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.
273 * nr_running and cpu_load should be in the same cacheline because
274 * remote CPUs use both these fields when doing load calculation.
276 unsigned long nr_running
;
277 #define CPU_LOAD_IDX_MAX 5
278 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
279 unsigned char idle_at_tick
;
281 unsigned char in_nohz_recently
;
283 /* capture load from *all* tasks on this cpu: */
284 struct load_weight load
;
285 unsigned long nr_load_updates
;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* list of leaf cfs_rq on this cpu: */
291 struct list_head leaf_cfs_rq_list
;
296 * This is part of a global counter where only the total sum
297 * over all CPUs matters. A task can increase this counter on
298 * one CPU and if it got migrated afterwards it may decrease
299 * it on another CPU. Always updated under the runqueue lock:
301 unsigned long nr_uninterruptible
;
303 struct task_struct
*curr
, *idle
;
304 unsigned long next_balance
;
305 struct mm_struct
*prev_mm
;
307 u64 clock
, prev_clock_raw
;
310 unsigned int clock_warps
, clock_overflows
;
312 unsigned int clock_deep_idle_events
;
318 struct sched_domain
*sd
;
320 /* For active balancing */
323 /* cpu of this runqueue: */
326 struct task_struct
*migration_thread
;
327 struct list_head migration_queue
;
330 #ifdef CONFIG_SCHEDSTATS
332 struct sched_info rq_sched_info
;
334 /* sys_sched_yield() stats */
335 unsigned int yld_exp_empty
;
336 unsigned int yld_act_empty
;
337 unsigned int yld_both_empty
;
338 unsigned int yld_count
;
340 /* schedule() stats */
341 unsigned int sched_switch
;
342 unsigned int sched_count
;
343 unsigned int sched_goidle
;
345 /* try_to_wake_up() stats */
346 unsigned int ttwu_count
;
347 unsigned int ttwu_local
;
350 unsigned int bkl_count
;
352 struct lock_class_key rq_lock_key
;
355 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
356 static DEFINE_MUTEX(sched_hotcpu_mutex
);
358 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
360 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
363 static inline int cpu_of(struct rq
*rq
)
373 * Update the per-runqueue clock, as finegrained as the platform can give
374 * us, but without assuming monotonicity, etc.:
376 static void __update_rq_clock(struct rq
*rq
)
378 u64 prev_raw
= rq
->prev_clock_raw
;
379 u64 now
= sched_clock();
380 s64 delta
= now
- prev_raw
;
381 u64 clock
= rq
->clock
;
383 #ifdef CONFIG_SCHED_DEBUG
384 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
387 * Protect against sched_clock() occasionally going backwards:
389 if (unlikely(delta
< 0)) {
394 * Catch too large forward jumps too:
396 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
397 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
398 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
401 rq
->clock_overflows
++;
403 if (unlikely(delta
> rq
->clock_max_delta
))
404 rq
->clock_max_delta
= delta
;
409 rq
->prev_clock_raw
= now
;
413 static void update_rq_clock(struct rq
*rq
)
415 if (likely(smp_processor_id() == cpu_of(rq
)))
416 __update_rq_clock(rq
);
420 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
421 * See detach_destroy_domains: synchronize_sched for details.
423 * The domain tree of any CPU may only be accessed from within
424 * preempt-disabled sections.
426 #define for_each_domain(cpu, __sd) \
427 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
429 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
430 #define this_rq() (&__get_cpu_var(runqueues))
431 #define task_rq(p) cpu_rq(task_cpu(p))
432 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
435 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
437 #ifdef CONFIG_SCHED_DEBUG
438 # define const_debug __read_mostly
440 # define const_debug static const
444 * Debugging: various feature bits
447 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
448 SCHED_FEAT_START_DEBIT
= 2,
449 SCHED_FEAT_TREE_AVG
= 4,
450 SCHED_FEAT_APPROX_AVG
= 8,
451 SCHED_FEAT_WAKEUP_PREEMPT
= 16,
452 SCHED_FEAT_PREEMPT_RESTRICT
= 32,
455 const_debug
unsigned int sysctl_sched_features
=
456 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
457 SCHED_FEAT_START_DEBIT
* 1 |
458 SCHED_FEAT_TREE_AVG
* 0 |
459 SCHED_FEAT_APPROX_AVG
* 0 |
460 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
461 SCHED_FEAT_PREEMPT_RESTRICT
* 1;
463 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
466 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
467 * clock constructed from sched_clock():
469 unsigned long long cpu_clock(int cpu
)
471 unsigned long long now
;
475 local_irq_save(flags
);
479 local_irq_restore(flags
);
483 EXPORT_SYMBOL_GPL(cpu_clock
);
485 #ifndef prepare_arch_switch
486 # define prepare_arch_switch(next) do { } while (0)
488 #ifndef finish_arch_switch
489 # define finish_arch_switch(prev) do { } while (0)
492 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
493 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
495 return rq
->curr
== p
;
498 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
502 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
504 #ifdef CONFIG_DEBUG_SPINLOCK
505 /* this is a valid case when another task releases the spinlock */
506 rq
->lock
.owner
= current
;
509 * If we are tracking spinlock dependencies then we have to
510 * fix up the runqueue lock - which gets 'carried over' from
513 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
515 spin_unlock_irq(&rq
->lock
);
518 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
519 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
524 return rq
->curr
== p
;
528 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
532 * We can optimise this out completely for !SMP, because the
533 * SMP rebalancing from interrupt is the only thing that cares
538 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
539 spin_unlock_irq(&rq
->lock
);
541 spin_unlock(&rq
->lock
);
545 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
549 * After ->oncpu is cleared, the task can be moved to a different CPU.
550 * We must ensure this doesn't happen until the switch is completely
556 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
560 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
563 * __task_rq_lock - lock the runqueue a given task resides on.
564 * Must be called interrupts disabled.
566 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
570 struct rq
*rq
= task_rq(p
);
571 spin_lock(&rq
->lock
);
572 if (likely(rq
== task_rq(p
)))
574 spin_unlock(&rq
->lock
);
579 * task_rq_lock - lock the runqueue a given task resides on and disable
580 * interrupts. Note the ordering: we can safely lookup the task_rq without
581 * explicitly disabling preemption.
583 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
589 local_irq_save(*flags
);
591 spin_lock(&rq
->lock
);
592 if (likely(rq
== task_rq(p
)))
594 spin_unlock_irqrestore(&rq
->lock
, *flags
);
598 static void __task_rq_unlock(struct rq
*rq
)
601 spin_unlock(&rq
->lock
);
604 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
607 spin_unlock_irqrestore(&rq
->lock
, *flags
);
611 * this_rq_lock - lock this runqueue and disable interrupts.
613 static struct rq
*this_rq_lock(void)
620 spin_lock(&rq
->lock
);
626 * We are going deep-idle (irqs are disabled):
628 void sched_clock_idle_sleep_event(void)
630 struct rq
*rq
= cpu_rq(smp_processor_id());
632 spin_lock(&rq
->lock
);
633 __update_rq_clock(rq
);
634 spin_unlock(&rq
->lock
);
635 rq
->clock_deep_idle_events
++;
637 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
640 * We just idled delta nanoseconds (called with irqs disabled):
642 void sched_clock_idle_wakeup_event(u64 delta_ns
)
644 struct rq
*rq
= cpu_rq(smp_processor_id());
645 u64 now
= sched_clock();
647 rq
->idle_clock
+= delta_ns
;
649 * Override the previous timestamp and ignore all
650 * sched_clock() deltas that occured while we idled,
651 * and use the PM-provided delta_ns to advance the
654 spin_lock(&rq
->lock
);
655 rq
->prev_clock_raw
= now
;
656 rq
->clock
+= delta_ns
;
657 spin_unlock(&rq
->lock
);
659 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
662 * resched_task - mark a task 'to be rescheduled now'.
664 * On UP this means the setting of the need_resched flag, on SMP it
665 * might also involve a cross-CPU call to trigger the scheduler on
670 #ifndef tsk_is_polling
671 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
674 static void resched_task(struct task_struct
*p
)
678 assert_spin_locked(&task_rq(p
)->lock
);
680 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
683 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
686 if (cpu
== smp_processor_id())
689 /* NEED_RESCHED must be visible before we test polling */
691 if (!tsk_is_polling(p
))
692 smp_send_reschedule(cpu
);
695 static void resched_cpu(int cpu
)
697 struct rq
*rq
= cpu_rq(cpu
);
700 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
702 resched_task(cpu_curr(cpu
));
703 spin_unlock_irqrestore(&rq
->lock
, flags
);
706 static inline void resched_task(struct task_struct
*p
)
708 assert_spin_locked(&task_rq(p
)->lock
);
709 set_tsk_need_resched(p
);
713 #if BITS_PER_LONG == 32
714 # define WMULT_CONST (~0UL)
716 # define WMULT_CONST (1UL << 32)
719 #define WMULT_SHIFT 32
722 * Shift right and round:
724 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
727 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
728 struct load_weight
*lw
)
732 if (unlikely(!lw
->inv_weight
))
733 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
735 tmp
= (u64
)delta_exec
* weight
;
737 * Check whether we'd overflow the 64-bit multiplication:
739 if (unlikely(tmp
> WMULT_CONST
))
740 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
743 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
745 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
748 static inline unsigned long
749 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
751 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
754 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
759 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
765 * To aid in avoiding the subversion of "niceness" due to uneven distribution
766 * of tasks with abnormal "nice" values across CPUs the contribution that
767 * each task makes to its run queue's load is weighted according to its
768 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
769 * scaled version of the new time slice allocation that they receive on time
773 #define WEIGHT_IDLEPRIO 2
774 #define WMULT_IDLEPRIO (1 << 31)
777 * Nice levels are multiplicative, with a gentle 10% change for every
778 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
779 * nice 1, it will get ~10% less CPU time than another CPU-bound task
780 * that remained on nice 0.
782 * The "10% effect" is relative and cumulative: from _any_ nice level,
783 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
784 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
785 * If a task goes up by ~10% and another task goes down by ~10% then
786 * the relative distance between them is ~25%.)
788 static const int prio_to_weight
[40] = {
789 /* -20 */ 88761, 71755, 56483, 46273, 36291,
790 /* -15 */ 29154, 23254, 18705, 14949, 11916,
791 /* -10 */ 9548, 7620, 6100, 4904, 3906,
792 /* -5 */ 3121, 2501, 1991, 1586, 1277,
793 /* 0 */ 1024, 820, 655, 526, 423,
794 /* 5 */ 335, 272, 215, 172, 137,
795 /* 10 */ 110, 87, 70, 56, 45,
796 /* 15 */ 36, 29, 23, 18, 15,
800 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
802 * In cases where the weight does not change often, we can use the
803 * precalculated inverse to speed up arithmetics by turning divisions
804 * into multiplications:
806 static const u32 prio_to_wmult
[40] = {
807 /* -20 */ 48388, 59856, 76040, 92818, 118348,
808 /* -15 */ 147320, 184698, 229616, 287308, 360437,
809 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
810 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
811 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
812 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
813 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
814 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
817 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
820 * runqueue iterator, to support SMP load-balancing between different
821 * scheduling classes, without having to expose their internal data
822 * structures to the load-balancing proper:
826 struct task_struct
*(*start
)(void *);
827 struct task_struct
*(*next
)(void *);
830 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
831 unsigned long max_nr_move
, unsigned long max_load_move
,
832 struct sched_domain
*sd
, enum cpu_idle_type idle
,
833 int *all_pinned
, unsigned long *load_moved
,
834 int *this_best_prio
, struct rq_iterator
*iterator
);
836 #include "sched_stats.h"
837 #include "sched_idletask.c"
838 #include "sched_fair.c"
839 #include "sched_rt.c"
840 #ifdef CONFIG_SCHED_DEBUG
841 # include "sched_debug.c"
844 #define sched_class_highest (&rt_sched_class)
847 * Update delta_exec, delta_fair fields for rq.
849 * delta_fair clock advances at a rate inversely proportional to
850 * total load (rq->load.weight) on the runqueue, while
851 * delta_exec advances at the same rate as wall-clock (provided
854 * delta_exec / delta_fair is a measure of the (smoothened) load on this
855 * runqueue over any given interval. This (smoothened) load is used
856 * during load balance.
858 * This function is called /before/ updating rq->load
859 * and when switching tasks.
861 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
863 update_load_add(&rq
->load
, p
->se
.load
.weight
);
866 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
868 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
871 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
877 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
883 static void set_load_weight(struct task_struct
*p
)
885 if (task_has_rt_policy(p
)) {
886 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
887 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
892 * SCHED_IDLE tasks get minimal weight:
894 if (p
->policy
== SCHED_IDLE
) {
895 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
896 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
900 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
901 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
904 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
906 sched_info_queued(p
);
907 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
911 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
913 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
918 * __normal_prio - return the priority that is based on the static prio
920 static inline int __normal_prio(struct task_struct
*p
)
922 return p
->static_prio
;
926 * Calculate the expected normal priority: i.e. priority
927 * without taking RT-inheritance into account. Might be
928 * boosted by interactivity modifiers. Changes upon fork,
929 * setprio syscalls, and whenever the interactivity
930 * estimator recalculates.
932 static inline int normal_prio(struct task_struct
*p
)
936 if (task_has_rt_policy(p
))
937 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
939 prio
= __normal_prio(p
);
944 * Calculate the current priority, i.e. the priority
945 * taken into account by the scheduler. This value might
946 * be boosted by RT tasks, or might be boosted by
947 * interactivity modifiers. Will be RT if the task got
948 * RT-boosted. If not then it returns p->normal_prio.
950 static int effective_prio(struct task_struct
*p
)
952 p
->normal_prio
= normal_prio(p
);
954 * If we are RT tasks or we were boosted to RT priority,
955 * keep the priority unchanged. Otherwise, update priority
956 * to the normal priority:
958 if (!rt_prio(p
->prio
))
959 return p
->normal_prio
;
964 * activate_task - move a task to the runqueue.
966 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
968 if (p
->state
== TASK_UNINTERRUPTIBLE
)
969 rq
->nr_uninterruptible
--;
971 enqueue_task(rq
, p
, wakeup
);
972 inc_nr_running(p
, rq
);
976 * deactivate_task - remove a task from the runqueue.
978 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
980 if (p
->state
== TASK_UNINTERRUPTIBLE
)
981 rq
->nr_uninterruptible
++;
983 dequeue_task(rq
, p
, sleep
);
984 dec_nr_running(p
, rq
);
988 * task_curr - is this task currently executing on a CPU?
989 * @p: the task in question.
991 inline int task_curr(const struct task_struct
*p
)
993 return cpu_curr(task_cpu(p
)) == p
;
996 /* Used instead of source_load when we know the type == 0 */
997 unsigned long weighted_cpuload(const int cpu
)
999 return cpu_rq(cpu
)->load
.weight
;
1002 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1005 task_thread_info(p
)->cpu
= cpu
;
1013 * Is this task likely cache-hot:
1016 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1020 if (p
->sched_class
!= &fair_sched_class
)
1023 if (sysctl_sched_migration_cost
== -1)
1025 if (sysctl_sched_migration_cost
== 0)
1028 delta
= now
- p
->se
.exec_start
;
1030 return delta
< (s64
)sysctl_sched_migration_cost
;
1034 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1036 int old_cpu
= task_cpu(p
);
1037 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1038 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1039 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1042 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1044 #ifdef CONFIG_SCHEDSTATS
1045 if (p
->se
.wait_start
)
1046 p
->se
.wait_start
-= clock_offset
;
1047 if (p
->se
.sleep_start
)
1048 p
->se
.sleep_start
-= clock_offset
;
1049 if (p
->se
.block_start
)
1050 p
->se
.block_start
-= clock_offset
;
1051 if (old_cpu
!= new_cpu
) {
1052 schedstat_inc(p
, se
.nr_migrations
);
1053 if (task_hot(p
, old_rq
->clock
, NULL
))
1054 schedstat_inc(p
, se
.nr_forced2_migrations
);
1057 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1058 new_cfsrq
->min_vruntime
;
1060 __set_task_cpu(p
, new_cpu
);
1063 struct migration_req
{
1064 struct list_head list
;
1066 struct task_struct
*task
;
1069 struct completion done
;
1073 * The task's runqueue lock must be held.
1074 * Returns true if you have to wait for migration thread.
1077 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1079 struct rq
*rq
= task_rq(p
);
1082 * If the task is not on a runqueue (and not running), then
1083 * it is sufficient to simply update the task's cpu field.
1085 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1086 set_task_cpu(p
, dest_cpu
);
1090 init_completion(&req
->done
);
1092 req
->dest_cpu
= dest_cpu
;
1093 list_add(&req
->list
, &rq
->migration_queue
);
1099 * wait_task_inactive - wait for a thread to unschedule.
1101 * The caller must ensure that the task *will* unschedule sometime soon,
1102 * else this function might spin for a *long* time. This function can't
1103 * be called with interrupts off, or it may introduce deadlock with
1104 * smp_call_function() if an IPI is sent by the same process we are
1105 * waiting to become inactive.
1107 void wait_task_inactive(struct task_struct
*p
)
1109 unsigned long flags
;
1115 * We do the initial early heuristics without holding
1116 * any task-queue locks at all. We'll only try to get
1117 * the runqueue lock when things look like they will
1123 * If the task is actively running on another CPU
1124 * still, just relax and busy-wait without holding
1127 * NOTE! Since we don't hold any locks, it's not
1128 * even sure that "rq" stays as the right runqueue!
1129 * But we don't care, since "task_running()" will
1130 * return false if the runqueue has changed and p
1131 * is actually now running somewhere else!
1133 while (task_running(rq
, p
))
1137 * Ok, time to look more closely! We need the rq
1138 * lock now, to be *sure*. If we're wrong, we'll
1139 * just go back and repeat.
1141 rq
= task_rq_lock(p
, &flags
);
1142 running
= task_running(rq
, p
);
1143 on_rq
= p
->se
.on_rq
;
1144 task_rq_unlock(rq
, &flags
);
1147 * Was it really running after all now that we
1148 * checked with the proper locks actually held?
1150 * Oops. Go back and try again..
1152 if (unlikely(running
)) {
1158 * It's not enough that it's not actively running,
1159 * it must be off the runqueue _entirely_, and not
1162 * So if it wa still runnable (but just not actively
1163 * running right now), it's preempted, and we should
1164 * yield - it could be a while.
