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 p
->se
.vruntime
-= old_rq
->cfs
.min_vruntime
- new_rq
->cfs
.min_vruntime
;
1057 __set_task_cpu(p
, new_cpu
);
1060 struct migration_req
{
1061 struct list_head list
;
1063 struct task_struct
*task
;
1066 struct completion done
;
1070 * The task's runqueue lock must be held.
1071 * Returns true if you have to wait for migration thread.
1074 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1076 struct rq
*rq
= task_rq(p
);
1079 * If the task is not on a runqueue (and not running), then
1080 * it is sufficient to simply update the task's cpu field.
1082 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1083 set_task_cpu(p
, dest_cpu
);
1087 init_completion(&req
->done
);
1089 req
->dest_cpu
= dest_cpu
;
1090 list_add(&req
->list
, &rq
->migration_queue
);
1096 * wait_task_inactive - wait for a thread to unschedule.
1098 * The caller must ensure that the task *will* unschedule sometime soon,
1099 * else this function might spin for a *long* time. This function can't
1100 * be called with interrupts off, or it may introduce deadlock with
1101 * smp_call_function() if an IPI is sent by the same process we are
1102 * waiting to become inactive.
1104 void wait_task_inactive(struct task_struct
*p
)
1106 unsigned long flags
;
1112 * We do the initial early heuristics without holding
1113 * any task-queue locks at all. We'll only try to get
1114 * the runqueue lock when things look like they will
1120 * If the task is actively running on another CPU
1121 * still, just relax and busy-wait without holding
1124 * NOTE! Since we don't hold any locks, it's not
1125 * even sure that "rq" stays as the right runqueue!
1126 * But we don't care, since "task_running()" will
1127 * return false if the runqueue has changed and p
1128 * is actually now running somewhere else!
1130 while (task_running(rq
, p
))
1134 * Ok, time to look more closely! We need the rq
1135 * lock now, to be *sure*. If we're wrong, we'll
1136 * just go back and repeat.
1138 rq
= task_rq_lock(p
, &flags
);
1139 running
= task_running(rq
, p
);
1140 on_rq
= p
->se
.on_rq
;
1141 task_rq_unlock(rq
, &flags
);
1144 * Was it really running after all now that we
1145 * checked with the proper locks actually held?
1147 * Oops. Go back and try again..
1149 if (unlikely(running
)) {
1155 * It's not enough that it's not actively running,
1156 * it must be off the runqueue _entirely_, and not
1159 * So if it wa still runnable (but just not actively
1160 * running right now), it's preempted, and we should
1161 * yield - it could be a while.
1163 if (unlikely(on_rq
)) {
1169 * Ahh, all good. It wasn't running, and it wasn't
1170 * runnable, which means that it will never become
1171 * running in the future either. We're all done!
1176 * kick_process - kick a running thread to enter/exit the kernel
1177 * @p: the to-be-kicked thread
1179 * Cause a process which is running on another CPU to enter
1180 * kernel-mode, without any delay. (to get signals handled.)
1182 * NOTE: this function doesnt have to take the runqueue lock,
1183 * because all it wants to ensure is that the remote task enters
1184 * the kernel. If the IPI races and the task has been migrated
1185 * to another CPU then no harm is done and the purpose has been
1188 void kick_process(struct task_struct
*p
)
1194 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1195 smp_send_reschedule(cpu
);
1200 * Return a low guess at the load of a migration-source cpu weighted
1201 * according to the scheduling class and "nice" value.
1203 * We want to under-estimate the load of migration sources, to
1204 * balance conservatively.
1206 static inline unsigned long source_load(int cpu
, int type
)
1208 struct rq
*rq
= cpu_rq(cpu
);
1209 unsigned long total
= weighted_cpuload(cpu
);
1214 return min(rq
->cpu_load
[type
-1], total
);
1218 * Return a high guess at the load of a migration-target cpu weighted
1219 * according to the scheduling class and "nice" value.
1221 static inline unsigned long target_load(int cpu
, int type
)
1223 struct rq
*rq
= cpu_rq(cpu
);
1224 unsigned long total
= weighted_cpuload(cpu
);
1229 return max(rq
->cpu_load
[type
-1], total
);
1233 * Return the average load per task on the cpu's run queue
1235 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1237 struct rq
*rq
= cpu_rq(cpu
);
1238 unsigned long total
= weighted_cpuload(cpu
);
1239 unsigned long n
= rq
->nr_running
;
1241 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1245 * find_idlest_group finds and returns the least busy CPU group within the
1248 static struct sched_group
*
1249 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1251 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1252 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1253 int load_idx
= sd
->forkexec_idx
;
1254 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1257 unsigned long load
, avg_load
;
1261 /* Skip over this group if it has no CPUs allowed */
1262 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1265 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1267 /* Tally up the load of all CPUs in the group */
1270 for_each_cpu_mask(i
, group
->cpumask
) {
1271 /* Bias balancing toward cpus of our domain */
1273 load
= source_load(i
, load_idx
);
1275 load
= target_load(i
, load_idx
);
1280 /* Adjust by relative CPU power of the group */
1281 avg_load
= sg_div_cpu_power(group
,
1282 avg_load
* SCHED_LOAD_SCALE
);
1285 this_load
= avg_load
;
1287 } else if (avg_load
< min_load
) {
1288 min_load
= avg_load
;
1292 group
= group
->next
;
1293 } while (group
!= sd
->groups
);
1295 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1301 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1304 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1307 unsigned long load
, min_load
= ULONG_MAX
;
1311 /* Traverse only the allowed CPUs */
1312 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1314 for_each_cpu_mask(i
, tmp
) {
1315 load
= weighted_cpuload(i
);
1317 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1327 * sched_balance_self: balance the current task (running on cpu) in domains
1328 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1331 * Balance, ie. select the least loaded group.
1333 * Returns the target CPU number, or the same CPU if no balancing is needed.
1335 * preempt must be disabled.
1337 static int sched_balance_self(int cpu
, int flag
)
1339 struct task_struct
*t
= current
;
1340 struct sched_domain
*tmp
, *sd
= NULL
;
1342 for_each_domain(cpu
, tmp
) {
1344 * If power savings logic is enabled for a domain, stop there.
1346 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1348 if (tmp
->flags
& flag
)
1354 struct sched_group
*group
;
1355 int new_cpu
, weight
;
1357 if (!(sd
->flags
& flag
)) {
1363 group
= find_idlest_group(sd
, t
, cpu
);
1369 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1370 if (new_cpu
== -1 || new_cpu
== cpu
) {
1371 /* Now try balancing at a lower domain level of cpu */
1376 /* Now try balancing at a lower domain level of new_cpu */
1379 weight
= cpus_weight(span
);
1380 for_each_domain(cpu
, tmp
) {
1381 if (weight
<= cpus_weight(tmp
->span
))
1383 if (tmp
->flags
& flag
)
1386 /* while loop will break here if sd == NULL */
1392 #endif /* CONFIG_SMP */
1395 * wake_idle() will wake a task on an idle cpu if task->cpu is
1396 * not idle and an idle cpu is available. The span of cpus to
1397 * search starts with cpus closest then further out as needed,
1398 * so we always favor a closer, idle cpu.
1400 * Returns the CPU we should wake onto.
1402 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1403 static int wake_idle(int cpu
, struct task_struct
*p
)
1406 struct sched_domain
*sd
;
1410 * If it is idle, then it is the best cpu to run this task.
1412 * This cpu is also the best, if it has more than one task already.
1413 * Siblings must be also busy(in most cases) as they didn't already
1414 * pickup the extra load from this cpu and hence we need not check
1415 * sibling runqueue info. This will avoid the checks and cache miss
1416 * penalities associated with that.
1418 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1421 for_each_domain(cpu
, sd
) {
1422 if (sd
->flags
& SD_WAKE_IDLE
) {
1423 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1424 for_each_cpu_mask(i
, tmp
) {
1435 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1442 * try_to_wake_up - wake up a thread
1443 * @p: the to-be-woken-up thread
1444 * @state: the mask of task states that can be woken
1445 * @sync: do a synchronous wakeup?
