4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak
)) sched_clock(void)
75 return (unsigned long long)jiffies
* (1000000000 / HZ
);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
116 * Since cpu_power is a 'constant', we can use a reciprocal divide.
118 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
120 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
124 * Each time a sched group cpu_power is changed,
125 * we must compute its reciprocal value
127 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
129 sg
->__cpu_power
+= val
;
130 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
134 static inline int rt_policy(int policy
)
136 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
141 static inline int task_has_rt_policy(struct task_struct
*p
)
143 return rt_policy(p
->policy
);
147 * This is the priority-queue data structure of the RT scheduling class:
149 struct rt_prio_array
{
150 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
151 struct list_head queue
[MAX_RT_PRIO
];
154 #ifdef CONFIG_FAIR_GROUP_SCHED
158 /* task group related information */
160 /* schedulable entities of this group on each cpu */
161 struct sched_entity
**se
;
162 /* runqueue "owned" by this group on each cpu */
163 struct cfs_rq
**cfs_rq
;
164 unsigned long shares
;
165 /* spinlock to serialize modification to shares */
169 /* Default task group's sched entity on each cpu */
170 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
171 /* Default task group's cfs_rq on each cpu */
172 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
174 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
175 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
177 /* Default task group.
178 * Every task in system belong to this group at bootup.
180 struct task_group init_task_group
= {
181 .se
= init_sched_entity_p
,
182 .cfs_rq
= init_cfs_rq_p
,
185 #ifdef CONFIG_FAIR_USER_SCHED
186 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
188 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
191 static int init_task_group_load
= INIT_TASK_GRP_LOAD
;
193 /* return group to which a task belongs */
194 static inline struct task_group
*task_group(struct task_struct
*p
)
196 struct task_group
*tg
;
198 #ifdef CONFIG_FAIR_USER_SCHED
201 tg
= &init_task_group
;
207 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
208 static inline void set_task_cfs_rq(struct task_struct
*p
)
210 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[task_cpu(p
)];
211 p
->se
.parent
= task_group(p
)->se
[task_cpu(p
)];
216 static inline void set_task_cfs_rq(struct task_struct
*p
) { }
218 #endif /* CONFIG_FAIR_GROUP_SCHED */
220 /* CFS-related fields in a runqueue */
222 struct load_weight load
;
223 unsigned long nr_running
;
228 struct rb_root tasks_timeline
;
229 struct rb_node
*rb_leftmost
;
230 struct rb_node
*rb_load_balance_curr
;
231 /* 'curr' points to currently running entity on this cfs_rq.
232 * It is set to NULL otherwise (i.e when none are currently running).
234 struct sched_entity
*curr
;
236 unsigned long nr_spread_over
;
238 #ifdef CONFIG_FAIR_GROUP_SCHED
239 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
241 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
242 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
243 * (like users, containers etc.)
245 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
246 * list is used during load balance.
248 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
249 struct task_group
*tg
; /* group that "owns" this runqueue */
254 /* Real-Time classes' related field in a runqueue: */
256 struct rt_prio_array active
;
257 int rt_load_balance_idx
;
258 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
262 * This is the main, per-CPU runqueue data structure.
264 * Locking rule: those places that want to lock multiple runqueues
265 * (such as the load balancing or the thread migration code), lock
266 * acquire operations must be ordered by ascending &runqueue.
269 spinlock_t lock
; /* runqueue lock */
272 * nr_running and cpu_load should be in the same cacheline because
273 * remote CPUs use both these fields when doing load calculation.
275 unsigned long nr_running
;
276 #define CPU_LOAD_IDX_MAX 5
277 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
278 unsigned char idle_at_tick
;
280 unsigned char in_nohz_recently
;
282 struct load_weight load
; /* capture load from *all* tasks on this cpu */
283 unsigned long nr_load_updates
;
287 #ifdef CONFIG_FAIR_GROUP_SCHED
288 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
293 * This is part of a global counter where only the total sum
294 * over all CPUs matters. A task can increase this counter on
295 * one CPU and if it got migrated afterwards it may decrease
296 * it on another CPU. Always updated under the runqueue lock:
298 unsigned long nr_uninterruptible
;
300 struct task_struct
*curr
, *idle
;
301 unsigned long next_balance
;
302 struct mm_struct
*prev_mm
;
304 u64 clock
, prev_clock_raw
;
307 unsigned int clock_warps
, clock_overflows
;
309 unsigned int clock_deep_idle_events
;
315 struct sched_domain
*sd
;
317 /* For active balancing */
320 int cpu
; /* cpu of this runqueue */
322 struct task_struct
*migration_thread
;
323 struct list_head migration_queue
;
326 #ifdef CONFIG_SCHEDSTATS
328 struct sched_info rq_sched_info
;
330 /* sys_sched_yield() stats */
331 unsigned long yld_exp_empty
;
332 unsigned long yld_act_empty
;
333 unsigned long yld_both_empty
;
334 unsigned long yld_count
;
336 /* schedule() stats */
337 unsigned long sched_switch
;
338 unsigned long sched_count
;
339 unsigned long sched_goidle
;
341 /* try_to_wake_up() stats */
342 unsigned long ttwu_count
;
343 unsigned long ttwu_local
;
346 unsigned long bkl_count
;
348 struct lock_class_key rq_lock_key
;
351 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
352 static DEFINE_MUTEX(sched_hotcpu_mutex
);
354 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
356 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
359 static inline int cpu_of(struct rq
*rq
)
369 * Update the per-runqueue clock, as finegrained as the platform can give
370 * us, but without assuming monotonicity, etc.:
372 static void __update_rq_clock(struct rq
*rq
)
374 u64 prev_raw
= rq
->prev_clock_raw
;
375 u64 now
= sched_clock();
376 s64 delta
= now
- prev_raw
;
377 u64 clock
= rq
->clock
;
379 #ifdef CONFIG_SCHED_DEBUG
380 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
383 * Protect against sched_clock() occasionally going backwards:
385 if (unlikely(delta
< 0)) {
390 * Catch too large forward jumps too:
392 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
393 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
394 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
397 rq
->clock_overflows
++;
399 if (unlikely(delta
> rq
->clock_max_delta
))
400 rq
->clock_max_delta
= delta
;
405 rq
->prev_clock_raw
= now
;
409 static void update_rq_clock(struct rq
*rq
)
411 if (likely(smp_processor_id() == cpu_of(rq
)))
412 __update_rq_clock(rq
);
416 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
417 * See detach_destroy_domains: synchronize_sched for details.
419 * The domain tree of any CPU may only be accessed from within
420 * preempt-disabled sections.
422 #define for_each_domain(cpu, __sd) \
423 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
425 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
426 #define this_rq() (&__get_cpu_var(runqueues))
427 #define task_rq(p) cpu_rq(task_cpu(p))
428 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
431 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
433 #ifdef CONFIG_SCHED_DEBUG
434 # define const_debug __read_mostly
436 # define const_debug static const
440 * Debugging: various feature bits
443 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
444 SCHED_FEAT_START_DEBIT
= 2,
445 SCHED_FEAT_TREE_AVG
= 4,
446 SCHED_FEAT_APPROX_AVG
= 8,
447 SCHED_FEAT_WAKEUP_PREEMPT
= 16,
448 SCHED_FEAT_PREEMPT_RESTRICT
= 32,
451 const_debug
unsigned int sysctl_sched_features
=
452 SCHED_FEAT_NEW_FAIR_SLEEPERS
*1 |
453 SCHED_FEAT_START_DEBIT
*1 |
454 SCHED_FEAT_TREE_AVG
*0 |
455 SCHED_FEAT_APPROX_AVG
*0 |
456 SCHED_FEAT_WAKEUP_PREEMPT
*1 |
457 SCHED_FEAT_PREEMPT_RESTRICT
*1;
459 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
462 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
463 * clock constructed from sched_clock():
465 unsigned long long cpu_clock(int cpu
)
467 unsigned long long now
;
471 local_irq_save(flags
);
475 local_irq_restore(flags
);
479 EXPORT_SYMBOL_GPL(cpu_clock
);
481 #ifndef prepare_arch_switch
482 # define prepare_arch_switch(next) do { } while (0)
484 #ifndef finish_arch_switch
485 # define finish_arch_switch(prev) do { } while (0)
488 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
489 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
491 return rq
->curr
== p
;
494 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
498 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
500 #ifdef CONFIG_DEBUG_SPINLOCK
501 /* this is a valid case when another task releases the spinlock */
502 rq
->lock
.owner
= current
;
505 * If we are tracking spinlock dependencies then we have to
506 * fix up the runqueue lock - which gets 'carried over' from
509 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
511 spin_unlock_irq(&rq
->lock
);
514 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
515 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
520 return rq
->curr
== p
;
524 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
528 * We can optimise this out completely for !SMP, because the
529 * SMP rebalancing from interrupt is the only thing that cares
534 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
535 spin_unlock_irq(&rq
->lock
);
537 spin_unlock(&rq
->lock
);
541 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
545 * After ->oncpu is cleared, the task can be moved to a different CPU.
546 * We must ensure this doesn't happen until the switch is completely
552 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
556 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
559 * __task_rq_lock - lock the runqueue a given task resides on.
560 * Must be called interrupts disabled.
562 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
566 struct rq
*rq
= task_rq(p
);
567 spin_lock(&rq
->lock
);
568 if (likely(rq
== task_rq(p
)))
570 spin_unlock(&rq
->lock
);
575 * task_rq_lock - lock the runqueue a given task resides on and disable
576 * interrupts. Note the ordering: we can safely lookup the task_rq without
577 * explicitly disabling preemption.
579 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
585 local_irq_save(*flags
);
587 spin_lock(&rq
->lock
);
588 if (likely(rq
== task_rq(p
)))
590 spin_unlock_irqrestore(&rq
->lock
, *flags
);
594 static void __task_rq_unlock(struct rq
*rq
)
597 spin_unlock(&rq
->lock
);
600 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
603 spin_unlock_irqrestore(&rq
->lock
, *flags
);
607 * this_rq_lock - lock this runqueue and disable interrupts.
609 static struct rq
*this_rq_lock(void)
616 spin_lock(&rq
->lock
);
622 * We are going deep-idle (irqs are disabled):
624 void sched_clock_idle_sleep_event(void)
626 struct rq
*rq
= cpu_rq(smp_processor_id());
628 spin_lock(&rq
->lock
);
629 __update_rq_clock(rq
);
630 spin_unlock(&rq
->lock
);
631 rq
->clock_deep_idle_events
++;
633 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
636 * We just idled delta nanoseconds (called with irqs disabled):
638 void sched_clock_idle_wakeup_event(u64 delta_ns
)
640 struct rq
*rq
= cpu_rq(smp_processor_id());
641 u64 now
= sched_clock();
643 rq
->idle_clock
+= delta_ns
;
645 * Override the previous timestamp and ignore all
646 * sched_clock() deltas that occured while we idled,
647 * and use the PM-provided delta_ns to advance the
650 spin_lock(&rq
->lock
);
651 rq
->prev_clock_raw
= now
;
652 rq
->clock
+= delta_ns
;
653 spin_unlock(&rq
->lock
);
655 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
658 * resched_task - mark a task 'to be rescheduled now'.
660 * On UP this means the setting of the need_resched flag, on SMP it
661 * might also involve a cross-CPU call to trigger the scheduler on
666 #ifndef tsk_is_polling
667 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
670 static void resched_task(struct task_struct
*p
)
674 assert_spin_locked(&task_rq(p
)->lock
);
676 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
679 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
682 if (cpu
== smp_processor_id())
685 /* NEED_RESCHED must be visible before we test polling */
687 if (!tsk_is_polling(p
))
688 smp_send_reschedule(cpu
);
691 static void resched_cpu(int cpu
)
693 struct rq
*rq
= cpu_rq(cpu
);
696 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
698 resched_task(cpu_curr(cpu
));
699 spin_unlock_irqrestore(&rq
->lock
, flags
);
702 static inline void resched_task(struct task_struct
*p
)
704 assert_spin_locked(&task_rq(p
)->lock
);
705 set_tsk_need_resched(p
);
709 #if BITS_PER_LONG == 32
710 # define WMULT_CONST (~0UL)
712 # define WMULT_CONST (1UL << 32)
715 #define WMULT_SHIFT 32
718 * Shift right and round:
720 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
723 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
724 struct load_weight
*lw
)
728 if (unlikely(!lw
->inv_weight
))
729 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
731 tmp
= (u64
)delta_exec
* weight
;
733 * Check whether we'd overflow the 64-bit multiplication:
735 if (unlikely(tmp
> WMULT_CONST
))
736 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
739 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
741 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
744 static inline unsigned long
745 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
747 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
750 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
755 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
761 * To aid in avoiding the subversion of "niceness" due to uneven distribution
762 * of tasks with abnormal "nice" values across CPUs the contribution that
763 * each task makes to its run queue's load is weighted according to its
764 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
765 * scaled version of the new time slice allocation that they receive on time
769 #define WEIGHT_IDLEPRIO 2
770 #define WMULT_IDLEPRIO (1 << 31)
773 * Nice levels are multiplicative, with a gentle 10% change for every
774 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
775 * nice 1, it will get ~10% less CPU time than another CPU-bound task
776 * that remained on nice 0.
778 * The "10% effect" is relative and cumulative: from _any_ nice level,
779 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
780 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
781 * If a task goes up by ~10% and another task goes down by ~10% then
782 * the relative distance between them is ~25%.)
784 static const int prio_to_weight
[40] = {
785 /* -20 */ 88761, 71755, 56483, 46273, 36291,
786 /* -15 */ 29154, 23254, 18705, 14949, 11916,
787 /* -10 */ 9548, 7620, 6100, 4904, 3906,
788 /* -5 */ 3121, 2501, 1991, 1586, 1277,
789 /* 0 */ 1024, 820, 655, 526, 423,
790 /* 5 */ 335, 272, 215, 172, 137,
791 /* 10 */ 110, 87, 70, 56, 45,
792 /* 15 */ 36, 29, 23, 18, 15,
796 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
798 * In cases where the weight does not change often, we can use the
799 * precalculated inverse to speed up arithmetics by turning divisions
800 * into multiplications:
802 static const u32 prio_to_wmult
[40] = {
803 /* -20 */ 48388, 59856, 76040, 92818, 118348,
804 /* -15 */ 147320, 184698, 229616, 287308, 360437,
805 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
806 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
807 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
808 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
809 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
810 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
813 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
816 * runqueue iterator, to support SMP load-balancing between different
817 * scheduling classes, without having to expose their internal data
818 * structures to the load-balancing proper:
822 struct task_struct
*(*start
)(void *);
823 struct task_struct
*(*next
)(void *);
826 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
827 unsigned long max_nr_move
, unsigned long max_load_move
,
828 struct sched_domain
*sd
, enum cpu_idle_type idle
,
829 int *all_pinned
, unsigned long *load_moved
,
830 int *this_best_prio
, struct rq_iterator
*iterator
);
832 #include "sched_stats.h"
833 #include "sched_idletask.c"
834 #include "sched_fair.c"
835 #include "sched_rt.c"
836 #ifdef CONFIG_SCHED_DEBUG
837 # include "sched_debug.c"
840 #define sched_class_highest (&rt_sched_class)
843 * Update delta_exec, delta_fair fields for rq.
845 * delta_fair clock advances at a rate inversely proportional to
846 * total load (rq->load.weight) on the runqueue, while
847 * delta_exec advances at the same rate as wall-clock (provided
850 * delta_exec / delta_fair is a measure of the (smoothened) load on this
851 * runqueue over any given interval. This (smoothened) load is used
852 * during load balance.
854 * This function is called /before/ updating rq->load
855 * and when switching tasks.
857 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
859 update_load_add(&rq
->load
, p
->se
.load
.weight
);
862 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
864 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
867 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
873 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
879 static void set_load_weight(struct task_struct
*p
)
881 if (task_has_rt_policy(p
)) {
882 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
883 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
888 * SCHED_IDLE tasks get minimal weight:
890 if (p
->policy
== SCHED_IDLE
) {
891 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
892 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
896 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
897 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
900 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
902 sched_info_queued(p
);
903 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
907 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
909 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
914 * __normal_prio - return the priority that is based on the static prio
916 static inline int __normal_prio(struct task_struct
*p
)
918 return p
->static_prio
;
922 * Calculate the expected normal priority: i.e. priority
923 * without taking RT-inheritance into account. Might be
924 * boosted by interactivity modifiers. Changes upon fork,
925 * setprio syscalls, and whenever the interactivity
926 * estimator recalculates.
928 static inline int normal_prio(struct task_struct
*p
)
932 if (task_has_rt_policy(p
))
933 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
935 prio
= __normal_prio(p
);
940 * Calculate the current priority, i.e. the priority
941 * taken into account by the scheduler. This value might
942 * be boosted by RT tasks, or might be boosted by
943 * interactivity modifiers. Will be RT if the task got
944 * RT-boosted. If not then it returns p->normal_prio.
946 static int effective_prio(struct task_struct
*p
)
948 p
->normal_prio
= normal_prio(p
);
950 * If we are RT tasks or we were boosted to RT priority,
951 * keep the priority unchanged. Otherwise, update priority
952 * to the normal priority:
954 if (!rt_prio(p
->prio
))
955 return p
->normal_prio
;
960 * activate_task - move a task to the runqueue.
962 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
964 if (p
->state
== TASK_UNINTERRUPTIBLE
)
965 rq
->nr_uninterruptible
--;
967 enqueue_task(rq
, p
, wakeup
);
968 inc_nr_running(p
, rq
);
972 * deactivate_task - remove a task from the runqueue.
974 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
976 if (p
->state
== TASK_UNINTERRUPTIBLE
)
977 rq
->nr_uninterruptible
++;
979 dequeue_task(rq
, p
, sleep
);
980 dec_nr_running(p
, rq
);
984 * task_curr - is this task currently executing on a CPU?
985 * @p: the task in question.
987 inline int task_curr(const struct task_struct
*p
)
989 return cpu_curr(task_cpu(p
)) == p
;
992 /* Used instead of source_load when we know the type == 0 */
993 unsigned long weighted_cpuload(const int cpu
)
995 return cpu_rq(cpu
)->load
.weight
;
998 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1001 task_thread_info(p
)->cpu
= cpu
;
1009 * Is this task likely cache-hot:
1012 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1016 if (p
->sched_class
!= &fair_sched_class
)
1019 delta
= now
- p
->se
.exec_start
;
1021 return delta
< (s64
)sysctl_sched_migration_cost
;
1025 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1027 int old_cpu
= task_cpu(p
);
1028 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1029 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1030 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1033 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1035 #ifdef CONFIG_SCHEDSTATS
1036 if (p
->se
.wait_start
)
1037 p
->se
.wait_start
-= clock_offset
;
1038 if (p
->se
.sleep_start
)
1039 p
->se
.sleep_start
-= clock_offset
;
1040 if (p
->se
.block_start
)
1041 p
->se
.block_start
-= clock_offset
;
1042 if (old_cpu
!= new_cpu
) {
1043 schedstat_inc(p
, se
.nr_migrations
);
1044 if (task_hot(p
, old_rq
->clock
, NULL
))
1045 schedstat_inc(p
, se
.nr_forced2_migrations
);
1048 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1049 new_cfsrq
->min_vruntime
;
1051 __set_task_cpu(p
, new_cpu
);
1054 struct migration_req
{
1055 struct list_head list
;
1057 struct task_struct
*task
;
1060 struct completion done
;
1064 * The task's runqueue lock must be held.
