 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5547 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5549 if (sd_degenerate(parent
))
5552 if (!cpus_equal(sd
->span
, parent
->span
))
5555 /* Does parent contain flags not in child? */
5556 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5557 if (cflags
& SD_WAKE_AFFINE
)
5558 pflags
&= ~SD_WAKE_BALANCE
;
5559 /* Flags needing groups don't count if only 1 group in parent */
5560 if (parent
->groups
== parent
->groups
->next
) {
5561 pflags
&= ~(SD_LOAD_BALANCE
|
5562 SD_BALANCE_NEWIDLE
|
5566 SD_SHARE_PKG_RESOURCES
);
5568 if (~cflags
& pflags
)
5575 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5576 * hold the hotplug lock.
5578 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5580 struct rq
*rq
= cpu_rq(cpu
);
5581 struct sched_domain
*tmp
;
5583 /* Remove the sched domains which do not contribute to scheduling. */
5584 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5585 struct sched_domain
*parent
= tmp
->parent
;
5588 if (sd_parent_degenerate(tmp
, parent
)) {
5589 tmp
->parent
= parent
->parent
;
5591 parent
->parent
->child
= tmp
;
5595 if (sd
&& sd_degenerate(sd
)) {
5601 sched_domain_debug(sd
, cpu
);
5603 rcu_assign_pointer(rq
->sd
, sd
);
5606 /* cpus with isolated domains */
5607 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5609 /* Setup the mask of cpus configured for isolated domains */
5610 static int __init
isolated_cpu_setup(char *str
)
5612 int ints
[NR_CPUS
], i
;
5614 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5615 cpus_clear(cpu_isolated_map
);
5616 for (i
= 1; i
<= ints
[0]; i
++)
5617 if (ints
[i
] < NR_CPUS
)
5618 cpu_set(ints
[i
], cpu_isolated_map
);
5622 __setup("isolcpus=", isolated_cpu_setup
);
5625 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5626 * to a function which identifies what group(along with sched group) a CPU
5627 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5628 * (due to the fact that we keep track of groups covered with a cpumask_t).
5630 * init_sched_build_groups will build a circular linked list of the groups
5631 * covered by the given span, and will set each group's ->cpumask correctly,
5632 * and ->cpu_power to 0.
5635 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5636 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5637 struct sched_group
**sg
))
5639 struct sched_group
*first
= NULL
, *last
= NULL
;
5640 cpumask_t covered
= CPU_MASK_NONE
;
5643 for_each_cpu_mask(i
, span
) {
5644 struct sched_group
*sg
;
5645 int group
= group_fn(i
, cpu_map
, &sg
);
5648 if (cpu_isset(i
, covered
))
5651 sg
->cpumask
= CPU_MASK_NONE
;
5652 sg
->__cpu_power
= 0;
5654 for_each_cpu_mask(j
, span
) {
5655 if (group_fn(j
, cpu_map
, NULL
) != group
)
5658 cpu_set(j
, covered
);
5659 cpu_set(j
, sg
->cpumask
);
5670 #define SD_NODES_PER_DOMAIN 16
5675 * find_next_best_node - find the next node to include in a sched_domain
5676 * @node: node whose sched_domain we're building
5677 * @used_nodes: nodes already in the sched_domain
5679 * Find the next node to include in a given scheduling domain. Simply
5680 * finds the closest node not already in the @used_nodes map.
5682 * Should use nodemask_t.
