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
)
368 static inline int is_migration_thread(struct task_struct
*p
, struct rq
*rq
)
371 return p
== rq
->migration_thread
;
378 * Update the per-runqueue clock, as finegrained as the platform can give
379 * us, but without assuming monotonicity, etc.:
381 static void __update_rq_clock(struct rq
*rq
)
383 u64 prev_raw
= rq
->prev_clock_raw
;
384 u64 now
= sched_clock();
385 s64 delta
= now
- prev_raw
;
386 u64 clock
= rq
->clock
;
388 #ifdef CONFIG_SCHED_DEBUG
389 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
392 * Protect against sched_clock() occasionally going backwards:
394 if (unlikely(delta
< 0)) {
399 * Catch too large forward jumps too:
401 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
402 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
403 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
406 rq
->clock_overflows
++;
408 if (unlikely(delta
> rq
->clock_max_delta
))
409 rq
->clock_max_delta
= delta
;
414 rq
->prev_clock_raw
= now
;
418 static void update_rq_clock(struct rq
*rq
)
420 if (likely(smp_processor_id() == cpu_of(rq
)))
421 __update_rq_clock(rq
);
425 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
426 * See detach_destroy_domains: synchronize_sched for details.
428 * The domain tree of any CPU may only be accessed from within
429 * preempt-disabled sections.
431 #define for_each_domain(cpu, __sd) \
432 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
434 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
435 #define this_rq() (&__get_cpu_var(runqueues))
436 #define task_rq(p) cpu_rq(task_cpu(p))
437 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
440 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
442 #ifdef CONFIG_SCHED_DEBUG
443 # define const_debug __read_mostly
445 # define const_debug static const
449 * Debugging: various feature bits
452 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
453 SCHED_FEAT_START_DEBIT
= 2,
454 SCHED_FEAT_TREE_AVG
= 4,
455 SCHED_FEAT_APPROX_AVG
= 8,
456 SCHED_FEAT_WAKEUP_PREEMPT
= 16,
457 SCHED_FEAT_PREEMPT_RESTRICT
= 32,
460 const_debug
unsigned int sysctl_sched_features
=
461 SCHED_FEAT_NEW_FAIR_SLEEPERS
*1 |
462 SCHED_FEAT_START_DEBIT
*1 |
463 SCHED_FEAT_TREE_AVG
*0 |
464 SCHED_FEAT_APPROX_AVG
*0 |
465 SCHED_FEAT_WAKEUP_PREEMPT
*1 |
466 SCHED_FEAT_PREEMPT_RESTRICT
*1;
468 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
471 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
472 * clock constructed from sched_clock():
474 unsigned long long cpu_clock(int cpu
)
476 unsigned long long now
;
480 local_irq_save(flags
);
484 local_irq_restore(flags
);
488 EXPORT_SYMBOL_GPL(cpu_clock
);
490 #ifndef prepare_arch_switch
491 # define prepare_arch_switch(next) do { } while (0)
493 #ifndef finish_arch_switch
494 # define finish_arch_switch(prev) do { } while (0)
497 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
498 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
500 return rq
->curr
== p
;
503 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
507 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
509 #ifdef CONFIG_DEBUG_SPINLOCK
510 /* this is a valid case when another task releases the spinlock */
511 rq
->lock
.owner
= current
;
514 * If we are tracking spinlock dependencies then we have to
515 * fix up the runqueue lock - which gets 'carried over' from
518 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
520 spin_unlock_irq(&rq
->lock
);
523 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
524 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
529 return rq
->curr
== p
;
533 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
537 * We can optimise this out completely for !SMP, because the
538 * SMP rebalancing from interrupt is the only thing that cares
543 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
544 spin_unlock_irq(&rq
->lock
);
546 spin_unlock(&rq
->lock
);
550 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
554 * After ->oncpu is cleared, the task can be moved to a different CPU.
555 * We must ensure this doesn't happen until the switch is completely
561 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
565 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
568 * __task_rq_lock - lock the runqueue a given task resides on.
569 * Must be called interrupts disabled.
571 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
575 struct rq
*rq
= task_rq(p
);
576 spin_lock(&rq
->lock
);
577 if (likely(rq
== task_rq(p
)))
579 spin_unlock(&rq
->lock
);
584 * task_rq_lock - lock the runqueue a given task resides on and disable
585 * interrupts. Note the ordering: we can safely lookup the task_rq without
586 * explicitly disabling preemption.
588 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
594 local_irq_save(*flags
);
596 spin_lock(&rq
->lock
);
597 if (likely(rq
== task_rq(p
)))
599 spin_unlock_irqrestore(&rq
->lock
, *flags
);
603 static void __task_rq_unlock(struct rq
*rq
)
606 spin_unlock(&rq
->lock
);
609 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
612 spin_unlock_irqrestore(&rq
->lock
, *flags
);
616 * this_rq_lock - lock this runqueue and disable interrupts.
618 static struct rq
*this_rq_lock(void)
625 spin_lock(&rq
->lock
);
631 * We are going deep-idle (irqs are disabled):
633 void sched_clock_idle_sleep_event(void)
635 struct rq
*rq
= cpu_rq(smp_processor_id());
637 spin_lock(&rq
->lock
);
638 __update_rq_clock(rq
);
639 spin_unlock(&rq
->lock
);
640 rq
->clock_deep_idle_events
++;
642 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
645 * We just idled delta nanoseconds (called with irqs disabled):
647 void sched_clock_idle_wakeup_event(u64 delta_ns
)
649 struct rq
*rq
= cpu_rq(smp_processor_id());
650 u64 now
= sched_clock();
652 rq
->idle_clock
+= delta_ns
;
654 * Override the previous timestamp and ignore all
655 * sched_clock() deltas that occured while we idled,
656 * and use the PM-provided delta_ns to advance the
659 spin_lock(&rq
->lock
);
660 rq
->prev_clock_raw
= now
;
661 rq
->clock
+= delta_ns
;
662 spin_unlock(&rq
->lock
);
664 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
667 * resched_task - mark a task 'to be rescheduled now'.
669 * On UP this means the setting of the need_resched flag, on SMP it
670 * might also involve a cross-CPU call to trigger the scheduler on
675 #ifndef tsk_is_polling
676 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
679 static void resched_task(struct task_struct
*p
)
683 assert_spin_locked(&task_rq(p
)->lock
);
685 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
688 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
691 if (cpu
== smp_processor_id())
694 /* NEED_RESCHED must be visible before we test polling */
696 if (!tsk_is_polling(p
))
697 smp_send_reschedule(cpu
);
700 static void resched_cpu(int cpu
)
702 struct rq
*rq
= cpu_rq(cpu
);
705 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
707 resched_task(cpu_curr(cpu
));
708 spin_unlock_irqrestore(&rq
->lock
, flags
);
711 static inline void resched_task(struct task_struct
*p
)
713 assert_spin_locked(&task_rq(p
)->lock
);
714 set_tsk_need_resched(p
);
718 #if BITS_PER_LONG == 32
719 # define WMULT_CONST (~0UL)
721 # define WMULT_CONST (1UL << 32)
724 #define WMULT_SHIFT 32
727 * Shift right and round:
729 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
732 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
733 struct load_weight
*lw
)
737 if (unlikely(!lw
->inv_weight
))
738 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
740 tmp
= (u64
)delta_exec
* weight
;
742 * Check whether we'd overflow the 64-bit multiplication:
744 if (unlikely(tmp
> WMULT_CONST
))
745 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
748 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
750 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
753 static inline unsigned long
754 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
756 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
759 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
764 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
770 * To aid in avoiding the subversion of "niceness" due to uneven distribution
771 * of tasks with abnormal "nice" values across CPUs the contribution that
772 * each task makes to its run queue's load is weighted according to its
773 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
774 * scaled version of the new time slice allocation that they receive on time
778 #define WEIGHT_IDLEPRIO 2
779 #define WMULT_IDLEPRIO (1 << 31)
782 * Nice levels are multiplicative, with a gentle 10% change for every
783 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
784 * nice 1, it will get ~10% less CPU time than another CPU-bound task
785 * that remained on nice 0.
787 * The "10% effect" is relative and cumulative: from _any_ nice level,
788 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
789 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
790 * If a task goes up by ~10% and another task goes down by ~10% then
791 * the relative distance between them is ~25%.)
793 static const int prio_to_weight
[40] = {
794 /* -20 */ 88761, 71755, 56483, 46273, 36291,
795 /* -15 */ 29154, 23254, 18705, 14949, 11916,
796 /* -10 */ 9548, 7620, 6100, 4904, 3906,
797 /* -5 */ 3121, 2501, 1991, 1586, 1277,
798 /* 0 */ 1024, 820, 655, 526, 423,
799 /* 5 */ 335, 272, 215, 172, 137,
800 /* 10 */ 110, 87, 70, 56, 45,
801 /* 15 */ 36, 29, 23, 18, 15,
805 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
807 * In cases where the weight does not change often, we can use the
808 * precalculated inverse to speed up arithmetics by turning divisions
809 * into multiplications:
811 static const u32 prio_to_wmult
[40] = {
812 /* -20 */ 48388, 59856, 76040, 92818, 118348,
813 /* -15 */ 147320, 184698, 229616, 287308, 360437,
814 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
815 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
816 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
817 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
818 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
819 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
822 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
825 * runqueue iterator, to support SMP load-balancing between different
826 * scheduling classes, without having to expose their internal data
827 * structures to the load-balancing proper:
831 struct task_struct
*(*start
)(void *);
832 struct task_struct
*(*next
)(void *);
835 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
836 unsigned long max_nr_move
, unsigned long max_load_move
,
837 struct sched_domain
*sd
, enum cpu_idle_type idle
,
838 int *all_pinned
, unsigned long *load_moved
,
839 int *this_best_prio
, struct rq_iterator
*iterator
);
841 #include "sched_stats.h"
842 #include "sched_idletask.c"
843 #include "sched_fair.c"
844 #include "sched_rt.c"
845 #ifdef CONFIG_SCHED_DEBUG
846 # include "sched_debug.c"
849 #define sched_class_highest (&rt_sched_class)
852 * Update delta_exec, delta_fair fields for rq.
854 * delta_fair clock advances at a rate inversely proportional to
855 * total load (rq->load.weight) on the runqueue, while
856 * delta_exec advances at the same rate as wall-clock (provided
859 * delta_exec / delta_fair is a measure of the (smoothened) load on this
860 * runqueue over any given interval. This (smoothened) load is used
861 * during load balance.
863 * This function is called /before/ updating rq->load
864 * and when switching tasks.
866 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
868 update_load_add(&rq
->load
, p
->se
.load
.weight
);
871 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
873 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
876 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
882 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
888 static void set_load_weight(struct task_struct
*p
)
890 if (task_has_rt_policy(p
)) {
891 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
892 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
897 * SCHED_IDLE tasks get minimal weight:
899 if (p
->policy
== SCHED_IDLE
) {
900 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
901 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
905 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
906 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
909 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
911 sched_info_queued(p
);
912 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
916 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
918 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
923 * __normal_prio - return the priority that is based on the static prio
925 static inline int __normal_prio(struct task_struct
*p
)
927 return p
->static_prio
;
931 * Calculate the expected normal priority: i.e. priority
932 * without taking RT-inheritance into account. Might be
933 * boosted by interactivity modifiers. Changes upon fork,
934 * setprio syscalls, and whenever the interactivity
935 * estimator recalculates.
937 static inline int normal_prio(struct task_struct
*p
)
941 if (task_has_rt_policy(p
))
942 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
944 prio
= __normal_prio(p
);
949 * Calculate the current priority, i.e. the priority
950 * taken into account by the scheduler. This value might
951 * be boosted by RT tasks, or might be boosted by
952 * interactivity modifiers. Will be RT if the task got
953 * RT-boosted. If not then it returns p->normal_prio.
955 static int effective_prio(struct task_struct
*p
)
957 p
->normal_prio
= normal_prio(p
);
959 * If we are RT tasks or we were boosted to RT priority,
960 * keep the priority unchanged. Otherwise, update priority
961 * to the normal priority:
963 if (!rt_prio(p
->prio
))
964 return p
->normal_prio
;
969 * activate_task - move a task to the runqueue.
971 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
973 if (p
->state
== TASK_UNINTERRUPTIBLE
)
974 rq
->nr_uninterruptible
--;
976 enqueue_task(rq
, p
, wakeup
);
977 inc_nr_running(p
, rq
);
981 * deactivate_task - remove a task from the runqueue.
983 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
985 if (p
->state
== TASK_UNINTERRUPTIBLE
)
986 rq
->nr_uninterruptible
++;
988 dequeue_task(rq
, p
, sleep
);
989 dec_nr_running(p
, rq
);
993 * task_curr - is this task currently executing on a CPU?
994 * @p: the task in question.
996 inline int task_curr(const struct task_struct
*p
)
998 return cpu_curr(task_cpu(p
)) == p
;
1001 /* Used instead of source_load when we know the type == 0 */
1002 unsigned long weighted_cpuload(const int cpu
)
1004 return cpu_rq(cpu
)->load
.weight
;
1007 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1010 task_thread_info(p
)->cpu
= cpu
;
1017 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1019 int old_cpu
= task_cpu(p
);
1020 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1021 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1022 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1025 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1027 #ifdef CONFIG_SCHEDSTATS
1028 if (p
->se
.wait_start
)
1029 p
->se
.wait_start
-= clock_offset
;
1030 if (p
->se
.sleep_start
)
1031 p
->se
.sleep_start
-= clock_offset
;
1032 if (p
->se
.block_start
)
1033 p
->se
.block_start
-= clock_offset
;
1035 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1036 new_cfsrq
->min_vruntime
;
1038 __set_task_cpu(p
, new_cpu
);
1041 struct migration_req
{
1042 struct list_head list
;
1044 struct task_struct
*task
;
1047 struct completion done
;
1051 * The task's runqueue lock must be held.
1052 * Returns true if you have to wait for migration thread.
1055 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1057 struct rq
*rq
= task_rq(p
);
1060 * If the task is not on a runqueue (and not running), then
1061 * it is sufficient to simply update the task's cpu field.
1063 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1064 set_task_cpu(p
, dest_cpu
);
1068 init_completion(&req
->done
);
1070 req
->dest_cpu
= dest_cpu
;
1071 list_add(&req
->list
, &rq
->migration_queue
);
1077 * wait_task_inactive - wait for a thread to unschedule.
1079 * The caller must ensure that the task *will* unschedule sometime soon,
1080 * else this function might spin for a *long* time. This function can't
1081 * be called with interrupts off, or it may introduce deadlock with
1082 * smp_call_function() if an IPI is sent by the same process we are
1083 * waiting to become inactive.
1085 void wait_task_inactive(struct task_struct
*p
)
1087 unsigned long flags
;
1093 * We do the initial early heuristics without holding
1094 * any task-queue locks at all. We'll only try to get
1095 * the runqueue lock when things look like they will
1101 * If the task is actively running on another CPU
1102 * still, just relax and busy-wait without holding
1105 * NOTE! Since we don't hold any locks, it's not
1106 * even sure that "rq" stays as the right runqueue!
1107 * But we don't care, since "task_running()" will
1108 * return false if the runqueue has changed and p
1109 * is actually now running somewhere else!
1111 while (task_running(rq
, p
))
1115 * Ok, time to look more closely! We need the rq
1116 * lock now, to be *sure*. If we're wrong, we'll
1117 * just go back and repeat.
1119 rq
= task_rq_lock(p
, &flags
);
1120 running
= task_running(rq
, p
);
1121 on_rq
= p
->se
.on_rq
;
1122 task_rq_unlock(rq
, &flags
);
1125 * Was it really running after all now that we
1126 * checked with the proper locks actually held?
1128 * Oops. Go back and try again..
1130 if (unlikely(running
)) {
1136 * It's not enough that it's not actively running,
1137 * it must be off the runqueue _entirely_, and not
1140 * So if it wa still runnable (but just not actively
1141 * running right now), it's preempted, and we should
1142 * yield - it could be a while.
1144 if (unlikely(on_rq
)) {
1145 schedule_timeout_uninterruptible(1);
1150 * Ahh, all good. It wasn't running, and it wasn't
1151 * runnable, which means that it will never become
1152 * running in the future either. We're all done!
1159 * kick_process - kick a running thread to enter/exit the kernel
1160 * @p: the to-be-kicked thread
1162 * Cause a process which is running on another CPU to enter
1163 * kernel-mode, without any delay. (to get signals handled.)
1165 * NOTE: this function doesnt have to take the runqueue lock,
1166 * because all it wants to ensure is that the remote task enters
1167 * the kernel. If the IPI races and the task has been migrated
1168 * to another CPU then no harm is done and the purpose has been
1171 void kick_process(struct task_struct
*p
)
1177 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1178 smp_send_reschedule(cpu
);
1183 * Return a low guess at the load of a migration-source cpu weighted
1184 * according to the scheduling class and "nice" value.
1186 * We want to under-estimate the load of migration sources, to
1187 * balance conservatively.
1189 static unsigned long source_load(int cpu
, int type
)
1191 struct rq
*rq
= cpu_rq(cpu
);
1192 unsigned long total
= weighted_cpuload(cpu
);
1197 return min(rq
->cpu_load
[type
-1], total
);
1201 * Return a high guess at the load of a migration-target cpu weighted
1202 * according to the scheduling class and "nice" value.
