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) ((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 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
122 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
131 sg
->__cpu_power
+= val
;
132 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
136 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio
)
145 if (static_prio
== NICE_TO_PRIO(19))
148 if (static_prio
< NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
151 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
154 static inline int rt_policy(int policy
)
156 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
161 static inline int task_has_rt_policy(struct task_struct
*p
)
163 return rt_policy(p
->policy
);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array
{
170 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
171 struct list_head queue
[MAX_RT_PRIO
];
174 #ifdef CONFIG_FAIR_GROUP_SCHED
176 #include <linux/container.h>
180 /* task group related information */
182 struct container_subsys_state css
;
183 /* schedulable entities of this group on each cpu */
184 struct sched_entity
**se
;
185 /* runqueue "owned" by this group on each cpu */
186 struct cfs_rq
**cfs_rq
;
187 unsigned long shares
;
190 /* Default task group's sched entity on each cpu */
191 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
192 /* Default task group's cfs_rq on each cpu */
193 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
195 static struct sched_entity
*init_sched_entity_p
[CONFIG_NR_CPUS
];
196 static struct cfs_rq
*init_cfs_rq_p
[CONFIG_NR_CPUS
];
198 /* Default task group.
199 * Every task in system belong to this group at bootup.
201 static struct task_grp init_task_grp
= {
202 .se
= init_sched_entity_p
,
203 .cfs_rq
= init_cfs_rq_p
,
206 /* return group to which a task belongs */
207 static inline struct task_grp
*task_grp(struct task_struct
*p
)
209 return container_of(task_subsys_state(p
, cpu_subsys_id
),
210 struct task_grp
, css
);
213 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
214 static inline void set_task_cfs_rq(struct task_struct
*p
)
216 p
->se
.cfs_rq
= task_grp(p
)->cfs_rq
[task_cpu(p
)];
217 p
->se
.parent
= task_grp(p
)->se
[task_cpu(p
)];
222 static inline void set_task_cfs_rq(struct task_struct
*p
) { }
224 #endif /* CONFIG_FAIR_GROUP_SCHED */
226 /* CFS-related fields in a runqueue */
228 struct load_weight load
;
229 unsigned long nr_running
;
234 struct rb_root tasks_timeline
;
235 struct rb_node
*rb_leftmost
;
236 struct rb_node
*rb_load_balance_curr
;
237 /* 'curr' points to currently running entity on this cfs_rq.
238 * It is set to NULL otherwise (i.e when none are currently running).
240 struct sched_entity
*curr
;
241 #ifdef CONFIG_FAIR_GROUP_SCHED
242 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
244 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
245 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
246 * (like users, containers etc.)
248 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
249 * list is used during load balance.
251 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
252 struct task_grp
*tg
; /* group that "owns" this runqueue */
256 /* Real-Time classes' related field in a runqueue: */
258 struct rt_prio_array active
;
259 int rt_load_balance_idx
;
260 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
264 * This is the main, per-CPU runqueue data structure.
266 * Locking rule: those places that want to lock multiple runqueues
267 * (such as the load balancing or the thread migration code), lock
268 * acquire operations must be ordered by ascending &runqueue.
271 spinlock_t lock
; /* runqueue lock */
274 * nr_running and cpu_load should be in the same cacheline because
275 * remote CPUs use both these fields when doing load calculation.
277 unsigned long nr_running
;
278 #define CPU_LOAD_IDX_MAX 5
279 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
280 unsigned char idle_at_tick
;
282 unsigned char in_nohz_recently
;
284 struct load_weight load
; /* capture load from *all* tasks on this cpu */
285 unsigned long nr_load_updates
;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
295 * This is part of a global counter where only the total sum
296 * over all CPUs matters. A task can increase this counter on
297 * one CPU and if it got migrated afterwards it may decrease
298 * it on another CPU. Always updated under the runqueue lock:
300 unsigned long nr_uninterruptible
;
302 struct task_struct
*curr
, *idle
;
303 unsigned long next_balance
;
304 struct mm_struct
*prev_mm
;
306 u64 clock
, prev_clock_raw
;
309 unsigned int clock_warps
, clock_overflows
;
311 unsigned int clock_deep_idle_events
;
317 struct sched_domain
*sd
;
319 /* For active balancing */
322 int cpu
; /* cpu of this runqueue */
324 struct task_struct
*migration_thread
;
325 struct list_head migration_queue
;
328 #ifdef CONFIG_SCHEDSTATS
330 struct sched_info rq_sched_info
;
332 /* sys_sched_yield() stats */
333 unsigned long yld_exp_empty
;
334 unsigned long yld_act_empty
;
335 unsigned long yld_both_empty
;
336 unsigned long yld_cnt
;
338 /* schedule() stats */
339 unsigned long sched_switch
;
340 unsigned long sched_cnt
;
341 unsigned long sched_goidle
;
343 /* try_to_wake_up() stats */
344 unsigned long ttwu_cnt
;
345 unsigned long ttwu_local
;
347 struct lock_class_key rq_lock_key
;
350 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
351 static DEFINE_MUTEX(sched_hotcpu_mutex
);
353 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
355 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
358 static inline int cpu_of(struct rq
*rq
)
368 * Update the per-runqueue clock, as finegrained as the platform can give
369 * us, but without assuming monotonicity, etc.:
371 static void __update_rq_clock(struct rq
*rq
)
373 u64 prev_raw
= rq
->prev_clock_raw
;
374 u64 now
= sched_clock();
375 s64 delta
= now
- prev_raw
;
376 u64 clock
= rq
->clock
;
378 #ifdef CONFIG_SCHED_DEBUG
379 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
382 * Protect against sched_clock() occasionally going backwards:
384 if (unlikely(delta
< 0)) {
389 * Catch too large forward jumps too:
391 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
392 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
393 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
396 rq
->clock_overflows
++;
398 if (unlikely(delta
> rq
->clock_max_delta
))
399 rq
->clock_max_delta
= delta
;
404 rq
->prev_clock_raw
= now
;
408 static void update_rq_clock(struct rq
*rq
)
410 if (likely(smp_processor_id() == cpu_of(rq
)))
411 __update_rq_clock(rq
);
415 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
416 * See detach_destroy_domains: synchronize_sched for details.
418 * The domain tree of any CPU may only be accessed from within
419 * preempt-disabled sections.
421 #define for_each_domain(cpu, __sd) \
422 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
424 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
425 #define this_rq() (&__get_cpu_var(runqueues))
426 #define task_rq(p) cpu_rq(task_cpu(p))
427 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
430 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
432 #ifdef CONFIG_SCHED_DEBUG
433 # define const_debug __read_mostly
435 # define const_debug static const
439 * Debugging: various feature bits
442 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
443 SCHED_FEAT_START_DEBIT
= 2,
444 SCHED_FEAT_USE_TREE_AVG
= 4,
445 SCHED_FEAT_APPROX_AVG
= 8,
448 const_debug
unsigned int sysctl_sched_features
=
449 SCHED_FEAT_NEW_FAIR_SLEEPERS
*1 |
450 SCHED_FEAT_START_DEBIT
*1 |
451 SCHED_FEAT_USE_TREE_AVG
*0 |
452 SCHED_FEAT_APPROX_AVG
*0;
454 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
457 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
458 * clock constructed from sched_clock():
460 unsigned long long cpu_clock(int cpu
)
462 unsigned long long now
;
466 local_irq_save(flags
);
470 local_irq_restore(flags
);
475 #ifndef prepare_arch_switch
476 # define prepare_arch_switch(next) do { } while (0)
478 #ifndef finish_arch_switch
479 # define finish_arch_switch(prev) do { } while (0)
482 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
483 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
485 return rq
->curr
== p
;
488 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
492 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
494 #ifdef CONFIG_DEBUG_SPINLOCK
495 /* this is a valid case when another task releases the spinlock */
496 rq
->lock
.owner
= current
;
499 * If we are tracking spinlock dependencies then we have to
500 * fix up the runqueue lock - which gets 'carried over' from
503 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
505 spin_unlock_irq(&rq
->lock
);
508 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
509 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
514 return rq
->curr
== p
;
518 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
522 * We can optimise this out completely for !SMP, because the
523 * SMP rebalancing from interrupt is the only thing that cares
528 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
529 spin_unlock_irq(&rq
->lock
);
531 spin_unlock(&rq
->lock
);
535 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
539 * After ->oncpu is cleared, the task can be moved to a different CPU.
540 * We must ensure this doesn't happen until the switch is completely
546 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
550 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
553 * __task_rq_lock - lock the runqueue a given task resides on.
554 * Must be called interrupts disabled.
556 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
563 spin_lock(&rq
->lock
);
564 if (unlikely(rq
!= task_rq(p
))) {
565 spin_unlock(&rq
->lock
);
566 goto repeat_lock_task
;
572 * task_rq_lock - lock the runqueue a given task resides on and disable
573 * interrupts. Note the ordering: we can safely lookup the task_rq without
574 * explicitly disabling preemption.
576 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
582 local_irq_save(*flags
);
584 spin_lock(&rq
->lock
);
585 if (unlikely(rq
!= task_rq(p
))) {
586 spin_unlock_irqrestore(&rq
->lock
, *flags
);
587 goto repeat_lock_task
;
592 static inline void __task_rq_unlock(struct rq
*rq
)
595 spin_unlock(&rq
->lock
);
598 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
601 spin_unlock_irqrestore(&rq
->lock
, *flags
);
605 * this_rq_lock - lock this runqueue and disable interrupts.
607 static inline struct rq
*this_rq_lock(void)
614 spin_lock(&rq
->lock
);
620 * We are going deep-idle (irqs are disabled):
622 void sched_clock_idle_sleep_event(void)
624 struct rq
*rq
= cpu_rq(smp_processor_id());
626 spin_lock(&rq
->lock
);
627 __update_rq_clock(rq
);
628 spin_unlock(&rq
->lock
);
629 rq
->clock_deep_idle_events
++;
631 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
634 * We just idled delta nanoseconds (called with irqs disabled):
636 void sched_clock_idle_wakeup_event(u64 delta_ns
)
638 struct rq
*rq
= cpu_rq(smp_processor_id());
639 u64 now
= sched_clock();
641 rq
->idle_clock
+= delta_ns
;
643 * Override the previous timestamp and ignore all
644 * sched_clock() deltas that occured while we idled,
645 * and use the PM-provided delta_ns to advance the
648 spin_lock(&rq
->lock
);
649 rq
->prev_clock_raw
= now
;
650 rq
->clock
+= delta_ns
;
651 spin_unlock(&rq
->lock
);
653 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
656 * resched_task - mark a task 'to be rescheduled now'.
658 * On UP this means the setting of the need_resched flag, on SMP it
659 * might also involve a cross-CPU call to trigger the scheduler on
664 #ifndef tsk_is_polling
665 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
668 static void resched_task(struct task_struct
*p
)
672 assert_spin_locked(&task_rq(p
)->lock
);
674 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
677 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
680 if (cpu
== smp_processor_id())
683 /* NEED_RESCHED must be visible before we test polling */
685 if (!tsk_is_polling(p
))
686 smp_send_reschedule(cpu
);
689 static void resched_cpu(int cpu
)
691 struct rq
*rq
= cpu_rq(cpu
);
694 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
696 resched_task(cpu_curr(cpu
));
697 spin_unlock_irqrestore(&rq
->lock
, flags
);
700 static inline void resched_task(struct task_struct
*p
)
702 assert_spin_locked(&task_rq(p
)->lock
);
703 set_tsk_need_resched(p
);
707 #if BITS_PER_LONG == 32
708 # define WMULT_CONST (~0UL)
710 # define WMULT_CONST (1UL << 32)
713 #define WMULT_SHIFT 32
716 * Shift right and round:
718 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
721 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
722 struct load_weight
*lw
)
726 if (unlikely(!lw
->inv_weight
))
727 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
729 tmp
= (u64
)delta_exec
* weight
;
731 * Check whether we'd overflow the 64-bit multiplication:
733 if (unlikely(tmp
> WMULT_CONST
))
734 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
737 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
739 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
742 static inline unsigned long
743 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
745 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
748 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
753 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
759 * To aid in avoiding the subversion of "niceness" due to uneven distribution
760 * of tasks with abnormal "nice" values across CPUs the contribution that
761 * each task makes to its run queue's load is weighted according to its
762 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
763 * scaled version of the new time slice allocation that they receive on time
767 #define WEIGHT_IDLEPRIO 2
768 #define WMULT_IDLEPRIO (1 << 31)
771 * Nice levels are multiplicative, with a gentle 10% change for every
772 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
773 * nice 1, it will get ~10% less CPU time than another CPU-bound task
774 * that remained on nice 0.
776 * The "10% effect" is relative and cumulative: from _any_ nice level,
777 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
778 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
779 * If a task goes up by ~10% and another task goes down by ~10% then
780 * the relative distance between them is ~25%.)
782 static const int prio_to_weight
[40] = {
783 /* -20 */ 88761, 71755, 56483, 46273, 36291,
784 /* -15 */ 29154, 23254, 18705, 14949, 11916,
785 /* -10 */ 9548, 7620, 6100, 4904, 3906,
786 /* -5 */ 3121, 2501, 1991, 1586, 1277,
787 /* 0 */ 1024, 820, 655, 526, 423,
788 /* 5 */ 335, 272, 215, 172, 137,
789 /* 10 */ 110, 87, 70, 56, 45,
790 /* 15 */ 36, 29, 23, 18, 15,
794 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
796 * In cases where the weight does not change often, we can use the
797 * precalculated inverse to speed up arithmetics by turning divisions
798 * into multiplications:
800 static const u32 prio_to_wmult
[40] = {
801 /* -20 */ 48388, 59856, 76040, 92818, 118348,
802 /* -15 */ 147320, 184698, 229616, 287308, 360437,
803 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
804 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
805 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
806 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
807 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
808 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
811 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
814 * runqueue iterator, to support SMP load-balancing between different
815 * scheduling classes, without having to expose their internal data
816 * structures to the load-balancing proper:
820 struct task_struct
*(*start
)(void *);
821 struct task_struct
*(*next
)(void *);
824 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
825 unsigned long max_nr_move
, unsigned long max_load_move
,
826 struct sched_domain
*sd
, enum cpu_idle_type idle
,
827 int *all_pinned
, unsigned long *load_moved
,
828 int *this_best_prio
, struct rq_iterator
*iterator
);
830 #include "sched_stats.h"
831 #include "sched_rt.c"
832 #include "sched_fair.c"
833 #include "sched_idletask.c"
834 #ifdef CONFIG_SCHED_DEBUG
835 # include "sched_debug.c"
838 #define sched_class_highest (&rt_sched_class)
841 * Update delta_exec, delta_fair fields for rq.
843 * delta_fair clock advances at a rate inversely proportional to
844 * total load (rq->load.weight) on the runqueue, while
845 * delta_exec advances at the same rate as wall-clock (provided
848 * delta_exec / delta_fair is a measure of the (smoothened) load on this
849 * runqueue over any given interval. This (smoothened) load is used
850 * during load balance.
852 * This function is called /before/ updating rq->load
853 * and when switching tasks.
855 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
857 update_load_add(&rq
->load
, p
->se
.load
.weight
);
860 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
862 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
865 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
871 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
877 static void set_load_weight(struct task_struct
*p
)
879 if (task_has_rt_policy(p
)) {
880 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
881 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
886 * SCHED_IDLE tasks get minimal weight:
888 if (p
->policy
== SCHED_IDLE
) {
889 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
890 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
894 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
895 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
898 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
900 sched_info_queued(p
);
901 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
905 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
907 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
912 * __normal_prio - return the priority that is based on the static prio
914 static inline int __normal_prio(struct task_struct
*p
)
916 return p
->static_prio
;
920 * Calculate the expected normal priority: i.e. priority
921 * without taking RT-inheritance into account. Might be
922 * boosted by interactivity modifiers. Changes upon fork,
923 * setprio syscalls, and whenever the interactivity
924 * estimator recalculates.
926 static inline int normal_prio(struct task_struct
*p
)
930 if (task_has_rt_policy(p
))
931 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
933 prio
= __normal_prio(p
);
938 * Calculate the current priority, i.e. the priority
939 * taken into account by the scheduler. This value might
940 * be boosted by RT tasks, or might be boosted by
941 * interactivity modifiers. Will be RT if the task got
942 * RT-boosted. If not then it returns p->normal_prio.
944 static int effective_prio(struct task_struct
*p
)
946 p
->normal_prio
= normal_prio(p
);
948 * If we are RT tasks or we were boosted to RT priority,
949 * keep the priority unchanged. Otherwise, update priority
950 * to the normal priority:
952 if (!rt_prio(p
->prio
))
953 return p
->normal_prio
;
958 * activate_task - move a task to the runqueue.
960 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
962 if (p
->state
== TASK_UNINTERRUPTIBLE
)
963 rq
->nr_uninterruptible
--;
965 enqueue_task(rq
, p
, wakeup
);
966 inc_nr_running(p
, rq
);
970 * activate_idle_task - move idle task to the _front_ of runqueue.
972 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
976 if (p
->state
== TASK_UNINTERRUPTIBLE
)
977 rq
->nr_uninterruptible
--;
979 enqueue_task(rq
, p
, 0);
980 inc_nr_running(p
, rq
);
984 * deactivate_task - remove a task from the runqueue.
986 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
988 if (p
->state
== TASK_UNINTERRUPTIBLE
)
989 rq
->nr_uninterruptible
++;
991 dequeue_task(rq
, p
, sleep
);
992 dec_nr_running(p
, rq
);
996 * task_curr - is this task currently executing on a CPU?
997 * @p: the task in question.
999 inline int task_curr(const struct task_struct
*p
)
1001 return cpu_curr(task_cpu(p
)) == p
;
1004 /* Used instead of source_load when we know the type == 0 */
1005 unsigned long weighted_cpuload(const int cpu
)
1007 return cpu_rq(cpu
)->load
.weight
;
1010 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1013 task_thread_info(p
)->cpu
= cpu
;
1020 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1022 int old_cpu
= task_cpu(p
);
1023 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1026 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1028 #ifdef CONFIG_SCHEDSTATS
1029 if (p
->se
.wait_start
)
1030 p
->se
.wait_start
-= clock_offset
;
1031 if (p
->se
.sleep_start
)
1032 p
->se
.sleep_start
-= clock_offset
;
1033 if (p
->se
.block_start
)
1034 p
->se
.block_start
-= clock_offset
;
1036 if (likely(new_rq
->cfs
.min_vruntime
))
1037 p
->se
.vruntime
-= old_rq
->cfs
.min_vruntime
-
1038 new_rq
->cfs
.min_vruntime
;
1040 __set_task_cpu(p
, new_cpu
);
1043 struct migration_req
{
1044 struct list_head list
;
1046 struct task_struct
*task
;
1049 struct completion done
;
1053 * The task's runqueue lock must be held.
1054 * Returns true if you have to wait for migration thread.