1065 * Returns true if you have to wait for migration thread.
1068 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1070 struct rq
*rq
= task_rq(p
);
1073 * If the task is not on a runqueue (and not running), then
1074 * it is sufficient to simply update the task's cpu field.
1076 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1077 set_task_cpu(p
, dest_cpu
);
1081 init_completion(&req
->done
);
1083 req
->dest_cpu
= dest_cpu
;
1084 list_add(&req
->list
, &rq
->migration_queue
);
1090 * wait_task_inactive - wait for a thread to unschedule.
1092 * The caller must ensure that the task *will* unschedule sometime soon,
1093 * else this function might spin for a *long* time. This function can't
1094 * be called with interrupts off, or it may introduce deadlock with
1095 * smp_call_function() if an IPI is sent by the same process we are
1096 * waiting to become inactive.
1098 void wait_task_inactive(struct task_struct
*p
)
1100 unsigned long flags
;
1106 * We do the initial early heuristics without holding
1107 * any task-queue locks at all. We'll only try to get
1108 * the runqueue lock when things look like they will
1114 * If the task is actively running on another CPU
1115 * still, just relax and busy-wait without holding
1118 * NOTE! Since we don't hold any locks, it's not
1119 * even sure that "rq" stays as the right runqueue!
1120 * But we don't care, since "task_running()" will
1121 * return false if the runqueue has changed and p
1122 * is actually now running somewhere else!
1124 while (task_running(rq
, p
))
1128 * Ok, time to look more closely! We need the rq
1129 * lock now, to be *sure*. If we're wrong, we'll
1130 * just go back and repeat.
1132 rq
= task_rq_lock(p
, &flags
);
1133 running
= task_running(rq
, p
);
1134 on_rq
= p
->se
.on_rq
;
1135 task_rq_unlock(rq
, &flags
);
1138 * Was it really running after all now that we
1139 * checked with the proper locks actually held?
1141 * Oops. Go back and try again..
1143 if (unlikely(running
)) {
1149 * It's not enough that it's not actively running,
1150 * it must be off the runqueue _entirely_, and not
1153 * So if it wa still runnable (but just not actively
1154 * running right now), it's preempted, and we should
1155 * yield - it could be a while.
1157 if (unlikely(on_rq
)) {
1158 schedule_timeout_uninterruptible(1);
1163 * Ahh, all good. It wasn't running, and it wasn't
1164 * runnable, which means that it will never become
1165 * running in the future either. We're all done!
1172 * kick_process - kick a running thread to enter/exit the kernel
1173 * @p: the to-be-kicked thread
1175 * Cause a process which is running on another CPU to enter
1176 * kernel-mode, without any delay. (to get signals handled.)
1178 * NOTE: this function doesnt have to take the runqueue lock,
1179 * because all it wants to ensure is that the remote task enters
1180 * the kernel. If the IPI races and the task has been migrated
1181 * to another CPU then no harm is done and the purpose has been
1184 void kick_process(struct task_struct
*p
)
1190 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1191 smp_send_reschedule(cpu
);
1196 * Return a low guess at the load of a migration-source cpu weighted
1197 * according to the scheduling class and "nice" value.
1199 * We want to under-estimate the load of migration sources, to
1200 * balance conservatively.
1202 static unsigned long source_load(int cpu
, int type
)
1204 struct rq
*rq
= cpu_rq(cpu
);
1205 unsigned long total
= weighted_cpuload(cpu
);
1210 return min(rq
->cpu_load
[type
-1], total
);
1214 * Return a high guess at the load of a migration-target cpu weighted
1215 * according to the scheduling class and "nice" value.
1217 static unsigned long target_load(int cpu
, int type
)
1219 struct rq
*rq
= cpu_rq(cpu
);
1220 unsigned long total
= weighted_cpuload(cpu
);
1225 return max(rq
->cpu_load
[type
-1], total
);
1229 * Return the average load per task on the cpu's run queue
1231 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1233 struct rq
*rq
= cpu_rq(cpu
);
1234 unsigned long total
= weighted_cpuload(cpu
);
1235 unsigned long n
= rq
->nr_running
;
1237 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1241 * find_idlest_group finds and returns the least busy CPU group within the
1244 static struct sched_group
*
1245 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1247 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1248 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1249 int load_idx
= sd
->forkexec_idx
;
1250 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1253 unsigned long load
, avg_load
;
1257 /* Skip over this group if it has no CPUs allowed */
1258 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1261 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1263 /* Tally up the load of all CPUs in the group */
1266 for_each_cpu_mask(i
, group
->cpumask
) {
1267 /* Bias balancing toward cpus of our domain */
1269 load
= source_load(i
, load_idx
);
1271 load
= target_load(i
, load_idx
);
1276 /* Adjust by relative CPU power of the group */
1277 avg_load
= sg_div_cpu_power(group
,
1278 avg_load
* SCHED_LOAD_SCALE
);
1281 this_load
= avg_load
;
1283 } else if (avg_load
< min_load
) {
1284 min_load
= avg_load
;
1287 } while (group
= group
->next
, group
!= sd
->groups
);
1289 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1295 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1298 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1301 unsigned long load
, min_load
= ULONG_MAX
;
1305 /* Traverse only the allowed CPUs */
1306 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1308 for_each_cpu_mask(i
, tmp
) {
1309 load
= weighted_cpuload(i
);
1311 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1321 * sched_balance_self: balance the current task (running on cpu) in domains
1322 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1325 * Balance, ie. select the least loaded group.
1327 * Returns the target CPU number, or the same CPU if no balancing is needed.
1329 * preempt must be disabled.
1331 static int sched_balance_self(int cpu
, int flag
)
1333 struct task_struct
*t
= current
;
1334 struct sched_domain
*tmp
, *sd
= NULL
;
1336 for_each_domain(cpu
, tmp
) {
1338 * If power savings logic is enabled for a domain, stop there.
1340 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1342 if (tmp
->flags
& flag
)
1348 struct sched_group
*group
;
1349 int new_cpu
, weight
;
1351 if (!(sd
->flags
& flag
)) {
1357 group
= find_idlest_group(sd
, t
, cpu
);
1363 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1364 if (new_cpu
== -1 || new_cpu
== cpu
) {
1365 /* Now try balancing at a lower domain level of cpu */
1370 /* Now try balancing at a lower domain level of new_cpu */
1373 weight
= cpus_weight(span
);
1374 for_each_domain(cpu
, tmp
) {
1375 if (weight
<= cpus_weight(tmp
->span
))
1377 if (tmp
->flags
& flag
)
1380 /* while loop will break here if sd == NULL */
1386 #endif /* CONFIG_SMP */
1389 * wake_idle() will wake a task on an idle cpu if task->cpu is
1390 * not idle and an idle cpu is available. The span of cpus to
1391 * search starts with cpus closest then further out as needed,
1392 * so we always favor a closer, idle cpu.
1394 * Returns the CPU we should wake onto.
1396 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1397 static int wake_idle(int cpu
, struct task_struct
*p
)
1400 struct sched_domain
*sd
;
1404 * If it is idle, then it is the best cpu to run this task.
1406 * This cpu is also the best, if it has more than one task already.
1407 * Siblings must be also busy(in most cases) as they didn't already
1408 * pickup the extra load from this cpu and hence we need not check
1409 * sibling runqueue info. This will avoid the checks and cache miss
1410 * penalities associated with that.
1412 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1415 for_each_domain(cpu
, sd
) {
1416 if (sd
->flags
& SD_WAKE_IDLE
) {
1417 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1418 for_each_cpu_mask(i
, tmp
) {
1420 if (i
!= task_cpu(p
)) {
1422 se
.nr_wakeups_idle
);
1434 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1441 * try_to_wake_up - wake up a thread
1442 * @p: the to-be-woken-up thread
1443 * @state: the mask of task states that can be woken
1444 * @sync: do a synchronous wakeup?
1446 * Put it on the run-queue if it's not already there. The "current"
1447 * thread is always on the run-queue (except when the actual
1448 * re-schedule is in progress), and as such you're allowed to do
1449 * the simpler "current->state = TASK_RUNNING" to mark yourself
1450 * runnable without the overhead of this.
1452 * returns failure only if the task is already active.
1454 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1456 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1457 unsigned long flags
;
1461 struct sched_domain
*sd
, *this_sd
= NULL
;
1462 unsigned long load
, this_load
;
1466 rq
= task_rq_lock(p
, &flags
);
1467 old_state
= p
->state
;
1468 if (!(old_state
& state
))
1476 this_cpu
= smp_processor_id();
1479 if (unlikely(task_running(rq
, p
)))
1484 schedstat_inc(rq
, ttwu_count
);
1485 if (cpu
== this_cpu
) {
1486 schedstat_inc(rq
, ttwu_local
);
1490 for_each_domain(this_cpu
, sd
) {
1491 if (cpu_isset(cpu
, sd
->span
)) {
1492 schedstat_inc(sd
, ttwu_wake_remote
);
1498 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1502 * Check for affine wakeup and passive balancing possibilities.
1505 int idx
= this_sd
->wake_idx
;
1506 unsigned int imbalance
;
1508 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1510 load
= source_load(cpu
, idx
);
1511 this_load
= target_load(this_cpu
, idx
);
1513 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1515 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1516 unsigned long tl
= this_load
;
1517 unsigned long tl_per_task
;
1519 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
1520 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1523 * If sync wakeup then subtract the (maximum possible)
1524 * effect of the currently running task from the load
1525 * of the current CPU:
1528 tl
-= current
->se
.load
.weight
;
1531 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1532 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1534 * This domain has SD_WAKE_AFFINE and
1535 * p is cache cold in this domain, and
1536 * there is no bad imbalance.
1538 schedstat_inc(this_sd
, ttwu_move_affine
);
1539 schedstat_inc(p
, se
.nr_wakeups_affine
);
1545 * Start passive balancing when half the imbalance_pct
1548 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1549 if (imbalance
*this_load
<= 100*load
) {
1550 schedstat_inc(this_sd
, ttwu_move_balance
);
1551 schedstat_inc(p
, se
.nr_wakeups_passive
);
1557 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1559 new_cpu
= wake_idle(new_cpu
, p
);
1560 if (new_cpu
!= cpu
) {
1561 set_task_cpu(p
, new_cpu
);
1562 task_rq_unlock(rq
, &flags
);
1563 /* might preempt at this point */
1564 rq
= task_rq_lock(p
, &flags
);
1565 old_state
= p
->state
;
1566 if (!(old_state
& state
))
1571 this_cpu
= smp_processor_id();
1576 #endif /* CONFIG_SMP */
1577 schedstat_inc(p
, se
.nr_wakeups
);
1579 schedstat_inc(p
, se
.nr_wakeups_sync
);
1580 if (orig_cpu
!= cpu
)
1581 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1582 if (cpu
== this_cpu
)
1583 schedstat_inc(p
, se
.nr_wakeups_local
);
1585 schedstat_inc(p
, se
.nr_wakeups_remote
);
1586 update_rq_clock(rq
);
1587 activate_task(rq
, p
, 1);
1589 * Sync wakeups (i.e. those types of wakeups where the waker
1590 * has indicated that it will leave the CPU in short order)
1591 * don't trigger a preemption, if the woken up task will run on
1592 * this cpu. (in this case the 'I will reschedule' promise of
1593 * the waker guarantees that the freshly woken up task is going
1594 * to be considered on this CPU.)
1596 if (!sync
|| cpu
!= this_cpu
)
1597 check_preempt_curr(rq
, p
);
1601 p
->state
= TASK_RUNNING
;
1603 task_rq_unlock(rq
, &flags
);
1608 int fastcall
wake_up_process(struct task_struct
*p
)
1610 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1611 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1613 EXPORT_SYMBOL(wake_up_process
);
1615 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1617 return try_to_wake_up(p
, state
, 0);
1621 * Perform scheduler related setup for a newly forked process p.
1622 * p is forked by current.
1624 * __sched_fork() is basic setup used by init_idle() too:
1626 static void __sched_fork(struct task_struct
*p
)
1628 p
->se
.exec_start
= 0;
1629 p
->se
.sum_exec_runtime
= 0;
1630 p
->se
.prev_sum_exec_runtime
= 0;
1632 #ifdef CONFIG_SCHEDSTATS
1633 p
->se
.wait_start
= 0;
1634 p
->se
.sum_sleep_runtime
= 0;
1635 p
->se
.sleep_start
= 0;
1636 p
->se
.block_start
= 0;
1637 p
->se
.sleep_max
= 0;
1638 p
->se
.block_max
= 0;
1640 p
->se
.slice_max
= 0;
1644 INIT_LIST_HEAD(&p
->run_list
);
1647 #ifdef CONFIG_PREEMPT_NOTIFIERS
1648 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1652 * We mark the process as running here, but have not actually
1653 * inserted it onto the runqueue yet. This guarantees that
1654 * nobody will actually run it, and a signal or other external
1655 * event cannot wake it up and insert it on the runqueue either.
1657 p
->state
= TASK_RUNNING
;
1661 * fork()/clone()-time setup:
1663 void sched_fork(struct task_struct
*p
, int clone_flags
)
1665 int cpu
= get_cpu();
1670 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1672 set_task_cpu(p
, cpu
);
1675 * Make sure we do not leak PI boosting priority to the child:
1677 p
->prio
= current
->normal_prio
;
1678 if (!rt_prio(p
->prio
))
1679 p
->sched_class
= &fair_sched_class
;
1681 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1682 if (likely(sched_info_on()))
1683 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1685 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1688 #ifdef CONFIG_PREEMPT
1689 /* Want to start with kernel preemption disabled. */
1690 task_thread_info(p
)->preempt_count
= 1;
1696 * wake_up_new_task - wake up a newly created task for the first time.
1698 * This function will do some initial scheduler statistics housekeeping
1699 * that must be done for every newly created context, then puts the task
1700 * on the runqueue and wakes it.
1702 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1704 unsigned long flags
;
1707 rq
= task_rq_lock(p
, &flags
);
1708 BUG_ON(p
->state
!= TASK_RUNNING
);
1709 update_rq_clock(rq
);
1711 p
->prio
= effective_prio(p
);
1713 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
|| !rq
->cfs
.curr
) {
1714 activate_task(rq
, p
, 0);
1717 * Let the scheduling class do new task startup
1718 * management (if any):
1720 p
->sched_class
->task_new(rq
, p
);
1721 inc_nr_running(p
, rq
);
1723 check_preempt_curr(rq
, p
);
1724 task_rq_unlock(rq
, &flags
);
1727 #ifdef CONFIG_PREEMPT_NOTIFIERS
1730 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1731 * @notifier: notifier struct to register
1733 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1735 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1737 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1740 * preempt_notifier_unregister - no longer interested in preemption notifications
1741 * @notifier: notifier struct to unregister
1743 * This is safe to call from within a preemption notifier.
1745 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1747 hlist_del(¬ifier
->link
);
1749 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1751 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1753 struct preempt_notifier
*notifier
;
1754 struct hlist_node
*node
;
1756 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1757 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1761 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1762 struct task_struct
*next
)
1764 struct preempt_notifier
*notifier
;
1765 struct hlist_node
*node
;
1767 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1768 notifier
->ops
->sched_out(notifier
, next
);
1773 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1778 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1779 struct task_struct
*next
)
1786 * prepare_task_switch - prepare to switch tasks
1787 * @rq: the runqueue preparing to switch
1788 * @prev: the current task that is being switched out
1789 * @next: the task we are going to switch to.
1791 * This is called with the rq lock held and interrupts off. It must
1792 * be paired with a subsequent finish_task_switch after the context
1795 * prepare_task_switch sets up locking and calls architecture specific
1799 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1800 struct task_struct
*next
)
1802 fire_sched_out_preempt_notifiers(prev
, next
);
1803 prepare_lock_switch(rq
, next
);
1804 prepare_arch_switch(next
);
1808 * finish_task_switch - clean up after a task-switch
1809 * @rq: runqueue associated with task-switch
1810 * @prev: the thread we just switched away from.
1812 * finish_task_switch must be called after the context switch, paired
1813 * with a prepare_task_switch call before the context switch.
1814 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1815 * and do any other architecture-specific cleanup actions.
1817 * Note that we may have delayed dropping an mm in context_switch(). If
1818 * so, we finish that here outside of the runqueue lock. (Doing it
1819 * with the lock held can cause deadlocks; see schedule() for
1822 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1823 __releases(rq
->lock
)
1825 struct mm_struct
*mm
= rq
->prev_mm
;
1831 * A task struct has one reference for the use as "current".
1832 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1833 * schedule one last time. The schedule call will never return, and
1834 * the scheduled task must drop that reference.
1835 * The test for TASK_DEAD must occur while the runqueue locks are
1836 * still held, otherwise prev could be scheduled on another cpu, die
1837 * there before we look at prev->state, and then the reference would
1839 * Manfred Spraul <manfred@colorfullife.com>
1841 prev_state
= prev
->state
;
1842 finish_arch_switch(prev
);
1843 finish_lock_switch(rq
, prev
);
1844 fire_sched_in_preempt_notifiers(current
);
1847 if (unlikely(prev_state
== TASK_DEAD
)) {
1849 * Remove function-return probe instances associated with this
1850 * task and put them back on the free list.
1852 kprobe_flush_task(prev
);
1853 put_task_struct(prev
);
1858 * schedule_tail - first thing a freshly forked thread must call.
1859 * @prev: the thread we just switched away from.
1861 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1862 __releases(rq
->lock
)
1864 struct rq
*rq
= this_rq();
1866 finish_task_switch(rq
, prev
);
1867 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1868 /* In this case, finish_task_switch does not reenable preemption */
1871 if (current
->set_child_tid
)
1872 put_user(current
->pid
, current
->set_child_tid
);
1876 * context_switch - switch to the new MM and the new
1877 * thread's register state.