1204 static unsigned long target_load(int cpu
, int type
)
1206 struct rq
*rq
= cpu_rq(cpu
);
1207 unsigned long total
= weighted_cpuload(cpu
);
1212 return max(rq
->cpu_load
[type
-1], total
);
1216 * Return the average load per task on the cpu's run queue
1218 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1220 struct rq
*rq
= cpu_rq(cpu
);
1221 unsigned long total
= weighted_cpuload(cpu
);
1222 unsigned long n
= rq
->nr_running
;
1224 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1228 * find_idlest_group finds and returns the least busy CPU group within the
1231 static struct sched_group
*
1232 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1234 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1235 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1236 int load_idx
= sd
->forkexec_idx
;
1237 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1240 unsigned long load
, avg_load
;
1244 /* Skip over this group if it has no CPUs allowed */
1245 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1248 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1250 /* Tally up the load of all CPUs in the group */
1253 for_each_cpu_mask(i
, group
->cpumask
) {
1254 /* Bias balancing toward cpus of our domain */
1256 load
= source_load(i
, load_idx
);
1258 load
= target_load(i
, load_idx
);
1263 /* Adjust by relative CPU power of the group */
1264 avg_load
= sg_div_cpu_power(group
,
1265 avg_load
* SCHED_LOAD_SCALE
);
1268 this_load
= avg_load
;
1270 } else if (avg_load
< min_load
) {
1271 min_load
= avg_load
;
1274 } while (group
= group
->next
, group
!= sd
->groups
);
1276 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1282 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1285 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1288 unsigned long load
, min_load
= ULONG_MAX
;
1292 /* Traverse only the allowed CPUs */
1293 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1295 for_each_cpu_mask(i
, tmp
) {
1296 load
= weighted_cpuload(i
);
1298 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1308 * sched_balance_self: balance the current task (running on cpu) in domains
1309 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1312 * Balance, ie. select the least loaded group.
1314 * Returns the target CPU number, or the same CPU if no balancing is needed.
1316 * preempt must be disabled.
1318 static int sched_balance_self(int cpu
, int flag
)
1320 struct task_struct
*t
= current
;
1321 struct sched_domain
*tmp
, *sd
= NULL
;
1323 for_each_domain(cpu
, tmp
) {
1325 * If power savings logic is enabled for a domain, stop there.
1327 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1329 if (tmp
->flags
& flag
)
1335 struct sched_group
*group
;
1336 int new_cpu
, weight
;
1338 if (!(sd
->flags
& flag
)) {
1344 group
= find_idlest_group(sd
, t
, cpu
);
1350 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1351 if (new_cpu
== -1 || new_cpu
== cpu
) {
1352 /* Now try balancing at a lower domain level of cpu */
1357 /* Now try balancing at a lower domain level of new_cpu */
1360 weight
= cpus_weight(span
);
1361 for_each_domain(cpu
, tmp
) {
1362 if (weight
<= cpus_weight(tmp
->span
))
1364 if (tmp
->flags
& flag
)
1367 /* while loop will break here if sd == NULL */
1373 #endif /* CONFIG_SMP */
1376 * wake_idle() will wake a task on an idle cpu if task->cpu is
1377 * not idle and an idle cpu is available. The span of cpus to
1378 * search starts with cpus closest then further out as needed,
1379 * so we always favor a closer, idle cpu.
1381 * Returns the CPU we should wake onto.
1383 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1384 static int wake_idle(int cpu
, struct task_struct
*p
)
1387 struct sched_domain
*sd
;
1391 * If it is idle, then it is the best cpu to run this task.
1393 * This cpu is also the best, if it has more than one task already.
1394 * Siblings must be also busy(in most cases) as they didn't already
1395 * pickup the extra load from this cpu and hence we need not check
1396 * sibling runqueue info. This will avoid the checks and cache miss
1397 * penalities associated with that.
1399 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1402 for_each_domain(cpu
, sd
) {
1403 if (sd
->flags
& SD_WAKE_IDLE
) {
1404 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1405 for_each_cpu_mask(i
, tmp
) {
1416 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1423 * try_to_wake_up - wake up a thread
1424 * @p: the to-be-woken-up thread
1425 * @state: the mask of task states that can be woken
1426 * @sync: do a synchronous wakeup?
1428 * Put it on the run-queue if it's not already there. The "current"
1429 * thread is always on the run-queue (except when the actual
1430 * re-schedule is in progress), and as such you're allowed to do
1431 * the simpler "current->state = TASK_RUNNING" to mark yourself
1432 * runnable without the overhead of this.
1434 * returns failure only if the task is already active.
1436 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1438 int cpu
, this_cpu
, success
= 0;
1439 unsigned long flags
;
1443 struct sched_domain
*sd
, *this_sd
= NULL
;
1444 unsigned long load
, this_load
;
1448 rq
= task_rq_lock(p
, &flags
);
1449 old_state
= p
->state
;
1450 if (!(old_state
& state
))
1457 this_cpu
= smp_processor_id();
1460 if (unlikely(task_running(rq
, p
)))
1465 schedstat_inc(rq
, ttwu_count
);
1466 if (cpu
== this_cpu
) {
1467 schedstat_inc(rq
, ttwu_local
);
1471 for_each_domain(this_cpu
, sd
) {
1472 if (cpu_isset(cpu
, sd
->span
)) {
1473 schedstat_inc(sd
, ttwu_wake_remote
);
1479 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1483 * Check for affine wakeup and passive balancing possibilities.
1486 int idx
= this_sd
->wake_idx
;
1487 unsigned int imbalance
;
1489 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1491 load
= source_load(cpu
, idx
);
1492 this_load
= target_load(this_cpu
, idx
);
1494 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1496 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1497 unsigned long tl
= this_load
;
1498 unsigned long tl_per_task
;
1500 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1503 * If sync wakeup then subtract the (maximum possible)
1504 * effect of the currently running task from the load
1505 * of the current CPU:
1508 tl
-= current
->se
.load
.weight
;
1511 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1512 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1514 * This domain has SD_WAKE_AFFINE and
1515 * p is cache cold in this domain, and
1516 * there is no bad imbalance.
1518 schedstat_inc(this_sd
, ttwu_move_affine
);
1524 * Start passive balancing when half the imbalance_pct
1527 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1528 if (imbalance
*this_load
<= 100*load
) {
1529 schedstat_inc(this_sd
, ttwu_move_balance
);
1535 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1537 new_cpu
= wake_idle(new_cpu
, p
);
1538 if (new_cpu
!= cpu
) {
1539 set_task_cpu(p
, new_cpu
);
1540 task_rq_unlock(rq
, &flags
);
1541 /* might preempt at this point */
1542 rq
= task_rq_lock(p
, &flags
);
1543 old_state
= p
->state
;
1544 if (!(old_state
& state
))
1549 this_cpu
= smp_processor_id();
1554 #endif /* CONFIG_SMP */
1555 update_rq_clock(rq
);
1556 activate_task(rq
, p
, 1);
1558 * Sync wakeups (i.e. those types of wakeups where the waker
1559 * has indicated that it will leave the CPU in short order)
1560 * don't trigger a preemption, if the woken up task will run on
1561 * this cpu. (in this case the 'I will reschedule' promise of
1562 * the waker guarantees that the freshly woken up task is going
1563 * to be considered on this CPU.)
1565 if (!sync
|| cpu
!= this_cpu
)
1566 check_preempt_curr(rq
, p
);
1570 p
->state
= TASK_RUNNING
;
1572 task_rq_unlock(rq
, &flags
);
1577 int fastcall
wake_up_process(struct task_struct
*p
)
1579 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1580 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1582 EXPORT_SYMBOL(wake_up_process
);
1584 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1586 return try_to_wake_up(p
, state
, 0);
1590 * Perform scheduler related setup for a newly forked process p.
1591 * p is forked by current.
1593 * __sched_fork() is basic setup used by init_idle() too:
1595 static void __sched_fork(struct task_struct
*p
)
1597 p
->se
.exec_start
= 0;
1598 p
->se
.sum_exec_runtime
= 0;
1599 p
->se
.prev_sum_exec_runtime
= 0;
1601 #ifdef CONFIG_SCHEDSTATS
1602 p
->se
.wait_start
= 0;
1603 p
->se
.sum_sleep_runtime
= 0;
1604 p
->se
.sleep_start
= 0;
1605 p
->se
.block_start
= 0;
1606 p
->se
.sleep_max
= 0;
1607 p
->se
.block_max
= 0;
1609 p
->se
.slice_max
= 0;
1613 INIT_LIST_HEAD(&p
->run_list
);
1616 #ifdef CONFIG_PREEMPT_NOTIFIERS
1617 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1621 * We mark the process as running here, but have not actually
1622 * inserted it onto the runqueue yet. This guarantees that
1623 * nobody will actually run it, and a signal or other external
1624 * event cannot wake it up and insert it on the runqueue either.
1626 p
->state
= TASK_RUNNING
;
1630 * fork()/clone()-time setup:
1632 void sched_fork(struct task_struct
*p
, int clone_flags
)
1634 int cpu
= get_cpu();
1639 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1641 set_task_cpu(p
, cpu
);
1644 * Make sure we do not leak PI boosting priority to the child:
1646 p
->prio
= current
->normal_prio
;
1647 if (!rt_prio(p
->prio
))
1648 p
->sched_class
= &fair_sched_class
;
1650 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1651 if (likely(sched_info_on()))
1652 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1654 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1657 #ifdef CONFIG_PREEMPT
1658 /* Want to start with kernel preemption disabled. */
1659 task_thread_info(p
)->preempt_count
= 1;
1665 * wake_up_new_task - wake up a newly created task for the first time.
1667 * This function will do some initial scheduler statistics housekeeping
1668 * that must be done for every newly created context, then puts the task
1669 * on the runqueue and wakes it.
1671 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1673 unsigned long flags
;
1676 rq
= task_rq_lock(p
, &flags
);
1677 BUG_ON(p
->state
!= TASK_RUNNING
);
1678 update_rq_clock(rq
);
1680 p
->prio
= effective_prio(p
);
1682 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
|| !rq
->cfs
.curr
) {
1683 activate_task(rq
, p
, 0);
1686 * Let the scheduling class do new task startup
1687 * management (if any):
1689 p
->sched_class
->task_new(rq
, p
);
1690 inc_nr_running(p
, rq
);
1692 check_preempt_curr(rq
, p
);
1693 task_rq_unlock(rq
, &flags
);
1696 #ifdef CONFIG_PREEMPT_NOTIFIERS
1699 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1700 * @notifier: notifier struct to register
1702 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1704 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1706 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1709 * preempt_notifier_unregister - no longer interested in preemption notifications
1710 * @notifier: notifier struct to unregister
1712 * This is safe to call from within a preemption notifier.
1714 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1716 hlist_del(¬ifier
->link
);
1718 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1720 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1722 struct preempt_notifier
*notifier
;
1723 struct hlist_node
*node
;
1725 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1726 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1730 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1731 struct task_struct
*next
)
1733 struct preempt_notifier
*notifier
;
1734 struct hlist_node
*node
;
1736 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1737 notifier
->ops
->sched_out(notifier
, next
);
1742 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1747 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1748 struct task_struct
*next
)
1755 * prepare_task_switch - prepare to switch tasks
1756 * @rq: the runqueue preparing to switch
1757 * @prev: the current task that is being switched out
1758 * @next: the task we are going to switch to.
1760 * This is called with the rq lock held and interrupts off. It must
1761 * be paired with a subsequent finish_task_switch after the context
1764 * prepare_task_switch sets up locking and calls architecture specific
1768 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1769 struct task_struct
*next
)
1771 fire_sched_out_preempt_notifiers(prev
, next
);
1772 prepare_lock_switch(rq
, next
);
1773 prepare_arch_switch(next
);
1777 * finish_task_switch - clean up after a task-switch
1778 * @rq: runqueue associated with task-switch
1779 * @prev: the thread we just switched away from.
1781 * finish_task_switch must be called after the context switch, paired
1782 * with a prepare_task_switch call before the context switch.
1783 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1784 * and do any other architecture-specific cleanup actions.
1786 * Note that we may have delayed dropping an mm in context_switch(). If
1787 * so, we finish that here outside of the runqueue lock. (Doing it
1788 * with the lock held can cause deadlocks; see schedule() for
1791 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1792 __releases(rq
->lock
)
1794 struct mm_struct
*mm
= rq
->prev_mm
;
1800 * A task struct has one reference for the use as "current".
1801 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1802 * schedule one last time. The schedule call will never return, and
1803 * the scheduled task must drop that reference.
1804 * The test for TASK_DEAD must occur while the runqueue locks are
1805 * still held, otherwise prev could be scheduled on another cpu, die
1806 * there before we look at prev->state, and then the reference would
1808 * Manfred Spraul <manfred@colorfullife.com>
1810 prev_state
= prev
->state
;
1811 finish_arch_switch(prev
);
1812 finish_lock_switch(rq
, prev
);
1813 fire_sched_in_preempt_notifiers(current
);
1816 if (unlikely(prev_state
== TASK_DEAD
)) {
1818 * Remove function-return probe instances associated with this
1819 * task and put them back on the free list.
1821 kprobe_flush_task(prev
);
1822 put_task_struct(prev
);
1827 * schedule_tail - first thing a freshly forked thread must call.
1828 * @prev: the thread we just switched away from.
1830 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1831 __releases(rq
->lock
)
1833 struct rq
*rq
= this_rq();
1835 finish_task_switch(rq
, prev
);
1836 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1837 /* In this case, finish_task_switch does not reenable preemption */
1840 if (current
->set_child_tid
)
1841 put_user(current
->pid
, current
->set_child_tid
);
1845 * context_switch - switch to the new MM and the new
1846 * thread's register state.
1849 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1850 struct task_struct
*next
)
1852 struct mm_struct
*mm
, *oldmm
;
1854 prepare_task_switch(rq
, prev
, next
);
1856 oldmm
= prev
->active_mm
;
1858 * For paravirt, this is coupled with an exit in switch_to to
1859 * combine the page table reload and the switch backend into
1862 arch_enter_lazy_cpu_mode();
1864 if (unlikely(!mm
)) {
1865 next
->active_mm
= oldmm
;
1866 atomic_inc(&oldmm
->mm_count
);
1867 enter_lazy_tlb(oldmm
, next
);
1869 switch_mm(oldmm
, mm
, next
);
1871 if (unlikely(!prev
->mm
)) {
1872 prev
->active_mm
= NULL
;
1873 rq
->prev_mm
= oldmm
;
1876 * Since the runqueue lock will be released by the next
1877 * task (which is an invalid locking op but in the case
1878 * of the scheduler it's an obvious special-case), so we
1879 * do an early lockdep release here:
1881 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1882 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1885 /* Here we just switch the register state and the stack. */
1886 switch_to(prev
, next
, prev
);
1890 * this_rq must be evaluated again because prev may have moved
1891 * CPUs since it called schedule(), thus the 'rq' on its stack
1892 * frame will be invalid.
1894 finish_task_switch(this_rq(), prev
);
1898 * nr_running, nr_uninterruptible and nr_context_switches:
1900 * externally visible scheduler statistics: current number of runnable
1901 * threads, current number of uninterruptible-sleeping threads, total
1902 * number of context switches performed since bootup.
1904 unsigned long nr_running(void)
1906 unsigned long i
, sum
= 0;
1908 for_each_online_cpu(i
)
1909 sum
+= cpu_rq(i
)->nr_running
;
1914 unsigned long nr_uninterruptible(void)
1916 unsigned long i
, sum
= 0;
1918 for_each_possible_cpu(i
)
1919 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1922 * Since we read the counters lockless, it might be slightly
1923 * inaccurate. Do not allow it to go below zero though:
1925 if (unlikely((long)sum
< 0))
1931 unsigned long long nr_context_switches(void)
1934 unsigned long long sum
= 0;
1936 for_each_possible_cpu(i
)
1937 sum
+= cpu_rq(i
)->nr_switches
;
1942 unsigned long nr_iowait(void)
1944 unsigned long i
, sum
= 0;
1946 for_each_possible_cpu(i
)
1947 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1952 unsigned long nr_active(void)
1954 unsigned long i
, running
= 0, uninterruptible
= 0;
1956 for_each_online_cpu(i
) {
1957 running
+= cpu_rq(i
)->nr_running
;
1958 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1961 if (unlikely((long)uninterruptible
< 0))
1962 uninterruptible
= 0;
1964 return running
+ uninterruptible
;
1968 * Update rq->cpu_load[] statistics. This function is usually called every
1969 * scheduler tick (TICK_NSEC).
1971 static void update_cpu_load(struct rq
*this_rq
)
1973 unsigned long this_load
= this_rq
->load
.weight
;
1976 this_rq
->nr_load_updates
++;
1978 /* Update our load: */
1979 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1980 unsigned long old_load
, new_load
;
1982 /* scale is effectively 1 << i now, and >> i divides by scale */
1984 old_load
= this_rq
->cpu_load
[i
];
1985 new_load
= this_load
;
1987 * Round up the averaging division if load is increasing. This
1988 * prevents us from getting stuck on 9 if the load is 10, for
1991 if (new_load
> old_load
)
1992 new_load
+= scale
-1;
1993 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2000 * double_rq_lock - safely lock two runqueues
2002 * Note this does not disable interrupts like task_rq_lock,
2003 * you need to do so manually before calling.
2005 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2006 __acquires(rq1
->lock
)
2007 __acquires(rq2
->lock
)
2009 BUG_ON(!irqs_disabled());
2011 spin_lock(&rq1
->lock
);
2012 __acquire(rq2
->lock
); /* Fake it out ;) */
2015 spin_lock(&rq1
->lock
);
2016 spin_lock(&rq2
->lock
);
2018 spin_lock(&rq2
->lock
);
2019 spin_lock(&rq1
->lock
);
2022 update_rq_clock(rq1
);
2023 update_rq_clock(rq2
);
2027 * double_rq_unlock - safely unlock two runqueues
2029 * Note this does not restore interrupts like task_rq_unlock,
2030 * you need to do so manually after calling.
2032 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2033 __releases(rq1
->lock
)
2034 __releases(rq2
->lock
)
2036 spin_unlock(&rq1
->lock
);
2038 spin_unlock(&rq2
->lock
);
2040 __release(rq2
->lock
);
2044 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2046 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2047 __releases(this_rq
->lock
)
2048 __acquires(busiest
->lock
)
2049 __acquires(this_rq
->lock
)
2051 if (unlikely(!irqs_disabled())) {
2052 /* printk() doesn't work good under rq->lock */
2053 spin_unlock(&this_rq
->lock
);
2056 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2057 if (busiest
< this_rq
) {
2058 spin_unlock(&this_rq
->lock
);
2059 spin_lock(&busiest
->lock
);
2060 spin_lock(&this_rq
->lock
);
2062 spin_lock(&busiest
->lock
);
2067 * If dest_cpu is allowed for this process, migrate the task to it.
