4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak
)) sched_clock(void)
75 return (unsigned long long)jiffies
* (1000000000 / HZ
);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
122 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
131 sg
->__cpu_power
+= val
;
132 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
136 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio
)
145 if (static_prio
== NICE_TO_PRIO(19))
148 if (static_prio
< NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
151 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
154 static inline int rt_policy(int policy
)
156 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
161 static inline int task_has_rt_policy(struct task_struct
*p
)
163 return rt_policy(p
->policy
);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array
{
170 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
171 struct list_head queue
[MAX_RT_PRIO
];
174 #ifdef CONFIG_FAIR_GROUP_SCHED
178 /* task group related information */
180 /* schedulable entities of this group on each cpu */
181 struct sched_entity
**se
;
182 /* runqueue "owned" by this group on each cpu */
183 struct cfs_rq
**cfs_rq
;
184 unsigned long shares
;
187 /* Default task group's sched entity on each cpu */
188 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
189 /* Default task group's cfs_rq on each cpu */
190 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
192 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
193 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
195 /* Default task group.
196 * Every task in system belong to this group at bootup.
198 struct task_grp init_task_grp
= {
199 .se
= init_sched_entity_p
,
200 .cfs_rq
= init_cfs_rq_p
,
203 #ifdef CONFIG_FAIR_USER_SCHED
204 #define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
206 #define INIT_TASK_GRP_LOAD NICE_0_LOAD
209 static int init_task_grp_load
= INIT_TASK_GRP_LOAD
;
211 /* return group to which a task belongs */
212 static inline struct task_grp
*task_grp(struct task_struct
*p
)
216 #ifdef CONFIG_FAIR_USER_SCHED
225 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
226 static inline void set_task_cfs_rq(struct task_struct
*p
)
228 p
->se
.cfs_rq
= task_grp(p
)->cfs_rq
[task_cpu(p
)];
229 p
->se
.parent
= task_grp(p
)->se
[task_cpu(p
)];
234 static inline void set_task_cfs_rq(struct task_struct
*p
) { }
236 #endif /* CONFIG_FAIR_GROUP_SCHED */
238 /* CFS-related fields in a runqueue */
240 struct load_weight load
;
241 unsigned long nr_running
;
246 struct rb_root tasks_timeline
;
247 struct rb_node
*rb_leftmost
;
248 struct rb_node
*rb_load_balance_curr
;
249 /* 'curr' points to currently running entity on this cfs_rq.
250 * It is set to NULL otherwise (i.e when none are currently running).
252 struct sched_entity
*curr
;
254 unsigned long nr_spread_over
;
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
259 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
260 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
261 * (like users, containers etc.)
263 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
264 * list is used during load balance.
266 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
267 struct task_grp
*tg
; /* group that "owns" this runqueue */
272 /* Real-Time classes' related field in a runqueue: */
274 struct rt_prio_array active
;
275 int rt_load_balance_idx
;
276 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
280 * This is the main, per-CPU runqueue data structure.
282 * Locking rule: those places that want to lock multiple runqueues
283 * (such as the load balancing or the thread migration code), lock
284 * acquire operations must be ordered by ascending &runqueue.
287 spinlock_t lock
; /* runqueue lock */
290 * nr_running and cpu_load should be in the same cacheline because
291 * remote CPUs use both these fields when doing load calculation.
293 unsigned long nr_running
;
294 #define CPU_LOAD_IDX_MAX 5
295 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
296 unsigned char idle_at_tick
;
298 unsigned char in_nohz_recently
;
300 struct load_weight load
; /* capture load from *all* tasks on this cpu */
301 unsigned long nr_load_updates
;
305 #ifdef CONFIG_FAIR_GROUP_SCHED
306 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
311 * This is part of a global counter where only the total sum
312 * over all CPUs matters. A task can increase this counter on
313 * one CPU and if it got migrated afterwards it may decrease
314 * it on another CPU. Always updated under the runqueue lock:
316 unsigned long nr_uninterruptible
;
318 struct task_struct
*curr
, *idle
;
319 unsigned long next_balance
;
320 struct mm_struct
*prev_mm
;
322 u64 clock
, prev_clock_raw
;
325 unsigned int clock_warps
, clock_overflows
;
327 unsigned int clock_deep_idle_events
;
333 struct sched_domain
*sd
;
335 /* For active balancing */
338 int cpu
; /* cpu of this runqueue */
340 struct task_struct
*migration_thread
;
341 struct list_head migration_queue
;
344 #ifdef CONFIG_SCHEDSTATS
346 struct sched_info rq_sched_info
;
348 /* sys_sched_yield() stats */
349 unsigned long yld_exp_empty
;
350 unsigned long yld_act_empty
;
351 unsigned long yld_both_empty
;
352 unsigned long yld_cnt
;
354 /* schedule() stats */
355 unsigned long sched_switch
;
356 unsigned long sched_cnt
;
357 unsigned long sched_goidle
;
359 /* try_to_wake_up() stats */
360 unsigned long ttwu_cnt
;
361 unsigned long ttwu_local
;
364 unsigned long bkl_cnt
;
366 struct lock_class_key rq_lock_key
;
369 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
370 static DEFINE_MUTEX(sched_hotcpu_mutex
);
372 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
374 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
377 static inline int cpu_of(struct rq
*rq
)
387 * Update the per-runqueue clock, as finegrained as the platform can give
388 * us, but without assuming monotonicity, etc.:
390 static void __update_rq_clock(struct rq
*rq
)
392 u64 prev_raw
= rq
->prev_clock_raw
;
393 u64 now
= sched_clock();
394 s64 delta
= now
- prev_raw
;
395 u64 clock
= rq
->clock
;
397 #ifdef CONFIG_SCHED_DEBUG
398 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
401 * Protect against sched_clock() occasionally going backwards:
403 if (unlikely(delta
< 0)) {
408 * Catch too large forward jumps too:
410 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
411 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
412 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
415 rq
->clock_overflows
++;
417 if (unlikely(delta
> rq
->clock_max_delta
))
418 rq
->clock_max_delta
= delta
;
423 rq
->prev_clock_raw
= now
;
427 static void update_rq_clock(struct rq
*rq
)
429 if (likely(smp_processor_id() == cpu_of(rq
)))
430 __update_rq_clock(rq
);
434 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
435 * See detach_destroy_domains: synchronize_sched for details.
437 * The domain tree of any CPU may only be accessed from within
438 * preempt-disabled sections.
440 #define for_each_domain(cpu, __sd) \
441 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
443 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
444 #define this_rq() (&__get_cpu_var(runqueues))
445 #define task_rq(p) cpu_rq(task_cpu(p))
446 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
449 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
451 #ifdef CONFIG_SCHED_DEBUG
452 # define const_debug __read_mostly
454 # define const_debug static const
458 * Debugging: various feature bits
461 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
462 SCHED_FEAT_START_DEBIT
= 2,
463 SCHED_FEAT_USE_TREE_AVG
= 4,
464 SCHED_FEAT_APPROX_AVG
= 8,
467 const_debug
unsigned int sysctl_sched_features
=
468 SCHED_FEAT_NEW_FAIR_SLEEPERS
*1 |
469 SCHED_FEAT_START_DEBIT
*1 |
470 SCHED_FEAT_USE_TREE_AVG
*0 |
471 SCHED_FEAT_APPROX_AVG
*0;
473 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
476 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
477 * clock constructed from sched_clock():
479 unsigned long long cpu_clock(int cpu
)
481 unsigned long long now
;
485 local_irq_save(flags
);
489 local_irq_restore(flags
);
494 #ifndef prepare_arch_switch
495 # define prepare_arch_switch(next) do { } while (0)
497 #ifndef finish_arch_switch
498 # define finish_arch_switch(prev) do { } while (0)
501 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
502 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
504 return rq
->curr
== p
;
507 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
511 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
513 #ifdef CONFIG_DEBUG_SPINLOCK
514 /* this is a valid case when another task releases the spinlock */
515 rq
->lock
.owner
= current
;
518 * If we are tracking spinlock dependencies then we have to
519 * fix up the runqueue lock - which gets 'carried over' from
522 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
524 spin_unlock_irq(&rq
->lock
);
527 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
528 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
533 return rq
->curr
== p
;
537 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
541 * We can optimise this out completely for !SMP, because the
542 * SMP rebalancing from interrupt is the only thing that cares
547 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
548 spin_unlock_irq(&rq
->lock
);
550 spin_unlock(&rq
->lock
);
554 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
558 * After ->oncpu is cleared, the task can be moved to a different CPU.
559 * We must ensure this doesn't happen until the switch is completely
565 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
569 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
572 * __task_rq_lock - lock the runqueue a given task resides on.
573 * Must be called interrupts disabled.
575 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
582 spin_lock(&rq
->lock
);
583 if (unlikely(rq
!= task_rq(p
))) {
584 spin_unlock(&rq
->lock
);
585 goto repeat_lock_task
;
591 * task_rq_lock - lock the runqueue a given task resides on and disable
592 * interrupts. Note the ordering: we can safely lookup the task_rq without
593 * explicitly disabling preemption.
595 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
601 local_irq_save(*flags
);
603 spin_lock(&rq
->lock
);
604 if (unlikely(rq
!= task_rq(p
))) {
605 spin_unlock_irqrestore(&rq
->lock
, *flags
);
606 goto repeat_lock_task
;
611 static inline void __task_rq_unlock(struct rq
*rq
)
614 spin_unlock(&rq
->lock
);
617 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
620 spin_unlock_irqrestore(&rq
->lock
, *flags
);
624 * this_rq_lock - lock this runqueue and disable interrupts.
626 static inline struct rq
*this_rq_lock(void)
633 spin_lock(&rq
->lock
);
639 * We are going deep-idle (irqs are disabled):
641 void sched_clock_idle_sleep_event(void)
643 struct rq
*rq
= cpu_rq(smp_processor_id());
645 spin_lock(&rq
->lock
);
646 __update_rq_clock(rq
);
647 spin_unlock(&rq
->lock
);
648 rq
->clock_deep_idle_events
++;
650 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
653 * We just idled delta nanoseconds (called with irqs disabled):
655 void sched_clock_idle_wakeup_event(u64 delta_ns
)
657 struct rq
*rq
= cpu_rq(smp_processor_id());
658 u64 now
= sched_clock();
660 rq
->idle_clock
+= delta_ns
;
662 * Override the previous timestamp and ignore all
663 * sched_clock() deltas that occured while we idled,
664 * and use the PM-provided delta_ns to advance the
667 spin_lock(&rq
->lock
);
668 rq
->prev_clock_raw
= now
;
669 rq
->clock
+= delta_ns
;
670 spin_unlock(&rq
->lock
);
672 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
675 * resched_task - mark a task 'to be rescheduled now'.
677 * On UP this means the setting of the need_resched flag, on SMP it
678 * might also involve a cross-CPU call to trigger the scheduler on
683 #ifndef tsk_is_polling
684 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
687 static void resched_task(struct task_struct
*p
)
691 assert_spin_locked(&task_rq(p
)->lock
);
693 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
696 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
699 if (cpu
== smp_processor_id())
702 /* NEED_RESCHED must be visible before we test polling */
704 if (!tsk_is_polling(p
))
705 smp_send_reschedule(cpu
);
708 static void resched_cpu(int cpu
)
710 struct rq
*rq
= cpu_rq(cpu
);
713 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
715 resched_task(cpu_curr(cpu
));
716 spin_unlock_irqrestore(&rq
->lock
, flags
);
719 static inline void resched_task(struct task_struct
*p
)
721 assert_spin_locked(&task_rq(p
)->lock
);
722 set_tsk_need_resched(p
);
726 #if BITS_PER_LONG == 32
727 # define WMULT_CONST (~0UL)
729 # define WMULT_CONST (1UL << 32)
732 #define WMULT_SHIFT 32
735 * Shift right and round:
737 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
740 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
741 struct load_weight
*lw
)
745 if (unlikely(!lw
->inv_weight
))
746 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
748 tmp
= (u64
)delta_exec
* weight
;
750 * Check whether we'd overflow the 64-bit multiplication:
752 if (unlikely(tmp
> WMULT_CONST
))
753 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
756 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
758 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
761 static inline unsigned long
762 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
764 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
767 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
772 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
778 * To aid in avoiding the subversion of "niceness" due to uneven distribution
779 * of tasks with abnormal "nice" values across CPUs the contribution that
780 * each task makes to its run queue's load is weighted according to its
781 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
782 * scaled version of the new time slice allocation that they receive on time
786 #define WEIGHT_IDLEPRIO 2
787 #define WMULT_IDLEPRIO (1 << 31)
790 * Nice levels are multiplicative, with a gentle 10% change for every
791 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
792 * nice 1, it will get ~10% less CPU time than another CPU-bound task
793 * that remained on nice 0.
795 * The "10% effect" is relative and cumulative: from _any_ nice level,
796 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
797 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
798 * If a task goes up by ~10% and another task goes down by ~10% then
799 * the relative distance between them is ~25%.)
801 static const int prio_to_weight
[40] = {
802 /* -20 */ 88761, 71755, 56483, 46273, 36291,
803 /* -15 */ 29154, 23254, 18705, 14949, 11916,
804 /* -10 */ 9548, 7620, 6100, 4904, 3906,
805 /* -5 */ 3121, 2501, 1991, 1586, 1277,
806 /* 0 */ 1024, 820, 655, 526, 423,
807 /* 5 */ 335, 272, 215, 172, 137,
808 /* 10 */ 110, 87, 70, 56, 45,
809 /* 15 */ 36, 29, 23, 18, 15,
813 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
815 * In cases where the weight does not change often, we can use the
816 * precalculated inverse to speed up arithmetics by turning divisions
817 * into multiplications:
819 static const u32 prio_to_wmult
[40] = {
820 /* -20 */ 48388, 59856, 76040, 92818, 118348,
821 /* -15 */ 147320, 184698, 229616, 287308, 360437,
822 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
823 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
824 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
825 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
826 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
827 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
830 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
833 * runqueue iterator, to support SMP load-balancing between different
834 * scheduling classes, without having to expose their internal data
835 * structures to the load-balancing proper:
839 struct task_struct
*(*start
)(void *);
840 struct task_struct
*(*next
)(void *);
843 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
844 unsigned long max_nr_move
, unsigned long max_load_move
,
845 struct sched_domain
*sd
, enum cpu_idle_type idle
,
846 int *all_pinned
, unsigned long *load_moved
,
847 int *this_best_prio
, struct rq_iterator
*iterator
);
849 #include "sched_stats.h"
850 #include "sched_rt.c"
851 #include "sched_fair.c"
852 #include "sched_idletask.c"
853 #ifdef CONFIG_SCHED_DEBUG
854 # include "sched_debug.c"
857 #define sched_class_highest (&rt_sched_class)
860 * Update delta_exec, delta_fair fields for rq.
862 * delta_fair clock advances at a rate inversely proportional to
863 * total load (rq->load.weight) on the runqueue, while
864 * delta_exec advances at the same rate as wall-clock (provided
867 * delta_exec / delta_fair is a measure of the (smoothened) load on this
868 * runqueue over any given interval. This (smoothened) load is used
869 * during load balance.
871 * This function is called /before/ updating rq->load
872 * and when switching tasks.
874 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
876 update_load_add(&rq
->load
, p
->se
.load
.weight
);
879 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
881 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
884 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
890 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
896 static void set_load_weight(struct task_struct
*p
)
898 if (task_has_rt_policy(p
)) {
899 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
900 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
905 * SCHED_IDLE tasks get minimal weight:
907 if (p
->policy
== SCHED_IDLE
) {
908 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
909 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
913 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
914 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
917 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
919 sched_info_queued(p
);
920 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
924 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
926 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
931 * __normal_prio - return the priority that is based on the static prio
933 static inline int __normal_prio(struct task_struct
*p
)
935 return p
->static_prio
;
939 * Calculate the expected normal priority: i.e. priority
940 * without taking RT-inheritance into account. Might be
941 * boosted by interactivity modifiers. Changes upon fork,
942 * setprio syscalls, and whenever the interactivity
943 * estimator recalculates.
945 static inline int normal_prio(struct task_struct
*p
)
949 if (task_has_rt_policy(p
))
950 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
952 prio
= __normal_prio(p
);
957 * Calculate the current priority, i.e. the priority
958 * taken into account by the scheduler. This value might
959 * be boosted by RT tasks, or might be boosted by
960 * interactivity modifiers. Will be RT if the task got
961 * RT-boosted. If not then it returns p->normal_prio.
963 static int effective_prio(struct task_struct
*p
)
965 p
->normal_prio
= normal_prio(p
);
967 * If we are RT tasks or we were boosted to RT priority,
968 * keep the priority unchanged. Otherwise, update priority
969 * to the normal priority:
971 if (!rt_prio(p
->prio
))
972 return p
->normal_prio
;
977 * activate_task - move a task to the runqueue.
979 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
981 if (p
->state
== TASK_UNINTERRUPTIBLE
)
982 rq
->nr_uninterruptible
--;
984 enqueue_task(rq
, p
, wakeup
);
985 inc_nr_running(p
, rq
);
989 * activate_idle_task - move idle task to the _front_ of runqueue.
991 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
995 if (p
->state
== TASK_UNINTERRUPTIBLE
)
996 rq
->nr_uninterruptible
--;
998 enqueue_task(rq
, p
, 0);
999 inc_nr_running(p
, rq
);
1003 * deactivate_task - remove a task from the runqueue.
1005 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1007 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1008 rq
->nr_uninterruptible
++;
1010 dequeue_task(rq
, p
, sleep
);
1011 dec_nr_running(p
, rq
);
1015 * task_curr - is this task currently executing on a CPU?
1016 * @p: the task in question.
1018 inline int task_curr(const struct task_struct
*p
)
1020 return cpu_curr(task_cpu(p
)) == p
;
1023 /* Used instead of source_load when we know the type == 0 */
1024 unsigned long weighted_cpuload(const int cpu
)
1026 return cpu_rq(cpu
)->load
.weight
;
1029 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1032 task_thread_info(p
)->cpu
= cpu
;
1039 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1041 int old_cpu
= task_cpu(p
);
1042 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1045 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1047 #ifdef CONFIG_SCHEDSTATS
1048 if (p
->se
.wait_start
)
1049 p
->se
.wait_start
-= clock_offset
;
1050 if (p
->se
.sleep_start
)
1051 p
->se
.sleep_start
-= clock_offset
;
1052 if (p
->se
.block_start
)
1053 p
->se
.block_start
-= clock_offset
;
1055 if (likely(new_rq
->cfs
.min_vruntime
))
1056 p
->se
.vruntime
-= old_rq
->cfs
.min_vruntime
-
1057 new_rq
->cfs
.min_vruntime
;
1059 __set_task_cpu(p
, new_cpu
);
1062 struct migration_req
{
1063 struct list_head list
;
1065 struct task_struct
*task
;
1068 struct completion done
;
1072 * The task's runqueue lock must be held.