1880 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1881 struct task_struct
*next
)
1883 struct mm_struct
*mm
, *oldmm
;
1885 prepare_task_switch(rq
, prev
, next
);
1887 oldmm
= prev
->active_mm
;
1889 * For paravirt, this is coupled with an exit in switch_to to
1890 * combine the page table reload and the switch backend into
1893 arch_enter_lazy_cpu_mode();
1895 if (unlikely(!mm
)) {
1896 next
->active_mm
= oldmm
;
1897 atomic_inc(&oldmm
->mm_count
);
1898 enter_lazy_tlb(oldmm
, next
);
1900 switch_mm(oldmm
, mm
, next
);
1902 if (unlikely(!prev
->mm
)) {
1903 prev
->active_mm
= NULL
;
1904 rq
->prev_mm
= oldmm
;
1907 * Since the runqueue lock will be released by the next
1908 * task (which is an invalid locking op but in the case
1909 * of the scheduler it's an obvious special-case), so we
1910 * do an early lockdep release here:
1912 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1913 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1916 /* Here we just switch the register state and the stack. */
1917 switch_to(prev
, next
, prev
);
1921 * this_rq must be evaluated again because prev may have moved
1922 * CPUs since it called schedule(), thus the 'rq' on its stack
1923 * frame will be invalid.
1925 finish_task_switch(this_rq(), prev
);
1929 * nr_running, nr_uninterruptible and nr_context_switches:
1931 * externally visible scheduler statistics: current number of runnable
1932 * threads, current number of uninterruptible-sleeping threads, total
1933 * number of context switches performed since bootup.
1935 unsigned long nr_running(void)
1937 unsigned long i
, sum
= 0;
1939 for_each_online_cpu(i
)
1940 sum
+= cpu_rq(i
)->nr_running
;
1945 unsigned long nr_uninterruptible(void)
1947 unsigned long i
, sum
= 0;
1949 for_each_possible_cpu(i
)
1950 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1953 * Since we read the counters lockless, it might be slightly
1954 * inaccurate. Do not allow it to go below zero though:
1956 if (unlikely((long)sum
< 0))
1962 unsigned long long nr_context_switches(void)
1965 unsigned long long sum
= 0;
1967 for_each_possible_cpu(i
)
1968 sum
+= cpu_rq(i
)->nr_switches
;
1973 unsigned long nr_iowait(void)
1975 unsigned long i
, sum
= 0;
1977 for_each_possible_cpu(i
)
1978 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1983 unsigned long nr_active(void)
1985 unsigned long i
, running
= 0, uninterruptible
= 0;
1987 for_each_online_cpu(i
) {
1988 running
+= cpu_rq(i
)->nr_running
;
1989 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1992 if (unlikely((long)uninterruptible
< 0))
1993 uninterruptible
= 0;
1995 return running
+ uninterruptible
;
1999 * Update rq->cpu_load[] statistics. This function is usually called every
2000 * scheduler tick (TICK_NSEC).
2002 static void update_cpu_load(struct rq
*this_rq
)
2004 unsigned long this_load
= this_rq
->load
.weight
;
2007 this_rq
->nr_load_updates
++;
2009 /* Update our load: */
2010 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2011 unsigned long old_load
, new_load
;
2013 /* scale is effectively 1 << i now, and >> i divides by scale */
2015 old_load
= this_rq
->cpu_load
[i
];
2016 new_load
= this_load
;
2018 * Round up the averaging division if load is increasing. This
2019 * prevents us from getting stuck on 9 if the load is 10, for
2022 if (new_load
> old_load
)
2023 new_load
+= scale
-1;
2024 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2031 * double_rq_lock - safely lock two runqueues
2033 * Note this does not disable interrupts like task_rq_lock,
2034 * you need to do so manually before calling.
2036 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2037 __acquires(rq1
->lock
)
2038 __acquires(rq2
->lock
)
2040 BUG_ON(!irqs_disabled());
2042 spin_lock(&rq1
->lock
);
2043 __acquire(rq2
->lock
); /* Fake it out ;) */
2046 spin_lock(&rq1
->lock
);
2047 spin_lock(&rq2
->lock
);
2049 spin_lock(&rq2
->lock
);
2050 spin_lock(&rq1
->lock
);
2053 update_rq_clock(rq1
);
2054 update_rq_clock(rq2
);
2058 * double_rq_unlock - safely unlock two runqueues
2060 * Note this does not restore interrupts like task_rq_unlock,
2061 * you need to do so manually after calling.
2063 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2064 __releases(rq1
->lock
)
2065 __releases(rq2
->lock
)
2067 spin_unlock(&rq1
->lock
);
2069 spin_unlock(&rq2
->lock
);
2071 __release(rq2
->lock
);
2075 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2077 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2078 __releases(this_rq
->lock
)
2079 __acquires(busiest
->lock
)
2080 __acquires(this_rq
->lock
)
2082 if (unlikely(!irqs_disabled())) {
2083 /* printk() doesn't work good under rq->lock */
2084 spin_unlock(&this_rq
->lock
);
2087 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2088 if (busiest
< this_rq
) {
2089 spin_unlock(&this_rq
->lock
);
2090 spin_lock(&busiest
->lock
);
2091 spin_lock(&this_rq
->lock
);
2093 spin_lock(&busiest
->lock
);
2098 * If dest_cpu is allowed for this process, migrate the task to it.
2099 * This is accomplished by forcing the cpu_allowed mask to only
2100 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2101 * the cpu_allowed mask is restored.
2103 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2105 struct migration_req req
;
2106 unsigned long flags
;
2109 rq
= task_rq_lock(p
, &flags
);
2110 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2111 || unlikely(cpu_is_offline(dest_cpu
)))
2114 /* force the process onto the specified CPU */
2115 if (migrate_task(p
, dest_cpu
, &req
)) {
2116 /* Need to wait for migration thread (might exit: take ref). */
2117 struct task_struct
*mt
= rq
->migration_thread
;
2119 get_task_struct(mt
);
2120 task_rq_unlock(rq
, &flags
);
2121 wake_up_process(mt
);
2122 put_task_struct(mt
);
2123 wait_for_completion(&req
.done
);
2128 task_rq_unlock(rq
, &flags
);
2132 * sched_exec - execve() is a valuable balancing opportunity, because at
2133 * this point the task has the smallest effective memory and cache footprint.
2135 void sched_exec(void)
2137 int new_cpu
, this_cpu
= get_cpu();
2138 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2140 if (new_cpu
!= this_cpu
)
2141 sched_migrate_task(current
, new_cpu
);
2145 * pull_task - move a task from a remote runqueue to the local runqueue.
2146 * Both runqueues must be locked.
2148 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2149 struct rq
*this_rq
, int this_cpu
)
2151 deactivate_task(src_rq
, p
, 0);
2152 set_task_cpu(p
, this_cpu
);
2153 activate_task(this_rq
, p
, 0);
2155 * Note that idle threads have a prio of MAX_PRIO, for this test
2156 * to be always true for them.
2158 check_preempt_curr(this_rq
, p
);
2162 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2165 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2166 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2170 * We do not migrate tasks that are:
2171 * 1) running (obviously), or
2172 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2173 * 3) are cache-hot on their current CPU.
2175 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2176 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2181 if (task_running(rq
, p
)) {
2182 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2187 * Aggressive migration if:
2188 * 1) task is cache cold, or
2189 * 2) too many balance attempts have failed.
2192 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2193 #ifdef CONFIG_SCHEDSTATS
2194 if (task_hot(p
, rq
->clock
, sd
)) {
2195 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2196 schedstat_inc(p
, se
.nr_forced_migrations
);
2202 if (task_hot(p
, rq
->clock
, sd
)) {
2203 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2209 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2210 unsigned long max_nr_move
, unsigned long max_load_move
,
2211 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2212 int *all_pinned
, unsigned long *load_moved
,
2213 int *this_best_prio
, struct rq_iterator
*iterator
)
2215 int pulled
= 0, pinned
= 0, skip_for_load
;
2216 struct task_struct
*p
;
2217 long rem_load_move
= max_load_move
;
2219 if (max_nr_move
== 0 || max_load_move
== 0)
2225 * Start the load-balancing iterator:
2227 p
= iterator
->start(iterator
->arg
);
2232 * To help distribute high priority tasks accross CPUs we don't
2233 * skip a task if it will be the highest priority task (i.e. smallest
2234 * prio value) on its new queue regardless of its load weight
2236 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2237 SCHED_LOAD_SCALE_FUZZ
;
2238 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2239 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2240 p
= iterator
->next(iterator
->arg
);
2244 pull_task(busiest
, p
, this_rq
, this_cpu
);
2246 rem_load_move
-= p
->se
.load
.weight
;
2249 * We only want to steal up to the prescribed number of tasks
2250 * and the prescribed amount of weighted load.
2252 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2253 if (p
->prio
< *this_best_prio
)
2254 *this_best_prio
= p
->prio
;
2255 p
= iterator
->next(iterator
->arg
);
2260 * Right now, this is the only place pull_task() is called,
2261 * so we can safely collect pull_task() stats here rather than
2262 * inside pull_task().
2264 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2267 *all_pinned
= pinned
;
2268 *load_moved
= max_load_move
- rem_load_move
;
2273 * move_tasks tries to move up to max_load_move weighted load from busiest to
2274 * this_rq, as part of a balancing operation within domain "sd".
2275 * Returns 1 if successful and 0 otherwise.
2277 * Called with both runqueues locked.
2279 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2280 unsigned long max_load_move
,
2281 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2284 const struct sched_class
*class = sched_class_highest
;
2285 unsigned long total_load_moved
= 0;
2286 int this_best_prio
= this_rq
->curr
->prio
;
2290 class->load_balance(this_rq
, this_cpu
, busiest
,
2291 ULONG_MAX
, max_load_move
- total_load_moved
,
2292 sd
, idle
, all_pinned
, &this_best_prio
);
2293 class = class->next
;
2294 } while (class && max_load_move
> total_load_moved
);
2296 return total_load_moved
> 0;
2300 * move_one_task tries to move exactly one task from busiest to this_rq, as
2301 * part of active balancing operations within "domain".
2302 * Returns 1 if successful and 0 otherwise.
2304 * Called with both runqueues locked.
2306 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2307 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2309 const struct sched_class
*class;
2310 int this_best_prio
= MAX_PRIO
;
2312 for (class = sched_class_highest
; class; class = class->next
)
2313 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2314 1, ULONG_MAX
, sd
, idle
, NULL
,
2322 * find_busiest_group finds and returns the busiest CPU group within the
2323 * domain. It calculates and returns the amount of weighted load which
2324 * should be moved to restore balance via the imbalance parameter.
2326 static struct sched_group
*
2327 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2328 unsigned long *imbalance
, enum cpu_idle_type idle
,
2329 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2331 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2332 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2333 unsigned long max_pull
;
2334 unsigned long busiest_load_per_task
, busiest_nr_running
;
2335 unsigned long this_load_per_task
, this_nr_running
;
2337 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2338 int power_savings_balance
= 1;
2339 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2340 unsigned long min_nr_running
= ULONG_MAX
;
2341 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2344 max_load
= this_load
= total_load
= total_pwr
= 0;
2345 busiest_load_per_task
= busiest_nr_running
= 0;
2346 this_load_per_task
= this_nr_running
= 0;
2347 if (idle
== CPU_NOT_IDLE
)
2348 load_idx
= sd
->busy_idx
;
2349 else if (idle
== CPU_NEWLY_IDLE
)
2350 load_idx
= sd
->newidle_idx
;
2352 load_idx
= sd
->idle_idx
;
2355 unsigned long load
, group_capacity
;
2358 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2359 unsigned long sum_nr_running
, sum_weighted_load
;
2361 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2364 balance_cpu
= first_cpu(group
->cpumask
);
2366 /* Tally up the load of all CPUs in the group */
2367 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2369 for_each_cpu_mask(i
, group
->cpumask
) {
2372 if (!cpu_isset(i
, *cpus
))
2377 if (*sd_idle
&& rq
->nr_running
)
2380 /* Bias balancing toward cpus of our domain */
2382 if (idle_cpu(i
) && !first_idle_cpu
) {
2387 load
= target_load(i
, load_idx
);
2389 load
= source_load(i
, load_idx
);
2392 sum_nr_running
+= rq
->nr_running
;
2393 sum_weighted_load
+= weighted_cpuload(i
);
2397 * First idle cpu or the first cpu(busiest) in this sched group
2398 * is eligible for doing load balancing at this and above
2399 * domains. In the newly idle case, we will allow all the cpu's
2400 * to do the newly idle load balance.
2402 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2403 balance_cpu
!= this_cpu
&& balance
) {
2408 total_load
+= avg_load
;
2409 total_pwr
+= group
->__cpu_power
;
2411 /* Adjust by relative CPU power of the group */
2412 avg_load
= sg_div_cpu_power(group
,
2413 avg_load
* SCHED_LOAD_SCALE
);
2415 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2418 this_load
= avg_load
;
2420 this_nr_running
= sum_nr_running
;
2421 this_load_per_task
= sum_weighted_load
;
2422 } else if (avg_load
> max_load
&&
2423 sum_nr_running
> group_capacity
) {
2424 max_load
= avg_load
;
2426 busiest_nr_running
= sum_nr_running
;
2427 busiest_load_per_task
= sum_weighted_load
;
2430 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2432 * Busy processors will not participate in power savings
2435 if (idle
== CPU_NOT_IDLE
||
2436 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2440 * If the local group is idle or completely loaded
2441 * no need to do power savings balance at this domain
2443 if (local_group
&& (this_nr_running
>= group_capacity
||
2445 power_savings_balance
= 0;
2448 * If a group is already running at full capacity or idle,
2449 * don't include that group in power savings calculations
2451 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2456 * Calculate the group which has the least non-idle load.
2457 * This is the group from where we need to pick up the load
2460 if ((sum_nr_running
< min_nr_running
) ||
2461 (sum_nr_running
== min_nr_running
&&
2462 first_cpu(group
->cpumask
) <
2463 first_cpu(group_min
->cpumask
))) {
2465 min_nr_running
= sum_nr_running
;
2466 min_load_per_task
= sum_weighted_load
/
2471 * Calculate the group which is almost near its
2472 * capacity but still has some space to pick up some load
2473 * from other group and save more power
2475 if (sum_nr_running
<= group_capacity
- 1) {
2476 if (sum_nr_running
> leader_nr_running
||
2477 (sum_nr_running
== leader_nr_running
&&
2478 first_cpu(group
->cpumask
) >
2479 first_cpu(group_leader
->cpumask
))) {
2480 group_leader
= group
;
2481 leader_nr_running
= sum_nr_running
;
2486 group
= group
->next
;
2487 } while (group
!= sd
->groups
);
2489 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2492 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2494 if (this_load
>= avg_load
||
2495 100*max_load
<= sd
->imbalance_pct
*this_load
)
2498 busiest_load_per_task
/= busiest_nr_running
;
2500 * We're trying to get all the cpus to the average_load, so we don't
2501 * want to push ourselves above the average load, nor do we wish to
2502 * reduce the max loaded cpu below the average load, as either of these
2503 * actions would just result in more rebalancing later, and ping-pong
2504 * tasks around. Thus we look for the minimum possible imbalance.
2505 * Negative imbalances (*we* are more loaded than anyone else) will
2506 * be counted as no imbalance for these purposes -- we can't fix that
2507 * by pulling tasks to us. Be careful of negative numbers as they'll
2508 * appear as very large values with unsigned longs.
2510 if (max_load
<= busiest_load_per_task
)
2514 * In the presence of smp nice balancing, certain scenarios can have
2515 * max load less than avg load(as we skip the groups at or below
2516 * its cpu_power, while calculating max_load..)
2518 if (max_load
< avg_load
) {
2520 goto small_imbalance
;
2523 /* Don't want to pull so many tasks that a group would go idle */
2524 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2526 /* How much load to actually move to equalise the imbalance */
2527 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2528 (avg_load
- this_load
) * this->__cpu_power
)
2532 * if *imbalance is less than the average load per runnable task
2533 * there is no gaurantee that any tasks will be moved so we'll have
2534 * a think about bumping its value to force at least one task to be
2537 if (*imbalance
< busiest_load_per_task
) {
2538 unsigned long tmp
, pwr_now
, pwr_move
;
2542 pwr_move
= pwr_now
= 0;
2544 if (this_nr_running
) {
2545 this_load_per_task
/= this_nr_running
;
2546 if (busiest_load_per_task
> this_load_per_task
)
2549 this_load_per_task
= SCHED_LOAD_SCALE
;
2551 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2552 busiest_load_per_task
* imbn
) {
2553 *imbalance
= busiest_load_per_task
;
2558 * OK, we don't have enough imbalance to justify moving tasks,
2559 * however we may be able to increase total CPU power used by
2563 pwr_now
+= busiest
->__cpu_power
*
2564 min(busiest_load_per_task
, max_load
);
2565 pwr_now
+= this->__cpu_power
*
2566 min(this_load_per_task
, this_load
);
2567 pwr_now
/= SCHED_LOAD_SCALE
;
2569 /* Amount of load we'd subtract */
2570 tmp
= sg_div_cpu_power(busiest
,
2571 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2573 pwr_move
+= busiest
->__cpu_power
*
2574 min(busiest_load_per_task
, max_load
- tmp
);
2576 /* Amount of load we'd add */
2577 if (max_load
* busiest
->__cpu_power
<
2578 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2579 tmp
= sg_div_cpu_power(this,
2580 max_load
* busiest
->__cpu_power
);
2582 tmp
= sg_div_cpu_power(this,
2583 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2584 pwr_move
+= this->__cpu_power
*
2585 min(this_load_per_task
, this_load
+ tmp
);
2586 pwr_move
/= SCHED_LOAD_SCALE
;
2588 /* Move if we gain throughput */
2589 if (pwr_move
> pwr_now
)
2590 *imbalance
= busiest_load_per_task
;
2596 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2597 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2600 if (this == group_leader
&& group_leader
!= group_min
) {
2601 *imbalance
= min_load_per_task
;
2611 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2614 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2615 unsigned long imbalance
, cpumask_t
*cpus
)
2617 struct rq
*busiest
= NULL
, *rq
;
2618 unsigned long max_load
= 0;
2621 for_each_cpu_mask(i
, group
->cpumask
) {
2624 if (!cpu_isset(i
, *cpus
))
2628 wl
= weighted_cpuload(i
);
2630 if (rq
->nr_running
== 1 && wl
> imbalance
)
2633 if (wl
> max_load
) {
2643 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2644 * so long as it is large enough.
2646 #define MAX_PINNED_INTERVAL 512
2649 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2650 * tasks if there is an imbalance.
2652 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2653 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2656 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2657 struct sched_group
*group
;
2658 unsigned long imbalance
;
2660 cpumask_t cpus
= CPU_MASK_ALL
;
2661 unsigned long flags
;
2664 * When power savings policy is enabled for the parent domain, idle
2665 * sibling can pick up load irrespective of busy siblings. In this case,
2666 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2667 * portraying it as CPU_NOT_IDLE.