5684 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5686 int i
, n
, val
, min_val
, best_node
= 0;
5690 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5691 /* Start at @node */
5692 n
= (node
+ i
) % MAX_NUMNODES
;
5694 if (!nr_cpus_node(n
))
5697 /* Skip already used nodes */
5698 if (test_bit(n
, used_nodes
))
5701 /* Simple min distance search */
5702 val
= node_distance(node
, n
);
5704 if (val
< min_val
) {
5710 set_bit(best_node
, used_nodes
);
5715 * sched_domain_node_span - get a cpumask for a node's sched_domain
5716 * @node: node whose cpumask we're constructing
5717 * @size: number of nodes to include in this span
5719 * Given a node, construct a good cpumask for its sched_domain to span. It
5720 * should be one that prevents unnecessary balancing, but also spreads tasks
5723 static cpumask_t
sched_domain_node_span(int node
)
5725 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5726 cpumask_t span
, nodemask
;
5730 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5732 nodemask
= node_to_cpumask(node
);
5733 cpus_or(span
, span
, nodemask
);
5734 set_bit(node
, used_nodes
);
5736 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5737 int next_node
= find_next_best_node(node
, used_nodes
);
5739 nodemask
= node_to_cpumask(next_node
);
5740 cpus_or(span
, span
, nodemask
);
5747 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5750 * SMT sched-domains:
5752 #ifdef CONFIG_SCHED_SMT
5753 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5754 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5756 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5757 struct sched_group
**sg
)
5760 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5766 * multi-core sched-domains:
5768 #ifdef CONFIG_SCHED_MC
5769 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5770 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5773 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5774 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5775 struct sched_group
**sg
)
5778 cpumask_t mask
= cpu_sibling_map
[cpu
];
5779 cpus_and(mask
, mask
, *cpu_map
);
5780 group
= first_cpu(mask
);
5782 *sg
= &per_cpu(sched_group_core
, group
);
5785 #elif defined(CONFIG_SCHED_MC)
5786 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5787 struct sched_group
**sg
)
5790 *sg
= &per_cpu(sched_group_core
, cpu
);
5795 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5796 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5798 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5799 struct sched_group
**sg
)
5802 #ifdef CONFIG_SCHED_MC
5803 cpumask_t mask
= cpu_coregroup_map(cpu
);
5804 cpus_and(mask
, mask
, *cpu_map
);
5805 group
= first_cpu(mask
);
5806 #elif defined(CONFIG_SCHED_SMT)
5807 cpumask_t mask
= cpu_sibling_map
[cpu
];
5808 cpus_and(mask
, mask
, *cpu_map
);
5809 group
= first_cpu(mask
);
5814 *sg
= &per_cpu(sched_group_phys
, group
);
5820 * The init_sched_build_groups can't handle what we want to do with node
5821 * groups, so roll our own. Now each node has its own list of groups which
5822 * gets dynamically allocated.
5824 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5825 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5827 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5828 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5830 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5831 struct sched_group
**sg
)
5833 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5836 cpus_and(nodemask
, nodemask
, *cpu_map
);
5837 group
= first_cpu(nodemask
);
5840 *sg
= &per_cpu(sched_group_allnodes
, group
);
5844 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5846 struct sched_group
*sg
= group_head
;
5852 for_each_cpu_mask(j
, sg
->cpumask
) {
5853 struct sched_domain
*sd
;
5855 sd
= &per_cpu(phys_domains
, j
);
5856 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5858 * Only add "power" once for each
5864 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5867 } while (sg
!= group_head
);
5872 /* Free memory allocated for various sched_group structures */
5873 static void free_sched_groups(const cpumask_t
*cpu_map
)
5877 for_each_cpu_mask(cpu
, *cpu_map
) {
5878 struct sched_group
**sched_group_nodes
5879 = sched_group_nodes_bycpu
[cpu
];
5881 if (!sched_group_nodes
)
5884 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5885 cpumask_t nodemask
= node_to_cpumask(i
);
5886 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5888 cpus_and(nodemask
, nodemask
, *cpu_map
);
5889 if (cpus_empty(nodemask
))
5899 if (oldsg
!= sched_group_nodes
[i
])
5902 kfree(sched_group_nodes
);
5903 sched_group_nodes_bycpu
[cpu
] = NULL
;
5907 static void free_sched_groups(const cpumask_t
*cpu_map
)
5913 * Initialize sched groups cpu_power.
5915 * cpu_power indicates the capacity of sched group, which is used while
5916 * distributing the load between different sched groups in a sched domain.
5917 * Typically cpu_power for all the groups in a sched domain will be same unless
5918 * there are asymmetries in the topology. If there are asymmetries, group
5919 * having more cpu_power will pickup more load compared to the group having
5922 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5923 * the maximum number of tasks a group can handle in the presence of other idle
5924 * or lightly loaded groups in the same sched domain.
5926 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5928 struct sched_domain
*child
;
5929 struct sched_group
*group
;
5931 WARN_ON(!sd
|| !sd
->groups
);
5933 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5938 sd
->groups
->__cpu_power
= 0;
5941 * For perf policy, if the groups in child domain share resources
5942 * (for example cores sharing some portions of the cache hierarchy
5943 * or SMT), then set this domain groups cpu_power such that each group
5944 * can handle only one task, when there are other idle groups in the
5945 * same sched domain.