1057 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1059 struct rq
*rq
= task_rq(p
);
1062 * If the task is not on a runqueue (and not running), then
1063 * it is sufficient to simply update the task's cpu field.
1065 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1066 set_task_cpu(p
, dest_cpu
);
1070 init_completion(&req
->done
);
1072 req
->dest_cpu
= dest_cpu
;
1073 list_add(&req
->list
, &rq
->migration_queue
);
1079 * wait_task_inactive - wait for a thread to unschedule.
1081 * The caller must ensure that the task *will* unschedule sometime soon,
1082 * else this function might spin for a *long* time. This function can't
1083 * be called with interrupts off, or it may introduce deadlock with
1084 * smp_call_function() if an IPI is sent by the same process we are
1085 * waiting to become inactive.
1087 void wait_task_inactive(struct task_struct
*p
)
1089 unsigned long flags
;
1095 * We do the initial early heuristics without holding
1096 * any task-queue locks at all. We'll only try to get
1097 * the runqueue lock when things look like they will
1103 * If the task is actively running on another CPU
1104 * still, just relax and busy-wait without holding
1107 * NOTE! Since we don't hold any locks, it's not
1108 * even sure that "rq" stays as the right runqueue!
1109 * But we don't care, since "task_running()" will
1110 * return false if the runqueue has changed and p
1111 * is actually now running somewhere else!
1113 while (task_running(rq
, p
))
1117 * Ok, time to look more closely! We need the rq
1118 * lock now, to be *sure*. If we're wrong, we'll
1119 * just go back and repeat.
1121 rq
= task_rq_lock(p
, &flags
);
1122 running
= task_running(rq
, p
);
1123 on_rq
= p
->se
.on_rq
;
1124 task_rq_unlock(rq
, &flags
);
1127 * Was it really running after all now that we
1128 * checked with the proper locks actually held?
1130 * Oops. Go back and try again..
1132 if (unlikely(running
)) {
1138 * It's not enough that it's not actively running,
1139 * it must be off the runqueue _entirely_, and not
1142 * So if it wa still runnable (but just not actively
1143 * running right now), it's preempted, and we should
1144 * yield - it could be a while.
1146 if (unlikely(on_rq
)) {
1152 * Ahh, all good. It wasn't running, and it wasn't
1153 * runnable, which means that it will never become
1154 * running in the future either. We're all done!
1159 * kick_process - kick a running thread to enter/exit the kernel
1160 * @p: the to-be-kicked thread
1162 * Cause a process which is running on another CPU to enter
1163 * kernel-mode, without any delay. (to get signals handled.)
1165 * NOTE: this function doesnt have to take the runqueue lock,
1166 * because all it wants to ensure is that the remote task enters
1167 * the kernel. If the IPI races and the task has been migrated
1168 * to another CPU then no harm is done and the purpose has been
1171 void kick_process(struct task_struct
*p
)
1177 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1178 smp_send_reschedule(cpu
);
1183 * Return a low guess at the load of a migration-source cpu weighted
1184 * according to the scheduling class and "nice" value.
1186 * We want to under-estimate the load of migration sources, to
1187 * balance conservatively.
1189 static inline unsigned long source_load(int cpu
, int type
)
1191 struct rq
*rq
= cpu_rq(cpu
);
1192 unsigned long total
= weighted_cpuload(cpu
);
1197 return min(rq
->cpu_load
[type
-1], total
);
1201 * Return a high guess at the load of a migration-target cpu weighted
1202 * according to the scheduling class and "nice" value.
1204 static inline unsigned long target_load(int cpu
, int type
)
1206 struct rq
*rq
= cpu_rq(cpu
);
1207 unsigned long total
= weighted_cpuload(cpu
);
1212 return max(rq
->cpu_load
[type
-1], total
);
1216 * Return the average load per task on the cpu's run queue
1218 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1220 struct rq
*rq
= cpu_rq(cpu
);
1221 unsigned long total
= weighted_cpuload(cpu
);
1222 unsigned long n
= rq
->nr_running
;
1224 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1228 * find_idlest_group finds and returns the least busy CPU group within the
1231 static struct sched_group
*
1232 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1234 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1235 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1236 int load_idx
= sd
->forkexec_idx
;
1237 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1240 unsigned long load
, avg_load
;
1244 /* Skip over this group if it has no CPUs allowed */
1245 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1248 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1250 /* Tally up the load of all CPUs in the group */
1253 for_each_cpu_mask(i
, group
->cpumask
) {
1254 /* Bias balancing toward cpus of our domain */
1256 load
= source_load(i
, load_idx
);
1258 load
= target_load(i
, load_idx
);
1263 /* Adjust by relative CPU power of the group */
1264 avg_load
= sg_div_cpu_power(group
,
1265 avg_load
* SCHED_LOAD_SCALE
);
1268 this_load
= avg_load
;
1270 } else if (avg_load
< min_load
) {
1271 min_load
= avg_load
;
1275 group
= group
->next
;
1276 } while (group
!= sd
->groups
);
1278 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1284 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1287 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1290 unsigned long load
, min_load
= ULONG_MAX
;
1294 /* Traverse only the allowed CPUs */
1295 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1297 for_each_cpu_mask(i
, tmp
) {
1298 load
= weighted_cpuload(i
);
1300 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1310 * sched_balance_self: balance the current task (running on cpu) in domains
1311 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1314 * Balance, ie. select the least loaded group.
1316 * Returns the target CPU number, or the same CPU if no balancing is needed.
1318 * preempt must be disabled.
1320 static int sched_balance_self(int cpu
, int flag
)
1322 struct task_struct
*t
= current
;
1323 struct sched_domain
*tmp
, *sd
= NULL
;
1325 for_each_domain(cpu
, tmp
) {
1327 * If power savings logic is enabled for a domain, stop there.
1329 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1331 if (tmp
->flags
& flag
)
1337 struct sched_group
*group
;
1338 int new_cpu
, weight
;
1340 if (!(sd
->flags
& flag
)) {
1346 group
= find_idlest_group(sd
, t
, cpu
);
1352 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1353 if (new_cpu
== -1 || new_cpu
== cpu
) {
1354 /* Now try balancing at a lower domain level of cpu */
1359 /* Now try balancing at a lower domain level of new_cpu */
1362 weight
= cpus_weight(span
);
1363 for_each_domain(cpu
, tmp
) {
1364 if (weight
<= cpus_weight(tmp
->span
))
1366 if (tmp
->flags
& flag
)
1369 /* while loop will break here if sd == NULL */
1375 #endif /* CONFIG_SMP */
1378 * wake_idle() will wake a task on an idle cpu if task->cpu is
1379 * not idle and an idle cpu is available. The span of cpus to
1380 * search starts with cpus closest then further out as needed,
1381 * so we always favor a closer, idle cpu.
1383 * Returns the CPU we should wake onto.
1385 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1386 static int wake_idle(int cpu
, struct task_struct
*p
)
1389 struct sched_domain
*sd
;
1393 * If it is idle, then it is the best cpu to run this task.
1395 * This cpu is also the best, if it has more than one task already.
1396 * Siblings must be also busy(in most cases) as they didn't already
1397 * pickup the extra load from this cpu and hence we need not check
1398 * sibling runqueue info. This will avoid the checks and cache miss
1399 * penalities associated with that.
1401 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1404 for_each_domain(cpu
, sd
) {
1405 if (sd
->flags
& SD_WAKE_IDLE
) {
1406 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1407 for_each_cpu_mask(i
, tmp
) {
1418 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1425 * try_to_wake_up - wake up a thread
1426 * @p: the to-be-woken-up thread
1427 * @state: the mask of task states that can be woken
1428 * @sync: do a synchronous wakeup?
1430 * Put it on the run-queue if it's not already there. The "current"
1431 * thread is always on the run-queue (except when the actual
1432 * re-schedule is in progress), and as such you're allowed to do
1433 * the simpler "current->state = TASK_RUNNING" to mark yourself
1434 * runnable without the overhead of this.
1436 * returns failure only if the task is already active.
1438 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1440 int cpu
, this_cpu
, success
= 0;
1441 unsigned long flags
;
1445 struct sched_domain
*sd
, *this_sd
= NULL
;
1446 unsigned long load
, this_load
;
1450 rq
= task_rq_lock(p
, &flags
);
1451 old_state
= p
->state
;
1452 if (!(old_state
& state
))
1459 this_cpu
= smp_processor_id();
1462 if (unlikely(task_running(rq
, p
)))
1467 schedstat_inc(rq
, ttwu_cnt
);
1468 if (cpu
== this_cpu
) {
1469 schedstat_inc(rq
, ttwu_local
);
1473 for_each_domain(this_cpu
, sd
) {
1474 if (cpu_isset(cpu
, sd
->span
)) {
1475 schedstat_inc(sd
, ttwu_wake_remote
);
1481 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1485 * Check for affine wakeup and passive balancing possibilities.
1488 int idx
= this_sd
->wake_idx
;
1489 unsigned int imbalance
;
1491 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1493 load
= source_load(cpu
, idx
);
1494 this_load
= target_load(this_cpu
, idx
);
1496 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1498 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1499 unsigned long tl
= this_load
;
1500 unsigned long tl_per_task
;
1502 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1505 * If sync wakeup then subtract the (maximum possible)
1506 * effect of the currently running task from the load
1507 * of the current CPU:
1510 tl
-= current
->se
.load
.weight
;
1513 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1514 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1516 * This domain has SD_WAKE_AFFINE and
1517 * p is cache cold in this domain, and
1518 * there is no bad imbalance.
1520 schedstat_inc(this_sd
, ttwu_move_affine
);
1526 * Start passive balancing when half the imbalance_pct
1529 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1530 if (imbalance
*this_load
<= 100*load
) {
1531 schedstat_inc(this_sd
, ttwu_move_balance
);
1537 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1539 new_cpu
= wake_idle(new_cpu
, p
);
1540 if (new_cpu
!= cpu
) {
1541 set_task_cpu(p
, new_cpu
);
1542 task_rq_unlock(rq
, &flags
);
1543 /* might preempt at this point */
1544 rq
= task_rq_lock(p
, &flags
);
1545 old_state
= p
->state
;
1546 if (!(old_state
& state
))
1551 this_cpu
= smp_processor_id();
1556 #endif /* CONFIG_SMP */
1557 update_rq_clock(rq
);
1558 activate_task(rq
, p
, 1);
1560 * Sync wakeups (i.e. those types of wakeups where the waker
1561 * has indicated that it will leave the CPU in short order)
1562 * don't trigger a preemption, if the woken up task will run on
1563 * this cpu. (in this case the 'I will reschedule' promise of
1564 * the waker guarantees that the freshly woken up task is going
1565 * to be considered on this CPU.)
1567 if (!sync
|| cpu
!= this_cpu
)
1568 check_preempt_curr(rq
, p
);
1572 p
->state
= TASK_RUNNING
;
1574 task_rq_unlock(rq
, &flags
);
1579 int fastcall
wake_up_process(struct task_struct
*p
)
1581 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1582 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1584 EXPORT_SYMBOL(wake_up_process
);
1586 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1588 return try_to_wake_up(p
, state
, 0);
1592 * Perform scheduler related setup for a newly forked process p.
1593 * p is forked by current.
1595 * __sched_fork() is basic setup used by init_idle() too:
1597 static void __sched_fork(struct task_struct
*p
)
1599 p
->se
.exec_start
= 0;
1600 p
->se
.sum_exec_runtime
= 0;
1601 p
->se
.prev_sum_exec_runtime
= 0;
1603 #ifdef CONFIG_SCHEDSTATS
1604 p
->se
.wait_start
= 0;
1605 p
->se
.sum_sleep_runtime
= 0;
1606 p
->se
.sleep_start
= 0;
1607 p
->se
.block_start
= 0;
1608 p
->se
.sleep_max
= 0;
1609 p
->se
.block_max
= 0;
1611 p
->se
.slice_max
= 0;
1615 INIT_LIST_HEAD(&p
->run_list
);
1618 #ifdef CONFIG_PREEMPT_NOTIFIERS
1619 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1623 * We mark the process as running here, but have not actually
1624 * inserted it onto the runqueue yet. This guarantees that
1625 * nobody will actually run it, and a signal or other external
1626 * event cannot wake it up and insert it on the runqueue either.
1628 p
->state
= TASK_RUNNING
;
1632 * fork()/clone()-time setup:
1634 void sched_fork(struct task_struct
*p
, int clone_flags
)
1636 int cpu
= get_cpu();
1641 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1643 __set_task_cpu(p
, cpu
);
1646 * Make sure we do not leak PI boosting priority to the child:
1648 p
->prio
= current
->normal_prio
;
1650 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1651 if (likely(sched_info_on()))
1652 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1654 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1657 #ifdef CONFIG_PREEMPT
1658 /* Want to start with kernel preemption disabled. */
1659 task_thread_info(p
)->preempt_count
= 1;
1665 * wake_up_new_task - wake up a newly created task for the first time.
1667 * This function will do some initial scheduler statistics housekeeping
1668 * that must be done for every newly created context, then puts the task
1669 * on the runqueue and wakes it.
1671 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1673 unsigned long flags
;
1677 rq
= task_rq_lock(p
, &flags
);
1678 BUG_ON(p
->state
!= TASK_RUNNING
);
1679 this_cpu
= smp_processor_id(); /* parent's CPU */
1680 update_rq_clock(rq
);
1682 p
->prio
= effective_prio(p
);
1684 if (rt_prio(p
->prio
))
1685 p
->sched_class
= &rt_sched_class
;
1687 p
->sched_class
= &fair_sched_class
;
1689 if (task_cpu(p
) != this_cpu
|| !p
->sched_class
->task_new
||
1690 !current
->se
.on_rq
) {
1691 activate_task(rq
, p
, 0);
1694 * Let the scheduling class do new task startup
1695 * management (if any):
1697 p
->sched_class
->task_new(rq
, p
);
1698 inc_nr_running(p
, rq
);
1700 check_preempt_curr(rq
, p
);
1701 task_rq_unlock(rq
, &flags
);
1704 #ifdef CONFIG_PREEMPT_NOTIFIERS
1707 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1708 * @notifier: notifier struct to register
1710 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1712 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1714 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1717 * preempt_notifier_unregister - no longer interested in preemption notifications
1718 * @notifier: notifier struct to unregister
1720 * This is safe to call from within a preemption notifier.
1722 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1724 hlist_del(¬ifier
->link
);
1726 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1728 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1730 struct preempt_notifier
*notifier
;
1731 struct hlist_node
*node
;
1733 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1734 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1738 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1739 struct task_struct
*next
)
1741 struct preempt_notifier
*notifier
;
1742 struct hlist_node
*node
;
1744 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1745 notifier
->ops
->sched_out(notifier
, next
);
1750 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1755 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1756 struct task_struct
*next
)
1763 * prepare_task_switch - prepare to switch tasks
1764 * @rq: the runqueue preparing to switch
1765 * @prev: the current task that is being switched out
1766 * @next: the task we are going to switch to.
1768 * This is called with the rq lock held and interrupts off. It must
1769 * be paired with a subsequent finish_task_switch after the context
1772 * prepare_task_switch sets up locking and calls architecture specific
1776 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1777 struct task_struct
*next
)
1779 fire_sched_out_preempt_notifiers(prev
, next
);
1780 prepare_lock_switch(rq
, next
);
1781 prepare_arch_switch(next
);
1785 * finish_task_switch - clean up after a task-switch
1786 * @rq: runqueue associated with task-switch
1787 * @prev: the thread we just switched away from.
1789 * finish_task_switch must be called after the context switch, paired
1790 * with a prepare_task_switch call before the context switch.
1791 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1792 * and do any other architecture-specific cleanup actions.
1794 * Note that we may have delayed dropping an mm in context_switch(). If
1795 * so, we finish that here outside of the runqueue lock. (Doing it
1796 * with the lock held can cause deadlocks; see schedule() for
1799 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1800 __releases(rq
->lock
)
1802 struct mm_struct
*mm
= rq
->prev_mm
;
1808 * A task struct has one reference for the use as "current".
1809 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1810 * schedule one last time. The schedule call will never return, and
1811 * the scheduled task must drop that reference.
1812 * The test for TASK_DEAD must occur while the runqueue locks are
1813 * still held, otherwise prev could be scheduled on another cpu, die
1814 * there before we look at prev->state, and then the reference would
1816 * Manfred Spraul <manfred@colorfullife.com>
1818 prev_state
= prev
->state
;
1819 finish_arch_switch(prev
);
1820 finish_lock_switch(rq
, prev
);
1821 fire_sched_in_preempt_notifiers(current
);
1824 if (unlikely(prev_state
== TASK_DEAD
)) {
1826 * Remove function-return probe instances associated with this
1827 * task and put them back on the free list.
1829 kprobe_flush_task(prev
);
1830 put_task_struct(prev
);
1835 * schedule_tail - first thing a freshly forked thread must call.
1836 * @prev: the thread we just switched away from.
1838 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1839 __releases(rq
->lock
)
1841 struct rq
*rq
= this_rq();
1843 finish_task_switch(rq
, prev
);
1844 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1845 /* In this case, finish_task_switch does not reenable preemption */
1848 if (current
->set_child_tid
)
1849 put_user(current
->pid
, current
->set_child_tid
);
1853 * context_switch - switch to the new MM and the new
1854 * thread's register state.
1857 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1858 struct task_struct
*next
)
1860 struct mm_struct
*mm
, *oldmm
;
1862 prepare_task_switch(rq
, prev
, next
);
1864 oldmm
= prev
->active_mm
;
1866 * For paravirt, this is coupled with an exit in switch_to to
1867 * combine the page table reload and the switch backend into
1870 arch_enter_lazy_cpu_mode();
1872 if (unlikely(!mm
)) {
1873 next
->active_mm
= oldmm
;
1874 atomic_inc(&oldmm
->mm_count
);
1875 enter_lazy_tlb(oldmm
, next
);
1877 switch_mm(oldmm
, mm
, next
);
1879 if (unlikely(!prev
->mm
)) {
1880 prev
->active_mm
= NULL
;
1881 rq
->prev_mm
= oldmm
;
1884 * Since the runqueue lock will be released by the next
1885 * task (which is an invalid locking op but in the case
1886 * of the scheduler it's an obvious special-case), so we
1887 * do an early lockdep release here:
1889 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1890 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1893 /* Here we just switch the register state and the stack. */
1894 switch_to(prev
, next
, prev
);
1898 * this_rq must be evaluated again because prev may have moved
1899 * CPUs since it called schedule(), thus the 'rq' on its stack
1900 * frame will be invalid.
1902 finish_task_switch(this_rq(), prev
);
1906 * nr_running, nr_uninterruptible and nr_context_switches:
1908 * externally visible scheduler statistics: current number of runnable
1909 * threads, current number of uninterruptible-sleeping threads, total
1910 * number of context switches performed since bootup.