2669 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2670 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2673 schedstat_inc(sd
, lb_count
[idle
]);
2676 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2683 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2687 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2689 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2693 BUG_ON(busiest
== this_rq
);
2695 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2698 if (busiest
->nr_running
> 1) {
2700 * Attempt to move tasks. If find_busiest_group has found
2701 * an imbalance but busiest->nr_running <= 1, the group is
2702 * still unbalanced. ld_moved simply stays zero, so it is
2703 * correctly treated as an imbalance.
2705 local_irq_save(flags
);
2706 double_rq_lock(this_rq
, busiest
);
2707 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2708 imbalance
, sd
, idle
, &all_pinned
);
2709 double_rq_unlock(this_rq
, busiest
);
2710 local_irq_restore(flags
);
2713 * some other cpu did the load balance for us.
2715 if (ld_moved
&& this_cpu
!= smp_processor_id())
2716 resched_cpu(this_cpu
);
2718 /* All tasks on this runqueue were pinned by CPU affinity */
2719 if (unlikely(all_pinned
)) {
2720 cpu_clear(cpu_of(busiest
), cpus
);
2721 if (!cpus_empty(cpus
))
2728 schedstat_inc(sd
, lb_failed
[idle
]);
2729 sd
->nr_balance_failed
++;
2731 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2733 spin_lock_irqsave(&busiest
->lock
, flags
);
2735 /* don't kick the migration_thread, if the curr
2736 * task on busiest cpu can't be moved to this_cpu
2738 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2739 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2741 goto out_one_pinned
;
2744 if (!busiest
->active_balance
) {
2745 busiest
->active_balance
= 1;
2746 busiest
->push_cpu
= this_cpu
;
2749 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2751 wake_up_process(busiest
->migration_thread
);
2754 * We've kicked active balancing, reset the failure
2757 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2760 sd
->nr_balance_failed
= 0;
2762 if (likely(!active_balance
)) {
2763 /* We were unbalanced, so reset the balancing interval */
2764 sd
->balance_interval
= sd
->min_interval
;
2767 * If we've begun active balancing, start to back off. This
2768 * case may not be covered by the all_pinned logic if there
2769 * is only 1 task on the busy runqueue (because we don't call
2772 if (sd
->balance_interval
< sd
->max_interval
)
2773 sd
->balance_interval
*= 2;
2776 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2777 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2782 schedstat_inc(sd
, lb_balanced
[idle
]);
2784 sd
->nr_balance_failed
= 0;
2787 /* tune up the balancing interval */
2788 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2789 (sd
->balance_interval
< sd
->max_interval
))
2790 sd
->balance_interval
*= 2;
2792 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2793 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2799 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2800 * tasks if there is an imbalance.
2802 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2803 * this_rq is locked.
2806 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2808 struct sched_group
*group
;
2809 struct rq
*busiest
= NULL
;
2810 unsigned long imbalance
;
2814 cpumask_t cpus
= CPU_MASK_ALL
;
2817 * When power savings policy is enabled for the parent domain, idle
2818 * sibling can pick up load irrespective of busy siblings. In this case,
2819 * let the state of idle sibling percolate up as IDLE, instead of
2820 * portraying it as CPU_NOT_IDLE.
2822 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2823 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2826 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2828 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2829 &sd_idle
, &cpus
, NULL
);
2831 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2835 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2838 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2842 BUG_ON(busiest
== this_rq
);
2844 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2847 if (busiest
->nr_running
> 1) {
2848 /* Attempt to move tasks */
2849 double_lock_balance(this_rq
, busiest
);
2850 /* this_rq->clock is already updated */
2851 update_rq_clock(busiest
);
2852 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2853 imbalance
, sd
, CPU_NEWLY_IDLE
,
2855 spin_unlock(&busiest
->lock
);
2857 if (unlikely(all_pinned
)) {
2858 cpu_clear(cpu_of(busiest
), cpus
);
2859 if (!cpus_empty(cpus
))
2865 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2866 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2867 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2870 sd
->nr_balance_failed
= 0;
2875 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2876 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2877 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2879 sd
->nr_balance_failed
= 0;
2885 * idle_balance is called by schedule() if this_cpu is about to become
2886 * idle. Attempts to pull tasks from other CPUs.
2888 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2890 struct sched_domain
*sd
;
2891 int pulled_task
= -1;
2892 unsigned long next_balance
= jiffies
+ HZ
;
2894 for_each_domain(this_cpu
, sd
) {
2895 unsigned long interval
;
2897 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2900 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2901 /* If we've pulled tasks over stop searching: */
2902 pulled_task
= load_balance_newidle(this_cpu
,
2905 interval
= msecs_to_jiffies(sd
->balance_interval
);
2906 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2907 next_balance
= sd
->last_balance
+ interval
;
2911 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2913 * We are going idle. next_balance may be set based on
2914 * a busy processor. So reset next_balance.
2916 this_rq
->next_balance
= next_balance
;
2921 * active_load_balance is run by migration threads. It pushes running tasks
2922 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2923 * running on each physical CPU where possible, and avoids physical /
2924 * logical imbalances.
2926 * Called with busiest_rq locked.
2928 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2930 int target_cpu
= busiest_rq
->push_cpu
;
2931 struct sched_domain
*sd
;
2932 struct rq
*target_rq
;
2934 /* Is there any task to move? */
2935 if (busiest_rq
->nr_running
<= 1)
2938 target_rq
= cpu_rq(target_cpu
);
2941 * This condition is "impossible", if it occurs
2942 * we need to fix it. Originally reported by
2943 * Bjorn Helgaas on a 128-cpu setup.
2945 BUG_ON(busiest_rq
== target_rq
);
2947 /* move a task from busiest_rq to target_rq */
2948 double_lock_balance(busiest_rq
, target_rq
);
2949 update_rq_clock(busiest_rq
);
2950 update_rq_clock(target_rq
);
2952 /* Search for an sd spanning us and the target CPU. */
2953 for_each_domain(target_cpu
, sd
) {
2954 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2955 cpu_isset(busiest_cpu
, sd
->span
))
2960 schedstat_inc(sd
, alb_count
);
2962 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2964 schedstat_inc(sd
, alb_pushed
);
2966 schedstat_inc(sd
, alb_failed
);
2968 spin_unlock(&target_rq
->lock
);
2973 atomic_t load_balancer
;
2975 } nohz ____cacheline_aligned
= {
2976 .load_balancer
= ATOMIC_INIT(-1),
2977 .cpu_mask
= CPU_MASK_NONE
,
2981 * This routine will try to nominate the ilb (idle load balancing)
2982 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2983 * load balancing on behalf of all those cpus. If all the cpus in the system
2984 * go into this tickless mode, then there will be no ilb owner (as there is
2985 * no need for one) and all the cpus will sleep till the next wakeup event
2988 * For the ilb owner, tick is not stopped. And this tick will be used
2989 * for idle load balancing. ilb owner will still be part of
2992 * While stopping the tick, this cpu will become the ilb owner if there
2993 * is no other owner. And will be the owner till that cpu becomes busy
2994 * or if all cpus in the system stop their ticks at which point
2995 * there is no need for ilb owner.
2997 * When the ilb owner becomes busy, it nominates another owner, during the
2998 * next busy scheduler_tick()
3000 int select_nohz_load_balancer(int stop_tick
)
3002 int cpu
= smp_processor_id();
3005 cpu_set(cpu
, nohz
.cpu_mask
);
3006 cpu_rq(cpu
)->in_nohz_recently
= 1;
3009 * If we are going offline and still the leader, give up!
3011 if (cpu_is_offline(cpu
) &&
3012 atomic_read(&nohz
.load_balancer
) == cpu
) {
3013 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3018 /* time for ilb owner also to sleep */
3019 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3020 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3021 atomic_set(&nohz
.load_balancer
, -1);
3025 if (atomic_read(&nohz
.load_balancer
) == -1) {
3026 /* make me the ilb owner */
3027 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3029 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3032 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3035 cpu_clear(cpu
, nohz
.cpu_mask
);
3037 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3038 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3045 static DEFINE_SPINLOCK(balancing
);
3048 * It checks each scheduling domain to see if it is due to be balanced,
3049 * and initiates a balancing operation if so.
3051 * Balancing parameters are set up in arch_init_sched_domains.
3053 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3056 struct rq
*rq
= cpu_rq(cpu
);
3057 unsigned long interval
;
3058 struct sched_domain
*sd
;
3059 /* Earliest time when we have to do rebalance again */
3060 unsigned long next_balance
= jiffies
+ 60*HZ
;
3061 int update_next_balance
= 0;
3063 for_each_domain(cpu
, sd
) {
3064 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3067 interval
= sd
->balance_interval
;
3068 if (idle
!= CPU_IDLE
)
3069 interval
*= sd
->busy_factor
;
3071 /* scale ms to jiffies */
3072 interval
= msecs_to_jiffies(interval
);
3073 if (unlikely(!interval
))
3075 if (interval
> HZ
*NR_CPUS
/10)
3076 interval
= HZ
*NR_CPUS
/10;
3079 if (sd
->flags
& SD_SERIALIZE
) {
3080 if (!spin_trylock(&balancing
))
3084 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3085 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3087 * We've pulled tasks over so either we're no
3088 * longer idle, or one of our SMT siblings is
3091 idle
= CPU_NOT_IDLE
;
3093 sd
->last_balance
= jiffies
;
3095 if (sd
->flags
& SD_SERIALIZE
)
3096 spin_unlock(&balancing
);
3098 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3099 next_balance
= sd
->last_balance
+ interval
;
3100 update_next_balance
= 1;
3104 * Stop the load balance at this level. There is another
3105 * CPU in our sched group which is doing load balancing more
3113 * next_balance will be updated only when there is a need.
3114 * When the cpu is attached to null domain for ex, it will not be
3117 if (likely(update_next_balance
))
3118 rq
->next_balance
= next_balance
;
3122 * run_rebalance_domains is triggered when needed from the scheduler tick.
3123 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3124 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3126 static void run_rebalance_domains(struct softirq_action
*h
)
3128 int this_cpu
= smp_processor_id();
3129 struct rq
*this_rq
= cpu_rq(this_cpu
);
3130 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3131 CPU_IDLE
: CPU_NOT_IDLE
;
3133 rebalance_domains(this_cpu
, idle
);
3137 * If this cpu is the owner for idle load balancing, then do the
3138 * balancing on behalf of the other idle cpus whose ticks are
3141 if (this_rq
->idle_at_tick
&&
3142 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3143 cpumask_t cpus
= nohz
.cpu_mask
;
3147 cpu_clear(this_cpu
, cpus
);
3148 for_each_cpu_mask(balance_cpu
, cpus
) {
3150 * If this cpu gets work to do, stop the load balancing
3151 * work being done for other cpus. Next load
3152 * balancing owner will pick it up.
3157 rebalance_domains(balance_cpu
, CPU_IDLE
);
3159 rq
= cpu_rq(balance_cpu
);
3160 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3161 this_rq
->next_balance
= rq
->next_balance
;
3168 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3170 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3171 * idle load balancing owner or decide to stop the periodic load balancing,
3172 * if the whole system is idle.
3174 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3178 * If we were in the nohz mode recently and busy at the current
3179 * scheduler tick, then check if we need to nominate new idle
3182 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3183 rq
->in_nohz_recently
= 0;
3185 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3186 cpu_clear(cpu
, nohz
.cpu_mask
);
3187 atomic_set(&nohz
.load_balancer
, -1);
3190 if (atomic_read(&nohz
.load_balancer
) == -1) {
3192 * simple selection for now: Nominate the
3193 * first cpu in the nohz list to be the next
3196 * TBD: Traverse the sched domains and nominate
3197 * the nearest cpu in the nohz.cpu_mask.
3199 int ilb
= first_cpu(nohz
.cpu_mask
);
3207 * If this cpu is idle and doing idle load balancing for all the
3208 * cpus with ticks stopped, is it time for that to stop?
3210 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3211 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3217 * If this cpu is idle and the idle load balancing is done by
3218 * someone else, then no need raise the SCHED_SOFTIRQ
3220 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3221 cpu_isset(cpu
, nohz
.cpu_mask
))
3224 if (time_after_eq(jiffies
, rq
->next_balance
))
3225 raise_softirq(SCHED_SOFTIRQ
);
3228 #else /* CONFIG_SMP */
3231 * on UP we do not need to balance between CPUs:
3233 static inline void idle_balance(int cpu
, struct rq
*rq
)
3237 /* Avoid "used but not defined" warning on UP */
3238 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3239 unsigned long max_nr_move
, unsigned long max_load_move
,
3240 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3241 int *all_pinned
, unsigned long *load_moved
,
3242 int *this_best_prio
, struct rq_iterator
*iterator
)
3251 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3253 EXPORT_PER_CPU_SYMBOL(kstat
);
3256 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3257 * that have not yet been banked in case the task is currently running.
3259 unsigned long long task_sched_runtime(struct task_struct
*p
)
3261 unsigned long flags
;
3265 rq
= task_rq_lock(p
, &flags
);
3266 ns
= p
->se
.sum_exec_runtime
;
3267 if (rq
->curr
== p
) {
3268 update_rq_clock(rq
);
3269 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3270 if ((s64
)delta_exec
> 0)
3273 task_rq_unlock(rq
, &flags
);
3279 * Account user cpu time to a process.
3280 * @p: the process that the cpu time gets accounted to
3281 * @hardirq_offset: the offset to subtract from hardirq_count()
3282 * @cputime: the cpu time spent in user space since the last update
3284 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3286 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3289 p
->utime
= cputime_add(p
->utime
, cputime
);
3291 /* Add user time to cpustat. */
3292 tmp
= cputime_to_cputime64(cputime
);
3293 if (TASK_NICE(p
) > 0)
3294 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3296 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3300 * Account system cpu time to a process.
3301 * @p: the process that the cpu time gets accounted to
3302 * @hardirq_offset: the offset to subtract from hardirq_count()
3303 * @cputime: the cpu time spent in kernel space since the last update
3305 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3308 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3309 struct rq
*rq
= this_rq();
3312 p
->stime
= cputime_add(p
->stime
, cputime
);
3314 /* Add system time to cpustat. */
3315 tmp
= cputime_to_cputime64(cputime
);
3316 if (hardirq_count() - hardirq_offset
)
3317 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3318 else if (softirq_count())
3319 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3320 else if (p
!= rq
->idle
)
3321 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3322 else if (atomic_read(&rq
->nr_iowait
) > 0)
3323 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3325 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3326 /* Account for system time used */
3327 acct_update_integrals(p
);
3331 * Account for involuntary wait time.
3332 * @p: the process from which the cpu time has been stolen
3333 * @steal: the cpu time spent in involuntary wait
3335 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3337 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3338 cputime64_t tmp
= cputime_to_cputime64(steal
);
3339 struct rq
*rq
= this_rq();
3341 if (p
== rq
->idle
) {
3342 p
->stime
= cputime_add(p
->stime
, steal
);
3343 if (atomic_read(&rq
->nr_iowait
) > 0)
3344 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3346 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3348 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3352 * This function gets called by the timer code, with HZ frequency.
3353 * We call it with interrupts disabled.
3355 * It also gets called by the fork code, when changing the parent's
3358 void scheduler_tick(void)
3360 int cpu
= smp_processor_id();
3361 struct rq
*rq
= cpu_rq(cpu
);
3362 struct task_struct
*curr
= rq
->curr
;
3363 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3365 spin_lock(&rq
->lock
);
3366 __update_rq_clock(rq
);
3368 * Let rq->clock advance by at least TICK_NSEC:
3370 if (unlikely(rq
->clock
< next_tick
))
3371 rq
->clock
= next_tick
;
3372 rq
->tick_timestamp
= rq
->clock
;
3373 update_cpu_load(rq
);
3374 if (curr
!= rq
->idle
) /* FIXME: needed? */
3375 curr
->sched_class
->task_tick(rq
, curr
);
3376 spin_unlock(&rq
->lock
);
3379 rq
->idle_at_tick
= idle_cpu(cpu
);
3380 trigger_load_balance(rq
, cpu
);
3384 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3386 void fastcall
add_preempt_count(int val
)
3391 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3393 preempt_count() += val
;
3395 * Spinlock count overflowing soon?
3397 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3400 EXPORT_SYMBOL(add_preempt_count
);
3402 void fastcall
sub_preempt_count(int val
)
3407 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3410 * Is the spinlock portion underflowing?
3412 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3413 !(preempt_count() & PREEMPT_MASK
)))
3416 preempt_count() -= val
;
3418 EXPORT_SYMBOL(sub_preempt_count
);
3423 * Print scheduling while atomic bug:
3425 static noinline
void __schedule_bug(struct task_struct
*prev
)
3427 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3428 prev
->comm
, preempt_count(), prev
->pid
);
3429 debug_show_held_locks(prev
);
3430 if (irqs_disabled())
3431 print_irqtrace_events(prev
);
3436 * Various schedule()-time debugging checks and statistics:
3438 static inline void schedule_debug(struct task_struct
*prev
)
3441 * Test if we are atomic. Since do_exit() needs to call into
3442 * schedule() atomically, we ignore that path for now.
3443 * Otherwise, whine if we are scheduling when we should not be.
3445 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3446 __schedule_bug(prev
);
3448 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3450 schedstat_inc(this_rq(), sched_count
);
3451 #ifdef CONFIG_SCHEDSTATS
3452 if (unlikely(prev
->lock_depth
>= 0)) {
3453 schedstat_inc(this_rq(), bkl_count
);
3454 schedstat_inc(prev
, sched_info
.bkl_count
);
3460 * Pick up the highest-prio task:
3462 static inline struct task_struct
*
3463 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3465 const struct sched_class
*class;
3466 struct task_struct
*p
;
3469 * Optimization: we know that if all tasks are in
3470 * the fair class we can call that function directly:
3472 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3473 p
= fair_sched_class
.pick_next_task(rq
);
3478 class = sched_class_highest
;
3480 p
= class->pick_next_task(rq
);
3484 * Will never be NULL as the idle class always
3485 * returns a non-NULL p:
3487 class = class->next
;
3492 * schedule() is the main scheduler function.