1447 * Put it on the run-queue if it's not already there. The "current"
1448 * thread is always on the run-queue (except when the actual
1449 * re-schedule is in progress), and as such you're allowed to do
1450 * the simpler "current->state = TASK_RUNNING" to mark yourself
1451 * runnable without the overhead of this.
1453 * returns failure only if the task is already active.
1455 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1457 int cpu
, this_cpu
, success
= 0;
1458 unsigned long flags
;
1462 struct sched_domain
*sd
, *this_sd
= NULL
;
1463 unsigned long load
, this_load
;
1467 rq
= task_rq_lock(p
, &flags
);
1468 old_state
= p
->state
;
1469 if (!(old_state
& state
))
1476 this_cpu
= smp_processor_id();
1479 if (unlikely(task_running(rq
, p
)))
1484 schedstat_inc(rq
, ttwu_cnt
);
1485 if (cpu
== this_cpu
) {
1486 schedstat_inc(rq
, ttwu_local
);
1490 for_each_domain(this_cpu
, sd
) {
1491 if (cpu_isset(cpu
, sd
->span
)) {
1492 schedstat_inc(sd
, ttwu_wake_remote
);
1498 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1502 * Check for affine wakeup and passive balancing possibilities.
1505 int idx
= this_sd
->wake_idx
;
1506 unsigned int imbalance
;
1508 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1510 load
= source_load(cpu
, idx
);
1511 this_load
= target_load(this_cpu
, idx
);
1513 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1515 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1516 unsigned long tl
= this_load
;
1517 unsigned long tl_per_task
;
1519 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1522 * If sync wakeup then subtract the (maximum possible)
1523 * effect of the currently running task from the load
1524 * of the current CPU:
1527 tl
-= current
->se
.load
.weight
;
1530 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1531 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1533 * This domain has SD_WAKE_AFFINE and
1534 * p is cache cold in this domain, and
1535 * there is no bad imbalance.
1537 schedstat_inc(this_sd
, ttwu_move_affine
);
1543 * Start passive balancing when half the imbalance_pct
1546 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1547 if (imbalance
*this_load
<= 100*load
) {
1548 schedstat_inc(this_sd
, ttwu_move_balance
);
1554 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1556 new_cpu
= wake_idle(new_cpu
, p
);
1557 if (new_cpu
!= cpu
) {
1558 set_task_cpu(p
, new_cpu
);
1559 task_rq_unlock(rq
, &flags
);
1560 /* might preempt at this point */
1561 rq
= task_rq_lock(p
, &flags
);
1562 old_state
= p
->state
;
1563 if (!(old_state
& state
))
1568 this_cpu
= smp_processor_id();
1573 #endif /* CONFIG_SMP */
1574 update_rq_clock(rq
);
1575 activate_task(rq
, p
, 1);
1577 * Sync wakeups (i.e. those types of wakeups where the waker
1578 * has indicated that it will leave the CPU in short order)
1579 * don't trigger a preemption, if the woken up task will run on
1580 * this cpu. (in this case the 'I will reschedule' promise of
1581 * the waker guarantees that the freshly woken up task is going
1582 * to be considered on this CPU.)
1584 if (!sync
|| cpu
!= this_cpu
)
1585 check_preempt_curr(rq
, p
);
1589 p
->state
= TASK_RUNNING
;
1591 task_rq_unlock(rq
, &flags
);
1596 int fastcall
wake_up_process(struct task_struct
*p
)
1598 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1599 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1601 EXPORT_SYMBOL(wake_up_process
);
1603 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1605 return try_to_wake_up(p
, state
, 0);
1609 * Perform scheduler related setup for a newly forked process p.
1610 * p is forked by current.
1612 * __sched_fork() is basic setup used by init_idle() too:
1614 static void __sched_fork(struct task_struct
*p
)
1616 p
->se
.exec_start
= 0;
1617 p
->se
.sum_exec_runtime
= 0;
1618 p
->se
.prev_sum_exec_runtime
= 0;
1620 #ifdef CONFIG_SCHEDSTATS
1621 p
->se
.wait_start
= 0;
1622 p
->se
.sum_sleep_runtime
= 0;
1623 p
->se
.sleep_start
= 0;
1624 p
->se
.block_start
= 0;
1625 p
->se
.sleep_max
= 0;
1626 p
->se
.block_max
= 0;
1628 p
->se
.slice_max
= 0;
1632 INIT_LIST_HEAD(&p
->run_list
);
1635 #ifdef CONFIG_PREEMPT_NOTIFIERS
1636 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1640 * We mark the process as running here, but have not actually
1641 * inserted it onto the runqueue yet. This guarantees that
1642 * nobody will actually run it, and a signal or other external
1643 * event cannot wake it up and insert it on the runqueue either.
1645 p
->state
= TASK_RUNNING
;
1649 * fork()/clone()-time setup:
1651 void sched_fork(struct task_struct
*p
, int clone_flags
)
1653 int cpu
= get_cpu();
1658 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1660 __set_task_cpu(p
, cpu
);
1663 * Make sure we do not leak PI boosting priority to the child:
1665 p
->prio
= current
->normal_prio
;
1667 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1668 if (likely(sched_info_on()))
1669 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1671 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1674 #ifdef CONFIG_PREEMPT
1675 /* Want to start with kernel preemption disabled. */
1676 task_thread_info(p
)->preempt_count
= 1;
1682 * wake_up_new_task - wake up a newly created task for the first time.
1684 * This function will do some initial scheduler statistics housekeeping
1685 * that must be done for every newly created context, then puts the task
1686 * on the runqueue and wakes it.
1688 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1690 unsigned long flags
;
1694 rq
= task_rq_lock(p
, &flags
);
1695 BUG_ON(p
->state
!= TASK_RUNNING
);
1696 this_cpu
= smp_processor_id(); /* parent's CPU */
1697 update_rq_clock(rq
);
1699 p
->prio
= effective_prio(p
);
1701 if (rt_prio(p
->prio
))
1702 p
->sched_class
= &rt_sched_class
;
1704 p
->sched_class
= &fair_sched_class
;
1706 if (task_cpu(p
) != this_cpu
|| !p
->sched_class
->task_new
||
1707 !current
->se
.on_rq
) {
1708 activate_task(rq
, p
, 0);
1711 * Let the scheduling class do new task startup
1712 * management (if any):
1714 p
->sched_class
->task_new(rq
, p
);
1715 inc_nr_running(p
, rq
);
1717 check_preempt_curr(rq
, p
);
1718 task_rq_unlock(rq
, &flags
);
1721 #ifdef CONFIG_PREEMPT_NOTIFIERS
1724 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1725 * @notifier: notifier struct to register
1727 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1729 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1731 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1734 * preempt_notifier_unregister - no longer interested in preemption notifications
1735 * @notifier: notifier struct to unregister
1737 * This is safe to call from within a preemption notifier.
1739 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1741 hlist_del(¬ifier
->link
);
1743 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1745 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1747 struct preempt_notifier
*notifier
;
1748 struct hlist_node
*node
;
1750 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1751 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1755 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1756 struct task_struct
*next
)
1758 struct preempt_notifier
*notifier
;
1759 struct hlist_node
*node
;
1761 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1762 notifier
->ops
->sched_out(notifier
, next
);
1767 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1772 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1773 struct task_struct
*next
)
1780 * prepare_task_switch - prepare to switch tasks
1781 * @rq: the runqueue preparing to switch
1782 * @prev: the current task that is being switched out
1783 * @next: the task we are going to switch to.