2068 * This is accomplished by forcing the cpu_allowed mask to only
2069 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2070 * the cpu_allowed mask is restored.
2072 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2074 struct migration_req req
;
2075 unsigned long flags
;
2078 rq
= task_rq_lock(p
, &flags
);
2079 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2080 || unlikely(cpu_is_offline(dest_cpu
)))
2083 /* force the process onto the specified CPU */
2084 if (migrate_task(p
, dest_cpu
, &req
)) {
2085 /* Need to wait for migration thread (might exit: take ref). */
2086 struct task_struct
*mt
= rq
->migration_thread
;
2088 get_task_struct(mt
);
2089 task_rq_unlock(rq
, &flags
);
2090 wake_up_process(mt
);
2091 put_task_struct(mt
);
2092 wait_for_completion(&req
.done
);
2097 task_rq_unlock(rq
, &flags
);
2101 * sched_exec - execve() is a valuable balancing opportunity, because at
2102 * this point the task has the smallest effective memory and cache footprint.
2104 void sched_exec(void)
2106 int new_cpu
, this_cpu
= get_cpu();
2107 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2109 if (new_cpu
!= this_cpu
)
2110 sched_migrate_task(current
, new_cpu
);
2114 * pull_task - move a task from a remote runqueue to the local runqueue.
2115 * Both runqueues must be locked.
2117 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2118 struct rq
*this_rq
, int this_cpu
)
2120 deactivate_task(src_rq
, p
, 0);
2121 set_task_cpu(p
, this_cpu
);
2122 activate_task(this_rq
, p
, 0);
2124 * Note that idle threads have a prio of MAX_PRIO, for this test
2125 * to be always true for them.
2127 check_preempt_curr(this_rq
, p
);
2131 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2134 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2135 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2139 * We do not migrate tasks that are:
2140 * 1) running (obviously), or
2141 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2142 * 3) are cache-hot on their current CPU.
2144 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2148 if (task_running(rq
, p
))
2154 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2155 unsigned long max_nr_move
, unsigned long max_load_move
,
2156 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2157 int *all_pinned
, unsigned long *load_moved
,
2158 int *this_best_prio
, struct rq_iterator
*iterator
)
2160 int pulled
= 0, pinned
= 0, skip_for_load
;
2161 struct task_struct
*p
;
2162 long rem_load_move
= max_load_move
;
2164 if (max_nr_move
== 0 || max_load_move
== 0)
2170 * Start the load-balancing iterator:
2172 p
= iterator
->start(iterator
->arg
);
2177 * To help distribute high priority tasks accross CPUs we don't
2178 * skip a task if it will be the highest priority task (i.e. smallest
2179 * prio value) on its new queue regardless of its load weight
2181 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2182 SCHED_LOAD_SCALE_FUZZ
;
2183 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2184 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2185 p
= iterator
->next(iterator
->arg
);
2189 pull_task(busiest
, p
, this_rq
, this_cpu
);
2191 rem_load_move
-= p
->se
.load
.weight
;
2194 * We only want to steal up to the prescribed number of tasks
2195 * and the prescribed amount of weighted load.
2197 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2198 if (p
->prio
< *this_best_prio
)
2199 *this_best_prio
= p
->prio
;
2200 p
= iterator
->next(iterator
->arg
);
2205 * Right now, this is the only place pull_task() is called,
2206 * so we can safely collect pull_task() stats here rather than
2207 * inside pull_task().
2209 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2212 *all_pinned
= pinned
;
2213 *load_moved
= max_load_move
- rem_load_move
;
2218 * move_tasks tries to move up to max_load_move weighted load from busiest to
2219 * this_rq, as part of a balancing operation within domain "sd".
2220 * Returns 1 if successful and 0 otherwise.
2222 * Called with both runqueues locked.
2224 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2225 unsigned long max_load_move
,
2226 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2229 const struct sched_class
*class = sched_class_highest
;
2230 unsigned long total_load_moved
= 0;
2231 int this_best_prio
= this_rq
->curr
->prio
;
2235 class->load_balance(this_rq
, this_cpu
, busiest
,
2236 ULONG_MAX
, max_load_move
- total_load_moved
,
2237 sd
, idle
, all_pinned
, &this_best_prio
);
2238 class = class->next
;
2239 } while (class && max_load_move
> total_load_moved
);
2241 return total_load_moved
> 0;
2245 * move_one_task tries to move exactly one task from busiest to this_rq, as
2246 * part of active balancing operations within "domain".
2247 * Returns 1 if successful and 0 otherwise.
2249 * Called with both runqueues locked.
2251 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2252 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2254 const struct sched_class
*class;
2255 int this_best_prio
= MAX_PRIO
;
2257 for (class = sched_class_highest
; class; class = class->next
)
2258 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2259 1, ULONG_MAX
, sd
, idle
, NULL
,
2267 * find_busiest_group finds and returns the busiest CPU group within the
2268 * domain. It calculates and returns the amount of weighted load which
2269 * should be moved to restore balance via the imbalance parameter.
2271 static struct sched_group
*
2272 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2273 unsigned long *imbalance
, enum cpu_idle_type idle
,
2274 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2276 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2277 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2278 unsigned long max_pull
;
2279 unsigned long busiest_load_per_task
, busiest_nr_running
;
2280 unsigned long this_load_per_task
, this_nr_running
;
2282 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2283 int power_savings_balance
= 1;
2284 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2285 unsigned long min_nr_running
= ULONG_MAX
;
2286 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2289 max_load
= this_load
= total_load
= total_pwr
= 0;
2290 busiest_load_per_task
= busiest_nr_running
= 0;
2291 this_load_per_task
= this_nr_running
= 0;
2292 if (idle
== CPU_NOT_IDLE
)
2293 load_idx
= sd
->busy_idx
;
2294 else if (idle
== CPU_NEWLY_IDLE
)
2295 load_idx
= sd
->newidle_idx
;
2297 load_idx
= sd
->idle_idx
;
2300 unsigned long load
, group_capacity
;
2303 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2304 unsigned long sum_nr_running
, sum_weighted_load
;
2306 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2309 balance_cpu
= first_cpu(group
->cpumask
);
2311 /* Tally up the load of all CPUs in the group */
2312 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2314 for_each_cpu_mask(i
, group
->cpumask
) {
2317 if (!cpu_isset(i
, *cpus
))
2322 if (*sd_idle
&& rq
->nr_running
)
2325 /* Bias balancing toward cpus of our domain */
2327 if (idle_cpu(i
) && !first_idle_cpu
) {
2332 load
= target_load(i
, load_idx
);
2334 load
= source_load(i
, load_idx
);
2337 sum_nr_running
+= rq
->nr_running
;
2338 sum_weighted_load
+= weighted_cpuload(i
);
2342 * First idle cpu or the first cpu(busiest) in this sched group
2343 * is eligible for doing load balancing at this and above
2344 * domains. In the newly idle case, we will allow all the cpu's
2345 * to do the newly idle load balance.
2347 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2348 balance_cpu
!= this_cpu
&& balance
) {
2353 total_load
+= avg_load
;
2354 total_pwr
+= group
->__cpu_power
;
2356 /* Adjust by relative CPU power of the group */
2357 avg_load
= sg_div_cpu_power(group
,
2358 avg_load
* SCHED_LOAD_SCALE
);
2360 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2363 this_load
= avg_load
;
2365 this_nr_running
= sum_nr_running
;
2366 this_load_per_task
= sum_weighted_load
;
2367 } else if (avg_load
> max_load
&&
2368 sum_nr_running
> group_capacity
) {
2369 max_load
= avg_load
;
2371 busiest_nr_running
= sum_nr_running
;
2372 busiest_load_per_task
= sum_weighted_load
;
2375 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2377 * Busy processors will not participate in power savings
2380 if (idle
== CPU_NOT_IDLE
||
2381 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2385 * If the local group is idle or completely loaded
2386 * no need to do power savings balance at this domain
2388 if (local_group
&& (this_nr_running
>= group_capacity
||
2390 power_savings_balance
= 0;
2393 * If a group is already running at full capacity or idle,
2394 * don't include that group in power savings calculations
2396 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2401 * Calculate the group which has the least non-idle load.
2402 * This is the group from where we need to pick up the load
2405 if ((sum_nr_running
< min_nr_running
) ||
2406 (sum_nr_running
== min_nr_running
&&
2407 first_cpu(group
->cpumask
) <
2408 first_cpu(group_min
->cpumask
))) {
2410 min_nr_running
= sum_nr_running
;
2411 min_load_per_task
= sum_weighted_load
/
2416 * Calculate the group which is almost near its
2417 * capacity but still has some space to pick up some load
2418 * from other group and save more power
2420 if (sum_nr_running
<= group_capacity
- 1) {
2421 if (sum_nr_running
> leader_nr_running
||
2422 (sum_nr_running
== leader_nr_running
&&
2423 first_cpu(group
->cpumask
) >
2424 first_cpu(group_leader
->cpumask
))) {
2425 group_leader
= group
;
2426 leader_nr_running
= sum_nr_running
;
2431 group
= group
->next
;
2432 } while (group
!= sd
->groups
);
2434 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2437 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2439 if (this_load
>= avg_load
||
2440 100*max_load
<= sd
->imbalance_pct
*this_load
)
2443 busiest_load_per_task
/= busiest_nr_running
;
2445 * We're trying to get all the cpus to the average_load, so we don't
2446 * want to push ourselves above the average load, nor do we wish to
2447 * reduce the max loaded cpu below the average load, as either of these
2448 * actions would just result in more rebalancing later, and ping-pong
2449 * tasks around. Thus we look for the minimum possible imbalance.
2450 * Negative imbalances (*we* are more loaded than anyone else) will
2451 * be counted as no imbalance for these purposes -- we can't fix that
2452 * by pulling tasks to us. Be careful of negative numbers as they'll
2453 * appear as very large values with unsigned longs.
2455 if (max_load
<= busiest_load_per_task
)
2459 * In the presence of smp nice balancing, certain scenarios can have
2460 * max load less than avg load(as we skip the groups at or below
2461 * its cpu_power, while calculating max_load..)
2463 if (max_load
< avg_load
) {
2465 goto small_imbalance
;
2468 /* Don't want to pull so many tasks that a group would go idle */
2469 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2471 /* How much load to actually move to equalise the imbalance */
2472 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2473 (avg_load
- this_load
) * this->__cpu_power
)
2477 * if *imbalance is less than the average load per runnable task
2478 * there is no gaurantee that any tasks will be moved so we'll have
2479 * a think about bumping its value to force at least one task to be
2482 if (*imbalance
< busiest_load_per_task
) {
2483 unsigned long tmp
, pwr_now
, pwr_move
;
2487 pwr_move
= pwr_now
= 0;
2489 if (this_nr_running
) {
2490 this_load_per_task
/= this_nr_running
;
2491 if (busiest_load_per_task
> this_load_per_task
)
2494 this_load_per_task
= SCHED_LOAD_SCALE
;
2496 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2497 busiest_load_per_task
* imbn
) {
2498 *imbalance
= busiest_load_per_task
;
2503 * OK, we don't have enough imbalance to justify moving tasks,
2504 * however we may be able to increase total CPU power used by
2508 pwr_now
+= busiest
->__cpu_power
*
2509 min(busiest_load_per_task
, max_load
);
2510 pwr_now
+= this->__cpu_power
*
2511 min(this_load_per_task
, this_load
);
2512 pwr_now
/= SCHED_LOAD_SCALE
;
2514 /* Amount of load we'd subtract */
2515 tmp
= sg_div_cpu_power(busiest
,
2516 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2518 pwr_move
+= busiest
->__cpu_power
*
2519 min(busiest_load_per_task
, max_load
- tmp
);
2521 /* Amount of load we'd add */
2522 if (max_load
* busiest
->__cpu_power
<
2523 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2524 tmp
= sg_div_cpu_power(this,
2525 max_load
* busiest
->__cpu_power
);
2527 tmp
= sg_div_cpu_power(this,
2528 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2529 pwr_move
+= this->__cpu_power
*
2530 min(this_load_per_task
, this_load
+ tmp
);
2531 pwr_move
/= SCHED_LOAD_SCALE
;
2533 /* Move if we gain throughput */
2534 if (pwr_move
> pwr_now
)
2535 *imbalance
= busiest_load_per_task
;
2541 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2542 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2545 if (this == group_leader
&& group_leader
!= group_min
) {
2546 *imbalance
= min_load_per_task
;
2556 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2559 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2560 unsigned long imbalance
, cpumask_t
*cpus
)
2562 struct rq
*busiest
= NULL
, *rq
;
2563 unsigned long max_load
= 0;
2566 for_each_cpu_mask(i
, group
->cpumask
) {
2569 if (!cpu_isset(i
, *cpus
))
2573 wl
= weighted_cpuload(i
);
2575 if (rq
->nr_running
== 1 && wl
> imbalance
)
2578 if (wl
> max_load
) {
2588 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2589 * so long as it is large enough.
2591 #define MAX_PINNED_INTERVAL 512
2594 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2595 * tasks if there is an imbalance.
2597 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2598 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2601 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2602 struct sched_group
*group
;
2603 unsigned long imbalance
;
2605 cpumask_t cpus
= CPU_MASK_ALL
;
2606 unsigned long flags
;
2609 * When power savings policy is enabled for the parent domain, idle
2610 * sibling can pick up load irrespective of busy siblings. In this case,
2611 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2612 * portraying it as CPU_NOT_IDLE.
2614 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2615 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2618 schedstat_inc(sd
, lb_count
[idle
]);
2621 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2628 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2632 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2634 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2638 BUG_ON(busiest
== this_rq
);
2640 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2643 if (busiest
->nr_running
> 1) {
2645 * Attempt to move tasks. If find_busiest_group has found
2646 * an imbalance but busiest->nr_running <= 1, the group is
2647 * still unbalanced. ld_moved simply stays zero, so it is
2648 * correctly treated as an imbalance.
2650 local_irq_save(flags
);
2651 double_rq_lock(this_rq
, busiest
);
2652 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2653 imbalance
, sd
, idle
, &all_pinned
);
2654 double_rq_unlock(this_rq
, busiest
);
2655 local_irq_restore(flags
);
2658 * some other cpu did the load balance for us.
2660 if (ld_moved
&& this_cpu
!= smp_processor_id())
2661 resched_cpu(this_cpu
);
2663 /* All tasks on this runqueue were pinned by CPU affinity */
2664 if (unlikely(all_pinned
)) {
2665 cpu_clear(cpu_of(busiest
), cpus
);
2666 if (!cpus_empty(cpus
))
2673 schedstat_inc(sd
, lb_failed
[idle
]);
2674 sd
->nr_balance_failed
++;
2676 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2678 spin_lock_irqsave(&busiest
->lock
, flags
);
2680 /* don't kick the migration_thread, if the curr
2681 * task on busiest cpu can't be moved to this_cpu
2683 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2684 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2686 goto out_one_pinned
;
2689 if (!busiest
->active_balance
) {
2690 busiest
->active_balance
= 1;
2691 busiest
->push_cpu
= this_cpu
;
2694 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2696 wake_up_process(busiest
->migration_thread
);
2699 * We've kicked active balancing, reset the failure
2702 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2705 sd
->nr_balance_failed
= 0;
2707 if (likely(!active_balance
)) {
2708 /* We were unbalanced, so reset the balancing interval */
2709 sd
->balance_interval
= sd
->min_interval
;
2712 * If we've begun active balancing, start to back off. This
2713 * case may not be covered by the all_pinned logic if there
2714 * is only 1 task on the busy runqueue (because we don't call
2717 if (sd
->balance_interval
< sd
->max_interval
)
2718 sd
->balance_interval
*= 2;
2721 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2722 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2727 schedstat_inc(sd
, lb_balanced
[idle
]);
2729 sd
->nr_balance_failed
= 0;
2732 /* tune up the balancing interval */
2733 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2734 (sd
->balance_interval
< sd
->max_interval
))
2735 sd
->balance_interval
*= 2;
2737 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2738 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2744 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2745 * tasks if there is an imbalance.
2747 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2748 * this_rq is locked.
2751 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2753 struct sched_group
*group
;
2754 struct rq
*busiest
= NULL
;
2755 unsigned long imbalance
;
2759 cpumask_t cpus
= CPU_MASK_ALL
;
2762 * When power savings policy is enabled for the parent domain, idle
2763 * sibling can pick up load irrespective of busy siblings. In this case,
2764 * let the state of idle sibling percolate up as IDLE, instead of
2765 * portraying it as CPU_NOT_IDLE.
2767 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2768 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2771 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2773 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2774 &sd_idle
, &cpus
, NULL
);
2776 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2780 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2783 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2787 BUG_ON(busiest
== this_rq
);
2789 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2792 if (busiest
->nr_running
> 1) {
2793 /* Attempt to move tasks */
2794 double_lock_balance(this_rq
, busiest
);
2795 /* this_rq->clock is already updated */
2796 update_rq_clock(busiest
);
2797 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2798 imbalance
, sd
, CPU_NEWLY_IDLE
,
2800 spin_unlock(&busiest
->lock
);
2802 if (unlikely(all_pinned
)) {
2803 cpu_clear(cpu_of(busiest
), cpus
);
2804 if (!cpus_empty(cpus
))
2810 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2811 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2812 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2815 sd
->nr_balance_failed
= 0;
2820 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2821 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2822 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2824 sd
->nr_balance_failed
= 0;
2830 * idle_balance is called by schedule() if this_cpu is about to become
2831 * idle. Attempts to pull tasks from other CPUs.