3494 asmlinkage
void __sched
schedule(void)
3496 struct task_struct
*prev
, *next
;
3503 cpu
= smp_processor_id();
3507 switch_count
= &prev
->nivcsw
;
3509 release_kernel_lock(prev
);
3510 need_resched_nonpreemptible
:
3512 schedule_debug(prev
);
3515 * Do the rq-clock update outside the rq lock:
3517 local_irq_disable();
3518 __update_rq_clock(rq
);
3519 spin_lock(&rq
->lock
);
3520 clear_tsk_need_resched(prev
);
3522 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3523 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3524 unlikely(signal_pending(prev
)))) {
3525 prev
->state
= TASK_RUNNING
;
3527 deactivate_task(rq
, prev
, 1);
3529 switch_count
= &prev
->nvcsw
;
3532 if (unlikely(!rq
->nr_running
))
3533 idle_balance(cpu
, rq
);
3535 prev
->sched_class
->put_prev_task(rq
, prev
);
3536 next
= pick_next_task(rq
, prev
);
3538 sched_info_switch(prev
, next
);
3540 if (likely(prev
!= next
)) {
3545 context_switch(rq
, prev
, next
); /* unlocks the rq */
3547 spin_unlock_irq(&rq
->lock
);
3549 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3550 cpu
= smp_processor_id();
3552 goto need_resched_nonpreemptible
;
3554 preempt_enable_no_resched();
3555 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3558 EXPORT_SYMBOL(schedule
);
3560 #ifdef CONFIG_PREEMPT
3562 * this is the entry point to schedule() from in-kernel preemption
3563 * off of preempt_enable. Kernel preemptions off return from interrupt
3564 * occur there and call schedule directly.
3566 asmlinkage
void __sched
preempt_schedule(void)
3568 struct thread_info
*ti
= current_thread_info();
3569 #ifdef CONFIG_PREEMPT_BKL
3570 struct task_struct
*task
= current
;
3571 int saved_lock_depth
;
3574 * If there is a non-zero preempt_count or interrupts are disabled,
3575 * we do not want to preempt the current task. Just return..
3577 if (likely(ti
->preempt_count
|| irqs_disabled()))
3581 add_preempt_count(PREEMPT_ACTIVE
);
3584 * We keep the big kernel semaphore locked, but we
3585 * clear ->lock_depth so that schedule() doesnt
3586 * auto-release the semaphore:
3588 #ifdef CONFIG_PREEMPT_BKL
3589 saved_lock_depth
= task
->lock_depth
;
3590 task
->lock_depth
= -1;
3593 #ifdef CONFIG_PREEMPT_BKL
3594 task
->lock_depth
= saved_lock_depth
;
3596 sub_preempt_count(PREEMPT_ACTIVE
);
3599 * Check again in case we missed a preemption opportunity
3600 * between schedule and now.
3603 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3605 EXPORT_SYMBOL(preempt_schedule
);
3608 * this is the entry point to schedule() from kernel preemption
3609 * off of irq context.
3610 * Note, that this is called and return with irqs disabled. This will
3611 * protect us against recursive calling from irq.
3613 asmlinkage
void __sched
preempt_schedule_irq(void)
3615 struct thread_info
*ti
= current_thread_info();
3616 #ifdef CONFIG_PREEMPT_BKL
3617 struct task_struct
*task
= current
;
3618 int saved_lock_depth
;
3620 /* Catch callers which need to be fixed */
3621 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3624 add_preempt_count(PREEMPT_ACTIVE
);
3627 * We keep the big kernel semaphore locked, but we
3628 * clear ->lock_depth so that schedule() doesnt
3629 * auto-release the semaphore:
3631 #ifdef CONFIG_PREEMPT_BKL
3632 saved_lock_depth
= task
->lock_depth
;
3633 task
->lock_depth
= -1;
3637 local_irq_disable();
3638 #ifdef CONFIG_PREEMPT_BKL
3639 task
->lock_depth
= saved_lock_depth
;
3641 sub_preempt_count(PREEMPT_ACTIVE
);
3644 * Check again in case we missed a preemption opportunity
3645 * between schedule and now.
3648 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3651 #endif /* CONFIG_PREEMPT */
3653 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3656 return try_to_wake_up(curr
->private, mode
, sync
);
3658 EXPORT_SYMBOL(default_wake_function
);
3661 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3662 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3663 * number) then we wake all the non-exclusive tasks and one exclusive task.
3665 * There are circumstances in which we can try to wake a task which has already
3666 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3667 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3669 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3670 int nr_exclusive
, int sync
, void *key
)
3672 wait_queue_t
*curr
, *next
;
3674 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3675 unsigned flags
= curr
->flags
;
3677 if (curr
->func(curr
, mode
, sync
, key
) &&
3678 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3684 * __wake_up - wake up threads blocked on a waitqueue.
3686 * @mode: which threads
3687 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3688 * @key: is directly passed to the wakeup function
3690 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3691 int nr_exclusive
, void *key
)
3693 unsigned long flags
;
3695 spin_lock_irqsave(&q
->lock
, flags
);
3696 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3697 spin_unlock_irqrestore(&q
->lock
, flags
);
3699 EXPORT_SYMBOL(__wake_up
);
3702 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3704 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3706 __wake_up_common(q
, mode
, 1, 0, NULL
);
3710 * __wake_up_sync - wake up threads blocked on a waitqueue.
3712 * @mode: which threads
3713 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3715 * The sync wakeup differs that the waker knows that it will schedule
3716 * away soon, so while the target thread will be woken up, it will not
3717 * be migrated to another CPU - ie. the two threads are 'synchronized'
3718 * with each other. This can prevent needless bouncing between CPUs.
3720 * On UP it can prevent extra preemption.
3723 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3725 unsigned long flags
;
3731 if (unlikely(!nr_exclusive
))
3734 spin_lock_irqsave(&q
->lock
, flags
);
3735 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3736 spin_unlock_irqrestore(&q
->lock
, flags
);
3738 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3740 void fastcall
complete(struct completion
*x
)
3742 unsigned long flags
;
3744 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3746 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3748 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3750 EXPORT_SYMBOL(complete
);
3752 void fastcall
complete_all(struct completion
*x
)
3754 unsigned long flags
;
3756 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3757 x
->done
+= UINT_MAX
/2;
3758 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3760 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3762 EXPORT_SYMBOL(complete_all
);
3764 static inline long __sched
3765 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3768 DECLARE_WAITQUEUE(wait
, current
);
3770 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3771 __add_wait_queue_tail(&x
->wait
, &wait
);
3773 if (state
== TASK_INTERRUPTIBLE
&&
3774 signal_pending(current
)) {
3775 __remove_wait_queue(&x
->wait
, &wait
);
3776 return -ERESTARTSYS
;
3778 __set_current_state(state
);
3779 spin_unlock_irq(&x
->wait
.lock
);
3780 timeout
= schedule_timeout(timeout
);
3781 spin_lock_irq(&x
->wait
.lock
);
3783 __remove_wait_queue(&x
->wait
, &wait
);
3787 __remove_wait_queue(&x
->wait
, &wait
);
3794 wait_for_common(struct completion
*x
, long timeout
, int state
)
3798 spin_lock_irq(&x
->wait
.lock
);
3799 timeout
= do_wait_for_common(x
, timeout
, state
);
3800 spin_unlock_irq(&x
->wait
.lock
);
3804 void fastcall __sched
wait_for_completion(struct completion
*x
)
3806 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3808 EXPORT_SYMBOL(wait_for_completion
);
3810 unsigned long fastcall __sched
3811 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3813 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3815 EXPORT_SYMBOL(wait_for_completion_timeout
);
3817 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3819 return wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3821 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3823 unsigned long fastcall __sched
3824 wait_for_completion_interruptible_timeout(struct completion
*x
,
3825 unsigned long timeout
)
3827 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3829 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3832 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3834 unsigned long flags
;
3837 init_waitqueue_entry(&wait
, current
);
3839 __set_current_state(state
);
3841 spin_lock_irqsave(&q
->lock
, flags
);
3842 __add_wait_queue(q
, &wait
);
3843 spin_unlock(&q
->lock
);
3844 timeout
= schedule_timeout(timeout
);
3845 spin_lock_irq(&q
->lock
);
3846 __remove_wait_queue(q
, &wait
);
3847 spin_unlock_irqrestore(&q
->lock
, flags
);
3852 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3854 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3856 EXPORT_SYMBOL(interruptible_sleep_on
);
3859 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3861 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3863 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3865 void __sched
sleep_on(wait_queue_head_t
*q
)
3867 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3869 EXPORT_SYMBOL(sleep_on
);
3871 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3873 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3875 EXPORT_SYMBOL(sleep_on_timeout
);
3877 #ifdef CONFIG_RT_MUTEXES
3880 * rt_mutex_setprio - set the current priority of a task
3882 * @prio: prio value (kernel-internal form)
3884 * This function changes the 'effective' priority of a task. It does
3885 * not touch ->normal_prio like __setscheduler().
3887 * Used by the rt_mutex code to implement priority inheritance logic.
3889 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3891 unsigned long flags
;
3892 int oldprio
, on_rq
, running
;
3895 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3897 rq
= task_rq_lock(p
, &flags
);
3898 update_rq_clock(rq
);
3901 on_rq
= p
->se
.on_rq
;
3902 running
= task_running(rq
, p
);
3904 dequeue_task(rq
, p
, 0);
3906 p
->sched_class
->put_prev_task(rq
, p
);
3910 p
->sched_class
= &rt_sched_class
;
3912 p
->sched_class
= &fair_sched_class
;
3918 p
->sched_class
->set_curr_task(rq
);
3919 enqueue_task(rq
, p
, 0);
3921 * Reschedule if we are currently running on this runqueue and
3922 * our priority decreased, or if we are not currently running on
3923 * this runqueue and our priority is higher than the current's
3926 if (p
->prio
> oldprio
)
3927 resched_task(rq
->curr
);
3929 check_preempt_curr(rq
, p
);
3932 task_rq_unlock(rq
, &flags
);
3937 void set_user_nice(struct task_struct
*p
, long nice
)
3939 int old_prio
, delta
, on_rq
;
3940 unsigned long flags
;
3943 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3946 * We have to be careful, if called from sys_setpriority(),
3947 * the task might be in the middle of scheduling on another CPU.
3949 rq
= task_rq_lock(p
, &flags
);
3950 update_rq_clock(rq
);
3952 * The RT priorities are set via sched_setscheduler(), but we still
3953 * allow the 'normal' nice value to be set - but as expected
3954 * it wont have any effect on scheduling until the task is
3955 * SCHED_FIFO/SCHED_RR:
3957 if (task_has_rt_policy(p
)) {
3958 p
->static_prio
= NICE_TO_PRIO(nice
);
3961 on_rq
= p
->se
.on_rq
;
3963 dequeue_task(rq
, p
, 0);
3967 p
->static_prio
= NICE_TO_PRIO(nice
);
3970 p
->prio
= effective_prio(p
);
3971 delta
= p
->prio
- old_prio
;
3974 enqueue_task(rq
, p
, 0);
3977 * If the task increased its priority or is running and
3978 * lowered its priority, then reschedule its CPU:
3980 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3981 resched_task(rq
->curr
);
3984 task_rq_unlock(rq
, &flags
);
3986 EXPORT_SYMBOL(set_user_nice
);
3989 * can_nice - check if a task can reduce its nice value
3993 int can_nice(const struct task_struct
*p
, const int nice
)
3995 /* convert nice value [19,-20] to rlimit style value [1,40] */
3996 int nice_rlim
= 20 - nice
;
3998 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3999 capable(CAP_SYS_NICE
));
4002 #ifdef __ARCH_WANT_SYS_NICE
4005 * sys_nice - change the priority of the current process.
4006 * @increment: priority increment
4008 * sys_setpriority is a more generic, but much slower function that
4009 * does similar things.
4011 asmlinkage
long sys_nice(int increment
)
4016 * Setpriority might change our priority at the same moment.
4017 * We don't have to worry. Conceptually one call occurs first
4018 * and we have a single winner.
4020 if (increment
< -40)
4025 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4031 if (increment
< 0 && !can_nice(current
, nice
))
4034 retval
= security_task_setnice(current
, nice
);
4038 set_user_nice(current
, nice
);
4045 * task_prio - return the priority value of a given task.
4046 * @p: the task in question.
4048 * This is the priority value as seen by users in /proc.
4049 * RT tasks are offset by -200. Normal tasks are centered
4050 * around 0, value goes from -16 to +15.
4052 int task_prio(const struct task_struct
*p
)
4054 return p
->prio
- MAX_RT_PRIO
;
4058 * task_nice - return the nice value of a given task.
4059 * @p: the task in question.
4061 int task_nice(const struct task_struct
*p
)
4063 return TASK_NICE(p
);
4065 EXPORT_SYMBOL_GPL(task_nice
);
4068 * idle_cpu - is a given cpu idle currently?
4069 * @cpu: the processor in question.
4071 int idle_cpu(int cpu
)
4073 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4077 * idle_task - return the idle task for a given cpu.
4078 * @cpu: the processor in question.
4080 struct task_struct
*idle_task(int cpu
)
4082 return cpu_rq(cpu
)->idle
;
4086 * find_process_by_pid - find a process with a matching PID value.
4087 * @pid: the pid in question.
4089 static struct task_struct
*find_process_by_pid(pid_t pid
)
4091 return pid
? find_task_by_pid(pid
) : current
;
4094 /* Actually do priority change: must hold rq lock. */
4096 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4098 BUG_ON(p
->se
.on_rq
);
4101 switch (p
->policy
) {
4105 p
->sched_class
= &fair_sched_class
;
4109 p
->sched_class
= &rt_sched_class
;
4113 p
->rt_priority
= prio
;
4114 p
->normal_prio
= normal_prio(p
);
4115 /* we are holding p->pi_lock already */
4116 p
->prio
= rt_mutex_getprio(p
);
4121 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4122 * @p: the task in question.
4123 * @policy: new policy.
4124 * @param: structure containing the new RT priority.
4126 * NOTE that the task may be already dead.
4128 int sched_setscheduler(struct task_struct
*p
, int policy
,
4129 struct sched_param
*param
)
4131 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4132 unsigned long flags
;
4135 /* may grab non-irq protected spin_locks */
4136 BUG_ON(in_interrupt());
4138 /* double check policy once rq lock held */
4140 policy
= oldpolicy
= p
->policy
;
4141 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4142 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4143 policy
!= SCHED_IDLE
)
4146 * Valid priorities for SCHED_FIFO and SCHED_RR are
4147 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4148 * SCHED_BATCH and SCHED_IDLE is 0.
4150 if (param
->sched_priority
< 0 ||
4151 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4152 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4154 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4158 * Allow unprivileged RT tasks to decrease priority:
4160 if (!capable(CAP_SYS_NICE
)) {
4161 if (rt_policy(policy
)) {
4162 unsigned long rlim_rtprio
;
4164 if (!lock_task_sighand(p
, &flags
))
4166 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4167 unlock_task_sighand(p
, &flags
);
4169 /* can't set/change the rt policy */
4170 if (policy
!= p
->policy
&& !rlim_rtprio
)
4173 /* can't increase priority */
4174 if (param
->sched_priority
> p
->rt_priority
&&
4175 param
->sched_priority
> rlim_rtprio
)
4179 * Like positive nice levels, dont allow tasks to
4180 * move out of SCHED_IDLE either:
4182 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4185 /* can't change other user's priorities */
4186 if ((current
->euid
!= p
->euid
) &&
4187 (current
->euid
!= p
->uid
))
4191 retval
= security_task_setscheduler(p
, policy
, param
);
4195 * make sure no PI-waiters arrive (or leave) while we are
4196 * changing the priority of the task:
4198 spin_lock_irqsave(&p
->pi_lock
, flags
);
4200 * To be able to change p->policy safely, the apropriate
4201 * runqueue lock must be held.
4203 rq
= __task_rq_lock(p
);
4204 /* recheck policy now with rq lock held */
4205 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4206 policy
= oldpolicy
= -1;
4207 __task_rq_unlock(rq
);
4208 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4211 update_rq_clock(rq
);
4212 on_rq
= p
->se
.on_rq
;
4213 running
= task_running(rq
, p
);
4215 deactivate_task(rq
, p
, 0);
4217 p
->sched_class
->put_prev_task(rq
, p
);
4221 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4225 p
->sched_class
->set_curr_task(rq
);
4226 activate_task(rq
, p
, 0);
4228 * Reschedule if we are currently running on this runqueue and
4229 * our priority decreased, or if we are not currently running on
4230 * this runqueue and our priority is higher than the current's
4233 if (p
->prio
> oldprio
)
4234 resched_task(rq
->curr
);
4236 check_preempt_curr(rq
, p
);
4239 __task_rq_unlock(rq
);
4240 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4242 rt_mutex_adjust_pi(p
);
4246 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4249 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4251 struct sched_param lparam
;
4252 struct task_struct
*p
;
4255 if (!param
|| pid
< 0)
4257 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4262 p
= find_process_by_pid(pid
);
4264 retval
= sched_setscheduler(p
, policy
, &lparam
);
4271 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4272 * @pid: the pid in question.
4273 * @policy: new policy.
4274 * @param: structure containing the new RT priority.
4276 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4277 struct sched_param __user
*param
)
4279 /* negative values for policy are not valid */
4283 return do_sched_setscheduler(pid
, policy
, param
);
4287 * sys_sched_setparam - set/change the RT priority of a thread
4288 * @pid: the pid in question.
4289 * @param: structure containing the new RT priority.
4291 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4293 return do_sched_setscheduler(pid
, -1, param
);
4297 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4298 * @pid: the pid in question.
4300 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4302 struct task_struct
*p
;
4309 read_lock(&tasklist_lock
);
4310 p
= find_process_by_pid(pid
);
4312 retval
= security_task_getscheduler(p
);
4316 read_unlock(&tasklist_lock
);
4321 * sys_sched_getscheduler - get the RT priority of a thread
4322 * @pid: the pid in question.
4323 * @param: structure containing the RT priority.
4325 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4327 struct sched_param lp
;
4328 struct task_struct
*p
;
4331 if (!param
|| pid
< 0)
4334 read_lock(&tasklist_lock
);
4335 p
= find_process_by_pid(pid
);
4340 retval
= security_task_getscheduler(p
);
4344 lp
.sched_priority
= p
->rt_priority
;
4345 read_unlock(&tasklist_lock
);
4348 * This one might sleep, we cannot do it with a spinlock held ...
4350 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4355 read_unlock(&tasklist_lock
);
4359 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4361 cpumask_t cpus_allowed
;
4362 struct task_struct
*p
;
4365 mutex_lock(&sched_hotcpu_mutex
);
4366 read_lock(&tasklist_lock
);
4368 p
= find_process_by_pid(pid
);
4370 read_unlock(&tasklist_lock
);
4371 mutex_unlock(&sched_hotcpu_mutex
);
4376 * It is not safe to call set_cpus_allowed with the
4377 * tasklist_lock held. We will bump the task_struct's
4378 * usage count and then drop tasklist_lock.