1785 * This is called with the rq lock held and interrupts off. It must
1786 * be paired with a subsequent finish_task_switch after the context
1789 * prepare_task_switch sets up locking and calls architecture specific
1793 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1794 struct task_struct
*next
)
1796 fire_sched_out_preempt_notifiers(prev
, next
);
1797 prepare_lock_switch(rq
, next
);
1798 prepare_arch_switch(next
);
1802 * finish_task_switch - clean up after a task-switch
1803 * @rq: runqueue associated with task-switch
1804 * @prev: the thread we just switched away from.
1806 * finish_task_switch must be called after the context switch, paired
1807 * with a prepare_task_switch call before the context switch.
1808 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1809 * and do any other architecture-specific cleanup actions.
1811 * Note that we may have delayed dropping an mm in context_switch(). If
1812 * so, we finish that here outside of the runqueue lock. (Doing it
1813 * with the lock held can cause deadlocks; see schedule() for
1816 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1817 __releases(rq
->lock
)
1819 struct mm_struct
*mm
= rq
->prev_mm
;
1825 * A task struct has one reference for the use as "current".
1826 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1827 * schedule one last time. The schedule call will never return, and
1828 * the scheduled task must drop that reference.
1829 * The test for TASK_DEAD must occur while the runqueue locks are
1830 * still held, otherwise prev could be scheduled on another cpu, die
1831 * there before we look at prev->state, and then the reference would
1833 * Manfred Spraul <manfred@colorfullife.com>
1835 prev_state
= prev
->state
;
1836 finish_arch_switch(prev
);
1837 finish_lock_switch(rq
, prev
);
1838 fire_sched_in_preempt_notifiers(current
);
1841 if (unlikely(prev_state
== TASK_DEAD
)) {
1843 * Remove function-return probe instances associated with this
1844 * task and put them back on the free list.
1846 kprobe_flush_task(prev
);
1847 put_task_struct(prev
);
1852 * schedule_tail - first thing a freshly forked thread must call.
1853 * @prev: the thread we just switched away from.
1855 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1856 __releases(rq
->lock
)
1858 struct rq
*rq
= this_rq();
1860 finish_task_switch(rq
, prev
);
1861 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1862 /* In this case, finish_task_switch does not reenable preemption */
1865 if (current
->set_child_tid
)
1866 put_user(current
->pid
, current
->set_child_tid
);
1870 * context_switch - switch to the new MM and the new
1871 * thread's register state.
1874 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1875 struct task_struct
*next
)
1877 struct mm_struct
*mm
, *oldmm
;
1879 prepare_task_switch(rq
, prev
, next
);
1881 oldmm
= prev
->active_mm
;
1883 * For paravirt, this is coupled with an exit in switch_to to
1884 * combine the page table reload and the switch backend into
1887 arch_enter_lazy_cpu_mode();
1889 if (unlikely(!mm
)) {
1890 next
->active_mm
= oldmm
;
1891 atomic_inc(&oldmm
->mm_count
);
1892 enter_lazy_tlb(oldmm
, next
);
1894 switch_mm(oldmm
, mm
, next
);
1896 if (unlikely(!prev
->mm
)) {
1897 prev
->active_mm
= NULL
;
1898 rq
->prev_mm
= oldmm
;
1901 * Since the runqueue lock will be released by the next
1902 * task (which is an invalid locking op but in the case
1903 * of the scheduler it's an obvious special-case), so we
1904 * do an early lockdep release here:
1906 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1907 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1910 /* Here we just switch the register state and the stack. */
1911 switch_to(prev
, next
, prev
);
1915 * this_rq must be evaluated again because prev may have moved
1916 * CPUs since it called schedule(), thus the 'rq' on its stack
1917 * frame will be invalid.
1919 finish_task_switch(this_rq(), prev
);
1923 * nr_running, nr_uninterruptible and nr_context_switches:
1925 * externally visible scheduler statistics: current number of runnable
1926 * threads, current number of uninterruptible-sleeping threads, total
1927 * number of context switches performed since bootup.
1929 unsigned long nr_running(void)
1931 unsigned long i
, sum
= 0;
1933 for_each_online_cpu(i
)
1934 sum
+= cpu_rq(i
)->nr_running
;
1939 unsigned long nr_uninterruptible(void)
1941 unsigned long i
, sum
= 0;
1943 for_each_possible_cpu(i
)
1944 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1947 * Since we read the counters lockless, it might be slightly
1948 * inaccurate. Do not allow it to go below zero though:
1950 if (unlikely((long)sum
< 0))
1956 unsigned long long nr_context_switches(void)
1959 unsigned long long sum
= 0;
1961 for_each_possible_cpu(i
)
1962 sum
+= cpu_rq(i
)->nr_switches
;
1967 unsigned long nr_iowait(void)
1969 unsigned long i
, sum
= 0;
1971 for_each_possible_cpu(i
)
1972 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1977 unsigned long nr_active(void)
1979 unsigned long i
, running
= 0, uninterruptible
= 0;
1981 for_each_online_cpu(i
) {
1982 running
+= cpu_rq(i
)->nr_running
;
1983 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1986 if (unlikely((long)uninterruptible
< 0))
1987 uninterruptible
= 0;
1989 return running
+ uninterruptible
;
1993 * Update rq->cpu_load[] statistics. This function is usually called every
1994 * scheduler tick (TICK_NSEC).
1996 static void update_cpu_load(struct rq
*this_rq
)
1998 unsigned long this_load
= this_rq
->load
.weight
;
2001 this_rq
->nr_load_updates
++;
2003 /* Update our load: */
2004 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2005 unsigned long old_load
, new_load
;
2007 /* scale is effectively 1 << i now, and >> i divides by scale */
2009 old_load
= this_rq
->cpu_load
[i
];
2010 new_load
= this_load
;
2012 * Round up the averaging division if load is increasing. This
2013 * prevents us from getting stuck on 9 if the load is 10, for
2016 if (new_load
> old_load
)
2017 new_load
+= scale
-1;
2018 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2025 * double_rq_lock - safely lock two runqueues
2027 * Note this does not disable interrupts like task_rq_lock,
2028 * you need to do so manually before calling.
2030 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2031 __acquires(rq1
->lock
)
2032 __acquires(rq2
->lock
)
2034 BUG_ON(!irqs_disabled());
2036 spin_lock(&rq1
->lock
);
2037 __acquire(rq2
->lock
); /* Fake it out ;) */
2040 spin_lock(&rq1
->lock
);
2041 spin_lock(&rq2
->lock
);
2043 spin_lock(&rq2
->lock
);
2044 spin_lock(&rq1
->lock
);
2047 update_rq_clock(rq1
);
2048 update_rq_clock(rq2
);
2052 * double_rq_unlock - safely unlock two runqueues
2054 * Note this does not restore interrupts like task_rq_unlock,
2055 * you need to do so manually after calling.
2057 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2058 __releases(rq1
->lock
)
2059 __releases(rq2
->lock
)
2061 spin_unlock(&rq1
->lock
);
2063 spin_unlock(&rq2
->lock
);
2065 __release(rq2
->lock
);
2069 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2071 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2072 __releases(this_rq
->lock
)
2073 __acquires(busiest
->lock
)
2074 __acquires(this_rq
->lock
)
2076 if (unlikely(!irqs_disabled())) {
2077 /* printk() doesn't work good under rq->lock */
2078 spin_unlock(&this_rq
->lock
);
2081 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2082 if (busiest
< this_rq
) {
2083 spin_unlock(&this_rq
->lock
);
2084 spin_lock(&busiest
->lock
);
2085 spin_lock(&this_rq
->lock
);
2087 spin_lock(&busiest
->lock
);
2092 * If dest_cpu is allowed for this process, migrate the task to it.
2093 * This is accomplished by forcing the cpu_allowed mask to only
2094 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2095 * the cpu_allowed mask is restored.