1912 unsigned long nr_running(void)
1914 unsigned long i
, sum
= 0;
1916 for_each_online_cpu(i
)
1917 sum
+= cpu_rq(i
)->nr_running
;
1922 unsigned long nr_uninterruptible(void)
1924 unsigned long i
, sum
= 0;
1926 for_each_possible_cpu(i
)
1927 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1930 * Since we read the counters lockless, it might be slightly
1931 * inaccurate. Do not allow it to go below zero though:
1933 if (unlikely((long)sum
< 0))
1939 unsigned long long nr_context_switches(void)
1942 unsigned long long sum
= 0;
1944 for_each_possible_cpu(i
)
1945 sum
+= cpu_rq(i
)->nr_switches
;
1950 unsigned long nr_iowait(void)
1952 unsigned long i
, sum
= 0;
1954 for_each_possible_cpu(i
)
1955 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1960 unsigned long nr_active(void)
1962 unsigned long i
, running
= 0, uninterruptible
= 0;
1964 for_each_online_cpu(i
) {
1965 running
+= cpu_rq(i
)->nr_running
;
1966 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1969 if (unlikely((long)uninterruptible
< 0))
1970 uninterruptible
= 0;
1972 return running
+ uninterruptible
;
1976 * Update rq->cpu_load[] statistics. This function is usually called every
1977 * scheduler tick (TICK_NSEC).
1979 static void update_cpu_load(struct rq
*this_rq
)
1981 unsigned long this_load
= this_rq
->load
.weight
;
1984 this_rq
->nr_load_updates
++;
1986 /* Update our load: */
1987 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1988 unsigned long old_load
, new_load
;
1990 /* scale is effectively 1 << i now, and >> i divides by scale */
1992 old_load
= this_rq
->cpu_load
[i
];
1993 new_load
= this_load
;
1995 * Round up the averaging division if load is increasing. This
1996 * prevents us from getting stuck on 9 if the load is 10, for
1999 if (new_load
> old_load
)
2000 new_load
+= scale
-1;
2001 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2008 * double_rq_lock - safely lock two runqueues
2010 * Note this does not disable interrupts like task_rq_lock,
2011 * you need to do so manually before calling.
2013 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2014 __acquires(rq1
->lock
)
2015 __acquires(rq2
->lock
)
2017 BUG_ON(!irqs_disabled());
2019 spin_lock(&rq1
->lock
);
2020 __acquire(rq2
->lock
); /* Fake it out ;) */
2023 spin_lock(&rq1
->lock
);
2024 spin_lock(&rq2
->lock
);
2026 spin_lock(&rq2
->lock
);
2027 spin_lock(&rq1
->lock
);
2030 update_rq_clock(rq1
);
2031 update_rq_clock(rq2
);
2035 * double_rq_unlock - safely unlock two runqueues
2037 * Note this does not restore interrupts like task_rq_unlock,
2038 * you need to do so manually after calling.
2040 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2041 __releases(rq1
->lock
)
2042 __releases(rq2
->lock
)
2044 spin_unlock(&rq1
->lock
);
2046 spin_unlock(&rq2
->lock
);
2048 __release(rq2
->lock
);
2052 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2054 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2055 __releases(this_rq
->lock
)
2056 __acquires(busiest
->lock
)
2057 __acquires(this_rq
->lock
)
2059 if (unlikely(!irqs_disabled())) {
2060 /* printk() doesn't work good under rq->lock */
2061 spin_unlock(&this_rq
->lock
);
2064 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2065 if (busiest
< this_rq
) {
2066 spin_unlock(&this_rq
->lock
);
2067 spin_lock(&busiest
->lock
);
2068 spin_lock(&this_rq
->lock
);
2070 spin_lock(&busiest
->lock
);
2075 * If dest_cpu is allowed for this process, migrate the task to it.
2076 * This is accomplished by forcing the cpu_allowed mask to only
2077 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2078 * the cpu_allowed mask is restored.
2080 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2082 struct migration_req req
;
2083 unsigned long flags
;
2086 rq
= task_rq_lock(p
, &flags
);
2087 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2088 || unlikely(cpu_is_offline(dest_cpu
)))
2091 /* force the process onto the specified CPU */
2092 if (migrate_task(p
, dest_cpu
, &req
)) {
2093 /* Need to wait for migration thread (might exit: take ref). */
2094 struct task_struct
*mt
= rq
->migration_thread
;
2096 get_task_struct(mt
);
2097 task_rq_unlock(rq
, &flags
);
2098 wake_up_process(mt
);
2099 put_task_struct(mt
);
2100 wait_for_completion(&req
.done
);
2105 task_rq_unlock(rq
, &flags
);
2109 * sched_exec - execve() is a valuable balancing opportunity, because at
2110 * this point the task has the smallest effective memory and cache footprint.
2112 void sched_exec(void)
2114 int new_cpu
, this_cpu
= get_cpu();
2115 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2117 if (new_cpu
!= this_cpu
)
2118 sched_migrate_task(current
, new_cpu
);
2122 * pull_task - move a task from a remote runqueue to the local runqueue.
2123 * Both runqueues must be locked.
2125 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2126 struct rq
*this_rq
, int this_cpu
)
2128 deactivate_task(src_rq
, p
, 0);
2129 set_task_cpu(p
, this_cpu
);
2130 activate_task(this_rq
, p
, 0);
2132 * Note that idle threads have a prio of MAX_PRIO, for this test
2133 * to be always true for them.
2135 check_preempt_curr(this_rq
, p
);
2139 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2142 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2143 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2147 * We do not migrate tasks that are:
2148 * 1) running (obviously), or
2149 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2150 * 3) are cache-hot on their current CPU.
2152 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2156 if (task_running(rq
, p
))
2162 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2163 unsigned long max_nr_move
, unsigned long max_load_move
,
2164 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2165 int *all_pinned
, unsigned long *load_moved
,
2166 int *this_best_prio
, struct rq_iterator
*iterator
)
2168 int pulled
= 0, pinned
= 0, skip_for_load
;
2169 struct task_struct
*p
;
2170 long rem_load_move
= max_load_move
;
2172 if (max_nr_move
== 0 || max_load_move
== 0)
2178 * Start the load-balancing iterator:
2180 p
= iterator
->start(iterator
->arg
);
2185 * To help distribute high priority tasks accross CPUs we don't
2186 * skip a task if it will be the highest priority task (i.e. smallest
2187 * prio value) on its new queue regardless of its load weight
2189 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2190 SCHED_LOAD_SCALE_FUZZ
;
2191 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2192 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2193 p
= iterator
->next(iterator
->arg
);
2197 pull_task(busiest
, p
, this_rq
, this_cpu
);
2199 rem_load_move
-= p
->se
.load
.weight
;
2202 * We only want to steal up to the prescribed number of tasks
2203 * and the prescribed amount of weighted load.
2205 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2206 if (p
->prio
< *this_best_prio
)
2207 *this_best_prio
= p
->prio
;
2208 p
= iterator
->next(iterator
->arg
);
2213 * Right now, this is the only place pull_task() is called,
2214 * so we can safely collect pull_task() stats here rather than
2215 * inside pull_task().
2217 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2220 *all_pinned
= pinned
;
2221 *load_moved
= max_load_move
- rem_load_move
;
2226 * move_tasks tries to move up to max_load_move weighted load from busiest to
2227 * this_rq, as part of a balancing operation within domain "sd".
2228 * Returns 1 if successful and 0 otherwise.
2230 * Called with both runqueues locked.
2232 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2233 unsigned long max_load_move
,
2234 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2237 struct sched_class
*class = sched_class_highest
;
2238 unsigned long total_load_moved
= 0;
2239 int this_best_prio
= this_rq
->curr
->prio
;
2243 class->load_balance(this_rq
, this_cpu
, busiest
,
2244 ULONG_MAX
, max_load_move
- total_load_moved
,
2245 sd
, idle
, all_pinned
, &this_best_prio
);
2246 class = class->next
;
2247 } while (class && max_load_move
> total_load_moved
);
2249 return total_load_moved
> 0;
2253 * move_one_task tries to move exactly one task from busiest to this_rq, as
2254 * part of active balancing operations within "domain".
2255 * Returns 1 if successful and 0 otherwise.
2257 * Called with both runqueues locked.
2259 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2260 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2262 struct sched_class
*class;
2263 int this_best_prio
= MAX_PRIO
;
2265 for (class = sched_class_highest
; class; class = class->next
)
2266 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2267 1, ULONG_MAX
, sd
, idle
, NULL
,
2275 * find_busiest_group finds and returns the busiest CPU group within the
2276 * domain. It calculates and returns the amount of weighted load which
2277 * should be moved to restore balance via the imbalance parameter.
2279 static struct sched_group
*
2280 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2281 unsigned long *imbalance
, enum cpu_idle_type idle
,
2282 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2284 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2285 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2286 unsigned long max_pull
;
2287 unsigned long busiest_load_per_task
, busiest_nr_running
;
2288 unsigned long this_load_per_task
, this_nr_running
;
2290 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2291 int power_savings_balance
= 1;
2292 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2293 unsigned long min_nr_running
= ULONG_MAX
;
2294 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2297 max_load
= this_load
= total_load
= total_pwr
= 0;
2298 busiest_load_per_task
= busiest_nr_running
= 0;
2299 this_load_per_task
= this_nr_running
= 0;
2300 if (idle
== CPU_NOT_IDLE
)
2301 load_idx
= sd
->busy_idx
;
2302 else if (idle
== CPU_NEWLY_IDLE
)
2303 load_idx
= sd
->newidle_idx
;
2305 load_idx
= sd
->idle_idx
;
2308 unsigned long load
, group_capacity
;
2311 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2312 unsigned long sum_nr_running
, sum_weighted_load
;
2314 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2317 balance_cpu
= first_cpu(group
->cpumask
);
2319 /* Tally up the load of all CPUs in the group */
2320 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2322 for_each_cpu_mask(i
, group
->cpumask
) {
2325 if (!cpu_isset(i
, *cpus
))
2330 if (*sd_idle
&& rq
->nr_running
)
2333 /* Bias balancing toward cpus of our domain */
2335 if (idle_cpu(i
) && !first_idle_cpu
) {
2340 load
= target_load(i
, load_idx
);
2342 load
= source_load(i
, load_idx
);
2345 sum_nr_running
+= rq
->nr_running
;
2346 sum_weighted_load
+= weighted_cpuload(i
);
2350 * First idle cpu or the first cpu(busiest) in this sched group
2351 * is eligible for doing load balancing at this and above
2352 * domains. In the newly idle case, we will allow all the cpu's
2353 * to do the newly idle load balance.
2355 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2356 balance_cpu
!= this_cpu
&& balance
) {
2361 total_load
+= avg_load
;
2362 total_pwr
+= group
->__cpu_power
;
2364 /* Adjust by relative CPU power of the group */
2365 avg_load
= sg_div_cpu_power(group
,
2366 avg_load
* SCHED_LOAD_SCALE
);
2368 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2371 this_load
= avg_load
;
2373 this_nr_running
= sum_nr_running
;
2374 this_load_per_task
= sum_weighted_load
;
2375 } else if (avg_load
> max_load
&&
2376 sum_nr_running
> group_capacity
) {
2377 max_load
= avg_load
;
2379 busiest_nr_running
= sum_nr_running
;
2380 busiest_load_per_task
= sum_weighted_load
;
2383 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2385 * Busy processors will not participate in power savings
2388 if (idle
== CPU_NOT_IDLE
||
2389 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2393 * If the local group is idle or completely loaded
2394 * no need to do power savings balance at this domain
2396 if (local_group
&& (this_nr_running
>= group_capacity
||
2398 power_savings_balance
= 0;
2401 * If a group is already running at full capacity or idle,
2402 * don't include that group in power savings calculations
2404 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2409 * Calculate the group which has the least non-idle load.
2410 * This is the group from where we need to pick up the load
2413 if ((sum_nr_running
< min_nr_running
) ||
2414 (sum_nr_running
== min_nr_running
&&
2415 first_cpu(group
->cpumask
) <
2416 first_cpu(group_min
->cpumask
))) {
2418 min_nr_running
= sum_nr_running
;
2419 min_load_per_task
= sum_weighted_load
/
2424 * Calculate the group which is almost near its
2425 * capacity but still has some space to pick up some load
2426 * from other group and save more power
2428 if (sum_nr_running
<= group_capacity
- 1) {
2429 if (sum_nr_running
> leader_nr_running
||
2430 (sum_nr_running
== leader_nr_running
&&
2431 first_cpu(group
->cpumask
) >
2432 first_cpu(group_leader
->cpumask
))) {
2433 group_leader
= group
;
2434 leader_nr_running
= sum_nr_running
;
2439 group
= group
->next
;
2440 } while (group
!= sd
->groups
);
2442 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2445 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2447 if (this_load
>= avg_load
||
2448 100*max_load
<= sd
->imbalance_pct
*this_load
)
2451 busiest_load_per_task
/= busiest_nr_running
;
2453 * We're trying to get all the cpus to the average_load, so we don't
2454 * want to push ourselves above the average load, nor do we wish to
2455 * reduce the max loaded cpu below the average load, as either of these
2456 * actions would just result in more rebalancing later, and ping-pong
2457 * tasks around. Thus we look for the minimum possible imbalance.
2458 * Negative imbalances (*we* are more loaded than anyone else) will
2459 * be counted as no imbalance for these purposes -- we can't fix that
2460 * by pulling tasks to us. Be careful of negative numbers as they'll
2461 * appear as very large values with unsigned longs.
2463 if (max_load
<= busiest_load_per_task
)
2467 * In the presence of smp nice balancing, certain scenarios can have
2468 * max load less than avg load(as we skip the groups at or below
2469 * its cpu_power, while calculating max_load..)
2471 if (max_load
< avg_load
) {
2473 goto small_imbalance
;
2476 /* Don't want to pull so many tasks that a group would go idle */
2477 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2479 /* How much load to actually move to equalise the imbalance */
2480 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2481 (avg_load
- this_load
) * this->__cpu_power
)
2485 * if *imbalance is less than the average load per runnable task
2486 * there is no gaurantee that any tasks will be moved so we'll have
2487 * a think about bumping its value to force at least one task to be
2490 if (*imbalance
< busiest_load_per_task
) {
2491 unsigned long tmp
, pwr_now
, pwr_move
;
2495 pwr_move
= pwr_now
= 0;
2497 if (this_nr_running
) {
2498 this_load_per_task
/= this_nr_running
;
2499 if (busiest_load_per_task
> this_load_per_task
)
2502 this_load_per_task
= SCHED_LOAD_SCALE
;
2504 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2505 busiest_load_per_task
* imbn
) {
2506 *imbalance
= busiest_load_per_task
;
2511 * OK, we don't have enough imbalance to justify moving tasks,
2512 * however we may be able to increase total CPU power used by
2516 pwr_now
+= busiest
->__cpu_power
*
2517 min(busiest_load_per_task
, max_load
);
2518 pwr_now
+= this->__cpu_power
*
2519 min(this_load_per_task
, this_load
);
2520 pwr_now
/= SCHED_LOAD_SCALE
;
2522 /* Amount of load we'd subtract */
2523 tmp
= sg_div_cpu_power(busiest
,
2524 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2526 pwr_move
+= busiest
->__cpu_power
*
2527 min(busiest_load_per_task
, max_load
- tmp
);
2529 /* Amount of load we'd add */
2530 if (max_load
* busiest
->__cpu_power
<
2531 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2532 tmp
= sg_div_cpu_power(this,
2533 max_load
* busiest
->__cpu_power
);
2535 tmp
= sg_div_cpu_power(this,
2536 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2537 pwr_move
+= this->__cpu_power
*
2538 min(this_load_per_task
, this_load
+ tmp
);
2539 pwr_move
/= SCHED_LOAD_SCALE
;
2541 /* Move if we gain throughput */
2542 if (pwr_move
> pwr_now
)
2543 *imbalance
= busiest_load_per_task
;
2549 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2550 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2553 if (this == group_leader
&& group_leader
!= group_min
) {
2554 *imbalance
= min_load_per_task
;
2564 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2567 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2568 unsigned long imbalance
, cpumask_t
*cpus
)
2570 struct rq
*busiest
= NULL
, *rq
;
2571 unsigned long max_load
= 0;
2574 for_each_cpu_mask(i
, group
->cpumask
) {
2577 if (!cpu_isset(i
, *cpus
))
2581 wl
= weighted_cpuload(i
);
2583 if (rq
->nr_running
== 1 && wl
> imbalance
)
2586 if (wl
> max_load
) {
2596 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2597 * so long as it is large enough.
2599 #define MAX_PINNED_INTERVAL 512
2602 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2603 * tasks if there is an imbalance.
2605 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2606 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2609 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2610 struct sched_group
*group
;
2611 unsigned long imbalance
;
2613 cpumask_t cpus
= CPU_MASK_ALL
;
2614 unsigned long flags
;
2617 * When power savings policy is enabled for the parent domain, idle
2618 * sibling can pick up load irrespective of busy siblings. In this case,
2619 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2620 * portraying it as CPU_NOT_IDLE.
2622 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2623 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2626 schedstat_inc(sd
, lb_cnt
[idle
]);
2629 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2636 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2640 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2642 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2646 BUG_ON(busiest
== this_rq
);
2648 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2651 if (busiest
->nr_running
> 1) {
2653 * Attempt to move tasks. If find_busiest_group has found
2654 * an imbalance but busiest->nr_running <= 1, the group is
2655 * still unbalanced. ld_moved simply stays zero, so it is
2656 * correctly treated as an imbalance.
2658 local_irq_save(flags
);
2659 double_rq_lock(this_rq
, busiest
);
2660 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2661 imbalance
, sd
, idle
, &all_pinned
);
2662 double_rq_unlock(this_rq
, busiest
);
2663 local_irq_restore(flags
);
2666 * some other cpu did the load balance for us.
2668 if (ld_moved
&& this_cpu
!= smp_processor_id())
2669 resched_cpu(this_cpu
);
2671 /* All tasks on this runqueue were pinned by CPU affinity */
2672 if (unlikely(all_pinned
)) {
2673 cpu_clear(cpu_of(busiest
), cpus
);
2674 if (!cpus_empty(cpus
))
2681 schedstat_inc(sd
, lb_failed
[idle
]);
2682 sd
->nr_balance_failed
++;
2684 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2686 spin_lock_irqsave(&busiest
->lock
, flags
);
2688 /* don't kick the migration_thread, if the curr
2689 * task on busiest cpu can't be moved to this_cpu
2691 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2692 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2694 goto out_one_pinned
;
2697 if (!busiest
->active_balance
) {
2698 busiest
->active_balance
= 1;
2699 busiest
->push_cpu
= this_cpu
;
2702 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2704 wake_up_process(busiest
->migration_thread
);
2707 * We've kicked active balancing, reset the failure
2710 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2713 sd
->nr_balance_failed
= 0;
2715 if (likely(!active_balance
)) {
2716 /* We were unbalanced, so reset the balancing interval */
2717 sd
->balance_interval
= sd
->min_interval
;
2720 * If we've begun active balancing, start to back off. This
2721 * case may not be covered by the all_pinned logic if there
2722 * is only 1 task on the busy runqueue (because we don't call
2725 if (sd
->balance_interval
< sd
->max_interval
)
2726 sd
->balance_interval
*= 2;
2729 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2730 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2735 schedstat_inc(sd
, lb_balanced
[idle
]);
2737 sd
->nr_balance_failed
= 0;
2740 /* tune up the balancing interval */
2741 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2742 (sd
->balance_interval
< sd
->max_interval
))
2743 sd
->balance_interval
*= 2;
2745 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2746 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2752 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2753 * tasks if there is an imbalance.