2833 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2835 struct sched_domain
*sd
;
2836 int pulled_task
= -1;
2837 unsigned long next_balance
= jiffies
+ HZ
;
2839 for_each_domain(this_cpu
, sd
) {
2840 unsigned long interval
;
2842 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2845 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2846 /* If we've pulled tasks over stop searching: */
2847 pulled_task
= load_balance_newidle(this_cpu
,
2850 interval
= msecs_to_jiffies(sd
->balance_interval
);
2851 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2852 next_balance
= sd
->last_balance
+ interval
;
2856 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2858 * We are going idle. next_balance may be set based on
2859 * a busy processor. So reset next_balance.
2861 this_rq
->next_balance
= next_balance
;
2866 * active_load_balance is run by migration threads. It pushes running tasks
2867 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2868 * running on each physical CPU where possible, and avoids physical /
2869 * logical imbalances.
2871 * Called with busiest_rq locked.
2873 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2875 int target_cpu
= busiest_rq
->push_cpu
;
2876 struct sched_domain
*sd
;
2877 struct rq
*target_rq
;
2879 /* Is there any task to move? */
2880 if (busiest_rq
->nr_running
<= 1)
2883 target_rq
= cpu_rq(target_cpu
);
2886 * This condition is "impossible", if it occurs
2887 * we need to fix it. Originally reported by
2888 * Bjorn Helgaas on a 128-cpu setup.
2890 BUG_ON(busiest_rq
== target_rq
);
2892 /* move a task from busiest_rq to target_rq */
2893 double_lock_balance(busiest_rq
, target_rq
);
2894 update_rq_clock(busiest_rq
);
2895 update_rq_clock(target_rq
);
2897 /* Search for an sd spanning us and the target CPU. */
2898 for_each_domain(target_cpu
, sd
) {
2899 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2900 cpu_isset(busiest_cpu
, sd
->span
))
2905 schedstat_inc(sd
, alb_count
);
2907 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2909 schedstat_inc(sd
, alb_pushed
);
2911 schedstat_inc(sd
, alb_failed
);
2913 spin_unlock(&target_rq
->lock
);
2918 atomic_t load_balancer
;
2920 } nohz ____cacheline_aligned
= {
2921 .load_balancer
= ATOMIC_INIT(-1),
2922 .cpu_mask
= CPU_MASK_NONE
,
2926 * This routine will try to nominate the ilb (idle load balancing)
2927 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2928 * load balancing on behalf of all those cpus. If all the cpus in the system
2929 * go into this tickless mode, then there will be no ilb owner (as there is
2930 * no need for one) and all the cpus will sleep till the next wakeup event
2933 * For the ilb owner, tick is not stopped. And this tick will be used
2934 * for idle load balancing. ilb owner will still be part of
2937 * While stopping the tick, this cpu will become the ilb owner if there
2938 * is no other owner. And will be the owner till that cpu becomes busy
2939 * or if all cpus in the system stop their ticks at which point
2940 * there is no need for ilb owner.
2942 * When the ilb owner becomes busy, it nominates another owner, during the
2943 * next busy scheduler_tick()
2945 int select_nohz_load_balancer(int stop_tick
)
2947 int cpu
= smp_processor_id();
2950 cpu_set(cpu
, nohz
.cpu_mask
);
2951 cpu_rq(cpu
)->in_nohz_recently
= 1;
2954 * If we are going offline and still the leader, give up!
2956 if (cpu_is_offline(cpu
) &&
2957 atomic_read(&nohz
.load_balancer
) == cpu
) {
2958 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2963 /* time for ilb owner also to sleep */
2964 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2965 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2966 atomic_set(&nohz
.load_balancer
, -1);
2970 if (atomic_read(&nohz
.load_balancer
) == -1) {
2971 /* make me the ilb owner */
2972 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2974 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2977 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2980 cpu_clear(cpu
, nohz
.cpu_mask
);
2982 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2983 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2990 static DEFINE_SPINLOCK(balancing
);
2993 * It checks each scheduling domain to see if it is due to be balanced,
2994 * and initiates a balancing operation if so.
2996 * Balancing parameters are set up in arch_init_sched_domains.
2998 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3001 struct rq
*rq
= cpu_rq(cpu
);
3002 unsigned long interval
;
3003 struct sched_domain
*sd
;
3004 /* Earliest time when we have to do rebalance again */
3005 unsigned long next_balance
= jiffies
+ 60*HZ
;
3006 int update_next_balance
= 0;
3008 for_each_domain(cpu
, sd
) {
3009 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3012 interval
= sd
->balance_interval
;
3013 if (idle
!= CPU_IDLE
)
3014 interval
*= sd
->busy_factor
;
3016 /* scale ms to jiffies */
3017 interval
= msecs_to_jiffies(interval
);
3018 if (unlikely(!interval
))
3020 if (interval
> HZ
*NR_CPUS
/10)
3021 interval
= HZ
*NR_CPUS
/10;
3024 if (sd
->flags
& SD_SERIALIZE
) {
3025 if (!spin_trylock(&balancing
))
3029 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3030 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3032 * We've pulled tasks over so either we're no
3033 * longer idle, or one of our SMT siblings is
3036 idle
= CPU_NOT_IDLE
;
3038 sd
->last_balance
= jiffies
;
3040 if (sd
->flags
& SD_SERIALIZE
)
3041 spin_unlock(&balancing
);
3043 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3044 next_balance
= sd
->last_balance
+ interval
;
3045 update_next_balance
= 1;
3049 * Stop the load balance at this level. There is another
3050 * CPU in our sched group which is doing load balancing more
3058 * next_balance will be updated only when there is a need.
3059 * When the cpu is attached to null domain for ex, it will not be
3062 if (likely(update_next_balance
))
3063 rq
->next_balance
= next_balance
;
3067 * run_rebalance_domains is triggered when needed from the scheduler tick.
3068 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3069 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3071 static void run_rebalance_domains(struct softirq_action
*h
)
3073 int this_cpu
= smp_processor_id();
3074 struct rq
*this_rq
= cpu_rq(this_cpu
);
3075 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3076 CPU_IDLE
: CPU_NOT_IDLE
;
3078 rebalance_domains(this_cpu
, idle
);
3082 * If this cpu is the owner for idle load balancing, then do the
3083 * balancing on behalf of the other idle cpus whose ticks are
3086 if (this_rq
->idle_at_tick
&&
3087 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3088 cpumask_t cpus
= nohz
.cpu_mask
;
3092 cpu_clear(this_cpu
, cpus
);
3093 for_each_cpu_mask(balance_cpu
, cpus
) {
3095 * If this cpu gets work to do, stop the load balancing
3096 * work being done for other cpus. Next load
3097 * balancing owner will pick it up.
3102 rebalance_domains(balance_cpu
, CPU_IDLE
);
3104 rq
= cpu_rq(balance_cpu
);
3105 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3106 this_rq
->next_balance
= rq
->next_balance
;
3113 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3115 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3116 * idle load balancing owner or decide to stop the periodic load balancing,
3117 * if the whole system is idle.
3119 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3123 * If we were in the nohz mode recently and busy at the current
3124 * scheduler tick, then check if we need to nominate new idle
3127 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3128 rq
->in_nohz_recently
= 0;
3130 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3131 cpu_clear(cpu
, nohz
.cpu_mask
);
3132 atomic_set(&nohz
.load_balancer
, -1);
3135 if (atomic_read(&nohz
.load_balancer
) == -1) {
3137 * simple selection for now: Nominate the
3138 * first cpu in the nohz list to be the next
3141 * TBD: Traverse the sched domains and nominate
3142 * the nearest cpu in the nohz.cpu_mask.
3144 int ilb
= first_cpu(nohz
.cpu_mask
);
3152 * If this cpu is idle and doing idle load balancing for all the
3153 * cpus with ticks stopped, is it time for that to stop?
3155 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3156 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3162 * If this cpu is idle and the idle load balancing is done by
3163 * someone else, then no need raise the SCHED_SOFTIRQ
3165 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3166 cpu_isset(cpu
, nohz
.cpu_mask
))
3169 if (time_after_eq(jiffies
, rq
->next_balance
))
3170 raise_softirq(SCHED_SOFTIRQ
);
3173 #else /* CONFIG_SMP */
3176 * on UP we do not need to balance between CPUs:
3178 static inline void idle_balance(int cpu
, struct rq
*rq
)
3182 /* Avoid "used but not defined" warning on UP */
3183 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3184 unsigned long max_nr_move
, unsigned long max_load_move
,
3185 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3186 int *all_pinned
, unsigned long *load_moved
,
3187 int *this_best_prio
, struct rq_iterator
*iterator
)
3196 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3198 EXPORT_PER_CPU_SYMBOL(kstat
);
3201 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3202 * that have not yet been banked in case the task is currently running.
3204 unsigned long long task_sched_runtime(struct task_struct
*p
)
3206 unsigned long flags
;
3210 rq
= task_rq_lock(p
, &flags
);
3211 ns
= p
->se
.sum_exec_runtime
;
3212 if (rq
->curr
== p
) {
3213 update_rq_clock(rq
);
3214 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3215 if ((s64
)delta_exec
> 0)
3218 task_rq_unlock(rq
, &flags
);
3224 * Account user cpu time to a process.
3225 * @p: the process that the cpu time gets accounted to
3226 * @hardirq_offset: the offset to subtract from hardirq_count()
3227 * @cputime: the cpu time spent in user space since the last update
3229 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3231 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3234 p
->utime
= cputime_add(p
->utime
, cputime
);
3236 /* Add user time to cpustat. */
3237 tmp
= cputime_to_cputime64(cputime
);
3238 if (TASK_NICE(p
) > 0)
3239 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3241 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3245 * Account system cpu time to a process.
3246 * @p: the process that the cpu time gets accounted to
3247 * @hardirq_offset: the offset to subtract from hardirq_count()
3248 * @cputime: the cpu time spent in kernel space since the last update
3250 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3253 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3254 struct rq
*rq
= this_rq();
3257 p
->stime
= cputime_add(p
->stime
, cputime
);
3259 /* Add system time to cpustat. */
3260 tmp
= cputime_to_cputime64(cputime
);
3261 if (hardirq_count() - hardirq_offset
)
3262 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3263 else if (softirq_count())
3264 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3265 else if (p
!= rq
->idle
)
3266 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3267 else if (atomic_read(&rq
->nr_iowait
) > 0)
3268 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3270 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3271 /* Account for system time used */
3272 acct_update_integrals(p
);
3276 * Account for involuntary wait time.
3277 * @p: the process from which the cpu time has been stolen
3278 * @steal: the cpu time spent in involuntary wait
3280 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3282 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3283 cputime64_t tmp
= cputime_to_cputime64(steal
);
3284 struct rq
*rq
= this_rq();
3286 if (p
== rq
->idle
) {
3287 p
->stime
= cputime_add(p
->stime
, steal
);
3288 if (atomic_read(&rq
->nr_iowait
) > 0)
3289 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3291 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3293 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3297 * This function gets called by the timer code, with HZ frequency.
3298 * We call it with interrupts disabled.
3300 * It also gets called by the fork code, when changing the parent's
3303 void scheduler_tick(void)
3305 int cpu
= smp_processor_id();
3306 struct rq
*rq
= cpu_rq(cpu
);
3307 struct task_struct
*curr
= rq
->curr
;
3308 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3310 spin_lock(&rq
->lock
);
3311 __update_rq_clock(rq
);
3313 * Let rq->clock advance by at least TICK_NSEC:
3315 if (unlikely(rq
->clock
< next_tick
))
3316 rq
->clock
= next_tick
;
3317 rq
->tick_timestamp
= rq
->clock
;
3318 update_cpu_load(rq
);
3319 if (curr
!= rq
->idle
) /* FIXME: needed? */
3320 curr
->sched_class
->task_tick(rq
, curr
);
3321 spin_unlock(&rq
->lock
);
3324 rq
->idle_at_tick
= idle_cpu(cpu
);
3325 trigger_load_balance(rq
, cpu
);
3329 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3331 void fastcall
add_preempt_count(int val
)
3336 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3338 preempt_count() += val
;
3340 * Spinlock count overflowing soon?
3342 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3345 EXPORT_SYMBOL(add_preempt_count
);
3347 void fastcall
sub_preempt_count(int val
)
3352 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3355 * Is the spinlock portion underflowing?
3357 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3358 !(preempt_count() & PREEMPT_MASK
)))
3361 preempt_count() -= val
;
3363 EXPORT_SYMBOL(sub_preempt_count
);
3368 * Print scheduling while atomic bug:
3370 static noinline
void __schedule_bug(struct task_struct
*prev
)
3372 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3373 prev
->comm
, preempt_count(), prev
->pid
);
3374 debug_show_held_locks(prev
);
3375 if (irqs_disabled())
3376 print_irqtrace_events(prev
);
3381 * Various schedule()-time debugging checks and statistics:
3383 static inline void schedule_debug(struct task_struct
*prev
)
3386 * Test if we are atomic. Since do_exit() needs to call into
3387 * schedule() atomically, we ignore that path for now.
3388 * Otherwise, whine if we are scheduling when we should not be.
3390 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3391 __schedule_bug(prev
);
3393 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3395 schedstat_inc(this_rq(), sched_count
);
3396 #ifdef CONFIG_SCHEDSTATS
3397 if (unlikely(prev
->lock_depth
>= 0)) {
3398 schedstat_inc(this_rq(), bkl_count
);
3399 schedstat_inc(prev
, sched_info
.bkl_count
);
3405 * Pick up the highest-prio task:
3407 static inline struct task_struct
*
3408 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3410 const struct sched_class
*class;
3411 struct task_struct
*p
;
3414 * Optimization: we know that if all tasks are in
3415 * the fair class we can call that function directly:
3417 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3418 p
= fair_sched_class
.pick_next_task(rq
);
3423 class = sched_class_highest
;
3425 p
= class->pick_next_task(rq
);
3429 * Will never be NULL as the idle class always
3430 * returns a non-NULL p:
3432 class = class->next
;
3437 * schedule() is the main scheduler function.
3439 asmlinkage
void __sched
schedule(void)
3441 struct task_struct
*prev
, *next
;
3448 cpu
= smp_processor_id();
3452 switch_count
= &prev
->nivcsw
;
3454 release_kernel_lock(prev
);
3455 need_resched_nonpreemptible
:
3457 schedule_debug(prev
);
3460 * Do the rq-clock update outside the rq lock:
3462 local_irq_disable();
3463 __update_rq_clock(rq
);
3464 spin_lock(&rq
->lock
);
3465 clear_tsk_need_resched(prev
);
3467 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3468 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3469 unlikely(signal_pending(prev
)))) {
3470 prev
->state
= TASK_RUNNING
;
3472 deactivate_task(rq
, prev
, 1);
3474 switch_count
= &prev
->nvcsw
;
3477 if (unlikely(!rq
->nr_running
))
3478 idle_balance(cpu
, rq
);
3480 prev
->sched_class
->put_prev_task(rq
, prev
);
3481 next
= pick_next_task(rq
, prev
);
3483 sched_info_switch(prev
, next
);
3485 if (likely(prev
!= next
)) {
3490 context_switch(rq
, prev
, next
); /* unlocks the rq */
3492 spin_unlock_irq(&rq
->lock
);
3494 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3495 cpu
= smp_processor_id();
3497 goto need_resched_nonpreemptible
;
3499 preempt_enable_no_resched();
3500 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3503 EXPORT_SYMBOL(schedule
);
3505 #ifdef CONFIG_PREEMPT
3507 * this is the entry point to schedule() from in-kernel preemption
3508 * off of preempt_enable. Kernel preemptions off return from interrupt
3509 * occur there and call schedule directly.
3511 asmlinkage
void __sched
preempt_schedule(void)
3513 struct thread_info
*ti
= current_thread_info();
3514 #ifdef CONFIG_PREEMPT_BKL
3515 struct task_struct
*task
= current
;
3516 int saved_lock_depth
;
3519 * If there is a non-zero preempt_count or interrupts are disabled,
3520 * we do not want to preempt the current task. Just return..
3522 if (likely(ti
->preempt_count
|| irqs_disabled()))
3526 add_preempt_count(PREEMPT_ACTIVE
);
3529 * We keep the big kernel semaphore locked, but we
3530 * clear ->lock_depth so that schedule() doesnt
3531 * auto-release the semaphore:
3533 #ifdef CONFIG_PREEMPT_BKL
3534 saved_lock_depth
= task
->lock_depth
;
3535 task
->lock_depth
= -1;
3538 #ifdef CONFIG_PREEMPT_BKL
3539 task
->lock_depth
= saved_lock_depth
;
3541 sub_preempt_count(PREEMPT_ACTIVE
);
3544 * Check again in case we missed a preemption opportunity
3545 * between schedule and now.
3548 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3550 EXPORT_SYMBOL(preempt_schedule
);
3553 * this is the entry point to schedule() from kernel preemption
3554 * off of irq context.
3555 * Note, that this is called and return with irqs disabled. This will
3556 * protect us against recursive calling from irq.
3558 asmlinkage
void __sched
preempt_schedule_irq(void)
3560 struct thread_info
*ti
= current_thread_info();
3561 #ifdef CONFIG_PREEMPT_BKL
3562 struct task_struct
*task
= current
;
3563 int saved_lock_depth
;
3565 /* Catch callers which need to be fixed */
3566 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3569 add_preempt_count(PREEMPT_ACTIVE
);
3572 * We keep the big kernel semaphore locked, but we
3573 * clear ->lock_depth so that schedule() doesnt
3574 * auto-release the semaphore:
3576 #ifdef CONFIG_PREEMPT_BKL
3577 saved_lock_depth
= task
->lock_depth
;
3578 task
->lock_depth
= -1;
3582 local_irq_disable();
3583 #ifdef CONFIG_PREEMPT_BKL
3584 task
->lock_depth
= saved_lock_depth
;
3586 sub_preempt_count(PREEMPT_ACTIVE
);
3589 * Check again in case we missed a preemption opportunity
3590 * between schedule and now.