2097 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2099 struct migration_req req
;
2100 unsigned long flags
;
2103 rq
= task_rq_lock(p
, &flags
);
2104 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2105 || unlikely(cpu_is_offline(dest_cpu
)))
2108 /* force the process onto the specified CPU */
2109 if (migrate_task(p
, dest_cpu
, &req
)) {
2110 /* Need to wait for migration thread (might exit: take ref). */
2111 struct task_struct
*mt
= rq
->migration_thread
;
2113 get_task_struct(mt
);
2114 task_rq_unlock(rq
, &flags
);
2115 wake_up_process(mt
);
2116 put_task_struct(mt
);
2117 wait_for_completion(&req
.done
);
2122 task_rq_unlock(rq
, &flags
);
2126 * sched_exec - execve() is a valuable balancing opportunity, because at
2127 * this point the task has the smallest effective memory and cache footprint.
2129 void sched_exec(void)
2131 int new_cpu
, this_cpu
= get_cpu();
2132 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2134 if (new_cpu
!= this_cpu
)
2135 sched_migrate_task(current
, new_cpu
);
2139 * pull_task - move a task from a remote runqueue to the local runqueue.
2140 * Both runqueues must be locked.
2142 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2143 struct rq
*this_rq
, int this_cpu
)
2145 deactivate_task(src_rq
, p
, 0);
2146 set_task_cpu(p
, this_cpu
);
2147 activate_task(this_rq
, p
, 0);
2149 * Note that idle threads have a prio of MAX_PRIO, for this test
2150 * to be always true for them.
2152 check_preempt_curr(this_rq
, p
);
2156 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2159 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2160 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2164 * We do not migrate tasks that are:
2165 * 1) running (obviously), or
2166 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2167 * 3) are cache-hot on their current CPU.
2169 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2173 if (task_running(rq
, p
))
2179 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2180 unsigned long max_nr_move
, unsigned long max_load_move
,
2181 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2182 int *all_pinned
, unsigned long *load_moved
,
2183 int *this_best_prio
, struct rq_iterator
*iterator
)
2185 int pulled
= 0, pinned
= 0, skip_for_load
;
2186 struct task_struct
*p
;
2187 long rem_load_move
= max_load_move
;
2189 if (max_nr_move
== 0 || max_load_move
== 0)
2195 * Start the load-balancing iterator:
2197 p
= iterator
->start(iterator
->arg
);
2202 * To help distribute high priority tasks accross CPUs we don't
2203 * skip a task if it will be the highest priority task (i.e. smallest
2204 * prio value) on its new queue regardless of its load weight
2206 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2207 SCHED_LOAD_SCALE_FUZZ
;
2208 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2209 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2210 p
= iterator
->next(iterator
->arg
);
2214 pull_task(busiest
, p
, this_rq
, this_cpu
);
2216 rem_load_move
-= p
->se
.load
.weight
;
2219 * We only want to steal up to the prescribed number of tasks
2220 * and the prescribed amount of weighted load.
2222 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2223 if (p
->prio
< *this_best_prio
)
2224 *this_best_prio
= p
->prio
;
2225 p
= iterator
->next(iterator
->arg
);
2230 * Right now, this is the only place pull_task() is called,
2231 * so we can safely collect pull_task() stats here rather than
2232 * inside pull_task().
2234 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2237 *all_pinned
= pinned
;
2238 *load_moved
= max_load_move
- rem_load_move
;
2243 * move_tasks tries to move up to max_load_move weighted load from busiest to
2244 * this_rq, as part of a balancing operation within domain "sd".
2245 * Returns 1 if successful and 0 otherwise.
2247 * Called with both runqueues locked.
2249 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2250 unsigned long max_load_move
,
2251 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2254 struct sched_class
*class = sched_class_highest
;
2255 unsigned long total_load_moved
= 0;
2256 int this_best_prio
= this_rq
->curr
->prio
;
2260 class->load_balance(this_rq
, this_cpu
, busiest
,
2261 ULONG_MAX
, max_load_move
- total_load_moved
,
2262 sd
, idle
, all_pinned
, &this_best_prio
);
2263 class = class->next
;
2264 } while (class && max_load_move
> total_load_moved
);
2266 return total_load_moved
> 0;
2270 * move_one_task tries to move exactly one task from busiest to this_rq, as
2271 * part of active balancing operations within "domain".
2272 * Returns 1 if successful and 0 otherwise.
2274 * Called with both runqueues locked.
2276 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2277 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2279 struct sched_class
*class;
2280 int this_best_prio
= MAX_PRIO
;
2282 for (class = sched_class_highest
; class; class = class->next
)
2283 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2284 1, ULONG_MAX
, sd
, idle
, NULL
,
2292 * find_busiest_group finds and returns the busiest CPU group within the
2293 * domain. It calculates and returns the amount of weighted load which
2294 * should be moved to restore balance via the imbalance parameter.
2296 static struct sched_group
*
2297 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2298 unsigned long *imbalance
, enum cpu_idle_type idle
,
2299 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2301 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2302 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2303 unsigned long max_pull
;
2304 unsigned long busiest_load_per_task
, busiest_nr_running
;
2305 unsigned long this_load_per_task
, this_nr_running
;
2307 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2308 int power_savings_balance
= 1;
2309 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2310 unsigned long min_nr_running
= ULONG_MAX
;
2311 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2314 max_load
= this_load
= total_load
= total_pwr
= 0;
2315 busiest_load_per_task
= busiest_nr_running
= 0;
2316 this_load_per_task
= this_nr_running
= 0;
2317 if (idle
== CPU_NOT_IDLE
)
2318 load_idx
= sd
->busy_idx
;
2319 else if (idle
== CPU_NEWLY_IDLE
)
2320 load_idx
= sd
->newidle_idx
;
2322 load_idx
= sd
->idle_idx
;
2325 unsigned long load
, group_capacity
;
2328 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2329 unsigned long sum_nr_running
, sum_weighted_load
;
2331 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2334 balance_cpu
= first_cpu(group
->cpumask
);
2336 /* Tally up the load of all CPUs in the group */
2337 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2339 for_each_cpu_mask(i
, group
->cpumask
) {
2342 if (!cpu_isset(i
, *cpus
))
2347 if (*sd_idle
&& rq
->nr_running
)
2350 /* Bias balancing toward cpus of our domain */
2352 if (idle_cpu(i
) && !first_idle_cpu
) {
2357 load
= target_load(i
, load_idx
);
2359 load
= source_load(i
, load_idx
);
2362 sum_nr_running
+= rq
->nr_running
;
2363 sum_weighted_load
+= weighted_cpuload(i
);
2367 * First idle cpu or the first cpu(busiest) in this sched group
2368 * is eligible for doing load balancing at this and above
2369 * domains. In the newly idle case, we will allow all the cpu's
2370 * to do the newly idle load balance.
2372 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2373 balance_cpu
!= this_cpu
&& balance
) {
2378 total_load
+= avg_load
;
2379 total_pwr
+= group
->__cpu_power
;
2381 /* Adjust by relative CPU power of the group */
2382 avg_load
= sg_div_cpu_power(group
,
2383 avg_load
* SCHED_LOAD_SCALE
);
2385 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2388 this_load
= avg_load
;
2390 this_nr_running
= sum_nr_running
;
2391 this_load_per_task
= sum_weighted_load
;
2392 } else if (avg_load
> max_load
&&
2393 sum_nr_running
> group_capacity
) {
2394 max_load
= avg_load
;
2396 busiest_nr_running
= sum_nr_running
;
2397 busiest_load_per_task
= sum_weighted_load
;
2400 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2402 * Busy processors will not participate in power savings
2405 if (idle
== CPU_NOT_IDLE
||
2406 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2410 * If the local group is idle or completely loaded
2411 * no need to do power savings balance at this domain
2413 if (local_group
&& (this_nr_running
>= group_capacity
||
2415 power_savings_balance
= 0;
2418 * If a group is already running at full capacity or idle,
2419 * don't include that group in power savings calculations
2421 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2426 * Calculate the group which has the least non-idle load.