2755 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2756 * this_rq is locked.
2759 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2761 struct sched_group
*group
;
2762 struct rq
*busiest
= NULL
;
2763 unsigned long imbalance
;
2767 cpumask_t cpus
= CPU_MASK_ALL
;
2770 * When power savings policy is enabled for the parent domain, idle
2771 * sibling can pick up load irrespective of busy siblings. In this case,
2772 * let the state of idle sibling percolate up as IDLE, instead of
2773 * portraying it as CPU_NOT_IDLE.
2775 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2776 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2779 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2781 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2782 &sd_idle
, &cpus
, NULL
);
2784 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2788 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2791 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2795 BUG_ON(busiest
== this_rq
);
2797 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2800 if (busiest
->nr_running
> 1) {
2801 /* Attempt to move tasks */
2802 double_lock_balance(this_rq
, busiest
);
2803 /* this_rq->clock is already updated */
2804 update_rq_clock(busiest
);
2805 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2806 imbalance
, sd
, CPU_NEWLY_IDLE
,
2808 spin_unlock(&busiest
->lock
);
2810 if (unlikely(all_pinned
)) {
2811 cpu_clear(cpu_of(busiest
), cpus
);
2812 if (!cpus_empty(cpus
))
2818 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2819 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2820 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2823 sd
->nr_balance_failed
= 0;
2828 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2829 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2830 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2832 sd
->nr_balance_failed
= 0;
2838 * idle_balance is called by schedule() if this_cpu is about to become
2839 * idle. Attempts to pull tasks from other CPUs.
2841 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2843 struct sched_domain
*sd
;
2844 int pulled_task
= -1;
2845 unsigned long next_balance
= jiffies
+ HZ
;
2847 for_each_domain(this_cpu
, sd
) {
2848 unsigned long interval
;
2850 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2853 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2854 /* If we've pulled tasks over stop searching: */
2855 pulled_task
= load_balance_newidle(this_cpu
,
2858 interval
= msecs_to_jiffies(sd
->balance_interval
);
2859 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2860 next_balance
= sd
->last_balance
+ interval
;
2864 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2866 * We are going idle. next_balance may be set based on
2867 * a busy processor. So reset next_balance.
2869 this_rq
->next_balance
= next_balance
;
2874 * active_load_balance is run by migration threads. It pushes running tasks
2875 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2876 * running on each physical CPU where possible, and avoids physical /
2877 * logical imbalances.
2879 * Called with busiest_rq locked.
2881 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2883 int target_cpu
= busiest_rq
->push_cpu
;
2884 struct sched_domain
*sd
;
2885 struct rq
*target_rq
;
2887 /* Is there any task to move? */
2888 if (busiest_rq
->nr_running
<= 1)
2891 target_rq
= cpu_rq(target_cpu
);
2894 * This condition is "impossible", if it occurs
2895 * we need to fix it. Originally reported by
2896 * Bjorn Helgaas on a 128-cpu setup.
2898 BUG_ON(busiest_rq
== target_rq
);
2900 /* move a task from busiest_rq to target_rq */
2901 double_lock_balance(busiest_rq
, target_rq
);
2902 update_rq_clock(busiest_rq
);
2903 update_rq_clock(target_rq
);
2905 /* Search for an sd spanning us and the target CPU. */
2906 for_each_domain(target_cpu
, sd
) {
2907 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2908 cpu_isset(busiest_cpu
, sd
->span
))
2913 schedstat_inc(sd
, alb_cnt
);
2915 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2917 schedstat_inc(sd
, alb_pushed
);
2919 schedstat_inc(sd
, alb_failed
);
2921 spin_unlock(&target_rq
->lock
);
2926 atomic_t load_balancer
;
2928 } nohz ____cacheline_aligned
= {
2929 .load_balancer
= ATOMIC_INIT(-1),
2930 .cpu_mask
= CPU_MASK_NONE
,
2934 * This routine will try to nominate the ilb (idle load balancing)
2935 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2936 * load balancing on behalf of all those cpus. If all the cpus in the system
2937 * go into this tickless mode, then there will be no ilb owner (as there is
2938 * no need for one) and all the cpus will sleep till the next wakeup event
2941 * For the ilb owner, tick is not stopped. And this tick will be used
2942 * for idle load balancing. ilb owner will still be part of
2945 * While stopping the tick, this cpu will become the ilb owner if there
2946 * is no other owner. And will be the owner till that cpu becomes busy
2947 * or if all cpus in the system stop their ticks at which point
2948 * there is no need for ilb owner.
2950 * When the ilb owner becomes busy, it nominates another owner, during the
2951 * next busy scheduler_tick()
2953 int select_nohz_load_balancer(int stop_tick
)
2955 int cpu
= smp_processor_id();
2958 cpu_set(cpu
, nohz
.cpu_mask
);
2959 cpu_rq(cpu
)->in_nohz_recently
= 1;
2962 * If we are going offline and still the leader, give up!
2964 if (cpu_is_offline(cpu
) &&
2965 atomic_read(&nohz
.load_balancer
) == cpu
) {
2966 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2971 /* time for ilb owner also to sleep */
2972 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2973 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2974 atomic_set(&nohz
.load_balancer
, -1);
2978 if (atomic_read(&nohz
.load_balancer
) == -1) {
2979 /* make me the ilb owner */
2980 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2982 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2985 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2988 cpu_clear(cpu
, nohz
.cpu_mask
);
2990 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2991 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2998 static DEFINE_SPINLOCK(balancing
);
3001 * It checks each scheduling domain to see if it is due to be balanced,
3002 * and initiates a balancing operation if so.
3004 * Balancing parameters are set up in arch_init_sched_domains.
3006 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3009 struct rq
*rq
= cpu_rq(cpu
);
3010 unsigned long interval
;
3011 struct sched_domain
*sd
;
3012 /* Earliest time when we have to do rebalance again */
3013 unsigned long next_balance
= jiffies
+ 60*HZ
;
3014 int update_next_balance
= 0;
3016 for_each_domain(cpu
, sd
) {
3017 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3020 interval
= sd
->balance_interval
;
3021 if (idle
!= CPU_IDLE
)
3022 interval
*= sd
->busy_factor
;
3024 /* scale ms to jiffies */
3025 interval
= msecs_to_jiffies(interval
);
3026 if (unlikely(!interval
))
3028 if (interval
> HZ
*NR_CPUS
/10)
3029 interval
= HZ
*NR_CPUS
/10;
3032 if (sd
->flags
& SD_SERIALIZE
) {
3033 if (!spin_trylock(&balancing
))
3037 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3038 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3040 * We've pulled tasks over so either we're no
3041 * longer idle, or one of our SMT siblings is
3044 idle
= CPU_NOT_IDLE
;
3046 sd
->last_balance
= jiffies
;
3048 if (sd
->flags
& SD_SERIALIZE
)
3049 spin_unlock(&balancing
);
3051 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3052 next_balance
= sd
->last_balance
+ interval
;
3053 update_next_balance
= 1;
3057 * Stop the load balance at this level. There is another
3058 * CPU in our sched group which is doing load balancing more
3066 * next_balance will be updated only when there is a need.
3067 * When the cpu is attached to null domain for ex, it will not be
3070 if (likely(update_next_balance
))
3071 rq
->next_balance
= next_balance
;
3075 * run_rebalance_domains is triggered when needed from the scheduler tick.
3076 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3077 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3079 static void run_rebalance_domains(struct softirq_action
*h
)
3081 int this_cpu
= smp_processor_id();
3082 struct rq
*this_rq
= cpu_rq(this_cpu
);
3083 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3084 CPU_IDLE
: CPU_NOT_IDLE
;
3086 rebalance_domains(this_cpu
, idle
);
3090 * If this cpu is the owner for idle load balancing, then do the
3091 * balancing on behalf of the other idle cpus whose ticks are
3094 if (this_rq
->idle_at_tick
&&
3095 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3096 cpumask_t cpus
= nohz
.cpu_mask
;
3100 cpu_clear(this_cpu
, cpus
);
3101 for_each_cpu_mask(balance_cpu
, cpus
) {
3103 * If this cpu gets work to do, stop the load balancing
3104 * work being done for other cpus. Next load
3105 * balancing owner will pick it up.
3110 rebalance_domains(balance_cpu
, CPU_IDLE
);
3112 rq
= cpu_rq(balance_cpu
);
3113 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3114 this_rq
->next_balance
= rq
->next_balance
;
3121 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3123 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3124 * idle load balancing owner or decide to stop the periodic load balancing,
3125 * if the whole system is idle.
3127 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3131 * If we were in the nohz mode recently and busy at the current
3132 * scheduler tick, then check if we need to nominate new idle
3135 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3136 rq
->in_nohz_recently
= 0;
3138 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3139 cpu_clear(cpu
, nohz
.cpu_mask
);
3140 atomic_set(&nohz
.load_balancer
, -1);
3143 if (atomic_read(&nohz
.load_balancer
) == -1) {
3145 * simple selection for now: Nominate the
3146 * first cpu in the nohz list to be the next
3149 * TBD: Traverse the sched domains and nominate
3150 * the nearest cpu in the nohz.cpu_mask.
3152 int ilb
= first_cpu(nohz
.cpu_mask
);
3160 * If this cpu is idle and doing idle load balancing for all the
3161 * cpus with ticks stopped, is it time for that to stop?
3163 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3164 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3170 * If this cpu is idle and the idle load balancing is done by
3171 * someone else, then no need raise the SCHED_SOFTIRQ
3173 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3174 cpu_isset(cpu
, nohz
.cpu_mask
))
3177 if (time_after_eq(jiffies
, rq
->next_balance
))
3178 raise_softirq(SCHED_SOFTIRQ
);
3181 #else /* CONFIG_SMP */
3184 * on UP we do not need to balance between CPUs:
3186 static inline void idle_balance(int cpu
, struct rq
*rq
)
3190 /* Avoid "used but not defined" warning on UP */
3191 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3192 unsigned long max_nr_move
, unsigned long max_load_move
,
3193 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3194 int *all_pinned
, unsigned long *load_moved
,
3195 int *this_best_prio
, struct rq_iterator
*iterator
)
3204 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3206 EXPORT_PER_CPU_SYMBOL(kstat
);
3209 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3210 * that have not yet been banked in case the task is currently running.
3212 unsigned long long task_sched_runtime(struct task_struct
*p
)
3214 unsigned long flags
;
3218 rq
= task_rq_lock(p
, &flags
);
3219 ns
= p
->se
.sum_exec_runtime
;
3220 if (rq
->curr
== p
) {
3221 update_rq_clock(rq
);
3222 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3223 if ((s64
)delta_exec
> 0)
3226 task_rq_unlock(rq
, &flags
);
3232 * Account user cpu time to a process.
3233 * @p: the process that the cpu time gets accounted to
3234 * @hardirq_offset: the offset to subtract from hardirq_count()
3235 * @cputime: the cpu time spent in user space since the last update
3237 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3239 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3242 p
->utime
= cputime_add(p
->utime
, cputime
);
3244 /* Add user time to cpustat. */
3245 tmp
= cputime_to_cputime64(cputime
);
3246 if (TASK_NICE(p
) > 0)
3247 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3249 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3253 * Account system cpu time to a process.
3254 * @p: the process that the cpu time gets accounted to
3255 * @hardirq_offset: the offset to subtract from hardirq_count()
3256 * @cputime: the cpu time spent in kernel space since the last update
3258 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3261 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3262 struct rq
*rq
= this_rq();
3265 p
->stime
= cputime_add(p
->stime
, cputime
);
3267 /* Add system time to cpustat. */
3268 tmp
= cputime_to_cputime64(cputime
);
3269 if (hardirq_count() - hardirq_offset
)
3270 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3271 else if (softirq_count())
3272 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3273 else if (p
!= rq
->idle
)
3274 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3275 else if (atomic_read(&rq
->nr_iowait
) > 0)
3276 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3278 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3279 /* Account for system time used */
3280 acct_update_integrals(p
);
3284 * Account for involuntary wait time.
3285 * @p: the process from which the cpu time has been stolen
3286 * @steal: the cpu time spent in involuntary wait
3288 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3290 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3291 cputime64_t tmp
= cputime_to_cputime64(steal
);
3292 struct rq
*rq
= this_rq();
3294 if (p
== rq
->idle
) {
3295 p
->stime
= cputime_add(p
->stime
, steal
);
3296 if (atomic_read(&rq
->nr_iowait
) > 0)
3297 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3299 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3301 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3305 * This function gets called by the timer code, with HZ frequency.
3306 * We call it with interrupts disabled.
3308 * It also gets called by the fork code, when changing the parent's
3311 void scheduler_tick(void)
3313 int cpu
= smp_processor_id();
3314 struct rq
*rq
= cpu_rq(cpu
);
3315 struct task_struct
*curr
= rq
->curr
;
3316 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3318 spin_lock(&rq
->lock
);
3319 __update_rq_clock(rq
);
3321 * Let rq->clock advance by at least TICK_NSEC:
3323 if (unlikely(rq
->clock
< next_tick
))
3324 rq
->clock
= next_tick
;
3325 rq
->tick_timestamp
= rq
->clock
;
3326 update_cpu_load(rq
);
3327 if (curr
!= rq
->idle
) /* FIXME: needed? */
3328 curr
->sched_class
->task_tick(rq
, curr
);
3329 spin_unlock(&rq
->lock
);
3332 rq
->idle_at_tick
= idle_cpu(cpu
);
3333 trigger_load_balance(rq
, cpu
);
3337 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3339 void fastcall
add_preempt_count(int val
)
3344 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3346 preempt_count() += val
;
3348 * Spinlock count overflowing soon?
3350 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3353 EXPORT_SYMBOL(add_preempt_count
);
3355 void fastcall
sub_preempt_count(int val
)
3360 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3363 * Is the spinlock portion underflowing?
3365 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3366 !(preempt_count() & PREEMPT_MASK
)))
3369 preempt_count() -= val
;
3371 EXPORT_SYMBOL(sub_preempt_count
);
3376 * Print scheduling while atomic bug:
3378 static noinline
void __schedule_bug(struct task_struct
*prev
)
3380 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3381 prev
->comm
, preempt_count(), prev
->pid
);
3382 debug_show_held_locks(prev
);
3383 if (irqs_disabled())
3384 print_irqtrace_events(prev
);
3389 * Various schedule()-time debugging checks and statistics:
3391 static inline void schedule_debug(struct task_struct
*prev
)
3394 * Test if we are atomic. Since do_exit() needs to call into
3395 * schedule() atomically, we ignore that path for now.
3396 * Otherwise, whine if we are scheduling when we should not be.
3398 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3399 __schedule_bug(prev
);
3401 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3403 schedstat_inc(this_rq(), sched_cnt
);
3407 * Pick up the highest-prio task:
3409 static inline struct task_struct
*
3410 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3412 struct sched_class
*class;
3413 struct task_struct
*p
;
3416 * Optimization: we know that if all tasks are in
3417 * the fair class we can call that function directly:
3419 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3420 p
= fair_sched_class
.pick_next_task(rq
);
3425 class = sched_class_highest
;
3427 p
= class->pick_next_task(rq
);
3431 * Will never be NULL as the idle class always
3432 * returns a non-NULL p:
3434 class = class->next
;
3439 * schedule() is the main scheduler function.
3441 asmlinkage
void __sched
schedule(void)
3443 struct task_struct
*prev
, *next
;
3450 cpu
= smp_processor_id();
3454 switch_count
= &prev
->nivcsw
;
3456 release_kernel_lock(prev
);
3457 need_resched_nonpreemptible
:
3459 schedule_debug(prev
);
3461 spin_lock_irq(&rq
->lock
);
3462 clear_tsk_need_resched(prev
);
3463 __update_rq_clock(rq
);
3465 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3466 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3467 unlikely(signal_pending(prev
)))) {
3468 prev
->state
= TASK_RUNNING
;
3470 deactivate_task(rq
, prev
, 1);
3472 switch_count
= &prev
->nvcsw
;
3475 if (unlikely(!rq
->nr_running
))
3476 idle_balance(cpu
, rq
);
3478 prev
->sched_class
->put_prev_task(rq
, prev
);
3479 next
= pick_next_task(rq
, prev
);
3481 sched_info_switch(prev
, next
);
3483 if (likely(prev
!= next
)) {
3488 context_switch(rq
, prev
, next
); /* unlocks the rq */
3490 spin_unlock_irq(&rq
->lock
);
3492 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3493 cpu
= smp_processor_id();
3495 goto need_resched_nonpreemptible
;
3497 preempt_enable_no_resched();
3498 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3501 EXPORT_SYMBOL(schedule
);
3503 #ifdef CONFIG_PREEMPT
3505 * this is the entry point to schedule() from in-kernel preemption
3506 * off of preempt_enable. Kernel preemptions off return from interrupt
3507 * occur there and call schedule directly.
3509 asmlinkage
void __sched
preempt_schedule(void)
3511 struct thread_info
*ti
= current_thread_info();
3512 #ifdef CONFIG_PREEMPT_BKL
3513 struct task_struct
*task
= current
;
3514 int saved_lock_depth
;
3517 * If there is a non-zero preempt_count or interrupts are disabled,
3518 * we do not want to preempt the current task. Just return..
3520 if (likely(ti
->preempt_count
|| irqs_disabled()))
3524 add_preempt_count(PREEMPT_ACTIVE
);
3526 * We keep the big kernel semaphore locked, but we
3527 * clear ->lock_depth so that schedule() doesnt
3528 * auto-release the semaphore:
3530 #ifdef CONFIG_PREEMPT_BKL
3531 saved_lock_depth
= task
->lock_depth
;
3532 task
->lock_depth
= -1;
3535 #ifdef CONFIG_PREEMPT_BKL
3536 task
->lock_depth
= saved_lock_depth
;
3538 sub_preempt_count(PREEMPT_ACTIVE
);
3540 /* we could miss a preemption opportunity between schedule and now */
3542 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3545 EXPORT_SYMBOL(preempt_schedule
);
3548 * this is the entry point to schedule() from kernel preemption
3549 * off of irq context.
3550 * Note, that this is called and return with irqs disabled. This will
3551 * protect us against recursive calling from irq.