1073 * Returns true if you have to wait for migration thread.
1076 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1078 struct rq
*rq
= task_rq(p
);
1081 * If the task is not on a runqueue (and not running), then
1082 * it is sufficient to simply update the task's cpu field.
1084 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1085 set_task_cpu(p
, dest_cpu
);
1089 init_completion(&req
->done
);
1091 req
->dest_cpu
= dest_cpu
;
1092 list_add(&req
->list
, &rq
->migration_queue
);
1098 * wait_task_inactive - wait for a thread to unschedule.
1100 * The caller must ensure that the task *will* unschedule sometime soon,
1101 * else this function might spin for a *long* time. This function can't
1102 * be called with interrupts off, or it may introduce deadlock with
1103 * smp_call_function() if an IPI is sent by the same process we are
1104 * waiting to become inactive.
1106 void wait_task_inactive(struct task_struct
*p
)
1108 unsigned long flags
;
1114 * We do the initial early heuristics without holding
1115 * any task-queue locks at all. We'll only try to get
1116 * the runqueue lock when things look like they will
1122 * If the task is actively running on another CPU
1123 * still, just relax and busy-wait without holding
1126 * NOTE! Since we don't hold any locks, it's not
1127 * even sure that "rq" stays as the right runqueue!
1128 * But we don't care, since "task_running()" will
1129 * return false if the runqueue has changed and p
1130 * is actually now running somewhere else!
1132 while (task_running(rq
, p
))
1136 * Ok, time to look more closely! We need the rq
1137 * lock now, to be *sure*. If we're wrong, we'll
1138 * just go back and repeat.
1140 rq
= task_rq_lock(p
, &flags
);
1141 running
= task_running(rq
, p
);
1142 on_rq
= p
->se
.on_rq
;
1143 task_rq_unlock(rq
, &flags
);
1146 * Was it really running after all now that we
1147 * checked with the proper locks actually held?
1149 * Oops. Go back and try again..
1151 if (unlikely(running
)) {
1157 * It's not enough that it's not actively running,
1158 * it must be off the runqueue _entirely_, and not
1161 * So if it wa still runnable (but just not actively
1162 * running right now), it's preempted, and we should
1163 * yield - it could be a while.
1165 if (unlikely(on_rq
)) {
1171 * Ahh, all good. It wasn't running, and it wasn't
1172 * runnable, which means that it will never become
1173 * running in the future either. We're all done!
1178 * kick_process - kick a running thread to enter/exit the kernel
1179 * @p: the to-be-kicked thread
1181 * Cause a process which is running on another CPU to enter
1182 * kernel-mode, without any delay. (to get signals handled.)
1184 * NOTE: this function doesnt have to take the runqueue lock,
1185 * because all it wants to ensure is that the remote task enters
1186 * the kernel. If the IPI races and the task has been migrated
1187 * to another CPU then no harm is done and the purpose has been
1190 void kick_process(struct task_struct
*p
)
1196 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1197 smp_send_reschedule(cpu
);
1202 * Return a low guess at the load of a migration-source cpu weighted
1203 * according to the scheduling class and "nice" value.
1205 * We want to under-estimate the load of migration sources, to
1206 * balance conservatively.
1208 static inline unsigned long source_load(int cpu
, int type
)
1210 struct rq
*rq
= cpu_rq(cpu
);
1211 unsigned long total
= weighted_cpuload(cpu
);
1216 return min(rq
->cpu_load
[type
-1], total
);
1220 * Return a high guess at the load of a migration-target cpu weighted
1221 * according to the scheduling class and "nice" value.
1223 static inline unsigned long target_load(int cpu
, int type
)
1225 struct rq
*rq
= cpu_rq(cpu
);
1226 unsigned long total
= weighted_cpuload(cpu
);
1231 return max(rq
->cpu_load
[type
-1], total
);
1235 * Return the average load per task on the cpu's run queue
1237 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1239 struct rq
*rq
= cpu_rq(cpu
);
1240 unsigned long total
= weighted_cpuload(cpu
);
1241 unsigned long n
= rq
->nr_running
;
1243 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1247 * find_idlest_group finds and returns the least busy CPU group within the
1250 static struct sched_group
*
1251 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1253 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1254 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1255 int load_idx
= sd
->forkexec_idx
;
1256 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1259 unsigned long load
, avg_load
;
1263 /* Skip over this group if it has no CPUs allowed */
1264 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1267 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1269 /* Tally up the load of all CPUs in the group */
1272 for_each_cpu_mask(i
, group
->cpumask
) {
1273 /* Bias balancing toward cpus of our domain */
1275 load
= source_load(i
, load_idx
);
1277 load
= target_load(i
, load_idx
);
1282 /* Adjust by relative CPU power of the group */
1283 avg_load
= sg_div_cpu_power(group
,
1284 avg_load
* SCHED_LOAD_SCALE
);
1287 this_load
= avg_load
;
1289 } else if (avg_load
< min_load
) {
1290 min_load
= avg_load
;
1294 group
= group
->next
;
1295 } while (group
!= sd
->groups
);
1297 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1303 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1306 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1309 unsigned long load
, min_load
= ULONG_MAX
;
1313 /* Traverse only the allowed CPUs */
1314 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1316 for_each_cpu_mask(i
, tmp
) {
1317 load
= weighted_cpuload(i
);
1319 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1329 * sched_balance_self: balance the current task (running on cpu) in domains
1330 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1333 * Balance, ie. select the least loaded group.
1335 * Returns the target CPU number, or the same CPU if no balancing is needed.
1337 * preempt must be disabled.
1339 static int sched_balance_self(int cpu
, int flag
)
1341 struct task_struct
*t
= current
;
1342 struct sched_domain
*tmp
, *sd
= NULL
;
1344 for_each_domain(cpu
, tmp
) {
1346 * If power savings logic is enabled for a domain, stop there.
1348 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1350 if (tmp
->flags
& flag
)
1356 struct sched_group
*group
;
1357 int new_cpu
, weight
;
1359 if (!(sd
->flags
& flag
)) {
1365 group
= find_idlest_group(sd
, t
, cpu
);
1371 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1372 if (new_cpu
== -1 || new_cpu
== cpu
) {
1373 /* Now try balancing at a lower domain level of cpu */
1378 /* Now try balancing at a lower domain level of new_cpu */
1381 weight
= cpus_weight(span
);
1382 for_each_domain(cpu
, tmp
) {
1383 if (weight
<= cpus_weight(tmp
->span
))
1385 if (tmp
->flags
& flag
)
1388 /* while loop will break here if sd == NULL */
1394 #endif /* CONFIG_SMP */
1397 * wake_idle() will wake a task on an idle cpu if task->cpu is
1398 * not idle and an idle cpu is available. The span of cpus to
1399 * search starts with cpus closest then further out as needed,
1400 * so we always favor a closer, idle cpu.
1402 * Returns the CPU we should wake onto.
1404 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1405 static int wake_idle(int cpu
, struct task_struct
*p
)
1408 struct sched_domain
*sd
;
1412 * If it is idle, then it is the best cpu to run this task.
1414 * This cpu is also the best, if it has more than one task already.
1415 * Siblings must be also busy(in most cases) as they didn't already
1416 * pickup the extra load from this cpu and hence we need not check
1417 * sibling runqueue info. This will avoid the checks and cache miss
1418 * penalities associated with that.
1420 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1423 for_each_domain(cpu
, sd
) {
1424 if (sd
->flags
& SD_WAKE_IDLE
) {
1425 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1426 for_each_cpu_mask(i
, tmp
) {
1437 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1444 * try_to_wake_up - wake up a thread
1445 * @p: the to-be-woken-up thread
1446 * @state: the mask of task states that can be woken
1447 * @sync: do a synchronous wakeup?
1449 * Put it on the run-queue if it's not already there. The "current"
1450 * thread is always on the run-queue (except when the actual
1451 * re-schedule is in progress), and as such you're allowed to do
1452 * the simpler "current->state = TASK_RUNNING" to mark yourself
1453 * runnable without the overhead of this.
1455 * returns failure only if the task is already active.
1457 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1459 int cpu
, this_cpu
, success
= 0;
1460 unsigned long flags
;
1464 struct sched_domain
*sd
, *this_sd
= NULL
;
1465 unsigned long load
, this_load
;
1469 rq
= task_rq_lock(p
, &flags
);
1470 old_state
= p
->state
;
1471 if (!(old_state
& state
))
1478 this_cpu
= smp_processor_id();
1481 if (unlikely(task_running(rq
, p
)))
1486 schedstat_inc(rq
, ttwu_cnt
);
1487 if (cpu
== this_cpu
) {
1488 schedstat_inc(rq
, ttwu_local
);
1492 for_each_domain(this_cpu
, sd
) {
1493 if (cpu_isset(cpu
, sd
->span
)) {
1494 schedstat_inc(sd
, ttwu_wake_remote
);
1500 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1504 * Check for affine wakeup and passive balancing possibilities.
1507 int idx
= this_sd
->wake_idx
;
1508 unsigned int imbalance
;
1510 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1512 load
= source_load(cpu
, idx
);
1513 this_load
= target_load(this_cpu
, idx
);
1515 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1517 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1518 unsigned long tl
= this_load
;
1519 unsigned long tl_per_task
;
1521 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1524 * If sync wakeup then subtract the (maximum possible)
1525 * effect of the currently running task from the load
1526 * of the current CPU:
1529 tl
-= current
->se
.load
.weight
;
1532 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1533 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1535 * This domain has SD_WAKE_AFFINE and
1536 * p is cache cold in this domain, and
1537 * there is no bad imbalance.
1539 schedstat_inc(this_sd
, ttwu_move_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
);
1556 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1558 new_cpu
= wake_idle(new_cpu
, p
);
1559 if (new_cpu
!= cpu
) {
1560 set_task_cpu(p
, new_cpu
);
1561 task_rq_unlock(rq
, &flags
);
1562 /* might preempt at this point */
1563 rq
= task_rq_lock(p
, &flags
);
1564 old_state
= p
->state
;
1565 if (!(old_state
& state
))
1570 this_cpu
= smp_processor_id();
1575 #endif /* CONFIG_SMP */
1576 update_rq_clock(rq
);
1577 activate_task(rq
, p
, 1);
1579 * Sync wakeups (i.e. those types of wakeups where the waker
1580 * has indicated that it will leave the CPU in short order)
1581 * don't trigger a preemption, if the woken up task will run on
1582 * this cpu. (in this case the 'I will reschedule' promise of
1583 * the waker guarantees that the freshly woken up task is going
1584 * to be considered on this CPU.)
1586 if (!sync
|| cpu
!= this_cpu
)
1587 check_preempt_curr(rq
, p
);
1591 p
->state
= TASK_RUNNING
;
1593 task_rq_unlock(rq
, &flags
);
1598 int fastcall
wake_up_process(struct task_struct
*p
)
1600 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1601 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1603 EXPORT_SYMBOL(wake_up_process
);
1605 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1607 return try_to_wake_up(p
, state
, 0);
1611 * Perform scheduler related setup for a newly forked process p.
1612 * p is forked by current.
1614 * __sched_fork() is basic setup used by init_idle() too:
1616 static void __sched_fork(struct task_struct
*p
)
1618 p
->se
.exec_start
= 0;
1619 p
->se
.sum_exec_runtime
= 0;
1620 p
->se
.prev_sum_exec_runtime
= 0;
1621 p
->se
.last_min_vruntime
= 0;
1623 #ifdef CONFIG_SCHEDSTATS
1624 p
->se
.wait_start
= 0;
1625 p
->se
.sum_sleep_runtime
= 0;
1626 p
->se
.sleep_start
= 0;
1627 p
->se
.block_start
= 0;
1628 p
->se
.sleep_max
= 0;
1629 p
->se
.block_max
= 0;
1631 p
->se
.slice_max
= 0;
1635 INIT_LIST_HEAD(&p
->run_list
);
1638 #ifdef CONFIG_PREEMPT_NOTIFIERS
1639 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1643 * We mark the process as running here, but have not actually
1644 * inserted it onto the runqueue yet. This guarantees that
1645 * nobody will actually run it, and a signal or other external
1646 * event cannot wake it up and insert it on the runqueue either.
1648 p
->state
= TASK_RUNNING
;
1652 * fork()/clone()-time setup:
1654 void sched_fork(struct task_struct
*p
, int clone_flags
)
1656 int cpu
= get_cpu();
1661 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1663 __set_task_cpu(p
, cpu
);
1666 * Make sure we do not leak PI boosting priority to the child:
1668 p
->prio
= current
->normal_prio
;
1670 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1671 if (likely(sched_info_on()))
1672 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1674 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1677 #ifdef CONFIG_PREEMPT
1678 /* Want to start with kernel preemption disabled. */
1679 task_thread_info(p
)->preempt_count
= 1;
1685 * wake_up_new_task - wake up a newly created task for the first time.
1687 * This function will do some initial scheduler statistics housekeeping
1688 * that must be done for every newly created context, then puts the task
1689 * on the runqueue and wakes it.
1691 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1693 unsigned long flags
;
1697 rq
= task_rq_lock(p
, &flags
);
1698 BUG_ON(p
->state
!= TASK_RUNNING
);
1699 this_cpu
= smp_processor_id(); /* parent's CPU */
1700 update_rq_clock(rq
);
1702 p
->prio
= effective_prio(p
);
1704 if (rt_prio(p
->prio
))
1705 p
->sched_class
= &rt_sched_class
;
1707 p
->sched_class
= &fair_sched_class
;
1709 if (task_cpu(p
) != this_cpu
|| !p
->sched_class
->task_new
||
1710 !current
->se
.on_rq
) {
1711 activate_task(rq
, p
, 0);
1714 * Let the scheduling class do new task startup
1715 * management (if any):
1717 p
->sched_class
->task_new(rq
, p
);
1718 inc_nr_running(p
, rq
);
1720 check_preempt_curr(rq
, p
);
1721 task_rq_unlock(rq
, &flags
);
1724 #ifdef CONFIG_PREEMPT_NOTIFIERS
1727 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1728 * @notifier: notifier struct to register
1730 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1732 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1734 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1737 * preempt_notifier_unregister - no longer interested in preemption notifications
1738 * @notifier: notifier struct to unregister
1740 * This is safe to call from within a preemption notifier.
1742 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1744 hlist_del(¬ifier
->link
);
1746 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1748 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1750 struct preempt_notifier
*notifier
;
1751 struct hlist_node
*node
;
1753 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1754 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1758 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1759 struct task_struct
*next
)
1761 struct preempt_notifier
*notifier
;
1762 struct hlist_node
*node
;
1764 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1765 notifier
->ops
->sched_out(notifier
, next
);
1770 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1775 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1776 struct task_struct
*next
)
1783 * prepare_task_switch - prepare to switch tasks
1784 * @rq: the runqueue preparing to switch
1785 * @prev: the current task that is being switched out
1786 * @next: the task we are going to switch to.
1788 * This is called with the rq lock held and interrupts off. It must
1789 * be paired with a subsequent finish_task_switch after the context
1792 * prepare_task_switch sets up locking and calls architecture specific
1796 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1797 struct task_struct
*next
)
1799 fire_sched_out_preempt_notifiers(prev
, next
);
1800 prepare_lock_switch(rq
, next
);
1801 prepare_arch_switch(next
);
1805 * finish_task_switch - clean up after a task-switch
1806 * @rq: runqueue associated with task-switch
1807 * @prev: the thread we just switched away from.
1809 * finish_task_switch must be called after the context switch, paired
1810 * with a prepare_task_switch call before the context switch.
1811 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1812 * and do any other architecture-specific cleanup actions.
1814 * Note that we may have delayed dropping an mm in context_switch(). If
1815 * so, we finish that here outside of the runqueue lock. (Doing it
1816 * with the lock held can cause deadlocks; see schedule() for
1819 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1820 __releases(rq
->lock
)
1822 struct mm_struct
*mm
= rq
->prev_mm
;
1828 * A task struct has one reference for the use as "current".
1829 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1830 * schedule one last time. The schedule call will never return, and
1831 * the scheduled task must drop that reference.
1832 * The test for TASK_DEAD must occur while the runqueue locks are
1833 * still held, otherwise prev could be scheduled on another cpu, die
1834 * there before we look at prev->state, and then the reference would
1836 * Manfred Spraul <manfred@colorfullife.com>
1838 prev_state
= prev
->state
;
1839 finish_arch_switch(prev
);
1840 finish_lock_switch(rq
, prev
);
1841 fire_sched_in_preempt_notifiers(current
);
1844 if (unlikely(prev_state
== TASK_DEAD
)) {
1846 * Remove function-return probe instances associated with this
1847 * task and put them back on the free list.
1849 kprobe_flush_task(prev
);
1850 put_task_struct(prev
);
1855 * schedule_tail - first thing a freshly forked thread must call.
1856 * @prev: the thread we just switched away from.
1858 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1859 __releases(rq
->lock
)
1861 struct rq
*rq
= this_rq();
1863 finish_task_switch(rq
, prev
);
1864 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1865 /* In this case, finish_task_switch does not reenable preemption */
1868 if (current
->set_child_tid
)
1869 put_user(current
->pid
, current
->set_child_tid
);
1873 * context_switch - switch to the new MM and the new
1874 * thread's register state.
1877 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1878 struct task_struct
*next
)
1880 struct mm_struct
*mm
, *oldmm
;
1882 prepare_task_switch(rq
, prev
, next
);
1884 oldmm
= prev
->active_mm
;
1886 * For paravirt, this is coupled with an exit in switch_to to
1887 * combine the page table reload and the switch backend into
1890 arch_enter_lazy_cpu_mode();
1892 if (unlikely(!mm
)) {
1893 next
->active_mm
= oldmm
;
1894 atomic_inc(&oldmm
->mm_count
);
1895 enter_lazy_tlb(oldmm
, next
);
1897 switch_mm(oldmm
, mm
, next
);
1899 if (unlikely(!prev
->mm
)) {
1900 prev
->active_mm
= NULL
;
1901 rq
->prev_mm
= oldmm
;
1904 * Since the runqueue lock will be released by the next
1905 * task (which is an invalid locking op but in the case
1906 * of the scheduler it's an obvious special-case), so we
1907 * do an early lockdep release here:
1909 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1910 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1913 /* Here we just switch the register state and the stack. */
1914 switch_to(prev
, next
, prev
);
1918 * this_rq must be evaluated again because prev may have moved
1919 * CPUs since it called schedule(), thus the 'rq' on its stack
1920 * frame will be invalid.
1922 finish_task_switch(this_rq(), prev
);
1926 * nr_running, nr_uninterruptible and nr_context_switches:
1928 * externally visible scheduler statistics: current number of runnable
1929 * threads, current number of uninterruptible-sleeping threads, total
1930 * number of context switches performed since bootup.