3593 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3596 #endif /* CONFIG_PREEMPT */
3598 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3601 return try_to_wake_up(curr
->private, mode
, sync
);
3603 EXPORT_SYMBOL(default_wake_function
);
3606 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3607 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3608 * number) then we wake all the non-exclusive tasks and one exclusive task.
3610 * There are circumstances in which we can try to wake a task which has already
3611 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3612 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3614 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3615 int nr_exclusive
, int sync
, void *key
)
3617 wait_queue_t
*curr
, *next
;
3619 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3620 unsigned flags
= curr
->flags
;
3622 if (curr
->func(curr
, mode
, sync
, key
) &&
3623 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3629 * __wake_up - wake up threads blocked on a waitqueue.
3631 * @mode: which threads
3632 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3633 * @key: is directly passed to the wakeup function
3635 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3636 int nr_exclusive
, void *key
)
3638 unsigned long flags
;
3640 spin_lock_irqsave(&q
->lock
, flags
);
3641 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3642 spin_unlock_irqrestore(&q
->lock
, flags
);
3644 EXPORT_SYMBOL(__wake_up
);
3647 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3649 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3651 __wake_up_common(q
, mode
, 1, 0, NULL
);
3655 * __wake_up_sync - wake up threads blocked on a waitqueue.
3657 * @mode: which threads
3658 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3660 * The sync wakeup differs that the waker knows that it will schedule
3661 * away soon, so while the target thread will be woken up, it will not
3662 * be migrated to another CPU - ie. the two threads are 'synchronized'
3663 * with each other. This can prevent needless bouncing between CPUs.
3665 * On UP it can prevent extra preemption.
3668 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3670 unsigned long flags
;
3676 if (unlikely(!nr_exclusive
))
3679 spin_lock_irqsave(&q
->lock
, flags
);
3680 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3681 spin_unlock_irqrestore(&q
->lock
, flags
);
3683 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3685 void fastcall
complete(struct completion
*x
)
3687 unsigned long flags
;
3689 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3691 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3693 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3695 EXPORT_SYMBOL(complete
);
3697 void fastcall
complete_all(struct completion
*x
)
3699 unsigned long flags
;
3701 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3702 x
->done
+= UINT_MAX
/2;
3703 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3705 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3707 EXPORT_SYMBOL(complete_all
);
3709 static inline long __sched
3710 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3713 DECLARE_WAITQUEUE(wait
, current
);
3715 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3716 __add_wait_queue_tail(&x
->wait
, &wait
);
3718 if (state
== TASK_INTERRUPTIBLE
&&
3719 signal_pending(current
)) {
3720 __remove_wait_queue(&x
->wait
, &wait
);
3721 return -ERESTARTSYS
;
3723 __set_current_state(state
);
3724 spin_unlock_irq(&x
->wait
.lock
);
3725 timeout
= schedule_timeout(timeout
);
3726 spin_lock_irq(&x
->wait
.lock
);
3728 __remove_wait_queue(&x
->wait
, &wait
);
3732 __remove_wait_queue(&x
->wait
, &wait
);
3739 wait_for_common(struct completion
*x
, long timeout
, int state
)
3743 spin_lock_irq(&x
->wait
.lock
);
3744 timeout
= do_wait_for_common(x
, timeout
, state
);
3745 spin_unlock_irq(&x
->wait
.lock
);
3749 void fastcall __sched
wait_for_completion(struct completion
*x
)
3751 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3753 EXPORT_SYMBOL(wait_for_completion
);
3755 unsigned long fastcall __sched
3756 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3758 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3760 EXPORT_SYMBOL(wait_for_completion_timeout
);
3762 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3764 return wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3766 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3768 unsigned long fastcall __sched
3769 wait_for_completion_interruptible_timeout(struct completion
*x
,
3770 unsigned long timeout
)
3772 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3774 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3777 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3779 unsigned long flags
;
3782 init_waitqueue_entry(&wait
, current
);
3784 __set_current_state(state
);
3786 spin_lock_irqsave(&q
->lock
, flags
);
3787 __add_wait_queue(q
, &wait
);
3788 spin_unlock(&q
->lock
);
3789 timeout
= schedule_timeout(timeout
);
3790 spin_lock_irq(&q
->lock
);
3791 __remove_wait_queue(q
, &wait
);
3792 spin_unlock_irqrestore(&q
->lock
, flags
);
3797 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3799 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3801 EXPORT_SYMBOL(interruptible_sleep_on
);
3804 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3806 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3808 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3810 void __sched
sleep_on(wait_queue_head_t
*q
)
3812 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3814 EXPORT_SYMBOL(sleep_on
);
3816 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3818 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3820 EXPORT_SYMBOL(sleep_on_timeout
);
3822 #ifdef CONFIG_RT_MUTEXES
3825 * rt_mutex_setprio - set the current priority of a task
3827 * @prio: prio value (kernel-internal form)
3829 * This function changes the 'effective' priority of a task. It does
3830 * not touch ->normal_prio like __setscheduler().
3832 * Used by the rt_mutex code to implement priority inheritance logic.
3834 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3836 unsigned long flags
;
3837 int oldprio
, on_rq
, running
;
3840 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3842 rq
= task_rq_lock(p
, &flags
);
3843 update_rq_clock(rq
);
3846 on_rq
= p
->se
.on_rq
;
3847 running
= task_running(rq
, p
);
3849 dequeue_task(rq
, p
, 0);
3851 p
->sched_class
->put_prev_task(rq
, p
);
3855 p
->sched_class
= &rt_sched_class
;
3857 p
->sched_class
= &fair_sched_class
;
3863 p
->sched_class
->set_curr_task(rq
);
3864 enqueue_task(rq
, p
, 0);
3866 * Reschedule if we are currently running on this runqueue and
3867 * our priority decreased, or if we are not currently running on
3868 * this runqueue and our priority is higher than the current's
3871 if (p
->prio
> oldprio
)
3872 resched_task(rq
->curr
);
3874 check_preempt_curr(rq
, p
);
3877 task_rq_unlock(rq
, &flags
);
3882 void set_user_nice(struct task_struct
*p
, long nice
)
3884 int old_prio
, delta
, on_rq
;
3885 unsigned long flags
;
3888 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3891 * We have to be careful, if called from sys_setpriority(),
3892 * the task might be in the middle of scheduling on another CPU.
3894 rq
= task_rq_lock(p
, &flags
);
3895 update_rq_clock(rq
);
3897 * The RT priorities are set via sched_setscheduler(), but we still
3898 * allow the 'normal' nice value to be set - but as expected
3899 * it wont have any effect on scheduling until the task is
3900 * SCHED_FIFO/SCHED_RR:
3902 if (task_has_rt_policy(p
)) {
3903 p
->static_prio
= NICE_TO_PRIO(nice
);
3906 on_rq
= p
->se
.on_rq
;
3908 dequeue_task(rq
, p
, 0);
3912 p
->static_prio
= NICE_TO_PRIO(nice
);
3915 p
->prio
= effective_prio(p
);
3916 delta
= p
->prio
- old_prio
;
3919 enqueue_task(rq
, p
, 0);
3922 * If the task increased its priority or is running and
3923 * lowered its priority, then reschedule its CPU:
3925 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3926 resched_task(rq
->curr
);
3929 task_rq_unlock(rq
, &flags
);
3931 EXPORT_SYMBOL(set_user_nice
);
3934 * can_nice - check if a task can reduce its nice value
3938 int can_nice(const struct task_struct
*p
, const int nice
)
3940 /* convert nice value [19,-20] to rlimit style value [1,40] */
3941 int nice_rlim
= 20 - nice
;
3943 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3944 capable(CAP_SYS_NICE
));
3947 #ifdef __ARCH_WANT_SYS_NICE
3950 * sys_nice - change the priority of the current process.
3951 * @increment: priority increment
3953 * sys_setpriority is a more generic, but much slower function that
3954 * does similar things.
3956 asmlinkage
long sys_nice(int increment
)
3961 * Setpriority might change our priority at the same moment.
3962 * We don't have to worry. Conceptually one call occurs first
3963 * and we have a single winner.
3965 if (increment
< -40)
3970 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3976 if (increment
< 0 && !can_nice(current
, nice
))
3979 retval
= security_task_setnice(current
, nice
);
3983 set_user_nice(current
, nice
);
3990 * task_prio - return the priority value of a given task.
3991 * @p: the task in question.
3993 * This is the priority value as seen by users in /proc.
3994 * RT tasks are offset by -200. Normal tasks are centered
3995 * around 0, value goes from -16 to +15.
3997 int task_prio(const struct task_struct
*p
)
3999 return p
->prio
- MAX_RT_PRIO
;
4003 * task_nice - return the nice value of a given task.
4004 * @p: the task in question.
4006 int task_nice(const struct task_struct
*p
)
4008 return TASK_NICE(p
);
4010 EXPORT_SYMBOL_GPL(task_nice
);
4013 * idle_cpu - is a given cpu idle currently?
4014 * @cpu: the processor in question.
4016 int idle_cpu(int cpu
)
4018 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4022 * idle_task - return the idle task for a given cpu.
4023 * @cpu: the processor in question.
4025 struct task_struct
*idle_task(int cpu
)
4027 return cpu_rq(cpu
)->idle
;
4031 * find_process_by_pid - find a process with a matching PID value.
4032 * @pid: the pid in question.
4034 static struct task_struct
*find_process_by_pid(pid_t pid
)
4036 return pid
? find_task_by_pid(pid
) : current
;
4039 /* Actually do priority change: must hold rq lock. */
4041 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4043 BUG_ON(p
->se
.on_rq
);
4046 switch (p
->policy
) {
4050 p
->sched_class
= &fair_sched_class
;
4054 p
->sched_class
= &rt_sched_class
;
4058 p
->rt_priority
= prio
;
4059 p
->normal_prio
= normal_prio(p
);
4060 /* we are holding p->pi_lock already */
4061 p
->prio
= rt_mutex_getprio(p
);
4066 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4067 * @p: the task in question.
4068 * @policy: new policy.
4069 * @param: structure containing the new RT priority.
4071 * NOTE that the task may be already dead.
4073 int sched_setscheduler(struct task_struct
*p
, int policy
,
4074 struct sched_param
*param
)
4076 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4077 unsigned long flags
;
4080 /* may grab non-irq protected spin_locks */
4081 BUG_ON(in_interrupt());
4083 /* double check policy once rq lock held */
4085 policy
= oldpolicy
= p
->policy
;
4086 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4087 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4088 policy
!= SCHED_IDLE
)
4091 * Valid priorities for SCHED_FIFO and SCHED_RR are
4092 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4093 * SCHED_BATCH and SCHED_IDLE is 0.
4095 if (param
->sched_priority
< 0 ||
4096 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4097 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4099 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4103 * Allow unprivileged RT tasks to decrease priority:
4105 if (!capable(CAP_SYS_NICE
)) {
4106 if (rt_policy(policy
)) {
4107 unsigned long rlim_rtprio
;
4109 if (!lock_task_sighand(p
, &flags
))
4111 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4112 unlock_task_sighand(p
, &flags
);
4114 /* can't set/change the rt policy */
4115 if (policy
!= p
->policy
&& !rlim_rtprio
)
4118 /* can't increase priority */
4119 if (param
->sched_priority
> p
->rt_priority
&&
4120 param
->sched_priority
> rlim_rtprio
)
4124 * Like positive nice levels, dont allow tasks to
4125 * move out of SCHED_IDLE either:
4127 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4130 /* can't change other user's priorities */
4131 if ((current
->euid
!= p
->euid
) &&
4132 (current
->euid
!= p
->uid
))
4136 retval
= security_task_setscheduler(p
, policy
, param
);
4140 * make sure no PI-waiters arrive (or leave) while we are
4141 * changing the priority of the task:
4143 spin_lock_irqsave(&p
->pi_lock
, flags
);
4145 * To be able to change p->policy safely, the apropriate
4146 * runqueue lock must be held.
4148 rq
= __task_rq_lock(p
);
4149 /* recheck policy now with rq lock held */
4150 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4151 policy
= oldpolicy
= -1;
4152 __task_rq_unlock(rq
);
4153 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4156 update_rq_clock(rq
);
4157 on_rq
= p
->se
.on_rq
;
4158 running
= task_running(rq
, p
);
4160 deactivate_task(rq
, p
, 0);
4162 p
->sched_class
->put_prev_task(rq
, p
);
4166 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4170 p
->sched_class
->set_curr_task(rq
);
4171 activate_task(rq
, p
, 0);
4173 * Reschedule if we are currently running on this runqueue and
4174 * our priority decreased, or if we are not currently running on
4175 * this runqueue and our priority is higher than the current's
4178 if (p
->prio
> oldprio
)
4179 resched_task(rq
->curr
);
4181 check_preempt_curr(rq
, p
);
4184 __task_rq_unlock(rq
);
4185 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4187 rt_mutex_adjust_pi(p
);
4191 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4194 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4196 struct sched_param lparam
;
4197 struct task_struct
*p
;
4200 if (!param
|| pid
< 0)
4202 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4207 p
= find_process_by_pid(pid
);
4209 retval
= sched_setscheduler(p
, policy
, &lparam
);
4216 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4217 * @pid: the pid in question.
4218 * @policy: new policy.
4219 * @param: structure containing the new RT priority.
4221 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4222 struct sched_param __user
*param
)
4224 /* negative values for policy are not valid */
4228 return do_sched_setscheduler(pid
, policy
, param
);
4232 * sys_sched_setparam - set/change the RT priority of a thread
4233 * @pid: the pid in question.
4234 * @param: structure containing the new RT priority.
4236 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4238 return do_sched_setscheduler(pid
, -1, param
);
4242 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4243 * @pid: the pid in question.
4245 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4247 struct task_struct
*p
;
4254 read_lock(&tasklist_lock
);
4255 p
= find_process_by_pid(pid
);
4257 retval
= security_task_getscheduler(p
);
4261 read_unlock(&tasklist_lock
);
4266 * sys_sched_getscheduler - get the RT priority of a thread
4267 * @pid: the pid in question.
4268 * @param: structure containing the RT priority.
4270 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4272 struct sched_param lp
;
4273 struct task_struct
*p
;
4276 if (!param
|| pid
< 0)
4279 read_lock(&tasklist_lock
);
4280 p
= find_process_by_pid(pid
);
4285 retval
= security_task_getscheduler(p
);
4289 lp
.sched_priority
= p
->rt_priority
;
4290 read_unlock(&tasklist_lock
);
4293 * This one might sleep, we cannot do it with a spinlock held ...
4295 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4300 read_unlock(&tasklist_lock
);
4304 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4306 cpumask_t cpus_allowed
;
4307 struct task_struct
*p
;
4310 mutex_lock(&sched_hotcpu_mutex
);
4311 read_lock(&tasklist_lock
);
4313 p
= find_process_by_pid(pid
);
4315 read_unlock(&tasklist_lock
);
4316 mutex_unlock(&sched_hotcpu_mutex
);
4321 * It is not safe to call set_cpus_allowed with the
4322 * tasklist_lock held. We will bump the task_struct's
4323 * usage count and then drop tasklist_lock.
4326 read_unlock(&tasklist_lock
);
4329 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4330 !capable(CAP_SYS_NICE
))
4333 retval
= security_task_setscheduler(p
, 0, NULL
);
4337 cpus_allowed
= cpuset_cpus_allowed(p
);
4338 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4339 retval
= set_cpus_allowed(p
, new_mask
);
4343 mutex_unlock(&sched_hotcpu_mutex
);
4347 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4348 cpumask_t
*new_mask
)
4350 if (len
< sizeof(cpumask_t
)) {
4351 memset(new_mask
, 0, sizeof(cpumask_t
));
4352 } else if (len
> sizeof(cpumask_t
)) {
4353 len
= sizeof(cpumask_t
);
4355 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4359 * sys_sched_setaffinity - set the cpu affinity of a process
4360 * @pid: pid of the process
4361 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4362 * @user_mask_ptr: user-space pointer to the new cpu mask
4364 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4365 unsigned long __user
*user_mask_ptr
)
4370 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4374 return sched_setaffinity(pid
, new_mask
);
4378 * Represents all cpu's present in the system
4379 * In systems capable of hotplug, this map could dynamically grow
4380 * as new cpu's are detected in the system via any platform specific
4381 * method, such as ACPI for e.g.
4384 cpumask_t cpu_present_map __read_mostly
;
4385 EXPORT_SYMBOL(cpu_present_map
);
4388 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4389 EXPORT_SYMBOL(cpu_online_map
);
4391 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4392 EXPORT_SYMBOL(cpu_possible_map
);
4395 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4397 struct task_struct
*p
;
4400 mutex_lock(&sched_hotcpu_mutex
);
4401 read_lock(&tasklist_lock
);
4404 p
= find_process_by_pid(pid
);
4408 retval
= security_task_getscheduler(p
);
4412 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4415 read_unlock(&tasklist_lock
);
4416 mutex_unlock(&sched_hotcpu_mutex
);
4422 * sys_sched_getaffinity - get the cpu affinity of a process
4423 * @pid: pid of the process
4424 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4425 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4427 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4428 unsigned long __user
*user_mask_ptr
)
4433 if (len
< sizeof(cpumask_t
))
4436 ret
= sched_getaffinity(pid
, &mask
);
4440 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4443 return sizeof(cpumask_t
);
4447 * sys_sched_yield - yield the current processor to other threads.
4449 * This function yields the current CPU to other tasks. If there are no
4450 * other threads running on this CPU then this function will return.