5947 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5949 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5950 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5955 * add cpu_power of each child group to this groups cpu_power
5957 group
= child
->groups
;
5959 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5960 group
= group
->next
;
5961 } while (group
!= child
->groups
);
5965 * Build sched domains for a given set of cpus and attach the sched domains
5966 * to the individual cpus
5968 static int build_sched_domains(const cpumask_t
*cpu_map
)
5972 struct sched_group
**sched_group_nodes
= NULL
;
5973 int sd_allnodes
= 0;
5976 * Allocate the per-node list of sched groups
5978 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5980 if (!sched_group_nodes
) {
5981 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5984 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5988 * Set up domains for cpus specified by the cpu_map.
5990 for_each_cpu_mask(i
, *cpu_map
) {
5991 struct sched_domain
*sd
= NULL
, *p
;
5992 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5994 cpus_and(nodemask
, nodemask
, *cpu_map
);
5997 if (cpus_weight(*cpu_map
) >
5998 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5999 sd
= &per_cpu(allnodes_domains
, i
);
6000 *sd
= SD_ALLNODES_INIT
;
6001 sd
->span
= *cpu_map
;
6002 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6008 sd
= &per_cpu(node_domains
, i
);
6010 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6014 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6018 sd
= &per_cpu(phys_domains
, i
);
6020 sd
->span
= nodemask
;
6024 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6026 #ifdef CONFIG_SCHED_MC
6028 sd
= &per_cpu(core_domains
, i
);
6030 sd
->span
= cpu_coregroup_map(i
);
6031 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6034 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6037 #ifdef CONFIG_SCHED_SMT
6039 sd
= &per_cpu(cpu_domains
, i
);
6040 *sd
= SD_SIBLING_INIT
;
6041 sd
->span
= cpu_sibling_map
[i
];
6042 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6045 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6049 #ifdef CONFIG_SCHED_SMT
6050 /* Set up CPU (sibling) groups */
6051 for_each_cpu_mask(i
, *cpu_map
) {
6052 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6053 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6054 if (i
!= first_cpu(this_sibling_map
))
6057 init_sched_build_groups(this_sibling_map
, cpu_map
,
6062 #ifdef CONFIG_SCHED_MC
6063 /* Set up multi-core groups */
6064 for_each_cpu_mask(i
, *cpu_map
) {
6065 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6066 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6067 if (i
!= first_cpu(this_core_map
))
6069 init_sched_build_groups(this_core_map
, cpu_map
,
6070 &cpu_to_core_group
);
6074 /* Set up physical groups */
6075 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6076 cpumask_t nodemask
= node_to_cpumask(i
);
6078 cpus_and(nodemask
, nodemask
, *cpu_map
);
6079 if (cpus_empty(nodemask
))
6082 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6086 /* Set up node groups */
6088 init_sched_build_groups(*cpu_map
, cpu_map
,
6089 &cpu_to_allnodes_group
);
6091 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6092 /* Set up node groups */
6093 struct sched_group
*sg
, *prev
;
6094 cpumask_t nodemask
= node_to_cpumask(i
);
6095 cpumask_t domainspan
;
6096 cpumask_t covered
= CPU_MASK_NONE
;
6099 cpus_and(nodemask
, nodemask
, *cpu_map
);
6100 if (cpus_empty(nodemask
)) {
6101 sched_group_nodes
[i
] = NULL
;
6105 domainspan
= sched_domain_node_span(i
);
6106 cpus_and(domainspan
, domainspan
, *cpu_map
);
6108 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6110 printk(KERN_WARNING
"Can not alloc domain group for "
6114 sched_group_nodes
[i
] = sg
;
6115 for_each_cpu_mask(j
, nodemask
) {
6116 struct sched_domain
*sd
;
6118 sd
= &per_cpu(node_domains
, j
);
6121 sg
->__cpu_power
= 0;
6122 sg
->cpumask
= nodemask
;
6124 cpus_or(covered
, covered
, nodemask
);
6127 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6128 cpumask_t tmp
, notcovered
;
6129 int n
= (i
+ j
) % MAX_NUMNODES
;
6131 cpus_complement(notcovered
, covered
);
6132 cpus_and(tmp
, notcovered
, *cpu_map
);
6133 cpus_and(tmp
, tmp
, domainspan
);
6134 if (cpus_empty(tmp
))
6137 nodemask
= node_to_cpumask(n
);
6138 cpus_and(tmp
, tmp
, nodemask
);
6139 if (cpus_empty(tmp
))
6142 sg
= kmalloc_node(sizeof(struct sched_group
),
6146 "Can not alloc domain group for node %d\n", j
);
6149 sg
->__cpu_power
= 0;
6151 sg
->next
= prev
->next
;
6152 cpus_or(covered
, covered
, tmp
);
6159 /* Calculate CPU power for physical packages and nodes */
6160 #ifdef CONFIG_SCHED_SMT