1932 unsigned long nr_running(void)
1934 unsigned long i
, sum
= 0;
1936 for_each_online_cpu(i
)
1937 sum
+= cpu_rq(i
)->nr_running
;
1942 unsigned long nr_uninterruptible(void)
1944 unsigned long i
, sum
= 0;
1946 for_each_possible_cpu(i
)
1947 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1950 * Since we read the counters lockless, it might be slightly
1951 * inaccurate. Do not allow it to go below zero though:
1953 if (unlikely((long)sum
< 0))
1959 unsigned long long nr_context_switches(void)
1962 unsigned long long sum
= 0;
1964 for_each_possible_cpu(i
)
1965 sum
+= cpu_rq(i
)->nr_switches
;
1970 unsigned long nr_iowait(void)
1972 unsigned long i
, sum
= 0;
1974 for_each_possible_cpu(i
)
1975 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1980 unsigned long nr_active(void)
1982 unsigned long i
, running
= 0, uninterruptible
= 0;
1984 for_each_online_cpu(i
) {
1985 running
+= cpu_rq(i
)->nr_running
;
1986 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1989 if (unlikely((long)uninterruptible
< 0))
1990 uninterruptible
= 0;
1992 return running
+ uninterruptible
;
1996 * Update rq->cpu_load[] statistics. This function is usually called every
1997 * scheduler tick (TICK_NSEC).
1999 static void update_cpu_load(struct rq
*this_rq
)
2001 unsigned long this_load
= this_rq
->load
.weight
;
2004 this_rq
->nr_load_updates
++;
2006 /* Update our load: */
2007 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2008 unsigned long old_load
, new_load
;
2010 /* scale is effectively 1 << i now, and >> i divides by scale */
2012 old_load
= this_rq
->cpu_load
[i
];
2013 new_load
= this_load
;
2015 * Round up the averaging division if load is increasing. This
2016 * prevents us from getting stuck on 9 if the load is 10, for
2019 if (new_load
> old_load
)
2020 new_load
+= scale
-1;
2021 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2028 * double_rq_lock - safely lock two runqueues
2030 * Note this does not disable interrupts like task_rq_lock,
2031 * you need to do so manually before calling.
2033 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2034 __acquires(rq1
->lock
)
2035 __acquires(rq2
->lock
)
2037 BUG_ON(!irqs_disabled());
2039 spin_lock(&rq1
->lock
);
2040 __acquire(rq2
->lock
); /* Fake it out ;) */
2043 spin_lock(&rq1
->lock
);
2044 spin_lock(&rq2
->lock
);
2046 spin_lock(&rq2
->lock
);
2047 spin_lock(&rq1
->lock
);
2050 update_rq_clock(rq1
);
2051 update_rq_clock(rq2
);
2055 * double_rq_unlock - safely unlock two runqueues
2057 * Note this does not restore interrupts like task_rq_unlock,
2058 * you need to do so manually after calling.
2060 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2061 __releases(rq1
->lock
)
2062 __releases(rq2
->lock
)
2064 spin_unlock(&rq1
->lock
);
2066 spin_unlock(&rq2
->lock
);
2068 __release(rq2
->lock
);
2072 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2074 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2075 __releases(this_rq
->lock
)
2076 __acquires(busiest
->lock
)
2077 __acquires(this_rq
->lock
)
2079 if (unlikely(!irqs_disabled())) {
2080 /* printk() doesn't work good under rq->lock */
2081 spin_unlock(&this_rq
->lock
);
2084 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2085 if (busiest
< this_rq
) {
2086 spin_unlock(&this_rq
->lock
);
2087 spin_lock(&busiest
->lock
);
2088 spin_lock(&this_rq
->lock
);
2090 spin_lock(&busiest
->lock
);
2095 * If dest_cpu is allowed for this process, migrate the task to it.
2096 * This is accomplished by forcing the cpu_allowed mask to only
2097 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2098 * the cpu_allowed mask is restored.
2100 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2102 struct migration_req req
;
2103 unsigned long flags
;
2106 rq
= task_rq_lock(p
, &flags
);
2107 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2108 || unlikely(cpu_is_offline(dest_cpu
)))
2111 /* force the process onto the specified CPU */
2112 if (migrate_task(p
, dest_cpu
, &req
)) {
2113 /* Need to wait for migration thread (might exit: take ref). */
2114 struct task_struct
*mt
= rq
->migration_thread
;
2116 get_task_struct(mt
);
2117 task_rq_unlock(rq
, &flags
);
2118 wake_up_process(mt
);
2119 put_task_struct(mt
);
2120 wait_for_completion(&req
.done
);
2125 task_rq_unlock(rq
, &flags
);
2129 * sched_exec - execve() is a valuable balancing opportunity, because at
2130 * this point the task has the smallest effective memory and cache footprint.
2132 void sched_exec(void)
2134 int new_cpu
, this_cpu
= get_cpu();
2135 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2137 if (new_cpu
!= this_cpu
)
2138 sched_migrate_task(current
, new_cpu
);
2142 * pull_task - move a task from a remote runqueue to the local runqueue.
2143 * Both runqueues must be locked.
2145 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2146 struct rq
*this_rq
, int this_cpu
)
2148 deactivate_task(src_rq
, p
, 0);
2149 set_task_cpu(p
, this_cpu
);
2150 activate_task(this_rq
, p
, 0);
2152 * Note that idle threads have a prio of MAX_PRIO, for this test
2153 * to be always true for them.
2155 check_preempt_curr(this_rq
, p
);
2159 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2162 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2163 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2167 * We do not migrate tasks that are:
2168 * 1) running (obviously), or
2169 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2170 * 3) are cache-hot on their current CPU.
2172 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2176 if (task_running(rq
, p
))
2182 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2183 unsigned long max_nr_move
, unsigned long max_load_move
,
2184 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2185 int *all_pinned
, unsigned long *load_moved
,
2186 int *this_best_prio
, struct rq_iterator
*iterator
)
2188 int pulled
= 0, pinned
= 0, skip_for_load
;
2189 struct task_struct
*p
;
2190 long rem_load_move
= max_load_move
;
2192 if (max_nr_move
== 0 || max_load_move
== 0)
2198 * Start the load-balancing iterator:
2200 p
= iterator
->start(iterator
->arg
);
2205 * To help distribute high priority tasks accross CPUs we don't
2206 * skip a task if it will be the highest priority task (i.e. smallest
2207 * prio value) on its new queue regardless of its load weight
2209 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2210 SCHED_LOAD_SCALE_FUZZ
;
2211 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2212 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2213 p
= iterator
->next(iterator
->arg
);
2217 pull_task(busiest
, p
, this_rq
, this_cpu
);
2219 rem_load_move
-= p
->se
.load
.weight
;
2222 * We only want to steal up to the prescribed number of tasks
2223 * and the prescribed amount of weighted load.
2225 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2226 if (p
->prio
< *this_best_prio
)
2227 *this_best_prio
= p
->prio
;
2228 p
= iterator
->next(iterator
->arg
);
2233 * Right now, this is the only place pull_task() is called,
2234 * so we can safely collect pull_task() stats here rather than
2235 * inside pull_task().
2237 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2240 *all_pinned
= pinned
;
2241 *load_moved
= max_load_move
- rem_load_move
;
2246 * move_tasks tries to move up to max_load_move weighted load from busiest to
2247 * this_rq, as part of a balancing operation within domain "sd".
2248 * Returns 1 if successful and 0 otherwise.
2250 * Called with both runqueues locked.
2252 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2253 unsigned long max_load_move
,
2254 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2257 struct sched_class
*class = sched_class_highest
;
2258 unsigned long total_load_moved
= 0;
2259 int this_best_prio
= this_rq
->curr
->prio
;
2263 class->load_balance(this_rq
, this_cpu
, busiest
,
2264 ULONG_MAX
, max_load_move
- total_load_moved
,
2265 sd
, idle
, all_pinned
, &this_best_prio
);
2266 class = class->next
;
2267 } while (class && max_load_move
> total_load_moved
);
2269 return total_load_moved
> 0;
2273 * move_one_task tries to move exactly one task from busiest to this_rq, as
2274 * part of active balancing operations within "domain".
2275 * Returns 1 if successful and 0 otherwise.
2277 * Called with both runqueues locked.
2279 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2280 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2282 struct sched_class
*class;
2283 int this_best_prio
= MAX_PRIO
;
2285 for (class = sched_class_highest
; class; class = class->next
)
2286 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2287 1, ULONG_MAX
, sd
, idle
, NULL
,
2295 * find_busiest_group finds and returns the busiest CPU group within the
2296 * domain. It calculates and returns the amount of weighted load which
2297 * should be moved to restore balance via the imbalance parameter.
2299 static struct sched_group
*
2300 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2301 unsigned long *imbalance
, enum cpu_idle_type idle
,
2302 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2304 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2305 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2306 unsigned long max_pull
;
2307 unsigned long busiest_load_per_task
, busiest_nr_running
;
2308 unsigned long this_load_per_task
, this_nr_running
;
2310 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2311 int power_savings_balance
= 1;
2312 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2313 unsigned long min_nr_running
= ULONG_MAX
;
2314 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2317 max_load
= this_load
= total_load
= total_pwr
= 0;
2318 busiest_load_per_task
= busiest_nr_running
= 0;
2319 this_load_per_task
= this_nr_running
= 0;
2320 if (idle
== CPU_NOT_IDLE
)
2321 load_idx
= sd
->busy_idx
;
2322 else if (idle
== CPU_NEWLY_IDLE
)
2323 load_idx
= sd
->newidle_idx
;
2325 load_idx
= sd
->idle_idx
;
2328 unsigned long load
, group_capacity
;
2331 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2332 unsigned long sum_nr_running
, sum_weighted_load
;
2334 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2337 balance_cpu
= first_cpu(group
->cpumask
);
2339 /* Tally up the load of all CPUs in the group */
2340 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2342 for_each_cpu_mask(i
, group
->cpumask
) {
2345 if (!cpu_isset(i
, *cpus
))
2350 if (*sd_idle
&& rq
->nr_running
)
2353 /* Bias balancing toward cpus of our domain */
2355 if (idle_cpu(i
) && !first_idle_cpu
) {
2360 load
= target_load(i
, load_idx
);
2362 load
= source_load(i
, load_idx
);
2365 sum_nr_running
+= rq
->nr_running
;
2366 sum_weighted_load
+= weighted_cpuload(i
);
2370 * First idle cpu or the first cpu(busiest) in this sched group
2371 * is eligible for doing load balancing at this and above
2372 * domains. In the newly idle case, we will allow all the cpu's
2373 * to do the newly idle load balance.
2375 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2376 balance_cpu
!= this_cpu
&& balance
) {
2381 total_load
+= avg_load
;
2382 total_pwr
+= group
->__cpu_power
;
2384 /* Adjust by relative CPU power of the group */
2385 avg_load
= sg_div_cpu_power(group
,
2386 avg_load
* SCHED_LOAD_SCALE
);
2388 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2391 this_load
= avg_load
;
2393 this_nr_running
= sum_nr_running
;
2394 this_load_per_task
= sum_weighted_load
;
2395 } else if (avg_load
> max_load
&&
2396 sum_nr_running
> group_capacity
) {
2397 max_load
= avg_load
;
2399 busiest_nr_running
= sum_nr_running
;
2400 busiest_load_per_task
= sum_weighted_load
;
2403 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2405 * Busy processors will not participate in power savings
2408 if (idle
== CPU_NOT_IDLE
||
2409 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2413 * If the local group is idle or completely loaded
2414 * no need to do power savings balance at this domain
2416 if (local_group
&& (this_nr_running
>= group_capacity
||
2418 power_savings_balance
= 0;
2421 * If a group is already running at full capacity or idle,
2422 * don't include that group in power savings calculations
2424 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2429 * Calculate the group which has the least non-idle load.
2430 * This is the group from where we need to pick up the load
2433 if ((sum_nr_running
< min_nr_running
) ||
2434 (sum_nr_running
== min_nr_running
&&
2435 first_cpu(group
->cpumask
) <
2436 first_cpu(group_min
->cpumask
))) {
2438 min_nr_running
= sum_nr_running
;
2439 min_load_per_task
= sum_weighted_load
/
2444 * Calculate the group which is almost near its
2445 * capacity but still has some space to pick up some load
2446 * from other group and save more power
2448 if (sum_nr_running
<= group_capacity
- 1) {
2449 if (sum_nr_running
> leader_nr_running
||
2450 (sum_nr_running
== leader_nr_running
&&
2451 first_cpu(group
->cpumask
) >
2452 first_cpu(group_leader
->cpumask
))) {
2453 group_leader
= group
;
2454 leader_nr_running
= sum_nr_running
;
2459 group
= group
->next
;
2460 } while (group
!= sd
->groups
);
2462 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2465 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2467 if (this_load
>= avg_load
||
2468 100*max_load
<= sd
->imbalance_pct
*this_load
)
2471 busiest_load_per_task
/= busiest_nr_running
;
2473 * We're trying to get all the cpus to the average_load, so we don't
2474 * want to push ourselves above the average load, nor do we wish to
2475 * reduce the max loaded cpu below the average load, as either of these
2476 * actions would just result in more rebalancing later, and ping-pong
2477 * tasks around. Thus we look for the minimum possible imbalance.
2478 * Negative imbalances (*we* are more loaded than anyone else) will
2479 * be counted as no imbalance for these purposes -- we can't fix that
2480 * by pulling tasks to us. Be careful of negative numbers as they'll
2481 * appear as very large values with unsigned longs.
2483 if (max_load
<= busiest_load_per_task
)
2487 * In the presence of smp nice balancing, certain scenarios can have
2488 * max load less than avg load(as we skip the groups at or below
2489 * its cpu_power, while calculating max_load..)
2491 if (max_load
< avg_load
) {
2493 goto small_imbalance
;
2496 /* Don't want to pull so many tasks that a group would go idle */
2497 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2499 /* How much load to actually move to equalise the imbalance */
2500 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2501 (avg_load
- this_load
) * this->__cpu_power
)
2505 * if *imbalance is less than the average load per runnable task
2506 * there is no gaurantee that any tasks will be moved so we'll have
2507 * a think about bumping its value to force at least one task to be
2510 if (*imbalance
< busiest_load_per_task
) {
2511 unsigned long tmp
, pwr_now
, pwr_move
;
2515 pwr_move
= pwr_now
= 0;
2517 if (this_nr_running
) {
2518 this_load_per_task
/= this_nr_running
;
2519 if (busiest_load_per_task
> this_load_per_task
)
2522 this_load_per_task
= SCHED_LOAD_SCALE
;
2524 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2525 busiest_load_per_task
* imbn
) {
2526 *imbalance
= busiest_load_per_task
;
2531 * OK, we don't have enough imbalance to justify moving tasks,
2532 * however we may be able to increase total CPU power used by
2536 pwr_now
+= busiest
->__cpu_power
*
2537 min(busiest_load_per_task
, max_load
);
2538 pwr_now
+= this->__cpu_power
*
2539 min(this_load_per_task
, this_load
);
2540 pwr_now
/= SCHED_LOAD_SCALE
;
2542 /* Amount of load we'd subtract */
2543 tmp
= sg_div_cpu_power(busiest
,
2544 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2546 pwr_move
+= busiest
->__cpu_power
*
2547 min(busiest_load_per_task
, max_load
- tmp
);
2549 /* Amount of load we'd add */
2550 if (max_load
* busiest
->__cpu_power
<
2551 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2552 tmp
= sg_div_cpu_power(this,
2553 max_load
* busiest
->__cpu_power
);
2555 tmp
= sg_div_cpu_power(this,
2556 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2557 pwr_move
+= this->__cpu_power
*
2558 min(this_load_per_task
, this_load
+ tmp
);
2559 pwr_move
/= SCHED_LOAD_SCALE
;
2561 /* Move if we gain throughput */
2562 if (pwr_move
> pwr_now
)
2563 *imbalance
= busiest_load_per_task
;
2569 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2570 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2573 if (this == group_leader
&& group_leader
!= group_min
) {
2574 *imbalance
= min_load_per_task
;
2584 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2587 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2588 unsigned long imbalance
, cpumask_t
*cpus
)
2590 struct rq
*busiest
= NULL
, *rq
;
2591 unsigned long max_load
= 0;
2594 for_each_cpu_mask(i
, group
->cpumask
) {
2597 if (!cpu_isset(i
, *cpus
))
2601 wl
= weighted_cpuload(i
);
2603 if (rq
->nr_running
== 1 && wl
> imbalance
)
2606 if (wl
> max_load
) {
2616 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2617 * so long as it is large enough.
2619 #define MAX_PINNED_INTERVAL 512
2622 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2623 * tasks if there is an imbalance.
2625 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2626 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2629 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2630 struct sched_group
*group
;
2631 unsigned long imbalance
;
2633 cpumask_t cpus
= CPU_MASK_ALL
;
2634 unsigned long flags
;
2637 * When power savings policy is enabled for the parent domain, idle
2638 * sibling can pick up load irrespective of busy siblings. In this case,
2639 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2640 * portraying it as CPU_NOT_IDLE.
2642 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2643 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2646 schedstat_inc(sd
, lb_cnt
[idle
]);
2649 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2656 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2660 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2662 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2666 BUG_ON(busiest
== this_rq
);
2668 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2671 if (busiest
->nr_running
> 1) {
2673 * Attempt to move tasks. If find_busiest_group has found
2674 * an imbalance but busiest->nr_running <= 1, the group is
2675 * still unbalanced. ld_moved simply stays zero, so it is
2676 * correctly treated as an imbalance.
2678 local_irq_save(flags
);
2679 double_rq_lock(this_rq
, busiest
);
2680 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2681 imbalance
, sd
, idle
, &all_pinned
);
2682 double_rq_unlock(this_rq
, busiest
);
2683 local_irq_restore(flags
);
2686 * some other cpu did the load balance for us.