2427 * This is the group from where we need to pick up the load
2430 if ((sum_nr_running
< min_nr_running
) ||
2431 (sum_nr_running
== min_nr_running
&&
2432 first_cpu(group
->cpumask
) <
2433 first_cpu(group_min
->cpumask
))) {
2435 min_nr_running
= sum_nr_running
;
2436 min_load_per_task
= sum_weighted_load
/
2441 * Calculate the group which is almost near its
2442 * capacity but still has some space to pick up some load
2443 * from other group and save more power
2445 if (sum_nr_running
<= group_capacity
- 1) {
2446 if (sum_nr_running
> leader_nr_running
||
2447 (sum_nr_running
== leader_nr_running
&&
2448 first_cpu(group
->cpumask
) >
2449 first_cpu(group_leader
->cpumask
))) {
2450 group_leader
= group
;
2451 leader_nr_running
= sum_nr_running
;
2456 group
= group
->next
;
2457 } while (group
!= sd
->groups
);
2459 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2462 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2464 if (this_load
>= avg_load
||
2465 100*max_load
<= sd
->imbalance_pct
*this_load
)
2468 busiest_load_per_task
/= busiest_nr_running
;
2470 * We're trying to get all the cpus to the average_load, so we don't
2471 * want to push ourselves above the average load, nor do we wish to
2472 * reduce the max loaded cpu below the average load, as either of these
2473 * actions would just result in more rebalancing later, and ping-pong
2474 * tasks around. Thus we look for the minimum possible imbalance.
2475 * Negative imbalances (*we* are more loaded than anyone else) will
2476 * be counted as no imbalance for these purposes -- we can't fix that
2477 * by pulling tasks to us. Be careful of negative numbers as they'll
2478 * appear as very large values with unsigned longs.
2480 if (max_load
<= busiest_load_per_task
)
2484 * In the presence of smp nice balancing, certain scenarios can have
2485 * max load less than avg load(as we skip the groups at or below
2486 * its cpu_power, while calculating max_load..)
2488 if (max_load
< avg_load
) {
2490 goto small_imbalance
;
2493 /* Don't want to pull so many tasks that a group would go idle */
2494 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2496 /* How much load to actually move to equalise the imbalance */
2497 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2498 (avg_load
- this_load
) * this->__cpu_power
)
2502 * if *imbalance is less than the average load per runnable task
2503 * there is no gaurantee that any tasks will be moved so we'll have
2504 * a think about bumping its value to force at least one task to be
2507 if (*imbalance
< busiest_load_per_task
) {
2508 unsigned long tmp
, pwr_now
, pwr_move
;
2512 pwr_move
= pwr_now
= 0;
2514 if (this_nr_running
) {
2515 this_load_per_task
/= this_nr_running
;
2516 if (busiest_load_per_task
> this_load_per_task
)
2519 this_load_per_task
= SCHED_LOAD_SCALE
;
2521 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2522 busiest_load_per_task
* imbn
) {
2523 *imbalance
= busiest_load_per_task
;
2528 * OK, we don't have enough imbalance to justify moving tasks,
2529 * however we may be able to increase total CPU power used by
2533 pwr_now
+= busiest
->__cpu_power
*
2534 min(busiest_load_per_task
, max_load
);
2535 pwr_now
+= this->__cpu_power
*
2536 min(this_load_per_task
, this_load
);
2537 pwr_now
/= SCHED_LOAD_SCALE
;
2539 /* Amount of load we'd subtract */
2540 tmp
= sg_div_cpu_power(busiest
,
2541 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2543 pwr_move
+= busiest
->__cpu_power
*
2544 min(busiest_load_per_task
, max_load
- tmp
);
2546 /* Amount of load we'd add */
2547 if (max_load
* busiest
->__cpu_power
<
2548 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2549 tmp
= sg_div_cpu_power(this,
2550 max_load
* busiest
->__cpu_power
);
2552 tmp
= sg_div_cpu_power(this,
2553 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2554 pwr_move
+= this->__cpu_power
*
2555 min(this_load_per_task
, this_load
+ tmp
);
2556 pwr_move
/= SCHED_LOAD_SCALE
;
2558 /* Move if we gain throughput */
2559 if (pwr_move
> pwr_now
)
2560 *imbalance
= busiest_load_per_task
;
2566 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2567 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2570 if (this == group_leader
&& group_leader
!= group_min
) {
2571 *imbalance
= min_load_per_task
;
2581 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2584 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2585 unsigned long imbalance
, cpumask_t
*cpus
)
2587 struct rq
*busiest
= NULL
, *rq
;
2588 unsigned long max_load
= 0;
2591 for_each_cpu_mask(i
, group
->cpumask
) {
2594 if (!cpu_isset(i
, *cpus
))
2598 wl
= weighted_cpuload(i
);
2600 if (rq
->nr_running
== 1 && wl
> imbalance
)
2603 if (wl
> max_load
) {
2613 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2614 * so long as it is large enough.
2616 #define MAX_PINNED_INTERVAL 512
2619 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2620 * tasks if there is an imbalance.
2622 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2623 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2626 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2627 struct sched_group
*group
;
2628 unsigned long imbalance
;
2630 cpumask_t cpus
= CPU_MASK_ALL
;
2631 unsigned long flags
;
2634 * When power savings policy is enabled for the parent domain, idle
2635 * sibling can pick up load irrespective of busy siblings. In this case,
2636 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2637 * portraying it as CPU_NOT_IDLE.
2639 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2640 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2643 schedstat_inc(sd
, lb_cnt
[idle
]);
2646 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2653 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2657 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2659 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2663 BUG_ON(busiest
== this_rq
);
2665 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2668 if (busiest
->nr_running
> 1) {
2670 * Attempt to move tasks. If find_busiest_group has found
2671 * an imbalance but busiest->nr_running <= 1, the group is
2672 * still unbalanced. ld_moved simply stays zero, so it is
2673 * correctly treated as an imbalance.
2675 local_irq_save(flags
);
2676 double_rq_lock(this_rq
, busiest
);
2677 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2678 imbalance
, sd
, idle
, &all_pinned
);
2679 double_rq_unlock(this_rq
, busiest
);
2680 local_irq_restore(flags
);
2683 * some other cpu did the load balance for us.
2685 if (ld_moved
&& this_cpu
!= smp_processor_id())
2686 resched_cpu(this_cpu
);
2688 /* All tasks on this runqueue were pinned by CPU affinity */
2689 if (unlikely(all_pinned
)) {
2690 cpu_clear(cpu_of(busiest
), cpus
);
2691 if (!cpus_empty(cpus
))
2698 schedstat_inc(sd
, lb_failed
[idle
]);
2699 sd
->nr_balance_failed
++;
2701 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2703 spin_lock_irqsave(&busiest
->lock
, flags
);
2705 /* don't kick the migration_thread, if the curr
2706 * task on busiest cpu can't be moved to this_cpu
2708 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2709 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2711 goto out_one_pinned
;
2714 if (!busiest
->active_balance
) {
2715 busiest
->active_balance
= 1;
2716 busiest
->push_cpu
= this_cpu
;
2719 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2721 wake_up_process(busiest
->migration_thread
);
2724 * We've kicked active balancing, reset the failure
2727 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2730 sd
->nr_balance_failed
= 0;
2732 if (likely(!active_balance
)) {
2733 /* We were unbalanced, so reset the balancing interval */
2734 sd
->balance_interval
= sd
->min_interval
;
2737 * If we've begun active balancing, start to back off. This
2738 * case may not be covered by the all_pinned logic if there
2739 * is only 1 task on the busy runqueue (because we don't call
2742 if (sd
->balance_interval
< sd
->max_interval
)
2743 sd
->balance_interval
*= 2;
2746 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2747 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2752 schedstat_inc(sd
, lb_balanced
[idle
]);
2754 sd
->nr_balance_failed
= 0;
2757 /* tune up the balancing interval */
2758 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2759 (sd
->balance_interval
< sd
->max_interval
))
2760 sd
->balance_interval
*= 2;
2762 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2763 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2769 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2770 * tasks if there is an imbalance.