4381 read_unlock(&tasklist_lock
);
4384 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4385 !capable(CAP_SYS_NICE
))
4388 retval
= security_task_setscheduler(p
, 0, NULL
);
4392 cpus_allowed
= cpuset_cpus_allowed(p
);
4393 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4394 retval
= set_cpus_allowed(p
, new_mask
);
4398 mutex_unlock(&sched_hotcpu_mutex
);
4402 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4403 cpumask_t
*new_mask
)
4405 if (len
< sizeof(cpumask_t
)) {
4406 memset(new_mask
, 0, sizeof(cpumask_t
));
4407 } else if (len
> sizeof(cpumask_t
)) {
4408 len
= sizeof(cpumask_t
);
4410 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4414 * sys_sched_setaffinity - set the cpu affinity of a process
4415 * @pid: pid of the process
4416 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4417 * @user_mask_ptr: user-space pointer to the new cpu mask
4419 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4420 unsigned long __user
*user_mask_ptr
)
4425 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4429 return sched_setaffinity(pid
, new_mask
);
4433 * Represents all cpu's present in the system
4434 * In systems capable of hotplug, this map could dynamically grow
4435 * as new cpu's are detected in the system via any platform specific
4436 * method, such as ACPI for e.g.
4439 cpumask_t cpu_present_map __read_mostly
;
4440 EXPORT_SYMBOL(cpu_present_map
);
4443 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4444 EXPORT_SYMBOL(cpu_online_map
);
4446 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4447 EXPORT_SYMBOL(cpu_possible_map
);
4450 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4452 struct task_struct
*p
;
4455 mutex_lock(&sched_hotcpu_mutex
);
4456 read_lock(&tasklist_lock
);
4459 p
= find_process_by_pid(pid
);
4463 retval
= security_task_getscheduler(p
);
4467 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4470 read_unlock(&tasklist_lock
);
4471 mutex_unlock(&sched_hotcpu_mutex
);
4477 * sys_sched_getaffinity - get the cpu affinity of a process
4478 * @pid: pid of the process
4479 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4480 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4482 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4483 unsigned long __user
*user_mask_ptr
)
4488 if (len
< sizeof(cpumask_t
))
4491 ret
= sched_getaffinity(pid
, &mask
);
4495 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4498 return sizeof(cpumask_t
);
4502 * sys_sched_yield - yield the current processor to other threads.
4504 * This function yields the current CPU to other tasks. If there are no
4505 * other threads running on this CPU then this function will return.
4507 asmlinkage
long sys_sched_yield(void)
4509 struct rq
*rq
= this_rq_lock();
4511 schedstat_inc(rq
, yld_count
);
4512 current
->sched_class
->yield_task(rq
);
4515 * Since we are going to call schedule() anyway, there's
4516 * no need to preempt or enable interrupts:
4518 __release(rq
->lock
);
4519 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4520 _raw_spin_unlock(&rq
->lock
);
4521 preempt_enable_no_resched();
4528 static void __cond_resched(void)
4530 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4531 __might_sleep(__FILE__
, __LINE__
);
4534 * The BKS might be reacquired before we have dropped
4535 * PREEMPT_ACTIVE, which could trigger a second
4536 * cond_resched() call.
4539 add_preempt_count(PREEMPT_ACTIVE
);
4541 sub_preempt_count(PREEMPT_ACTIVE
);
4542 } while (need_resched());
4545 int __sched
cond_resched(void)
4547 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4548 system_state
== SYSTEM_RUNNING
) {
4554 EXPORT_SYMBOL(cond_resched
);
4557 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4558 * call schedule, and on return reacquire the lock.
4560 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4561 * operations here to prevent schedule() from being called twice (once via
4562 * spin_unlock(), once by hand).
4564 int cond_resched_lock(spinlock_t
*lock
)
4568 if (need_lockbreak(lock
)) {
4574 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4575 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4576 _raw_spin_unlock(lock
);
4577 preempt_enable_no_resched();
4584 EXPORT_SYMBOL(cond_resched_lock
);
4586 int __sched
cond_resched_softirq(void)
4588 BUG_ON(!in_softirq());
4590 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4598 EXPORT_SYMBOL(cond_resched_softirq
);
4601 * yield - yield the current processor to other threads.
4603 * This is a shortcut for kernel-space yielding - it marks the
4604 * thread runnable and calls sys_sched_yield().
4606 void __sched
yield(void)
4608 set_current_state(TASK_RUNNING
);
4611 EXPORT_SYMBOL(yield
);
4614 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4615 * that process accounting knows that this is a task in IO wait state.
4617 * But don't do that if it is a deliberate, throttling IO wait (this task
4618 * has set its backing_dev_info: the queue against which it should throttle)
4620 void __sched
io_schedule(void)
4622 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4624 delayacct_blkio_start();
4625 atomic_inc(&rq
->nr_iowait
);
4627 atomic_dec(&rq
->nr_iowait
);
4628 delayacct_blkio_end();
4630 EXPORT_SYMBOL(io_schedule
);
4632 long __sched
io_schedule_timeout(long timeout
)
4634 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4637 delayacct_blkio_start();
4638 atomic_inc(&rq
->nr_iowait
);
4639 ret
= schedule_timeout(timeout
);
4640 atomic_dec(&rq
->nr_iowait
);
4641 delayacct_blkio_end();
4646 * sys_sched_get_priority_max - return maximum RT priority.
4647 * @policy: scheduling class.
4649 * this syscall returns the maximum rt_priority that can be used
4650 * by a given scheduling class.
4652 asmlinkage
long sys_sched_get_priority_max(int policy
)
4659 ret
= MAX_USER_RT_PRIO
-1;
4671 * sys_sched_get_priority_min - return minimum RT priority.
4672 * @policy: scheduling class.
4674 * this syscall returns the minimum rt_priority that can be used
4675 * by a given scheduling class.
4677 asmlinkage
long sys_sched_get_priority_min(int policy
)
4695 * sys_sched_rr_get_interval - return the default timeslice of a process.
4696 * @pid: pid of the process.
4697 * @interval: userspace pointer to the timeslice value.
4699 * this syscall writes the default timeslice value of a given process
4700 * into the user-space timespec buffer. A value of '0' means infinity.
4703 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4705 struct task_struct
*p
;
4706 unsigned int time_slice
;
4714 read_lock(&tasklist_lock
);
4715 p
= find_process_by_pid(pid
);
4719 retval
= security_task_getscheduler(p
);
4723 if (p
->policy
== SCHED_FIFO
)
4725 else if (p
->policy
== SCHED_RR
)
4726 time_slice
= DEF_TIMESLICE
;
4728 struct sched_entity
*se
= &p
->se
;
4729 unsigned long flags
;
4732 rq
= task_rq_lock(p
, &flags
);
4733 time_slice
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
4734 task_rq_unlock(rq
, &flags
);
4736 read_unlock(&tasklist_lock
);
4737 jiffies_to_timespec(time_slice
, &t
);
4738 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4742 read_unlock(&tasklist_lock
);
4746 static const char stat_nam
[] = "RSDTtZX";
4748 static void show_task(struct task_struct
*p
)
4750 unsigned long free
= 0;
4753 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4754 printk("%-13.13s %c", p
->comm
,
4755 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4756 #if BITS_PER_LONG == 32
4757 if (state
== TASK_RUNNING
)
4758 printk(" running ");
4760 printk(" %08lx ", thread_saved_pc(p
));
4762 if (state
== TASK_RUNNING
)
4763 printk(" running task ");
4765 printk(" %016lx ", thread_saved_pc(p
));
4767 #ifdef CONFIG_DEBUG_STACK_USAGE
4769 unsigned long *n
= end_of_stack(p
);
4772 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4775 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4777 if (state
!= TASK_RUNNING
)
4778 show_stack(p
, NULL
);
4781 void show_state_filter(unsigned long state_filter
)
4783 struct task_struct
*g
, *p
;
4785 #if BITS_PER_LONG == 32
4787 " task PC stack pid father\n");
4790 " task PC stack pid father\n");
4792 read_lock(&tasklist_lock
);
4793 do_each_thread(g
, p
) {
4795 * reset the NMI-timeout, listing all files on a slow
4796 * console might take alot of time:
4798 touch_nmi_watchdog();
4799 if (!state_filter
|| (p
->state
& state_filter
))
4801 } while_each_thread(g
, p
);
4803 touch_all_softlockup_watchdogs();
4805 #ifdef CONFIG_SCHED_DEBUG
4806 sysrq_sched_debug_show();
4808 read_unlock(&tasklist_lock
);
4810 * Only show locks if all tasks are dumped:
4812 if (state_filter
== -1)
4813 debug_show_all_locks();
4816 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4818 idle
->sched_class
= &idle_sched_class
;
4822 * init_idle - set up an idle thread for a given CPU
4823 * @idle: task in question
4824 * @cpu: cpu the idle task belongs to
4826 * NOTE: this function does not set the idle thread's NEED_RESCHED
4827 * flag, to make booting more robust.
4829 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4831 struct rq
*rq
= cpu_rq(cpu
);
4832 unsigned long flags
;
4835 idle
->se
.exec_start
= sched_clock();
4837 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4838 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4839 __set_task_cpu(idle
, cpu
);
4841 spin_lock_irqsave(&rq
->lock
, flags
);
4842 rq
->curr
= rq
->idle
= idle
;
4843 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4846 spin_unlock_irqrestore(&rq
->lock
, flags
);
4848 /* Set the preempt count _outside_ the spinlocks! */
4849 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4850 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4852 task_thread_info(idle
)->preempt_count
= 0;
4855 * The idle tasks have their own, simple scheduling class:
4857 idle
->sched_class
= &idle_sched_class
;
4861 * In a system that switches off the HZ timer nohz_cpu_mask
4862 * indicates which cpus entered this state. This is used
4863 * in the rcu update to wait only for active cpus. For system
4864 * which do not switch off the HZ timer nohz_cpu_mask should
4865 * always be CPU_MASK_NONE.
4867 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4871 * This is how migration works:
4873 * 1) we queue a struct migration_req structure in the source CPU's
4874 * runqueue and wake up that CPU's migration thread.
4875 * 2) we down() the locked semaphore => thread blocks.
4876 * 3) migration thread wakes up (implicitly it forces the migrated
4877 * thread off the CPU)
4878 * 4) it gets the migration request and checks whether the migrated
4879 * task is still in the wrong runqueue.
4880 * 5) if it's in the wrong runqueue then the migration thread removes
4881 * it and puts it into the right queue.
4882 * 6) migration thread up()s the semaphore.
4883 * 7) we wake up and the migration is done.
4887 * Change a given task's CPU affinity. Migrate the thread to a
4888 * proper CPU and schedule it away if the CPU it's executing on
4889 * is removed from the allowed bitmask.
4891 * NOTE: the caller must have a valid reference to the task, the
4892 * task must not exit() & deallocate itself prematurely. The
4893 * call is not atomic; no spinlocks may be held.
4895 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4897 struct migration_req req
;
4898 unsigned long flags
;
4902 rq
= task_rq_lock(p
, &flags
);
4903 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4908 p
->cpus_allowed
= new_mask
;
4909 /* Can the task run on the task's current CPU? If so, we're done */
4910 if (cpu_isset(task_cpu(p
), new_mask
))
4913 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4914 /* Need help from migration thread: drop lock and wait. */
4915 task_rq_unlock(rq
, &flags
);
4916 wake_up_process(rq
->migration_thread
);
4917 wait_for_completion(&req
.done
);
4918 tlb_migrate_finish(p
->mm
);
4922 task_rq_unlock(rq
, &flags
);
4926 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4929 * Move (not current) task off this cpu, onto dest cpu. We're doing
4930 * this because either it can't run here any more (set_cpus_allowed()
4931 * away from this CPU, or CPU going down), or because we're
4932 * attempting to rebalance this task on exec (sched_exec).
4934 * So we race with normal scheduler movements, but that's OK, as long
4935 * as the task is no longer on this CPU.
4937 * Returns non-zero if task was successfully migrated.
4939 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4941 struct rq
*rq_dest
, *rq_src
;
4944 if (unlikely(cpu_is_offline(dest_cpu
)))
4947 rq_src
= cpu_rq(src_cpu
);
4948 rq_dest
= cpu_rq(dest_cpu
);
4950 double_rq_lock(rq_src
, rq_dest
);
4951 /* Already moved. */
4952 if (task_cpu(p
) != src_cpu
)
4954 /* Affinity changed (again). */
4955 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4958 on_rq
= p
->se
.on_rq
;
4960 deactivate_task(rq_src
, p
, 0);
4962 set_task_cpu(p
, dest_cpu
);
4964 activate_task(rq_dest
, p
, 0);
4965 check_preempt_curr(rq_dest
, p
);
4969 double_rq_unlock(rq_src
, rq_dest
);
4974 * migration_thread - this is a highprio system thread that performs
4975 * thread migration by bumping thread off CPU then 'pushing' onto
4978 static int migration_thread(void *data
)
4980 int cpu
= (long)data
;
4984 BUG_ON(rq
->migration_thread
!= current
);
4986 set_current_state(TASK_INTERRUPTIBLE
);
4987 while (!kthread_should_stop()) {
4988 struct migration_req
*req
;
4989 struct list_head
*head
;
4991 spin_lock_irq(&rq
->lock
);
4993 if (cpu_is_offline(cpu
)) {
4994 spin_unlock_irq(&rq
->lock
);
4998 if (rq
->active_balance
) {
4999 active_load_balance(rq
, cpu
);
5000 rq
->active_balance
= 0;
5003 head
= &rq
->migration_queue
;
5005 if (list_empty(head
)) {
5006 spin_unlock_irq(&rq
->lock
);
5008 set_current_state(TASK_INTERRUPTIBLE
);
5011 req
= list_entry(head
->next
, struct migration_req
, list
);
5012 list_del_init(head
->next
);
5014 spin_unlock(&rq
->lock
);
5015 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5018 complete(&req
->done
);
5020 __set_current_state(TASK_RUNNING
);
5024 /* Wait for kthread_stop */
5025 set_current_state(TASK_INTERRUPTIBLE
);
5026 while (!kthread_should_stop()) {
5028 set_current_state(TASK_INTERRUPTIBLE
);
5030 __set_current_state(TASK_RUNNING
);
5034 #ifdef CONFIG_HOTPLUG_CPU
5036 * Figure out where task on dead CPU should go, use force if neccessary.
5037 * NOTE: interrupts should be disabled by the caller
5039 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5041 unsigned long flags
;
5048 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5049 cpus_and(mask
, mask
, p
->cpus_allowed
);
5050 dest_cpu
= any_online_cpu(mask
);
5052 /* On any allowed CPU? */
5053 if (dest_cpu
== NR_CPUS
)
5054 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5056 /* No more Mr. Nice Guy. */
5057 if (dest_cpu
== NR_CPUS
) {
5058 rq
= task_rq_lock(p
, &flags
);
5059 cpus_setall(p
->cpus_allowed
);
5060 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5061 task_rq_unlock(rq
, &flags
);
5064 * Don't tell them about moving exiting tasks or
5065 * kernel threads (both mm NULL), since they never
5068 if (p
->mm
&& printk_ratelimit())
5069 printk(KERN_INFO
"process %d (%s) no "
5070 "longer affine to cpu%d\n",
5071 p
->pid
, p
->comm
, dead_cpu
);
5073 } while (!__migrate_task(p
, dead_cpu
, dest_cpu
));
5077 * While a dead CPU has no uninterruptible tasks queued at this point,
5078 * it might still have a nonzero ->nr_uninterruptible counter, because
5079 * for performance reasons the counter is not stricly tracking tasks to
5080 * their home CPUs. So we just add the counter to another CPU's counter,
5081 * to keep the global sum constant after CPU-down:
5083 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5085 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5086 unsigned long flags
;
5088 local_irq_save(flags
);
5089 double_rq_lock(rq_src
, rq_dest
);
5090 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5091 rq_src
->nr_uninterruptible
= 0;
5092 double_rq_unlock(rq_src
, rq_dest
);
5093 local_irq_restore(flags
);
5096 /* Run through task list and migrate tasks from the dead cpu. */
5097 static void migrate_live_tasks(int src_cpu
)
5099 struct task_struct
*p
, *t
;
5101 write_lock_irq(&tasklist_lock
);
5103 do_each_thread(t
, p
) {
5107 if (task_cpu(p
) == src_cpu
)
5108 move_task_off_dead_cpu(src_cpu
, p
);
5109 } while_each_thread(t
, p
);
5111 write_unlock_irq(&tasklist_lock
);
5115 * activate_idle_task - move idle task to the _front_ of runqueue.
5117 static void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
5119 update_rq_clock(rq
);
5121 if (p
->state
== TASK_UNINTERRUPTIBLE
)
5122 rq
->nr_uninterruptible
--;
5124 enqueue_task(rq
, p
, 0);
5125 inc_nr_running(p
, rq
);
5129 * Schedules idle task to be the next runnable task on current CPU.
5130 * It does so by boosting its priority to highest possible and adding it to
5131 * the _front_ of the runqueue. Used by CPU offline code.
5133 void sched_idle_next(void)
5135 int this_cpu
= smp_processor_id();
5136 struct rq
*rq
= cpu_rq(this_cpu
);
5137 struct task_struct
*p
= rq
->idle
;
5138 unsigned long flags
;
5140 /* cpu has to be offline */
5141 BUG_ON(cpu_online(this_cpu
));
5144 * Strictly not necessary since rest of the CPUs are stopped by now
5145 * and interrupts disabled on the current cpu.