2688 if (ld_moved
&& this_cpu
!= smp_processor_id())
2689 resched_cpu(this_cpu
);
2691 /* All tasks on this runqueue were pinned by CPU affinity */
2692 if (unlikely(all_pinned
)) {
2693 cpu_clear(cpu_of(busiest
), cpus
);
2694 if (!cpus_empty(cpus
))
2701 schedstat_inc(sd
, lb_failed
[idle
]);
2702 sd
->nr_balance_failed
++;
2704 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2706 spin_lock_irqsave(&busiest
->lock
, flags
);
2708 /* don't kick the migration_thread, if the curr
2709 * task on busiest cpu can't be moved to this_cpu
2711 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2712 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2714 goto out_one_pinned
;
2717 if (!busiest
->active_balance
) {
2718 busiest
->active_balance
= 1;
2719 busiest
->push_cpu
= this_cpu
;
2722 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2724 wake_up_process(busiest
->migration_thread
);
2727 * We've kicked active balancing, reset the failure
2730 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2733 sd
->nr_balance_failed
= 0;
2735 if (likely(!active_balance
)) {
2736 /* We were unbalanced, so reset the balancing interval */
2737 sd
->balance_interval
= sd
->min_interval
;
2740 * If we've begun active balancing, start to back off. This
2741 * case may not be covered by the all_pinned logic if there
2742 * is only 1 task on the busy runqueue (because we don't call
2745 if (sd
->balance_interval
< sd
->max_interval
)
2746 sd
->balance_interval
*= 2;
2749 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2750 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2755 schedstat_inc(sd
, lb_balanced
[idle
]);
2757 sd
->nr_balance_failed
= 0;
2760 /* tune up the balancing interval */
2761 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2762 (sd
->balance_interval
< sd
->max_interval
))
2763 sd
->balance_interval
*= 2;
2765 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2766 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2772 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2773 * tasks if there is an imbalance.
2775 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2776 * this_rq is locked.
2779 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2781 struct sched_group
*group
;
2782 struct rq
*busiest
= NULL
;
2783 unsigned long imbalance
;
2787 cpumask_t cpus
= CPU_MASK_ALL
;
2790 * When power savings policy is enabled for the parent domain, idle
2791 * sibling can pick up load irrespective of busy siblings. In this case,
2792 * let the state of idle sibling percolate up as IDLE, instead of
2793 * portraying it as CPU_NOT_IDLE.
2795 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2796 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2799 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2801 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2802 &sd_idle
, &cpus
, NULL
);
2804 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2808 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2811 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2815 BUG_ON(busiest
== this_rq
);
2817 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2820 if (busiest
->nr_running
> 1) {
2821 /* Attempt to move tasks */
2822 double_lock_balance(this_rq
, busiest
);
2823 /* this_rq->clock is already updated */
2824 update_rq_clock(busiest
);
2825 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2826 imbalance
, sd
, CPU_NEWLY_IDLE
,
2828 spin_unlock(&busiest
->lock
);
2830 if (unlikely(all_pinned
)) {
2831 cpu_clear(cpu_of(busiest
), cpus
);
2832 if (!cpus_empty(cpus
))
2838 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2839 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2840 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2843 sd
->nr_balance_failed
= 0;
2848 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2849 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2850 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2852 sd
->nr_balance_failed
= 0;
2858 * idle_balance is called by schedule() if this_cpu is about to become
2859 * idle. Attempts to pull tasks from other CPUs.
2861 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2863 struct sched_domain
*sd
;
2864 int pulled_task
= -1;
2865 unsigned long next_balance
= jiffies
+ HZ
;
2867 for_each_domain(this_cpu
, sd
) {
2868 unsigned long interval
;
2870 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2873 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2874 /* If we've pulled tasks over stop searching: */
2875 pulled_task
= load_balance_newidle(this_cpu
,
2878 interval
= msecs_to_jiffies(sd
->balance_interval
);
2879 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2880 next_balance
= sd
->last_balance
+ interval
;
2884 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2886 * We are going idle. next_balance may be set based on
2887 * a busy processor. So reset next_balance.
2889 this_rq
->next_balance
= next_balance
;
2894 * active_load_balance is run by migration threads. It pushes running tasks
2895 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2896 * running on each physical CPU where possible, and avoids physical /
2897 * logical imbalances.
2899 * Called with busiest_rq locked.
2901 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2903 int target_cpu
= busiest_rq
->push_cpu
;
2904 struct sched_domain
*sd
;
2905 struct rq
*target_rq
;
2907 /* Is there any task to move? */
2908 if (busiest_rq
->nr_running
<= 1)
2911 target_rq
= cpu_rq(target_cpu
);
2914 * This condition is "impossible", if it occurs
2915 * we need to fix it. Originally reported by
2916 * Bjorn Helgaas on a 128-cpu setup.
2918 BUG_ON(busiest_rq
== target_rq
);
2920 /* move a task from busiest_rq to target_rq */
2921 double_lock_balance(busiest_rq
, target_rq
);
2922 update_rq_clock(busiest_rq
);
2923 update_rq_clock(target_rq
);
2925 /* Search for an sd spanning us and the target CPU. */
2926 for_each_domain(target_cpu
, sd
) {
2927 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2928 cpu_isset(busiest_cpu
, sd
->span
))
2933 schedstat_inc(sd
, alb_cnt
);
2935 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2937 schedstat_inc(sd
, alb_pushed
);
2939 schedstat_inc(sd
, alb_failed
);
2941 spin_unlock(&target_rq
->lock
);
2946 atomic_t load_balancer
;
2948 } nohz ____cacheline_aligned
= {
2949 .load_balancer
= ATOMIC_INIT(-1),
2950 .cpu_mask
= CPU_MASK_NONE
,
2954 * This routine will try to nominate the ilb (idle load balancing)
2955 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2956 * load balancing on behalf of all those cpus. If all the cpus in the system
2957 * go into this tickless mode, then there will be no ilb owner (as there is
2958 * no need for one) and all the cpus will sleep till the next wakeup event
2961 * For the ilb owner, tick is not stopped. And this tick will be used
2962 * for idle load balancing. ilb owner will still be part of
2965 * While stopping the tick, this cpu will become the ilb owner if there
2966 * is no other owner. And will be the owner till that cpu becomes busy
2967 * or if all cpus in the system stop their ticks at which point
2968 * there is no need for ilb owner.
2970 * When the ilb owner becomes busy, it nominates another owner, during the
2971 * next busy scheduler_tick()
2973 int select_nohz_load_balancer(int stop_tick
)
2975 int cpu
= smp_processor_id();
2978 cpu_set(cpu
, nohz
.cpu_mask
);
2979 cpu_rq(cpu
)->in_nohz_recently
= 1;
2982 * If we are going offline and still the leader, give up!
2984 if (cpu_is_offline(cpu
) &&
2985 atomic_read(&nohz
.load_balancer
) == cpu
) {
2986 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2991 /* time for ilb owner also to sleep */
2992 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2993 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2994 atomic_set(&nohz
.load_balancer
, -1);
2998 if (atomic_read(&nohz
.load_balancer
) == -1) {
2999 /* make me the ilb owner */
3000 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3002 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3005 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3008 cpu_clear(cpu
, nohz
.cpu_mask
);
3010 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3011 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3018 static DEFINE_SPINLOCK(balancing
);
3021 * It checks each scheduling domain to see if it is due to be balanced,
3022 * and initiates a balancing operation if so.
3024 * Balancing parameters are set up in arch_init_sched_domains.
3026 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3029 struct rq
*rq
= cpu_rq(cpu
);
3030 unsigned long interval
;
3031 struct sched_domain
*sd
;
3032 /* Earliest time when we have to do rebalance again */
3033 unsigned long next_balance
= jiffies
+ 60*HZ
;
3034 int update_next_balance
= 0;
3036 for_each_domain(cpu
, sd
) {
3037 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3040 interval
= sd
->balance_interval
;
3041 if (idle
!= CPU_IDLE
)
3042 interval
*= sd
->busy_factor
;
3044 /* scale ms to jiffies */
3045 interval
= msecs_to_jiffies(interval
);
3046 if (unlikely(!interval
))
3048 if (interval
> HZ
*NR_CPUS
/10)
3049 interval
= HZ
*NR_CPUS
/10;
3052 if (sd
->flags
& SD_SERIALIZE
) {
3053 if (!spin_trylock(&balancing
))
3057 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3058 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3060 * We've pulled tasks over so either we're no
3061 * longer idle, or one of our SMT siblings is
3064 idle
= CPU_NOT_IDLE
;
3066 sd
->last_balance
= jiffies
;
3068 if (sd
->flags
& SD_SERIALIZE
)
3069 spin_unlock(&balancing
);
3071 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3072 next_balance
= sd
->last_balance
+ interval
;
3073 update_next_balance
= 1;
3077 * Stop the load balance at this level. There is another
3078 * CPU in our sched group which is doing load balancing more
3086 * next_balance will be updated only when there is a need.
3087 * When the cpu is attached to null domain for ex, it will not be
3090 if (likely(update_next_balance
))
3091 rq
->next_balance
= next_balance
;
3095 * run_rebalance_domains is triggered when needed from the scheduler tick.
3096 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3097 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3099 static void run_rebalance_domains(struct softirq_action
*h
)
3101 int this_cpu
= smp_processor_id();
3102 struct rq
*this_rq
= cpu_rq(this_cpu
);
3103 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3104 CPU_IDLE
: CPU_NOT_IDLE
;
3106 rebalance_domains(this_cpu
, idle
);
3110 * If this cpu is the owner for idle load balancing, then do the
3111 * balancing on behalf of the other idle cpus whose ticks are
3114 if (this_rq
->idle_at_tick
&&
3115 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3116 cpumask_t cpus
= nohz
.cpu_mask
;
3120 cpu_clear(this_cpu
, cpus
);
3121 for_each_cpu_mask(balance_cpu
, cpus
) {
3123 * If this cpu gets work to do, stop the load balancing
3124 * work being done for other cpus. Next load
3125 * balancing owner will pick it up.
3130 rebalance_domains(balance_cpu
, CPU_IDLE
);
3132 rq
= cpu_rq(balance_cpu
);
3133 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3134 this_rq
->next_balance
= rq
->next_balance
;
3141 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3143 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3144 * idle load balancing owner or decide to stop the periodic load balancing,
3145 * if the whole system is idle.
3147 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3151 * If we were in the nohz mode recently and busy at the current
3152 * scheduler tick, then check if we need to nominate new idle
3155 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3156 rq
->in_nohz_recently
= 0;
3158 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3159 cpu_clear(cpu
, nohz
.cpu_mask
);
3160 atomic_set(&nohz
.load_balancer
, -1);
3163 if (atomic_read(&nohz
.load_balancer
) == -1) {
3165 * simple selection for now: Nominate the
3166 * first cpu in the nohz list to be the next
3169 * TBD: Traverse the sched domains and nominate
3170 * the nearest cpu in the nohz.cpu_mask.
3172 int ilb
= first_cpu(nohz
.cpu_mask
);
3180 * If this cpu is idle and doing idle load balancing for all the
3181 * cpus with ticks stopped, is it time for that to stop?
3183 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3184 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3190 * If this cpu is idle and the idle load balancing is done by
3191 * someone else, then no need raise the SCHED_SOFTIRQ
3193 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3194 cpu_isset(cpu
, nohz
.cpu_mask
))
3197 if (time_after_eq(jiffies
, rq
->next_balance
))
3198 raise_softirq(SCHED_SOFTIRQ
);
3201 #else /* CONFIG_SMP */
3204 * on UP we do not need to balance between CPUs:
3206 static inline void idle_balance(int cpu
, struct rq
*rq
)
3210 /* Avoid "used but not defined" warning on UP */
3211 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3212 unsigned long max_nr_move
, unsigned long max_load_move
,
3213 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3214 int *all_pinned
, unsigned long *load_moved
,
3215 int *this_best_prio
, struct rq_iterator
*iterator
)
3224 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3226 EXPORT_PER_CPU_SYMBOL(kstat
);
3229 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3230 * that have not yet been banked in case the task is currently running.
3232 unsigned long long task_sched_runtime(struct task_struct
*p
)
3234 unsigned long flags
;
3238 rq
= task_rq_lock(p
, &flags
);
3239 ns
= p
->se
.sum_exec_runtime
;
3240 if (rq
->curr
== p
) {
3241 update_rq_clock(rq
);
3242 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3243 if ((s64
)delta_exec
> 0)
3246 task_rq_unlock(rq
, &flags
);
3252 * Account user cpu time to a process.
3253 * @p: the process that the cpu time gets accounted to
3254 * @hardirq_offset: the offset to subtract from hardirq_count()
3255 * @cputime: the cpu time spent in user space since the last update
3257 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3259 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3262 p
->utime
= cputime_add(p
->utime
, cputime
);
3264 /* Add user time to cpustat. */
3265 tmp
= cputime_to_cputime64(cputime
);
3266 if (TASK_NICE(p
) > 0)
3267 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3269 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3273 * Account system cpu time to a process.
3274 * @p: the process that the cpu time gets accounted to
3275 * @hardirq_offset: the offset to subtract from hardirq_count()
3276 * @cputime: the cpu time spent in kernel space since the last update
3278 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3281 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3282 struct rq
*rq
= this_rq();
3285 p
->stime
= cputime_add(p
->stime
, cputime
);
3287 /* Add system time to cpustat. */
3288 tmp
= cputime_to_cputime64(cputime
);
3289 if (hardirq_count() - hardirq_offset
)
3290 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3291 else if (softirq_count())
3292 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3293 else if (p
!= rq
->idle
)
3294 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3295 else if (atomic_read(&rq
->nr_iowait
) > 0)
3296 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3298 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3299 /* Account for system time used */
3300 acct_update_integrals(p
);
3304 * Account for involuntary wait time.
3305 * @p: the process from which the cpu time has been stolen
3306 * @steal: the cpu time spent in involuntary wait
3308 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3310 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3311 cputime64_t tmp
= cputime_to_cputime64(steal
);
3312 struct rq
*rq
= this_rq();
3314 if (p
== rq
->idle
) {
3315 p
->stime
= cputime_add(p
->stime
, steal
);
3316 if (atomic_read(&rq
->nr_iowait
) > 0)
3317 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3319 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3321 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3325 * This function gets called by the timer code, with HZ frequency.
3326 * We call it with interrupts disabled.
3328 * It also gets called by the fork code, when changing the parent's
3331 void scheduler_tick(void)
3333 int cpu
= smp_processor_id();
3334 struct rq
*rq
= cpu_rq(cpu
);
3335 struct task_struct
*curr
= rq
->curr
;
3336 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3338 spin_lock(&rq
->lock
);
3339 __update_rq_clock(rq
);
3341 * Let rq->clock advance by at least TICK_NSEC:
3343 if (unlikely(rq
->clock
< next_tick
))
3344 rq
->clock
= next_tick
;
3345 rq
->tick_timestamp
= rq
->clock
;
3346 update_cpu_load(rq
);
3347 if (curr
!= rq
->idle
) /* FIXME: needed? */
3348 curr
->sched_class
->task_tick(rq
, curr
);
3349 spin_unlock(&rq
->lock
);
3352 rq
->idle_at_tick
= idle_cpu(cpu
);
3353 trigger_load_balance(rq
, cpu
);
3357 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3359 void fastcall
add_preempt_count(int val
)
3364 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3366 preempt_count() += val
;
3368 * Spinlock count overflowing soon?
3370 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3373 EXPORT_SYMBOL(add_preempt_count
);
3375 void fastcall
sub_preempt_count(int val
)
3380 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3383 * Is the spinlock portion underflowing?
3385 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3386 !(preempt_count() & PREEMPT_MASK
)))
3389 preempt_count() -= val
;
3391 EXPORT_SYMBOL(sub_preempt_count
);
3396 * Print scheduling while atomic bug:
3398 static noinline
void __schedule_bug(struct task_struct
*prev
)
3400 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3401 prev
->comm
, preempt_count(), prev
->pid
);
3402 debug_show_held_locks(prev
);
3403 if (irqs_disabled())
3404 print_irqtrace_events(prev
);
3409 * Various schedule()-time debugging checks and statistics:
3411 static inline void schedule_debug(struct task_struct
*prev
)
3414 * Test if we are atomic. Since do_exit() needs to call into
3415 * schedule() atomically, we ignore that path for now.
3416 * Otherwise, whine if we are scheduling when we should not be.
3418 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3419 __schedule_bug(prev
);
3421 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3423 schedstat_inc(this_rq(), sched_cnt
);
3424 #ifdef CONFIG_SCHEDSTATS
3425 if (unlikely(prev
->lock_depth
>= 0)) {
3426 schedstat_inc(this_rq(), bkl_cnt
);
3427 schedstat_inc(prev
, sched_info
.bkl_cnt
);
3433 * Pick up the highest-prio task:
3435 static inline struct task_struct
*
3436 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3438 struct sched_class
*class;
3439 struct task_struct
*p
;
3442 * Optimization: we know that if all tasks are in
3443 * the fair class we can call that function directly:
3445 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3446 p
= fair_sched_class
.pick_next_task(rq
);
3451 class = sched_class_highest
;
3453 p
= class->pick_next_task(rq
);
3457 * Will never be NULL as the idle class always
3458 * returns a non-NULL p:
3460 class = class->next
;
3465 * schedule() is the main scheduler function.
3467 asmlinkage
void __sched
schedule(void)
3469 struct task_struct
*prev
, *next
;
3476 cpu
= smp_processor_id();
3480 switch_count
= &prev
->nivcsw
;
3482 release_kernel_lock(prev
);
3483 need_resched_nonpreemptible
:
3485 schedule_debug(prev
);
3487 spin_lock_irq(&rq
->lock
);
3488 clear_tsk_need_resched(prev
);
3489 __update_rq_clock(rq
);
3491 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3492 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3493 unlikely(signal_pending(prev
)))) {
3494 prev
->state
= TASK_RUNNING
;
3496 deactivate_task(rq
, prev
, 1);
3498 switch_count
= &prev
->nvcsw
;
3501 if (unlikely(!rq
->nr_running
))
3502 idle_balance(cpu
, rq
);
3504 prev
->sched_class
->put_prev_task(rq
, prev
);
3505 next
= pick_next_task(rq
, prev
);
3507 sched_info_switch(prev
, next
);
3509 if (likely(prev
!= next
)) {
3514 context_switch(rq
, prev
, next
); /* unlocks the rq */
3516 spin_unlock_irq(&rq
->lock
);
3518 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3519 cpu
= smp_processor_id();
3521 goto need_resched_nonpreemptible
;
3523 preempt_enable_no_resched();
3524 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3527 EXPORT_SYMBOL(schedule
);
3529 #ifdef CONFIG_PREEMPT
3531 * this is the entry point to schedule() from in-kernel preemption
3532 * off of preempt_enable. Kernel preemptions off return from interrupt
3533 * occur there and call schedule directly.
3535 asmlinkage
void __sched
preempt_schedule(void)
3537 struct thread_info
*ti
= current_thread_info();
3538 #ifdef CONFIG_PREEMPT_BKL
3539 struct task_struct
*task
= current
;
3540 int saved_lock_depth
;
3543 * If there is a non-zero preempt_count or interrupts are disabled,
3544 * we do not want to preempt the current task. Just return..