6161 for_each_cpu_mask(i
, *cpu_map
) {
6162 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6164 init_sched_groups_power(i
, sd
);
6167 #ifdef CONFIG_SCHED_MC
6168 for_each_cpu_mask(i
, *cpu_map
) {
6169 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6171 init_sched_groups_power(i
, sd
);
6175 for_each_cpu_mask(i
, *cpu_map
) {
6176 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6178 init_sched_groups_power(i
, sd
);
6182 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6183 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6186 struct sched_group
*sg
;
6188 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6189 init_numa_sched_groups_power(sg
);
6193 /* Attach the domains */
6194 for_each_cpu_mask(i
, *cpu_map
) {
6195 struct sched_domain
*sd
;
6196 #ifdef CONFIG_SCHED_SMT
6197 sd
= &per_cpu(cpu_domains
, i
);
6198 #elif defined(CONFIG_SCHED_MC)
6199 sd
= &per_cpu(core_domains
, i
);
6201 sd
= &per_cpu(phys_domains
, i
);
6203 cpu_attach_domain(sd
, i
);
6210 free_sched_groups(cpu_map
);
6215 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6217 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6219 cpumask_t cpu_default_map
;
6223 * Setup mask for cpus without special case scheduling requirements.
6224 * For now this just excludes isolated cpus, but could be used to
6225 * exclude other special cases in the future.
6227 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6229 err
= build_sched_domains(&cpu_default_map
);
6234 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6236 free_sched_groups(cpu_map
);
6240 * Detach sched domains from a group of cpus specified in cpu_map
6241 * These cpus will now be attached to the NULL domain
6243 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6247 for_each_cpu_mask(i
, *cpu_map
)
6248 cpu_attach_domain(NULL
, i
);
6249 synchronize_sched();
6250 arch_destroy_sched_domains(cpu_map
);
6254 * Partition sched domains as specified by the cpumasks below.
6255 * This attaches all cpus from the cpumasks to the NULL domain,
6256 * waits for a RCU quiescent period, recalculates sched
6257 * domain information and then attaches them back to the
6258 * correct sched domains
6259 * Call with hotplug lock held
6261 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6263 cpumask_t change_map
;
6266 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6267 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6268 cpus_or(change_map
, *partition1
, *partition2
);
6270 /* Detach sched domains from all of the affected cpus */
6271 detach_destroy_domains(&change_map
);
6272 if (!cpus_empty(*partition1
))
6273 err
= build_sched_domains(partition1
);
6274 if (!err
&& !cpus_empty(*partition2
))
6275 err
= build_sched_domains(partition2
);
6280 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6281 static int arch_reinit_sched_domains(void)
6285 mutex_lock(&sched_hotcpu_mutex
);
6286 detach_destroy_domains(&cpu_online_map
);
6287 err
= arch_init_sched_domains(&cpu_online_map
);
6288 mutex_unlock(&sched_hotcpu_mutex
);
6293 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6297 if (buf
[0] != '0' && buf
[0] != '1')
6301 sched_smt_power_savings
= (buf
[0] == '1');
6303 sched_mc_power_savings
= (buf
[0] == '1');
6305 ret
= arch_reinit_sched_domains();
6307 return ret
? ret
: count
;
6310 #ifdef CONFIG_SCHED_MC
6311 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6313 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6315 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6316 const char *buf
, size_t count
)
6318 return sched_power_savings_store(buf
, count
, 0);
6320 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6321 sched_mc_power_savings_store
);
6324 #ifdef CONFIG_SCHED_SMT
6325 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6327 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6329 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6330 const char *buf
, size_t count
)
6332 return sched_power_savings_store(buf
, count
, 1);
6334 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6335 sched_smt_power_savings_store
);
6338 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6342 #ifdef CONFIG_SCHED_SMT
6344 err
= sysfs_create_file(&cls
->kset
.kobj
,
6345 &attr_sched_smt_power_savings
.attr
);
6347 #ifdef CONFIG_SCHED_MC
6348 if (!err
&& mc_capable())
6349 err
= sysfs_create_file(&cls
->kset
.kobj
,
6350 &attr_sched_mc_power_savings
.attr
);
6357 * Force a reinitialization of the sched domains hierarchy. The domains
6358 * and groups cannot be updated in place without racing with the balancing
6359 * code, so we temporarily attach all running cpus to the NULL domain
6360 * which will prevent rebalancing while the sched domains are recalculated.