3553 asmlinkage
void __sched
preempt_schedule_irq(void)
3555 struct thread_info
*ti
= current_thread_info();
3556 #ifdef CONFIG_PREEMPT_BKL
3557 struct task_struct
*task
= current
;
3558 int saved_lock_depth
;
3560 /* Catch callers which need to be fixed */
3561 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3564 add_preempt_count(PREEMPT_ACTIVE
);
3566 * We keep the big kernel semaphore locked, but we
3567 * clear ->lock_depth so that schedule() doesnt
3568 * auto-release the semaphore:
3570 #ifdef CONFIG_PREEMPT_BKL
3571 saved_lock_depth
= task
->lock_depth
;
3572 task
->lock_depth
= -1;
3576 local_irq_disable();
3577 #ifdef CONFIG_PREEMPT_BKL
3578 task
->lock_depth
= saved_lock_depth
;
3580 sub_preempt_count(PREEMPT_ACTIVE
);
3582 /* we could miss a preemption opportunity between schedule and now */
3584 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3588 #endif /* CONFIG_PREEMPT */
3590 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3593 return try_to_wake_up(curr
->private, mode
, sync
);
3595 EXPORT_SYMBOL(default_wake_function
);
3598 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3599 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3600 * number) then we wake all the non-exclusive tasks and one exclusive task.
3602 * There are circumstances in which we can try to wake a task which has already
3603 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3604 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3606 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3607 int nr_exclusive
, int sync
, void *key
)
3609 wait_queue_t
*curr
, *next
;
3611 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3612 unsigned flags
= curr
->flags
;
3614 if (curr
->func(curr
, mode
, sync
, key
) &&
3615 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3621 * __wake_up - wake up threads blocked on a waitqueue.
3623 * @mode: which threads
3624 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3625 * @key: is directly passed to the wakeup function
3627 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3628 int nr_exclusive
, void *key
)
3630 unsigned long flags
;
3632 spin_lock_irqsave(&q
->lock
, flags
);
3633 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3634 spin_unlock_irqrestore(&q
->lock
, flags
);
3636 EXPORT_SYMBOL(__wake_up
);
3639 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3641 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3643 __wake_up_common(q
, mode
, 1, 0, NULL
);
3647 * __wake_up_sync - wake up threads blocked on a waitqueue.
3649 * @mode: which threads
3650 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3652 * The sync wakeup differs that the waker knows that it will schedule
3653 * away soon, so while the target thread will be woken up, it will not
3654 * be migrated to another CPU - ie. the two threads are 'synchronized'
3655 * with each other. This can prevent needless bouncing between CPUs.
3657 * On UP it can prevent extra preemption.
3660 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3662 unsigned long flags
;
3668 if (unlikely(!nr_exclusive
))
3671 spin_lock_irqsave(&q
->lock
, flags
);
3672 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3673 spin_unlock_irqrestore(&q
->lock
, flags
);
3675 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3677 void fastcall
complete(struct completion
*x
)
3679 unsigned long flags
;
3681 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3683 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3685 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3687 EXPORT_SYMBOL(complete
);
3689 void fastcall
complete_all(struct completion
*x
)
3691 unsigned long flags
;
3693 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3694 x
->done
+= UINT_MAX
/2;
3695 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3697 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3699 EXPORT_SYMBOL(complete_all
);
3701 void fastcall __sched
wait_for_completion(struct completion
*x
)
3705 spin_lock_irq(&x
->wait
.lock
);
3707 DECLARE_WAITQUEUE(wait
, current
);
3709 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3710 __add_wait_queue_tail(&x
->wait
, &wait
);
3712 __set_current_state(TASK_UNINTERRUPTIBLE
);
3713 spin_unlock_irq(&x
->wait
.lock
);
3715 spin_lock_irq(&x
->wait
.lock
);
3717 __remove_wait_queue(&x
->wait
, &wait
);
3720 spin_unlock_irq(&x
->wait
.lock
);
3722 EXPORT_SYMBOL(wait_for_completion
);
3724 unsigned long fastcall __sched
3725 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3729 spin_lock_irq(&x
->wait
.lock
);
3731 DECLARE_WAITQUEUE(wait
, current
);
3733 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3734 __add_wait_queue_tail(&x
->wait
, &wait
);
3736 __set_current_state(TASK_UNINTERRUPTIBLE
);
3737 spin_unlock_irq(&x
->wait
.lock
);
3738 timeout
= schedule_timeout(timeout
);
3739 spin_lock_irq(&x
->wait
.lock
);
3741 __remove_wait_queue(&x
->wait
, &wait
);
3745 __remove_wait_queue(&x
->wait
, &wait
);
3749 spin_unlock_irq(&x
->wait
.lock
);
3752 EXPORT_SYMBOL(wait_for_completion_timeout
);
3754 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3760 spin_lock_irq(&x
->wait
.lock
);
3762 DECLARE_WAITQUEUE(wait
, current
);
3764 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3765 __add_wait_queue_tail(&x
->wait
, &wait
);
3767 if (signal_pending(current
)) {
3769 __remove_wait_queue(&x
->wait
, &wait
);
3772 __set_current_state(TASK_INTERRUPTIBLE
);
3773 spin_unlock_irq(&x
->wait
.lock
);
3775 spin_lock_irq(&x
->wait
.lock
);
3777 __remove_wait_queue(&x
->wait
, &wait
);
3781 spin_unlock_irq(&x
->wait
.lock
);
3785 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3787 unsigned long fastcall __sched
3788 wait_for_completion_interruptible_timeout(struct completion
*x
,
3789 unsigned long timeout
)
3793 spin_lock_irq(&x
->wait
.lock
);
3795 DECLARE_WAITQUEUE(wait
, current
);
3797 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3798 __add_wait_queue_tail(&x
->wait
, &wait
);
3800 if (signal_pending(current
)) {
3801 timeout
= -ERESTARTSYS
;
3802 __remove_wait_queue(&x
->wait
, &wait
);
3805 __set_current_state(TASK_INTERRUPTIBLE
);
3806 spin_unlock_irq(&x
->wait
.lock
);
3807 timeout
= schedule_timeout(timeout
);
3808 spin_lock_irq(&x
->wait
.lock
);
3810 __remove_wait_queue(&x
->wait
, &wait
);
3814 __remove_wait_queue(&x
->wait
, &wait
);
3818 spin_unlock_irq(&x
->wait
.lock
);
3821 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3824 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3826 spin_lock_irqsave(&q
->lock
, *flags
);
3827 __add_wait_queue(q
, wait
);
3828 spin_unlock(&q
->lock
);
3832 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3834 spin_lock_irq(&q
->lock
);
3835 __remove_wait_queue(q
, wait
);
3836 spin_unlock_irqrestore(&q
->lock
, *flags
);
3839 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3841 unsigned long flags
;
3844 init_waitqueue_entry(&wait
, current
);
3846 current
->state
= TASK_INTERRUPTIBLE
;
3848 sleep_on_head(q
, &wait
, &flags
);
3850 sleep_on_tail(q
, &wait
, &flags
);
3852 EXPORT_SYMBOL(interruptible_sleep_on
);
3855 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3857 unsigned long flags
;
3860 init_waitqueue_entry(&wait
, current
);
3862 current
->state
= TASK_INTERRUPTIBLE
;
3864 sleep_on_head(q
, &wait
, &flags
);
3865 timeout
= schedule_timeout(timeout
);
3866 sleep_on_tail(q
, &wait
, &flags
);
3870 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3872 void __sched
sleep_on(wait_queue_head_t
*q
)
3874 unsigned long flags
;
3877 init_waitqueue_entry(&wait
, current
);
3879 current
->state
= TASK_UNINTERRUPTIBLE
;
3881 sleep_on_head(q
, &wait
, &flags
);
3883 sleep_on_tail(q
, &wait
, &flags
);
3885 EXPORT_SYMBOL(sleep_on
);
3887 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3889 unsigned long flags
;
3892 init_waitqueue_entry(&wait
, current
);
3894 current
->state
= TASK_UNINTERRUPTIBLE
;
3896 sleep_on_head(q
, &wait
, &flags
);
3897 timeout
= schedule_timeout(timeout
);
3898 sleep_on_tail(q
, &wait
, &flags
);
3902 EXPORT_SYMBOL(sleep_on_timeout
);
3904 #ifdef CONFIG_RT_MUTEXES
3907 * rt_mutex_setprio - set the current priority of a task
3909 * @prio: prio value (kernel-internal form)
3911 * This function changes the 'effective' priority of a task. It does
3912 * not touch ->normal_prio like __setscheduler().
3914 * Used by the rt_mutex code to implement priority inheritance logic.
3916 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3918 unsigned long flags
;
3922 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3924 rq
= task_rq_lock(p
, &flags
);
3925 update_rq_clock(rq
);
3928 on_rq
= p
->se
.on_rq
;
3930 dequeue_task(rq
, p
, 0);
3933 p
->sched_class
= &rt_sched_class
;
3935 p
->sched_class
= &fair_sched_class
;
3940 enqueue_task(rq
, p
, 0);
3942 * Reschedule if we are currently running on this runqueue and
3943 * our priority decreased, or if we are not currently running on
3944 * this runqueue and our priority is higher than the current's
3946 if (task_running(rq
, p
)) {
3947 if (p
->prio
> oldprio
)
3948 resched_task(rq
->curr
);
3950 check_preempt_curr(rq
, p
);
3953 task_rq_unlock(rq
, &flags
);
3958 void set_user_nice(struct task_struct
*p
, long nice
)
3960 int old_prio
, delta
, on_rq
;
3961 unsigned long flags
;
3964 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3967 * We have to be careful, if called from sys_setpriority(),
3968 * the task might be in the middle of scheduling on another CPU.
3970 rq
= task_rq_lock(p
, &flags
);
3971 update_rq_clock(rq
);
3973 * The RT priorities are set via sched_setscheduler(), but we still
3974 * allow the 'normal' nice value to be set - but as expected
3975 * it wont have any effect on scheduling until the task is
3976 * SCHED_FIFO/SCHED_RR:
3978 if (task_has_rt_policy(p
)) {
3979 p
->static_prio
= NICE_TO_PRIO(nice
);
3982 on_rq
= p
->se
.on_rq
;
3984 dequeue_task(rq
, p
, 0);
3988 p
->static_prio
= NICE_TO_PRIO(nice
);
3991 p
->prio
= effective_prio(p
);
3992 delta
= p
->prio
- old_prio
;
3995 enqueue_task(rq
, p
, 0);
3998 * If the task increased its priority or is running and
3999 * lowered its priority, then reschedule its CPU:
4001 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4002 resched_task(rq
->curr
);
4005 task_rq_unlock(rq
, &flags
);
4007 EXPORT_SYMBOL(set_user_nice
);
4010 * can_nice - check if a task can reduce its nice value
4014 int can_nice(const struct task_struct
*p
, const int nice
)
4016 /* convert nice value [19,-20] to rlimit style value [1,40] */
4017 int nice_rlim
= 20 - nice
;
4019 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4020 capable(CAP_SYS_NICE
));
4023 #ifdef __ARCH_WANT_SYS_NICE
4026 * sys_nice - change the priority of the current process.
4027 * @increment: priority increment
4029 * sys_setpriority is a more generic, but much slower function that
4030 * does similar things.
4032 asmlinkage
long sys_nice(int increment
)
4037 * Setpriority might change our priority at the same moment.
4038 * We don't have to worry. Conceptually one call occurs first
4039 * and we have a single winner.
4041 if (increment
< -40)
4046 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4052 if (increment
< 0 && !can_nice(current
, nice
))
4055 retval
= security_task_setnice(current
, nice
);
4059 set_user_nice(current
, nice
);
4066 * task_prio - return the priority value of a given task.
4067 * @p: the task in question.
4069 * This is the priority value as seen by users in /proc.
4070 * RT tasks are offset by -200. Normal tasks are centered
4071 * around 0, value goes from -16 to +15.
4073 int task_prio(const struct task_struct
*p
)
4075 return p
->prio
- MAX_RT_PRIO
;
4079 * task_nice - return the nice value of a given task.
4080 * @p: the task in question.
4082 int task_nice(const struct task_struct
*p
)
4084 return TASK_NICE(p
);
4086 EXPORT_SYMBOL_GPL(task_nice
);
4089 * idle_cpu - is a given cpu idle currently?
4090 * @cpu: the processor in question.
4092 int idle_cpu(int cpu
)
4094 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4098 * idle_task - return the idle task for a given cpu.
4099 * @cpu: the processor in question.
4101 struct task_struct
*idle_task(int cpu
)
4103 return cpu_rq(cpu
)->idle
;
4107 * find_process_by_pid - find a process with a matching PID value.
4108 * @pid: the pid in question.
4110 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4112 return pid
? find_task_by_pid(pid
) : current
;
4115 /* Actually do priority change: must hold rq lock. */
4117 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4119 BUG_ON(p
->se
.on_rq
);
4122 switch (p
->policy
) {
4126 p
->sched_class
= &fair_sched_class
;
4130 p
->sched_class
= &rt_sched_class
;
4134 p
->rt_priority
= prio
;
4135 p
->normal_prio
= normal_prio(p
);
4136 /* we are holding p->pi_lock already */
4137 p
->prio
= rt_mutex_getprio(p
);
4142 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4143 * @p: the task in question.
4144 * @policy: new policy.
4145 * @param: structure containing the new RT priority.
4147 * NOTE that the task may be already dead.
4149 int sched_setscheduler(struct task_struct
*p
, int policy
,
4150 struct sched_param
*param
)
4152 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4153 unsigned long flags
;
4156 /* may grab non-irq protected spin_locks */
4157 BUG_ON(in_interrupt());
4159 /* double check policy once rq lock held */
4161 policy
= oldpolicy
= p
->policy
;
4162 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4163 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4164 policy
!= SCHED_IDLE
)
4167 * Valid priorities for SCHED_FIFO and SCHED_RR are
4168 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4169 * SCHED_BATCH and SCHED_IDLE is 0.
4171 if (param
->sched_priority
< 0 ||
4172 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4173 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4175 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4179 * Allow unprivileged RT tasks to decrease priority:
4181 if (!capable(CAP_SYS_NICE
)) {
4182 if (rt_policy(policy
)) {
4183 unsigned long rlim_rtprio
;
4185 if (!lock_task_sighand(p
, &flags
))
4187 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4188 unlock_task_sighand(p
, &flags
);
4190 /* can't set/change the rt policy */
4191 if (policy
!= p
->policy
&& !rlim_rtprio
)
4194 /* can't increase priority */
4195 if (param
->sched_priority
> p
->rt_priority
&&
4196 param
->sched_priority
> rlim_rtprio
)
4200 * Like positive nice levels, dont allow tasks to
4201 * move out of SCHED_IDLE either:
4203 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4206 /* can't change other user's priorities */
4207 if ((current
->euid
!= p
->euid
) &&
4208 (current
->euid
!= p
->uid
))
4212 retval
= security_task_setscheduler(p
, policy
, param
);
4216 * make sure no PI-waiters arrive (or leave) while we are
4217 * changing the priority of the task:
4219 spin_lock_irqsave(&p
->pi_lock
, flags
);
4221 * To be able to change p->policy safely, the apropriate
4222 * runqueue lock must be held.
4224 rq
= __task_rq_lock(p
);
4225 /* recheck policy now with rq lock held */
4226 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4227 policy
= oldpolicy
= -1;
4228 __task_rq_unlock(rq
);
4229 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4232 update_rq_clock(rq
);
4233 on_rq
= p
->se
.on_rq
;
4235 deactivate_task(rq
, p
, 0);
4238 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4241 activate_task(rq
, p
, 0);
4243 * Reschedule if we are currently running on this runqueue and
4244 * our priority decreased, or if we are not currently running on
4245 * this runqueue and our priority is higher than the current's
4247 if (task_running(rq
, p
)) {
4248 if (p
->prio
> oldprio
)
4249 resched_task(rq
->curr
);
4251 check_preempt_curr(rq
, p
);
4254 __task_rq_unlock(rq
);
4255 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4257 rt_mutex_adjust_pi(p
);
4261 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4264 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4266 struct sched_param lparam
;
4267 struct task_struct
*p
;
4270 if (!param
|| pid
< 0)
4272 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4277 p
= find_process_by_pid(pid
);
4279 retval
= sched_setscheduler(p
, policy
, &lparam
);
4286 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4287 * @pid: the pid in question.
4288 * @policy: new policy.
4289 * @param: structure containing the new RT priority.
4291 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4292 struct sched_param __user
*param
)
4294 /* negative values for policy are not valid */
4298 return do_sched_setscheduler(pid
, policy
, param
);
4302 * sys_sched_setparam - set/change the RT priority of a thread
4303 * @pid: the pid in question.
4304 * @param: structure containing the new RT priority.
4306 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4308 return do_sched_setscheduler(pid
, -1, param
);
4312 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4313 * @pid: the pid in question.
4315 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4317 struct task_struct
*p
;
4318 int retval
= -EINVAL
;
4324 read_lock(&tasklist_lock
);
4325 p
= find_process_by_pid(pid
);
4327 retval
= security_task_getscheduler(p
);
4331 read_unlock(&tasklist_lock
);
4338 * sys_sched_getscheduler - get the RT priority of a thread
4339 * @pid: the pid in question.
4340 * @param: structure containing the RT priority.
4342 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4344 struct sched_param lp
;
4345 struct task_struct
*p
;
4346 int retval
= -EINVAL
;
4348 if (!param
|| pid
< 0)
4351 read_lock(&tasklist_lock
);
4352 p
= find_process_by_pid(pid
);
4357 retval
= security_task_getscheduler(p
);
4361 lp
.sched_priority
= p
->rt_priority
;
4362 read_unlock(&tasklist_lock
);
4365 * This one might sleep, we cannot do it with a spinlock held ...
4367 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4373 read_unlock(&tasklist_lock
);
4377 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4379 cpumask_t cpus_allowed
;
4380 struct task_struct
*p
;
4383 mutex_lock(&sched_hotcpu_mutex
);
4384 read_lock(&tasklist_lock
);
4386 p
= find_process_by_pid(pid
);
4388 read_unlock(&tasklist_lock
);
4389 mutex_unlock(&sched_hotcpu_mutex
);
4394 * It is not safe to call set_cpus_allowed with the
4395 * tasklist_lock held. We will bump the task_struct's
4396 * usage count and then drop tasklist_lock.
4399 read_unlock(&tasklist_lock
);
4402 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4403 !capable(CAP_SYS_NICE
))
4406 retval
= security_task_setscheduler(p
, 0, NULL
);
4410 cpus_allowed
= cpuset_cpus_allowed(p
);
4411 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4412 retval
= set_cpus_allowed(p
, new_mask
);
4416 mutex_unlock(&sched_hotcpu_mutex
);
4420 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4421 cpumask_t
*new_mask
)
4423 if (len
< sizeof(cpumask_t
)) {
4424 memset(new_mask
, 0, sizeof(cpumask_t
));
4425 } else if (len
> sizeof(cpumask_t
)) {
4426 len
= sizeof(cpumask_t
);
4428 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4432 * sys_sched_setaffinity - set the cpu affinity of a process
4433 * @pid: pid of the process
4434 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4435 * @user_mask_ptr: user-space pointer to the new cpu mask
4437 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4438 unsigned long __user
*user_mask_ptr
)
4443 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4447 return sched_setaffinity(pid
, new_mask
);
4451 * Represents all cpu's present in the system
4452 * In systems capable of hotplug, this map could dynamically grow
4453 * as new cpu's are detected in the system via any platform specific
4454 * method, such as ACPI for e.g.