3546 if (likely(ti
->preempt_count
|| irqs_disabled()))
3550 add_preempt_count(PREEMPT_ACTIVE
);
3552 * We keep the big kernel semaphore locked, but we
3553 * clear ->lock_depth so that schedule() doesnt
3554 * auto-release the semaphore:
3556 #ifdef CONFIG_PREEMPT_BKL
3557 saved_lock_depth
= task
->lock_depth
;
3558 task
->lock_depth
= -1;
3561 #ifdef CONFIG_PREEMPT_BKL
3562 task
->lock_depth
= saved_lock_depth
;
3564 sub_preempt_count(PREEMPT_ACTIVE
);
3566 /* we could miss a preemption opportunity between schedule and now */
3568 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3571 EXPORT_SYMBOL(preempt_schedule
);
3574 * this is the entry point to schedule() from kernel preemption
3575 * off of irq context.
3576 * Note, that this is called and return with irqs disabled. This will
3577 * protect us against recursive calling from irq.
3579 asmlinkage
void __sched
preempt_schedule_irq(void)
3581 struct thread_info
*ti
= current_thread_info();
3582 #ifdef CONFIG_PREEMPT_BKL
3583 struct task_struct
*task
= current
;
3584 int saved_lock_depth
;
3586 /* Catch callers which need to be fixed */
3587 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3590 add_preempt_count(PREEMPT_ACTIVE
);
3592 * We keep the big kernel semaphore locked, but we
3593 * clear ->lock_depth so that schedule() doesnt
3594 * auto-release the semaphore:
3596 #ifdef CONFIG_PREEMPT_BKL
3597 saved_lock_depth
= task
->lock_depth
;
3598 task
->lock_depth
= -1;
3602 local_irq_disable();
3603 #ifdef CONFIG_PREEMPT_BKL
3604 task
->lock_depth
= saved_lock_depth
;
3606 sub_preempt_count(PREEMPT_ACTIVE
);
3608 /* we could miss a preemption opportunity between schedule and now */
3610 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3614 #endif /* CONFIG_PREEMPT */
3616 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3619 return try_to_wake_up(curr
->private, mode
, sync
);
3621 EXPORT_SYMBOL(default_wake_function
);
3624 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3625 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3626 * number) then we wake all the non-exclusive tasks and one exclusive task.
3628 * There are circumstances in which we can try to wake a task which has already
3629 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3630 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3632 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3633 int nr_exclusive
, int sync
, void *key
)
3635 wait_queue_t
*curr
, *next
;
3637 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3638 unsigned flags
= curr
->flags
;
3640 if (curr
->func(curr
, mode
, sync
, key
) &&
3641 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3647 * __wake_up - wake up threads blocked on a waitqueue.
3649 * @mode: which threads
3650 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3651 * @key: is directly passed to the wakeup function
3653 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3654 int nr_exclusive
, void *key
)
3656 unsigned long flags
;
3658 spin_lock_irqsave(&q
->lock
, flags
);
3659 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3660 spin_unlock_irqrestore(&q
->lock
, flags
);
3662 EXPORT_SYMBOL(__wake_up
);
3665 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3667 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3669 __wake_up_common(q
, mode
, 1, 0, NULL
);
3673 * __wake_up_sync - wake up threads blocked on a waitqueue.
3675 * @mode: which threads
3676 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3678 * The sync wakeup differs that the waker knows that it will schedule
3679 * away soon, so while the target thread will be woken up, it will not
3680 * be migrated to another CPU - ie. the two threads are 'synchronized'
3681 * with each other. This can prevent needless bouncing between CPUs.
3683 * On UP it can prevent extra preemption.
3686 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3688 unsigned long flags
;
3694 if (unlikely(!nr_exclusive
))
3697 spin_lock_irqsave(&q
->lock
, flags
);
3698 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3699 spin_unlock_irqrestore(&q
->lock
, flags
);
3701 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3703 void fastcall
complete(struct completion
*x
)
3705 unsigned long flags
;
3707 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3709 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3711 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3713 EXPORT_SYMBOL(complete
);
3715 void fastcall
complete_all(struct completion
*x
)
3717 unsigned long flags
;
3719 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3720 x
->done
+= UINT_MAX
/2;
3721 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3723 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3725 EXPORT_SYMBOL(complete_all
);
3727 void fastcall __sched
wait_for_completion(struct completion
*x
)
3731 spin_lock_irq(&x
->wait
.lock
);
3733 DECLARE_WAITQUEUE(wait
, current
);
3735 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3736 __add_wait_queue_tail(&x
->wait
, &wait
);
3738 __set_current_state(TASK_UNINTERRUPTIBLE
);
3739 spin_unlock_irq(&x
->wait
.lock
);
3741 spin_lock_irq(&x
->wait
.lock
);
3743 __remove_wait_queue(&x
->wait
, &wait
);
3746 spin_unlock_irq(&x
->wait
.lock
);
3748 EXPORT_SYMBOL(wait_for_completion
);
3750 unsigned long fastcall __sched
3751 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3755 spin_lock_irq(&x
->wait
.lock
);
3757 DECLARE_WAITQUEUE(wait
, current
);
3759 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3760 __add_wait_queue_tail(&x
->wait
, &wait
);
3762 __set_current_state(TASK_UNINTERRUPTIBLE
);
3763 spin_unlock_irq(&x
->wait
.lock
);
3764 timeout
= schedule_timeout(timeout
);
3765 spin_lock_irq(&x
->wait
.lock
);
3767 __remove_wait_queue(&x
->wait
, &wait
);
3771 __remove_wait_queue(&x
->wait
, &wait
);
3775 spin_unlock_irq(&x
->wait
.lock
);
3778 EXPORT_SYMBOL(wait_for_completion_timeout
);
3780 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3786 spin_lock_irq(&x
->wait
.lock
);
3788 DECLARE_WAITQUEUE(wait
, current
);
3790 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3791 __add_wait_queue_tail(&x
->wait
, &wait
);
3793 if (signal_pending(current
)) {
3795 __remove_wait_queue(&x
->wait
, &wait
);
3798 __set_current_state(TASK_INTERRUPTIBLE
);
3799 spin_unlock_irq(&x
->wait
.lock
);
3801 spin_lock_irq(&x
->wait
.lock
);
3803 __remove_wait_queue(&x
->wait
, &wait
);
3807 spin_unlock_irq(&x
->wait
.lock
);
3811 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3813 unsigned long fastcall __sched
3814 wait_for_completion_interruptible_timeout(struct completion
*x
,
3815 unsigned long timeout
)
3819 spin_lock_irq(&x
->wait
.lock
);
3821 DECLARE_WAITQUEUE(wait
, current
);
3823 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3824 __add_wait_queue_tail(&x
->wait
, &wait
);
3826 if (signal_pending(current
)) {
3827 timeout
= -ERESTARTSYS
;
3828 __remove_wait_queue(&x
->wait
, &wait
);
3831 __set_current_state(TASK_INTERRUPTIBLE
);
3832 spin_unlock_irq(&x
->wait
.lock
);
3833 timeout
= schedule_timeout(timeout
);
3834 spin_lock_irq(&x
->wait
.lock
);
3836 __remove_wait_queue(&x
->wait
, &wait
);
3840 __remove_wait_queue(&x
->wait
, &wait
);
3844 spin_unlock_irq(&x
->wait
.lock
);
3847 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3850 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3852 spin_lock_irqsave(&q
->lock
, *flags
);
3853 __add_wait_queue(q
, wait
);
3854 spin_unlock(&q
->lock
);
3858 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3860 spin_lock_irq(&q
->lock
);
3861 __remove_wait_queue(q
, wait
);
3862 spin_unlock_irqrestore(&q
->lock
, *flags
);
3865 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3867 unsigned long flags
;
3870 init_waitqueue_entry(&wait
, current
);
3872 current
->state
= TASK_INTERRUPTIBLE
;
3874 sleep_on_head(q
, &wait
, &flags
);
3876 sleep_on_tail(q
, &wait
, &flags
);
3878 EXPORT_SYMBOL(interruptible_sleep_on
);
3881 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3883 unsigned long flags
;
3886 init_waitqueue_entry(&wait
, current
);
3888 current
->state
= TASK_INTERRUPTIBLE
;
3890 sleep_on_head(q
, &wait
, &flags
);
3891 timeout
= schedule_timeout(timeout
);
3892 sleep_on_tail(q
, &wait
, &flags
);
3896 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3898 void __sched
sleep_on(wait_queue_head_t
*q
)
3900 unsigned long flags
;
3903 init_waitqueue_entry(&wait
, current
);
3905 current
->state
= TASK_UNINTERRUPTIBLE
;
3907 sleep_on_head(q
, &wait
, &flags
);
3909 sleep_on_tail(q
, &wait
, &flags
);
3911 EXPORT_SYMBOL(sleep_on
);
3913 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3915 unsigned long flags
;
3918 init_waitqueue_entry(&wait
, current
);
3920 current
->state
= TASK_UNINTERRUPTIBLE
;
3922 sleep_on_head(q
, &wait
, &flags
);
3923 timeout
= schedule_timeout(timeout
);
3924 sleep_on_tail(q
, &wait
, &flags
);
3928 EXPORT_SYMBOL(sleep_on_timeout
);
3930 #ifdef CONFIG_RT_MUTEXES
3933 * rt_mutex_setprio - set the current priority of a task
3935 * @prio: prio value (kernel-internal form)
3937 * This function changes the 'effective' priority of a task. It does
3938 * not touch ->normal_prio like __setscheduler().
3940 * Used by the rt_mutex code to implement priority inheritance logic.
3942 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3944 unsigned long flags
;
3945 int oldprio
, on_rq
, running
;
3948 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3950 rq
= task_rq_lock(p
, &flags
);
3951 update_rq_clock(rq
);
3954 on_rq
= p
->se
.on_rq
;
3955 running
= task_running(rq
, p
);
3957 dequeue_task(rq
, p
, 0);
3959 p
->sched_class
->put_prev_task(rq
, p
);
3963 p
->sched_class
= &rt_sched_class
;
3965 p
->sched_class
= &fair_sched_class
;
3971 p
->sched_class
->set_curr_task(rq
);
3972 enqueue_task(rq
, p
, 0);
3974 * Reschedule if we are currently running on this runqueue and
3975 * our priority decreased, or if we are not currently running on
3976 * this runqueue and our priority is higher than the current's
3979 if (p
->prio
> oldprio
)
3980 resched_task(rq
->curr
);
3982 check_preempt_curr(rq
, p
);
3985 task_rq_unlock(rq
, &flags
);
3990 void set_user_nice(struct task_struct
*p
, long nice
)
3992 int old_prio
, delta
, on_rq
;
3993 unsigned long flags
;
3996 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3999 * We have to be careful, if called from sys_setpriority(),
4000 * the task might be in the middle of scheduling on another CPU.
4002 rq
= task_rq_lock(p
, &flags
);
4003 update_rq_clock(rq
);
4005 * The RT priorities are set via sched_setscheduler(), but we still
4006 * allow the 'normal' nice value to be set - but as expected
4007 * it wont have any effect on scheduling until the task is
4008 * SCHED_FIFO/SCHED_RR:
4010 if (task_has_rt_policy(p
)) {
4011 p
->static_prio
= NICE_TO_PRIO(nice
);
4014 on_rq
= p
->se
.on_rq
;
4016 dequeue_task(rq
, p
, 0);
4020 p
->static_prio
= NICE_TO_PRIO(nice
);
4023 p
->prio
= effective_prio(p
);
4024 delta
= p
->prio
- old_prio
;
4027 enqueue_task(rq
, p
, 0);
4030 * If the task increased its priority or is running and
4031 * lowered its priority, then reschedule its CPU:
4033 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4034 resched_task(rq
->curr
);
4037 task_rq_unlock(rq
, &flags
);
4039 EXPORT_SYMBOL(set_user_nice
);
4042 * can_nice - check if a task can reduce its nice value
4046 int can_nice(const struct task_struct
*p
, const int nice
)
4048 /* convert nice value [19,-20] to rlimit style value [1,40] */
4049 int nice_rlim
= 20 - nice
;
4051 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4052 capable(CAP_SYS_NICE
));
4055 #ifdef __ARCH_WANT_SYS_NICE
4058 * sys_nice - change the priority of the current process.
4059 * @increment: priority increment
4061 * sys_setpriority is a more generic, but much slower function that
4062 * does similar things.
4064 asmlinkage
long sys_nice(int increment
)
4069 * Setpriority might change our priority at the same moment.
4070 * We don't have to worry. Conceptually one call occurs first
4071 * and we have a single winner.
4073 if (increment
< -40)
4078 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4084 if (increment
< 0 && !can_nice(current
, nice
))
4087 retval
= security_task_setnice(current
, nice
);
4091 set_user_nice(current
, nice
);
4098 * task_prio - return the priority value of a given task.
4099 * @p: the task in question.
4101 * This is the priority value as seen by users in /proc.
4102 * RT tasks are offset by -200. Normal tasks are centered
4103 * around 0, value goes from -16 to +15.
4105 int task_prio(const struct task_struct
*p
)
4107 return p
->prio
- MAX_RT_PRIO
;
4111 * task_nice - return the nice value of a given task.
4112 * @p: the task in question.
4114 int task_nice(const struct task_struct
*p
)
4116 return TASK_NICE(p
);
4118 EXPORT_SYMBOL_GPL(task_nice
);
4121 * idle_cpu - is a given cpu idle currently?
4122 * @cpu: the processor in question.
4124 int idle_cpu(int cpu
)
4126 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4130 * idle_task - return the idle task for a given cpu.
4131 * @cpu: the processor in question.
4133 struct task_struct
*idle_task(int cpu
)
4135 return cpu_rq(cpu
)->idle
;
4139 * find_process_by_pid - find a process with a matching PID value.
4140 * @pid: the pid in question.
4142 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4144 return pid
? find_task_by_pid(pid
) : current
;
4147 /* Actually do priority change: must hold rq lock. */
4149 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4151 BUG_ON(p
->se
.on_rq
);
4154 switch (p
->policy
) {
4158 p
->sched_class
= &fair_sched_class
;
4162 p
->sched_class
= &rt_sched_class
;
4166 p
->rt_priority
= prio
;
4167 p
->normal_prio
= normal_prio(p
);
4168 /* we are holding p->pi_lock already */
4169 p
->prio
= rt_mutex_getprio(p
);
4174 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4175 * @p: the task in question.
4176 * @policy: new policy.
4177 * @param: structure containing the new RT priority.
4179 * NOTE that the task may be already dead.
4181 int sched_setscheduler(struct task_struct
*p
, int policy
,
4182 struct sched_param
*param
)
4184 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4185 unsigned long flags
;
4188 /* may grab non-irq protected spin_locks */
4189 BUG_ON(in_interrupt());
4191 /* double check policy once rq lock held */
4193 policy
= oldpolicy
= p
->policy
;
4194 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4195 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4196 policy
!= SCHED_IDLE
)
4199 * Valid priorities for SCHED_FIFO and SCHED_RR are
4200 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4201 * SCHED_BATCH and SCHED_IDLE is 0.
4203 if (param
->sched_priority
< 0 ||
4204 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4205 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4207 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4211 * Allow unprivileged RT tasks to decrease priority:
4213 if (!capable(CAP_SYS_NICE
)) {
4214 if (rt_policy(policy
)) {
4215 unsigned long rlim_rtprio
;
4217 if (!lock_task_sighand(p
, &flags
))
4219 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4220 unlock_task_sighand(p
, &flags
);
4222 /* can't set/change the rt policy */
4223 if (policy
!= p
->policy
&& !rlim_rtprio
)
4226 /* can't increase priority */
4227 if (param
->sched_priority
> p
->rt_priority
&&
4228 param
->sched_priority
> rlim_rtprio
)
4232 * Like positive nice levels, dont allow tasks to
4233 * move out of SCHED_IDLE either:
4235 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4238 /* can't change other user's priorities */
4239 if ((current
->euid
!= p
->euid
) &&
4240 (current
->euid
!= p
->uid
))
4244 retval
= security_task_setscheduler(p
, policy
, param
);
4248 * make sure no PI-waiters arrive (or leave) while we are
4249 * changing the priority of the task:
4251 spin_lock_irqsave(&p
->pi_lock
, flags
);
4253 * To be able to change p->policy safely, the apropriate
4254 * runqueue lock must be held.
4256 rq
= __task_rq_lock(p
);
4257 /* recheck policy now with rq lock held */
4258 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4259 policy
= oldpolicy
= -1;
4260 __task_rq_unlock(rq
);
4261 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4264 update_rq_clock(rq
);
4265 on_rq
= p
->se
.on_rq
;
4266 running
= task_running(rq
, p
);
4268 deactivate_task(rq
, p
, 0);
4270 p
->sched_class
->put_prev_task(rq
, p
);
4274 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4278 p
->sched_class
->set_curr_task(rq
);
4279 activate_task(rq
, p
, 0);
4281 * Reschedule if we are currently running on this runqueue and
4282 * our priority decreased, or if we are not currently running on
4283 * this runqueue and our priority is higher than the current's
4286 if (p
->prio
> oldprio
)
4287 resched_task(rq
->curr
);
4289 check_preempt_curr(rq
, p
);
4292 __task_rq_unlock(rq
);
4293 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4295 rt_mutex_adjust_pi(p
);
4299 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4302 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4304 struct sched_param lparam
;
4305 struct task_struct
*p
;
4308 if (!param
|| pid
< 0)
4310 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4315 p
= find_process_by_pid(pid
);
4317 retval
= sched_setscheduler(p
, policy
, &lparam
);
4324 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4325 * @pid: the pid in question.
4326 * @policy: new policy.
4327 * @param: structure containing the new RT priority.
4329 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4330 struct sched_param __user
*param
)
4332 /* negative values for policy are not valid */
4336 return do_sched_setscheduler(pid
, policy
, param
);
4340 * sys_sched_setparam - set/change the RT priority of a thread
4341 * @pid: the pid in question.
4342 * @param: structure containing the new RT priority.
4344 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4346 return do_sched_setscheduler(pid
, -1, param
);
4350 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4351 * @pid: the pid in question.
4353 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4355 struct task_struct
*p
;
4356 int retval
= -EINVAL
;
4362 read_lock(&tasklist_lock
);
4363 p
= find_process_by_pid(pid
);
4365 retval
= security_task_getscheduler(p
);
4369 read_unlock(&tasklist_lock
);
4376 * sys_sched_getscheduler - get the RT priority of a thread
4377 * @pid: the pid in question.
4378 * @param: structure containing the RT priority.
4380 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4382 struct sched_param lp
;
4383 struct task_struct
*p
;
4384 int retval
= -EINVAL
;
4386 if (!param
|| pid
< 0)
4389 read_lock(&tasklist_lock
);
4390 p
= find_process_by_pid(pid
);
4395 retval
= security_task_getscheduler(p
);
4399 lp
.sched_priority
= p
->rt_priority
;
4400 read_unlock(&tasklist_lock
);
4403 * This one might sleep, we cannot do it with a spinlock held ...