1166 if (unlikely(on_rq
)) {
1167 schedule_timeout_uninterruptible(1);
1172 * Ahh, all good. It wasn't running, and it wasn't
1173 * runnable, which means that it will never become
1174 * running in the future either. We're all done!
1181 * kick_process - kick a running thread to enter/exit the kernel
1182 * @p: the to-be-kicked thread
1184 * Cause a process which is running on another CPU to enter
1185 * kernel-mode, without any delay. (to get signals handled.)
1187 * NOTE: this function doesnt have to take the runqueue lock,
1188 * because all it wants to ensure is that the remote task enters
1189 * the kernel. If the IPI races and the task has been migrated
1190 * to another CPU then no harm is done and the purpose has been
1193 void kick_process(struct task_struct
*p
)
1199 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1200 smp_send_reschedule(cpu
);
1205 * Return a low guess at the load of a migration-source cpu weighted
1206 * according to the scheduling class and "nice" value.
1208 * We want to under-estimate the load of migration sources, to
1209 * balance conservatively.
1211 static unsigned long source_load(int cpu
, int type
)
1213 struct rq
*rq
= cpu_rq(cpu
);
1214 unsigned long total
= weighted_cpuload(cpu
);
1219 return min(rq
->cpu_load
[type
-1], total
);
1223 * Return a high guess at the load of a migration-target cpu weighted
1224 * according to the scheduling class and "nice" value.
1226 static unsigned long target_load(int cpu
, int type
)
1228 struct rq
*rq
= cpu_rq(cpu
);
1229 unsigned long total
= weighted_cpuload(cpu
);
1234 return max(rq
->cpu_load
[type
-1], total
);
1238 * Return the average load per task on the cpu's run queue
1240 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1242 struct rq
*rq
= cpu_rq(cpu
);
1243 unsigned long total
= weighted_cpuload(cpu
);
1244 unsigned long n
= rq
->nr_running
;
1246 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1250 * find_idlest_group finds and returns the least busy CPU group within the
1253 static struct sched_group
*
1254 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1256 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1257 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1258 int load_idx
= sd
->forkexec_idx
;
1259 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1262 unsigned long load
, avg_load
;
1266 /* Skip over this group if it has no CPUs allowed */
1267 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1270 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1272 /* Tally up the load of all CPUs in the group */
1275 for_each_cpu_mask(i
, group
->cpumask
) {
1276 /* Bias balancing toward cpus of our domain */
1278 load
= source_load(i
, load_idx
);
1280 load
= target_load(i
, load_idx
);
1285 /* Adjust by relative CPU power of the group */
1286 avg_load
= sg_div_cpu_power(group
,
1287 avg_load
* SCHED_LOAD_SCALE
);
1290 this_load
= avg_load
;
1292 } else if (avg_load
< min_load
) {
1293 min_load
= avg_load
;
1296 } while (group
= group
->next
, group
!= sd
->groups
);
1298 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1304 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1307 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1310 unsigned long load
, min_load
= ULONG_MAX
;
1314 /* Traverse only the allowed CPUs */
1315 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1317 for_each_cpu_mask(i
, tmp
) {
1318 load
= weighted_cpuload(i
);
1320 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1330 * sched_balance_self: balance the current task (running on cpu) in domains
1331 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1334 * Balance, ie. select the least loaded group.
1336 * Returns the target CPU number, or the same CPU if no balancing is needed.
1338 * preempt must be disabled.
1340 static int sched_balance_self(int cpu
, int flag
)
1342 struct task_struct
*t
= current
;
1343 struct sched_domain
*tmp
, *sd
= NULL
;
1345 for_each_domain(cpu
, tmp
) {
1347 * If power savings logic is enabled for a domain, stop there.
1349 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1351 if (tmp
->flags
& flag
)
1357 struct sched_group
*group
;
1358 int new_cpu
, weight
;
1360 if (!(sd
->flags
& flag
)) {
1366 group
= find_idlest_group(sd
, t
, cpu
);
1372 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1373 if (new_cpu
== -1 || new_cpu
== cpu
) {
1374 /* Now try balancing at a lower domain level of cpu */
1379 /* Now try balancing at a lower domain level of new_cpu */
1382 weight
= cpus_weight(span
);
1383 for_each_domain(cpu
, tmp
) {
1384 if (weight
<= cpus_weight(tmp
->span
))
1386 if (tmp
->flags
& flag
)
1389 /* while loop will break here if sd == NULL */
1395 #endif /* CONFIG_SMP */
1398 * wake_idle() will wake a task on an idle cpu if task->cpu is
1399 * not idle and an idle cpu is available. The span of cpus to
1400 * search starts with cpus closest then further out as needed,
1401 * so we always favor a closer, idle cpu.
1403 * Returns the CPU we should wake onto.
1405 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1406 static int wake_idle(int cpu
, struct task_struct
*p
)
1409 struct sched_domain
*sd
;
1413 * If it is idle, then it is the best cpu to run this task.
1415 * This cpu is also the best, if it has more than one task already.
1416 * Siblings must be also busy(in most cases) as they didn't already
1417 * pickup the extra load from this cpu and hence we need not check
1418 * sibling runqueue info. This will avoid the checks and cache miss
1419 * penalities associated with that.
1421 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1424 for_each_domain(cpu
, sd
) {
1425 if (sd
->flags
& SD_WAKE_IDLE
) {
1426 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1427 for_each_cpu_mask(i
, tmp
) {
1429 if (i
!= task_cpu(p
)) {
1431 se
.nr_wakeups_idle
);
1443 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1450 * try_to_wake_up - wake up a thread
1451 * @p: the to-be-woken-up thread
1452 * @state: the mask of task states that can be woken
1453 * @sync: do a synchronous wakeup?
1455 * Put it on the run-queue if it's not already there. The "current"
1456 * thread is always on the run-queue (except when the actual
1457 * re-schedule is in progress), and as such you're allowed to do
1458 * the simpler "current->state = TASK_RUNNING" to mark yourself
1459 * runnable without the overhead of this.
1461 * returns failure only if the task is already active.
1463 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1465 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1466 unsigned long flags
;
1470 struct sched_domain
*sd
, *this_sd
= NULL
;
1471 unsigned long load
, this_load
;
1475 rq
= task_rq_lock(p
, &flags
);
1476 old_state
= p
->state
;
1477 if (!(old_state
& state
))
1485 this_cpu
= smp_processor_id();
1488 if (unlikely(task_running(rq
, p
)))
1493 schedstat_inc(rq
, ttwu_count
);
1494 if (cpu
== this_cpu
) {
1495 schedstat_inc(rq
, ttwu_local
);
1499 for_each_domain(this_cpu
, sd
) {
1500 if (cpu_isset(cpu
, sd
->span
)) {
1501 schedstat_inc(sd
, ttwu_wake_remote
);
1507 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1511 * Check for affine wakeup and passive balancing possibilities.
1514 int idx
= this_sd
->wake_idx
;
1515 unsigned int imbalance
;
1517 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1519 load
= source_load(cpu
, idx
);
1520 this_load
= target_load(this_cpu
, idx
);
1522 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1524 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1525 unsigned long tl
= this_load
;
1526 unsigned long tl_per_task
;
1529 * Attract cache-cold tasks on sync wakeups:
1531 if (sync
&& !task_hot(p
, rq
->clock
, this_sd
))
1534 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
1535 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1538 * If sync wakeup then subtract the (maximum possible)
1539 * effect of the currently running task from the load
1540 * of the current CPU:
1543 tl
-= current
->se
.load
.weight
;
1546 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1547 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1549 * This domain has SD_WAKE_AFFINE and
1550 * p is cache cold in this domain, and
1551 * there is no bad imbalance.
1553 schedstat_inc(this_sd
, ttwu_move_affine
);
1554 schedstat_inc(p
, se
.nr_wakeups_affine
);
1560 * Start passive balancing when half the imbalance_pct
1563 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1564 if (imbalance
*this_load
<= 100*load
) {
1565 schedstat_inc(this_sd
, ttwu_move_balance
);
1566 schedstat_inc(p
, se
.nr_wakeups_passive
);
1572 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1574 new_cpu
= wake_idle(new_cpu
, p
);
1575 if (new_cpu
!= cpu
) {
1576 set_task_cpu(p
, new_cpu
);
1577 task_rq_unlock(rq
, &flags
);
1578 /* might preempt at this point */
1579 rq
= task_rq_lock(p
, &flags
);
1580 old_state
= p
->state
;
1581 if (!(old_state
& state
))
1586 this_cpu
= smp_processor_id();
1591 #endif /* CONFIG_SMP */
1592 schedstat_inc(p
, se
.nr_wakeups
);
1594 schedstat_inc(p
, se
.nr_wakeups_sync
);
1595 if (orig_cpu
!= cpu
)
1596 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1597 if (cpu
== this_cpu
)
1598 schedstat_inc(p
, se
.nr_wakeups_local
);
1600 schedstat_inc(p
, se
.nr_wakeups_remote
);
1601 update_rq_clock(rq
);
1602 activate_task(rq
, p
, 1);
1603 check_preempt_curr(rq
, p
);
1607 p
->state
= TASK_RUNNING
;
1609 task_rq_unlock(rq
, &flags
);
1614 int fastcall
wake_up_process(struct task_struct
*p
)
1616 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1617 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1619 EXPORT_SYMBOL(wake_up_process
);
1621 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1623 return try_to_wake_up(p
, state
, 0);
1627 * Perform scheduler related setup for a newly forked process p.
1628 * p is forked by current.
1630 * __sched_fork() is basic setup used by init_idle() too:
1632 static void __sched_fork(struct task_struct
*p
)
1634 p
->se
.exec_start
= 0;
1635 p
->se
.sum_exec_runtime
= 0;
1636 p
->se
.prev_sum_exec_runtime
= 0;
1638 #ifdef CONFIG_SCHEDSTATS
1639 p
->se
.wait_start
= 0;
1640 p
->se
.sum_sleep_runtime
= 0;
1641 p
->se
.sleep_start
= 0;
1642 p
->se
.block_start
= 0;
1643 p
->se
.sleep_max
= 0;
1644 p
->se
.block_max
= 0;
1646 p
->se
.slice_max
= 0;
1650 INIT_LIST_HEAD(&p
->run_list
);
1653 #ifdef CONFIG_PREEMPT_NOTIFIERS
1654 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1658 * We mark the process as running here, but have not actually
1659 * inserted it onto the runqueue yet. This guarantees that
1660 * nobody will actually run it, and a signal or other external
1661 * event cannot wake it up and insert it on the runqueue either.
1663 p
->state
= TASK_RUNNING
;
1667 * fork()/clone()-time setup:
1669 void sched_fork(struct task_struct
*p
, int clone_flags
)
1671 int cpu
= get_cpu();
1676 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1678 set_task_cpu(p
, cpu
);
1681 * Make sure we do not leak PI boosting priority to the child:
1683 p
->prio
= current
->normal_prio
;
1684 if (!rt_prio(p
->prio
))
1685 p
->sched_class
= &fair_sched_class
;
1687 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1688 if (likely(sched_info_on()))
1689 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1691 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1694 #ifdef CONFIG_PREEMPT
1695 /* Want to start with kernel preemption disabled. */
1696 task_thread_info(p
)->preempt_count
= 1;
1702 * wake_up_new_task - wake up a newly created task for the first time.
1704 * This function will do some initial scheduler statistics housekeeping
1705 * that must be done for every newly created context, then puts the task
1706 * on the runqueue and wakes it.
1708 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1710 unsigned long flags
;
1713 rq
= task_rq_lock(p
, &flags
);
1714 BUG_ON(p
->state
!= TASK_RUNNING
);
1715 update_rq_clock(rq
);
1717 p
->prio
= effective_prio(p
);
1719 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1720 activate_task(rq
, p
, 0);
1723 * Let the scheduling class do new task startup
1724 * management (if any):
1726 p
->sched_class
->task_new(rq
, p
);
1727 inc_nr_running(p
, rq
);
1729 check_preempt_curr(rq
, p
);
1730 task_rq_unlock(rq
, &flags
);
1733 #ifdef CONFIG_PREEMPT_NOTIFIERS
1736 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1737 * @notifier: notifier struct to register
1739 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1741 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1743 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1746 * preempt_notifier_unregister - no longer interested in preemption notifications
1747 * @notifier: notifier struct to unregister
1749 * This is safe to call from within a preemption notifier.
1751 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1753 hlist_del(¬ifier
->link
);
1755 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1757 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1759 struct preempt_notifier
*notifier
;
1760 struct hlist_node
*node
;
1762 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1763 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1767 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1768 struct task_struct
*next
)
1770 struct preempt_notifier
*notifier
;
1771 struct hlist_node
*node
;
1773 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1774 notifier
->ops
->sched_out(notifier
, next
);
1779 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1784 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1785 struct task_struct
*next
)
1792 * prepare_task_switch - prepare to switch tasks
1793 * @rq: the runqueue preparing to switch
1794 * @prev: the current task that is being switched out
1795 * @next: the task we are going to switch to.
1797 * This is called with the rq lock held and interrupts off. It must
1798 * be paired with a subsequent finish_task_switch after the context
1801 * prepare_task_switch sets up locking and calls architecture specific
1805 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1806 struct task_struct
*next
)
1808 fire_sched_out_preempt_notifiers(prev
, next
);
1809 prepare_lock_switch(rq
, next
);
1810 prepare_arch_switch(next
);
1814 * finish_task_switch - clean up after a task-switch
1815 * @rq: runqueue associated with task-switch
1816 * @prev: the thread we just switched away from.
1818 * finish_task_switch must be called after the context switch, paired
1819 * with a prepare_task_switch call before the context switch.
1820 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1821 * and do any other architecture-specific cleanup actions.
1823 * Note that we may have delayed dropping an mm in context_switch(). If
1824 * so, we finish that here outside of the runqueue lock. (Doing it
1825 * with the lock held can cause deadlocks; see schedule() for
1828 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1829 __releases(rq
->lock
)
1831 struct mm_struct
*mm
= rq
->prev_mm
;
1837 * A task struct has one reference for the use as "current".
1838 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1839 * schedule one last time. The schedule call will never return, and
1840 * the scheduled task must drop that reference.
1841 * The test for TASK_DEAD must occur while the runqueue locks are
1842 * still held, otherwise prev could be scheduled on another cpu, die
1843 * there before we look at prev->state, and then the reference would
1845 * Manfred Spraul <manfred@colorfullife.com>
1847 prev_state
= prev
->state
;
1848 finish_arch_switch(prev
);
1849 finish_lock_switch(rq
, prev
);
1850 fire_sched_in_preempt_notifiers(current
);
1853 if (unlikely(prev_state
== TASK_DEAD
)) {
1855 * Remove function-return probe instances associated with this
1856 * task and put them back on the free list.
1858 kprobe_flush_task(prev
);
1859 put_task_struct(prev
);
1864 * schedule_tail - first thing a freshly forked thread must call.
1865 * @prev: the thread we just switched away from.
1867 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1868 __releases(rq
->lock
)
1870 struct rq
*rq
= this_rq();
1872 finish_task_switch(rq
, prev
);
1873 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1874 /* In this case, finish_task_switch does not reenable preemption */
1877 if (current
->set_child_tid
)
1878 put_user(current
->pid
, current
->set_child_tid
);
1882 * context_switch - switch to the new MM and the new
1883 * thread's register state.
1886 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1887 struct task_struct
*next
)
1889 struct mm_struct
*mm
, *oldmm
;
1891 prepare_task_switch(rq
, prev
, next
);
1893 oldmm
= prev
->active_mm
;
1895 * For paravirt, this is coupled with an exit in switch_to to
1896 * combine the page table reload and the switch backend into
1899 arch_enter_lazy_cpu_mode();
1901 if (unlikely(!mm
)) {
1902 next
->active_mm
= oldmm
;
1903 atomic_inc(&oldmm
->mm_count
);
1904 enter_lazy_tlb(oldmm
, next
);
1906 switch_mm(oldmm
, mm
, next
);
1908 if (unlikely(!prev
->mm
)) {
1909 prev
->active_mm
= NULL
;
1910 rq
->prev_mm
= oldmm
;
1913 * Since the runqueue lock will be released by the next
1914 * task (which is an invalid locking op but in the case
1915 * of the scheduler it's an obvious special-case), so we
1916 * do an early lockdep release here:
1918 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1919 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1922 /* Here we just switch the register state and the stack. */
1923 switch_to(prev
, next
, prev
);
1927 * this_rq must be evaluated again because prev may have moved
1928 * CPUs since it called schedule(), thus the 'rq' on its stack
1929 * frame will be invalid.
1931 finish_task_switch(this_rq(), prev
);
1935 * nr_running, nr_uninterruptible and nr_context_switches:
1937 * externally visible scheduler statistics: current number of runnable
1938 * threads, current number of uninterruptible-sleeping threads, total
1939 * number of context switches performed since bootup.
1941 unsigned long nr_running(void)
1943 unsigned long i
, sum
= 0;
1945 for_each_online_cpu(i
)
1946 sum
+= cpu_rq(i
)->nr_running
;
1951 unsigned long nr_uninterruptible(void)
1953 unsigned long i
, sum
= 0;
1955 for_each_possible_cpu(i
)
1956 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1959 * Since we read the counters lockless, it might be slightly
1960 * inaccurate. Do not allow it to go below zero though:
1962 if (unlikely((long)sum
< 0))
1968 unsigned long long nr_context_switches(void)
1971 unsigned long long sum
= 0;
1973 for_each_possible_cpu(i
)
1974 sum
+= cpu_rq(i
)->nr_switches
;
1979 unsigned long nr_iowait(void)
1981 unsigned long i
, sum
= 0;
1983 for_each_possible_cpu(i
)
1984 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1989 unsigned long nr_active(void)
1991 unsigned long i
, running
= 0, uninterruptible
= 0;
1993 for_each_online_cpu(i
) {
1994 running
+= cpu_rq(i
)->nr_running
;
1995 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1998 if (unlikely((long)uninterruptible
< 0))
1999 uninterruptible
= 0;
2001 return running
+ uninterruptible
;
2005 * Update rq->cpu_load[] statistics. This function is usually called every
2006 * scheduler tick (TICK_NSEC).