2772 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2773 * this_rq is locked.
2776 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2778 struct sched_group
*group
;
2779 struct rq
*busiest
= NULL
;
2780 unsigned long imbalance
;
2784 cpumask_t cpus
= CPU_MASK_ALL
;
2787 * When power savings policy is enabled for the parent domain, idle
2788 * sibling can pick up load irrespective of busy siblings. In this case,
2789 * let the state of idle sibling percolate up as IDLE, instead of
2790 * portraying it as CPU_NOT_IDLE.
2792 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2793 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2796 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2798 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2799 &sd_idle
, &cpus
, NULL
);
2801 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2805 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2808 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2812 BUG_ON(busiest
== this_rq
);
2814 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2817 if (busiest
->nr_running
> 1) {
2818 /* Attempt to move tasks */
2819 double_lock_balance(this_rq
, busiest
);
2820 /* this_rq->clock is already updated */
2821 update_rq_clock(busiest
);
2822 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2823 imbalance
, sd
, CPU_NEWLY_IDLE
,
2825 spin_unlock(&busiest
->lock
);
2827 if (unlikely(all_pinned
)) {
2828 cpu_clear(cpu_of(busiest
), cpus
);
2829 if (!cpus_empty(cpus
))
2835 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2836 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2837 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2840 sd
->nr_balance_failed
= 0;
2845 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2846 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2847 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2849 sd
->nr_balance_failed
= 0;
2855 * idle_balance is called by schedule() if this_cpu is about to become
2856 * idle. Attempts to pull tasks from other CPUs.
2858 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2860 struct sched_domain
*sd
;
2861 int pulled_task
= -1;
2862 unsigned long next_balance
= jiffies
+ HZ
;
2864 for_each_domain(this_cpu
, sd
) {
2865 unsigned long interval
;
2867 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2870 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2871 /* If we've pulled tasks over stop searching: */
2872 pulled_task
= load_balance_newidle(this_cpu
,
2875 interval
= msecs_to_jiffies(sd
->balance_interval
);
2876 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2877 next_balance
= sd
->last_balance
+ interval
;
2881 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2883 * We are going idle. next_balance may be set based on
2884 * a busy processor. So reset next_balance.
2886 this_rq
->next_balance
= next_balance
;
2891 * active_load_balance is run by migration threads. It pushes running tasks
2892 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2893 * running on each physical CPU where possible, and avoids physical /
2894 * logical imbalances.
2896 * Called with busiest_rq locked.
2898 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2900 int target_cpu
= busiest_rq
->push_cpu
;
2901 struct sched_domain
*sd
;
2902 struct rq
*target_rq
;
2904 /* Is there any task to move? */
2905 if (busiest_rq
->nr_running
<= 1)
2908 target_rq
= cpu_rq(target_cpu
);
2911 * This condition is "impossible", if it occurs
2912 * we need to fix it. Originally reported by
2913 * Bjorn Helgaas on a 128-cpu setup.
2915 BUG_ON(busiest_rq
== target_rq
);
2917 /* move a task from busiest_rq to target_rq */
2918 double_lock_balance(busiest_rq
, target_rq
);
2919 update_rq_clock(busiest_rq
);
2920 update_rq_clock(target_rq
);
2922 /* Search for an sd spanning us and the target CPU. */
2923 for_each_domain(target_cpu
, sd
) {
2924 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2925 cpu_isset(busiest_cpu
, sd
->span
))
2930 schedstat_inc(sd
, alb_cnt
);
2932 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2934 schedstat_inc(sd
, alb_pushed
);
2936 schedstat_inc(sd
, alb_failed
);
2938 spin_unlock(&target_rq
->lock
);
2943 atomic_t load_balancer
;
2945 } nohz ____cacheline_aligned
= {
2946 .load_balancer
= ATOMIC_INIT(-1),
2947 .cpu_mask
= CPU_MASK_NONE
,
2951 * This routine will try to nominate the ilb (idle load balancing)
2952 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2953 * load balancing on behalf of all those cpus. If all the cpus in the system
2954 * go into this tickless mode, then there will be no ilb owner (as there is
2955 * no need for one) and all the cpus will sleep till the next wakeup event
2958 * For the ilb owner, tick is not stopped. And this tick will be used
2959 * for idle load balancing. ilb owner will still be part of
2962 * While stopping the tick, this cpu will become the ilb owner if there
2963 * is no other owner. And will be the owner till that cpu becomes busy
2964 * or if all cpus in the system stop their ticks at which point
2965 * there is no need for ilb owner.
2967 * When the ilb owner becomes busy, it nominates another owner, during the
2968 * next busy scheduler_tick()
2970 int select_nohz_load_balancer(int stop_tick
)
2972 int cpu
= smp_processor_id();
2975 cpu_set(cpu
, nohz
.cpu_mask
);
2976 cpu_rq(cpu
)->in_nohz_recently
= 1;
2979 * If we are going offline and still the leader, give up!
2981 if (cpu_is_offline(cpu
) &&
2982 atomic_read(&nohz
.load_balancer
) == cpu
) {
2983 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2988 /* time for ilb owner also to sleep */
2989 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2990 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2991 atomic_set(&nohz
.load_balancer
, -1);
2995 if (atomic_read(&nohz
.load_balancer
) == -1) {
2996 /* make me the ilb owner */
2997 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2999 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3002 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3005 cpu_clear(cpu
, nohz
.cpu_mask
);
3007 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3008 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3015 static DEFINE_SPINLOCK(balancing
);
3018 * It checks each scheduling domain to see if it is due to be balanced,
3019 * and initiates a balancing operation if so.
3021 * Balancing parameters are set up in arch_init_sched_domains.
3023 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3026 struct rq
*rq
= cpu_rq(cpu
);
3027 unsigned long interval
;
3028 struct sched_domain
*sd
;
3029 /* Earliest time when we have to do rebalance again */
3030 unsigned long next_balance
= jiffies
+ 60*HZ
;
3031 int update_next_balance
= 0;
3033 for_each_domain(cpu
, sd
) {
3034 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3037 interval
= sd
->balance_interval
;
3038 if (idle
!= CPU_IDLE
)
3039 interval
*= sd
->busy_factor
;
3041 /* scale ms to jiffies */
3042 interval
= msecs_to_jiffies(interval
);
3043 if (unlikely(!interval
))
3045 if (interval
> HZ
*NR_CPUS
/10)
3046 interval
= HZ
*NR_CPUS
/10;
3049 if (sd
->flags
& SD_SERIALIZE
) {
3050 if (!spin_trylock(&balancing
))
3054 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3055 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3057 * We've pulled tasks over so either we're no
3058 * longer idle, or one of our SMT siblings is
3061 idle
= CPU_NOT_IDLE
;
3063 sd
->last_balance
= jiffies
;
3065 if (sd
->flags
& SD_SERIALIZE
)
3066 spin_unlock(&balancing
);
3068 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3069 next_balance
= sd
->last_balance
+ interval
;
3070 update_next_balance
= 1;
3074 * Stop the load balance at this level. There is another
3075 * CPU in our sched group which is doing load balancing more
3083 * next_balance will be updated only when there is a need.
3084 * When the cpu is attached to null domain for ex, it will not be
3087 if (likely(update_next_balance
))
3088 rq
->next_balance
= next_balance
;
3092 * run_rebalance_domains is triggered when needed from the scheduler tick.
3093 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3094 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3096 static void run_rebalance_domains(struct softirq_action
*h
)
3098 int this_cpu
= smp_processor_id();
3099 struct rq
*this_rq
= cpu_rq(this_cpu
);
3100 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3101 CPU_IDLE
: CPU_NOT_IDLE
;
3103 rebalance_domains(this_cpu
, idle
);
3107 * If this cpu is the owner for idle load balancing, then do the
3108 * balancing on behalf of the other idle cpus whose ticks are
3111 if (this_rq
->idle_at_tick
&&
3112 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3113 cpumask_t cpus
= nohz
.cpu_mask
;
3117 cpu_clear(this_cpu
, cpus
);
3118 for_each_cpu_mask(balance_cpu
, cpus
) {
3120 * If this cpu gets work to do, stop the load balancing
3121 * work being done for other cpus. Next load
3122 * balancing owner will pick it up.