4405 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4411 read_unlock(&tasklist_lock
);
4415 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4417 cpumask_t cpus_allowed
;
4418 struct task_struct
*p
;
4421 mutex_lock(&sched_hotcpu_mutex
);
4422 read_lock(&tasklist_lock
);
4424 p
= find_process_by_pid(pid
);
4426 read_unlock(&tasklist_lock
);
4427 mutex_unlock(&sched_hotcpu_mutex
);
4432 * It is not safe to call set_cpus_allowed with the
4433 * tasklist_lock held. We will bump the task_struct's
4434 * usage count and then drop tasklist_lock.
4437 read_unlock(&tasklist_lock
);
4440 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4441 !capable(CAP_SYS_NICE
))
4444 retval
= security_task_setscheduler(p
, 0, NULL
);
4448 cpus_allowed
= cpuset_cpus_allowed(p
);
4449 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4450 retval
= set_cpus_allowed(p
, new_mask
);
4454 mutex_unlock(&sched_hotcpu_mutex
);
4458 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4459 cpumask_t
*new_mask
)
4461 if (len
< sizeof(cpumask_t
)) {
4462 memset(new_mask
, 0, sizeof(cpumask_t
));
4463 } else if (len
> sizeof(cpumask_t
)) {
4464 len
= sizeof(cpumask_t
);
4466 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4470 * sys_sched_setaffinity - set the cpu affinity of a process
4471 * @pid: pid of the process
4472 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4473 * @user_mask_ptr: user-space pointer to the new cpu mask
4475 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4476 unsigned long __user
*user_mask_ptr
)
4481 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4485 return sched_setaffinity(pid
, new_mask
);
4489 * Represents all cpu's present in the system
4490 * In systems capable of hotplug, this map could dynamically grow
4491 * as new cpu's are detected in the system via any platform specific
4492 * method, such as ACPI for e.g.
4495 cpumask_t cpu_present_map __read_mostly
;
4496 EXPORT_SYMBOL(cpu_present_map
);
4499 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4500 EXPORT_SYMBOL(cpu_online_map
);
4502 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4503 EXPORT_SYMBOL(cpu_possible_map
);
4506 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4508 struct task_struct
*p
;
4511 mutex_lock(&sched_hotcpu_mutex
);
4512 read_lock(&tasklist_lock
);
4515 p
= find_process_by_pid(pid
);
4519 retval
= security_task_getscheduler(p
);
4523 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4526 read_unlock(&tasklist_lock
);
4527 mutex_unlock(&sched_hotcpu_mutex
);
4533 * sys_sched_getaffinity - get the cpu affinity of a process
4534 * @pid: pid of the process
4535 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4536 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4538 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4539 unsigned long __user
*user_mask_ptr
)
4544 if (len
< sizeof(cpumask_t
))
4547 ret
= sched_getaffinity(pid
, &mask
);
4551 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4554 return sizeof(cpumask_t
);
4558 * sys_sched_yield - yield the current processor to other threads.
4560 * This function yields the current CPU to other tasks. If there are no
4561 * other threads running on this CPU then this function will return.
4563 asmlinkage
long sys_sched_yield(void)
4565 struct rq
*rq
= this_rq_lock();
4567 schedstat_inc(rq
, yld_cnt
);
4568 current
->sched_class
->yield_task(rq
);
4571 * Since we are going to call schedule() anyway, there's
4572 * no need to preempt or enable interrupts:
4574 __release(rq
->lock
);
4575 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4576 _raw_spin_unlock(&rq
->lock
);
4577 preempt_enable_no_resched();
4584 static void __cond_resched(void)
4586 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4587 __might_sleep(__FILE__
, __LINE__
);
4590 * The BKS might be reacquired before we have dropped
4591 * PREEMPT_ACTIVE, which could trigger a second
4592 * cond_resched() call.
4595 add_preempt_count(PREEMPT_ACTIVE
);
4597 sub_preempt_count(PREEMPT_ACTIVE
);
4598 } while (need_resched());
4601 int __sched
cond_resched(void)
4603 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4604 system_state
== SYSTEM_RUNNING
) {
4610 EXPORT_SYMBOL(cond_resched
);
4613 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4614 * call schedule, and on return reacquire the lock.
4616 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4617 * operations here to prevent schedule() from being called twice (once via
4618 * spin_unlock(), once by hand).
4620 int cond_resched_lock(spinlock_t
*lock
)
4624 if (need_lockbreak(lock
)) {
4630 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4631 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4632 _raw_spin_unlock(lock
);
4633 preempt_enable_no_resched();
4640 EXPORT_SYMBOL(cond_resched_lock
);
4642 int __sched
cond_resched_softirq(void)
4644 BUG_ON(!in_softirq());
4646 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4654 EXPORT_SYMBOL(cond_resched_softirq
);
4657 * yield - yield the current processor to other threads.
4659 * This is a shortcut for kernel-space yielding - it marks the
4660 * thread runnable and calls sys_sched_yield().
4662 void __sched
yield(void)
4664 set_current_state(TASK_RUNNING
);
4667 EXPORT_SYMBOL(yield
);
4670 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4671 * that process accounting knows that this is a task in IO wait state.
4673 * But don't do that if it is a deliberate, throttling IO wait (this task
4674 * has set its backing_dev_info: the queue against which it should throttle)
4676 void __sched
io_schedule(void)
4678 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4680 delayacct_blkio_start();
4681 atomic_inc(&rq
->nr_iowait
);
4683 atomic_dec(&rq
->nr_iowait
);
4684 delayacct_blkio_end();
4686 EXPORT_SYMBOL(io_schedule
);
4688 long __sched
io_schedule_timeout(long timeout
)
4690 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4693 delayacct_blkio_start();
4694 atomic_inc(&rq
->nr_iowait
);
4695 ret
= schedule_timeout(timeout
);
4696 atomic_dec(&rq
->nr_iowait
);
4697 delayacct_blkio_end();
4702 * sys_sched_get_priority_max - return maximum RT priority.
4703 * @policy: scheduling class.
4705 * this syscall returns the maximum rt_priority that can be used
4706 * by a given scheduling class.
4708 asmlinkage
long sys_sched_get_priority_max(int policy
)
4715 ret
= MAX_USER_RT_PRIO
-1;
4727 * sys_sched_get_priority_min - return minimum RT priority.
4728 * @policy: scheduling class.
4730 * this syscall returns the minimum rt_priority that can be used
4731 * by a given scheduling class.
4733 asmlinkage
long sys_sched_get_priority_min(int policy
)
4751 * sys_sched_rr_get_interval - return the default timeslice of a process.
4752 * @pid: pid of the process.
4753 * @interval: userspace pointer to the timeslice value.
4755 * this syscall writes the default timeslice value of a given process
4756 * into the user-space timespec buffer. A value of '0' means infinity.
4759 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4761 struct task_struct
*p
;
4762 int retval
= -EINVAL
;
4769 read_lock(&tasklist_lock
);
4770 p
= find_process_by_pid(pid
);
4774 retval
= security_task_getscheduler(p
);
4778 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4779 0 : static_prio_timeslice(p
->static_prio
), &t
);
4780 read_unlock(&tasklist_lock
);
4781 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4785 read_unlock(&tasklist_lock
);
4789 static const char stat_nam
[] = "RSDTtZX";
4791 static void show_task(struct task_struct
*p
)
4793 unsigned long free
= 0;
4796 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4797 printk("%-13.13s %c", p
->comm
,
4798 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4799 #if BITS_PER_LONG == 32
4800 if (state
== TASK_RUNNING
)
4801 printk(" running ");
4803 printk(" %08lx ", thread_saved_pc(p
));
4805 if (state
== TASK_RUNNING
)
4806 printk(" running task ");
4808 printk(" %016lx ", thread_saved_pc(p
));
4810 #ifdef CONFIG_DEBUG_STACK_USAGE
4812 unsigned long *n
= end_of_stack(p
);
4815 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4818 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4820 if (state
!= TASK_RUNNING
)
4821 show_stack(p
, NULL
);
4824 void show_state_filter(unsigned long state_filter
)
4826 struct task_struct
*g
, *p
;
4828 #if BITS_PER_LONG == 32
4830 " task PC stack pid father\n");
4833 " task PC stack pid father\n");
4835 read_lock(&tasklist_lock
);
4836 do_each_thread(g
, p
) {
4838 * reset the NMI-timeout, listing all files on a slow
4839 * console might take alot of time:
4841 touch_nmi_watchdog();
4842 if (!state_filter
|| (p
->state
& state_filter
))
4844 } while_each_thread(g
, p
);
4846 touch_all_softlockup_watchdogs();
4848 #ifdef CONFIG_SCHED_DEBUG
4849 sysrq_sched_debug_show();
4851 read_unlock(&tasklist_lock
);
4853 * Only show locks if all tasks are dumped:
4855 if (state_filter
== -1)
4856 debug_show_all_locks();
4859 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4861 idle
->sched_class
= &idle_sched_class
;
4865 * init_idle - set up an idle thread for a given CPU
4866 * @idle: task in question
4867 * @cpu: cpu the idle task belongs to
4869 * NOTE: this function does not set the idle thread's NEED_RESCHED
4870 * flag, to make booting more robust.
4872 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4874 struct rq
*rq
= cpu_rq(cpu
);
4875 unsigned long flags
;
4878 idle
->se
.exec_start
= sched_clock();
4880 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4881 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4882 __set_task_cpu(idle
, cpu
);
4884 spin_lock_irqsave(&rq
->lock
, flags
);
4885 rq
->curr
= rq
->idle
= idle
;
4886 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4889 spin_unlock_irqrestore(&rq
->lock
, flags
);
4891 /* Set the preempt count _outside_ the spinlocks! */
4892 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4893 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4895 task_thread_info(idle
)->preempt_count
= 0;
4898 * The idle tasks have their own, simple scheduling class:
4900 idle
->sched_class
= &idle_sched_class
;
4904 * In a system that switches off the HZ timer nohz_cpu_mask
4905 * indicates which cpus entered this state. This is used
4906 * in the rcu update to wait only for active cpus. For system
4907 * which do not switch off the HZ timer nohz_cpu_mask should
4908 * always be CPU_MASK_NONE.
4910 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4914 * This is how migration works:
4916 * 1) we queue a struct migration_req structure in the source CPU's
4917 * runqueue and wake up that CPU's migration thread.
4918 * 2) we down() the locked semaphore => thread blocks.
4919 * 3) migration thread wakes up (implicitly it forces the migrated
4920 * thread off the CPU)
4921 * 4) it gets the migration request and checks whether the migrated
4922 * task is still in the wrong runqueue.
4923 * 5) if it's in the wrong runqueue then the migration thread removes
4924 * it and puts it into the right queue.
4925 * 6) migration thread up()s the semaphore.
4926 * 7) we wake up and the migration is done.
4930 * Change a given task's CPU affinity. Migrate the thread to a
4931 * proper CPU and schedule it away if the CPU it's executing on
4932 * is removed from the allowed bitmask.
4934 * NOTE: the caller must have a valid reference to the task, the
4935 * task must not exit() & deallocate itself prematurely. The
4936 * call is not atomic; no spinlocks may be held.
4938 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4940 struct migration_req req
;
4941 unsigned long flags
;
4945 rq
= task_rq_lock(p
, &flags
);
4946 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4951 p
->cpus_allowed
= new_mask
;
4952 /* Can the task run on the task's current CPU? If so, we're done */
4953 if (cpu_isset(task_cpu(p
), new_mask
))
4956 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4957 /* Need help from migration thread: drop lock and wait. */
4958 task_rq_unlock(rq
, &flags
);
4959 wake_up_process(rq
->migration_thread
);
4960 wait_for_completion(&req
.done
);
4961 tlb_migrate_finish(p
->mm
);
4965 task_rq_unlock(rq
, &flags
);
4969 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4972 * Move (not current) task off this cpu, onto dest cpu. We're doing
4973 * this because either it can't run here any more (set_cpus_allowed()
4974 * away from this CPU, or CPU going down), or because we're
4975 * attempting to rebalance this task on exec (sched_exec).
4977 * So we race with normal scheduler movements, but that's OK, as long
4978 * as the task is no longer on this CPU.
4980 * Returns non-zero if task was successfully migrated.
4982 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4984 struct rq
*rq_dest
, *rq_src
;
4987 if (unlikely(cpu_is_offline(dest_cpu
)))
4990 rq_src
= cpu_rq(src_cpu
);
4991 rq_dest
= cpu_rq(dest_cpu
);
4993 double_rq_lock(rq_src
, rq_dest
);
4994 /* Already moved. */
4995 if (task_cpu(p
) != src_cpu
)
4997 /* Affinity changed (again). */
4998 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5001 on_rq
= p
->se
.on_rq
;
5003 deactivate_task(rq_src
, p
, 0);
5005 set_task_cpu(p
, dest_cpu
);
5007 activate_task(rq_dest
, p
, 0);
5008 check_preempt_curr(rq_dest
, p
);
5012 double_rq_unlock(rq_src
, rq_dest
);
5017 * migration_thread - this is a highprio system thread that performs
5018 * thread migration by bumping thread off CPU then 'pushing' onto
5021 static int migration_thread(void *data
)
5023 int cpu
= (long)data
;
5027 BUG_ON(rq
->migration_thread
!= current
);
5029 set_current_state(TASK_INTERRUPTIBLE
);
5030 while (!kthread_should_stop()) {
5031 struct migration_req
*req
;
5032 struct list_head
*head
;
5034 spin_lock_irq(&rq
->lock
);
5036 if (cpu_is_offline(cpu
)) {
5037 spin_unlock_irq(&rq
->lock
);
5041 if (rq
->active_balance
) {
5042 active_load_balance(rq
, cpu
);
5043 rq
->active_balance
= 0;
5046 head
= &rq
->migration_queue
;
5048 if (list_empty(head
)) {
5049 spin_unlock_irq(&rq
->lock
);
5051 set_current_state(TASK_INTERRUPTIBLE
);
5054 req
= list_entry(head
->next
, struct migration_req
, list
);
5055 list_del_init(head
->next
);
5057 spin_unlock(&rq
->lock
);
5058 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5061 complete(&req
->done
);
5063 __set_current_state(TASK_RUNNING
);
5067 /* Wait for kthread_stop */
5068 set_current_state(TASK_INTERRUPTIBLE
);
5069 while (!kthread_should_stop()) {
5071 set_current_state(TASK_INTERRUPTIBLE
);
5073 __set_current_state(TASK_RUNNING
);
5077 #ifdef CONFIG_HOTPLUG_CPU
5079 * Figure out where task on dead CPU should go, use force if neccessary.
5080 * NOTE: interrupts should be disabled by the caller
5082 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5084 unsigned long flags
;
5091 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5092 cpus_and(mask
, mask
, p
->cpus_allowed
);
5093 dest_cpu
= any_online_cpu(mask
);
5095 /* On any allowed CPU? */
5096 if (dest_cpu
== NR_CPUS
)
5097 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5099 /* No more Mr. Nice Guy. */
5100 if (dest_cpu
== NR_CPUS
) {
5101 rq
= task_rq_lock(p
, &flags
);
5102 cpus_setall(p
->cpus_allowed
);
5103 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5104 task_rq_unlock(rq
, &flags
);
5107 * Don't tell them about moving exiting tasks or
5108 * kernel threads (both mm NULL), since they never
5111 if (p
->mm
&& printk_ratelimit())
5112 printk(KERN_INFO
"process %d (%s) no "
5113 "longer affine to cpu%d\n",
5114 p
->pid
, p
->comm
, dead_cpu
);
5116 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5121 * While a dead CPU has no uninterruptible tasks queued at this point,
5122 * it might still have a nonzero ->nr_uninterruptible counter, because
5123 * for performance reasons the counter is not stricly tracking tasks to
5124 * their home CPUs. So we just add the counter to another CPU's counter,
5125 * to keep the global sum constant after CPU-down:
5127 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5129 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5130 unsigned long flags
;
5132 local_irq_save(flags
);
5133 double_rq_lock(rq_src
, rq_dest
);
5134 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5135 rq_src
->nr_uninterruptible
= 0;
5136 double_rq_unlock(rq_src
, rq_dest
);
5137 local_irq_restore(flags
);
5140 /* Run through task list and migrate tasks from the dead cpu. */
5141 static void migrate_live_tasks(int src_cpu
)
5143 struct task_struct
*p
, *t
;
5145 write_lock_irq(&tasklist_lock
);
5147 do_each_thread(t
, p
) {
5151 if (task_cpu(p
) == src_cpu
)
5152 move_task_off_dead_cpu(src_cpu
, p
);
5153 } while_each_thread(t
, p
);
5155 write_unlock_irq(&tasklist_lock
);
5159 * Schedules idle task to be the next runnable task on current CPU.
5160 * It does so by boosting its priority to highest possible and adding it to
5161 * the _front_ of the runqueue. Used by CPU offline code.
5163 void sched_idle_next(void)
5165 int this_cpu
= smp_processor_id();
5166 struct rq
*rq
= cpu_rq(this_cpu
);
5167 struct task_struct
*p
= rq
->idle
;
5168 unsigned long flags
;
5170 /* cpu has to be offline */
5171 BUG_ON(cpu_online(this_cpu
));
5174 * Strictly not necessary since rest of the CPUs are stopped by now
5175 * and interrupts disabled on the current cpu.