6362 static int update_sched_domains(struct notifier_block
*nfb
,
6363 unsigned long action
, void *hcpu
)
6366 case CPU_UP_PREPARE
:
6367 case CPU_UP_PREPARE_FROZEN
:
6368 case CPU_DOWN_PREPARE
:
6369 case CPU_DOWN_PREPARE_FROZEN
:
6370 detach_destroy_domains(&cpu_online_map
);
6373 case CPU_UP_CANCELED
:
6374 case CPU_UP_CANCELED_FROZEN
:
6375 case CPU_DOWN_FAILED
:
6376 case CPU_DOWN_FAILED_FROZEN
:
6378 case CPU_ONLINE_FROZEN
:
6380 case CPU_DEAD_FROZEN
:
6382 * Fall through and re-initialise the domains.
6389 /* The hotplug lock is already held by cpu_up/cpu_down */
6390 arch_init_sched_domains(&cpu_online_map
);
6395 void __init
sched_init_smp(void)
6397 cpumask_t non_isolated_cpus
;
6399 mutex_lock(&sched_hotcpu_mutex
);
6400 arch_init_sched_domains(&cpu_online_map
);
6401 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6402 if (cpus_empty(non_isolated_cpus
))
6403 cpu_set(smp_processor_id(), non_isolated_cpus
);
6404 mutex_unlock(&sched_hotcpu_mutex
);
6405 /* XXX: Theoretical race here - CPU may be hotplugged now */
6406 hotcpu_notifier(update_sched_domains
, 0);
6408 init_sched_domain_sysctl();
6410 /* Move init over to a non-isolated CPU */
6411 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6415 void __init
sched_init_smp(void)
6418 #endif /* CONFIG_SMP */
6420 int in_sched_functions(unsigned long addr
)
6422 /* Linker adds these: start and end of __sched functions */
6423 extern char __sched_text_start
[], __sched_text_end
[];
6425 return in_lock_functions(addr
) ||
6426 (addr
>= (unsigned long)__sched_text_start
6427 && addr
< (unsigned long)__sched_text_end
);
6430 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6432 cfs_rq
->tasks_timeline
= RB_ROOT
;
6433 #ifdef CONFIG_FAIR_GROUP_SCHED
6436 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6439 void __init
sched_init(void)
6441 int highest_cpu
= 0;
6444 for_each_possible_cpu(i
) {
6445 struct rt_prio_array
*array
;
6449 spin_lock_init(&rq
->lock
);
6450 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6453 init_cfs_rq(&rq
->cfs
, rq
);
6454 #ifdef CONFIG_FAIR_GROUP_SCHED
6455 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6457 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6458 struct sched_entity
*se
=
6459 &per_cpu(init_sched_entity
, i
);
6461 init_cfs_rq_p
[i
] = cfs_rq
;
6462 init_cfs_rq(cfs_rq
, rq
);
6463 cfs_rq
->tg
= &init_task_group
;
6464 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6465 &rq
->leaf_cfs_rq_list
);
6467 init_sched_entity_p
[i
] = se
;
6468 se
->cfs_rq
= &rq
->cfs
;
6470 se
->load
.weight
= init_task_group_load
;
6471 se
->load
.inv_weight
=
6472 div64_64(1ULL<<32, init_task_group_load
);
6475 init_task_group
.shares
= init_task_group_load
;
6476 spin_lock_init(&init_task_group
.lock
);
6479 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6480 rq
->cpu_load
[j
] = 0;
6483 rq
->active_balance
= 0;
6484 rq
->next_balance
= jiffies
;
6487 rq
->migration_thread
= NULL
;
6488 INIT_LIST_HEAD(&rq
->migration_queue
);
6490 atomic_set(&rq
->nr_iowait
, 0);
6492 array
= &rq
->rt
.active
;
6493 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6494 INIT_LIST_HEAD(array
->queue
+ j
);
6495 __clear_bit(j
, array
->bitmap
);
6498 /* delimiter for bitsearch: */
6499 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6502 set_load_weight(&init_task
);
6504 #ifdef CONFIG_PREEMPT_NOTIFIERS
6505 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6509 nr_cpu_ids
= highest_cpu
+ 1;
6510 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6513 #ifdef CONFIG_RT_MUTEXES
6514 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6518 * The boot idle thread does lazy MMU switching as well:
6520 atomic_inc(&init_mm
.mm_count
);
6521 enter_lazy_tlb(&init_mm
, current
);
6524 * Make us the idle thread. Technically, schedule() should not be
6525 * called from this thread, however somewhere below it might be,
6526 * but because we are the idle thread, we just pick up running again
6527 * when this runqueue becomes "idle".