5147 spin_lock_irqsave(&rq
->lock
, flags
);
5149 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5151 /* Add idle task to the _front_ of its priority queue: */
5152 activate_idle_task(p
, rq
);
5154 spin_unlock_irqrestore(&rq
->lock
, flags
);
5158 * Ensures that the idle task is using init_mm right before its cpu goes
5161 void idle_task_exit(void)
5163 struct mm_struct
*mm
= current
->active_mm
;
5165 BUG_ON(cpu_online(smp_processor_id()));
5168 switch_mm(mm
, &init_mm
, current
);
5172 /* called under rq->lock with disabled interrupts */
5173 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5175 struct rq
*rq
= cpu_rq(dead_cpu
);
5177 /* Must be exiting, otherwise would be on tasklist. */
5178 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5180 /* Cannot have done final schedule yet: would have vanished. */
5181 BUG_ON(p
->state
== TASK_DEAD
);
5186 * Drop lock around migration; if someone else moves it,
5187 * that's OK. No task can be added to this CPU, so iteration is
5189 * NOTE: interrupts should be left disabled --dev@
5191 spin_unlock(&rq
->lock
);
5192 move_task_off_dead_cpu(dead_cpu
, p
);
5193 spin_lock(&rq
->lock
);
5198 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5199 static void migrate_dead_tasks(unsigned int dead_cpu
)
5201 struct rq
*rq
= cpu_rq(dead_cpu
);
5202 struct task_struct
*next
;
5205 if (!rq
->nr_running
)
5207 update_rq_clock(rq
);
5208 next
= pick_next_task(rq
, rq
->curr
);
5211 migrate_dead(dead_cpu
, next
);
5215 #endif /* CONFIG_HOTPLUG_CPU */
5217 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5219 static struct ctl_table sd_ctl_dir
[] = {
5221 .procname
= "sched_domain",
5227 static struct ctl_table sd_ctl_root
[] = {
5229 .ctl_name
= CTL_KERN
,
5230 .procname
= "kernel",
5232 .child
= sd_ctl_dir
,
5237 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5239 struct ctl_table
*entry
=
5240 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5243 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5249 set_table_entry(struct ctl_table
*entry
,
5250 const char *procname
, void *data
, int maxlen
,
5251 mode_t mode
, proc_handler
*proc_handler
)
5253 entry
->procname
= procname
;
5255 entry
->maxlen
= maxlen
;
5257 entry
->proc_handler
= proc_handler
;
5260 static struct ctl_table
*
5261 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5263 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5265 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5266 sizeof(long), 0644, proc_doulongvec_minmax
);
5267 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5268 sizeof(long), 0644, proc_doulongvec_minmax
);
5269 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5270 sizeof(int), 0644, proc_dointvec_minmax
);
5271 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5272 sizeof(int), 0644, proc_dointvec_minmax
);
5273 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5274 sizeof(int), 0644, proc_dointvec_minmax
);
5275 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5276 sizeof(int), 0644, proc_dointvec_minmax
);
5277 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5278 sizeof(int), 0644, proc_dointvec_minmax
);
5279 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5280 sizeof(int), 0644, proc_dointvec_minmax
);
5281 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5282 sizeof(int), 0644, proc_dointvec_minmax
);
5283 set_table_entry(&table
[9], "cache_nice_tries",
5284 &sd
->cache_nice_tries
,
5285 sizeof(int), 0644, proc_dointvec_minmax
);
5286 set_table_entry(&table
[10], "flags", &sd
->flags
,
5287 sizeof(int), 0644, proc_dointvec_minmax
);
5292 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5294 struct ctl_table
*entry
, *table
;
5295 struct sched_domain
*sd
;
5296 int domain_num
= 0, i
;
5299 for_each_domain(cpu
, sd
)
5301 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5304 for_each_domain(cpu
, sd
) {
5305 snprintf(buf
, 32, "domain%d", i
);
5306 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5308 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5315 static struct ctl_table_header
*sd_sysctl_header
;
5316 static void init_sched_domain_sysctl(void)
5318 int i
, cpu_num
= num_online_cpus();
5319 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5322 sd_ctl_dir
[0].child
= entry
;
5324 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5325 snprintf(buf
, 32, "cpu%d", i
);
5326 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5328 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5330 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5333 static void init_sched_domain_sysctl(void)
5339 * migration_call - callback that gets triggered when a CPU is added.
5340 * Here we can start up the necessary migration thread for the new CPU.
5342 static int __cpuinit
5343 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5345 struct task_struct
*p
;
5346 int cpu
= (long)hcpu
;
5347 unsigned long flags
;
5351 case CPU_LOCK_ACQUIRE
:
5352 mutex_lock(&sched_hotcpu_mutex
);
5355 case CPU_UP_PREPARE
:
5356 case CPU_UP_PREPARE_FROZEN
:
5357 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5360 kthread_bind(p
, cpu
);
5361 /* Must be high prio: stop_machine expects to yield to it. */
5362 rq
= task_rq_lock(p
, &flags
);
5363 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5364 task_rq_unlock(rq
, &flags
);
5365 cpu_rq(cpu
)->migration_thread
= p
;
5369 case CPU_ONLINE_FROZEN
:
5370 /* Strictly unneccessary, as first user will wake it. */
5371 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5374 #ifdef CONFIG_HOTPLUG_CPU
5375 case CPU_UP_CANCELED
:
5376 case CPU_UP_CANCELED_FROZEN
:
5377 if (!cpu_rq(cpu
)->migration_thread
)
5379 /* Unbind it from offline cpu so it can run. Fall thru. */
5380 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5381 any_online_cpu(cpu_online_map
));
5382 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5383 cpu_rq(cpu
)->migration_thread
= NULL
;
5387 case CPU_DEAD_FROZEN
:
5388 migrate_live_tasks(cpu
);
5390 kthread_stop(rq
->migration_thread
);
5391 rq
->migration_thread
= NULL
;
5392 /* Idle task back to normal (off runqueue, low prio) */
5393 rq
= task_rq_lock(rq
->idle
, &flags
);
5394 update_rq_clock(rq
);
5395 deactivate_task(rq
, rq
->idle
, 0);
5396 rq
->idle
->static_prio
= MAX_PRIO
;
5397 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5398 rq
->idle
->sched_class
= &idle_sched_class
;
5399 migrate_dead_tasks(cpu
);
5400 task_rq_unlock(rq
, &flags
);
5401 migrate_nr_uninterruptible(rq
);
5402 BUG_ON(rq
->nr_running
!= 0);
5404 /* No need to migrate the tasks: it was best-effort if
5405 * they didn't take sched_hotcpu_mutex. Just wake up
5406 * the requestors. */
5407 spin_lock_irq(&rq
->lock
);
5408 while (!list_empty(&rq
->migration_queue
)) {
5409 struct migration_req
*req
;
5411 req
= list_entry(rq
->migration_queue
.next
,
5412 struct migration_req
, list
);
5413 list_del_init(&req
->list
);
5414 complete(&req
->done
);
5416 spin_unlock_irq(&rq
->lock
);
5419 case CPU_LOCK_RELEASE
:
5420 mutex_unlock(&sched_hotcpu_mutex
);
5426 /* Register at highest priority so that task migration (migrate_all_tasks)
5427 * happens before everything else.
5429 static struct notifier_block __cpuinitdata migration_notifier
= {
5430 .notifier_call
= migration_call
,
5434 int __init
migration_init(void)
5436 void *cpu
= (void *)(long)smp_processor_id();
5439 /* Start one for the boot CPU: */
5440 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5441 BUG_ON(err
== NOTIFY_BAD
);
5442 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5443 register_cpu_notifier(&migration_notifier
);
5451 /* Number of possible processor ids */
5452 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5453 EXPORT_SYMBOL(nr_cpu_ids
);
5455 #ifdef CONFIG_SCHED_DEBUG
5456 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5461 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5465 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5470 struct sched_group
*group
= sd
->groups
;
5471 cpumask_t groupmask
;
5473 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5474 cpus_clear(groupmask
);
5477 for (i
= 0; i
< level
+ 1; i
++)
5479 printk("domain %d: ", level
);
5481 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5482 printk("does not load-balance\n");
5484 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5489 printk("span %s\n", str
);
5491 if (!cpu_isset(cpu
, sd
->span
))
5492 printk(KERN_ERR
"ERROR: domain->span does not contain "
5494 if (!cpu_isset(cpu
, group
->cpumask
))
5495 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5499 for (i
= 0; i
< level
+ 2; i
++)
5505 printk(KERN_ERR
"ERROR: group is NULL\n");
5509 if (!group
->__cpu_power
) {
5511 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5516 if (!cpus_weight(group
->cpumask
)) {
5518 printk(KERN_ERR
"ERROR: empty group\n");
5522 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5524 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5528 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5530 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5533 group
= group
->next
;
5534 } while (group
!= sd
->groups
);
5537 if (!cpus_equal(sd
->span
, groupmask
))
5538 printk(KERN_ERR
"ERROR: groups don't span "
5546 if (!cpus_subset(groupmask
, sd
->span
))
5547 printk(KERN_ERR
"ERROR: parent span is not a superset "
5548 "of domain->span\n");
5553 # define sched_domain_debug(sd, cpu) do { } while (0)
5556 static int sd_degenerate(struct sched_domain
*sd
)
5558 if (cpus_weight(sd
->span
) == 1)
5561 /* Following flags need at least 2 groups */
5562 if (sd
->flags
& (SD_LOAD_BALANCE
|
5563 SD_BALANCE_NEWIDLE
|
5567 SD_SHARE_PKG_RESOURCES
)) {
5568 if (sd
->groups
!= sd
->groups
->next
)
5572 /* Following flags don't use groups */
5573 if (sd
->flags
& (SD_WAKE_IDLE
|
5582 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5584 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5586 if (sd_degenerate(parent
))
5589 if (!cpus_equal(sd
->span
, parent
->span
))
5592 /* Does parent contain flags not in child? */
5593 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5594 if (cflags
& SD_WAKE_AFFINE
)
5595 pflags
&= ~SD_WAKE_BALANCE
;
5596 /* Flags needing groups don't count if only 1 group in parent */
5597 if (parent
->groups
== parent
->groups
->next
) {
5598 pflags
&= ~(SD_LOAD_BALANCE
|
5599 SD_BALANCE_NEWIDLE
|
5603 SD_SHARE_PKG_RESOURCES
);
5605 if (~cflags
& pflags
)
5612 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5613 * hold the hotplug lock.
5615 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5617 struct rq
*rq
= cpu_rq(cpu
);
5618 struct sched_domain
*tmp
;
5620 /* Remove the sched domains which do not contribute to scheduling. */
5621 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5622 struct sched_domain
*parent
= tmp
->parent
;
5625 if (sd_parent_degenerate(tmp
, parent
)) {
5626 tmp
->parent
= parent
->parent
;
5628 parent
->parent
->child
= tmp
;
5632 if (sd
&& sd_degenerate(sd
)) {
5638 sched_domain_debug(sd
, cpu
);
5640 rcu_assign_pointer(rq
->sd
, sd
);
5643 /* cpus with isolated domains */
5644 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5646 /* Setup the mask of cpus configured for isolated domains */
5647 static int __init
isolated_cpu_setup(char *str
)
5649 int ints
[NR_CPUS
], i
;
5651 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5652 cpus_clear(cpu_isolated_map
);
5653 for (i
= 1; i
<= ints
[0]; i
++)
5654 if (ints
[i
] < NR_CPUS
)
5655 cpu_set(ints
[i
], cpu_isolated_map
);
5659 __setup("isolcpus=", isolated_cpu_setup
);
5662 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5663 * to a function which identifies what group(along with sched group) a CPU
5664 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5665 * (due to the fact that we keep track of groups covered with a cpumask_t).
5667 * init_sched_build_groups will build a circular linked list of the groups
5668 * covered by the given span, and will set each group's ->cpumask correctly,
5669 * and ->cpu_power to 0.
5672 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5673 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5674 struct sched_group
**sg
))
5676 struct sched_group
*first
= NULL
, *last
= NULL
;
5677 cpumask_t covered
= CPU_MASK_NONE
;
5680 for_each_cpu_mask(i
, span
) {
5681 struct sched_group
*sg
;
5682 int group
= group_fn(i
, cpu_map
, &sg
);
5685 if (cpu_isset(i
, covered
))
5688 sg
->cpumask
= CPU_MASK_NONE
;
5689 sg
->__cpu_power
= 0;
5691 for_each_cpu_mask(j
, span
) {
5692 if (group_fn(j
, cpu_map
, NULL
) != group
)
5695 cpu_set(j
, covered
);
5696 cpu_set(j
, sg
->cpumask
);
5707 #define SD_NODES_PER_DOMAIN 16
5712 * find_next_best_node - find the next node to include in a sched_domain
5713 * @node: node whose sched_domain we're building
5714 * @used_nodes: nodes already in the sched_domain
5716 * Find the next node to include in a given scheduling domain. Simply
5717 * finds the closest node not already in the @used_nodes map.
5719 * Should use nodemask_t.
5721 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5723 int i
, n
, val
, min_val
, best_node
= 0;
5727 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5728 /* Start at @node */
5729 n
= (node
+ i
) % MAX_NUMNODES
;
5731 if (!nr_cpus_node(n
))
5734 /* Skip already used nodes */
5735 if (test_bit(n
, used_nodes
))
5738 /* Simple min distance search */
5739 val
= node_distance(node
, n
);
5741 if (val
< min_val
) {
5747 set_bit(best_node
, used_nodes
);
5752 * sched_domain_node_span - get a cpumask for a node's sched_domain
5753 * @node: node whose cpumask we're constructing
5754 * @size: number of nodes to include in this span
5756 * Given a node, construct a good cpumask for its sched_domain to span. It
5757 * should be one that prevents unnecessary balancing, but also spreads tasks
5760 static cpumask_t
sched_domain_node_span(int node
)
5762 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5763 cpumask_t span
, nodemask
;
5767 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5769 nodemask
= node_to_cpumask(node
);
5770 cpus_or(span
, span
, nodemask
);
5771 set_bit(node
, used_nodes
);
5773 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5774 int next_node
= find_next_best_node(node
, used_nodes
);
5776 nodemask
= node_to_cpumask(next_node
);
5777 cpus_or(span
, span
, nodemask
);
5784 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5787 * SMT sched-domains:
5789 #ifdef CONFIG_SCHED_SMT
5790 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5791 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5793 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5794 struct sched_group
**sg
)
5797 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5803 * multi-core sched-domains:
5805 #ifdef CONFIG_SCHED_MC
5806 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5807 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5810 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5811 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5812 struct sched_group
**sg
)
5815 cpumask_t mask
= cpu_sibling_map
[cpu
];
5816 cpus_and(mask
, mask
, *cpu_map
);
5817 group
= first_cpu(mask
);
5819 *sg
= &per_cpu(sched_group_core
, group
);
5822 #elif defined(CONFIG_SCHED_MC)
5823 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5824 struct sched_group
**sg
)
5827 *sg
= &per_cpu(sched_group_core
, cpu
);
5832 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5833 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5835 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5836 struct sched_group
**sg
)
5839 #ifdef CONFIG_SCHED_MC
5840 cpumask_t mask
= cpu_coregroup_map(cpu
);
5841 cpus_and(mask
, mask
, *cpu_map
);
5842 group
= first_cpu(mask
);
5843 #elif defined(CONFIG_SCHED_SMT)
5844 cpumask_t mask
= cpu_sibling_map
[cpu
];
5845 cpus_and(mask
, mask
, *cpu_map
);
5846 group
= first_cpu(mask
);
5851 *sg
= &per_cpu(sched_group_phys
, group
);
5857 * The init_sched_build_groups can't handle what we want to do with node
5858 * groups, so roll our own. Now each node has its own list of groups which
5859 * gets dynamically allocated.
5861 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5862 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5864 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5865 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5867 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5868 struct sched_group
**sg
)
5870 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5873 cpus_and(nodemask
, nodemask
, *cpu_map
);
5874 group
= first_cpu(nodemask
);
5877 *sg
= &per_cpu(sched_group_allnodes
, group
);
5881 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5883 struct sched_group
*sg
= group_head
;
5889 for_each_cpu_mask(j
, sg
->cpumask
) {
5890 struct sched_domain
*sd
;
5892 sd
= &per_cpu(phys_domains
, j
);
5893 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5895 * Only add "power" once for each
5901 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5904 } while (sg
!= group_head
);
5909 /* Free memory allocated for various sched_group structures */
5910 static void free_sched_groups(const cpumask_t
*cpu_map
)
5914 for_each_cpu_mask(cpu
, *cpu_map
) {
5915 struct sched_group
**sched_group_nodes
5916 = sched_group_nodes_bycpu
[cpu
];
5918 if (!sched_group_nodes
)
5921 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5922 cpumask_t nodemask
= node_to_cpumask(i
);
5923 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5925 cpus_and(nodemask
, nodemask
, *cpu_map
);
5926 if (cpus_empty(nodemask
))
5936 if (oldsg
!= sched_group_nodes
[i
])
5939 kfree(sched_group_nodes
);
5940 sched_group_nodes_bycpu
[cpu
] = NULL
;
5944 static void free_sched_groups(const cpumask_t
*cpu_map
)
5950 * Initialize sched groups cpu_power.
5952 * cpu_power indicates the capacity of sched group, which is used while
5953 * distributing the load between different sched groups in a sched domain.
5954 * Typically cpu_power for all the groups in a sched domain will be same unless
5955 * there are asymmetries in the topology. If there are asymmetries, group
5956 * having more cpu_power will pickup more load compared to the group having
5959 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5960 * the maximum number of tasks a group can handle in the presence of other idle
5961 * or lightly loaded groups in the same sched domain.
5963 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5965 struct sched_domain
*child
;
5966 struct sched_group
*group
;
5968 WARN_ON(!sd
|| !sd
->groups
);
5970 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5975 sd
->groups
->__cpu_power
= 0;
5978 * For perf policy, if the groups in child domain share resources
5979 * (for example cores sharing some portions of the cache hierarchy
5980 * or SMT), then set this domain groups cpu_power such that each group
5981 * can handle only one task, when there are other idle groups in the
5982 * same sched domain.
5984 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5986 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5987 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5992 * add cpu_power of each child group to this groups cpu_power
5994 group
= child
->groups
;
5996 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5997 group
= group
->next
;
5998 } while (group
!= child
->groups
);
6002 * Build sched domains for a given set of cpus and attach the sched domains
6003 * to the individual cpus
6005 static int build_sched_domains(const cpumask_t
*cpu_map
)
6009 struct sched_group
**sched_group_nodes
= NULL
;
6010 int sd_allnodes
= 0;
6013 * Allocate the per-node list of sched groups
6015 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6017 if (!sched_group_nodes
) {
6018 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6021 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6025 * Set up domains for cpus specified by the cpu_map.