4457 cpumask_t cpu_present_map __read_mostly
;
4458 EXPORT_SYMBOL(cpu_present_map
);
4461 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4462 EXPORT_SYMBOL(cpu_online_map
);
4464 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4465 EXPORT_SYMBOL(cpu_possible_map
);
4468 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4470 struct task_struct
*p
;
4473 mutex_lock(&sched_hotcpu_mutex
);
4474 read_lock(&tasklist_lock
);
4477 p
= find_process_by_pid(pid
);
4481 retval
= security_task_getscheduler(p
);
4485 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4488 read_unlock(&tasklist_lock
);
4489 mutex_unlock(&sched_hotcpu_mutex
);
4495 * sys_sched_getaffinity - get the cpu affinity of a process
4496 * @pid: pid of the process
4497 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4498 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4500 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4501 unsigned long __user
*user_mask_ptr
)
4506 if (len
< sizeof(cpumask_t
))
4509 ret
= sched_getaffinity(pid
, &mask
);
4513 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4516 return sizeof(cpumask_t
);
4520 * sys_sched_yield - yield the current processor to other threads.
4522 * This function yields the current CPU to other tasks. If there are no
4523 * other threads running on this CPU then this function will return.
4525 asmlinkage
long sys_sched_yield(void)
4527 struct rq
*rq
= this_rq_lock();
4529 schedstat_inc(rq
, yld_cnt
);
4530 current
->sched_class
->yield_task(rq
);
4533 * Since we are going to call schedule() anyway, there's
4534 * no need to preempt or enable interrupts:
4536 __release(rq
->lock
);
4537 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4538 _raw_spin_unlock(&rq
->lock
);
4539 preempt_enable_no_resched();
4546 static void __cond_resched(void)
4548 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4549 __might_sleep(__FILE__
, __LINE__
);
4552 * The BKS might be reacquired before we have dropped
4553 * PREEMPT_ACTIVE, which could trigger a second
4554 * cond_resched() call.
4557 add_preempt_count(PREEMPT_ACTIVE
);
4559 sub_preempt_count(PREEMPT_ACTIVE
);
4560 } while (need_resched());
4563 int __sched
cond_resched(void)
4565 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4566 system_state
== SYSTEM_RUNNING
) {
4572 EXPORT_SYMBOL(cond_resched
);
4575 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4576 * call schedule, and on return reacquire the lock.
4578 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4579 * operations here to prevent schedule() from being called twice (once via
4580 * spin_unlock(), once by hand).
4582 int cond_resched_lock(spinlock_t
*lock
)
4586 if (need_lockbreak(lock
)) {
4592 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4593 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4594 _raw_spin_unlock(lock
);
4595 preempt_enable_no_resched();
4602 EXPORT_SYMBOL(cond_resched_lock
);
4604 int __sched
cond_resched_softirq(void)
4606 BUG_ON(!in_softirq());
4608 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4616 EXPORT_SYMBOL(cond_resched_softirq
);
4619 * yield - yield the current processor to other threads.
4621 * This is a shortcut for kernel-space yielding - it marks the
4622 * thread runnable and calls sys_sched_yield().
4624 void __sched
yield(void)
4626 set_current_state(TASK_RUNNING
);
4629 EXPORT_SYMBOL(yield
);
4632 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4633 * that process accounting knows that this is a task in IO wait state.
4635 * But don't do that if it is a deliberate, throttling IO wait (this task
4636 * has set its backing_dev_info: the queue against which it should throttle)
4638 void __sched
io_schedule(void)
4640 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4642 delayacct_blkio_start();
4643 atomic_inc(&rq
->nr_iowait
);
4645 atomic_dec(&rq
->nr_iowait
);
4646 delayacct_blkio_end();
4648 EXPORT_SYMBOL(io_schedule
);
4650 long __sched
io_schedule_timeout(long timeout
)
4652 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4655 delayacct_blkio_start();
4656 atomic_inc(&rq
->nr_iowait
);
4657 ret
= schedule_timeout(timeout
);
4658 atomic_dec(&rq
->nr_iowait
);
4659 delayacct_blkio_end();
4664 * sys_sched_get_priority_max - return maximum RT priority.
4665 * @policy: scheduling class.
4667 * this syscall returns the maximum rt_priority that can be used
4668 * by a given scheduling class.
4670 asmlinkage
long sys_sched_get_priority_max(int policy
)
4677 ret
= MAX_USER_RT_PRIO
-1;
4689 * sys_sched_get_priority_min - return minimum RT priority.
4690 * @policy: scheduling class.
4692 * this syscall returns the minimum rt_priority that can be used
4693 * by a given scheduling class.
4695 asmlinkage
long sys_sched_get_priority_min(int policy
)
4713 * sys_sched_rr_get_interval - return the default timeslice of a process.
4714 * @pid: pid of the process.
4715 * @interval: userspace pointer to the timeslice value.
4717 * this syscall writes the default timeslice value of a given process
4718 * into the user-space timespec buffer. A value of '0' means infinity.
4721 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4723 struct task_struct
*p
;
4724 int retval
= -EINVAL
;
4731 read_lock(&tasklist_lock
);
4732 p
= find_process_by_pid(pid
);
4736 retval
= security_task_getscheduler(p
);
4740 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4741 0 : static_prio_timeslice(p
->static_prio
), &t
);
4742 read_unlock(&tasklist_lock
);
4743 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4747 read_unlock(&tasklist_lock
);
4751 static const char stat_nam
[] = "RSDTtZX";
4753 static void show_task(struct task_struct
*p
)
4755 unsigned long free
= 0;
4758 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4759 printk("%-13.13s %c", p
->comm
,
4760 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4761 #if BITS_PER_LONG == 32
4762 if (state
== TASK_RUNNING
)
4763 printk(" running ");
4765 printk(" %08lx ", thread_saved_pc(p
));
4767 if (state
== TASK_RUNNING
)
4768 printk(" running task ");
4770 printk(" %016lx ", thread_saved_pc(p
));
4772 #ifdef CONFIG_DEBUG_STACK_USAGE
4774 unsigned long *n
= end_of_stack(p
);
4777 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4780 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4782 if (state
!= TASK_RUNNING
)
4783 show_stack(p
, NULL
);
4786 void show_state_filter(unsigned long state_filter
)
4788 struct task_struct
*g
, *p
;
4790 #if BITS_PER_LONG == 32
4792 " task PC stack pid father\n");
4795 " task PC stack pid father\n");
4797 read_lock(&tasklist_lock
);
4798 do_each_thread(g
, p
) {
4800 * reset the NMI-timeout, listing all files on a slow
4801 * console might take alot of time:
4803 touch_nmi_watchdog();
4804 if (!state_filter
|| (p
->state
& state_filter
))
4806 } while_each_thread(g
, p
);
4808 touch_all_softlockup_watchdogs();
4810 #ifdef CONFIG_SCHED_DEBUG
4811 sysrq_sched_debug_show();
4813 read_unlock(&tasklist_lock
);
4815 * Only show locks if all tasks are dumped:
4817 if (state_filter
== -1)
4818 debug_show_all_locks();
4821 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4823 idle
->sched_class
= &idle_sched_class
;
4827 * init_idle - set up an idle thread for a given CPU
4828 * @idle: task in question
4829 * @cpu: cpu the idle task belongs to
4831 * NOTE: this function does not set the idle thread's NEED_RESCHED
4832 * flag, to make booting more robust.
4834 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4836 struct rq
*rq
= cpu_rq(cpu
);
4837 unsigned long flags
;
4840 idle
->se
.exec_start
= sched_clock();
4842 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4843 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4844 __set_task_cpu(idle
, cpu
);
4846 spin_lock_irqsave(&rq
->lock
, flags
);
4847 rq
->curr
= rq
->idle
= idle
;
4848 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4851 spin_unlock_irqrestore(&rq
->lock
, flags
);
4853 /* Set the preempt count _outside_ the spinlocks! */
4854 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4855 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4857 task_thread_info(idle
)->preempt_count
= 0;
4860 * The idle tasks have their own, simple scheduling class:
4862 idle
->sched_class
= &idle_sched_class
;
4866 * In a system that switches off the HZ timer nohz_cpu_mask
4867 * indicates which cpus entered this state. This is used
4868 * in the rcu update to wait only for active cpus. For system
4869 * which do not switch off the HZ timer nohz_cpu_mask should
4870 * always be CPU_MASK_NONE.
4872 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4876 * This is how migration works:
4878 * 1) we queue a struct migration_req structure in the source CPU's
4879 * runqueue and wake up that CPU's migration thread.
4880 * 2) we down() the locked semaphore => thread blocks.
4881 * 3) migration thread wakes up (implicitly it forces the migrated
4882 * thread off the CPU)
4883 * 4) it gets the migration request and checks whether the migrated
4884 * task is still in the wrong runqueue.
4885 * 5) if it's in the wrong runqueue then the migration thread removes
4886 * it and puts it into the right queue.
4887 * 6) migration thread up()s the semaphore.
4888 * 7) we wake up and the migration is done.
4892 * Change a given task's CPU affinity. Migrate the thread to a
4893 * proper CPU and schedule it away if the CPU it's executing on
4894 * is removed from the allowed bitmask.
4896 * NOTE: the caller must have a valid reference to the task, the
4897 * task must not exit() & deallocate itself prematurely. The
4898 * call is not atomic; no spinlocks may be held.
4900 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4902 struct migration_req req
;
4903 unsigned long flags
;
4907 rq
= task_rq_lock(p
, &flags
);
4908 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4913 p
->cpus_allowed
= new_mask
;
4914 /* Can the task run on the task's current CPU? If so, we're done */
4915 if (cpu_isset(task_cpu(p
), new_mask
))
4918 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4919 /* Need help from migration thread: drop lock and wait. */
4920 task_rq_unlock(rq
, &flags
);
4921 wake_up_process(rq
->migration_thread
);
4922 wait_for_completion(&req
.done
);
4923 tlb_migrate_finish(p
->mm
);
4927 task_rq_unlock(rq
, &flags
);
4931 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4934 * Move (not current) task off this cpu, onto dest cpu. We're doing
4935 * this because either it can't run here any more (set_cpus_allowed()
4936 * away from this CPU, or CPU going down), or because we're
4937 * attempting to rebalance this task on exec (sched_exec).
4939 * So we race with normal scheduler movements, but that's OK, as long
4940 * as the task is no longer on this CPU.
4942 * Returns non-zero if task was successfully migrated.
4944 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4946 struct rq
*rq_dest
, *rq_src
;
4949 if (unlikely(cpu_is_offline(dest_cpu
)))
4952 rq_src
= cpu_rq(src_cpu
);
4953 rq_dest
= cpu_rq(dest_cpu
);
4955 double_rq_lock(rq_src
, rq_dest
);
4956 /* Already moved. */
4957 if (task_cpu(p
) != src_cpu
)
4959 /* Affinity changed (again). */
4960 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4963 on_rq
= p
->se
.on_rq
;
4965 deactivate_task(rq_src
, p
, 0);
4967 set_task_cpu(p
, dest_cpu
);
4969 activate_task(rq_dest
, p
, 0);
4970 check_preempt_curr(rq_dest
, p
);
4974 double_rq_unlock(rq_src
, rq_dest
);
4979 * migration_thread - this is a highprio system thread that performs
4980 * thread migration by bumping thread off CPU then 'pushing' onto
4983 static int migration_thread(void *data
)
4985 int cpu
= (long)data
;
4989 BUG_ON(rq
->migration_thread
!= current
);
4991 set_current_state(TASK_INTERRUPTIBLE
);
4992 while (!kthread_should_stop()) {
4993 struct migration_req
*req
;
4994 struct list_head
*head
;
4996 spin_lock_irq(&rq
->lock
);
4998 if (cpu_is_offline(cpu
)) {
4999 spin_unlock_irq(&rq
->lock
);
5003 if (rq
->active_balance
) {
5004 active_load_balance(rq
, cpu
);
5005 rq
->active_balance
= 0;
5008 head
= &rq
->migration_queue
;
5010 if (list_empty(head
)) {
5011 spin_unlock_irq(&rq
->lock
);
5013 set_current_state(TASK_INTERRUPTIBLE
);
5016 req
= list_entry(head
->next
, struct migration_req
, list
);
5017 list_del_init(head
->next
);
5019 spin_unlock(&rq
->lock
);
5020 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5023 complete(&req
->done
);
5025 __set_current_state(TASK_RUNNING
);
5029 /* Wait for kthread_stop */
5030 set_current_state(TASK_INTERRUPTIBLE
);
5031 while (!kthread_should_stop()) {
5033 set_current_state(TASK_INTERRUPTIBLE
);
5035 __set_current_state(TASK_RUNNING
);
5039 #ifdef CONFIG_HOTPLUG_CPU
5041 * Figure out where task on dead CPU should go, use force if neccessary.
5042 * NOTE: interrupts should be disabled by the caller
5044 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5046 unsigned long flags
;
5053 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5054 cpus_and(mask
, mask
, p
->cpus_allowed
);
5055 dest_cpu
= any_online_cpu(mask
);
5057 /* On any allowed CPU? */
5058 if (dest_cpu
== NR_CPUS
)
5059 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5061 /* No more Mr. Nice Guy. */
5062 if (dest_cpu
== NR_CPUS
) {
5063 rq
= task_rq_lock(p
, &flags
);
5064 cpus_setall(p
->cpus_allowed
);
5065 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5066 task_rq_unlock(rq
, &flags
);
5069 * Don't tell them about moving exiting tasks or
5070 * kernel threads (both mm NULL), since they never
5073 if (p
->mm
&& printk_ratelimit())
5074 printk(KERN_INFO
"process %d (%s) no "
5075 "longer affine to cpu%d\n",
5076 p
->pid
, p
->comm
, dead_cpu
);
5078 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5083 * While a dead CPU has no uninterruptible tasks queued at this point,
5084 * it might still have a nonzero ->nr_uninterruptible counter, because
5085 * for performance reasons the counter is not stricly tracking tasks to
5086 * their home CPUs. So we just add the counter to another CPU's counter,
5087 * to keep the global sum constant after CPU-down:
5089 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5091 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5092 unsigned long flags
;
5094 local_irq_save(flags
);
5095 double_rq_lock(rq_src
, rq_dest
);
5096 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5097 rq_src
->nr_uninterruptible
= 0;
5098 double_rq_unlock(rq_src
, rq_dest
);
5099 local_irq_restore(flags
);
5102 /* Run through task list and migrate tasks from the dead cpu. */
5103 static void migrate_live_tasks(int src_cpu
)
5105 struct task_struct
*p
, *t
;
5107 write_lock_irq(&tasklist_lock
);
5109 do_each_thread(t
, p
) {
5113 if (task_cpu(p
) == src_cpu
)
5114 move_task_off_dead_cpu(src_cpu
, p
);
5115 } while_each_thread(t
, p
);
5117 write_unlock_irq(&tasklist_lock
);
5121 * Schedules idle task to be the next runnable task on current CPU.
5122 * It does so by boosting its priority to highest possible and adding it to
5123 * the _front_ of the runqueue. Used by CPU offline code.
5125 void sched_idle_next(void)
5127 int this_cpu
= smp_processor_id();
5128 struct rq
*rq
= cpu_rq(this_cpu
);
5129 struct task_struct
*p
= rq
->idle
;
5130 unsigned long flags
;
5132 /* cpu has to be offline */
5133 BUG_ON(cpu_online(this_cpu
));
5136 * Strictly not necessary since rest of the CPUs are stopped by now
5137 * and interrupts disabled on the current cpu.
5139 spin_lock_irqsave(&rq
->lock
, flags
);
5141 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5143 /* Add idle task to the _front_ of its priority queue: */
5144 activate_idle_task(p
, rq
);
5146 spin_unlock_irqrestore(&rq
->lock
, flags
);
5150 * Ensures that the idle task is using init_mm right before its cpu goes
5153 void idle_task_exit(void)
5155 struct mm_struct
*mm
= current
->active_mm
;
5157 BUG_ON(cpu_online(smp_processor_id()));
5160 switch_mm(mm
, &init_mm
, current
);
5164 /* called under rq->lock with disabled interrupts */
5165 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5167 struct rq
*rq
= cpu_rq(dead_cpu
);
5169 /* Must be exiting, otherwise would be on tasklist. */
5170 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5172 /* Cannot have done final schedule yet: would have vanished. */
5173 BUG_ON(p
->state
== TASK_DEAD
);
5178 * Drop lock around migration; if someone else moves it,
5179 * that's OK. No task can be added to this CPU, so iteration is
5181 * NOTE: interrupts should be left disabled --dev@
5183 spin_unlock(&rq
->lock
);
5184 move_task_off_dead_cpu(dead_cpu
, p
);
5185 spin_lock(&rq
->lock
);
5190 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5191 static void migrate_dead_tasks(unsigned int dead_cpu
)
5193 struct rq
*rq
= cpu_rq(dead_cpu
);
5194 struct task_struct
*next
;
5197 if (!rq
->nr_running
)
5199 update_rq_clock(rq
);
5200 next
= pick_next_task(rq
, rq
->curr
);
5203 migrate_dead(dead_cpu
, next
);
5207 #endif /* CONFIG_HOTPLUG_CPU */
5209 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5211 static struct ctl_table sd_ctl_dir
[] = {
5213 .procname
= "sched_domain",
5219 static struct ctl_table sd_ctl_root
[] = {
5221 .ctl_name
= CTL_KERN
,
5222 .procname
= "kernel",
5224 .child
= sd_ctl_dir
,
5229 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5231 struct ctl_table
*entry
=
5232 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5235 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5241 set_table_entry(struct ctl_table
*entry
,
5242 const char *procname
, void *data
, int maxlen
,
5243 mode_t mode
, proc_handler
*proc_handler
)
5245 entry
->procname
= procname
;
5247 entry
->maxlen
= maxlen
;
5249 entry
->proc_handler
= proc_handler
;
5252 static struct ctl_table
*
5253 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5255 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5257 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5258 sizeof(long), 0644, proc_doulongvec_minmax
);
5259 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5260 sizeof(long), 0644, proc_doulongvec_minmax
);
5261 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5262 sizeof(int), 0644, proc_dointvec_minmax
);
5263 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5264 sizeof(int), 0644, proc_dointvec_minmax
);
5265 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5266 sizeof(int), 0644, proc_dointvec_minmax
);
5267 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5268 sizeof(int), 0644, proc_dointvec_minmax
);
5269 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5270 sizeof(int), 0644, proc_dointvec_minmax
);
5271 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5272 sizeof(int), 0644, proc_dointvec_minmax
);
5273 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5274 sizeof(int), 0644, proc_dointvec_minmax
);
5275 set_table_entry(&table
[10], "cache_nice_tries",
5276 &sd
->cache_nice_tries
,
5277 sizeof(int), 0644, proc_dointvec_minmax
);
5278 set_table_entry(&table
[12], "flags", &sd
->flags
,
5279 sizeof(int), 0644, proc_dointvec_minmax
);
5284 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5286 struct ctl_table
*entry
, *table
;
5287 struct sched_domain
*sd
;
5288 int domain_num
= 0, i
;
5291 for_each_domain(cpu
, sd
)
5293 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5296 for_each_domain(cpu
, sd
) {
5297 snprintf(buf
, 32, "domain%d", i
);
5298 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5300 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5307 static struct ctl_table_header
*sd_sysctl_header
;
5308 static void init_sched_domain_sysctl(void)
5310 int i
, cpu_num
= num_online_cpus();
5311 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5314 sd_ctl_dir
[0].child
= entry
;
5316 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5317 snprintf(buf
, 32, "cpu%d", i
);
5318 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5320 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5322 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5325 static void init_sched_domain_sysctl(void)
5331 * migration_call - callback that gets triggered when a CPU is added.