6529 init_idle(current
, smp_processor_id());
6531 * During early bootup we pretend to be a normal task:
6533 current
->sched_class
= &fair_sched_class
;
6536 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6537 void __might_sleep(char *file
, int line
)
6540 static unsigned long prev_jiffy
; /* ratelimiting */
6542 if ((in_atomic() || irqs_disabled()) &&
6543 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6544 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6546 prev_jiffy
= jiffies
;
6547 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6548 " context at %s:%d\n", file
, line
);
6549 printk("in_atomic():%d, irqs_disabled():%d\n",
6550 in_atomic(), irqs_disabled());
6551 debug_show_held_locks(current
);
6552 if (irqs_disabled())
6553 print_irqtrace_events(current
);
6558 EXPORT_SYMBOL(__might_sleep
);
6561 #ifdef CONFIG_MAGIC_SYSRQ
6562 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6565 update_rq_clock(rq
);
6566 on_rq
= p
->se
.on_rq
;
6568 deactivate_task(rq
, p
, 0);
6569 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6571 activate_task(rq
, p
, 0);
6572 resched_task(rq
->curr
);
6576 void normalize_rt_tasks(void)
6578 struct task_struct
*g
, *p
;
6579 unsigned long flags
;
6582 read_lock_irq(&tasklist_lock
);
6583 do_each_thread(g
, p
) {
6585 * Only normalize user tasks:
6590 p
->se
.exec_start
= 0;
6591 #ifdef CONFIG_SCHEDSTATS
6592 p
->se
.wait_start
= 0;
6593 p
->se
.sleep_start
= 0;
6594 p
->se
.block_start
= 0;
6596 task_rq(p
)->clock
= 0;
6600 * Renice negative nice level userspace
6603 if (TASK_NICE(p
) < 0 && p
->mm
)
6604 set_user_nice(p
, 0);
6608 spin_lock_irqsave(&p
->pi_lock
, flags
);
6609 rq
= __task_rq_lock(p
);
6611 normalize_task(rq
, p
);
6613 __task_rq_unlock(rq
);
6614 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6615 } while_each_thread(g
, p
);
6617 read_unlock_irq(&tasklist_lock
);
6620 #endif /* CONFIG_MAGIC_SYSRQ */
6624 * These functions are only useful for the IA64 MCA handling.
6626 * They can only be called when the whole system has been
6627 * stopped - every CPU needs to be quiescent, and no scheduling
6628 * activity can take place. Using them for anything else would
6629 * be a serious bug, and as a result, they aren't even visible
6630 * under any other configuration.
6634 * curr_task - return the current task for a given cpu.
6635 * @cpu: the processor in question.
6637 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6639 struct task_struct
*curr_task(int cpu
)
6641 return cpu_curr(cpu
);
6645 * set_curr_task - set the current task for a given cpu.
6646 * @cpu: the processor in question.
6647 * @p: the task pointer to set.