3127 rebalance_domains(balance_cpu
, CPU_IDLE
);
3129 rq
= cpu_rq(balance_cpu
);
3130 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3131 this_rq
->next_balance
= rq
->next_balance
;
3138 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3140 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3141 * idle load balancing owner or decide to stop the periodic load balancing,
3142 * if the whole system is idle.
3144 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3148 * If we were in the nohz mode recently and busy at the current
3149 * scheduler tick, then check if we need to nominate new idle
3152 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3153 rq
->in_nohz_recently
= 0;
3155 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3156 cpu_clear(cpu
, nohz
.cpu_mask
);
3157 atomic_set(&nohz
.load_balancer
, -1);
3160 if (atomic_read(&nohz
.load_balancer
) == -1) {
3162 * simple selection for now: Nominate the
3163 * first cpu in the nohz list to be the next
3166 * TBD: Traverse the sched domains and nominate
3167 * the nearest cpu in the nohz.cpu_mask.
3169 int ilb
= first_cpu(nohz
.cpu_mask
);
3177 * If this cpu is idle and doing idle load balancing for all the
3178 * cpus with ticks stopped, is it time for that to stop?
3180 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3181 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3187 * If this cpu is idle and the idle load balancing is done by
3188 * someone else, then no need raise the SCHED_SOFTIRQ
3190 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3191 cpu_isset(cpu
, nohz
.cpu_mask
))
3194 if (time_after_eq(jiffies
, rq
->next_balance
))
3195 raise_softirq(SCHED_SOFTIRQ
);
3198 #else /* CONFIG_SMP */
3201 * on UP we do not need to balance between CPUs:
3203 static inline void idle_balance(int cpu
, struct rq
*rq
)
3207 /* Avoid "used but not defined" warning on UP */
3208 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3209 unsigned long max_nr_move
, unsigned long max_load_move
,
3210 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3211 int *all_pinned
, unsigned long *load_moved
,
3212 int *this_best_prio
, struct rq_iterator
*iterator
)
3221 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3223 EXPORT_PER_CPU_SYMBOL(kstat
);
3226 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3227 * that have not yet been banked in case the task is currently running.
3229 unsigned long long task_sched_runtime(struct task_struct
*p
)
3231 unsigned long flags
;
3235 rq
= task_rq_lock(p
, &flags
);
3236 ns
= p
->se
.sum_exec_runtime
;
3237 if (rq
->curr
== p
) {
3238 update_rq_clock(rq
);
3239 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3240 if ((s64
)delta_exec
> 0)
3243 task_rq_unlock(rq
, &flags
);
3249 * Account user cpu time to a process.
3250 * @p: the process that the cpu time gets accounted to
3251 * @hardirq_offset: the offset to subtract from hardirq_count()
3252 * @cputime: the cpu time spent in user space since the last update
3254 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3256 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3259 p
->utime
= cputime_add(p
->utime
, cputime
);
3261 /* Add user time to cpustat. */
3262 tmp
= cputime_to_cputime64(cputime
);
3263 if (TASK_NICE(p
) > 0)
3264 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3266 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3270 * Account system cpu time to a process.
3271 * @p: the process that the cpu time gets accounted to
3272 * @hardirq_offset: the offset to subtract from hardirq_count()
3273 * @cputime: the cpu time spent in kernel space since the last update
3275 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3278 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3279 struct rq
*rq
= this_rq();
3282 p
->stime
= cputime_add(p
->stime
, cputime
);
3284 /* Add system time to cpustat. */
3285 tmp
= cputime_to_cputime64(cputime
);
3286 if (hardirq_count() - hardirq_offset
)
3287 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3288 else if (softirq_count())
3289 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3290 else if (p
!= rq
->idle
)
3291 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3292 else if (atomic_read(&rq
->nr_iowait
) > 0)
3293 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3295 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3296 /* Account for system time used */
3297 acct_update_integrals(p
);
3301 * Account for involuntary wait time.
3302 * @p: the process from which the cpu time has been stolen
3303 * @steal: the cpu time spent in involuntary wait
3305 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3307 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3308 cputime64_t tmp
= cputime_to_cputime64(steal
);
3309 struct rq
*rq
= this_rq();
3311 if (p
== rq
->idle
) {
3312 p
->stime
= cputime_add(p
->stime
, steal
);
3313 if (atomic_read(&rq
->nr_iowait
) > 0)
3314 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3316 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3318 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3322 * This function gets called by the timer code, with HZ frequency.
3323 * We call it with interrupts disabled.
3325 * It also gets called by the fork code, when changing the parent's
3328 void scheduler_tick(void)
3330 int cpu
= smp_processor_id();
3331 struct rq
*rq
= cpu_rq(cpu
);
3332 struct task_struct
*curr
= rq
->curr
;
3333 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3335 spin_lock(&rq
->lock
);
3336 __update_rq_clock(rq
);
3338 * Let rq->clock advance by at least TICK_NSEC:
3340 if (unlikely(rq
->clock
< next_tick
))
3341 rq
->clock
= next_tick
;
3342 rq
->tick_timestamp
= rq
->clock
;
3343 update_cpu_load(rq
);
3344 if (curr
!= rq
->idle
) /* FIXME: needed? */
3345 curr
->sched_class
->task_tick(rq
, curr
);
3346 spin_unlock(&rq
->lock
);
3349 rq
->idle_at_tick
= idle_cpu(cpu
);
3350 trigger_load_balance(rq
, cpu
);
3354 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3356 void fastcall
add_preempt_count(int val
)
3361 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3363 preempt_count() += val
;
3365 * Spinlock count overflowing soon?
3367 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3370 EXPORT_SYMBOL(add_preempt_count
);
3372 void fastcall
sub_preempt_count(int val
)
3377 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3380 * Is the spinlock portion underflowing?
3382 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3383 !(preempt_count() & PREEMPT_MASK
)))
3386 preempt_count() -= val
;
3388 EXPORT_SYMBOL(sub_preempt_count
);
3393 * Print scheduling while atomic bug:
3395 static noinline
void __schedule_bug(struct task_struct
*prev
)
3397 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3398 prev
->comm
, preempt_count(), prev
->pid
);
3399 debug_show_held_locks(prev
);
3400 if (irqs_disabled())
3401 print_irqtrace_events(prev
);
3406 * Various schedule()-time debugging checks and statistics:
3408 static inline void schedule_debug(struct task_struct
*prev
)
3411 * Test if we are atomic. Since do_exit() needs to call into
3412 * schedule() atomically, we ignore that path for now.
3413 * Otherwise, whine if we are scheduling when we should not be.
3415 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3416 __schedule_bug(prev
);
3418 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3420 schedstat_inc(this_rq(), sched_cnt
);
3421 #ifdef CONFIG_SCHEDSTATS
3422 if (unlikely(prev
->lock_depth
>= 0)) {
3423 schedstat_inc(this_rq(), bkl_cnt
);
3424 schedstat_inc(prev
, sched_info
.bkl_cnt
);
3430 * Pick up the highest-prio task:
3432 static inline struct task_struct
*
3433 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3435 struct sched_class
*class;
3436 struct task_struct
*p
;
3439 * Optimization: we know that if all tasks are in
3440 * the fair class we can call that function directly:
3442 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3443 p
= fair_sched_class
.pick_next_task(rq
);
3448 class = sched_class_highest
;
3450 p
= class->pick_next_task(rq
);
3454 * Will never be NULL as the idle class always
3455 * returns a non-NULL p:
3457 class = class->next
;
3462 * schedule() is the main scheduler function.