5332 * Here we can start up the necessary migration thread for the new CPU.
5334 static int __cpuinit
5335 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5337 struct task_struct
*p
;
5338 int cpu
= (long)hcpu
;
5339 unsigned long flags
;
5343 case CPU_LOCK_ACQUIRE
:
5344 mutex_lock(&sched_hotcpu_mutex
);
5347 case CPU_UP_PREPARE
:
5348 case CPU_UP_PREPARE_FROZEN
:
5349 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5352 kthread_bind(p
, cpu
);
5353 /* Must be high prio: stop_machine expects to yield to it. */
5354 rq
= task_rq_lock(p
, &flags
);
5355 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5356 task_rq_unlock(rq
, &flags
);
5357 cpu_rq(cpu
)->migration_thread
= p
;
5361 case CPU_ONLINE_FROZEN
:
5362 /* Strictly unneccessary, as first user will wake it. */
5363 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5366 #ifdef CONFIG_HOTPLUG_CPU
5367 case CPU_UP_CANCELED
:
5368 case CPU_UP_CANCELED_FROZEN
:
5369 if (!cpu_rq(cpu
)->migration_thread
)
5371 /* Unbind it from offline cpu so it can run. Fall thru. */
5372 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5373 any_online_cpu(cpu_online_map
));
5374 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5375 cpu_rq(cpu
)->migration_thread
= NULL
;
5379 case CPU_DEAD_FROZEN
:
5380 migrate_live_tasks(cpu
);
5382 kthread_stop(rq
->migration_thread
);
5383 rq
->migration_thread
= NULL
;
5384 /* Idle task back to normal (off runqueue, low prio) */
5385 rq
= task_rq_lock(rq
->idle
, &flags
);
5386 update_rq_clock(rq
);
5387 deactivate_task(rq
, rq
->idle
, 0);
5388 rq
->idle
->static_prio
= MAX_PRIO
;
5389 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5390 rq
->idle
->sched_class
= &idle_sched_class
;
5391 migrate_dead_tasks(cpu
);
5392 task_rq_unlock(rq
, &flags
);
5393 migrate_nr_uninterruptible(rq
);
5394 BUG_ON(rq
->nr_running
!= 0);
5396 /* No need to migrate the tasks: it was best-effort if
5397 * they didn't take sched_hotcpu_mutex. Just wake up
5398 * the requestors. */
5399 spin_lock_irq(&rq
->lock
);
5400 while (!list_empty(&rq
->migration_queue
)) {
5401 struct migration_req
*req
;
5403 req
= list_entry(rq
->migration_queue
.next
,
5404 struct migration_req
, list
);
5405 list_del_init(&req
->list
);
5406 complete(&req
->done
);
5408 spin_unlock_irq(&rq
->lock
);
5411 case CPU_LOCK_RELEASE
:
5412 mutex_unlock(&sched_hotcpu_mutex
);
5418 /* Register at highest priority so that task migration (migrate_all_tasks)
5419 * happens before everything else.
5421 static struct notifier_block __cpuinitdata migration_notifier
= {
5422 .notifier_call
= migration_call
,
5426 int __init
migration_init(void)
5428 void *cpu
= (void *)(long)smp_processor_id();
5431 /* Start one for the boot CPU: */
5432 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5433 BUG_ON(err
== NOTIFY_BAD
);
5434 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5435 register_cpu_notifier(&migration_notifier
);
5443 /* Number of possible processor ids */
5444 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5445 EXPORT_SYMBOL(nr_cpu_ids
);
5447 #undef SCHED_DOMAIN_DEBUG
5448 #ifdef SCHED_DOMAIN_DEBUG
5449 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5454 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5458 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5463 struct sched_group
*group
= sd
->groups
;
5464 cpumask_t groupmask
;
5466 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5467 cpus_clear(groupmask
);
5470 for (i
= 0; i
< level
+ 1; i
++)
5472 printk("domain %d: ", level
);
5474 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5475 printk("does not load-balance\n");
5477 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5482 printk("span %s\n", str
);
5484 if (!cpu_isset(cpu
, sd
->span
))
5485 printk(KERN_ERR
"ERROR: domain->span does not contain "
5487 if (!cpu_isset(cpu
, group
->cpumask
))
5488 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5492 for (i
= 0; i
< level
+ 2; i
++)
5498 printk(KERN_ERR
"ERROR: group is NULL\n");
5502 if (!group
->__cpu_power
) {
5504 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5508 if (!cpus_weight(group
->cpumask
)) {
5510 printk(KERN_ERR
"ERROR: empty group\n");
5513 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5515 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5518 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5520 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5523 group
= group
->next
;
5524 } while (group
!= sd
->groups
);
5527 if (!cpus_equal(sd
->span
, groupmask
))
5528 printk(KERN_ERR
"ERROR: groups don't span "
5536 if (!cpus_subset(groupmask
, sd
->span
))
5537 printk(KERN_ERR
"ERROR: parent span is not a superset "
5538 "of domain->span\n");
5543 # define sched_domain_debug(sd, cpu) do { } while (0)
5546 static int sd_degenerate(struct sched_domain
*sd
)
5548 if (cpus_weight(sd
->span
) == 1)
5551 /* Following flags need at least 2 groups */
5552 if (sd
->flags
& (SD_LOAD_BALANCE
|
5553 SD_BALANCE_NEWIDLE
|
5557 SD_SHARE_PKG_RESOURCES
)) {
5558 if (sd
->groups
!= sd
->groups
->next
)
5562 /* Following flags don't use groups */
5563 if (sd
->flags
& (SD_WAKE_IDLE
|
5572 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5574 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5576 if (sd_degenerate(parent
))
5579 if (!cpus_equal(sd
->span
, parent
->span
))
5582 /* Does parent contain flags not in child? */
5583 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5584 if (cflags
& SD_WAKE_AFFINE
)
5585 pflags
&= ~SD_WAKE_BALANCE
;
5586 /* Flags needing groups don't count if only 1 group in parent */
5587 if (parent
->groups
== parent
->groups
->next
) {
5588 pflags
&= ~(SD_LOAD_BALANCE
|
5589 SD_BALANCE_NEWIDLE
|
5593 SD_SHARE_PKG_RESOURCES
);
5595 if (~cflags
& pflags
)
5602 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5603 * hold the hotplug lock.
5605 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5607 struct rq
*rq
= cpu_rq(cpu
);
5608 struct sched_domain
*tmp
;
5610 /* Remove the sched domains which do not contribute to scheduling. */
5611 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5612 struct sched_domain
*parent
= tmp
->parent
;
5615 if (sd_parent_degenerate(tmp
, parent
)) {
5616 tmp
->parent
= parent
->parent
;
5618 parent
->parent
->child
= tmp
;
5622 if (sd
&& sd_degenerate(sd
)) {
5628 sched_domain_debug(sd
, cpu
);
5630 rcu_assign_pointer(rq
->sd
, sd
);
5633 /* cpus with isolated domains */
5634 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5636 /* Setup the mask of cpus configured for isolated domains */
5637 static int __init
isolated_cpu_setup(char *str
)
5639 int ints
[NR_CPUS
], i
;
5641 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5642 cpus_clear(cpu_isolated_map
);
5643 for (i
= 1; i
<= ints
[0]; i
++)
5644 if (ints
[i
] < NR_CPUS
)
5645 cpu_set(ints
[i
], cpu_isolated_map
);
5649 __setup ("isolcpus=", isolated_cpu_setup
);
5652 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5653 * to a function which identifies what group(along with sched group) a CPU
5654 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5655 * (due to the fact that we keep track of groups covered with a cpumask_t).
5657 * init_sched_build_groups will build a circular linked list of the groups
5658 * covered by the given span, and will set each group's ->cpumask correctly,
5659 * and ->cpu_power to 0.
5662 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5663 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5664 struct sched_group
**sg
))
5666 struct sched_group
*first
= NULL
, *last
= NULL
;
5667 cpumask_t covered
= CPU_MASK_NONE
;
5670 for_each_cpu_mask(i
, span
) {
5671 struct sched_group
*sg
;
5672 int group
= group_fn(i
, cpu_map
, &sg
);
5675 if (cpu_isset(i
, covered
))
5678 sg
->cpumask
= CPU_MASK_NONE
;
5679 sg
->__cpu_power
= 0;
5681 for_each_cpu_mask(j
, span
) {
5682 if (group_fn(j
, cpu_map
, NULL
) != group
)
5685 cpu_set(j
, covered
);
5686 cpu_set(j
, sg
->cpumask
);
5697 #define SD_NODES_PER_DOMAIN 16
5702 * find_next_best_node - find the next node to include in a sched_domain
5703 * @node: node whose sched_domain we're building
5704 * @used_nodes: nodes already in the sched_domain
5706 * Find the next node to include in a given scheduling domain. Simply
5707 * finds the closest node not already in the @used_nodes map.
5709 * Should use nodemask_t.
5711 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5713 int i
, n
, val
, min_val
, best_node
= 0;
5717 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5718 /* Start at @node */
5719 n
= (node
+ i
) % MAX_NUMNODES
;
5721 if (!nr_cpus_node(n
))
5724 /* Skip already used nodes */
5725 if (test_bit(n
, used_nodes
))
5728 /* Simple min distance search */
5729 val
= node_distance(node
, n
);
5731 if (val
< min_val
) {
5737 set_bit(best_node
, used_nodes
);
5742 * sched_domain_node_span - get a cpumask for a node's sched_domain
5743 * @node: node whose cpumask we're constructing
5744 * @size: number of nodes to include in this span
5746 * Given a node, construct a good cpumask for its sched_domain to span. It
5747 * should be one that prevents unnecessary balancing, but also spreads tasks
5750 static cpumask_t
sched_domain_node_span(int node
)
5752 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5753 cpumask_t span
, nodemask
;
5757 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5759 nodemask
= node_to_cpumask(node
);
5760 cpus_or(span
, span
, nodemask
);
5761 set_bit(node
, used_nodes
);
5763 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5764 int next_node
= find_next_best_node(node
, used_nodes
);
5766 nodemask
= node_to_cpumask(next_node
);
5767 cpus_or(span
, span
, nodemask
);
5774 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5777 * SMT sched-domains:
5779 #ifdef CONFIG_SCHED_SMT
5780 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5781 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5783 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5784 struct sched_group
**sg
)
5787 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5793 * multi-core sched-domains:
5795 #ifdef CONFIG_SCHED_MC
5796 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5797 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5800 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5801 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5802 struct sched_group
**sg
)
5805 cpumask_t mask
= cpu_sibling_map
[cpu
];
5806 cpus_and(mask
, mask
, *cpu_map
);
5807 group
= first_cpu(mask
);
5809 *sg
= &per_cpu(sched_group_core
, group
);
5812 #elif defined(CONFIG_SCHED_MC)
5813 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5814 struct sched_group
**sg
)
5817 *sg
= &per_cpu(sched_group_core
, cpu
);
5822 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5823 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5825 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5826 struct sched_group
**sg
)
5829 #ifdef CONFIG_SCHED_MC
5830 cpumask_t mask
= cpu_coregroup_map(cpu
);
5831 cpus_and(mask
, mask
, *cpu_map
);
5832 group
= first_cpu(mask
);
5833 #elif defined(CONFIG_SCHED_SMT)
5834 cpumask_t mask
= cpu_sibling_map
[cpu
];
5835 cpus_and(mask
, mask
, *cpu_map
);
5836 group
= first_cpu(mask
);
5841 *sg
= &per_cpu(sched_group_phys
, group
);
5847 * The init_sched_build_groups can't handle what we want to do with node
5848 * groups, so roll our own. Now each node has its own list of groups which
5849 * gets dynamically allocated.
5851 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5852 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5854 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5855 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5857 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5858 struct sched_group
**sg
)
5860 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5863 cpus_and(nodemask
, nodemask
, *cpu_map
);
5864 group
= first_cpu(nodemask
);
5867 *sg
= &per_cpu(sched_group_allnodes
, group
);
5871 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5873 struct sched_group
*sg
= group_head
;
5879 for_each_cpu_mask(j
, sg
->cpumask
) {
5880 struct sched_domain
*sd
;
5882 sd
= &per_cpu(phys_domains
, j
);
5883 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5885 * Only add "power" once for each
5891 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5894 if (sg
!= group_head
)
5900 /* Free memory allocated for various sched_group structures */
5901 static void free_sched_groups(const cpumask_t
*cpu_map
)
5905 for_each_cpu_mask(cpu
, *cpu_map
) {
5906 struct sched_group
**sched_group_nodes
5907 = sched_group_nodes_bycpu
[cpu
];
5909 if (!sched_group_nodes
)
5912 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5913 cpumask_t nodemask
= node_to_cpumask(i
);
5914 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5916 cpus_and(nodemask
, nodemask
, *cpu_map
);
5917 if (cpus_empty(nodemask
))
5927 if (oldsg
!= sched_group_nodes
[i
])
5930 kfree(sched_group_nodes
);
5931 sched_group_nodes_bycpu
[cpu
] = NULL
;
5935 static void free_sched_groups(const cpumask_t
*cpu_map
)
5941 * Initialize sched groups cpu_power.
5943 * cpu_power indicates the capacity of sched group, which is used while
5944 * distributing the load between different sched groups in a sched domain.
5945 * Typically cpu_power for all the groups in a sched domain will be same unless
5946 * there are asymmetries in the topology. If there are asymmetries, group
5947 * having more cpu_power will pickup more load compared to the group having
5950 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5951 * the maximum number of tasks a group can handle in the presence of other idle
5952 * or lightly loaded groups in the same sched domain.
5954 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5956 struct sched_domain
*child
;
5957 struct sched_group
*group
;
5959 WARN_ON(!sd
|| !sd
->groups
);
5961 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5966 sd
->groups
->__cpu_power
= 0;
5969 * For perf policy, if the groups in child domain share resources
5970 * (for example cores sharing some portions of the cache hierarchy
5971 * or SMT), then set this domain groups cpu_power such that each group
5972 * can handle only one task, when there are other idle groups in the
5973 * same sched domain.
5975 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5977 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5978 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5983 * add cpu_power of each child group to this groups cpu_power
5985 group
= child
->groups
;
5987 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5988 group
= group
->next
;
5989 } while (group
!= child
->groups
);
5993 * Build sched domains for a given set of cpus and attach the sched domains
5994 * to the individual cpus
5996 static int build_sched_domains(const cpumask_t
*cpu_map
)
6000 struct sched_group
**sched_group_nodes
= NULL
;
6001 int sd_allnodes
= 0;
6004 * Allocate the per-node list of sched groups
6006 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6008 if (!sched_group_nodes
) {
6009 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6012 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6016 * Set up domains for cpus specified by the cpu_map.
6018 for_each_cpu_mask(i
, *cpu_map
) {
6019 struct sched_domain
*sd
= NULL
, *p
;
6020 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6022 cpus_and(nodemask
, nodemask
, *cpu_map
);
6025 if (cpus_weight(*cpu_map
) >
6026 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6027 sd
= &per_cpu(allnodes_domains
, i
);
6028 *sd
= SD_ALLNODES_INIT
;
6029 sd
->span
= *cpu_map
;
6030 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6036 sd
= &per_cpu(node_domains
, i
);
6038 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6042 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6046 sd
= &per_cpu(phys_domains
, i
);
6048 sd
->span
= nodemask
;
6052 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6054 #ifdef CONFIG_SCHED_MC
6056 sd
= &per_cpu(core_domains
, i
);
6058 sd
->span
= cpu_coregroup_map(i
);
6059 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6062 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6065 #ifdef CONFIG_SCHED_SMT
6067 sd
= &per_cpu(cpu_domains
, i
);
6068 *sd
= SD_SIBLING_INIT
;
6069 sd
->span
= cpu_sibling_map
[i
];
6070 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6073 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6077 #ifdef CONFIG_SCHED_SMT
6078 /* Set up CPU (sibling) groups */
6079 for_each_cpu_mask(i
, *cpu_map
) {
6080 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6081 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6082 if (i
!= first_cpu(this_sibling_map
))
6085 init_sched_build_groups(this_sibling_map
, cpu_map
,
6090 #ifdef CONFIG_SCHED_MC
6091 /* Set up multi-core groups */
6092 for_each_cpu_mask(i
, *cpu_map
) {
6093 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6094 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6095 if (i
!= first_cpu(this_core_map
))
6097 init_sched_build_groups(this_core_map
, cpu_map
,
6098 &cpu_to_core_group
);
6102 /* Set up physical groups */
6103 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6104 cpumask_t nodemask
= node_to_cpumask(i
);
6106 cpus_and(nodemask
, nodemask
, *cpu_map
);
6107 if (cpus_empty(nodemask
))
6110 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6114 /* Set up node groups */
6116 init_sched_build_groups(*cpu_map
, cpu_map
,
6117 &cpu_to_allnodes_group
);
6119 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6120 /* Set up node groups */
6121 struct sched_group
*sg
, *prev
;
6122 cpumask_t nodemask
= node_to_cpumask(i
);
6123 cpumask_t domainspan
;
6124 cpumask_t covered
= CPU_MASK_NONE
;
6127 cpus_and(nodemask
, nodemask
, *cpu_map
);
6128 if (cpus_empty(nodemask
)) {
6129 sched_group_nodes
[i
] = NULL
;
6133 domainspan
= sched_domain_node_span(i
);
6134 cpus_and(domainspan
, domainspan
, *cpu_map
);
6136 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6138 printk(KERN_WARNING
"Can not alloc domain group for "
6142 sched_group_nodes
[i
] = sg
;
6143 for_each_cpu_mask(j
, nodemask
) {
6144 struct sched_domain
*sd
;
6146 sd
= &per_cpu(node_domains
, j
);
6149 sg
->__cpu_power
= 0;
6150 sg
->cpumask
= nodemask
;
6152 cpus_or(covered
, covered
, nodemask
);
6155 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6156 cpumask_t tmp
, notcovered
;
6157 int n
= (i
+ j
) % MAX_NUMNODES
;
6159 cpus_complement(notcovered
, covered
);
6160 cpus_and(tmp
, notcovered
, *cpu_map
);
6161 cpus_and(tmp
, tmp
, domainspan
);
6162 if (cpus_empty(tmp
))
6165 nodemask
= node_to_cpumask(n
);
6166 cpus_and(tmp
, tmp
, nodemask
);
6167 if (cpus_empty(tmp
))
6170 sg
= kmalloc_node(sizeof(struct sched_group
),
6174 "Can not alloc domain group for node %d\n", j
);
6177 sg
->__cpu_power
= 0;
6179 sg
->next
= prev
->next
;
6180 cpus_or(covered
, covered
, tmp
);
6187 /* Calculate CPU power for physical packages and nodes */
6188 #ifdef CONFIG_SCHED_SMT
6189 for_each_cpu_mask(i
, *cpu_map
) {
6190 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6192 init_sched_groups_power(i
, sd
);
6195 #ifdef CONFIG_SCHED_MC
6196 for_each_cpu_mask(i
, *cpu_map
) {
6197 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6199 init_sched_groups_power(i
, sd
);
6203 for_each_cpu_mask(i
, *cpu_map
) {
6204 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6206 init_sched_groups_power(i
, sd
);
6210 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6211 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6214 struct sched_group
*sg
;
6216 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6217 init_numa_sched_groups_power(sg
);
6221 /* Attach the domains */
6222 for_each_cpu_mask(i
, *cpu_map
) {
6223 struct sched_domain
*sd
;
6224 #ifdef CONFIG_SCHED_SMT
6225 sd
= &per_cpu(cpu_domains
, i
);
6226 #elif defined(CONFIG_SCHED_MC)
6227 sd
= &per_cpu(core_domains
, i
);
6229 sd
= &per_cpu(phys_domains
, i
);
6231 cpu_attach_domain(sd
, i
);
6238 free_sched_groups(cpu_map
);
6243 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6245 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6247 cpumask_t cpu_default_map
;
6251 * Setup mask for cpus without special case scheduling requirements.