6027 for_each_cpu_mask(i
, *cpu_map
) {
6028 struct sched_domain
*sd
= NULL
, *p
;
6029 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6031 cpus_and(nodemask
, nodemask
, *cpu_map
);
6034 if (cpus_weight(*cpu_map
) >
6035 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6036 sd
= &per_cpu(allnodes_domains
, i
);
6037 *sd
= SD_ALLNODES_INIT
;
6038 sd
->span
= *cpu_map
;
6039 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6045 sd
= &per_cpu(node_domains
, i
);
6047 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6051 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6055 sd
= &per_cpu(phys_domains
, i
);
6057 sd
->span
= nodemask
;
6061 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6063 #ifdef CONFIG_SCHED_MC
6065 sd
= &per_cpu(core_domains
, i
);
6067 sd
->span
= cpu_coregroup_map(i
);
6068 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6071 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6074 #ifdef CONFIG_SCHED_SMT
6076 sd
= &per_cpu(cpu_domains
, i
);
6077 *sd
= SD_SIBLING_INIT
;
6078 sd
->span
= cpu_sibling_map
[i
];
6079 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6082 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6086 #ifdef CONFIG_SCHED_SMT
6087 /* Set up CPU (sibling) groups */
6088 for_each_cpu_mask(i
, *cpu_map
) {
6089 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6090 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6091 if (i
!= first_cpu(this_sibling_map
))
6094 init_sched_build_groups(this_sibling_map
, cpu_map
,
6099 #ifdef CONFIG_SCHED_MC
6100 /* Set up multi-core groups */
6101 for_each_cpu_mask(i
, *cpu_map
) {
6102 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6103 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6104 if (i
!= first_cpu(this_core_map
))
6106 init_sched_build_groups(this_core_map
, cpu_map
,
6107 &cpu_to_core_group
);
6111 /* Set up physical groups */
6112 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6113 cpumask_t nodemask
= node_to_cpumask(i
);
6115 cpus_and(nodemask
, nodemask
, *cpu_map
);
6116 if (cpus_empty(nodemask
))
6119 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6123 /* Set up node groups */
6125 init_sched_build_groups(*cpu_map
, cpu_map
,
6126 &cpu_to_allnodes_group
);
6128 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6129 /* Set up node groups */
6130 struct sched_group
*sg
, *prev
;
6131 cpumask_t nodemask
= node_to_cpumask(i
);
6132 cpumask_t domainspan
;
6133 cpumask_t covered
= CPU_MASK_NONE
;
6136 cpus_and(nodemask
, nodemask
, *cpu_map
);
6137 if (cpus_empty(nodemask
)) {
6138 sched_group_nodes
[i
] = NULL
;
6142 domainspan
= sched_domain_node_span(i
);
6143 cpus_and(domainspan
, domainspan
, *cpu_map
);
6145 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6147 printk(KERN_WARNING
"Can not alloc domain group for "
6151 sched_group_nodes
[i
] = sg
;
6152 for_each_cpu_mask(j
, nodemask
) {
6153 struct sched_domain
*sd
;
6155 sd
= &per_cpu(node_domains
, j
);
6158 sg
->__cpu_power
= 0;
6159 sg
->cpumask
= nodemask
;
6161 cpus_or(covered
, covered
, nodemask
);
6164 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6165 cpumask_t tmp
, notcovered
;
6166 int n
= (i
+ j
) % MAX_NUMNODES
;
6168 cpus_complement(notcovered
, covered
);
6169 cpus_and(tmp
, notcovered
, *cpu_map
);
6170 cpus_and(tmp
, tmp
, domainspan
);
6171 if (cpus_empty(tmp
))
6174 nodemask
= node_to_cpumask(n
);
6175 cpus_and(tmp
, tmp
, nodemask
);
6176 if (cpus_empty(tmp
))
6179 sg
= kmalloc_node(sizeof(struct sched_group
),
6183 "Can not alloc domain group for node %d\n", j
);
6186 sg
->__cpu_power
= 0;
6188 sg
->next
= prev
->next
;
6189 cpus_or(covered
, covered
, tmp
);
6196 /* Calculate CPU power for physical packages and nodes */
6197 #ifdef CONFIG_SCHED_SMT
6198 for_each_cpu_mask(i
, *cpu_map
) {
6199 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6201 init_sched_groups_power(i
, sd
);
6204 #ifdef CONFIG_SCHED_MC
6205 for_each_cpu_mask(i
, *cpu_map
) {
6206 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6208 init_sched_groups_power(i
, sd
);
6212 for_each_cpu_mask(i
, *cpu_map
) {
6213 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6215 init_sched_groups_power(i
, sd
);
6219 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6220 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6223 struct sched_group
*sg
;
6225 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6226 init_numa_sched_groups_power(sg
);
6230 /* Attach the domains */
6231 for_each_cpu_mask(i
, *cpu_map
) {
6232 struct sched_domain
*sd
;
6233 #ifdef CONFIG_SCHED_SMT
6234 sd
= &per_cpu(cpu_domains
, i
);
6235 #elif defined(CONFIG_SCHED_MC)
6236 sd
= &per_cpu(core_domains
, i
);
6238 sd
= &per_cpu(phys_domains
, i
);
6240 cpu_attach_domain(sd
, i
);
6247 free_sched_groups(cpu_map
);
6252 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6254 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6256 cpumask_t cpu_default_map
;
6260 * Setup mask for cpus without special case scheduling requirements.
6261 * For now this just excludes isolated cpus, but could be used to
6262 * exclude other special cases in the future.
6264 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6266 err
= build_sched_domains(&cpu_default_map
);
6271 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6273 free_sched_groups(cpu_map
);
6277 * Detach sched domains from a group of cpus specified in cpu_map
6278 * These cpus will now be attached to the NULL domain
6280 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6284 for_each_cpu_mask(i
, *cpu_map
)
6285 cpu_attach_domain(NULL
, i
);
6286 synchronize_sched();
6287 arch_destroy_sched_domains(cpu_map
);
6291 * Partition sched domains as specified by the cpumasks below.
6292 * This attaches all cpus from the cpumasks to the NULL domain,
6293 * waits for a RCU quiescent period, recalculates sched
6294 * domain information and then attaches them back to the
6295 * correct sched domains
6296 * Call with hotplug lock held
6298 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6300 cpumask_t change_map
;
6303 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6304 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6305 cpus_or(change_map
, *partition1
, *partition2
);
6307 /* Detach sched domains from all of the affected cpus */
6308 detach_destroy_domains(&change_map
);
6309 if (!cpus_empty(*partition1
))
6310 err
= build_sched_domains(partition1
);
6311 if (!err
&& !cpus_empty(*partition2
))
6312 err
= build_sched_domains(partition2
);
6317 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6318 static int arch_reinit_sched_domains(void)
6322 mutex_lock(&sched_hotcpu_mutex
);
6323 detach_destroy_domains(&cpu_online_map
);
6324 err
= arch_init_sched_domains(&cpu_online_map
);
6325 mutex_unlock(&sched_hotcpu_mutex
);
6330 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6334 if (buf
[0] != '0' && buf
[0] != '1')
6338 sched_smt_power_savings
= (buf
[0] == '1');
6340 sched_mc_power_savings
= (buf
[0] == '1');
6342 ret
= arch_reinit_sched_domains();
6344 return ret
? ret
: count
;
6347 #ifdef CONFIG_SCHED_MC
6348 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6350 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6352 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6353 const char *buf
, size_t count
)
6355 return sched_power_savings_store(buf
, count
, 0);
6357 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6358 sched_mc_power_savings_store
);
6361 #ifdef CONFIG_SCHED_SMT
6362 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6364 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6366 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6367 const char *buf
, size_t count
)
6369 return sched_power_savings_store(buf
, count
, 1);
6371 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6372 sched_smt_power_savings_store
);
6375 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6379 #ifdef CONFIG_SCHED_SMT
6381 err
= sysfs_create_file(&cls
->kset
.kobj
,
6382 &attr_sched_smt_power_savings
.attr
);
6384 #ifdef CONFIG_SCHED_MC
6385 if (!err
&& mc_capable())
6386 err
= sysfs_create_file(&cls
->kset
.kobj
,
6387 &attr_sched_mc_power_savings
.attr
);
6394 * Force a reinitialization of the sched domains hierarchy. The domains
6395 * and groups cannot be updated in place without racing with the balancing
6396 * code, so we temporarily attach all running cpus to the NULL domain
6397 * which will prevent rebalancing while the sched domains are recalculated.
6399 static int update_sched_domains(struct notifier_block
*nfb
,
6400 unsigned long action
, void *hcpu
)
6403 case CPU_UP_PREPARE
:
6404 case CPU_UP_PREPARE_FROZEN
:
6405 case CPU_DOWN_PREPARE
:
6406 case CPU_DOWN_PREPARE_FROZEN
:
6407 detach_destroy_domains(&cpu_online_map
);
6410 case CPU_UP_CANCELED
:
6411 case CPU_UP_CANCELED_FROZEN
:
6412 case CPU_DOWN_FAILED
:
6413 case CPU_DOWN_FAILED_FROZEN
:
6415 case CPU_ONLINE_FROZEN
:
6417 case CPU_DEAD_FROZEN
:
6419 * Fall through and re-initialise the domains.
6426 /* The hotplug lock is already held by cpu_up/cpu_down */
6427 arch_init_sched_domains(&cpu_online_map
);
6432 void __init
sched_init_smp(void)
6434 cpumask_t non_isolated_cpus
;
6436 mutex_lock(&sched_hotcpu_mutex
);
6437 arch_init_sched_domains(&cpu_online_map
);
6438 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6439 if (cpus_empty(non_isolated_cpus
))
6440 cpu_set(smp_processor_id(), non_isolated_cpus
);
6441 mutex_unlock(&sched_hotcpu_mutex
);
6442 /* XXX: Theoretical race here - CPU may be hotplugged now */
6443 hotcpu_notifier(update_sched_domains
, 0);
6445 init_sched_domain_sysctl();
6447 /* Move init over to a non-isolated CPU */
6448 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6452 void __init
sched_init_smp(void)
6455 #endif /* CONFIG_SMP */
6457 int in_sched_functions(unsigned long addr
)
6459 /* Linker adds these: start and end of __sched functions */
6460 extern char __sched_text_start
[], __sched_text_end
[];
6462 return in_lock_functions(addr
) ||
6463 (addr
>= (unsigned long)__sched_text_start
6464 && addr
< (unsigned long)__sched_text_end
);
6467 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6469 cfs_rq
->tasks_timeline
= RB_ROOT
;
6470 #ifdef CONFIG_FAIR_GROUP_SCHED
6473 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6476 void __init
sched_init(void)
6478 int highest_cpu
= 0;
6481 for_each_possible_cpu(i
) {
6482 struct rt_prio_array
*array
;
6486 spin_lock_init(&rq
->lock
);
6487 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6490 init_cfs_rq(&rq
->cfs
, rq
);
6491 #ifdef CONFIG_FAIR_GROUP_SCHED
6492 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6494 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6495 struct sched_entity
*se
=
6496 &per_cpu(init_sched_entity
, i
);
6498 init_cfs_rq_p
[i
] = cfs_rq
;
6499 init_cfs_rq(cfs_rq
, rq
);
6500 cfs_rq
->tg
= &init_task_group
;
6501 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6502 &rq
->leaf_cfs_rq_list
);
6504 init_sched_entity_p
[i
] = se
;
6505 se
->cfs_rq
= &rq
->cfs
;
6507 se
->load
.weight
= init_task_group_load
;
6508 se
->load
.inv_weight
=
6509 div64_64(1ULL<<32, init_task_group_load
);
6512 init_task_group
.shares
= init_task_group_load
;
6513 spin_lock_init(&init_task_group
.lock
);
6516 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6517 rq
->cpu_load
[j
] = 0;
6520 rq
->active_balance
= 0;
6521 rq
->next_balance
= jiffies
;
6524 rq
->migration_thread
= NULL
;
6525 INIT_LIST_HEAD(&rq
->migration_queue
);
6527 atomic_set(&rq
->nr_iowait
, 0);
6529 array
= &rq
->rt
.active
;
6530 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6531 INIT_LIST_HEAD(array
->queue
+ j
);
6532 __clear_bit(j
, array
->bitmap
);
6535 /* delimiter for bitsearch: */
6536 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6539 set_load_weight(&init_task
);
6541 #ifdef CONFIG_PREEMPT_NOTIFIERS
6542 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6546 nr_cpu_ids
= highest_cpu
+ 1;
6547 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6550 #ifdef CONFIG_RT_MUTEXES
6551 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6555 * The boot idle thread does lazy MMU switching as well:
6557 atomic_inc(&init_mm
.mm_count
);
6558 enter_lazy_tlb(&init_mm
, current
);
6561 * Make us the idle thread. Technically, schedule() should not be
6562 * called from this thread, however somewhere below it might be,
6563 * but because we are the idle thread, we just pick up running again
6564 * when this runqueue becomes "idle".
6566 init_idle(current
, smp_processor_id());
6568 * During early bootup we pretend to be a normal task:
6570 current
->sched_class
= &fair_sched_class
;
6573 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6574 void __might_sleep(char *file
, int line
)
6577 static unsigned long prev_jiffy
; /* ratelimiting */
6579 if ((in_atomic() || irqs_disabled()) &&
6580 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6581 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6583 prev_jiffy
= jiffies
;
6584 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6585 " context at %s:%d\n", file
, line
);
6586 printk("in_atomic():%d, irqs_disabled():%d\n",
6587 in_atomic(), irqs_disabled());
6588 debug_show_held_locks(current
);
6589 if (irqs_disabled())
6590 print_irqtrace_events(current
);
6595 EXPORT_SYMBOL(__might_sleep
);
6598 #ifdef CONFIG_MAGIC_SYSRQ
6599 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6602 update_rq_clock(rq
);
6603 on_rq
= p
->se
.on_rq
;
6605 deactivate_task(rq
, p
, 0);
6606 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6608 activate_task(rq
, p
, 0);
6609 resched_task(rq
->curr
);
6613 void normalize_rt_tasks(void)
6615 struct task_struct
*g
, *p
;
6616 unsigned long flags
;
6619 read_lock_irq(&tasklist_lock
);
6620 do_each_thread(g
, p
) {
6622 * Only normalize user tasks:
6627 p
->se
.exec_start
= 0;
6628 #ifdef CONFIG_SCHEDSTATS
6629 p
->se
.wait_start
= 0;
6630 p
->se
.sleep_start
= 0;
6631 p
->se
.block_start
= 0;
6633 task_rq(p
)->clock
= 0;
6637 * Renice negative nice level userspace
6640 if (TASK_NICE(p
) < 0 && p
->mm
)
6641 set_user_nice(p
, 0);
6645 spin_lock_irqsave(&p
->pi_lock
, flags
);
6646 rq
= __task_rq_lock(p
);
6648 normalize_task(rq
, p
);
6650 __task_rq_unlock(rq
);
6651 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6652 } while_each_thread(g
, p
);
6654 read_unlock_irq(&tasklist_lock
);
6657 #endif /* CONFIG_MAGIC_SYSRQ */
6661 * These functions are only useful for the IA64 MCA handling.
6663 * They can only be called when the whole system has been
6664 * stopped - every CPU needs to be quiescent, and no scheduling
6665 * activity can take place. Using them for anything else would
6666 * be a serious bug, and as a result, they aren't even visible
6667 * under any other configuration.
6671 * curr_task - return the current task for a given cpu.
6672 * @cpu: the processor in question.
6674 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6676 struct task_struct
*curr_task(int cpu
)
6678 return cpu_curr(cpu
);
6682 * set_curr_task - set the current task for a given cpu.
6683 * @cpu: the processor in question.
6684 * @p: the task pointer to set.
6686 * Description: This function must only be used when non-maskable interrupts
6687 * are serviced on a separate stack. It allows the architecture to switch the
6688 * notion of the current task on a cpu in a non-blocking manner. This function
6689 * must be called with all CPU's synchronized, and interrupts disabled, the
6690 * and caller must save the original value of the current task (see
6691 * curr_task() above) and restore that value before reenabling interrupts and
6692 * re-starting the system.
6694 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6696 void set_curr_task(int cpu
, struct task_struct
*p
)
6703 #ifdef CONFIG_FAIR_GROUP_SCHED
6705 /* allocate runqueue etc for a new task group */
6706 struct task_group
*sched_create_group(void)
6708 struct task_group
*tg
;
6709 struct cfs_rq
*cfs_rq
;
6710 struct sched_entity
*se
;
6714 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6716 return ERR_PTR(-ENOMEM
);
6718 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
6721 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
6725 for_each_possible_cpu(i
) {
6728 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
6733 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
6738 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
6739 memset(se
, 0, sizeof(struct sched_entity
));
6741 tg
->cfs_rq
[i
] = cfs_rq
;
6742 init_cfs_rq(cfs_rq
, rq
);
6746 se
->cfs_rq
= &rq
->cfs
;
6748 se
->load
.weight
= NICE_0_LOAD
;
6749 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6753 for_each_possible_cpu(i
) {
6755 cfs_rq
= tg
->cfs_rq
[i
];
6756 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6759 tg
->shares
= NICE_0_LOAD
;
6760 spin_lock_init(&tg
->lock
);
6765 for_each_possible_cpu(i
) {
6767 kfree(tg
->cfs_rq
[i
]);
6775 return ERR_PTR(-ENOMEM
);
6778 /* rcu callback to free various structures associated with a task group */
6779 static void free_sched_group(struct rcu_head
*rhp
)
6781 struct cfs_rq
*cfs_rq
= container_of(rhp
, struct cfs_rq
, rcu
);
6782 struct task_group
*tg
= cfs_rq
->tg
;
6783 struct sched_entity
*se
;
6786 /* now it should be safe to free those cfs_rqs */
6787 for_each_possible_cpu(i
) {
6788 cfs_rq
= tg
->cfs_rq
[i
];
6800 /* Destroy runqueue etc associated with a task group */
6801 void sched_destroy_group(struct task_group
*tg
)
6803 struct cfs_rq
*cfs_rq
;
6806 for_each_possible_cpu(i
) {
6807 cfs_rq
= tg
->cfs_rq
[i
];
6808 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
6811 cfs_rq
= tg
->cfs_rq
[0];
6813 /* wait for possible concurrent references to cfs_rqs complete */
6814 call_rcu(&cfs_rq
->rcu
, free_sched_group
);
6817 /* change task's runqueue when it moves between groups.
6818 * The caller of this function should have put the task in its new group
6819 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6820 * reflect its new group.
6822 void sched_move_task(struct task_struct
*tsk
)
6825 unsigned long flags
;
6828 rq
= task_rq_lock(tsk
, &flags
);
6830 if (tsk
->sched_class
!= &fair_sched_class
)
6833 update_rq_clock(rq
);
6835 running
= task_running(rq
, tsk
);
6836 on_rq
= tsk
->se
.on_rq
;
6839 dequeue_task(rq
, tsk
, 0);
6840 if (unlikely(running
))
6841 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6844 set_task_cfs_rq(tsk
);
6847 if (unlikely(running
))
6848 tsk
->sched_class
->set_curr_task(rq
);
6849 enqueue_task(rq
, tsk
, 0);
6853 task_rq_unlock(rq
, &flags
);
6856 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
6858 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
6859 struct rq
*rq
= cfs_rq
->rq
;
6862 spin_lock_irq(&rq
->lock
);
6866 dequeue_entity(cfs_rq
, se
, 0);
6868 se
->load
.weight
= shares
;
6869 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
6872 enqueue_entity(cfs_rq
, se
, 0);
6874 spin_unlock_irq(&rq
->lock
);
6877 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6881 spin_lock(&tg
->lock
);
6882 if (tg
->shares
== shares
)
6885 tg
->shares
= shares
;
6886 for_each_possible_cpu(i
)
6887 set_se_shares(tg
->se
[i
], shares
);
6890 spin_unlock(&tg
->lock
);
6894 unsigned long sched_group_shares(struct task_group
*tg
)
6899 #endif /* CONFIG_FAIR_GROUP_SCHED */