5177 spin_lock_irqsave(&rq
->lock
, flags
);
5179 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5181 /* Add idle task to the _front_ of its priority queue: */
5182 activate_idle_task(p
, rq
);
5184 spin_unlock_irqrestore(&rq
->lock
, flags
);
5188 * Ensures that the idle task is using init_mm right before its cpu goes
5191 void idle_task_exit(void)
5193 struct mm_struct
*mm
= current
->active_mm
;
5195 BUG_ON(cpu_online(smp_processor_id()));
5198 switch_mm(mm
, &init_mm
, current
);
5202 /* called under rq->lock with disabled interrupts */
5203 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5205 struct rq
*rq
= cpu_rq(dead_cpu
);
5207 /* Must be exiting, otherwise would be on tasklist. */
5208 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5210 /* Cannot have done final schedule yet: would have vanished. */
5211 BUG_ON(p
->state
== TASK_DEAD
);
5216 * Drop lock around migration; if someone else moves it,
5217 * that's OK. No task can be added to this CPU, so iteration is
5219 * NOTE: interrupts should be left disabled --dev@
5221 spin_unlock(&rq
->lock
);
5222 move_task_off_dead_cpu(dead_cpu
, p
);
5223 spin_lock(&rq
->lock
);
5228 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5229 static void migrate_dead_tasks(unsigned int dead_cpu
)
5231 struct rq
*rq
= cpu_rq(dead_cpu
);
5232 struct task_struct
*next
;
5235 if (!rq
->nr_running
)
5237 update_rq_clock(rq
);
5238 next
= pick_next_task(rq
, rq
->curr
);
5241 migrate_dead(dead_cpu
, next
);
5245 #endif /* CONFIG_HOTPLUG_CPU */
5247 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5249 static struct ctl_table sd_ctl_dir
[] = {
5251 .procname
= "sched_domain",
5257 static struct ctl_table sd_ctl_root
[] = {
5259 .ctl_name
= CTL_KERN
,
5260 .procname
= "kernel",
5262 .child
= sd_ctl_dir
,
5267 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5269 struct ctl_table
*entry
=
5270 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5273 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5279 set_table_entry(struct ctl_table
*entry
,
5280 const char *procname
, void *data
, int maxlen
,
5281 mode_t mode
, proc_handler
*proc_handler
)
5283 entry
->procname
= procname
;
5285 entry
->maxlen
= maxlen
;
5287 entry
->proc_handler
= proc_handler
;
5290 static struct ctl_table
*
5291 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5293 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5295 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5296 sizeof(long), 0644, proc_doulongvec_minmax
);
5297 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5298 sizeof(long), 0644, proc_doulongvec_minmax
);
5299 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5300 sizeof(int), 0644, proc_dointvec_minmax
);
5301 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5302 sizeof(int), 0644, proc_dointvec_minmax
);
5303 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5304 sizeof(int), 0644, proc_dointvec_minmax
);
5305 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5306 sizeof(int), 0644, proc_dointvec_minmax
);
5307 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5308 sizeof(int), 0644, proc_dointvec_minmax
);
5309 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5310 sizeof(int), 0644, proc_dointvec_minmax
);
5311 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5312 sizeof(int), 0644, proc_dointvec_minmax
);
5313 set_table_entry(&table
[10], "cache_nice_tries",
5314 &sd
->cache_nice_tries
,
5315 sizeof(int), 0644, proc_dointvec_minmax
);
5316 set_table_entry(&table
[12], "flags", &sd
->flags
,
5317 sizeof(int), 0644, proc_dointvec_minmax
);
5322 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5324 struct ctl_table
*entry
, *table
;
5325 struct sched_domain
*sd
;
5326 int domain_num
= 0, i
;
5329 for_each_domain(cpu
, sd
)
5331 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5334 for_each_domain(cpu
, sd
) {
5335 snprintf(buf
, 32, "domain%d", i
);
5336 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5338 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5345 static struct ctl_table_header
*sd_sysctl_header
;
5346 static void init_sched_domain_sysctl(void)
5348 int i
, cpu_num
= num_online_cpus();
5349 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5352 sd_ctl_dir
[0].child
= entry
;
5354 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5355 snprintf(buf
, 32, "cpu%d", i
);
5356 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5358 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5360 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5363 static void init_sched_domain_sysctl(void)
5369 * migration_call - callback that gets triggered when a CPU is added.
5370 * Here we can start up the necessary migration thread for the new CPU.
5372 static int __cpuinit
5373 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5375 struct task_struct
*p
;
5376 int cpu
= (long)hcpu
;
5377 unsigned long flags
;
5381 case CPU_LOCK_ACQUIRE
:
5382 mutex_lock(&sched_hotcpu_mutex
);
5385 case CPU_UP_PREPARE
:
5386 case CPU_UP_PREPARE_FROZEN
:
5387 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5390 kthread_bind(p
, cpu
);
5391 /* Must be high prio: stop_machine expects to yield to it. */
5392 rq
= task_rq_lock(p
, &flags
);
5393 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5394 task_rq_unlock(rq
, &flags
);
5395 cpu_rq(cpu
)->migration_thread
= p
;
5399 case CPU_ONLINE_FROZEN
:
5400 /* Strictly unneccessary, as first user will wake it. */
5401 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5404 #ifdef CONFIG_HOTPLUG_CPU
5405 case CPU_UP_CANCELED
:
5406 case CPU_UP_CANCELED_FROZEN
:
5407 if (!cpu_rq(cpu
)->migration_thread
)
5409 /* Unbind it from offline cpu so it can run. Fall thru. */
5410 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5411 any_online_cpu(cpu_online_map
));
5412 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5413 cpu_rq(cpu
)->migration_thread
= NULL
;
5417 case CPU_DEAD_FROZEN
:
5418 migrate_live_tasks(cpu
);
5420 kthread_stop(rq
->migration_thread
);
5421 rq
->migration_thread
= NULL
;
5422 /* Idle task back to normal (off runqueue, low prio) */
5423 rq
= task_rq_lock(rq
->idle
, &flags
);
5424 update_rq_clock(rq
);
5425 deactivate_task(rq
, rq
->idle
, 0);
5426 rq
->idle
->static_prio
= MAX_PRIO
;
5427 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5428 rq
->idle
->sched_class
= &idle_sched_class
;
5429 migrate_dead_tasks(cpu
);
5430 task_rq_unlock(rq
, &flags
);
5431 migrate_nr_uninterruptible(rq
);
5432 BUG_ON(rq
->nr_running
!= 0);
5434 /* No need to migrate the tasks: it was best-effort if
5435 * they didn't take sched_hotcpu_mutex. Just wake up
5436 * the requestors. */
5437 spin_lock_irq(&rq
->lock
);
5438 while (!list_empty(&rq
->migration_queue
)) {
5439 struct migration_req
*req
;
5441 req
= list_entry(rq
->migration_queue
.next
,
5442 struct migration_req
, list
);
5443 list_del_init(&req
->list
);
5444 complete(&req
->done
);
5446 spin_unlock_irq(&rq
->lock
);
5449 case CPU_LOCK_RELEASE
:
5450 mutex_unlock(&sched_hotcpu_mutex
);
5456 /* Register at highest priority so that task migration (migrate_all_tasks)
5457 * happens before everything else.
5459 static struct notifier_block __cpuinitdata migration_notifier
= {
5460 .notifier_call
= migration_call
,
5464 int __init
migration_init(void)
5466 void *cpu
= (void *)(long)smp_processor_id();
5469 /* Start one for the boot CPU: */
5470 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5471 BUG_ON(err
== NOTIFY_BAD
);
5472 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5473 register_cpu_notifier(&migration_notifier
);
5481 /* Number of possible processor ids */
5482 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5483 EXPORT_SYMBOL(nr_cpu_ids
);
5485 #undef SCHED_DOMAIN_DEBUG
5486 #ifdef SCHED_DOMAIN_DEBUG
5487 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5492 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5496 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5501 struct sched_group
*group
= sd
->groups
;
5502 cpumask_t groupmask
;
5504 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5505 cpus_clear(groupmask
);
5508 for (i
= 0; i
< level
+ 1; i
++)
5510 printk("domain %d: ", level
);
5512 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5513 printk("does not load-balance\n");
5515 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5520 printk("span %s\n", str
);
5522 if (!cpu_isset(cpu
, sd
->span
))
5523 printk(KERN_ERR
"ERROR: domain->span does not contain "
5525 if (!cpu_isset(cpu
, group
->cpumask
))
5526 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5530 for (i
= 0; i
< level
+ 2; i
++)
5536 printk(KERN_ERR
"ERROR: group is NULL\n");
5540 if (!group
->__cpu_power
) {
5542 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5546 if (!cpus_weight(group
->cpumask
)) {
5548 printk(KERN_ERR
"ERROR: empty group\n");
5551 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5553 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5556 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5558 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5561 group
= group
->next
;
5562 } while (group
!= sd
->groups
);
5565 if (!cpus_equal(sd
->span
, groupmask
))
5566 printk(KERN_ERR
"ERROR: groups don't span "
5574 if (!cpus_subset(groupmask
, sd
->span
))
5575 printk(KERN_ERR
"ERROR: parent span is not a superset "
5576 "of domain->span\n");
5581 # define sched_domain_debug(sd, cpu) do { } while (0)
5584 static int sd_degenerate(struct sched_domain
*sd
)
5586 if (cpus_weight(sd
->span
) == 1)
5589 /* Following flags need at least 2 groups */
5590 if (sd
->flags
& (SD_LOAD_BALANCE
|
5591 SD_BALANCE_NEWIDLE
|
5595 SD_SHARE_PKG_RESOURCES
)) {
5596 if (sd
->groups
!= sd
->groups
->next
)
5600 /* Following flags don't use groups */
5601 if (sd
->flags
& (SD_WAKE_IDLE
|
5610 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5612 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5614 if (sd_degenerate(parent
))
5617 if (!cpus_equal(sd
->span
, parent
->span
))
5620 /* Does parent contain flags not in child? */
5621 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5622 if (cflags
& SD_WAKE_AFFINE
)
5623 pflags
&= ~SD_WAKE_BALANCE
;
5624 /* Flags needing groups don't count if only 1 group in parent */
5625 if (parent
->groups
== parent
->groups
->next
) {
5626 pflags
&= ~(SD_LOAD_BALANCE
|
5627 SD_BALANCE_NEWIDLE
|
5631 SD_SHARE_PKG_RESOURCES
);
5633 if (~cflags
& pflags
)
5640 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5641 * hold the hotplug lock.
5643 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5645 struct rq
*rq
= cpu_rq(cpu
);
5646 struct sched_domain
*tmp
;
5648 /* Remove the sched domains which do not contribute to scheduling. */
5649 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5650 struct sched_domain
*parent
= tmp
->parent
;
5653 if (sd_parent_degenerate(tmp
, parent
)) {
5654 tmp
->parent
= parent
->parent
;
5656 parent
->parent
->child
= tmp
;
5660 if (sd
&& sd_degenerate(sd
)) {
5666 sched_domain_debug(sd
, cpu
);
5668 rcu_assign_pointer(rq
->sd
, sd
);
5671 /* cpus with isolated domains */
5672 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5674 /* Setup the mask of cpus configured for isolated domains */
5675 static int __init
isolated_cpu_setup(char *str
)
5677 int ints
[NR_CPUS
], i
;
5679 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5680 cpus_clear(cpu_isolated_map
);
5681 for (i
= 1; i
<= ints
[0]; i
++)
5682 if (ints
[i
] < NR_CPUS
)
5683 cpu_set(ints
[i
], cpu_isolated_map
);
5687 __setup ("isolcpus=", isolated_cpu_setup
);
5690 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5691 * to a function which identifies what group(along with sched group) a CPU
5692 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5693 * (due to the fact that we keep track of groups covered with a cpumask_t).
5695 * init_sched_build_groups will build a circular linked list of the groups
5696 * covered by the given span, and will set each group's ->cpumask correctly,
5697 * and ->cpu_power to 0.
5700 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5701 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5702 struct sched_group
**sg
))
5704 struct sched_group
*first
= NULL
, *last
= NULL
;
5705 cpumask_t covered
= CPU_MASK_NONE
;
5708 for_each_cpu_mask(i
, span
) {
5709 struct sched_group
*sg
;
5710 int group
= group_fn(i
, cpu_map
, &sg
);
5713 if (cpu_isset(i
, covered
))
5716 sg
->cpumask
= CPU_MASK_NONE
;
5717 sg
->__cpu_power
= 0;
5719 for_each_cpu_mask(j
, span
) {
5720 if (group_fn(j
, cpu_map
, NULL
) != group
)
5723 cpu_set(j
, covered
);
5724 cpu_set(j
, sg
->cpumask
);
5735 #define SD_NODES_PER_DOMAIN 16
5740 * find_next_best_node - find the next node to include in a sched_domain
5741 * @node: node whose sched_domain we're building
5742 * @used_nodes: nodes already in the sched_domain
5744 * Find the next node to include in a given scheduling domain. Simply
5745 * finds the closest node not already in the @used_nodes map.
5747 * Should use nodemask_t.
5749 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5751 int i
, n
, val
, min_val
, best_node
= 0;
5755 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5756 /* Start at @node */
5757 n
= (node
+ i
) % MAX_NUMNODES
;
5759 if (!nr_cpus_node(n
))
5762 /* Skip already used nodes */
5763 if (test_bit(n
, used_nodes
))
5766 /* Simple min distance search */
5767 val
= node_distance(node
, n
);
5769 if (val
< min_val
) {
5775 set_bit(best_node
, used_nodes
);
5780 * sched_domain_node_span - get a cpumask for a node's sched_domain
5781 * @node: node whose cpumask we're constructing
5782 * @size: number of nodes to include in this span
5784 * Given a node, construct a good cpumask for its sched_domain to span. It
5785 * should be one that prevents unnecessary balancing, but also spreads tasks
5788 static cpumask_t
sched_domain_node_span(int node
)
5790 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5791 cpumask_t span
, nodemask
;
5795 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5797 nodemask
= node_to_cpumask(node
);
5798 cpus_or(span
, span
, nodemask
);
5799 set_bit(node
, used_nodes
);
5801 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5802 int next_node
= find_next_best_node(node
, used_nodes
);
5804 nodemask
= node_to_cpumask(next_node
);
5805 cpus_or(span
, span
, nodemask
);
5812 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5815 * SMT sched-domains:
5817 #ifdef CONFIG_SCHED_SMT
5818 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5819 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5821 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5822 struct sched_group
**sg
)
5825 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5831 * multi-core sched-domains:
5833 #ifdef CONFIG_SCHED_MC
5834 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5835 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5838 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5839 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5840 struct sched_group
**sg
)
5843 cpumask_t mask
= cpu_sibling_map
[cpu
];
5844 cpus_and(mask
, mask
, *cpu_map
);
5845 group
= first_cpu(mask
);
5847 *sg
= &per_cpu(sched_group_core
, group
);
5850 #elif defined(CONFIG_SCHED_MC)
5851 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5852 struct sched_group
**sg
)
5855 *sg
= &per_cpu(sched_group_core
, cpu
);
5860 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5861 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5863 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5864 struct sched_group
**sg
)
5867 #ifdef CONFIG_SCHED_MC
5868 cpumask_t mask
= cpu_coregroup_map(cpu
);
5869 cpus_and(mask
, mask
, *cpu_map
);
5870 group
= first_cpu(mask
);
5871 #elif defined(CONFIG_SCHED_SMT)
5872 cpumask_t mask
= cpu_sibling_map
[cpu
];
5873 cpus_and(mask
, mask
, *cpu_map
);
5874 group
= first_cpu(mask
);
5879 *sg
= &per_cpu(sched_group_phys
, group
);
5885 * The init_sched_build_groups can't handle what we want to do with node
5886 * groups, so roll our own. Now each node has its own list of groups which
5887 * gets dynamically allocated.
5889 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5890 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5892 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5893 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5895 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5896 struct sched_group
**sg
)
5898 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5901 cpus_and(nodemask
, nodemask
, *cpu_map
);
5902 group
= first_cpu(nodemask
);
5905 *sg
= &per_cpu(sched_group_allnodes
, group
);
5909 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5911 struct sched_group
*sg
= group_head
;
5917 for_each_cpu_mask(j
, sg
->cpumask
) {
5918 struct sched_domain
*sd
;
5920 sd
= &per_cpu(phys_domains
, j
);
5921 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5923 * Only add "power" once for each
5929 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5932 if (sg
!= group_head
)
5938 /* Free memory allocated for various sched_group structures */
5939 static void free_sched_groups(const cpumask_t
*cpu_map
)
5943 for_each_cpu_mask(cpu
, *cpu_map
) {
5944 struct sched_group
**sched_group_nodes
5945 = sched_group_nodes_bycpu
[cpu
];
5947 if (!sched_group_nodes
)
5950 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5951 cpumask_t nodemask
= node_to_cpumask(i
);
5952 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5954 cpus_and(nodemask
, nodemask
, *cpu_map
);
5955 if (cpus_empty(nodemask
))
5965 if (oldsg
!= sched_group_nodes
[i
])
5968 kfree(sched_group_nodes
);
5969 sched_group_nodes_bycpu
[cpu
] = NULL
;
5973 static void free_sched_groups(const cpumask_t
*cpu_map
)
5979 * Initialize sched groups cpu_power.
5981 * cpu_power indicates the capacity of sched group, which is used while
5982 * distributing the load between different sched groups in a sched domain.
5983 * Typically cpu_power for all the groups in a sched domain will be same unless
5984 * there are asymmetries in the topology. If there are asymmetries, group
5985 * having more cpu_power will pickup more load compared to the group having
5988 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5989 * the maximum number of tasks a group can handle in the presence of other idle
5990 * or lightly loaded groups in the same sched domain.
5992 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5994 struct sched_domain
*child
;
5995 struct sched_group
*group
;
5997 WARN_ON(!sd
|| !sd
->groups
);
5999 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6004 sd
->groups
->__cpu_power
= 0;
6007 * For perf policy, if the groups in child domain share resources
6008 * (for example cores sharing some portions of the cache hierarchy
6009 * or SMT), then set this domain groups cpu_power such that each group
6010 * can handle only one task, when there are other idle groups in the
6011 * same sched domain.
6013 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6015 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6016 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6021 * add cpu_power of each child group to this groups cpu_power
6023 group
= child
->groups
;
6025 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6026 group
= group
->next
;
6027 } while (group
!= child
->groups
);
6031 * Build sched domains for a given set of cpus and attach the sched domains
6032 * to the individual cpus
6034 static int build_sched_domains(const cpumask_t
*cpu_map
)
6038 struct sched_group
**sched_group_nodes
= NULL
;
6039 int sd_allnodes
= 0;
6042 * Allocate the per-node list of sched groups
6044 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6046 if (!sched_group_nodes
) {
6047 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6050 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6054 * Set up domains for cpus specified by the cpu_map.