4452 asmlinkage
long sys_sched_yield(void)
4454 struct rq
*rq
= this_rq_lock();
4456 schedstat_inc(rq
, yld_count
);
4457 current
->sched_class
->yield_task(rq
);
4460 * Since we are going to call schedule() anyway, there's
4461 * no need to preempt or enable interrupts:
4463 __release(rq
->lock
);
4464 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4465 _raw_spin_unlock(&rq
->lock
);
4466 preempt_enable_no_resched();
4473 static void __cond_resched(void)
4475 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4476 __might_sleep(__FILE__
, __LINE__
);
4479 * The BKS might be reacquired before we have dropped
4480 * PREEMPT_ACTIVE, which could trigger a second
4481 * cond_resched() call.
4484 add_preempt_count(PREEMPT_ACTIVE
);
4486 sub_preempt_count(PREEMPT_ACTIVE
);
4487 } while (need_resched());
4490 int __sched
cond_resched(void)
4492 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4493 system_state
== SYSTEM_RUNNING
) {
4499 EXPORT_SYMBOL(cond_resched
);
4502 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4503 * call schedule, and on return reacquire the lock.
4505 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4506 * operations here to prevent schedule() from being called twice (once via
4507 * spin_unlock(), once by hand).
4509 int cond_resched_lock(spinlock_t
*lock
)
4513 if (need_lockbreak(lock
)) {
4519 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4520 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4521 _raw_spin_unlock(lock
);
4522 preempt_enable_no_resched();
4529 EXPORT_SYMBOL(cond_resched_lock
);
4531 int __sched
cond_resched_softirq(void)
4533 BUG_ON(!in_softirq());
4535 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4543 EXPORT_SYMBOL(cond_resched_softirq
);
4546 * yield - yield the current processor to other threads.
4548 * This is a shortcut for kernel-space yielding - it marks the
4549 * thread runnable and calls sys_sched_yield().
4551 void __sched
yield(void)
4553 set_current_state(TASK_RUNNING
);
4556 EXPORT_SYMBOL(yield
);
4559 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4560 * that process accounting knows that this is a task in IO wait state.
4562 * But don't do that if it is a deliberate, throttling IO wait (this task
4563 * has set its backing_dev_info: the queue against which it should throttle)
4565 void __sched
io_schedule(void)
4567 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4569 delayacct_blkio_start();
4570 atomic_inc(&rq
->nr_iowait
);
4572 atomic_dec(&rq
->nr_iowait
);
4573 delayacct_blkio_end();
4575 EXPORT_SYMBOL(io_schedule
);
4577 long __sched
io_schedule_timeout(long timeout
)
4579 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4582 delayacct_blkio_start();
4583 atomic_inc(&rq
->nr_iowait
);
4584 ret
= schedule_timeout(timeout
);
4585 atomic_dec(&rq
->nr_iowait
);
4586 delayacct_blkio_end();
4591 * sys_sched_get_priority_max - return maximum RT priority.
4592 * @policy: scheduling class.
4594 * this syscall returns the maximum rt_priority that can be used
4595 * by a given scheduling class.
4597 asmlinkage
long sys_sched_get_priority_max(int policy
)
4604 ret
= MAX_USER_RT_PRIO
-1;
4616 * sys_sched_get_priority_min - return minimum RT priority.
4617 * @policy: scheduling class.
4619 * this syscall returns the minimum rt_priority that can be used
4620 * by a given scheduling class.
4622 asmlinkage
long sys_sched_get_priority_min(int policy
)
4640 * sys_sched_rr_get_interval - return the default timeslice of a process.
4641 * @pid: pid of the process.
4642 * @interval: userspace pointer to the timeslice value.
4644 * this syscall writes the default timeslice value of a given process
4645 * into the user-space timespec buffer. A value of '0' means infinity.
4648 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4650 struct task_struct
*p
;
4651 unsigned int time_slice
;
4659 read_lock(&tasklist_lock
);
4660 p
= find_process_by_pid(pid
);
4664 retval
= security_task_getscheduler(p
);
4668 if (p
->policy
== SCHED_FIFO
)
4670 else if (p
->policy
== SCHED_RR
)
4671 time_slice
= DEF_TIMESLICE
;
4673 struct sched_entity
*se
= &p
->se
;
4674 unsigned long flags
;
4677 rq
= task_rq_lock(p
, &flags
);
4678 time_slice
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
4679 task_rq_unlock(rq
, &flags
);
4681 read_unlock(&tasklist_lock
);
4682 jiffies_to_timespec(time_slice
, &t
);
4683 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4687 read_unlock(&tasklist_lock
);
4691 static const char stat_nam
[] = "RSDTtZX";
4693 static void show_task(struct task_struct
*p
)
4695 unsigned long free
= 0;
4698 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4699 printk("%-13.13s %c", p
->comm
,
4700 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4701 #if BITS_PER_LONG == 32
4702 if (state
== TASK_RUNNING
)
4703 printk(" running ");
4705 printk(" %08lx ", thread_saved_pc(p
));
4707 if (state
== TASK_RUNNING
)
4708 printk(" running task ");
4710 printk(" %016lx ", thread_saved_pc(p
));
4712 #ifdef CONFIG_DEBUG_STACK_USAGE
4714 unsigned long *n
= end_of_stack(p
);
4717 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4720 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4722 if (state
!= TASK_RUNNING
)
4723 show_stack(p
, NULL
);
4726 void show_state_filter(unsigned long state_filter
)
4728 struct task_struct
*g
, *p
;
4730 #if BITS_PER_LONG == 32
4732 " task PC stack pid father\n");
4735 " task PC stack pid father\n");
4737 read_lock(&tasklist_lock
);
4738 do_each_thread(g
, p
) {
4740 * reset the NMI-timeout, listing all files on a slow
4741 * console might take alot of time:
4743 touch_nmi_watchdog();
4744 if (!state_filter
|| (p
->state
& state_filter
))
4746 } while_each_thread(g
, p
);
4748 touch_all_softlockup_watchdogs();
4750 #ifdef CONFIG_SCHED_DEBUG
4751 sysrq_sched_debug_show();
4753 read_unlock(&tasklist_lock
);
4755 * Only show locks if all tasks are dumped:
4757 if (state_filter
== -1)
4758 debug_show_all_locks();
4761 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4763 idle
->sched_class
= &idle_sched_class
;
4767 * init_idle - set up an idle thread for a given CPU
4768 * @idle: task in question
4769 * @cpu: cpu the idle task belongs to
4771 * NOTE: this function does not set the idle thread's NEED_RESCHED
4772 * flag, to make booting more robust.
4774 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4776 struct rq
*rq
= cpu_rq(cpu
);
4777 unsigned long flags
;
4780 idle
->se
.exec_start
= sched_clock();
4782 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4783 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4784 __set_task_cpu(idle
, cpu
);
4786 spin_lock_irqsave(&rq
->lock
, flags
);
4787 rq
->curr
= rq
->idle
= idle
;
4788 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4791 spin_unlock_irqrestore(&rq
->lock
, flags
);
4793 /* Set the preempt count _outside_ the spinlocks! */
4794 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4795 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4797 task_thread_info(idle
)->preempt_count
= 0;
4800 * The idle tasks have their own, simple scheduling class:
4802 idle
->sched_class
= &idle_sched_class
;
4806 * In a system that switches off the HZ timer nohz_cpu_mask
4807 * indicates which cpus entered this state. This is used
4808 * in the rcu update to wait only for active cpus. For system
4809 * which do not switch off the HZ timer nohz_cpu_mask should
4810 * always be CPU_MASK_NONE.
4812 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4816 * This is how migration works:
4818 * 1) we queue a struct migration_req structure in the source CPU's
4819 * runqueue and wake up that CPU's migration thread.
4820 * 2) we down() the locked semaphore => thread blocks.
4821 * 3) migration thread wakes up (implicitly it forces the migrated
4822 * thread off the CPU)
4823 * 4) it gets the migration request and checks whether the migrated
4824 * task is still in the wrong runqueue.
4825 * 5) if it's in the wrong runqueue then the migration thread removes
4826 * it and puts it into the right queue.
4827 * 6) migration thread up()s the semaphore.
4828 * 7) we wake up and the migration is done.
4832 * Change a given task's CPU affinity. Migrate the thread to a
4833 * proper CPU and schedule it away if the CPU it's executing on
4834 * is removed from the allowed bitmask.
4836 * NOTE: the caller must have a valid reference to the task, the
4837 * task must not exit() & deallocate itself prematurely. The
4838 * call is not atomic; no spinlocks may be held.
4840 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4842 struct migration_req req
;
4843 unsigned long flags
;
4847 rq
= task_rq_lock(p
, &flags
);
4848 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4853 p
->cpus_allowed
= new_mask
;
4854 /* Can the task run on the task's current CPU? If so, we're done */
4855 if (cpu_isset(task_cpu(p
), new_mask
))
4858 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4859 /* Need help from migration thread: drop lock and wait. */
4860 task_rq_unlock(rq
, &flags
);
4861 wake_up_process(rq
->migration_thread
);
4862 wait_for_completion(&req
.done
);
4863 tlb_migrate_finish(p
->mm
);
4867 task_rq_unlock(rq
, &flags
);
4871 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4874 * Move (not current) task off this cpu, onto dest cpu. We're doing
4875 * this because either it can't run here any more (set_cpus_allowed()
4876 * away from this CPU, or CPU going down), or because we're
4877 * attempting to rebalance this task on exec (sched_exec).
4879 * So we race with normal scheduler movements, but that's OK, as long
4880 * as the task is no longer on this CPU.
4882 * Returns non-zero if task was successfully migrated.
4884 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4886 struct rq
*rq_dest
, *rq_src
;
4889 if (unlikely(cpu_is_offline(dest_cpu
)))
4892 rq_src
= cpu_rq(src_cpu
);
4893 rq_dest
= cpu_rq(dest_cpu
);
4895 double_rq_lock(rq_src
, rq_dest
);
4896 /* Already moved. */
4897 if (task_cpu(p
) != src_cpu
)
4899 /* Affinity changed (again). */
4900 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4903 on_rq
= p
->se
.on_rq
;
4905 deactivate_task(rq_src
, p
, 0);
4907 set_task_cpu(p
, dest_cpu
);
4909 activate_task(rq_dest
, p
, 0);
4910 check_preempt_curr(rq_dest
, p
);
4914 double_rq_unlock(rq_src
, rq_dest
);
4919 * migration_thread - this is a highprio system thread that performs
4920 * thread migration by bumping thread off CPU then 'pushing' onto
4923 static int migration_thread(void *data
)
4925 int cpu
= (long)data
;
4929 BUG_ON(rq
->migration_thread
!= current
);
4931 set_current_state(TASK_INTERRUPTIBLE
);
4932 while (!kthread_should_stop()) {
4933 struct migration_req
*req
;
4934 struct list_head
*head
;
4936 spin_lock_irq(&rq
->lock
);
4938 if (cpu_is_offline(cpu
)) {
4939 spin_unlock_irq(&rq
->lock
);
4943 if (rq
->active_balance
) {
4944 active_load_balance(rq
, cpu
);
4945 rq
->active_balance
= 0;
4948 head
= &rq
->migration_queue
;
4950 if (list_empty(head
)) {
4951 spin_unlock_irq(&rq
->lock
);
4953 set_current_state(TASK_INTERRUPTIBLE
);
4956 req
= list_entry(head
->next
, struct migration_req
, list
);
4957 list_del_init(head
->next
);
4959 spin_unlock(&rq
->lock
);
4960 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4963 complete(&req
->done
);
4965 __set_current_state(TASK_RUNNING
);
4969 /* Wait for kthread_stop */
4970 set_current_state(TASK_INTERRUPTIBLE
);
4971 while (!kthread_should_stop()) {
4973 set_current_state(TASK_INTERRUPTIBLE
);
4975 __set_current_state(TASK_RUNNING
);
4979 #ifdef CONFIG_HOTPLUG_CPU
4981 * Figure out where task on dead CPU should go, use force if neccessary.
4982 * NOTE: interrupts should be disabled by the caller
4984 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
4986 unsigned long flags
;
4993 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4994 cpus_and(mask
, mask
, p
->cpus_allowed
);
4995 dest_cpu
= any_online_cpu(mask
);
4997 /* On any allowed CPU? */
4998 if (dest_cpu
== NR_CPUS
)
4999 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5001 /* No more Mr. Nice Guy. */
5002 if (dest_cpu
== NR_CPUS
) {
5003 rq
= task_rq_lock(p
, &flags
);
5004 cpus_setall(p
->cpus_allowed
);
5005 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5006 task_rq_unlock(rq
, &flags
);
5009 * Don't tell them about moving exiting tasks or
5010 * kernel threads (both mm NULL), since they never
5013 if (p
->mm
&& printk_ratelimit())
5014 printk(KERN_INFO
"process %d (%s) no "
5015 "longer affine to cpu%d\n",
5016 p
->pid
, p
->comm
, dead_cpu
);
5018 } while (!__migrate_task(p
, dead_cpu
, dest_cpu
));
5022 * While a dead CPU has no uninterruptible tasks queued at this point,
5023 * it might still have a nonzero ->nr_uninterruptible counter, because
5024 * for performance reasons the counter is not stricly tracking tasks to
5025 * their home CPUs. So we just add the counter to another CPU's counter,
5026 * to keep the global sum constant after CPU-down:
5028 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5030 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5031 unsigned long flags
;
5033 local_irq_save(flags
);
5034 double_rq_lock(rq_src
, rq_dest
);
5035 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5036 rq_src
->nr_uninterruptible
= 0;
5037 double_rq_unlock(rq_src
, rq_dest
);
5038 local_irq_restore(flags
);
5041 /* Run through task list and migrate tasks from the dead cpu. */
5042 static void migrate_live_tasks(int src_cpu
)
5044 struct task_struct
*p
, *t
;
5046 write_lock_irq(&tasklist_lock
);
5048 do_each_thread(t
, p
) {
5052 if (task_cpu(p
) == src_cpu
)
5053 move_task_off_dead_cpu(src_cpu
, p
);
5054 } while_each_thread(t
, p
);
5056 write_unlock_irq(&tasklist_lock
);
5060 * activate_idle_task - move idle task to the _front_ of runqueue.
5062 static void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
5064 update_rq_clock(rq
);
5066 if (p
->state
== TASK_UNINTERRUPTIBLE
)
5067 rq
->nr_uninterruptible
--;
5069 enqueue_task(rq
, p
, 0);
5070 inc_nr_running(p
, rq
);
5074 * Schedules idle task to be the next runnable task on current CPU.
5075 * It does so by boosting its priority to highest possible and adding it to
5076 * the _front_ of the runqueue. Used by CPU offline code.
5078 void sched_idle_next(void)
5080 int this_cpu
= smp_processor_id();
5081 struct rq
*rq
= cpu_rq(this_cpu
);
5082 struct task_struct
*p
= rq
->idle
;
5083 unsigned long flags
;
5085 /* cpu has to be offline */
5086 BUG_ON(cpu_online(this_cpu
));
5089 * Strictly not necessary since rest of the CPUs are stopped by now
5090 * and interrupts disabled on the current cpu.
5092 spin_lock_irqsave(&rq
->lock
, flags
);
5094 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5096 /* Add idle task to the _front_ of its priority queue: */
5097 activate_idle_task(p
, rq
);
5099 spin_unlock_irqrestore(&rq
->lock
, flags
);
5103 * Ensures that the idle task is using init_mm right before its cpu goes
5106 void idle_task_exit(void)
5108 struct mm_struct
*mm
= current
->active_mm
;
5110 BUG_ON(cpu_online(smp_processor_id()));
5113 switch_mm(mm
, &init_mm
, current
);
5117 /* called under rq->lock with disabled interrupts */
5118 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5120 struct rq
*rq
= cpu_rq(dead_cpu
);
5122 /* Must be exiting, otherwise would be on tasklist. */
5123 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5125 /* Cannot have done final schedule yet: would have vanished. */
5126 BUG_ON(p
->state
== TASK_DEAD
);
5131 * Drop lock around migration; if someone else moves it,
5132 * that's OK. No task can be added to this CPU, so iteration is
5134 * NOTE: interrupts should be left disabled --dev@
5136 spin_unlock(&rq
->lock
);
5137 move_task_off_dead_cpu(dead_cpu
, p
);
5138 spin_lock(&rq
->lock
);
5143 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5144 static void migrate_dead_tasks(unsigned int dead_cpu
)
5146 struct rq
*rq
= cpu_rq(dead_cpu
);
5147 struct task_struct
*next
;
5150 if (!rq
->nr_running
)
5152 update_rq_clock(rq
);
5153 next
= pick_next_task(rq
, rq
->curr
);
5156 migrate_dead(dead_cpu
, next
);
5160 #endif /* CONFIG_HOTPLUG_CPU */
5162 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5164 static struct ctl_table sd_ctl_dir
[] = {
5166 .procname
= "sched_domain",
5172 static struct ctl_table sd_ctl_root
[] = {
5174 .ctl_name
= CTL_KERN
,
5175 .procname
= "kernel",
5177 .child
= sd_ctl_dir
,
5182 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5184 struct ctl_table
*entry
=
5185 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5188 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5194 set_table_entry(struct ctl_table
*entry
,
5195 const char *procname
, void *data
, int maxlen
,
5196 mode_t mode
, proc_handler
*proc_handler
)
5198 entry
->procname
= procname
;
5200 entry
->maxlen
= maxlen
;
5202 entry
->proc_handler
= proc_handler
;
5205 static struct ctl_table
*
5206 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5208 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5210 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5211 sizeof(long), 0644, proc_doulongvec_minmax
);
5212 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5213 sizeof(long), 0644, proc_doulongvec_minmax
);
5214 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5215 sizeof(int), 0644, proc_dointvec_minmax
);
5216 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5217 sizeof(int), 0644, proc_dointvec_minmax
);
5218 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5219 sizeof(int), 0644, proc_dointvec_minmax
);
5220 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5221 sizeof(int), 0644, proc_dointvec_minmax
);
5222 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5223 sizeof(int), 0644, proc_dointvec_minmax
);
5224 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5225 sizeof(int), 0644, proc_dointvec_minmax
);
5226 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5227 sizeof(int), 0644, proc_dointvec_minmax
);
5228 set_table_entry(&table
[9], "cache_nice_tries",
5229 &sd
->cache_nice_tries
,
5230 sizeof(int), 0644, proc_dointvec_minmax
);
5231 set_table_entry(&table
[10], "flags", &sd
->flags
,
5232 sizeof(int), 0644, proc_dointvec_minmax
);
5237 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5239 struct ctl_table
*entry
, *table
;
5240 struct sched_domain
*sd
;
5241 int domain_num
= 0, i
;
5244 for_each_domain(cpu
, sd
)
5246 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5249 for_each_domain(cpu
, sd
) {
5250 snprintf(buf
, 32, "domain%d", i
);
5251 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5253 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5260 static struct ctl_table_header
*sd_sysctl_header
;
5261 static void init_sched_domain_sysctl(void)
5263 int i
, cpu_num
= num_online_cpus();
5264 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5267 sd_ctl_dir
[0].child
= entry
;
5269 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5270 snprintf(buf
, 32, "cpu%d", i
);
5271 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5273 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5275 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5278 static void init_sched_domain_sysctl(void)
5284 * migration_call - callback that gets triggered when a CPU is added.