2008 static void update_cpu_load(struct rq
*this_rq
)
2010 unsigned long this_load
= this_rq
->load
.weight
;
2013 this_rq
->nr_load_updates
++;
2015 /* Update our load: */
2016 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2017 unsigned long old_load
, new_load
;
2019 /* scale is effectively 1 << i now, and >> i divides by scale */
2021 old_load
= this_rq
->cpu_load
[i
];
2022 new_load
= this_load
;
2024 * Round up the averaging division if load is increasing. This
2025 * prevents us from getting stuck on 9 if the load is 10, for
2028 if (new_load
> old_load
)
2029 new_load
+= scale
-1;
2030 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2037 * double_rq_lock - safely lock two runqueues
2039 * Note this does not disable interrupts like task_rq_lock,
2040 * you need to do so manually before calling.
2042 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2043 __acquires(rq1
->lock
)
2044 __acquires(rq2
->lock
)
2046 BUG_ON(!irqs_disabled());
2048 spin_lock(&rq1
->lock
);
2049 __acquire(rq2
->lock
); /* Fake it out ;) */
2052 spin_lock(&rq1
->lock
);
2053 spin_lock(&rq2
->lock
);
2055 spin_lock(&rq2
->lock
);
2056 spin_lock(&rq1
->lock
);
2059 update_rq_clock(rq1
);
2060 update_rq_clock(rq2
);
2064 * double_rq_unlock - safely unlock two runqueues
2066 * Note this does not restore interrupts like task_rq_unlock,
2067 * you need to do so manually after calling.
2069 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2070 __releases(rq1
->lock
)
2071 __releases(rq2
->lock
)
2073 spin_unlock(&rq1
->lock
);
2075 spin_unlock(&rq2
->lock
);
2077 __release(rq2
->lock
);
2081 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2083 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2084 __releases(this_rq
->lock
)
2085 __acquires(busiest
->lock
)
2086 __acquires(this_rq
->lock
)
2088 if (unlikely(!irqs_disabled())) {
2089 /* printk() doesn't work good under rq->lock */
2090 spin_unlock(&this_rq
->lock
);
2093 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2094 if (busiest
< this_rq
) {
2095 spin_unlock(&this_rq
->lock
);
2096 spin_lock(&busiest
->lock
);
2097 spin_lock(&this_rq
->lock
);
2099 spin_lock(&busiest
->lock
);
2104 * If dest_cpu is allowed for this process, migrate the task to it.
2105 * This is accomplished by forcing the cpu_allowed mask to only
2106 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2107 * the cpu_allowed mask is restored.
2109 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2111 struct migration_req req
;
2112 unsigned long flags
;
2115 rq
= task_rq_lock(p
, &flags
);
2116 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2117 || unlikely(cpu_is_offline(dest_cpu
)))
2120 /* force the process onto the specified CPU */
2121 if (migrate_task(p
, dest_cpu
, &req
)) {
2122 /* Need to wait for migration thread (might exit: take ref). */
2123 struct task_struct
*mt
= rq
->migration_thread
;
2125 get_task_struct(mt
);
2126 task_rq_unlock(rq
, &flags
);
2127 wake_up_process(mt
);
2128 put_task_struct(mt
);
2129 wait_for_completion(&req
.done
);
2134 task_rq_unlock(rq
, &flags
);
2138 * sched_exec - execve() is a valuable balancing opportunity, because at
2139 * this point the task has the smallest effective memory and cache footprint.
2141 void sched_exec(void)
2143 int new_cpu
, this_cpu
= get_cpu();
2144 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2146 if (new_cpu
!= this_cpu
)
2147 sched_migrate_task(current
, new_cpu
);
2151 * pull_task - move a task from a remote runqueue to the local runqueue.
2152 * Both runqueues must be locked.
2154 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2155 struct rq
*this_rq
, int this_cpu
)
2157 deactivate_task(src_rq
, p
, 0);
2158 set_task_cpu(p
, this_cpu
);
2159 activate_task(this_rq
, p
, 0);
2161 * Note that idle threads have a prio of MAX_PRIO, for this test
2162 * to be always true for them.
2164 check_preempt_curr(this_rq
, p
);
2168 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2171 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2172 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2176 * We do not migrate tasks that are:
2177 * 1) running (obviously), or
2178 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2179 * 3) are cache-hot on their current CPU.
2181 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2182 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2187 if (task_running(rq
, p
)) {
2188 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2193 * Aggressive migration if:
2194 * 1) task is cache cold, or
2195 * 2) too many balance attempts have failed.
2198 if (!task_hot(p
, rq
->clock
, sd
) ||
2199 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2200 #ifdef CONFIG_SCHEDSTATS
2201 if (task_hot(p
, rq
->clock
, sd
)) {
2202 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2203 schedstat_inc(p
, se
.nr_forced_migrations
);
2209 if (task_hot(p
, rq
->clock
, sd
)) {
2210 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2216 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2217 unsigned long max_nr_move
, unsigned long max_load_move
,
2218 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2219 int *all_pinned
, unsigned long *load_moved
,
2220 int *this_best_prio
, struct rq_iterator
*iterator
)
2222 int pulled
= 0, pinned
= 0, skip_for_load
;
2223 struct task_struct
*p
;
2224 long rem_load_move
= max_load_move
;
2226 if (max_nr_move
== 0 || max_load_move
== 0)
2232 * Start the load-balancing iterator:
2234 p
= iterator
->start(iterator
->arg
);
2239 * To help distribute high priority tasks accross CPUs we don't
2240 * skip a task if it will be the highest priority task (i.e. smallest
2241 * prio value) on its new queue regardless of its load weight
2243 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2244 SCHED_LOAD_SCALE_FUZZ
;
2245 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2246 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2247 p
= iterator
->next(iterator
->arg
);
2251 pull_task(busiest
, p
, this_rq
, this_cpu
);
2253 rem_load_move
-= p
->se
.load
.weight
;
2256 * We only want to steal up to the prescribed number of tasks
2257 * and the prescribed amount of weighted load.
2259 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2260 if (p
->prio
< *this_best_prio
)
2261 *this_best_prio
= p
->prio
;
2262 p
= iterator
->next(iterator
->arg
);
2267 * Right now, this is the only place pull_task() is called,
2268 * so we can safely collect pull_task() stats here rather than
2269 * inside pull_task().
2271 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2274 *all_pinned
= pinned
;
2275 *load_moved
= max_load_move
- rem_load_move
;
2280 * move_tasks tries to move up to max_load_move weighted load from busiest to
2281 * this_rq, as part of a balancing operation within domain "sd".
2282 * Returns 1 if successful and 0 otherwise.
2284 * Called with both runqueues locked.
2286 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2287 unsigned long max_load_move
,
2288 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2291 const struct sched_class
*class = sched_class_highest
;
2292 unsigned long total_load_moved
= 0;
2293 int this_best_prio
= this_rq
->curr
->prio
;
2297 class->load_balance(this_rq
, this_cpu
, busiest
,
2298 ULONG_MAX
, max_load_move
- total_load_moved
,
2299 sd
, idle
, all_pinned
, &this_best_prio
);
2300 class = class->next
;
2301 } while (class && max_load_move
> total_load_moved
);
2303 return total_load_moved
> 0;
2307 * move_one_task tries to move exactly one task from busiest to this_rq, as
2308 * part of active balancing operations within "domain".
2309 * Returns 1 if successful and 0 otherwise.
2311 * Called with both runqueues locked.
2313 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2314 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2316 const struct sched_class
*class;
2317 int this_best_prio
= MAX_PRIO
;
2319 for (class = sched_class_highest
; class; class = class->next
)
2320 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2321 1, ULONG_MAX
, sd
, idle
, NULL
,
2329 * find_busiest_group finds and returns the busiest CPU group within the
2330 * domain. It calculates and returns the amount of weighted load which
2331 * should be moved to restore balance via the imbalance parameter.
2333 static struct sched_group
*
2334 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2335 unsigned long *imbalance
, enum cpu_idle_type idle
,
2336 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2338 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2339 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2340 unsigned long max_pull
;
2341 unsigned long busiest_load_per_task
, busiest_nr_running
;
2342 unsigned long this_load_per_task
, this_nr_running
;
2343 int load_idx
, group_imb
= 0;
2344 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2345 int power_savings_balance
= 1;
2346 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2347 unsigned long min_nr_running
= ULONG_MAX
;
2348 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2351 max_load
= this_load
= total_load
= total_pwr
= 0;
2352 busiest_load_per_task
= busiest_nr_running
= 0;
2353 this_load_per_task
= this_nr_running
= 0;
2354 if (idle
== CPU_NOT_IDLE
)
2355 load_idx
= sd
->busy_idx
;
2356 else if (idle
== CPU_NEWLY_IDLE
)
2357 load_idx
= sd
->newidle_idx
;
2359 load_idx
= sd
->idle_idx
;
2362 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2365 int __group_imb
= 0;
2366 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2367 unsigned long sum_nr_running
, sum_weighted_load
;
2369 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2372 balance_cpu
= first_cpu(group
->cpumask
);
2374 /* Tally up the load of all CPUs in the group */
2375 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2377 min_cpu_load
= ~0UL;
2379 for_each_cpu_mask(i
, group
->cpumask
) {
2382 if (!cpu_isset(i
, *cpus
))
2387 if (*sd_idle
&& rq
->nr_running
)
2390 /* Bias balancing toward cpus of our domain */
2392 if (idle_cpu(i
) && !first_idle_cpu
) {
2397 load
= target_load(i
, load_idx
);
2399 load
= source_load(i
, load_idx
);
2400 if (load
> max_cpu_load
)
2401 max_cpu_load
= load
;
2402 if (min_cpu_load
> load
)
2403 min_cpu_load
= load
;
2407 sum_nr_running
+= rq
->nr_running
;
2408 sum_weighted_load
+= weighted_cpuload(i
);
2412 * First idle cpu or the first cpu(busiest) in this sched group
2413 * is eligible for doing load balancing at this and above
2414 * domains. In the newly idle case, we will allow all the cpu's
2415 * to do the newly idle load balance.
2417 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2418 balance_cpu
!= this_cpu
&& balance
) {
2423 total_load
+= avg_load
;
2424 total_pwr
+= group
->__cpu_power
;
2426 /* Adjust by relative CPU power of the group */
2427 avg_load
= sg_div_cpu_power(group
,
2428 avg_load
* SCHED_LOAD_SCALE
);
2430 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2433 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2436 this_load
= avg_load
;
2438 this_nr_running
= sum_nr_running
;
2439 this_load_per_task
= sum_weighted_load
;
2440 } else if (avg_load
> max_load
&&
2441 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2442 max_load
= avg_load
;
2444 busiest_nr_running
= sum_nr_running
;
2445 busiest_load_per_task
= sum_weighted_load
;
2446 group_imb
= __group_imb
;
2449 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2451 * Busy processors will not participate in power savings
2454 if (idle
== CPU_NOT_IDLE
||
2455 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2459 * If the local group is idle or completely loaded
2460 * no need to do power savings balance at this domain
2462 if (local_group
&& (this_nr_running
>= group_capacity
||
2464 power_savings_balance
= 0;
2467 * If a group is already running at full capacity or idle,
2468 * don't include that group in power savings calculations
2470 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2475 * Calculate the group which has the least non-idle load.
2476 * This is the group from where we need to pick up the load
2479 if ((sum_nr_running
< min_nr_running
) ||
2480 (sum_nr_running
== min_nr_running
&&
2481 first_cpu(group
->cpumask
) <
2482 first_cpu(group_min
->cpumask
))) {
2484 min_nr_running
= sum_nr_running
;
2485 min_load_per_task
= sum_weighted_load
/
2490 * Calculate the group which is almost near its
2491 * capacity but still has some space to pick up some load
2492 * from other group and save more power
2494 if (sum_nr_running
<= group_capacity
- 1) {
2495 if (sum_nr_running
> leader_nr_running
||
2496 (sum_nr_running
== leader_nr_running
&&
2497 first_cpu(group
->cpumask
) >
2498 first_cpu(group_leader
->cpumask
))) {
2499 group_leader
= group
;
2500 leader_nr_running
= sum_nr_running
;
2505 group
= group
->next
;
2506 } while (group
!= sd
->groups
);
2508 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2511 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2513 if (this_load
>= avg_load
||
2514 100*max_load
<= sd
->imbalance_pct
*this_load
)
2517 busiest_load_per_task
/= busiest_nr_running
;
2519 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2522 * We're trying to get all the cpus to the average_load, so we don't
2523 * want to push ourselves above the average load, nor do we wish to
2524 * reduce the max loaded cpu below the average load, as either of these
2525 * actions would just result in more rebalancing later, and ping-pong
2526 * tasks around. Thus we look for the minimum possible imbalance.
2527 * Negative imbalances (*we* are more loaded than anyone else) will
2528 * be counted as no imbalance for these purposes -- we can't fix that
2529 * by pulling tasks to us. Be careful of negative numbers as they'll
2530 * appear as very large values with unsigned longs.
2532 if (max_load
<= busiest_load_per_task
)
2536 * In the presence of smp nice balancing, certain scenarios can have
2537 * max load less than avg load(as we skip the groups at or below
2538 * its cpu_power, while calculating max_load..)
2540 if (max_load
< avg_load
) {
2542 goto small_imbalance
;
2545 /* Don't want to pull so many tasks that a group would go idle */
2546 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2548 /* How much load to actually move to equalise the imbalance */
2549 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2550 (avg_load
- this_load
) * this->__cpu_power
)
2554 * if *imbalance is less than the average load per runnable task
2555 * there is no gaurantee that any tasks will be moved so we'll have
2556 * a think about bumping its value to force at least one task to be
2559 if (*imbalance
< busiest_load_per_task
) {
2560 unsigned long tmp
, pwr_now
, pwr_move
;
2564 pwr_move
= pwr_now
= 0;
2566 if (this_nr_running
) {
2567 this_load_per_task
/= this_nr_running
;
2568 if (busiest_load_per_task
> this_load_per_task
)
2571 this_load_per_task
= SCHED_LOAD_SCALE
;
2573 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2574 busiest_load_per_task
* imbn
) {
2575 *imbalance
= busiest_load_per_task
;
2580 * OK, we don't have enough imbalance to justify moving tasks,
2581 * however we may be able to increase total CPU power used by
2585 pwr_now
+= busiest
->__cpu_power
*
2586 min(busiest_load_per_task
, max_load
);
2587 pwr_now
+= this->__cpu_power
*
2588 min(this_load_per_task
, this_load
);
2589 pwr_now
/= SCHED_LOAD_SCALE
;
2591 /* Amount of load we'd subtract */
2592 tmp
= sg_div_cpu_power(busiest
,
2593 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2595 pwr_move
+= busiest
->__cpu_power
*
2596 min(busiest_load_per_task
, max_load
- tmp
);
2598 /* Amount of load we'd add */
2599 if (max_load
* busiest
->__cpu_power
<
2600 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2601 tmp
= sg_div_cpu_power(this,
2602 max_load
* busiest
->__cpu_power
);
2604 tmp
= sg_div_cpu_power(this,
2605 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2606 pwr_move
+= this->__cpu_power
*
2607 min(this_load_per_task
, this_load
+ tmp
);
2608 pwr_move
/= SCHED_LOAD_SCALE
;
2610 /* Move if we gain throughput */
2611 if (pwr_move
> pwr_now
)
2612 *imbalance
= busiest_load_per_task
;
2618 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2619 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2622 if (this == group_leader
&& group_leader
!= group_min
) {
2623 *imbalance
= min_load_per_task
;
2633 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2636 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2637 unsigned long imbalance
, cpumask_t
*cpus
)
2639 struct rq
*busiest
= NULL
, *rq
;
2640 unsigned long max_load
= 0;
2643 for_each_cpu_mask(i
, group
->cpumask
) {
2646 if (!cpu_isset(i
, *cpus
))
2650 wl
= weighted_cpuload(i
);
2652 if (rq
->nr_running
== 1 && wl
> imbalance
)
2655 if (wl
> max_load
) {
2665 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2666 * so long as it is large enough.
2668 #define MAX_PINNED_INTERVAL 512
2671 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2672 * tasks if there is an imbalance.
2674 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2675 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2678 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2679 struct sched_group
*group
;
2680 unsigned long imbalance
;
2682 cpumask_t cpus
= CPU_MASK_ALL
;
2683 unsigned long flags
;
2686 * When power savings policy is enabled for the parent domain, idle
2687 * sibling can pick up load irrespective of busy siblings. In this case,
2688 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2689 * portraying it as CPU_NOT_IDLE.
2691 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2692 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2695 schedstat_inc(sd
, lb_count
[idle
]);
2698 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2705 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2709 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2711 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2715 BUG_ON(busiest
== this_rq
);
2717 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2720 if (busiest
->nr_running
> 1) {
2722 * Attempt to move tasks. If find_busiest_group has found
2723 * an imbalance but busiest->nr_running <= 1, the group is
2724 * still unbalanced. ld_moved simply stays zero, so it is
2725 * correctly treated as an imbalance.
2727 local_irq_save(flags
);
2728 double_rq_lock(this_rq
, busiest
);
2729 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2730 imbalance
, sd
, idle
, &all_pinned
);
2731 double_rq_unlock(this_rq
, busiest
);
2732 local_irq_restore(flags
);
2735 * some other cpu did the load balance for us.
2737 if (ld_moved
&& this_cpu
!= smp_processor_id())
2738 resched_cpu(this_cpu
);
2740 /* All tasks on this runqueue were pinned by CPU affinity */
2741 if (unlikely(all_pinned
)) {
2742 cpu_clear(cpu_of(busiest
), cpus
);
2743 if (!cpus_empty(cpus
))
2750 schedstat_inc(sd
, lb_failed
[idle
]);
2751 sd
->nr_balance_failed
++;
2753 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2755 spin_lock_irqsave(&busiest
->lock
, flags
);
2757 /* don't kick the migration_thread, if the curr
2758 * task on busiest cpu can't be moved to this_cpu
2760 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2761 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2763 goto out_one_pinned
;
2766 if (!busiest
->active_balance
) {
2767 busiest
->active_balance
= 1;
2768 busiest
->push_cpu
= this_cpu
;
2771 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2773 wake_up_process(busiest
->migration_thread
);
2776 * We've kicked active balancing, reset the failure
2779 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2782 sd
->nr_balance_failed
= 0;
2784 if (likely(!active_balance
)) {
2785 /* We were unbalanced, so reset the balancing interval */
2786 sd
->balance_interval
= sd
->min_interval
;
2789 * If we've begun active balancing, start to back off. This
2790 * case may not be covered by the all_pinned logic if there
2791 * is only 1 task on the busy runqueue (because we don't call
2794 if (sd
->balance_interval
< sd
->max_interval
)
2795 sd
->balance_interval
*= 2;
2798 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2799 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2804 schedstat_inc(sd
, lb_balanced
[idle
]);
2806 sd
->nr_balance_failed
= 0;
2809 /* tune up the balancing interval */
2810 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2811 (sd
->balance_interval
< sd
->max_interval
))
2812 sd
->balance_interval
*= 2;
2814 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2815 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2821 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2822 * tasks if there is an imbalance.