6056 for_each_cpu_mask(i
, *cpu_map
) {
6057 struct sched_domain
*sd
= NULL
, *p
;
6058 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6060 cpus_and(nodemask
, nodemask
, *cpu_map
);
6063 if (cpus_weight(*cpu_map
) >
6064 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6065 sd
= &per_cpu(allnodes_domains
, i
);
6066 *sd
= SD_ALLNODES_INIT
;
6067 sd
->span
= *cpu_map
;
6068 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6074 sd
= &per_cpu(node_domains
, i
);
6076 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6080 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6084 sd
= &per_cpu(phys_domains
, i
);
6086 sd
->span
= nodemask
;
6090 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6092 #ifdef CONFIG_SCHED_MC
6094 sd
= &per_cpu(core_domains
, i
);
6096 sd
->span
= cpu_coregroup_map(i
);
6097 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6100 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6103 #ifdef CONFIG_SCHED_SMT
6105 sd
= &per_cpu(cpu_domains
, i
);
6106 *sd
= SD_SIBLING_INIT
;
6107 sd
->span
= cpu_sibling_map
[i
];
6108 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6111 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6115 #ifdef CONFIG_SCHED_SMT
6116 /* Set up CPU (sibling) groups */
6117 for_each_cpu_mask(i
, *cpu_map
) {
6118 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6119 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6120 if (i
!= first_cpu(this_sibling_map
))
6123 init_sched_build_groups(this_sibling_map
, cpu_map
,
6128 #ifdef CONFIG_SCHED_MC
6129 /* Set up multi-core groups */
6130 for_each_cpu_mask(i
, *cpu_map
) {
6131 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6132 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6133 if (i
!= first_cpu(this_core_map
))
6135 init_sched_build_groups(this_core_map
, cpu_map
,
6136 &cpu_to_core_group
);
6140 /* Set up physical groups */
6141 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6142 cpumask_t nodemask
= node_to_cpumask(i
);
6144 cpus_and(nodemask
, nodemask
, *cpu_map
);
6145 if (cpus_empty(nodemask
))
6148 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6152 /* Set up node groups */
6154 init_sched_build_groups(*cpu_map
, cpu_map
,
6155 &cpu_to_allnodes_group
);
6157 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6158 /* Set up node groups */
6159 struct sched_group
*sg
, *prev
;
6160 cpumask_t nodemask
= node_to_cpumask(i
);
6161 cpumask_t domainspan
;
6162 cpumask_t covered
= CPU_MASK_NONE
;
6165 cpus_and(nodemask
, nodemask
, *cpu_map
);
6166 if (cpus_empty(nodemask
)) {
6167 sched_group_nodes
[i
] = NULL
;
6171 domainspan
= sched_domain_node_span(i
);
6172 cpus_and(domainspan
, domainspan
, *cpu_map
);
6174 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6176 printk(KERN_WARNING
"Can not alloc domain group for "
6180 sched_group_nodes
[i
] = sg
;
6181 for_each_cpu_mask(j
, nodemask
) {
6182 struct sched_domain
*sd
;
6184 sd
= &per_cpu(node_domains
, j
);
6187 sg
->__cpu_power
= 0;
6188 sg
->cpumask
= nodemask
;
6190 cpus_or(covered
, covered
, nodemask
);
6193 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6194 cpumask_t tmp
, notcovered
;
6195 int n
= (i
+ j
) % MAX_NUMNODES
;
6197 cpus_complement(notcovered
, covered
);
6198 cpus_and(tmp
, notcovered
, *cpu_map
);
6199 cpus_and(tmp
, tmp
, domainspan
);
6200 if (cpus_empty(tmp
))
6203 nodemask
= node_to_cpumask(n
);
6204 cpus_and(tmp
, tmp
, nodemask
);
6205 if (cpus_empty(tmp
))
6208 sg
= kmalloc_node(sizeof(struct sched_group
),
6212 "Can not alloc domain group for node %d\n", j
);
6215 sg
->__cpu_power
= 0;
6217 sg
->next
= prev
->next
;
6218 cpus_or(covered
, covered
, tmp
);
6225 /* Calculate CPU power for physical packages and nodes */
6226 #ifdef CONFIG_SCHED_SMT
6227 for_each_cpu_mask(i
, *cpu_map
) {
6228 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6230 init_sched_groups_power(i
, sd
);
6233 #ifdef CONFIG_SCHED_MC
6234 for_each_cpu_mask(i
, *cpu_map
) {
6235 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6237 init_sched_groups_power(i
, sd
);
6241 for_each_cpu_mask(i
, *cpu_map
) {
6242 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6244 init_sched_groups_power(i
, sd
);
6248 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6249 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6252 struct sched_group
*sg
;
6254 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6255 init_numa_sched_groups_power(sg
);
6259 /* Attach the domains */
6260 for_each_cpu_mask(i
, *cpu_map
) {
6261 struct sched_domain
*sd
;
6262 #ifdef CONFIG_SCHED_SMT
6263 sd
= &per_cpu(cpu_domains
, i
);
6264 #elif defined(CONFIG_SCHED_MC)
6265 sd
= &per_cpu(core_domains
, i
);
6267 sd
= &per_cpu(phys_domains
, i
);
6269 cpu_attach_domain(sd
, i
);
6276 free_sched_groups(cpu_map
);
6281 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6283 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6285 cpumask_t cpu_default_map
;
6289 * Setup mask for cpus without special case scheduling requirements.
6290 * For now this just excludes isolated cpus, but could be used to
6291 * exclude other special cases in the future.
6293 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6295 err
= build_sched_domains(&cpu_default_map
);
6300 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6302 free_sched_groups(cpu_map
);
6306 * Detach sched domains from a group of cpus specified in cpu_map
6307 * These cpus will now be attached to the NULL domain
6309 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6313 for_each_cpu_mask(i
, *cpu_map
)
6314 cpu_attach_domain(NULL
, i
);
6315 synchronize_sched();
6316 arch_destroy_sched_domains(cpu_map
);
6320 * Partition sched domains as specified by the cpumasks below.
6321 * This attaches all cpus from the cpumasks to the NULL domain,
6322 * waits for a RCU quiescent period, recalculates sched
6323 * domain information and then attaches them back to the
6324 * correct sched domains
6325 * Call with hotplug lock held
6327 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6329 cpumask_t change_map
;
6332 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6333 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6334 cpus_or(change_map
, *partition1
, *partition2
);
6336 /* Detach sched domains from all of the affected cpus */
6337 detach_destroy_domains(&change_map
);
6338 if (!cpus_empty(*partition1
))
6339 err
= build_sched_domains(partition1
);
6340 if (!err
&& !cpus_empty(*partition2
))
6341 err
= build_sched_domains(partition2
);
6346 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6347 static int arch_reinit_sched_domains(void)
6351 mutex_lock(&sched_hotcpu_mutex
);
6352 detach_destroy_domains(&cpu_online_map
);
6353 err
= arch_init_sched_domains(&cpu_online_map
);
6354 mutex_unlock(&sched_hotcpu_mutex
);
6359 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6363 if (buf
[0] != '0' && buf
[0] != '1')
6367 sched_smt_power_savings
= (buf
[0] == '1');
6369 sched_mc_power_savings
= (buf
[0] == '1');
6371 ret
= arch_reinit_sched_domains();
6373 return ret
? ret
: count
;
6376 #ifdef CONFIG_SCHED_MC
6377 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6379 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6381 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6382 const char *buf
, size_t count
)
6384 return sched_power_savings_store(buf
, count
, 0);
6386 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6387 sched_mc_power_savings_store
);
6390 #ifdef CONFIG_SCHED_SMT
6391 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6393 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6395 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6396 const char *buf
, size_t count
)
6398 return sched_power_savings_store(buf
, count
, 1);
6400 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6401 sched_smt_power_savings_store
);
6404 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6408 #ifdef CONFIG_SCHED_SMT
6410 err
= sysfs_create_file(&cls
->kset
.kobj
,
6411 &attr_sched_smt_power_savings
.attr
);
6413 #ifdef CONFIG_SCHED_MC
6414 if (!err
&& mc_capable())
6415 err
= sysfs_create_file(&cls
->kset
.kobj
,
6416 &attr_sched_mc_power_savings
.attr
);
6423 * Force a reinitialization of the sched domains hierarchy. The domains
6424 * and groups cannot be updated in place without racing with the balancing
6425 * code, so we temporarily attach all running cpus to the NULL domain
6426 * which will prevent rebalancing while the sched domains are recalculated.
6428 static int update_sched_domains(struct notifier_block
*nfb
,
6429 unsigned long action
, void *hcpu
)
6432 case CPU_UP_PREPARE
:
6433 case CPU_UP_PREPARE_FROZEN
:
6434 case CPU_DOWN_PREPARE
:
6435 case CPU_DOWN_PREPARE_FROZEN
:
6436 detach_destroy_domains(&cpu_online_map
);
6439 case CPU_UP_CANCELED
:
6440 case CPU_UP_CANCELED_FROZEN
:
6441 case CPU_DOWN_FAILED
:
6442 case CPU_DOWN_FAILED_FROZEN
:
6444 case CPU_ONLINE_FROZEN
:
6446 case CPU_DEAD_FROZEN
:
6448 * Fall through and re-initialise the domains.
6455 /* The hotplug lock is already held by cpu_up/cpu_down */
6456 arch_init_sched_domains(&cpu_online_map
);
6461 void __init
sched_init_smp(void)
6463 cpumask_t non_isolated_cpus
;
6465 mutex_lock(&sched_hotcpu_mutex
);
6466 arch_init_sched_domains(&cpu_online_map
);
6467 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6468 if (cpus_empty(non_isolated_cpus
))
6469 cpu_set(smp_processor_id(), non_isolated_cpus
);
6470 mutex_unlock(&sched_hotcpu_mutex
);
6471 /* XXX: Theoretical race here - CPU may be hotplugged now */
6472 hotcpu_notifier(update_sched_domains
, 0);
6474 init_sched_domain_sysctl();
6476 /* Move init over to a non-isolated CPU */
6477 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6481 void __init
sched_init_smp(void)
6484 #endif /* CONFIG_SMP */
6486 int in_sched_functions(unsigned long addr
)
6488 /* Linker adds these: start and end of __sched functions */
6489 extern char __sched_text_start
[], __sched_text_end
[];
6491 return in_lock_functions(addr
) ||
6492 (addr
>= (unsigned long)__sched_text_start
6493 && addr
< (unsigned long)__sched_text_end
);
6496 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6498 cfs_rq
->tasks_timeline
= RB_ROOT
;
6499 #ifdef CONFIG_FAIR_GROUP_SCHED
6502 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6505 void __init
sched_init(void)
6507 int highest_cpu
= 0;
6511 * Link up the scheduling class hierarchy:
6513 rt_sched_class
.next
= &fair_sched_class
;
6514 fair_sched_class
.next
= &idle_sched_class
;
6515 idle_sched_class
.next
= NULL
;
6517 for_each_possible_cpu(i
) {
6518 struct rt_prio_array
*array
;
6522 spin_lock_init(&rq
->lock
);
6523 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6526 init_cfs_rq(&rq
->cfs
, rq
);
6527 #ifdef CONFIG_FAIR_GROUP_SCHED
6528 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6530 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6531 struct sched_entity
*se
=
6532 &per_cpu(init_sched_entity
, i
);
6534 init_cfs_rq_p
[i
] = cfs_rq
;
6535 init_cfs_rq(cfs_rq
, rq
);
6536 cfs_rq
->tg
= &init_task_grp
;
6537 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6538 &rq
->leaf_cfs_rq_list
);
6540 init_sched_entity_p
[i
] = se
;
6541 se
->cfs_rq
= &rq
->cfs
;
6543 se
->load
.weight
= init_task_grp_load
;
6544 se
->load
.inv_weight
=
6545 div64_64(1ULL<<32, init_task_grp_load
);
6548 init_task_grp
.shares
= init_task_grp_load
;
6551 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6552 rq
->cpu_load
[j
] = 0;
6555 rq
->active_balance
= 0;
6556 rq
->next_balance
= jiffies
;
6559 rq
->migration_thread
= NULL
;
6560 INIT_LIST_HEAD(&rq
->migration_queue
);
6562 atomic_set(&rq
->nr_iowait
, 0);
6564 array
= &rq
->rt
.active
;
6565 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6566 INIT_LIST_HEAD(array
->queue
+ j
);
6567 __clear_bit(j
, array
->bitmap
);
6570 /* delimiter for bitsearch: */
6571 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6574 set_load_weight(&init_task
);
6576 #ifdef CONFIG_PREEMPT_NOTIFIERS
6577 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6581 nr_cpu_ids
= highest_cpu
+ 1;
6582 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6585 #ifdef CONFIG_RT_MUTEXES
6586 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6590 * The boot idle thread does lazy MMU switching as well:
6592 atomic_inc(&init_mm
.mm_count
);
6593 enter_lazy_tlb(&init_mm
, current
);
6596 * Make us the idle thread. Technically, schedule() should not be
6597 * called from this thread, however somewhere below it might be,
6598 * but because we are the idle thread, we just pick up running again
6599 * when this runqueue becomes "idle".
6601 init_idle(current
, smp_processor_id());
6603 * During early bootup we pretend to be a normal task:
6605 current
->sched_class
= &fair_sched_class
;
6608 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6609 void __might_sleep(char *file
, int line
)
6612 static unsigned long prev_jiffy
; /* ratelimiting */
6614 if ((in_atomic() || irqs_disabled()) &&
6615 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6616 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6618 prev_jiffy
= jiffies
;
6619 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6620 " context at %s:%d\n", file
, line
);
6621 printk("in_atomic():%d, irqs_disabled():%d\n",
6622 in_atomic(), irqs_disabled());
6623 debug_show_held_locks(current
);
6624 if (irqs_disabled())
6625 print_irqtrace_events(current
);
6630 EXPORT_SYMBOL(__might_sleep
);
6633 #ifdef CONFIG_MAGIC_SYSRQ
6634 void normalize_rt_tasks(void)
6636 struct task_struct
*g
, *p
;
6637 unsigned long flags
;
6641 read_lock_irq(&tasklist_lock
);
6642 do_each_thread(g
, p
) {
6643 p
->se
.exec_start
= 0;
6644 #ifdef CONFIG_SCHEDSTATS
6645 p
->se
.wait_start
= 0;
6646 p
->se
.sleep_start
= 0;
6647 p
->se
.block_start
= 0;
6649 task_rq(p
)->clock
= 0;
6653 * Renice negative nice level userspace
6656 if (TASK_NICE(p
) < 0 && p
->mm
)
6657 set_user_nice(p
, 0);
6661 spin_lock_irqsave(&p
->pi_lock
, flags
);
6662 rq
= __task_rq_lock(p
);
6665 * Do not touch the migration thread:
6667 if (p
== rq
->migration_thread
)
6671 update_rq_clock(rq
);
6672 on_rq
= p
->se
.on_rq
;
6674 deactivate_task(rq
, p
, 0);
6675 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6677 activate_task(rq
, p
, 0);
6678 resched_task(rq
->curr
);
6683 __task_rq_unlock(rq
);
6684 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6685 } while_each_thread(g
, p
);
6687 read_unlock_irq(&tasklist_lock
);
6690 #endif /* CONFIG_MAGIC_SYSRQ */
6694 * These functions are only useful for the IA64 MCA handling.
6696 * They can only be called when the whole system has been
6697 * stopped - every CPU needs to be quiescent, and no scheduling
6698 * activity can take place. Using them for anything else would
6699 * be a serious bug, and as a result, they aren't even visible
6700 * under any other configuration.
6704 * curr_task - return the current task for a given cpu.
6705 * @cpu: the processor in question.
6707 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6709 struct task_struct
*curr_task(int cpu
)
6711 return cpu_curr(cpu
);
6715 * set_curr_task - set the current task for a given cpu.
6716 * @cpu: the processor in question.
6717 * @p: the task pointer to set.
6719 * Description: This function must only be used when non-maskable interrupts
6720 * are serviced on a separate stack. It allows the architecture to switch the
6721 * notion of the current task on a cpu in a non-blocking manner. This function
6722 * must be called with all CPU's synchronized, and interrupts disabled, the
6723 * and caller must save the original value of the current task (see
6724 * curr_task() above) and restore that value before reenabling interrupts and
6725 * re-starting the system.
6727 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6729 void set_curr_task(int cpu
, struct task_struct
*p
)
6736 #ifdef CONFIG_FAIR_GROUP_SCHED
6738 /* allocate runqueue etc for a new task group */
6739 struct task_grp
*sched_create_group(void)
6741 struct task_grp
*tg
;
6742 struct cfs_rq
*cfs_rq
;
6743 struct sched_entity
*se
;
6747 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6749 return ERR_PTR(-ENOMEM
);
6751 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
6754 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
6758 for_each_possible_cpu(i
) {
6761 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
6766 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
6771 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
6772 memset(se
, 0, sizeof(struct sched_entity
));
6774 tg
->cfs_rq
[i
] = cfs_rq
;
6775 init_cfs_rq(cfs_rq
, rq
);
6779 se
->cfs_rq
= &rq
->cfs
;
6781 se
->load
.weight
= NICE_0_LOAD
;
6782 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6786 for_each_possible_cpu(i
) {
6788 cfs_rq
= tg
->cfs_rq
[i
];
6789 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6792 tg
->shares
= NICE_0_LOAD
;
6797 for_each_possible_cpu(i
) {
6798 if (tg
->cfs_rq
&& tg
->cfs_rq
[i
])
6799 kfree(tg
->cfs_rq
[i
]);
6800 if (tg
->se
&& tg
->se
[i
])
6810 return ERR_PTR(-ENOMEM
);
6813 /* rcu callback to free various structures associated with a task group */
6814 static void free_sched_group(struct rcu_head
*rhp
)
6816 struct cfs_rq
*cfs_rq
= container_of(rhp
, struct cfs_rq
, rcu
);
6817 struct task_grp
*tg
= cfs_rq
->tg
;
6818 struct sched_entity
*se
;
6821 /* now it should be safe to free those cfs_rqs */
6822 for_each_possible_cpu(i
) {
6823 cfs_rq
= tg
->cfs_rq
[i
];
6835 /* Destroy runqueue etc associated with a task group */
6836 void sched_destroy_group(struct task_grp
*tg
)
6838 struct cfs_rq
*cfs_rq
;
6841 for_each_possible_cpu(i
) {
6842 cfs_rq
= tg
->cfs_rq
[i
];
6843 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
6846 cfs_rq
= tg
->cfs_rq
[0];
6848 /* wait for possible concurrent references to cfs_rqs complete */
6849 call_rcu(&cfs_rq
->rcu
, free_sched_group
);
6852 /* change task's runqueue when it moves between groups.
6853 * The caller of this function should have put the task in its new group
6854 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6855 * reflect its new group.
6857 void sched_move_task(struct task_struct
*tsk
)
6860 unsigned long flags
;
6863 rq
= task_rq_lock(tsk
, &flags
);
6865 if (tsk
->sched_class
!= &fair_sched_class
)
6868 update_rq_clock(rq
);
6870 running
= task_running(rq
, tsk
);
6871 on_rq
= tsk
->se
.on_rq
;
6874 dequeue_task(rq
, tsk
, 0);
6875 if (unlikely(running
))
6876 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6879 set_task_cfs_rq(tsk
);
6882 if (unlikely(running
))
6883 tsk
->sched_class
->set_curr_task(rq
);
6884 enqueue_task(rq
, tsk
, 0);
6888 task_rq_unlock(rq
, &flags
);
6891 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
6893 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
6894 struct rq
*rq
= cfs_rq
->rq
;
6897 spin_lock_irq(&rq
->lock
);
6901 dequeue_entity(cfs_rq
, se
, 0);
6903 se
->load
.weight
= shares
;
6904 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
6907 enqueue_entity(cfs_rq
, se
, 0);
6909 spin_unlock_irq(&rq
->lock
);
6912 int sched_group_set_shares(struct task_grp
*tg
, unsigned long shares
)
6916 if (tg
->shares
== shares
)
6919 /* return -EINVAL if the new value is not sane */
6921 tg
->shares
= shares
;
6922 for_each_possible_cpu(i
)
6923 set_se_shares(tg
->se
[i
], shares
);
6928 #endif /* CONFIG_FAIR_GROUP_SCHED */