5285 * Here we can start up the necessary migration thread for the new CPU.
5287 static int __cpuinit
5288 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5290 struct task_struct
*p
;
5291 int cpu
= (long)hcpu
;
5292 unsigned long flags
;
5296 case CPU_LOCK_ACQUIRE
:
5297 mutex_lock(&sched_hotcpu_mutex
);
5300 case CPU_UP_PREPARE
:
5301 case CPU_UP_PREPARE_FROZEN
:
5302 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5305 kthread_bind(p
, cpu
);
5306 /* Must be high prio: stop_machine expects to yield to it. */
5307 rq
= task_rq_lock(p
, &flags
);
5308 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5309 task_rq_unlock(rq
, &flags
);
5310 cpu_rq(cpu
)->migration_thread
= p
;
5314 case CPU_ONLINE_FROZEN
:
5315 /* Strictly unneccessary, as first user will wake it. */
5316 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5319 #ifdef CONFIG_HOTPLUG_CPU
5320 case CPU_UP_CANCELED
:
5321 case CPU_UP_CANCELED_FROZEN
:
5322 if (!cpu_rq(cpu
)->migration_thread
)
5324 /* Unbind it from offline cpu so it can run. Fall thru. */
5325 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5326 any_online_cpu(cpu_online_map
));
5327 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5328 cpu_rq(cpu
)->migration_thread
= NULL
;
5332 case CPU_DEAD_FROZEN
:
5333 migrate_live_tasks(cpu
);
5335 kthread_stop(rq
->migration_thread
);
5336 rq
->migration_thread
= NULL
;
5337 /* Idle task back to normal (off runqueue, low prio) */
5338 rq
= task_rq_lock(rq
->idle
, &flags
);
5339 update_rq_clock(rq
);
5340 deactivate_task(rq
, rq
->idle
, 0);
5341 rq
->idle
->static_prio
= MAX_PRIO
;
5342 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5343 rq
->idle
->sched_class
= &idle_sched_class
;
5344 migrate_dead_tasks(cpu
);
5345 task_rq_unlock(rq
, &flags
);
5346 migrate_nr_uninterruptible(rq
);
5347 BUG_ON(rq
->nr_running
!= 0);
5349 /* No need to migrate the tasks: it was best-effort if
5350 * they didn't take sched_hotcpu_mutex. Just wake up
5351 * the requestors. */
5352 spin_lock_irq(&rq
->lock
);
5353 while (!list_empty(&rq
->migration_queue
)) {
5354 struct migration_req
*req
;
5356 req
= list_entry(rq
->migration_queue
.next
,
5357 struct migration_req
, list
);
5358 list_del_init(&req
->list
);
5359 complete(&req
->done
);
5361 spin_unlock_irq(&rq
->lock
);
5364 case CPU_LOCK_RELEASE
:
5365 mutex_unlock(&sched_hotcpu_mutex
);
5371 /* Register at highest priority so that task migration (migrate_all_tasks)
5372 * happens before everything else.
5374 static struct notifier_block __cpuinitdata migration_notifier
= {
5375 .notifier_call
= migration_call
,
5379 int __init
migration_init(void)
5381 void *cpu
= (void *)(long)smp_processor_id();
5384 /* Start one for the boot CPU: */
5385 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5386 BUG_ON(err
== NOTIFY_BAD
);
5387 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5388 register_cpu_notifier(&migration_notifier
);
5396 /* Number of possible processor ids */
5397 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5398 EXPORT_SYMBOL(nr_cpu_ids
);
5400 #ifdef CONFIG_SCHED_DEBUG
5401 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5406 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5410 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5415 struct sched_group
*group
= sd
->groups
;
5416 cpumask_t groupmask
;
5418 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5419 cpus_clear(groupmask
);
5422 for (i
= 0; i
< level
+ 1; i
++)
5424 printk("domain %d: ", level
);
5426 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5427 printk("does not load-balance\n");
5429 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5434 printk("span %s\n", str
);
5436 if (!cpu_isset(cpu
, sd
->span
))
5437 printk(KERN_ERR
"ERROR: domain->span does not contain "
5439 if (!cpu_isset(cpu
, group
->cpumask
))
5440 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5444 for (i
= 0; i
< level
+ 2; i
++)
5450 printk(KERN_ERR
"ERROR: group is NULL\n");
5454 if (!group
->__cpu_power
) {
5456 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5461 if (!cpus_weight(group
->cpumask
)) {
5463 printk(KERN_ERR
"ERROR: empty group\n");
5467 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5469 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5473 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5475 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5478 group
= group
->next
;
5479 } while (group
!= sd
->groups
);
5482 if (!cpus_equal(sd
->span
, groupmask
))
5483 printk(KERN_ERR
"ERROR: groups don't span "
5491 if (!cpus_subset(groupmask
, sd
->span
))
5492 printk(KERN_ERR
"ERROR: parent span is not a superset "
5493 "of domain->span\n");
5498 # define sched_domain_debug(sd, cpu) do { } while (0)
5501 static int sd_degenerate(struct sched_domain
*sd
)
5503 if (cpus_weight(sd
->span
) == 1)
5506 /* Following flags need at least 2 groups */
5507 if (sd
->flags
& (SD_LOAD_BALANCE
|
5508 SD_BALANCE_NEWIDLE
|
5512 SD_SHARE_PKG_RESOURCES
)) {
5513 if (sd
->groups
!= sd
->groups
->next
)
5517 /* Following flags don't use groups */
5518 if (sd
->flags
& (SD_WAKE_IDLE
|
5527 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5529 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5531 if (sd_degenerate(parent
))
5534 if (!cpus_equal(sd
->span
, parent
->span
))
5537 /* Does parent contain flags not in child? */
5538 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5539 if (cflags
& SD_WAKE_AFFINE
)
5540 pflags
&= ~SD_WAKE_BALANCE
;
5541 /* Flags needing groups don't count if only 1 group in parent */
5542 if (parent
->groups
== parent
->groups
->next
) {
5543 pflags
&= ~(SD_LOAD_BALANCE
|
5544 SD_BALANCE_NEWIDLE
|
5548 SD_SHARE_PKG_RESOURCES
);
5550 if (~cflags
& pflags
)
5557 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5558 * hold the hotplug lock.
5560 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5562 struct rq
*rq
= cpu_rq(cpu
);
5563 struct sched_domain
*tmp
;
5565 /* Remove the sched domains which do not contribute to scheduling. */
5566 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5567 struct sched_domain
*parent
= tmp
->parent
;
5570 if (sd_parent_degenerate(tmp
, parent
)) {
5571 tmp
->parent
= parent
->parent
;
5573 parent
->parent
->child
= tmp
;
5577 if (sd
&& sd_degenerate(sd
)) {
5583 sched_domain_debug(sd
, cpu
);
5585 rcu_assign_pointer(rq
->sd
, sd
);
5588 /* cpus with isolated domains */
5589 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5591 /* Setup the mask of cpus configured for isolated domains */
5592 static int __init
isolated_cpu_setup(char *str
)
5594 int ints
[NR_CPUS
], i
;
5596 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5597 cpus_clear(cpu_isolated_map
);
5598 for (i
= 1; i
<= ints
[0]; i
++)
5599 if (ints
[i
] < NR_CPUS
)
5600 cpu_set(ints
[i
], cpu_isolated_map
);
5604 __setup("isolcpus=", isolated_cpu_setup
);
5607 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5608 * to a function which identifies what group(along with sched group) a CPU
5609 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5610 * (due to the fact that we keep track of groups covered with a cpumask_t).
5612 * init_sched_build_groups will build a circular linked list of the groups
5613 * covered by the given span, and will set each group's ->cpumask correctly,
5614 * and ->cpu_power to 0.
5617 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5618 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5619 struct sched_group
**sg
))
5621 struct sched_group
*first
= NULL
, *last
= NULL
;
5622 cpumask_t covered
= CPU_MASK_NONE
;
5625 for_each_cpu_mask(i
, span
) {
5626 struct sched_group
*sg
;
5627 int group
= group_fn(i
, cpu_map
, &sg
);
5630 if (cpu_isset(i
, covered
))
5633 sg
->cpumask
= CPU_MASK_NONE
;
5634 sg
->__cpu_power
= 0;
5636 for_each_cpu_mask(j
, span
) {
5637 if (group_fn(j
, cpu_map
, NULL
) != group
)
5640 cpu_set(j
, covered
);
5641 cpu_set(j
, sg
->cpumask
);
5652 #define SD_NODES_PER_DOMAIN 16
5657 * find_next_best_node - find the next node to include in a sched_domain
5658 * @node: node whose sched_domain we're building
5659 * @used_nodes: nodes already in the sched_domain
5661 * Find the next node to include in a given scheduling domain. Simply
5662 * finds the closest node not already in the @used_nodes map.
5664 * Should use nodemask_t.
5666 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5668 int i
, n
, val
, min_val
, best_node
= 0;
5672 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5673 /* Start at @node */
5674 n
= (node
+ i
) % MAX_NUMNODES
;
5676 if (!nr_cpus_node(n
))
5679 /* Skip already used nodes */
5680 if (test_bit(n
, used_nodes
))
5683 /* Simple min distance search */
5684 val
= node_distance(node
, n
);
5686 if (val
< min_val
) {
5692 set_bit(best_node
, used_nodes
);
5697 * sched_domain_node_span - get a cpumask for a node's sched_domain
5698 * @node: node whose cpumask we're constructing
5699 * @size: number of nodes to include in this span
5701 * Given a node, construct a good cpumask for its sched_domain to span. It
5702 * should be one that prevents unnecessary balancing, but also spreads tasks
5705 static cpumask_t
sched_domain_node_span(int node
)
5707 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5708 cpumask_t span
, nodemask
;
5712 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5714 nodemask
= node_to_cpumask(node
);
5715 cpus_or(span
, span
, nodemask
);
5716 set_bit(node
, used_nodes
);
5718 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5719 int next_node
= find_next_best_node(node
, used_nodes
);
5721 nodemask
= node_to_cpumask(next_node
);
5722 cpus_or(span
, span
, nodemask
);
5729 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5732 * SMT sched-domains:
5734 #ifdef CONFIG_SCHED_SMT
5735 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5736 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5738 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5739 struct sched_group
**sg
)
5742 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5748 * multi-core sched-domains:
5750 #ifdef CONFIG_SCHED_MC
5751 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5752 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5755 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5756 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5757 struct sched_group
**sg
)
5760 cpumask_t mask
= cpu_sibling_map
[cpu
];
5761 cpus_and(mask
, mask
, *cpu_map
);
5762 group
= first_cpu(mask
);
5764 *sg
= &per_cpu(sched_group_core
, group
);
5767 #elif defined(CONFIG_SCHED_MC)
5768 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5769 struct sched_group
**sg
)
5772 *sg
= &per_cpu(sched_group_core
, cpu
);
5777 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5778 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5780 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5781 struct sched_group
**sg
)
5784 #ifdef CONFIG_SCHED_MC
5785 cpumask_t mask
= cpu_coregroup_map(cpu
);
5786 cpus_and(mask
, mask
, *cpu_map
);
5787 group
= first_cpu(mask
);
5788 #elif defined(CONFIG_SCHED_SMT)
5789 cpumask_t mask
= cpu_sibling_map
[cpu
];
5790 cpus_and(mask
, mask
, *cpu_map
);
5791 group
= first_cpu(mask
);
5796 *sg
= &per_cpu(sched_group_phys
, group
);
5802 * The init_sched_build_groups can't handle what we want to do with node
5803 * groups, so roll our own. Now each node has its own list of groups which
5804 * gets dynamically allocated.
5806 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5807 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5809 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5810 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5812 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5813 struct sched_group
**sg
)
5815 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5818 cpus_and(nodemask
, nodemask
, *cpu_map
);
5819 group
= first_cpu(nodemask
);
5822 *sg
= &per_cpu(sched_group_allnodes
, group
);
5826 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5828 struct sched_group
*sg
= group_head
;
5834 for_each_cpu_mask(j
, sg
->cpumask
) {
5835 struct sched_domain
*sd
;
5837 sd
= &per_cpu(phys_domains
, j
);
5838 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5840 * Only add "power" once for each
5846 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5849 } while (sg
!= group_head
);
5854 /* Free memory allocated for various sched_group structures */
5855 static void free_sched_groups(const cpumask_t
*cpu_map
)
5859 for_each_cpu_mask(cpu
, *cpu_map
) {
5860 struct sched_group
**sched_group_nodes
5861 = sched_group_nodes_bycpu
[cpu
];
5863 if (!sched_group_nodes
)
5866 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5867 cpumask_t nodemask
= node_to_cpumask(i
);
5868 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5870 cpus_and(nodemask
, nodemask
, *cpu_map
);
5871 if (cpus_empty(nodemask
))
5881 if (oldsg
!= sched_group_nodes
[i
])
5884 kfree(sched_group_nodes
);
5885 sched_group_nodes_bycpu
[cpu
] = NULL
;
5889 static void free_sched_groups(const cpumask_t
*cpu_map
)
5895 * Initialize sched groups cpu_power.
5897 * cpu_power indicates the capacity of sched group, which is used while
5898 * distributing the load between different sched groups in a sched domain.
5899 * Typically cpu_power for all the groups in a sched domain will be same unless
5900 * there are asymmetries in the topology. If there are asymmetries, group
5901 * having more cpu_power will pickup more load compared to the group having
5904 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5905 * the maximum number of tasks a group can handle in the presence of other idle
5906 * or lightly loaded groups in the same sched domain.
5908 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5910 struct sched_domain
*child
;
5911 struct sched_group
*group
;
5913 WARN_ON(!sd
|| !sd
->groups
);
5915 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5920 sd
->groups
->__cpu_power
= 0;
5923 * For perf policy, if the groups in child domain share resources
5924 * (for example cores sharing some portions of the cache hierarchy
5925 * or SMT), then set this domain groups cpu_power such that each group
5926 * can handle only one task, when there are other idle groups in the
5927 * same sched domain.
5929 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5931 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5932 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5937 * add cpu_power of each child group to this groups cpu_power
5939 group
= child
->groups
;
5941 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5942 group
= group
->next
;
5943 } while (group
!= child
->groups
);
5947 * Build sched domains for a given set of cpus and attach the sched domains
5948 * to the individual cpus
5950 static int build_sched_domains(const cpumask_t
*cpu_map
)
5954 struct sched_group
**sched_group_nodes
= NULL
;
5955 int sd_allnodes
= 0;
5958 * Allocate the per-node list of sched groups
5960 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5962 if (!sched_group_nodes
) {
5963 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5966 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5970 * Set up domains for cpus specified by the cpu_map.