2824 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2825 * this_rq is locked.
2828 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2830 struct sched_group
*group
;
2831 struct rq
*busiest
= NULL
;
2832 unsigned long imbalance
;
2836 cpumask_t cpus
= CPU_MASK_ALL
;
2839 * When power savings policy is enabled for the parent domain, idle
2840 * sibling can pick up load irrespective of busy siblings. In this case,
2841 * let the state of idle sibling percolate up as IDLE, instead of
2842 * portraying it as CPU_NOT_IDLE.
2844 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2845 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2848 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2850 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2851 &sd_idle
, &cpus
, NULL
);
2853 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2857 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2860 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2864 BUG_ON(busiest
== this_rq
);
2866 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2869 if (busiest
->nr_running
> 1) {
2870 /* Attempt to move tasks */
2871 double_lock_balance(this_rq
, busiest
);
2872 /* this_rq->clock is already updated */
2873 update_rq_clock(busiest
);
2874 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2875 imbalance
, sd
, CPU_NEWLY_IDLE
,
2877 spin_unlock(&busiest
->lock
);
2879 if (unlikely(all_pinned
)) {
2880 cpu_clear(cpu_of(busiest
), cpus
);
2881 if (!cpus_empty(cpus
))
2887 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2888 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2889 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2892 sd
->nr_balance_failed
= 0;
2897 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2898 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2899 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2901 sd
->nr_balance_failed
= 0;
2907 * idle_balance is called by schedule() if this_cpu is about to become
2908 * idle. Attempts to pull tasks from other CPUs.
2910 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2912 struct sched_domain
*sd
;
2913 int pulled_task
= -1;
2914 unsigned long next_balance
= jiffies
+ HZ
;
2916 for_each_domain(this_cpu
, sd
) {
2917 unsigned long interval
;
2919 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2922 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2923 /* If we've pulled tasks over stop searching: */
2924 pulled_task
= load_balance_newidle(this_cpu
,
2927 interval
= msecs_to_jiffies(sd
->balance_interval
);
2928 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2929 next_balance
= sd
->last_balance
+ interval
;
2933 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2935 * We are going idle. next_balance may be set based on
2936 * a busy processor. So reset next_balance.
2938 this_rq
->next_balance
= next_balance
;
2943 * active_load_balance is run by migration threads. It pushes running tasks
2944 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2945 * running on each physical CPU where possible, and avoids physical /
2946 * logical imbalances.
2948 * Called with busiest_rq locked.
2950 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2952 int target_cpu
= busiest_rq
->push_cpu
;
2953 struct sched_domain
*sd
;
2954 struct rq
*target_rq
;
2956 /* Is there any task to move? */
2957 if (busiest_rq
->nr_running
<= 1)
2960 target_rq
= cpu_rq(target_cpu
);
2963 * This condition is "impossible", if it occurs
2964 * we need to fix it. Originally reported by
2965 * Bjorn Helgaas on a 128-cpu setup.
2967 BUG_ON(busiest_rq
== target_rq
);
2969 /* move a task from busiest_rq to target_rq */
2970 double_lock_balance(busiest_rq
, target_rq
);
2971 update_rq_clock(busiest_rq
);
2972 update_rq_clock(target_rq
);
2974 /* Search for an sd spanning us and the target CPU. */
2975 for_each_domain(target_cpu
, sd
) {
2976 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2977 cpu_isset(busiest_cpu
, sd
->span
))
2982 schedstat_inc(sd
, alb_count
);
2984 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2986 schedstat_inc(sd
, alb_pushed
);
2988 schedstat_inc(sd
, alb_failed
);
2990 spin_unlock(&target_rq
->lock
);
2995 atomic_t load_balancer
;
2997 } nohz ____cacheline_aligned
= {
2998 .load_balancer
= ATOMIC_INIT(-1),
2999 .cpu_mask
= CPU_MASK_NONE
,
3003 * This routine will try to nominate the ilb (idle load balancing)
3004 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3005 * load balancing on behalf of all those cpus. If all the cpus in the system
3006 * go into this tickless mode, then there will be no ilb owner (as there is
3007 * no need for one) and all the cpus will sleep till the next wakeup event
3010 * For the ilb owner, tick is not stopped. And this tick will be used
3011 * for idle load balancing. ilb owner will still be part of
3014 * While stopping the tick, this cpu will become the ilb owner if there
3015 * is no other owner. And will be the owner till that cpu becomes busy
3016 * or if all cpus in the system stop their ticks at which point
3017 * there is no need for ilb owner.
3019 * When the ilb owner becomes busy, it nominates another owner, during the
3020 * next busy scheduler_tick()
3022 int select_nohz_load_balancer(int stop_tick
)
3024 int cpu
= smp_processor_id();
3027 cpu_set(cpu
, nohz
.cpu_mask
);
3028 cpu_rq(cpu
)->in_nohz_recently
= 1;
3031 * If we are going offline and still the leader, give up!
3033 if (cpu_is_offline(cpu
) &&
3034 atomic_read(&nohz
.load_balancer
) == cpu
) {
3035 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3040 /* time for ilb owner also to sleep */
3041 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3042 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3043 atomic_set(&nohz
.load_balancer
, -1);
3047 if (atomic_read(&nohz
.load_balancer
) == -1) {
3048 /* make me the ilb owner */
3049 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3051 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3054 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3057 cpu_clear(cpu
, nohz
.cpu_mask
);
3059 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3060 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3067 static DEFINE_SPINLOCK(balancing
);
3070 * It checks each scheduling domain to see if it is due to be balanced,
3071 * and initiates a balancing operation if so.
3073 * Balancing parameters are set up in arch_init_sched_domains.
3075 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3078 struct rq
*rq
= cpu_rq(cpu
);
3079 unsigned long interval
;
3080 struct sched_domain
*sd
;
3081 /* Earliest time when we have to do rebalance again */
3082 unsigned long next_balance
= jiffies
+ 60*HZ
;
3083 int update_next_balance
= 0;
3085 for_each_domain(cpu
, sd
) {
3086 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3089 interval
= sd
->balance_interval
;
3090 if (idle
!= CPU_IDLE
)
3091 interval
*= sd
->busy_factor
;
3093 /* scale ms to jiffies */
3094 interval
= msecs_to_jiffies(interval
);
3095 if (unlikely(!interval
))
3097 if (interval
> HZ
*NR_CPUS
/10)
3098 interval
= HZ
*NR_CPUS
/10;
3101 if (sd
->flags
& SD_SERIALIZE
) {
3102 if (!spin_trylock(&balancing
))
3106 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3107 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3109 * We've pulled tasks over so either we're no
3110 * longer idle, or one of our SMT siblings is
3113 idle
= CPU_NOT_IDLE
;
3115 sd
->last_balance
= jiffies
;
3117 if (sd
->flags
& SD_SERIALIZE
)
3118 spin_unlock(&balancing
);
3120 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3121 next_balance
= sd
->last_balance
+ interval
;
3122 update_next_balance
= 1;
3126 * Stop the load balance at this level. There is another
3127 * CPU in our sched group which is doing load balancing more
3135 * next_balance will be updated only when there is a need.
3136 * When the cpu is attached to null domain for ex, it will not be
3139 if (likely(update_next_balance
))
3140 rq
->next_balance
= next_balance
;
3144 * run_rebalance_domains is triggered when needed from the scheduler tick.
3145 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3146 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3148 static void run_rebalance_domains(struct softirq_action
*h
)
3150 int this_cpu
= smp_processor_id();
3151 struct rq
*this_rq
= cpu_rq(this_cpu
);
3152 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3153 CPU_IDLE
: CPU_NOT_IDLE
;
3155 rebalance_domains(this_cpu
, idle
);
3159 * If this cpu is the owner for idle load balancing, then do the
3160 * balancing on behalf of the other idle cpus whose ticks are
3163 if (this_rq
->idle_at_tick
&&
3164 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3165 cpumask_t cpus
= nohz
.cpu_mask
;
3169 cpu_clear(this_cpu
, cpus
);
3170 for_each_cpu_mask(balance_cpu
, cpus
) {
3172 * If this cpu gets work to do, stop the load balancing
3173 * work being done for other cpus. Next load
3174 * balancing owner will pick it up.
3179 rebalance_domains(balance_cpu
, CPU_IDLE
);
3181 rq
= cpu_rq(balance_cpu
);
3182 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3183 this_rq
->next_balance
= rq
->next_balance
;
3190 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3192 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3193 * idle load balancing owner or decide to stop the periodic load balancing,
3194 * if the whole system is idle.
3196 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3200 * If we were in the nohz mode recently and busy at the current
3201 * scheduler tick, then check if we need to nominate new idle
3204 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3205 rq
->in_nohz_recently
= 0;
3207 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3208 cpu_clear(cpu
, nohz
.cpu_mask
);
3209 atomic_set(&nohz
.load_balancer
, -1);
3212 if (atomic_read(&nohz
.load_balancer
) == -1) {
3214 * simple selection for now: Nominate the
3215 * first cpu in the nohz list to be the next
3218 * TBD: Traverse the sched domains and nominate
3219 * the nearest cpu in the nohz.cpu_mask.
3221 int ilb
= first_cpu(nohz
.cpu_mask
);
3229 * If this cpu is idle and doing idle load balancing for all the
3230 * cpus with ticks stopped, is it time for that to stop?
3232 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3233 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3239 * If this cpu is idle and the idle load balancing is done by
3240 * someone else, then no need raise the SCHED_SOFTIRQ
3242 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3243 cpu_isset(cpu
, nohz
.cpu_mask
))
3246 if (time_after_eq(jiffies
, rq
->next_balance
))
3247 raise_softirq(SCHED_SOFTIRQ
);
3250 #else /* CONFIG_SMP */
3253 * on UP we do not need to balance between CPUs:
3255 static inline void idle_balance(int cpu
, struct rq
*rq
)
3259 /* Avoid "used but not defined" warning on UP */
3260 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3261 unsigned long max_nr_move
, unsigned long max_load_move
,
3262 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3263 int *all_pinned
, unsigned long *load_moved
,
3264 int *this_best_prio
, struct rq_iterator
*iterator
)
3273 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3275 EXPORT_PER_CPU_SYMBOL(kstat
);
3278 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3279 * that have not yet been banked in case the task is currently running.
3281 unsigned long long task_sched_runtime(struct task_struct
*p
)
3283 unsigned long flags
;
3287 rq
= task_rq_lock(p
, &flags
);
3288 ns
= p
->se
.sum_exec_runtime
;
3289 if (rq
->curr
== p
) {
3290 update_rq_clock(rq
);
3291 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3292 if ((s64
)delta_exec
> 0)
3295 task_rq_unlock(rq
, &flags
);
3301 * Account user cpu time to a process.
3302 * @p: the process that the cpu time gets accounted to
3303 * @hardirq_offset: the offset to subtract from hardirq_count()
3304 * @cputime: the cpu time spent in user space since the last update
3306 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3308 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3311 p
->utime
= cputime_add(p
->utime
, cputime
);
3313 /* Add user time to cpustat. */
3314 tmp
= cputime_to_cputime64(cputime
);
3315 if (TASK_NICE(p
) > 0)
3316 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3318 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3322 * Account guest cpu time to a process.
3323 * @p: the process that the cpu time gets accounted to
3324 * @cputime: the cpu time spent in virtual machine since the last update
3326 void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3329 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3331 tmp
= cputime_to_cputime64(cputime
);
3333 p
->utime
= cputime_add(p
->utime
, cputime
);
3334 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3336 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3337 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3341 * Account system cpu time to a process.
3342 * @p: the process that the cpu time gets accounted to
3343 * @hardirq_offset: the offset to subtract from hardirq_count()
3344 * @cputime: the cpu time spent in kernel space since the last update
3346 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3349 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3350 struct rq
*rq
= this_rq();
3353 if (p
->flags
& PF_VCPU
) {
3354 account_guest_time(p
, cputime
);
3355 p
->flags
&= ~PF_VCPU
;
3359 p
->stime
= cputime_add(p
->stime
, cputime
);
3361 /* Add system time to cpustat. */
3362 tmp
= cputime_to_cputime64(cputime
);
3363 if (hardirq_count() - hardirq_offset
)
3364 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3365 else if (softirq_count())
3366 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3367 else if (p
!= rq
->idle
)
3368 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3369 else if (atomic_read(&rq
->nr_iowait
) > 0)
3370 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3372 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3373 /* Account for system time used */
3374 acct_update_integrals(p
);
3378 * Account for involuntary wait time.
3379 * @p: the process from which the cpu time has been stolen
3380 * @steal: the cpu time spent in involuntary wait
3382 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3384 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3385 cputime64_t tmp
= cputime_to_cputime64(steal
);
3386 struct rq
*rq
= this_rq();
3388 if (p
== rq
->idle
) {
3389 p
->stime
= cputime_add(p
->stime
, steal
);
3390 if (atomic_read(&rq
->nr_iowait
) > 0)
3391 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3393 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3395 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3399 * This function gets called by the timer code, with HZ frequency.
3400 * We call it with interrupts disabled.
3402 * It also gets called by the fork code, when changing the parent's
3405 void scheduler_tick(void)
3407 int cpu
= smp_processor_id();
3408 struct rq
*rq
= cpu_rq(cpu
);
3409 struct task_struct
*curr
= rq
->curr
;
3410 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3412 spin_lock(&rq
->lock
);
3413 __update_rq_clock(rq
);
3415 * Let rq->clock advance by at least TICK_NSEC:
3417 if (unlikely(rq
->clock
< next_tick
))
3418 rq
->clock
= next_tick
;
3419 rq
->tick_timestamp
= rq
->clock
;
3420 update_cpu_load(rq
);
3421 if (curr
!= rq
->idle
) /* FIXME: needed? */
3422 curr
->sched_class
->task_tick(rq
, curr
);
3423 spin_unlock(&rq
->lock
);
3426 rq
->idle_at_tick
= idle_cpu(cpu
);
3427 trigger_load_balance(rq
, cpu
);
3431 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3433 void fastcall
add_preempt_count(int val
)
3438 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3440 preempt_count() += val
;
3442 * Spinlock count overflowing soon?
3444 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3447 EXPORT_SYMBOL(add_preempt_count
);
3449 void fastcall
sub_preempt_count(int val
)
3454 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3457 * Is the spinlock portion underflowing?
3459 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3460 !(preempt_count() & PREEMPT_MASK
)))
3463 preempt_count() -= val
;
3465 EXPORT_SYMBOL(sub_preempt_count
);
3470 * Print scheduling while atomic bug:
3472 static noinline
void __schedule_bug(struct task_struct
*prev
)
3474 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3475 prev
->comm
, preempt_count(), prev
->pid
);
3476 debug_show_held_locks(prev
);
3477 if (irqs_disabled())
3478 print_irqtrace_events(prev
);
3483 * Various schedule()-time debugging checks and statistics:
3485 static inline void schedule_debug(struct task_struct
*prev
)
3488 * Test if we are atomic. Since do_exit() needs to call into
3489 * schedule() atomically, we ignore that path for now.
3490 * Otherwise, whine if we are scheduling when we should not be.
3492 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3493 __schedule_bug(prev
);
3495 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3497 schedstat_inc(this_rq(), sched_count
);
3498 #ifdef CONFIG_SCHEDSTATS
3499 if (unlikely(prev
->lock_depth
>= 0)) {
3500 schedstat_inc(this_rq(), bkl_count
);
3501 schedstat_inc(prev
, sched_info
.bkl_count
);
3507 * Pick up the highest-prio task:
3509 static inline struct task_struct
*
3510 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3512 const struct sched_class
*class;
3513 struct task_struct
*p
;
3516 * Optimization: we know that if all tasks are in
3517 * the fair class we can call that function directly:
3519 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3520 p
= fair_sched_class
.pick_next_task(rq
);
3525 class = sched_class_highest
;
3527 p
= class->pick_next_task(rq
);
3531 * Will never be NULL as the idle class always
3532 * returns a non-NULL p:
3534 class = class->next
;
3539 * schedule() is the main scheduler function.
3541 asmlinkage
void __sched
schedule(void)
3543 struct task_struct
*prev
, *next
;
3550 cpu
= smp_processor_id();
3554 switch_count
= &prev
->nivcsw
;
3556 release_kernel_lock(prev
);
3557 need_resched_nonpreemptible
:
3559 schedule_debug(prev
);
3562 * Do the rq-clock update outside the rq lock:
3564 local_irq_disable();
3565 __update_rq_clock(rq
);
3566 spin_lock(&rq
->lock
);
3567 clear_tsk_need_resched(prev
);
3569 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3570 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3571 unlikely(signal_pending(prev
)))) {
3572 prev
->state
= TASK_RUNNING
;
3574 deactivate_task(rq
, prev
, 1);
3576 switch_count
= &prev
->nvcsw
;
3579 if (unlikely(!rq
->nr_running
))
3580 idle_balance(cpu
, rq
);
3582 prev
->sched_class
->put_prev_task(rq
, prev
);
3583 next
= pick_next_task(rq
, prev
);
3585 sched_info_switch(prev
, next
);
3587 if (likely(prev
!= next
)) {
3592 context_switch(rq
, prev
, next
); /* unlocks the rq */
3594 spin_unlock_irq(&rq
->lock
);
3596 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3597 cpu
= smp_processor_id();
3599 goto need_resched_nonpreemptible
;
3601 preempt_enable_no_resched();
3602 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3605 EXPORT_SYMBOL(schedule
);
3607 #ifdef CONFIG_PREEMPT
3609 * this is the entry point to schedule() from in-kernel preemption
3610 * off of preempt_enable. Kernel preemptions off return from interrupt
3611 * occur there and call schedule directly.
3613 asmlinkage
void __sched
preempt_schedule(void)
3615 struct thread_info
*ti
= current_thread_info();
3616 #ifdef CONFIG_PREEMPT_BKL
3617 struct task_struct
*task
= current
;
3618 int saved_lock_depth
;
3621 * If there is a non-zero preempt_count or interrupts are disabled,
3622 * we do not want to preempt the current task. Just return..