6649 * Description: This function must only be used when non-maskable interrupts
6650 * are serviced on a separate stack. It allows the architecture to switch the
6651 * notion of the current task on a cpu in a non-blocking manner. This function
6652 * must be called with all CPU's synchronized, and interrupts disabled, the
6653 * and caller must save the original value of the current task (see
6654 * curr_task() above) and restore that value before reenabling interrupts and
6655 * re-starting the system.
6657 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6659 void set_curr_task(int cpu
, struct task_struct
*p
)
6666 #ifdef CONFIG_FAIR_GROUP_SCHED
6668 /* allocate runqueue etc for a new task group */
6669 struct task_group
*sched_create_group(void)
6671 struct task_group
*tg
;
6672 struct cfs_rq
*cfs_rq
;
6673 struct sched_entity
*se
;
6677 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6679 return ERR_PTR(-ENOMEM
);
6681 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
6684 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
6688 for_each_possible_cpu(i
) {
6691 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
6696 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
6701 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
6702 memset(se
, 0, sizeof(struct sched_entity
));
6704 tg
->cfs_rq
[i
] = cfs_rq
;
6705 init_cfs_rq(cfs_rq
, rq
);
6709 se
->cfs_rq
= &rq
->cfs
;
6711 se
->load
.weight
= NICE_0_LOAD
;
6712 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6716 for_each_possible_cpu(i
) {
6718 cfs_rq
= tg
->cfs_rq
[i
];
6719 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6722 tg
->shares
= NICE_0_LOAD
;
6723 spin_lock_init(&tg
->lock
);
6728 for_each_possible_cpu(i
) {
6730 kfree(tg
->cfs_rq
[i
]);
6738 return ERR_PTR(-ENOMEM
);
6741 /* rcu callback to free various structures associated with a task group */
6742 static void free_sched_group(struct rcu_head
*rhp
)
6744 struct cfs_rq
*cfs_rq
= container_of(rhp
, struct cfs_rq
, rcu
);
6745 struct task_group
*tg
= cfs_rq
->tg
;
6746 struct sched_entity
*se
;
6749 /* now it should be safe to free those cfs_rqs */
6750 for_each_possible_cpu(i
) {
6751 cfs_rq
= tg
->cfs_rq
[i
];
6763 /* Destroy runqueue etc associated with a task group */
6764 void sched_destroy_group(struct task_group
*tg
)
6766 struct cfs_rq
*cfs_rq
;
6769 for_each_possible_cpu(i
) {
6770 cfs_rq
= tg
->cfs_rq
[i
];
6771 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
6774 cfs_rq
= tg
->cfs_rq
[0];
6776 /* wait for possible concurrent references to cfs_rqs complete */
6777 call_rcu(&cfs_rq
->rcu
, free_sched_group
);
6780 /* change task's runqueue when it moves between groups.
6781 * The caller of this function should have put the task in its new group
6782 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6783 * reflect its new group.
6785 void sched_move_task(struct task_struct
*tsk
)
6788 unsigned long flags
;
6791 rq
= task_rq_lock(tsk
, &flags
);
6793 if (tsk
->sched_class
!= &fair_sched_class
)
6796 update_rq_clock(rq
);
6798 running
= task_running(rq
, tsk
);
6799 on_rq
= tsk
->se
.on_rq
;
6802 dequeue_task(rq
, tsk
, 0);
6803 if (unlikely(running
))
6804 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6807 set_task_cfs_rq(tsk
);
6810 if (unlikely(running
))
6811 tsk
->sched_class
->set_curr_task(rq
);
6812 enqueue_task(rq
, tsk
, 0);
6816 task_rq_unlock(rq
, &flags
);
6819 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
6821 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
6822 struct rq
*rq
= cfs_rq
->rq
;
6825 spin_lock_irq(&rq
->lock
);
6829 dequeue_entity(cfs_rq
, se
, 0);
6831 se
->load
.weight
= shares
;
6832 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
6835 enqueue_entity(cfs_rq
, se
, 0);
6837 spin_unlock_irq(&rq
->lock
);
6840 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6844 spin_lock(&tg
->lock
);
6845 if (tg
->shares
== shares
)
6848 tg
->shares
= shares
;
6849 for_each_possible_cpu(i
)
6850 set_se_shares(tg
->se
[i
], shares
);
6853 spin_unlock(&tg
->lock
);
6857 unsigned long sched_group_shares(struct task_group
*tg
)
6862 #endif /* CONFIG_FAIR_GROUP_SCHED */