5972 for_each_cpu_mask(i
, *cpu_map
) {
5973 struct sched_domain
*sd
= NULL
, *p
;
5974 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5976 cpus_and(nodemask
, nodemask
, *cpu_map
);
5979 if (cpus_weight(*cpu_map
) >
5980 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5981 sd
= &per_cpu(allnodes_domains
, i
);
5982 *sd
= SD_ALLNODES_INIT
;
5983 sd
->span
= *cpu_map
;
5984 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
5990 sd
= &per_cpu(node_domains
, i
);
5992 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5996 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6000 sd
= &per_cpu(phys_domains
, i
);
6002 sd
->span
= nodemask
;
6006 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6008 #ifdef CONFIG_SCHED_MC
6010 sd
= &per_cpu(core_domains
, i
);
6012 sd
->span
= cpu_coregroup_map(i
);
6013 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6016 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6019 #ifdef CONFIG_SCHED_SMT
6021 sd
= &per_cpu(cpu_domains
, i
);
6022 *sd
= SD_SIBLING_INIT
;
6023 sd
->span
= cpu_sibling_map
[i
];
6024 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6027 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6031 #ifdef CONFIG_SCHED_SMT
6032 /* Set up CPU (sibling) groups */
6033 for_each_cpu_mask(i
, *cpu_map
) {
6034 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6035 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6036 if (i
!= first_cpu(this_sibling_map
))
6039 init_sched_build_groups(this_sibling_map
, cpu_map
,
6044 #ifdef CONFIG_SCHED_MC
6045 /* Set up multi-core groups */
6046 for_each_cpu_mask(i
, *cpu_map
) {
6047 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6048 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6049 if (i
!= first_cpu(this_core_map
))
6051 init_sched_build_groups(this_core_map
, cpu_map
,
6052 &cpu_to_core_group
);
6056 /* Set up physical groups */
6057 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6058 cpumask_t nodemask
= node_to_cpumask(i
);
6060 cpus_and(nodemask
, nodemask
, *cpu_map
);
6061 if (cpus_empty(nodemask
))
6064 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6068 /* Set up node groups */
6070 init_sched_build_groups(*cpu_map
, cpu_map
,
6071 &cpu_to_allnodes_group
);
6073 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6074 /* Set up node groups */
6075 struct sched_group
*sg
, *prev
;
6076 cpumask_t nodemask
= node_to_cpumask(i
);
6077 cpumask_t domainspan
;
6078 cpumask_t covered
= CPU_MASK_NONE
;
6081 cpus_and(nodemask
, nodemask
, *cpu_map
);
6082 if (cpus_empty(nodemask
)) {
6083 sched_group_nodes
[i
] = NULL
;
6087 domainspan
= sched_domain_node_span(i
);
6088 cpus_and(domainspan
, domainspan
, *cpu_map
);
6090 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6092 printk(KERN_WARNING
"Can not alloc domain group for "
6096 sched_group_nodes
[i
] = sg
;
6097 for_each_cpu_mask(j
, nodemask
) {
6098 struct sched_domain
*sd
;
6100 sd
= &per_cpu(node_domains
, j
);
6103 sg
->__cpu_power
= 0;
6104 sg
->cpumask
= nodemask
;
6106 cpus_or(covered
, covered
, nodemask
);
6109 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6110 cpumask_t tmp
, notcovered
;
6111 int n
= (i
+ j
) % MAX_NUMNODES
;
6113 cpus_complement(notcovered
, covered
);
6114 cpus_and(tmp
, notcovered
, *cpu_map
);
6115 cpus_and(tmp
, tmp
, domainspan
);
6116 if (cpus_empty(tmp
))
6119 nodemask
= node_to_cpumask(n
);
6120 cpus_and(tmp
, tmp
, nodemask
);
6121 if (cpus_empty(tmp
))
6124 sg
= kmalloc_node(sizeof(struct sched_group
),
6128 "Can not alloc domain group for node %d\n", j
);
6131 sg
->__cpu_power
= 0;
6133 sg
->next
= prev
->next
;
6134 cpus_or(covered
, covered
, tmp
);
6141 /* Calculate CPU power for physical packages and nodes */
6142 #ifdef CONFIG_SCHED_SMT
6143 for_each_cpu_mask(i
, *cpu_map
) {
6144 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6146 init_sched_groups_power(i
, sd
);
6149 #ifdef CONFIG_SCHED_MC
6150 for_each_cpu_mask(i
, *cpu_map
) {
6151 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6153 init_sched_groups_power(i
, sd
);
6157 for_each_cpu_mask(i
, *cpu_map
) {
6158 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6160 init_sched_groups_power(i
, sd
);
6164 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6165 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6168 struct sched_group
*sg
;
6170 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6171 init_numa_sched_groups_power(sg
);
6175 /* Attach the domains */
6176 for_each_cpu_mask(i
, *cpu_map
) {
6177 struct sched_domain
*sd
;
6178 #ifdef CONFIG_SCHED_SMT
6179 sd
= &per_cpu(cpu_domains
, i
);
6180 #elif defined(CONFIG_SCHED_MC)
6181 sd
= &per_cpu(core_domains
, i
);
6183 sd
= &per_cpu(phys_domains
, i
);
6185 cpu_attach_domain(sd
, i
);
6192 free_sched_groups(cpu_map
);
6197 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6199 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6201 cpumask_t cpu_default_map
;
6205 * Setup mask for cpus without special case scheduling requirements.
6206 * For now this just excludes isolated cpus, but could be used to
6207 * exclude other special cases in the future.
6209 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6211 err
= build_sched_domains(&cpu_default_map
);
6216 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6218 free_sched_groups(cpu_map
);
6222 * Detach sched domains from a group of cpus specified in cpu_map
6223 * These cpus will now be attached to the NULL domain
6225 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6229 for_each_cpu_mask(i
, *cpu_map
)
6230 cpu_attach_domain(NULL
, i
);
6231 synchronize_sched();
6232 arch_destroy_sched_domains(cpu_map
);
6236 * Partition sched domains as specified by the cpumasks below.
6237 * This attaches all cpus from the cpumasks to the NULL domain,
6238 * waits for a RCU quiescent period, recalculates sched
6239 * domain information and then attaches them back to the
6240 * correct sched domains
6241 * Call with hotplug lock held
6243 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6245 cpumask_t change_map
;
6248 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6249 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6250 cpus_or(change_map
, *partition1
, *partition2
);
6252 /* Detach sched domains from all of the affected cpus */
6253 detach_destroy_domains(&change_map
);
6254 if (!cpus_empty(*partition1
))
6255 err
= build_sched_domains(partition1
);
6256 if (!err
&& !cpus_empty(*partition2
))
6257 err
= build_sched_domains(partition2
);
6262 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6263 static int arch_reinit_sched_domains(void)
6267 mutex_lock(&sched_hotcpu_mutex
);
6268 detach_destroy_domains(&cpu_online_map
);
6269 err
= arch_init_sched_domains(&cpu_online_map
);
6270 mutex_unlock(&sched_hotcpu_mutex
);
6275 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6279 if (buf
[0] != '0' && buf
[0] != '1')
6283 sched_smt_power_savings
= (buf
[0] == '1');
6285 sched_mc_power_savings
= (buf
[0] == '1');
6287 ret
= arch_reinit_sched_domains();
6289 return ret
? ret
: count
;
6292 #ifdef CONFIG_SCHED_MC
6293 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6295 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6297 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6298 const char *buf
, size_t count
)
6300 return sched_power_savings_store(buf
, count
, 0);
6302 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6303 sched_mc_power_savings_store
);
6306 #ifdef CONFIG_SCHED_SMT
6307 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6309 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6311 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6312 const char *buf
, size_t count
)
6314 return sched_power_savings_store(buf
, count
, 1);
6316 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6317 sched_smt_power_savings_store
);
6320 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6324 #ifdef CONFIG_SCHED_SMT
6326 err
= sysfs_create_file(&cls
->kset
.kobj
,
6327 &attr_sched_smt_power_savings
.attr
);
6329 #ifdef CONFIG_SCHED_MC
6330 if (!err
&& mc_capable())
6331 err
= sysfs_create_file(&cls
->kset
.kobj
,
6332 &attr_sched_mc_power_savings
.attr
);
6339 * Force a reinitialization of the sched domains hierarchy. The domains
6340 * and groups cannot be updated in place without racing with the balancing
6341 * code, so we temporarily attach all running cpus to the NULL domain
6342 * which will prevent rebalancing while the sched domains are recalculated.
6344 static int update_sched_domains(struct notifier_block
*nfb
,
6345 unsigned long action
, void *hcpu
)
6348 case CPU_UP_PREPARE
:
6349 case CPU_UP_PREPARE_FROZEN
:
6350 case CPU_DOWN_PREPARE
:
6351 case CPU_DOWN_PREPARE_FROZEN
:
6352 detach_destroy_domains(&cpu_online_map
);
6355 case CPU_UP_CANCELED
:
6356 case CPU_UP_CANCELED_FROZEN
:
6357 case CPU_DOWN_FAILED
:
6358 case CPU_DOWN_FAILED_FROZEN
:
6360 case CPU_ONLINE_FROZEN
:
6362 case CPU_DEAD_FROZEN
:
6364 * Fall through and re-initialise the domains.
6371 /* The hotplug lock is already held by cpu_up/cpu_down */
6372 arch_init_sched_domains(&cpu_online_map
);
6377 void __init
sched_init_smp(void)
6379 cpumask_t non_isolated_cpus
;
6381 mutex_lock(&sched_hotcpu_mutex
);
6382 arch_init_sched_domains(&cpu_online_map
);
6383 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6384 if (cpus_empty(non_isolated_cpus
))
6385 cpu_set(smp_processor_id(), non_isolated_cpus
);
6386 mutex_unlock(&sched_hotcpu_mutex
);
6387 /* XXX: Theoretical race here - CPU may be hotplugged now */
6388 hotcpu_notifier(update_sched_domains
, 0);
6390 init_sched_domain_sysctl();
6392 /* Move init over to a non-isolated CPU */
6393 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6397 void __init
sched_init_smp(void)
6400 #endif /* CONFIG_SMP */
6402 int in_sched_functions(unsigned long addr
)
6404 /* Linker adds these: start and end of __sched functions */
6405 extern char __sched_text_start
[], __sched_text_end
[];
6407 return in_lock_functions(addr
) ||
6408 (addr
>= (unsigned long)__sched_text_start
6409 && addr
< (unsigned long)__sched_text_end
);
6412 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6414 cfs_rq
->tasks_timeline
= RB_ROOT
;
6415 #ifdef CONFIG_FAIR_GROUP_SCHED
6418 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6421 void __init
sched_init(void)
6423 int highest_cpu
= 0;
6426 for_each_possible_cpu(i
) {
6427 struct rt_prio_array
*array
;
6431 spin_lock_init(&rq
->lock
);
6432 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6435 init_cfs_rq(&rq
->cfs
, rq
);
6436 #ifdef CONFIG_FAIR_GROUP_SCHED
6437 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6439 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6440 struct sched_entity
*se
=
6441 &per_cpu(init_sched_entity
, i
);
6443 init_cfs_rq_p
[i
] = cfs_rq
;
6444 init_cfs_rq(cfs_rq
, rq
);
6445 cfs_rq
->tg
= &init_task_group
;
6446 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6447 &rq
->leaf_cfs_rq_list
);
6449 init_sched_entity_p
[i
] = se
;
6450 se
->cfs_rq
= &rq
->cfs
;
6452 se
->load
.weight
= init_task_group_load
;
6453 se
->load
.inv_weight
=
6454 div64_64(1ULL<<32, init_task_group_load
);
6457 init_task_group
.shares
= init_task_group_load
;
6458 spin_lock_init(&init_task_group
.lock
);
6461 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6462 rq
->cpu_load
[j
] = 0;
6465 rq
->active_balance
= 0;
6466 rq
->next_balance
= jiffies
;
6469 rq
->migration_thread
= NULL
;
6470 INIT_LIST_HEAD(&rq
->migration_queue
);
6472 atomic_set(&rq
->nr_iowait
, 0);
6474 array
= &rq
->rt
.active
;
6475 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6476 INIT_LIST_HEAD(array
->queue
+ j
);
6477 __clear_bit(j
, array
->bitmap
);
6480 /* delimiter for bitsearch: */
6481 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6484 set_load_weight(&init_task
);
6486 #ifdef CONFIG_PREEMPT_NOTIFIERS
6487 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6491 nr_cpu_ids
= highest_cpu
+ 1;
6492 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6495 #ifdef CONFIG_RT_MUTEXES
6496 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6500 * The boot idle thread does lazy MMU switching as well:
6502 atomic_inc(&init_mm
.mm_count
);
6503 enter_lazy_tlb(&init_mm
, current
);
6506 * Make us the idle thread. Technically, schedule() should not be
6507 * called from this thread, however somewhere below it might be,
6508 * but because we are the idle thread, we just pick up running again
6509 * when this runqueue becomes "idle".
6511 init_idle(current
, smp_processor_id());
6513 * During early bootup we pretend to be a normal task:
6515 current
->sched_class
= &fair_sched_class
;
6518 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6519 void __might_sleep(char *file
, int line
)
6522 static unsigned long prev_jiffy
; /* ratelimiting */
6524 if ((in_atomic() || irqs_disabled()) &&
6525 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6526 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6528 prev_jiffy
= jiffies
;
6529 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6530 " context at %s:%d\n", file
, line
);
6531 printk("in_atomic():%d, irqs_disabled():%d\n",
6532 in_atomic(), irqs_disabled());
6533 debug_show_held_locks(current
);
6534 if (irqs_disabled())
6535 print_irqtrace_events(current
);
6540 EXPORT_SYMBOL(__might_sleep
);
6543 #ifdef CONFIG_MAGIC_SYSRQ
6544 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6547 update_rq_clock(rq
);
6548 on_rq
= p
->se
.on_rq
;
6550 deactivate_task(rq
, p
, 0);
6551 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6553 activate_task(rq
, p
, 0);
6554 resched_task(rq
->curr
);
6558 void normalize_rt_tasks(void)
6560 struct task_struct
*g
, *p
;
6561 unsigned long flags
;
6564 read_lock_irq(&tasklist_lock
);
6565 do_each_thread(g
, p
) {
6566 p
->se
.exec_start
= 0;
6567 #ifdef CONFIG_SCHEDSTATS
6568 p
->se
.wait_start
= 0;
6569 p
->se
.sleep_start
= 0;
6570 p
->se
.block_start
= 0;
6572 task_rq(p
)->clock
= 0;
6576 * Renice negative nice level userspace
6579 if (TASK_NICE(p
) < 0 && p
->mm
)
6580 set_user_nice(p
, 0);
6584 spin_lock_irqsave(&p
->pi_lock
, flags
);
6585 rq
= __task_rq_lock(p
);
6587 if (!is_migration_thread(p
, rq
))
6588 normalize_task(rq
, p
);
6590 __task_rq_unlock(rq
);
6591 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6592 } while_each_thread(g
, p
);
6594 read_unlock_irq(&tasklist_lock
);
6597 #endif /* CONFIG_MAGIC_SYSRQ */
6601 * These functions are only useful for the IA64 MCA handling.
6603 * They can only be called when the whole system has been
6604 * stopped - every CPU needs to be quiescent, and no scheduling
6605 * activity can take place. Using them for anything else would
6606 * be a serious bug, and as a result, they aren't even visible
6607 * under any other configuration.
6611 * curr_task - return the current task for a given cpu.
6612 * @cpu: the processor in question.
6614 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6616 struct task_struct
*curr_task(int cpu
)
6618 return cpu_curr(cpu
);
6622 * set_curr_task - set the current task for a given cpu.
6623 * @cpu: the processor in question.
6624 * @p: the task pointer to set.
6626 * Description: This function must only be used when non-maskable interrupts
6627 * are serviced on a separate stack. It allows the architecture to switch the
6628 * notion of the current task on a cpu in a non-blocking manner. This function
6629 * must be called with all CPU's synchronized, and interrupts disabled, the
6630 * and caller must save the original value of the current task (see
6631 * curr_task() above) and restore that value before reenabling interrupts and
6632 * re-starting the system.
6634 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6636 void set_curr_task(int cpu
, struct task_struct
*p
)
6643 #ifdef CONFIG_FAIR_GROUP_SCHED
6645 /* allocate runqueue etc for a new task group */
6646 struct task_group
*sched_create_group(void)
6648 struct task_group
*tg
;
6649 struct cfs_rq
*cfs_rq
;
6650 struct sched_entity
*se
;
6654 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6656 return ERR_PTR(-ENOMEM
);
6658 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
6661 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
6665 for_each_possible_cpu(i
) {
6668 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
6673 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
6678 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
6679 memset(se
, 0, sizeof(struct sched_entity
));
6681 tg
->cfs_rq
[i
] = cfs_rq
;
6682 init_cfs_rq(cfs_rq
, rq
);
6686 se
->cfs_rq
= &rq
->cfs
;
6688 se
->load
.weight
= NICE_0_LOAD
;
6689 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6693 for_each_possible_cpu(i
) {
6695 cfs_rq
= tg
->cfs_rq
[i
];
6696 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6699 tg
->shares
= NICE_0_LOAD
;
6700 spin_lock_init(&tg
->lock
);
6705 for_each_possible_cpu(i
) {
6707 kfree(tg
->cfs_rq
[i
]);
6715 return ERR_PTR(-ENOMEM
);
6718 /* rcu callback to free various structures associated with a task group */
6719 static void free_sched_group(struct rcu_head
*rhp
)
6721 struct cfs_rq
*cfs_rq
= container_of(rhp
, struct cfs_rq
, rcu
);
6722 struct task_group
*tg
= cfs_rq
->tg
;
6723 struct sched_entity
*se
;
6726 /* now it should be safe to free those cfs_rqs */
6727 for_each_possible_cpu(i
) {
6728 cfs_rq
= tg
->cfs_rq
[i
];
6740 /* Destroy runqueue etc associated with a task group */
6741 void sched_destroy_group(struct task_group
*tg
)
6743 struct cfs_rq
*cfs_rq
;
6746 for_each_possible_cpu(i
) {
6747 cfs_rq
= tg
->cfs_rq
[i
];
6748 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
6751 cfs_rq
= tg
->cfs_rq
[0];
6753 /* wait for possible concurrent references to cfs_rqs complete */
6754 call_rcu(&cfs_rq
->rcu
, free_sched_group
);
6757 /* change task's runqueue when it moves between groups.
6758 * The caller of this function should have put the task in its new group
6759 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6760 * reflect its new group.
6762 void sched_move_task(struct task_struct
*tsk
)
6765 unsigned long flags
;
6768 rq
= task_rq_lock(tsk
, &flags
);
6770 if (tsk
->sched_class
!= &fair_sched_class
)
6773 update_rq_clock(rq
);
6775 running
= task_running(rq
, tsk
);
6776 on_rq
= tsk
->se
.on_rq
;
6779 dequeue_task(rq
, tsk
, 0);
6780 if (unlikely(running
))
6781 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6784 set_task_cfs_rq(tsk
);
6787 if (unlikely(running
))
6788 tsk
->sched_class
->set_curr_task(rq
);
6789 enqueue_task(rq
, tsk
, 0);
6793 task_rq_unlock(rq
, &flags
);
6796 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
6798 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
6799 struct rq
*rq
= cfs_rq
->rq
;
6802 spin_lock_irq(&rq
->lock
);
6806 dequeue_entity(cfs_rq
, se
, 0);
6808 se
->load
.weight
= shares
;
6809 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
6812 enqueue_entity(cfs_rq
, se
, 0);
6814 spin_unlock_irq(&rq
->lock
);
6817 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6821 spin_lock(&tg
->lock
);
6822 if (tg
->shares
== shares
)
6825 tg
->shares
= shares
;
6826 for_each_possible_cpu(i
)
6827 set_se_shares(tg
->se
[i
], shares
);
6830 spin_unlock(&tg
->lock
);
6834 unsigned long sched_group_shares(struct task_group
*tg
)
6839 #endif /* CONFIG_FAIR_GROUP_SCHED */