3624 if (likely(ti
->preempt_count
|| irqs_disabled()))
3628 add_preempt_count(PREEMPT_ACTIVE
);
3631 * We keep the big kernel semaphore locked, but we
3632 * clear ->lock_depth so that schedule() doesnt
3633 * auto-release the semaphore:
3635 #ifdef CONFIG_PREEMPT_BKL
3636 saved_lock_depth
= task
->lock_depth
;
3637 task
->lock_depth
= -1;
3640 #ifdef CONFIG_PREEMPT_BKL
3641 task
->lock_depth
= saved_lock_depth
;
3643 sub_preempt_count(PREEMPT_ACTIVE
);
3646 * Check again in case we missed a preemption opportunity
3647 * between schedule and now.
3650 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3652 EXPORT_SYMBOL(preempt_schedule
);
3655 * this is the entry point to schedule() from kernel preemption
3656 * off of irq context.
3657 * Note, that this is called and return with irqs disabled. This will
3658 * protect us against recursive calling from irq.
3660 asmlinkage
void __sched
preempt_schedule_irq(void)
3662 struct thread_info
*ti
= current_thread_info();
3663 #ifdef CONFIG_PREEMPT_BKL
3664 struct task_struct
*task
= current
;
3665 int saved_lock_depth
;
3667 /* Catch callers which need to be fixed */
3668 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3671 add_preempt_count(PREEMPT_ACTIVE
);
3674 * We keep the big kernel semaphore locked, but we
3675 * clear ->lock_depth so that schedule() doesnt
3676 * auto-release the semaphore:
3678 #ifdef CONFIG_PREEMPT_BKL
3679 saved_lock_depth
= task
->lock_depth
;
3680 task
->lock_depth
= -1;
3684 local_irq_disable();
3685 #ifdef CONFIG_PREEMPT_BKL
3686 task
->lock_depth
= saved_lock_depth
;
3688 sub_preempt_count(PREEMPT_ACTIVE
);
3691 * Check again in case we missed a preemption opportunity
3692 * between schedule and now.
3695 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3698 #endif /* CONFIG_PREEMPT */
3700 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3703 return try_to_wake_up(curr
->private, mode
, sync
);
3705 EXPORT_SYMBOL(default_wake_function
);
3708 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3709 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3710 * number) then we wake all the non-exclusive tasks and one exclusive task.
3712 * There are circumstances in which we can try to wake a task which has already
3713 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3714 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3716 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3717 int nr_exclusive
, int sync
, void *key
)
3719 wait_queue_t
*curr
, *next
;
3721 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3722 unsigned flags
= curr
->flags
;
3724 if (curr
->func(curr
, mode
, sync
, key
) &&
3725 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3731 * __wake_up - wake up threads blocked on a waitqueue.
3733 * @mode: which threads
3734 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3735 * @key: is directly passed to the wakeup function
3737 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3738 int nr_exclusive
, void *key
)
3740 unsigned long flags
;
3742 spin_lock_irqsave(&q
->lock
, flags
);
3743 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3744 spin_unlock_irqrestore(&q
->lock
, flags
);
3746 EXPORT_SYMBOL(__wake_up
);
3749 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3751 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3753 __wake_up_common(q
, mode
, 1, 0, NULL
);
3757 * __wake_up_sync - wake up threads blocked on a waitqueue.
3759 * @mode: which threads
3760 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3762 * The sync wakeup differs that the waker knows that it will schedule
3763 * away soon, so while the target thread will be woken up, it will not
3764 * be migrated to another CPU - ie. the two threads are 'synchronized'
3765 * with each other. This can prevent needless bouncing between CPUs.
3767 * On UP it can prevent extra preemption.
3770 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3772 unsigned long flags
;
3778 if (unlikely(!nr_exclusive
))
3781 spin_lock_irqsave(&q
->lock
, flags
);
3782 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3783 spin_unlock_irqrestore(&q
->lock
, flags
);
3785 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3787 void fastcall
complete(struct completion
*x
)
3789 unsigned long flags
;
3791 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3793 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3795 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3797 EXPORT_SYMBOL(complete
);
3799 void fastcall
complete_all(struct completion
*x
)
3801 unsigned long flags
;
3803 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3804 x
->done
+= UINT_MAX
/2;
3805 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3807 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3809 EXPORT_SYMBOL(complete_all
);
3811 static inline long __sched
3812 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3815 DECLARE_WAITQUEUE(wait
, current
);
3817 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3818 __add_wait_queue_tail(&x
->wait
, &wait
);
3820 if (state
== TASK_INTERRUPTIBLE
&&
3821 signal_pending(current
)) {
3822 __remove_wait_queue(&x
->wait
, &wait
);
3823 return -ERESTARTSYS
;
3825 __set_current_state(state
);
3826 spin_unlock_irq(&x
->wait
.lock
);
3827 timeout
= schedule_timeout(timeout
);
3828 spin_lock_irq(&x
->wait
.lock
);
3830 __remove_wait_queue(&x
->wait
, &wait
);
3834 __remove_wait_queue(&x
->wait
, &wait
);
3841 wait_for_common(struct completion
*x
, long timeout
, int state
)
3845 spin_lock_irq(&x
->wait
.lock
);
3846 timeout
= do_wait_for_common(x
, timeout
, state
);
3847 spin_unlock_irq(&x
->wait
.lock
);
3851 void fastcall __sched
wait_for_completion(struct completion
*x
)
3853 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3855 EXPORT_SYMBOL(wait_for_completion
);
3857 unsigned long fastcall __sched
3858 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3860 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3862 EXPORT_SYMBOL(wait_for_completion_timeout
);
3864 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3866 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3867 if (t
== -ERESTARTSYS
)
3871 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3873 unsigned long fastcall __sched
3874 wait_for_completion_interruptible_timeout(struct completion
*x
,
3875 unsigned long timeout
)
3877 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3879 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3882 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3884 unsigned long flags
;
3887 init_waitqueue_entry(&wait
, current
);
3889 __set_current_state(state
);
3891 spin_lock_irqsave(&q
->lock
, flags
);
3892 __add_wait_queue(q
, &wait
);
3893 spin_unlock(&q
->lock
);
3894 timeout
= schedule_timeout(timeout
);
3895 spin_lock_irq(&q
->lock
);
3896 __remove_wait_queue(q
, &wait
);
3897 spin_unlock_irqrestore(&q
->lock
, flags
);
3902 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3904 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3906 EXPORT_SYMBOL(interruptible_sleep_on
);
3909 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3911 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3913 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3915 void __sched
sleep_on(wait_queue_head_t
*q
)
3917 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3919 EXPORT_SYMBOL(sleep_on
);
3921 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3923 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3925 EXPORT_SYMBOL(sleep_on_timeout
);
3927 #ifdef CONFIG_RT_MUTEXES
3930 * rt_mutex_setprio - set the current priority of a task
3932 * @prio: prio value (kernel-internal form)
3934 * This function changes the 'effective' priority of a task. It does
3935 * not touch ->normal_prio like __setscheduler().
3937 * Used by the rt_mutex code to implement priority inheritance logic.
3939 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3941 unsigned long flags
;
3942 int oldprio
, on_rq
, running
;
3945 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3947 rq
= task_rq_lock(p
, &flags
);
3948 update_rq_clock(rq
);
3951 on_rq
= p
->se
.on_rq
;
3952 running
= task_running(rq
, p
);
3954 dequeue_task(rq
, p
, 0);
3956 p
->sched_class
->put_prev_task(rq
, p
);
3960 p
->sched_class
= &rt_sched_class
;
3962 p
->sched_class
= &fair_sched_class
;
3968 p
->sched_class
->set_curr_task(rq
);
3969 enqueue_task(rq
, p
, 0);
3971 * Reschedule if we are currently running on this runqueue and
3972 * our priority decreased, or if we are not currently running on
3973 * this runqueue and our priority is higher than the current's
3976 if (p
->prio
> oldprio
)
3977 resched_task(rq
->curr
);
3979 check_preempt_curr(rq
, p
);
3982 task_rq_unlock(rq
, &flags
);
3987 void set_user_nice(struct task_struct
*p
, long nice
)
3989 int old_prio
, delta
, on_rq
;
3990 unsigned long flags
;
3993 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3996 * We have to be careful, if called from sys_setpriority(),
3997 * the task might be in the middle of scheduling on another CPU.
3999 rq
= task_rq_lock(p
, &flags
);
4000 update_rq_clock(rq
);
4002 * The RT priorities are set via sched_setscheduler(), but we still
4003 * allow the 'normal' nice value to be set - but as expected
4004 * it wont have any effect on scheduling until the task is
4005 * SCHED_FIFO/SCHED_RR:
4007 if (task_has_rt_policy(p
)) {
4008 p
->static_prio
= NICE_TO_PRIO(nice
);
4011 on_rq
= p
->se
.on_rq
;
4013 dequeue_task(rq
, p
, 0);
4017 p
->static_prio
= NICE_TO_PRIO(nice
);
4020 p
->prio
= effective_prio(p
);
4021 delta
= p
->prio
- old_prio
;
4024 enqueue_task(rq
, p
, 0);
4027 * If the task increased its priority or is running and
4028 * lowered its priority, then reschedule its CPU:
4030 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4031 resched_task(rq
->curr
);
4034 task_rq_unlock(rq
, &flags
);
4036 EXPORT_SYMBOL(set_user_nice
);
4039 * can_nice - check if a task can reduce its nice value
4043 int can_nice(const struct task_struct
*p
, const int nice
)
4045 /* convert nice value [19,-20] to rlimit style value [1,40] */
4046 int nice_rlim
= 20 - nice
;
4048 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4049 capable(CAP_SYS_NICE
));
4052 #ifdef __ARCH_WANT_SYS_NICE
4055 * sys_nice - change the priority of the current process.
4056 * @increment: priority increment
4058 * sys_setpriority is a more generic, but much slower function that
4059 * does similar things.
4061 asmlinkage
long sys_nice(int increment
)
4066 * Setpriority might change our priority at the same moment.
4067 * We don't have to worry. Conceptually one call occurs first
4068 * and we have a single winner.
4070 if (increment
< -40)
4075 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4081 if (increment
< 0 && !can_nice(current
, nice
))
4084 retval
= security_task_setnice(current
, nice
);
4088 set_user_nice(current
, nice
);
4095 * task_prio - return the priority value of a given task.
4096 * @p: the task in question.
4098 * This is the priority value as seen by users in /proc.
4099 * RT tasks are offset by -200. Normal tasks are centered
4100 * around 0, value goes from -16 to +15.
4102 int task_prio(const struct task_struct
*p
)
4104 return p
->prio
- MAX_RT_PRIO
;
4108 * task_nice - return the nice value of a given task.
4109 * @p: the task in question.
4111 int task_nice(const struct task_struct
*p
)
4113 return TASK_NICE(p
);
4115 EXPORT_SYMBOL_GPL(task_nice
);
4118 * idle_cpu - is a given cpu idle currently?
4119 * @cpu: the processor in question.
4121 int idle_cpu(int cpu
)
4123 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4127 * idle_task - return the idle task for a given cpu.
4128 * @cpu: the processor in question.
4130 struct task_struct
*idle_task(int cpu
)
4132 return cpu_rq(cpu
)->idle
;
4136 * find_process_by_pid - find a process with a matching PID value.
4137 * @pid: the pid in question.
4139 static struct task_struct
*find_process_by_pid(pid_t pid
)
4141 return pid
? find_task_by_pid(pid
) : current
;
4144 /* Actually do priority change: must hold rq lock. */
4146 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4148 BUG_ON(p
->se
.on_rq
);
4151 switch (p
->policy
) {
4155 p
->sched_class
= &fair_sched_class
;
4159 p
->sched_class
= &rt_sched_class
;
4163 p
->rt_priority
= prio
;
4164 p
->normal_prio
= normal_prio(p
);
4165 /* we are holding p->pi_lock already */
4166 p
->prio
= rt_mutex_getprio(p
);
4171 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4172 * @p: the task in question.
4173 * @policy: new policy.
4174 * @param: structure containing the new RT priority.
4176 * NOTE that the task may be already dead.
4178 int sched_setscheduler(struct task_struct
*p
, int policy
,
4179 struct sched_param
*param
)
4181 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4182 unsigned long flags
;
4185 /* may grab non-irq protected spin_locks */
4186 BUG_ON(in_interrupt());
4188 /* double check policy once rq lock held */
4190 policy
= oldpolicy
= p
->policy
;
4191 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4192 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4193 policy
!= SCHED_IDLE
)
4196 * Valid priorities for SCHED_FIFO and SCHED_RR are
4197 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4198 * SCHED_BATCH and SCHED_IDLE is 0.
4200 if (param
->sched_priority
< 0 ||
4201 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4202 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4204 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4208 * Allow unprivileged RT tasks to decrease priority:
4210 if (!capable(CAP_SYS_NICE
)) {
4211 if (rt_policy(policy
)) {
4212 unsigned long rlim_rtprio
;
4214 if (!lock_task_sighand(p
, &flags
))
4216 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4217 unlock_task_sighand(p
, &flags
);
4219 /* can't set/change the rt policy */
4220 if (policy
!= p
->policy
&& !rlim_rtprio
)
4223 /* can't increase priority */
4224 if (param
->sched_priority
> p
->rt_priority
&&
4225 param
->sched_priority
> rlim_rtprio
)
4229 * Like positive nice levels, dont allow tasks to
4230 * move out of SCHED_IDLE either:
4232 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4235 /* can't change other user's priorities */
4236 if ((current
->euid
!= p
->euid
) &&
4237 (current
->euid
!= p
->uid
))
4241 retval
= security_task_setscheduler(p
, policy
, param
);
4245 * make sure no PI-waiters arrive (or leave) while we are
4246 * changing the priority of the task:
4248 spin_lock_irqsave(&p
->pi_lock
, flags
);
4250 * To be able to change p->policy safely, the apropriate
4251 * runqueue lock must be held.
4253 rq
= __task_rq_lock(p
);
4254 /* recheck policy now with rq lock held */
4255 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4256 policy
= oldpolicy
= -1;
4257 __task_rq_unlock(rq
);
4258 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4261 update_rq_clock(rq
);
4262 on_rq
= p
->se
.on_rq
;
4263 running
= task_running(rq
, p
);
4265 deactivate_task(rq
, p
, 0);
4267 p
->sched_class
->put_prev_task(rq
, p
);
4271 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4275 p
->sched_class
->set_curr_task(rq
);
4276 activate_task(rq
, p
, 0);
4278 * Reschedule if we are currently running on this runqueue and
4279 * our priority decreased, or if we are not currently running on
4280 * this runqueue and our priority is higher than the current's
4283 if (p
->prio
> oldprio
)
4284 resched_task(rq
->curr
);
4286 check_preempt_curr(rq
, p
);
4289 __task_rq_unlock(rq
);
4290 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4292 rt_mutex_adjust_pi(p
);
4296 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4299 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4301 struct sched_param lparam
;
4302 struct task_struct
*p
;
4305 if (!param
|| pid
< 0)
4307 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4312 p
= find_process_by_pid(pid
);
4314 retval
= sched_setscheduler(p
, policy
, &lparam
);
4321 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4322 * @pid: the pid in question.
4323 * @policy: new policy.
4324 * @param: structure containing the new RT priority.
4326 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4327 struct sched_param __user
*param
)
4329 /* negative values for policy are not valid */
4333 return do_sched_setscheduler(pid
, policy
, param
);
4337 * sys_sched_setparam - set/change the RT priority of a thread
4338 * @pid: the pid in question.
4339 * @param: structure containing the new RT priority.
4341 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4343 return do_sched_setscheduler(pid
, -1, param
);
4347 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4348 * @pid: the pid in question.
4350 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4352 struct task_struct
*p
;
4359 read_lock(&tasklist_lock
);
4360 p
= find_process_by_pid(pid
);
4362 retval
= security_task_getscheduler(p
);
4366 read_unlock(&tasklist_lock
);
4371 * sys_sched_getscheduler - get the RT priority of a thread
4372 * @pid: the pid in question.
4373 * @param: structure containing the RT priority.
4375 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4377 struct sched_param lp
;
4378 struct task_struct
*p
;
4381 if (!param
|| pid
< 0)
4384 read_lock(&tasklist_lock
);
4385 p
= find_process_by_pid(pid
);
4390 retval
= security_task_getscheduler(p
);
4394 lp
.sched_priority
= p
->rt_priority
;
4395 read_unlock(&tasklist_lock
);
4398 * This one might sleep, we cannot do it with a spinlock held ...
4400 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4405 read_unlock(&tasklist_lock
);
4409 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4411 cpumask_t cpus_allowed
;
4412 struct task_struct
*p
;
4415 mutex_lock(&sched_hotcpu_mutex
);
4416 read_lock(&tasklist_lock
);
4418 p
= find_process_by_pid(pid
);
4420 read_unlock(&tasklist_lock
);
4421 mutex_unlock(&sched_hotcpu_mutex
);
4426 * It is not safe to call set_cpus_allowed with the
4427 * tasklist_lock held. We will bump the task_struct's
4428 * usage count and then drop tasklist_lock.
4431 read_unlock(&tasklist_lock
);
4434 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4435 !capable(CAP_SYS_NICE
))
4438 retval
= security_task_setscheduler(p
, 0, NULL
);
4442 cpus_allowed
= cpuset_cpus_allowed(p
);
4443 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4444 retval
= set_cpus_allowed(p
, new_mask
);
4448 mutex_unlock(&sched_hotcpu_mutex
);
4452 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4453 cpumask_t
*new_mask
)
4455 if (len
< sizeof(cpumask_t
)) {
4456 memset(new_mask
, 0, sizeof(cpumask_t
));
4457 } else if (len
> sizeof(cpumask_t
)) {
4458 len
= sizeof(cpumask_t
);
4460 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4464 * sys_sched_setaffinity - set the cpu affinity of a process
4465 * @pid: pid of the process
4466 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4467 * @user_mask_ptr: user-space pointer to the new cpu mask
4469 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4470 unsigned long __user
*user_mask_ptr
)
4475 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4479 return sched_setaffinity(pid
, new_mask
);
4483 * Represents all cpu's present in the system
4484 * In systems capable of hotplug, this map could dynamically grow
4485 * as new cpu's are detected in the system via any platform specific
4486 * method, such as ACPI for e.g.