6252 * For now this just excludes isolated cpus, but could be used to
6253 * exclude other special cases in the future.
6255 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6257 err
= build_sched_domains(&cpu_default_map
);
6262 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6264 free_sched_groups(cpu_map
);
6268 * Detach sched domains from a group of cpus specified in cpu_map
6269 * These cpus will now be attached to the NULL domain
6271 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6275 for_each_cpu_mask(i
, *cpu_map
)
6276 cpu_attach_domain(NULL
, i
);
6277 synchronize_sched();
6278 arch_destroy_sched_domains(cpu_map
);
6282 * Partition sched domains as specified by the cpumasks below.
6283 * This attaches all cpus from the cpumasks to the NULL domain,
6284 * waits for a RCU quiescent period, recalculates sched
6285 * domain information and then attaches them back to the
6286 * correct sched domains
6287 * Call with hotplug lock held
6289 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6291 cpumask_t change_map
;
6294 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6295 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6296 cpus_or(change_map
, *partition1
, *partition2
);
6298 /* Detach sched domains from all of the affected cpus */
6299 detach_destroy_domains(&change_map
);
6300 if (!cpus_empty(*partition1
))
6301 err
= build_sched_domains(partition1
);
6302 if (!err
&& !cpus_empty(*partition2
))
6303 err
= build_sched_domains(partition2
);
6308 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6309 static int arch_reinit_sched_domains(void)
6313 mutex_lock(&sched_hotcpu_mutex
);
6314 detach_destroy_domains(&cpu_online_map
);
6315 err
= arch_init_sched_domains(&cpu_online_map
);
6316 mutex_unlock(&sched_hotcpu_mutex
);
6321 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6325 if (buf
[0] != '0' && buf
[0] != '1')
6329 sched_smt_power_savings
= (buf
[0] == '1');
6331 sched_mc_power_savings
= (buf
[0] == '1');
6333 ret
= arch_reinit_sched_domains();
6335 return ret
? ret
: count
;
6338 #ifdef CONFIG_SCHED_MC
6339 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6341 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6343 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6344 const char *buf
, size_t count
)
6346 return sched_power_savings_store(buf
, count
, 0);
6348 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6349 sched_mc_power_savings_store
);
6352 #ifdef CONFIG_SCHED_SMT
6353 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6355 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6357 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6358 const char *buf
, size_t count
)
6360 return sched_power_savings_store(buf
, count
, 1);
6362 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6363 sched_smt_power_savings_store
);
6366 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6370 #ifdef CONFIG_SCHED_SMT
6372 err
= sysfs_create_file(&cls
->kset
.kobj
,
6373 &attr_sched_smt_power_savings
.attr
);
6375 #ifdef CONFIG_SCHED_MC
6376 if (!err
&& mc_capable())
6377 err
= sysfs_create_file(&cls
->kset
.kobj
,
6378 &attr_sched_mc_power_savings
.attr
);
6385 * Force a reinitialization of the sched domains hierarchy. The domains
6386 * and groups cannot be updated in place without racing with the balancing
6387 * code, so we temporarily attach all running cpus to the NULL domain
6388 * which will prevent rebalancing while the sched domains are recalculated.
6390 static int update_sched_domains(struct notifier_block
*nfb
,
6391 unsigned long action
, void *hcpu
)
6394 case CPU_UP_PREPARE
:
6395 case CPU_UP_PREPARE_FROZEN
:
6396 case CPU_DOWN_PREPARE
:
6397 case CPU_DOWN_PREPARE_FROZEN
:
6398 detach_destroy_domains(&cpu_online_map
);
6401 case CPU_UP_CANCELED
:
6402 case CPU_UP_CANCELED_FROZEN
:
6403 case CPU_DOWN_FAILED
:
6404 case CPU_DOWN_FAILED_FROZEN
:
6406 case CPU_ONLINE_FROZEN
:
6408 case CPU_DEAD_FROZEN
:
6410 * Fall through and re-initialise the domains.
6417 /* The hotplug lock is already held by cpu_up/cpu_down */
6418 arch_init_sched_domains(&cpu_online_map
);
6423 void __init
sched_init_smp(void)
6425 cpumask_t non_isolated_cpus
;
6427 mutex_lock(&sched_hotcpu_mutex
);
6428 arch_init_sched_domains(&cpu_online_map
);
6429 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6430 if (cpus_empty(non_isolated_cpus
))
6431 cpu_set(smp_processor_id(), non_isolated_cpus
);
6432 mutex_unlock(&sched_hotcpu_mutex
);
6433 /* XXX: Theoretical race here - CPU may be hotplugged now */
6434 hotcpu_notifier(update_sched_domains
, 0);
6436 init_sched_domain_sysctl();
6438 /* Move init over to a non-isolated CPU */
6439 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6443 void __init
sched_init_smp(void)
6446 #endif /* CONFIG_SMP */
6448 int in_sched_functions(unsigned long addr
)
6450 /* Linker adds these: start and end of __sched functions */
6451 extern char __sched_text_start
[], __sched_text_end
[];
6453 return in_lock_functions(addr
) ||
6454 (addr
>= (unsigned long)__sched_text_start
6455 && addr
< (unsigned long)__sched_text_end
);
6458 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6460 cfs_rq
->tasks_timeline
= RB_ROOT
;
6461 #ifdef CONFIG_FAIR_GROUP_SCHED
6466 void __init
sched_init(void)
6468 int highest_cpu
= 0;
6472 * Link up the scheduling class hierarchy:
6474 rt_sched_class
.next
= &fair_sched_class
;
6475 fair_sched_class
.next
= &idle_sched_class
;
6476 idle_sched_class
.next
= NULL
;
6478 for_each_possible_cpu(i
) {
6479 struct rt_prio_array
*array
;
6483 spin_lock_init(&rq
->lock
);
6484 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6487 init_cfs_rq(&rq
->cfs
, rq
);
6488 #ifdef CONFIG_FAIR_GROUP_SCHED
6489 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6491 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6492 struct sched_entity
*se
=
6493 &per_cpu(init_sched_entity
, i
);
6495 init_cfs_rq_p
[i
] = cfs_rq
;
6496 init_cfs_rq(cfs_rq
, rq
);
6497 cfs_rq
->tg
= &init_task_grp
;
6498 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6499 &rq
->leaf_cfs_rq_list
);
6501 init_sched_entity_p
[i
] = se
;
6502 se
->cfs_rq
= &rq
->cfs
;
6504 se
->load
.weight
= NICE_0_LOAD
;
6505 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6508 init_task_grp
.shares
= NICE_0_LOAD
;
6511 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6512 rq
->cpu_load
[j
] = 0;
6515 rq
->active_balance
= 0;
6516 rq
->next_balance
= jiffies
;
6519 rq
->migration_thread
= NULL
;
6520 INIT_LIST_HEAD(&rq
->migration_queue
);
6522 atomic_set(&rq
->nr_iowait
, 0);
6524 array
= &rq
->rt
.active
;
6525 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6526 INIT_LIST_HEAD(array
->queue
+ j
);
6527 __clear_bit(j
, array
->bitmap
);
6530 /* delimiter for bitsearch: */
6531 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6534 set_load_weight(&init_task
);
6536 #ifdef CONFIG_PREEMPT_NOTIFIERS
6537 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6541 nr_cpu_ids
= highest_cpu
+ 1;
6542 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6545 #ifdef CONFIG_RT_MUTEXES
6546 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6550 * The boot idle thread does lazy MMU switching as well:
6552 atomic_inc(&init_mm
.mm_count
);
6553 enter_lazy_tlb(&init_mm
, current
);
6556 * Make us the idle thread. Technically, schedule() should not be
6557 * called from this thread, however somewhere below it might be,
6558 * but because we are the idle thread, we just pick up running again
6559 * when this runqueue becomes "idle".
6561 init_idle(current
, smp_processor_id());
6563 * During early bootup we pretend to be a normal task:
6565 current
->sched_class
= &fair_sched_class
;
6568 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6569 void __might_sleep(char *file
, int line
)
6572 static unsigned long prev_jiffy
; /* ratelimiting */
6574 if ((in_atomic() || irqs_disabled()) &&
6575 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6576 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6578 prev_jiffy
= jiffies
;
6579 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6580 " context at %s:%d\n", file
, line
);
6581 printk("in_atomic():%d, irqs_disabled():%d\n",
6582 in_atomic(), irqs_disabled());
6583 debug_show_held_locks(current
);
6584 if (irqs_disabled())
6585 print_irqtrace_events(current
);
6590 EXPORT_SYMBOL(__might_sleep
);
6593 #ifdef CONFIG_MAGIC_SYSRQ
6594 void normalize_rt_tasks(void)
6596 struct task_struct
*g
, *p
;
6597 unsigned long flags
;
6601 read_lock_irq(&tasklist_lock
);
6602 do_each_thread(g
, p
) {
6603 p
->se
.exec_start
= 0;
6604 #ifdef CONFIG_SCHEDSTATS
6605 p
->se
.wait_start
= 0;
6606 p
->se
.sleep_start
= 0;
6607 p
->se
.block_start
= 0;
6609 task_rq(p
)->clock
= 0;
6613 * Renice negative nice level userspace
6616 if (TASK_NICE(p
) < 0 && p
->mm
)
6617 set_user_nice(p
, 0);
6621 spin_lock_irqsave(&p
->pi_lock
, flags
);
6622 rq
= __task_rq_lock(p
);
6625 * Do not touch the migration thread:
6627 if (p
== rq
->migration_thread
)
6631 update_rq_clock(rq
);
6632 on_rq
= p
->se
.on_rq
;
6634 deactivate_task(rq
, p
, 0);
6635 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6637 activate_task(rq
, p
, 0);
6638 resched_task(rq
->curr
);
6643 __task_rq_unlock(rq
);
6644 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6645 } while_each_thread(g
, p
);
6647 read_unlock_irq(&tasklist_lock
);
6650 #endif /* CONFIG_MAGIC_SYSRQ */
6654 * These functions are only useful for the IA64 MCA handling.
6656 * They can only be called when the whole system has been
6657 * stopped - every CPU needs to be quiescent, and no scheduling
6658 * activity can take place. Using them for anything else would
6659 * be a serious bug, and as a result, they aren't even visible
6660 * under any other configuration.
6664 * curr_task - return the current task for a given cpu.
6665 * @cpu: the processor in question.
6667 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6669 struct task_struct
*curr_task(int cpu
)
6671 return cpu_curr(cpu
);
6675 * set_curr_task - set the current task for a given cpu.
6676 * @cpu: the processor in question.
6677 * @p: the task pointer to set.
6679 * Description: This function must only be used when non-maskable interrupts
6680 * are serviced on a separate stack. It allows the architecture to switch the
6681 * notion of the current task on a cpu in a non-blocking manner. This function
6682 * must be called with all CPU's synchronized, and interrupts disabled, the
6683 * and caller must save the original value of the current task (see
6684 * curr_task() above) and restore that value before reenabling interrupts and
6685 * re-starting the system.
6687 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6689 void set_curr_task(int cpu
, struct task_struct
*p
)
6696 #ifdef CONFIG_FAIR_GROUP_SCHED
6698 /* return corresponding task_grp object of a container */
6699 static inline struct task_grp
*container_tg(struct container
*cont
)
6701 return container_of(container_subsys_state(cont
, cpu_subsys_id
),
6702 struct task_grp
, css
);
6705 /* allocate runqueue etc for a new task group */
6706 static struct container_subsys_state
*
6707 sched_create_group(struct container_subsys
*ss
, struct container
*cont
)
6709 struct task_grp
*tg
;
6710 struct cfs_rq
*cfs_rq
;
6711 struct sched_entity
*se
;
6714 if (!cont
->parent
) {
6715 /* This is early initialization for the top container */
6716 init_task_grp
.css
.container
= cont
;
6717 return &init_task_grp
.css
;
6720 /* we support only 1-level deep hierarchical scheduler atm */
6721 if (cont
->parent
->parent
)
6722 return ERR_PTR(-EINVAL
);
6724 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6726 return ERR_PTR(-ENOMEM
);
6728 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * num_possible_cpus(), GFP_KERNEL
);
6731 tg
->se
= kzalloc(sizeof(se
) * num_possible_cpus(), GFP_KERNEL
);
6735 for_each_possible_cpu(i
) {
6736 struct rq
*rq
= cpu_rq(i
);
6738 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
6743 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
6748 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
6749 memset(se
, 0, sizeof(struct sched_entity
));
6751 tg
->cfs_rq
[i
] = cfs_rq
;
6752 init_cfs_rq(cfs_rq
, rq
);
6754 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6757 se
->cfs_rq
= &rq
->cfs
;
6759 se
->load
.weight
= NICE_0_LOAD
;
6760 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6764 tg
->shares
= NICE_0_LOAD
;
6766 /* Bind the container to task_grp object we just created */
6767 tg
->css
.container
= cont
;
6772 for_each_possible_cpu(i
) {
6773 if (tg
->cfs_rq
&& tg
->cfs_rq
[i
])
6774 kfree(tg
->cfs_rq
[i
]);
6775 if (tg
->se
&& tg
->se
[i
])
6785 return ERR_PTR(-ENOMEM
);
6789 /* destroy runqueue etc associated with a task group */
6790 static void sched_destroy_group(struct container_subsys
*ss
,
6791 struct container
*cont
)
6793 struct task_grp
*tg
= container_tg(cont
);
6794 struct cfs_rq
*cfs_rq
;
6795 struct sched_entity
*se
;
6798 for_each_possible_cpu(i
) {
6799 cfs_rq
= tg
->cfs_rq
[i
];
6800 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
6803 /* wait for possible concurrent references to cfs_rqs complete */
6804 synchronize_sched();
6806 /* now it should be safe to free those cfs_rqs */
6807 for_each_possible_cpu(i
) {
6808 cfs_rq
= tg
->cfs_rq
[i
];
6820 static int sched_can_attach(struct container_subsys
*ss
,
6821 struct container
*cont
, struct task_struct
*tsk
)
6823 /* We don't support RT-tasks being in separate groups */
6824 if (tsk
->sched_class
!= &fair_sched_class
)
6830 /* change task's runqueue when it moves between groups */
6831 static void sched_move_task(struct container_subsys
*ss
, struct container
*cont
,
6832 struct container
*old_cont
, struct task_struct
*tsk
)
6835 unsigned long flags
;
6838 rq
= task_rq_lock(tsk
, &flags
);
6840 if (tsk
->sched_class
!= &fair_sched_class
)
6843 update_rq_clock(rq
);
6845 running
= task_running(rq
, tsk
);
6846 on_rq
= tsk
->se
.on_rq
;
6849 dequeue_task(rq
, tsk
, 0);
6851 set_task_cfs_rq(tsk
);
6854 enqueue_task(rq
, tsk
, 0);
6857 task_rq_unlock(rq
, &flags
);
6860 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
6862 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
6863 struct rq
*rq
= cfs_rq
->rq
;
6866 spin_lock_irq(&rq
->lock
);
6870 dequeue_entity(cfs_rq
, se
, 0);
6872 se
->load
.weight
= shares
;
6873 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
6876 enqueue_entity(cfs_rq
, se
, 0);
6878 spin_unlock_irq(&rq
->lock
);
6881 static ssize_t
cpu_shares_write(struct container
*cont
, struct cftype
*cftype
,
6882 struct file
*file
, const char __user
*userbuf
,
6883 size_t nbytes
, loff_t
*ppos
)
6886 unsigned long shareval
;
6887 struct task_grp
*tg
= container_tg(cont
);
6888 char buffer
[2*sizeof(unsigned long) + 1];
6890 if (nbytes
> 2*sizeof(unsigned long)) /* safety check */
6893 if (copy_from_user(buffer
, userbuf
, nbytes
))
6896 buffer
[nbytes
] = 0; /* nul-terminate */
6897 shareval
= simple_strtoul(buffer
, NULL
, 10);
6899 tg
->shares
= shareval
;
6900 for_each_possible_cpu(i
)
6901 set_se_shares(tg
->se
[i
], shareval
);
6906 static u64
cpu_shares_read_uint(struct container
*cont
, struct cftype
*cft
)
6908 struct task_grp
*tg
= container_tg(cont
);
6910 return (u64
) tg
->shares
;
6913 struct cftype cpuctl_share
= {
6915 .read_uint
= cpu_shares_read_uint
,
6916 .write
= cpu_shares_write
,
6919 static int sched_populate(struct container_subsys
*ss
, struct container
*cont
)
6921 return container_add_file(cont
, ss
, &cpuctl_share
);
6924 struct container_subsys cpu_subsys
= {
6926 .create
= sched_create_group
,
6927 .destroy
= sched_destroy_group
,
6928 .can_attach
= sched_can_attach
,
6929 .attach
= sched_move_task
,
6930 .populate
= sched_populate
,
6931 .subsys_id
= cpu_subsys_id
,
6935 #endif /* CONFIG_FAIR_GROUP_SCHED */