4489 cpumask_t cpu_present_map __read_mostly
;
4490 EXPORT_SYMBOL(cpu_present_map
);
4493 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4494 EXPORT_SYMBOL(cpu_online_map
);
4496 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4497 EXPORT_SYMBOL(cpu_possible_map
);
4500 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4502 struct task_struct
*p
;
4505 mutex_lock(&sched_hotcpu_mutex
);
4506 read_lock(&tasklist_lock
);
4509 p
= find_process_by_pid(pid
);
4513 retval
= security_task_getscheduler(p
);
4517 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4520 read_unlock(&tasklist_lock
);
4521 mutex_unlock(&sched_hotcpu_mutex
);
4527 * sys_sched_getaffinity - get the cpu affinity of a process
4528 * @pid: pid of the process
4529 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4530 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4532 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4533 unsigned long __user
*user_mask_ptr
)
4538 if (len
< sizeof(cpumask_t
))
4541 ret
= sched_getaffinity(pid
, &mask
);
4545 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4548 return sizeof(cpumask_t
);
4552 * sys_sched_yield - yield the current processor to other threads.
4554 * This function yields the current CPU to other tasks. If there are no
4555 * other threads running on this CPU then this function will return.
4557 asmlinkage
long sys_sched_yield(void)
4559 struct rq
*rq
= this_rq_lock();
4561 schedstat_inc(rq
, yld_count
);
4562 current
->sched_class
->yield_task(rq
);
4565 * Since we are going to call schedule() anyway, there's
4566 * no need to preempt or enable interrupts:
4568 __release(rq
->lock
);
4569 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4570 _raw_spin_unlock(&rq
->lock
);
4571 preempt_enable_no_resched();
4578 static void __cond_resched(void)
4580 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4581 __might_sleep(__FILE__
, __LINE__
);
4584 * The BKS might be reacquired before we have dropped
4585 * PREEMPT_ACTIVE, which could trigger a second
4586 * cond_resched() call.
4589 add_preempt_count(PREEMPT_ACTIVE
);
4591 sub_preempt_count(PREEMPT_ACTIVE
);
4592 } while (need_resched());
4595 int __sched
cond_resched(void)
4597 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4598 system_state
== SYSTEM_RUNNING
) {
4604 EXPORT_SYMBOL(cond_resched
);
4607 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4608 * call schedule, and on return reacquire the lock.
4610 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4611 * operations here to prevent schedule() from being called twice (once via
4612 * spin_unlock(), once by hand).
4614 int cond_resched_lock(spinlock_t
*lock
)
4618 if (need_lockbreak(lock
)) {
4624 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4625 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4626 _raw_spin_unlock(lock
);
4627 preempt_enable_no_resched();
4634 EXPORT_SYMBOL(cond_resched_lock
);
4636 int __sched
cond_resched_softirq(void)
4638 BUG_ON(!in_softirq());
4640 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4648 EXPORT_SYMBOL(cond_resched_softirq
);
4651 * yield - yield the current processor to other threads.
4653 * This is a shortcut for kernel-space yielding - it marks the
4654 * thread runnable and calls sys_sched_yield().
4656 void __sched
yield(void)
4658 set_current_state(TASK_RUNNING
);
4661 EXPORT_SYMBOL(yield
);
4664 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4665 * that process accounting knows that this is a task in IO wait state.
4667 * But don't do that if it is a deliberate, throttling IO wait (this task
4668 * has set its backing_dev_info: the queue against which it should throttle)
4670 void __sched
io_schedule(void)
4672 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4674 delayacct_blkio_start();
4675 atomic_inc(&rq
->nr_iowait
);
4677 atomic_dec(&rq
->nr_iowait
);
4678 delayacct_blkio_end();
4680 EXPORT_SYMBOL(io_schedule
);
4682 long __sched
io_schedule_timeout(long timeout
)
4684 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4687 delayacct_blkio_start();
4688 atomic_inc(&rq
->nr_iowait
);
4689 ret
= schedule_timeout(timeout
);
4690 atomic_dec(&rq
->nr_iowait
);
4691 delayacct_blkio_end();
4696 * sys_sched_get_priority_max - return maximum RT priority.
4697 * @policy: scheduling class.
4699 * this syscall returns the maximum rt_priority that can be used
4700 * by a given scheduling class.
4702 asmlinkage
long sys_sched_get_priority_max(int policy
)
4709 ret
= MAX_USER_RT_PRIO
-1;
4721 * sys_sched_get_priority_min - return minimum RT priority.
4722 * @policy: scheduling class.
4724 * this syscall returns the minimum rt_priority that can be used
4725 * by a given scheduling class.
4727 asmlinkage
long sys_sched_get_priority_min(int policy
)
4745 * sys_sched_rr_get_interval - return the default timeslice of a process.
4746 * @pid: pid of the process.
4747 * @interval: userspace pointer to the timeslice value.
4749 * this syscall writes the default timeslice value of a given process
4750 * into the user-space timespec buffer. A value of '0' means infinity.
4753 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4755 struct task_struct
*p
;
4756 unsigned int time_slice
;
4764 read_lock(&tasklist_lock
);
4765 p
= find_process_by_pid(pid
);
4769 retval
= security_task_getscheduler(p
);
4773 if (p
->policy
== SCHED_FIFO
)
4775 else if (p
->policy
== SCHED_RR
)
4776 time_slice
= DEF_TIMESLICE
;
4778 struct sched_entity
*se
= &p
->se
;
4779 unsigned long flags
;
4782 rq
= task_rq_lock(p
, &flags
);
4783 time_slice
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
4784 task_rq_unlock(rq
, &flags
);
4786 read_unlock(&tasklist_lock
);
4787 jiffies_to_timespec(time_slice
, &t
);
4788 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4792 read_unlock(&tasklist_lock
);
4796 static const char stat_nam
[] = "RSDTtZX";
4798 static void show_task(struct task_struct
*p
)
4800 unsigned long free
= 0;
4803 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4804 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
4805 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4806 #if BITS_PER_LONG == 32
4807 if (state
== TASK_RUNNING
)
4808 printk(KERN_CONT
" running ");
4810 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4812 if (state
== TASK_RUNNING
)
4813 printk(KERN_CONT
" running task ");
4815 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4817 #ifdef CONFIG_DEBUG_STACK_USAGE
4819 unsigned long *n
= end_of_stack(p
);
4822 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4825 printk(KERN_CONT
"%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4827 if (state
!= TASK_RUNNING
)
4828 show_stack(p
, NULL
);
4831 void show_state_filter(unsigned long state_filter
)
4833 struct task_struct
*g
, *p
;
4835 #if BITS_PER_LONG == 32
4837 " task PC stack pid father\n");
4840 " task PC stack pid father\n");
4842 read_lock(&tasklist_lock
);
4843 do_each_thread(g
, p
) {
4845 * reset the NMI-timeout, listing all files on a slow
4846 * console might take alot of time:
4848 touch_nmi_watchdog();
4849 if (!state_filter
|| (p
->state
& state_filter
))
4851 } while_each_thread(g
, p
);
4853 touch_all_softlockup_watchdogs();
4855 #ifdef CONFIG_SCHED_DEBUG
4856 sysrq_sched_debug_show();
4858 read_unlock(&tasklist_lock
);
4860 * Only show locks if all tasks are dumped:
4862 if (state_filter
== -1)
4863 debug_show_all_locks();
4866 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4868 idle
->sched_class
= &idle_sched_class
;
4872 * init_idle - set up an idle thread for a given CPU
4873 * @idle: task in question
4874 * @cpu: cpu the idle task belongs to
4876 * NOTE: this function does not set the idle thread's NEED_RESCHED
4877 * flag, to make booting more robust.
4879 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4881 struct rq
*rq
= cpu_rq(cpu
);
4882 unsigned long flags
;
4885 idle
->se
.exec_start
= sched_clock();
4887 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4888 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4889 __set_task_cpu(idle
, cpu
);
4891 spin_lock_irqsave(&rq
->lock
, flags
);
4892 rq
->curr
= rq
->idle
= idle
;
4893 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4896 spin_unlock_irqrestore(&rq
->lock
, flags
);
4898 /* Set the preempt count _outside_ the spinlocks! */
4899 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4900 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4902 task_thread_info(idle
)->preempt_count
= 0;
4905 * The idle tasks have their own, simple scheduling class:
4907 idle
->sched_class
= &idle_sched_class
;
4911 * In a system that switches off the HZ timer nohz_cpu_mask
4912 * indicates which cpus entered this state. This is used
4913 * in the rcu update to wait only for active cpus. For system
4914 * which do not switch off the HZ timer nohz_cpu_mask should
4915 * always be CPU_MASK_NONE.
4917 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4921 * This is how migration works:
4923 * 1) we queue a struct migration_req structure in the source CPU's
4924 * runqueue and wake up that CPU's migration thread.
4925 * 2) we down() the locked semaphore => thread blocks.
4926 * 3) migration thread wakes up (implicitly it forces the migrated
4927 * thread off the CPU)
4928 * 4) it gets the migration request and checks whether the migrated
4929 * task is still in the wrong runqueue.
4930 * 5) if it's in the wrong runqueue then the migration thread removes
4931 * it and puts it into the right queue.
4932 * 6) migration thread up()s the semaphore.
4933 * 7) we wake up and the migration is done.
4937 * Change a given task's CPU affinity. Migrate the thread to a
4938 * proper CPU and schedule it away if the CPU it's executing on
4939 * is removed from the allowed bitmask.
4941 * NOTE: the caller must have a valid reference to the task, the
4942 * task must not exit() & deallocate itself prematurely. The
4943 * call is not atomic; no spinlocks may be held.
4945 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4947 struct migration_req req
;
4948 unsigned long flags
;
4952 rq
= task_rq_lock(p
, &flags
);
4953 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4958 p
->cpus_allowed
= new_mask
;
4959 /* Can the task run on the task's current CPU? If so, we're done */
4960 if (cpu_isset(task_cpu(p
), new_mask
))
4963 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4964 /* Need help from migration thread: drop lock and wait. */
4965 task_rq_unlock(rq
, &flags
);
4966 wake_up_process(rq
->migration_thread
);
4967 wait_for_completion(&req
.done
);
4968 tlb_migrate_finish(p
->mm
);
4972 task_rq_unlock(rq
, &flags
);
4976 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4979 * Move (not current) task off this cpu, onto dest cpu. We're doing
4980 * this because either it can't run here any more (set_cpus_allowed()
4981 * away from this CPU, or CPU going down), or because we're
4982 * attempting to rebalance this task on exec (sched_exec).
4984 * So we race with normal scheduler movements, but that's OK, as long
4985 * as the task is no longer on this CPU.
4987 * Returns non-zero if task was successfully migrated.
4989 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4991 struct rq
*rq_dest
, *rq_src
;
4994 if (unlikely(cpu_is_offline(dest_cpu
)))
4997 rq_src
= cpu_rq(src_cpu
);
4998 rq_dest
= cpu_rq(dest_cpu
);
5000 double_rq_lock(rq_src
, rq_dest
);
5001 /* Already moved. */
5002 if (task_cpu(p
) != src_cpu
)
5004 /* Affinity changed (again). */
5005 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5008 on_rq
= p
->se
.on_rq
;
5010 deactivate_task(rq_src
, p
, 0);
5012 set_task_cpu(p
, dest_cpu
);
5014 activate_task(rq_dest
, p
, 0);
5015 check_preempt_curr(rq_dest
, p
);
5019 double_rq_unlock(rq_src
, rq_dest
);
5024 * migration_thread - this is a highprio system thread that performs
5025 * thread migration by bumping thread off CPU then 'pushing' onto
5028 static int migration_thread(void *data
)
5030 int cpu
= (long)data
;
5034 BUG_ON(rq
->migration_thread
!= current
);
5036 set_current_state(TASK_INTERRUPTIBLE
);
5037 while (!kthread_should_stop()) {
5038 struct migration_req
*req
;
5039 struct list_head
*head
;
5041 spin_lock_irq(&rq
->lock
);
5043 if (cpu_is_offline(cpu
)) {
5044 spin_unlock_irq(&rq
->lock
);
5048 if (rq
->active_balance
) {
5049 active_load_balance(rq
, cpu
);
5050 rq
->active_balance
= 0;
5053 head
= &rq
->migration_queue
;
5055 if (list_empty(head
)) {
5056 spin_unlock_irq(&rq
->lock
);
5058 set_current_state(TASK_INTERRUPTIBLE
);
5061 req
= list_entry(head
->next
, struct migration_req
, list
);
5062 list_del_init(head
->next
);
5064 spin_unlock(&rq
->lock
);
5065 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5068 complete(&req
->done
);
5070 __set_current_state(TASK_RUNNING
);
5074 /* Wait for kthread_stop */
5075 set_current_state(TASK_INTERRUPTIBLE
);
5076 while (!kthread_should_stop()) {
5078 set_current_state(TASK_INTERRUPTIBLE
);
5080 __set_current_state(TASK_RUNNING
);
5084 #ifdef CONFIG_HOTPLUG_CPU
5086 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5090 local_irq_disable();
5091 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5097 * Figure out where task on dead CPU should go, use force if neccessary.
5098 * NOTE: interrupts should be disabled by the caller
5100 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5102 unsigned long flags
;
5109 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5110 cpus_and(mask
, mask
, p
->cpus_allowed
);
5111 dest_cpu
= any_online_cpu(mask
);
5113 /* On any allowed CPU? */
5114 if (dest_cpu
== NR_CPUS
)
5115 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5117 /* No more Mr. Nice Guy. */
5118 if (dest_cpu
== NR_CPUS
) {
5119 rq
= task_rq_lock(p
, &flags
);
5120 cpus_setall(p
->cpus_allowed
);
5121 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5122 task_rq_unlock(rq
, &flags
);
5125 * Don't tell them about moving exiting tasks or
5126 * kernel threads (both mm NULL), since they never
5129 if (p
->mm
&& printk_ratelimit())
5130 printk(KERN_INFO
"process %d (%s) no "
5131 "longer affine to cpu%d\n",
5132 p
->pid
, p
->comm
, dead_cpu
);
5134 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5138 * While a dead CPU has no uninterruptible tasks queued at this point,
5139 * it might still have a nonzero ->nr_uninterruptible counter, because
5140 * for performance reasons the counter is not stricly tracking tasks to
5141 * their home CPUs. So we just add the counter to another CPU's counter,
5142 * to keep the global sum constant after CPU-down:
5144 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5146 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5147 unsigned long flags
;
5149 local_irq_save(flags
);
5150 double_rq_lock(rq_src
, rq_dest
);
5151 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5152 rq_src
->nr_uninterruptible
= 0;
5153 double_rq_unlock(rq_src
, rq_dest
);
5154 local_irq_restore(flags
);
5157 /* Run through task list and migrate tasks from the dead cpu. */
5158 static void migrate_live_tasks(int src_cpu
)
5160 struct task_struct
*p
, *t
;
5162 read_lock(&tasklist_lock
);
5164 do_each_thread(t
, p
) {
5168 if (task_cpu(p
) == src_cpu
)
5169 move_task_off_dead_cpu(src_cpu
, p
);
5170 } while_each_thread(t
, p
);
5172 read_unlock(&tasklist_lock
);
5176 * activate_idle_task - move idle task to the _front_ of runqueue.
5178 static void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
5180 update_rq_clock(rq
);
5182 if (p
->state
== TASK_UNINTERRUPTIBLE
)
5183 rq
->nr_uninterruptible
--;
5185 enqueue_task(rq
, p
, 0);
5186 inc_nr_running(p
, rq
);
5190 * Schedules idle task to be the next runnable task on current CPU.
5191 * It does so by boosting its priority to highest possible and adding it to
5192 * the _front_ of the runqueue. Used by CPU offline code.
5194 void sched_idle_next(void)
5196 int this_cpu
= smp_processor_id();
5197 struct rq
*rq
= cpu_rq(this_cpu
);
5198 struct task_struct
*p
= rq
->idle
;
5199 unsigned long flags
;
5201 /* cpu has to be offline */
5202 BUG_ON(cpu_online(this_cpu
));
5205 * Strictly not necessary since rest of the CPUs are stopped by now
5206 * and interrupts disabled on the current cpu.
5208 spin_lock_irqsave(&rq
->lock
, flags
);
5210 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5212 /* Add idle task to the _front_ of its priority queue: */
5213 activate_idle_task(p
, rq
);
5215 spin_unlock_irqrestore(&rq
->lock
, flags
);
5219 * Ensures that the idle task is using init_mm right before its cpu goes
5222 void idle_task_exit(void)
5224 struct mm_struct
*mm
= current
->active_mm
;
5226 BUG_ON(cpu_online(smp_processor_id()));
5229 switch_mm(mm
, &init_mm
, current
);
5233 /* called under rq->lock with disabled interrupts */
5234 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5236 struct rq
*rq
= cpu_rq(dead_cpu
);
5238 /* Must be exiting, otherwise would be on tasklist. */
5239 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5241 /* Cannot have done final schedule yet: would have vanished. */
5242 BUG_ON(p
->state
== TASK_DEAD
);
5247 * Drop lock around migration; if someone else moves it,
5248 * that's OK. No task can be added to this CPU, so iteration is
5251 spin_unlock_irq(&rq
->lock
);
5252 move_task_off_dead_cpu(dead_cpu
, p
);
5253 spin_lock_irq(&rq
->lock
);
5258 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5259 static void migrate_dead_tasks(unsigned int dead_cpu
)
5261 struct rq
*rq
= cpu_rq(dead_cpu
);
5262 struct task_struct
*next
;
5265 if (!rq
->nr_running
)
5267 update_rq_clock(rq
);
5268 next
= pick_next_task(rq
, rq
->curr
);
5271 migrate_dead(dead_cpu
, next
);
5275 #endif /* CONFIG_HOTPLUG_CPU */
5277 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5279 static struct ctl_table sd_ctl_dir
[] = {
5281 .procname
= "sched_domain",
5287 static struct ctl_table sd_ctl_root
[] = {
5289 .ctl_name
= CTL_KERN
,
5290 .procname
= "kernel",
5292 .child
= sd_ctl_dir
,
5297 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5299 struct ctl_table
*entry
=
5300 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5305 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5307 struct ctl_table
*entry
;
5310 * In the intermediate directories, both the child directory and
5311 * procname are dynamically allocated and could fail but the mode
5312 * will always be set. In the lowest directory the names are
5313 * static strings and all have proc handlers.
5315 for (entry
= *tablep
; entry
->mode
; entry
++) {
5317 sd_free_ctl_entry(&entry
->child
);
5318 if (entry
->proc_handler
== NULL
)
5319 kfree(entry
->procname
);
5327 set_table_entry(struct ctl_table
*entry
,
5328 const char *procname
, void *data
, int maxlen
,
5329 mode_t mode
, proc_handler
*proc_handler
)
5331 entry
->procname
= procname
;
5333 entry
->maxlen
= maxlen
;
5335 entry
->proc_handler
= proc_handler
;
5338 static struct ctl_table
*
5339 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5341 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5346 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5347 sizeof(long), 0644, proc_doulongvec_minmax
);
5348 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5349 sizeof(long), 0644, proc_doulongvec_minmax
);
5350 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5351 sizeof(int), 0644, proc_dointvec_minmax
);
5352 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5353 sizeof(int), 0644, proc_dointvec_minmax
);
5354 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5355 sizeof(int), 0644, proc_dointvec_minmax
);
5356 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5357 sizeof(int), 0644, proc_dointvec_minmax
);
5358 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5359 sizeof(int), 0644, proc_dointvec_minmax
);
5360 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5361 sizeof(int), 0644, proc_dointvec_minmax
);
5362 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5363 sizeof(int), 0644, proc_dointvec_minmax
);
5364 set_table_entry(&table
[9], "cache_nice_tries",
5365 &sd
->cache_nice_tries
,
5366 sizeof(int), 0644, proc_dointvec_minmax
);
5367 set_table_entry(&table
[10], "flags", &sd
->flags
,
5368 sizeof(int), 0644, proc_dointvec_minmax
);
5369 /* &table[11] is terminator */
5374 static ctl_table
* sd_alloc_ctl_cpu_table(int cpu
)
5376 struct ctl_table
*entry
, *table
;
5377 struct sched_domain
*sd
;
5378 int domain_num
= 0, i
;
5381 for_each_domain(cpu
, sd
)
5383 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5388 for_each_domain(cpu
, sd
) {
5389 snprintf(buf
, 32, "domain%d", i
);
5390 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5392 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5399 static struct ctl_table_header
*sd_sysctl_header
;
5400 static void register_sched_domain_sysctl(void)
5402 int i
, cpu_num
= num_online_cpus();
5403 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5409 sd_ctl_dir
[0].child
= entry
;
5411 for_each_online_cpu(i
) {
5412 snprintf(buf
, 32, "cpu%d", i
);
5413 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5415 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5418 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5421 static void unregister_sched_domain_sysctl(void)
5423 unregister_sysctl_table(sd_sysctl_header
);
5424 sd_sysctl_header
= NULL
;
5425 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5428 static void register_sched_domain_sysctl(void)
5431 static void unregister_sched_domain_sysctl(void)
5437 * migration_call - callback that gets triggered when a CPU is added.
5438 * Here we can start up the necessary migration thread for the new CPU.
5440 static int __cpuinit
5441 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5443 struct task_struct
*p
;
5444 int cpu
= (long)hcpu
;
5445 unsigned long flags
;
5449 case CPU_LOCK_ACQUIRE
:
5450 mutex_lock(&sched_hotcpu_mutex
);
5453 case CPU_UP_PREPARE
:
5454 case CPU_UP_PREPARE_FROZEN
:
5455 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5458 kthread_bind(p
, cpu
);
5459 /* Must be high prio: stop_machine expects to yield to it. */
5460 rq
= task_rq_lock(p
, &flags
);
5461 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5462 task_rq_unlock(rq
, &flags
);
5463 cpu_rq(cpu
)->migration_thread
= p
;
5467 case CPU_ONLINE_FROZEN
:
5468 /* Strictly unneccessary, as first user will wake it. */
5469 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5472 #ifdef CONFIG_HOTPLUG_CPU
5473 case CPU_UP_CANCELED
:
5474 case CPU_UP_CANCELED_FROZEN
:
5475 if (!cpu_rq(cpu
)->migration_thread
)
5477 /* Unbind it from offline cpu so it can run. Fall thru. */
5478 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5479 any_online_cpu(cpu_online_map
));
5480 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5481 cpu_rq(cpu
)->migration_thread
= NULL
;
5485 case CPU_DEAD_FROZEN
:
5486 migrate_live_tasks(cpu
);
5488 kthread_stop(rq
->migration_thread
);
5489 rq
->migration_thread
= NULL
;
5490 /* Idle task back to normal (off runqueue, low prio) */
5491 spin_lock_irq(&rq
->lock
);
5492 update_rq_clock(rq
);
5493 deactivate_task(rq
, rq
->idle
, 0);
5494 rq
->idle
->static_prio
= MAX_PRIO
;
5495 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5496 rq
->idle
->sched_class
= &idle_sched_class
;
5497 migrate_dead_tasks(cpu
);
5498 spin_unlock_irq(&rq
->lock
);
5499 migrate_nr_uninterruptible(rq
);
5500 BUG_ON(rq
->nr_running
!= 0);
5502 /* No need to migrate the tasks: it was best-effort if
5503 * they didn't take sched_hotcpu_mutex. Just wake up
5504 * the requestors. */
5505 spin_lock_irq(&rq
->lock
);
5506 while (!list_empty(&rq
->migration_queue
)) {
5507 struct migration_req
*req
;
5509 req
= list_entry(rq
->migration_queue
.next
,
5510 struct migration_req
, list
);
5511 list_del_init(&req
->list
);
5512 complete(&req
->done
);
5514 spin_unlock_irq(&rq
->lock
);
5517 case CPU_LOCK_RELEASE
:
5518 mutex_unlock(&sched_hotcpu_mutex
);
5524 /* Register at highest priority so that task migration (migrate_all_tasks)
5525 * happens before everything else.
5527 static struct notifier_block __cpuinitdata migration_notifier
= {
5528 .notifier_call
= migration_call
,
5532 int __init
migration_init(void)
5534 void *cpu
= (void *)(long)smp_processor_id();
5537 /* Start one for the boot CPU: */
5538 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5539 BUG_ON(err
== NOTIFY_BAD
);
5540 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5541 register_cpu_notifier(&migration_notifier
);
5549 /* Number of possible processor ids */
5550 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5551 EXPORT_SYMBOL(nr_cpu_ids
);
5553 #ifdef CONFIG_SCHED_DEBUG
5554 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5559 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5563 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5568 struct sched_group
*group
= sd
->groups
;
5569 cpumask_t groupmask
;
5571 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5572 cpus_clear(groupmask
);
5575 for (i
= 0; i
< level
+ 1; i
++)
5577 printk("domain %d: ", level
);
5579 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5580 printk("does not load-balance\n");
5582 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5587 printk("span %s\n", str
);
5589 if (!cpu_isset(cpu
, sd
->span
))
5590 printk(KERN_ERR
"ERROR: domain->span does not contain "
5592 if (!cpu_isset(cpu
, group
->cpumask
))
5593 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5597 for (i
= 0; i
< level
+ 2; i
++)
5603 printk(KERN_ERR
"ERROR: group is NULL\n");
5607 if (!group
->__cpu_power
) {
5608 printk(KERN_CONT
"\n");
5609 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5614 if (!cpus_weight(group
->cpumask
)) {
5615 printk(KERN_CONT
"\n");
5616 printk(KERN_ERR
"ERROR: empty group\n");
5620 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5621 printk(KERN_CONT
"\n");
5622 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5626 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5628 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5629 printk(KERN_CONT
" %s", str
);
5631 group
= group
->next
;
5632 } while (group
!= sd
->groups
);
5633 printk(KERN_CONT
"\n");
5635 if (!cpus_equal(sd
->span
, groupmask
))
5636 printk(KERN_ERR
"ERROR: groups don't span "
5644 if (!cpus_subset(groupmask
, sd
->span
))
5645 printk(KERN_ERR
"ERROR: parent span is not a superset "
5646 "of domain->span\n");
5651 # define sched_domain_debug(sd, cpu) do { } while (0)
5654 static int sd_degenerate(struct sched_domain
*sd
)
5656 if (cpus_weight(sd
->span
) == 1)
5659 /* Following flags need at least 2 groups */
5660 if (sd
->flags
& (SD_LOAD_BALANCE
|
5661 SD_BALANCE_NEWIDLE
|
5665 SD_SHARE_PKG_RESOURCES
)) {
5666 if (sd
->groups
!= sd
->groups
->next
)
5670 /* Following flags don't use groups */
5671 if (sd
->flags
& (SD_WAKE_IDLE
|
5680 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5682 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5684 if (sd_degenerate(parent
))
5687 if (!cpus_equal(sd
->span
, parent
->span
))
5690 /* Does parent contain flags not in child? */
5691 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5692 if (cflags
& SD_WAKE_AFFINE
)
5693 pflags
&= ~SD_WAKE_BALANCE
;
5694 /* Flags needing groups don't count if only 1 group in parent */
5695 if (parent
->groups
== parent
->groups
->next
) {
5696 pflags
&= ~(SD_LOAD_BALANCE
|
5697 SD_BALANCE_NEWIDLE
|
5701 SD_SHARE_PKG_RESOURCES
);
5703 if (~cflags
& pflags
)
5710 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5711 * hold the hotplug lock.
5713 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5715 struct rq
*rq
= cpu_rq(cpu
);
5716 struct sched_domain
*tmp
;
5718 /* Remove the sched domains which do not contribute to scheduling. */
5719 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5720 struct sched_domain
*parent
= tmp
->parent
;
5723 if (sd_parent_degenerate(tmp
, parent
)) {
5724 tmp
->parent
= parent
->parent
;
5726 parent
->parent
->child
= tmp
;
5730 if (sd
&& sd_degenerate(sd
)) {
5736 sched_domain_debug(sd
, cpu
);
5738 rcu_assign_pointer(rq
->sd
, sd
);
5741 /* cpus with isolated domains */
5742 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5744 /* Setup the mask of cpus configured for isolated domains */
5745 static int __init
isolated_cpu_setup(char *str
)
5747 int ints
[NR_CPUS
], i
;
5749 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5750 cpus_clear(cpu_isolated_map
);
5751 for (i
= 1; i
<= ints
[0]; i
++)
5752 if (ints
[i
] < NR_CPUS
)
5753 cpu_set(ints
[i
], cpu_isolated_map
);
5757 __setup("isolcpus=", isolated_cpu_setup
);
5760 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5761 * to a function which identifies what group(along with sched group) a CPU
5762 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5763 * (due to the fact that we keep track of groups covered with a cpumask_t).
5765 * init_sched_build_groups will build a circular linked list of the groups
5766 * covered by the given span, and will set each group's ->cpumask correctly,
5767 * and ->cpu_power to 0.
5770 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5771 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5772 struct sched_group
**sg
))
5774 struct sched_group
*first
= NULL
, *last
= NULL
;
5775 cpumask_t covered
= CPU_MASK_NONE
;
5778 for_each_cpu_mask(i
, span
) {
5779 struct sched_group
*sg
;
5780 int group
= group_fn(i
, cpu_map
, &sg
);
5783 if (cpu_isset(i
, covered
))
5786 sg
->cpumask
= CPU_MASK_NONE
;
5787 sg
->__cpu_power
= 0;
5789 for_each_cpu_mask(j
, span
) {
5790 if (group_fn(j
, cpu_map
, NULL
) != group
)
5793 cpu_set(j
, covered
);
5794 cpu_set(j
, sg
->cpumask
);
5805 #define SD_NODES_PER_DOMAIN 16
5810 * find_next_best_node - find the next node to include in a sched_domain
5811 * @node: node whose sched_domain we're building
5812 * @used_nodes: nodes already in the sched_domain
5814 * Find the next node to include in a given scheduling domain. Simply
5815 * finds the closest node not already in the @used_nodes map.
5817 * Should use nodemask_t.
5819 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5821 int i
, n
, val
, min_val
, best_node
= 0;
5825 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5826 /* Start at @node */
5827 n
= (node
+ i
) % MAX_NUMNODES
;
5829 if (!nr_cpus_node(n
))
5832 /* Skip already used nodes */
5833 if (test_bit(n
, used_nodes
))
5836 /* Simple min distance search */
5837 val
= node_distance(node
, n
);
5839 if (val
< min_val
) {
5845 set_bit(best_node
, used_nodes
);
5850 * sched_domain_node_span - get a cpumask for a node's sched_domain
5851 * @node: node whose cpumask we're constructing
5852 * @size: number of nodes to include in this span
5854 * Given a node, construct a good cpumask for its sched_domain to span. It
5855 * should be one that prevents unnecessary balancing, but also spreads tasks
5858 static cpumask_t
sched_domain_node_span(int node
)
5860 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5861 cpumask_t span
, nodemask
;
5865 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5867 nodemask
= node_to_cpumask(node
);
5868 cpus_or(span
, span
, nodemask
);
5869 set_bit(node
, used_nodes
);
5871 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5872 int next_node
= find_next_best_node(node
, used_nodes
);
5874 nodemask
= node_to_cpumask(next_node
);
5875 cpus_or(span
, span
, nodemask
);
5882 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5885 * SMT sched-domains:
5887 #ifdef CONFIG_SCHED_SMT
5888 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5889 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5891 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5892 struct sched_group
**sg
)
5895 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5901 * multi-core sched-domains:
5903 #ifdef CONFIG_SCHED_MC
5904 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5905 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5908 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5909 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5910 struct sched_group
**sg
)
5913 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
5914 cpus_and(mask
, mask
, *cpu_map
);
5915 group
= first_cpu(mask
);
5917 *sg
= &per_cpu(sched_group_core
, group
);
5920 #elif defined(CONFIG_SCHED_MC)
5921 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5922 struct sched_group
**sg
)
5925 *sg
= &per_cpu(sched_group_core
, cpu
);
5930 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5931 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5933 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5934 struct sched_group
**sg
)
5937 #ifdef CONFIG_SCHED_MC
5938 cpumask_t mask
= cpu_coregroup_map(cpu
);
5939 cpus_and(mask
, mask
, *cpu_map
);
5940 group
= first_cpu(mask
);
5941 #elif defined(CONFIG_SCHED_SMT)
5942 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
5943 cpus_and(mask
, mask
, *cpu_map
);
5944 group
= first_cpu(mask
);
5949 *sg
= &per_cpu(sched_group_phys
, group
);
5955 * The init_sched_build_groups can't handle what we want to do with node
5956 * groups, so roll our own. Now each node has its own list of groups which
5957 * gets dynamically allocated.
5959 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5960 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5962 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5963 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5965 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5966 struct sched_group
**sg
)
5968 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5971 cpus_and(nodemask
, nodemask
, *cpu_map
);
5972 group
= first_cpu(nodemask
);
5975 *sg
= &per_cpu(sched_group_allnodes
, group
);
5979 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5981 struct sched_group
*sg
= group_head
;
5987 for_each_cpu_mask(j
, sg
->cpumask
) {
5988 struct sched_domain
*sd
;
5990 sd
= &per_cpu(phys_domains
, j
);
5991 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5993 * Only add "power" once for each
5999 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6002 } while (sg
!= group_head
);
6007 /* Free memory allocated for various sched_group structures */
6008 static void free_sched_groups(const cpumask_t
*cpu_map
)
6012 for_each_cpu_mask(cpu
, *cpu_map
) {
6013 struct sched_group
**sched_group_nodes
6014 = sched_group_nodes_bycpu
[cpu
];
6016 if (!sched_group_nodes
)
6019 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6020 cpumask_t nodemask
= node_to_cpumask(i
);
6021 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6023 cpus_and(nodemask
, nodemask
, *cpu_map
);
6024 if (cpus_empty(nodemask
))
6034 if (oldsg
!= sched_group_nodes
[i
])
6037 kfree(sched_group_nodes
);
6038 sched_group_nodes_bycpu
[cpu
] = NULL
;
6042 static void free_sched_groups(const cpumask_t
*cpu_map
)
6048 * Initialize sched groups cpu_power.
6050 * cpu_power indicates the capacity of sched group, which is used while
6051 * distributing the load between different sched groups in a sched domain.
6052 * Typically cpu_power for all the groups in a sched domain will be same unless
6053 * there are asymmetries in the topology. If there are asymmetries, group
6054 * having more cpu_power will pickup more load compared to the group having
6057 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6058 * the maximum number of tasks a group can handle in the presence of other idle
6059 * or lightly loaded groups in the same sched domain.
6061 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6063 struct sched_domain
*child
;
6064 struct sched_group
*group
;
6066 WARN_ON(!sd
|| !sd
->groups
);
6068 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6073 sd
->groups
->__cpu_power
= 0;
6076 * For perf policy, if the groups in child domain share resources
6077 * (for example cores sharing some portions of the cache hierarchy
6078 * or SMT), then set this domain groups cpu_power such that each group
6079 * can handle only one task, when there are other idle groups in the
6080 * same sched domain.
6082 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6084 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6085 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6090 * add cpu_power of each child group to this groups cpu_power
6092 group
= child
->groups
;
6094 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6095 group
= group
->next
;
6096 } while (group
!= child
->groups
);
6100 * Build sched domains for a given set of cpus and attach the sched domains
6101 * to the individual cpus
6103 static int build_sched_domains(const cpumask_t
*cpu_map
)
6107 struct sched_group
**sched_group_nodes
= NULL
;
6108 int sd_allnodes
= 0;
6111 * Allocate the per-node list of sched groups
6113 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6115 if (!sched_group_nodes
) {
6116 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6119 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6123 * Set up domains for cpus specified by the cpu_map.
6125 for_each_cpu_mask(i
, *cpu_map
) {
6126 struct sched_domain
*sd
= NULL
, *p
;
6127 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6129 cpus_and(nodemask
, nodemask
, *cpu_map
);
6132 if (cpus_weight(*cpu_map
) >
6133 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6134 sd
= &per_cpu(allnodes_domains
, i
);
6135 *sd
= SD_ALLNODES_INIT
;
6136 sd
->span
= *cpu_map
;
6137 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6143 sd
= &per_cpu(node_domains
, i
);
6145 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6149 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6153 sd
= &per_cpu(phys_domains
, i
);
6155 sd
->span
= nodemask
;
6159 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6161 #ifdef CONFIG_SCHED_MC
6163 sd
= &per_cpu(core_domains
, i
);
6165 sd
->span
= cpu_coregroup_map(i
);
6166 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6169 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6172 #ifdef CONFIG_SCHED_SMT
6174 sd
= &per_cpu(cpu_domains
, i
);
6175 *sd
= SD_SIBLING_INIT
;
6176 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6177 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6180 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6184 #ifdef CONFIG_SCHED_SMT
6185 /* Set up CPU (sibling) groups */
6186 for_each_cpu_mask(i
, *cpu_map
) {
6187 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6188 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6189 if (i
!= first_cpu(this_sibling_map
))
6192 init_sched_build_groups(this_sibling_map
, cpu_map
,
6197 #ifdef CONFIG_SCHED_MC
6198 /* Set up multi-core groups */
6199 for_each_cpu_mask(i
, *cpu_map
) {
6200 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6201 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6202 if (i
!= first_cpu(this_core_map
))
6204 init_sched_build_groups(this_core_map
, cpu_map
,
6205 &cpu_to_core_group
);
6209 /* Set up physical groups */
6210 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6211 cpumask_t nodemask
= node_to_cpumask(i
);
6213 cpus_and(nodemask
, nodemask
, *cpu_map
);
6214 if (cpus_empty(nodemask
))
6217 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6221 /* Set up node groups */
6223 init_sched_build_groups(*cpu_map
, cpu_map
,
6224 &cpu_to_allnodes_group
);
6226 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6227 /* Set up node groups */
6228 struct sched_group
*sg
, *prev
;
6229 cpumask_t nodemask
= node_to_cpumask(i
);
6230 cpumask_t domainspan
;
6231 cpumask_t covered
= CPU_MASK_NONE
;
6234 cpus_and(nodemask
, nodemask
, *cpu_map
);
6235 if (cpus_empty(nodemask
)) {
6236 sched_group_nodes
[i
] = NULL
;
6240 domainspan
= sched_domain_node_span(i
);
6241 cpus_and(domainspan
, domainspan
, *cpu_map
);
6243 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6245 printk(KERN_WARNING
"Can not alloc domain group for "
6249 sched_group_nodes
[i
] = sg
;
6250 for_each_cpu_mask(j
, nodemask
) {
6251 struct sched_domain
*sd
;
6253 sd
= &per_cpu(node_domains
, j
);
6256 sg
->__cpu_power
= 0;
6257 sg
->cpumask
= nodemask
;
6259 cpus_or(covered
, covered
, nodemask
);
6262 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6263 cpumask_t tmp
, notcovered
;
6264 int n
= (i
+ j
) % MAX_NUMNODES
;
6266 cpus_complement(notcovered
, covered
);
6267 cpus_and(tmp
, notcovered
, *cpu_map
);
6268 cpus_and(tmp
, tmp
, domainspan
);
6269 if (cpus_empty(tmp
))
6272 nodemask
= node_to_cpumask(n
);
6273 cpus_and(tmp
, tmp
, nodemask
);
6274 if (cpus_empty(tmp
))
6277 sg
= kmalloc_node(sizeof(struct sched_group
),
6281 "Can not alloc domain group for node %d\n", j
);
6284 sg
->__cpu_power
= 0;
6286 sg
->next
= prev
->next
;
6287 cpus_or(covered
, covered
, tmp
);
6294 /* Calculate CPU power for physical packages and nodes */
6295 #ifdef CONFIG_SCHED_SMT
6296 for_each_cpu_mask(i
, *cpu_map
) {
6297 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6299 init_sched_groups_power(i
, sd
);
6302 #ifdef CONFIG_SCHED_MC
6303 for_each_cpu_mask(i
, *cpu_map
) {
6304 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6306 init_sched_groups_power(i
, sd
);
6310 for_each_cpu_mask(i
, *cpu_map
) {
6311 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6313 init_sched_groups_power(i
, sd
);
6317 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6318 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6321 struct sched_group
*sg
;
6323 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6324 init_numa_sched_groups_power(sg
);
6328 /* Attach the domains */
6329 for_each_cpu_mask(i
, *cpu_map
) {
6330 struct sched_domain
*sd
;
6331 #ifdef CONFIG_SCHED_SMT
6332 sd
= &per_cpu(cpu_domains
, i
);
6333 #elif defined(CONFIG_SCHED_MC)
6334 sd
= &per_cpu(core_domains
, i
);
6336 sd
= &per_cpu(phys_domains
, i
);
6338 cpu_attach_domain(sd
, i
);
6345 free_sched_groups(cpu_map
);
6350 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6352 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6354 cpumask_t cpu_default_map
;
6358 * Setup mask for cpus without special case scheduling requirements.
6359 * For now this just excludes isolated cpus, but could be used to
6360 * exclude other special cases in the future.
6362 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6364 err
= build_sched_domains(&cpu_default_map
);
6366 register_sched_domain_sysctl();
6371 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6373 free_sched_groups(cpu_map
);
6377 * Detach sched domains from a group of cpus specified in cpu_map
6378 * These cpus will now be attached to the NULL domain
6380 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6384 unregister_sched_domain_sysctl();
6386 for_each_cpu_mask(i
, *cpu_map
)
6387 cpu_attach_domain(NULL
, i
);
6388 synchronize_sched();
6389 arch_destroy_sched_domains(cpu_map
);
6392 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6393 static int arch_reinit_sched_domains(void)
6397 mutex_lock(&sched_hotcpu_mutex
);
6398 detach_destroy_domains(&cpu_online_map
);
6399 err
= arch_init_sched_domains(&cpu_online_map
);
6400 mutex_unlock(&sched_hotcpu_mutex
);
6405 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6409 if (buf
[0] != '0' && buf
[0] != '1')
6413 sched_smt_power_savings
= (buf
[0] == '1');
6415 sched_mc_power_savings
= (buf
[0] == '1');
6417 ret
= arch_reinit_sched_domains();
6419 return ret
? ret
: count
;
6422 #ifdef CONFIG_SCHED_MC
6423 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6425 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6427 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6428 const char *buf
, size_t count
)
6430 return sched_power_savings_store(buf
, count
, 0);
6432 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6433 sched_mc_power_savings_store
);
6436 #ifdef CONFIG_SCHED_SMT
6437 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6439 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6441 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6442 const char *buf
, size_t count
)
6444 return sched_power_savings_store(buf
, count
, 1);
6446 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6447 sched_smt_power_savings_store
);
6450 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6454 #ifdef CONFIG_SCHED_SMT
6456 err
= sysfs_create_file(&cls
->kset
.kobj
,
6457 &attr_sched_smt_power_savings
.attr
);
6459 #ifdef CONFIG_SCHED_MC
6460 if (!err
&& mc_capable())
6461 err
= sysfs_create_file(&cls
->kset
.kobj
,
6462 &attr_sched_mc_power_savings
.attr
);
6469 * Force a reinitialization of the sched domains hierarchy. The domains
6470 * and groups cannot be updated in place without racing with the balancing
6471 * code, so we temporarily attach all running cpus to the NULL domain
6472 * which will prevent rebalancing while the sched domains are recalculated.
6474 static int update_sched_domains(struct notifier_block
*nfb
,
6475 unsigned long action
, void *hcpu
)
6478 case CPU_UP_PREPARE
:
6479 case CPU_UP_PREPARE_FROZEN
:
6480 case CPU_DOWN_PREPARE
:
6481 case CPU_DOWN_PREPARE_FROZEN
:
6482 detach_destroy_domains(&cpu_online_map
);
6485 case CPU_UP_CANCELED
:
6486 case CPU_UP_CANCELED_FROZEN
:
6487 case CPU_DOWN_FAILED
:
6488 case CPU_DOWN_FAILED_FROZEN
:
6490 case CPU_ONLINE_FROZEN
:
6492 case CPU_DEAD_FROZEN
:
6494 * Fall through and re-initialise the domains.
6501 /* The hotplug lock is already held by cpu_up/cpu_down */
6502 arch_init_sched_domains(&cpu_online_map
);
6507 void __init
sched_init_smp(void)
6509 cpumask_t non_isolated_cpus
;
6511 mutex_lock(&sched_hotcpu_mutex
);
6512 arch_init_sched_domains(&cpu_online_map
);
6513 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6514 if (cpus_empty(non_isolated_cpus
))
6515 cpu_set(smp_processor_id(), non_isolated_cpus
);
6516 mutex_unlock(&sched_hotcpu_mutex
);
6517 /* XXX: Theoretical race here - CPU may be hotplugged now */
6518 hotcpu_notifier(update_sched_domains
, 0);
6520 /* Move init over to a non-isolated CPU */
6521 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6525 void __init
sched_init_smp(void)
6528 #endif /* CONFIG_SMP */
6530 int in_sched_functions(unsigned long addr
)
6532 /* Linker adds these: start and end of __sched functions */
6533 extern char __sched_text_start
[], __sched_text_end
[];
6535 return in_lock_functions(addr
) ||
6536 (addr
>= (unsigned long)__sched_text_start
6537 && addr
< (unsigned long)__sched_text_end
);
6540 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6542 cfs_rq
->tasks_timeline
= RB_ROOT
;
6543 #ifdef CONFIG_FAIR_GROUP_SCHED
6546 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6549 void __init
sched_init(void)
6551 int highest_cpu
= 0;
6554 for_each_possible_cpu(i
) {
6555 struct rt_prio_array
*array
;
6559 spin_lock_init(&rq
->lock
);
6560 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6563 init_cfs_rq(&rq
->cfs
, rq
);
6564 #ifdef CONFIG_FAIR_GROUP_SCHED
6565 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6567 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6568 struct sched_entity
*se
=
6569 &per_cpu(init_sched_entity
, i
);
6571 init_cfs_rq_p
[i
] = cfs_rq
;
6572 init_cfs_rq(cfs_rq
, rq
);
6573 cfs_rq
->tg
= &init_task_group
;
6574 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6575 &rq
->leaf_cfs_rq_list
);
6577 init_sched_entity_p
[i
] = se
;
6578 se
->cfs_rq
= &rq
->cfs
;
6580 se
->load
.weight
= init_task_group_load
;
6581 se
->load
.inv_weight
=
6582 div64_64(1ULL<<32, init_task_group_load
);
6585 init_task_group
.shares
= init_task_group_load
;
6586 spin_lock_init(&init_task_group
.lock
);
6589 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6590 rq
->cpu_load
[j
] = 0;
6593 rq
->active_balance
= 0;
6594 rq
->next_balance
= jiffies
;
6597 rq
->migration_thread
= NULL
;
6598 INIT_LIST_HEAD(&rq
->migration_queue
);
6600 atomic_set(&rq
->nr_iowait
, 0);
6602 array
= &rq
->rt
.active
;
6603 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6604 INIT_LIST_HEAD(array
->queue
+ j
);
6605 __clear_bit(j
, array
->bitmap
);
6608 /* delimiter for bitsearch: */
6609 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6612 set_load_weight(&init_task
);
6614 #ifdef CONFIG_PREEMPT_NOTIFIERS
6615 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6619 nr_cpu_ids
= highest_cpu
+ 1;
6620 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6623 #ifdef CONFIG_RT_MUTEXES
6624 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6628 * The boot idle thread does lazy MMU switching as well:
6630 atomic_inc(&init_mm
.mm_count
);
6631 enter_lazy_tlb(&init_mm
, current
);
6634 * Make us the idle thread. Technically, schedule() should not be
6635 * called from this thread, however somewhere below it might be,
6636 * but because we are the idle thread, we just pick up running again
6637 * when this runqueue becomes "idle".
6639 init_idle(current
, smp_processor_id());
6641 * During early bootup we pretend to be a normal task:
6643 current
->sched_class
= &fair_sched_class
;
6646 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6647 void __might_sleep(char *file
, int line
)
6650 static unsigned long prev_jiffy
; /* ratelimiting */
6652 if ((in_atomic() || irqs_disabled()) &&
6653 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6654 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6656 prev_jiffy
= jiffies
;
6657 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6658 " context at %s:%d\n", file
, line
);
6659 printk("in_atomic():%d, irqs_disabled():%d\n",
6660 in_atomic(), irqs_disabled());
6661 debug_show_held_locks(current
);
6662 if (irqs_disabled())
6663 print_irqtrace_events(current
);
6668 EXPORT_SYMBOL(__might_sleep
);
6671 #ifdef CONFIG_MAGIC_SYSRQ
6672 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6675 update_rq_clock(rq
);
6676 on_rq
= p
->se
.on_rq
;
6678 deactivate_task(rq
, p
, 0);
6679 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6681 activate_task(rq
, p
, 0);
6682 resched_task(rq
->curr
);
6686 void normalize_rt_tasks(void)
6688 struct task_struct
*g
, *p
;
6689 unsigned long flags
;
6692 read_lock_irq(&tasklist_lock
);
6693 do_each_thread(g
, p
) {
6695 * Only normalize user tasks:
6700 p
->se
.exec_start
= 0;
6701 #ifdef CONFIG_SCHEDSTATS
6702 p
->se
.wait_start
= 0;
6703 p
->se
.sleep_start
= 0;
6704 p
->se
.block_start
= 0;
6706 task_rq(p
)->clock
= 0;
6710 * Renice negative nice level userspace
6713 if (TASK_NICE(p
) < 0 && p
->mm
)
6714 set_user_nice(p
, 0);
6718 spin_lock_irqsave(&p
->pi_lock
, flags
);
6719 rq
= __task_rq_lock(p
);
6721 normalize_task(rq
, p
);
6723 __task_rq_unlock(rq
);
6724 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6725 } while_each_thread(g
, p
);
6727 read_unlock_irq(&tasklist_lock
);
6730 #endif /* CONFIG_MAGIC_SYSRQ */
6734 * These functions are only useful for the IA64 MCA handling.
6736 * They can only be called when the whole system has been
6737 * stopped - every CPU needs to be quiescent, and no scheduling
6738 * activity can take place. Using them for anything else would
6739 * be a serious bug, and as a result, they aren't even visible
6740 * under any other configuration.
6744 * curr_task - return the current task for a given cpu.
6745 * @cpu: the processor in question.
6747 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6749 struct task_struct
*curr_task(int cpu
)
6751 return cpu_curr(cpu
);
6755 * set_curr_task - set the current task for a given cpu.
6756 * @cpu: the processor in question.
6757 * @p: the task pointer to set.
6759 * Description: This function must only be used when non-maskable interrupts
6760 * are serviced on a separate stack. It allows the architecture to switch the
6761 * notion of the current task on a cpu in a non-blocking manner. This function
6762 * must be called with all CPU's synchronized, and interrupts disabled, the
6763 * and caller must save the original value of the current task (see
6764 * curr_task() above) and restore that value before reenabling interrupts and
6765 * re-starting the system.
6767 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6769 void set_curr_task(int cpu
, struct task_struct
*p
)
6776 #ifdef CONFIG_FAIR_GROUP_SCHED
6778 /* allocate runqueue etc for a new task group */
6779 struct task_group
*sched_create_group(void)
6781 struct task_group
*tg
;
6782 struct cfs_rq
*cfs_rq
;
6783 struct sched_entity
*se
;
6787 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6789 return ERR_PTR(-ENOMEM
);
6791 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
6794 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
6798 for_each_possible_cpu(i
) {
6801 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
6806 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
6811 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
6812 memset(se
, 0, sizeof(struct sched_entity
));
6814 tg
->cfs_rq
[i
] = cfs_rq
;
6815 init_cfs_rq(cfs_rq
, rq
);
6819 se
->cfs_rq
= &rq
->cfs
;
6821 se
->load
.weight
= NICE_0_LOAD
;
6822 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6826 for_each_possible_cpu(i
) {
6828 cfs_rq
= tg
->cfs_rq
[i
];
6829 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6832 tg
->shares
= NICE_0_LOAD
;
6833 spin_lock_init(&tg
->lock
);
6838 for_each_possible_cpu(i
) {
6840 kfree(tg
->cfs_rq
[i
]);
6848 return ERR_PTR(-ENOMEM
);
6851 /* rcu callback to free various structures associated with a task group */
6852 static void free_sched_group(struct rcu_head
*rhp
)
6854 struct cfs_rq
*cfs_rq
= container_of(rhp
, struct cfs_rq
, rcu
);
6855 struct task_group
*tg
= cfs_rq
->tg
;
6856 struct sched_entity
*se
;
6859 /* now it should be safe to free those cfs_rqs */
6860 for_each_possible_cpu(i
) {
6861 cfs_rq
= tg
->cfs_rq
[i
];
6873 /* Destroy runqueue etc associated with a task group */
6874 void sched_destroy_group(struct task_group
*tg
)
6876 struct cfs_rq
*cfs_rq
;
6879 for_each_possible_cpu(i
) {
6880 cfs_rq
= tg
->cfs_rq
[i
];
6881 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
6884 cfs_rq
= tg
->cfs_rq
[0];
6886 /* wait for possible concurrent references to cfs_rqs complete */
6887 call_rcu(&cfs_rq
->rcu
, free_sched_group
);
6890 /* change task's runqueue when it moves between groups.
6891 * The caller of this function should have put the task in its new group
6892 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6893 * reflect its new group.
6895 void sched_move_task(struct task_struct
*tsk
)
6898 unsigned long flags
;
6901 rq
= task_rq_lock(tsk
, &flags
);
6903 if (tsk
->sched_class
!= &fair_sched_class
)
6906 update_rq_clock(rq
);
6908 running
= task_running(rq
, tsk
);
6909 on_rq
= tsk
->se
.on_rq
;
6912 dequeue_task(rq
, tsk
, 0);
6913 if (unlikely(running
))
6914 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6917 set_task_cfs_rq(tsk
);
6920 if (unlikely(running
))
6921 tsk
->sched_class
->set_curr_task(rq
);
6922 enqueue_task(rq
, tsk
, 0);
6926 task_rq_unlock(rq
, &flags
);
6929 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
6931 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
6932 struct rq
*rq
= cfs_rq
->rq
;
6935 spin_lock_irq(&rq
->lock
);
6939 dequeue_entity(cfs_rq
, se
, 0);
6941 se
->load
.weight
= shares
;
6942 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
6945 enqueue_entity(cfs_rq
, se
, 0);
6947 spin_unlock_irq(&rq
->lock
);
6950 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6954 spin_lock(&tg
->lock
);
6955 if (tg
->shares
== shares
)
6958 tg
->shares
= shares
;
6959 for_each_possible_cpu(i
)
6960 set_se_shares(tg
->se
[i
], shares
);
6963 spin_unlock(&tg
->lock
);
6967 unsigned long sched_group_shares(struct task_group
*tg
)
6972 #endif /* CONFIG_FAIR_GROUP_SCHED */