3464 asmlinkage
void __sched
schedule(void)
3466 struct task_struct
*prev
, *next
;
3473 cpu
= smp_processor_id();
3477 switch_count
= &prev
->nivcsw
;
3479 release_kernel_lock(prev
);
3480 need_resched_nonpreemptible
:
3482 schedule_debug(prev
);
3484 spin_lock_irq(&rq
->lock
);
3485 clear_tsk_need_resched(prev
);
3486 __update_rq_clock(rq
);
3488 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3489 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3490 unlikely(signal_pending(prev
)))) {
3491 prev
->state
= TASK_RUNNING
;
3493 deactivate_task(rq
, prev
, 1);
3495 switch_count
= &prev
->nvcsw
;
3498 if (unlikely(!rq
->nr_running
))
3499 idle_balance(cpu
, rq
);
3501 prev
->sched_class
->put_prev_task(rq
, prev
);
3502 next
= pick_next_task(rq
, prev
);
3504 sched_info_switch(prev
, next
);
3506 if (likely(prev
!= next
)) {
3511 context_switch(rq
, prev
, next
); /* unlocks the rq */
3513 spin_unlock_irq(&rq
->lock
);
3515 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3516 cpu
= smp_processor_id();
3518 goto need_resched_nonpreemptible
;
3520 preempt_enable_no_resched();
3521 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3524 EXPORT_SYMBOL(schedule
);
3526 #ifdef CONFIG_PREEMPT
3528 * this is the entry point to schedule() from in-kernel preemption
3529 * off of preempt_enable. Kernel preemptions off return from interrupt
3530 * occur there and call schedule directly.
3532 asmlinkage
void __sched
preempt_schedule(void)
3534 struct thread_info
*ti
= current_thread_info();
3535 #ifdef CONFIG_PREEMPT_BKL
3536 struct task_struct
*task
= current
;
3537 int saved_lock_depth
;
3540 * If there is a non-zero preempt_count or interrupts are disabled,
3541 * we do not want to preempt the current task. Just return..
3543 if (likely(ti
->preempt_count
|| irqs_disabled()))
3547 add_preempt_count(PREEMPT_ACTIVE
);
3549 * We keep the big kernel semaphore locked, but we
3550 * clear ->lock_depth so that schedule() doesnt
3551 * auto-release the semaphore:
3553 #ifdef CONFIG_PREEMPT_BKL
3554 saved_lock_depth
= task
->lock_depth
;
3555 task
->lock_depth
= -1;
3558 #ifdef CONFIG_PREEMPT_BKL
3559 task
->lock_depth
= saved_lock_depth
;
3561 sub_preempt_count(PREEMPT_ACTIVE
);
3563 /* we could miss a preemption opportunity between schedule and now */
3565 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3568 EXPORT_SYMBOL(preempt_schedule
);
3571 * this is the entry point to schedule() from kernel preemption
3572 * off of irq context.
3573 * Note, that this is called and return with irqs disabled. This will
3574 * protect us against recursive calling from irq.
3576 asmlinkage
void __sched
preempt_schedule_irq(void)
3578 struct thread_info
*ti
= current_thread_info();
3579 #ifdef CONFIG_PREEMPT_BKL
3580 struct task_struct
*task
= current
;
3581 int saved_lock_depth
;
3583 /* Catch callers which need to be fixed */
3584 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3587 add_preempt_count(PREEMPT_ACTIVE
);
3589 * We keep the big kernel semaphore locked, but we
3590 * clear ->lock_depth so that schedule() doesnt
3591 * auto-release the semaphore:
3593 #ifdef CONFIG_PREEMPT_BKL
3594 saved_lock_depth
= task
->lock_depth
;
3595 task
->lock_depth
= -1;
3599 local_irq_disable();
3600 #ifdef CONFIG_PREEMPT_BKL
3601 task
->lock_depth
= saved_lock_depth
;
3603 sub_preempt_count(PREEMPT_ACTIVE
);
3605 /* we could miss a preemption opportunity between schedule and now */
3607 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3611 #endif /* CONFIG_PREEMPT */
3613 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3616 return try_to_wake_up(curr
->private, mode
, sync
);
3618 EXPORT_SYMBOL(default_wake_function
);
3621 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3622 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3623 * number) then we wake all the non-exclusive tasks and one exclusive task.
3625 * There are circumstances in which we can try to wake a task which has already
3626 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3627 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3629 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3630 int nr_exclusive
, int sync
, void *key
)
3632 wait_queue_t
*curr
, *next
;
3634 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3635 unsigned flags
= curr
->flags
;
3637 if (curr
->func(curr
, mode
, sync
, key
) &&
3638 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3644 * __wake_up - wake up threads blocked on a waitqueue.
3646 * @mode: which threads
3647 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3648 * @key: is directly passed to the wakeup function
3650 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3651 int nr_exclusive
, void *key
)
3653 unsigned long flags
;
3655 spin_lock_irqsave(&q
->lock
, flags
);
3656 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3657 spin_unlock_irqrestore(&q
->lock
, flags
);
3659 EXPORT_SYMBOL(__wake_up
);
3662 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3664 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3666 __wake_up_common(q
, mode
, 1, 0, NULL
);
3670 * __wake_up_sync - wake up threads blocked on a waitqueue.
3672 * @mode: which threads
3673 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3675 * The sync wakeup differs that the waker knows that it will schedule
3676 * away soon, so while the target thread will be woken up, it will not
3677 * be migrated to another CPU - ie. the two threads are 'synchronized'
3678 * with each other. This can prevent needless bouncing between CPUs.
3680 * On UP it can prevent extra preemption.
3683 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3685 unsigned long flags
;
3691 if (unlikely(!nr_exclusive
))
3694 spin_lock_irqsave(&q
->lock
, flags
);
3695 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3696 spin_unlock_irqrestore(&q
->lock
, flags
);
3698 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3700 void fastcall
complete(struct completion
*x
)
3702 unsigned long flags
;
3704 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3706 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3708 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3710 EXPORT_SYMBOL(complete
);
3712 void fastcall
complete_all(struct completion
*x
)
3714 unsigned long flags
;
3716 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3717 x
->done
+= UINT_MAX
/2;
3718 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3720 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3722 EXPORT_SYMBOL(complete_all
);
3724 void fastcall __sched
wait_for_completion(struct completion
*x
)
3728 spin_lock_irq(&x
->wait
.lock
);
3730 DECLARE_WAITQUEUE(wait
, current
);
3732 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3733 __add_wait_queue_tail(&x
->wait
, &wait
);
3735 __set_current_state(TASK_UNINTERRUPTIBLE
);
3736 spin_unlock_irq(&x
->wait
.lock
);
3738 spin_lock_irq(&x
->wait
.lock
);
3740 __remove_wait_queue(&x
->wait
, &wait
);
3743 spin_unlock_irq(&x
->wait
.lock
);
3745 EXPORT_SYMBOL(wait_for_completion
);
3747 unsigned long fastcall __sched
3748 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3752 spin_lock_irq(&x
->wait
.lock
);
3754 DECLARE_WAITQUEUE(wait
, current
);
3756 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3757 __add_wait_queue_tail(&x
->wait
, &wait
);
3759 __set_current_state(TASK_UNINTERRUPTIBLE
);
3760 spin_unlock_irq(&x
->wait
.lock
);
3761 timeout
= schedule_timeout(timeout
);
3762 spin_lock_irq(&x
->wait
.lock
);
3764 __remove_wait_queue(&x
->wait
, &wait
);
3768 __remove_wait_queue(&x
->wait
, &wait
);
3772 spin_unlock_irq(&x
->wait
.lock
);
3775 EXPORT_SYMBOL(wait_for_completion_timeout
);
3777 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3783 spin_lock_irq(&x
->wait
.lock
);
3785 DECLARE_WAITQUEUE(wait
, current
);
3787 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3788 __add_wait_queue_tail(&x
->wait
, &wait
);
3790 if (signal_pending(current
)) {
3792 __remove_wait_queue(&x
->wait
, &wait
);
3795 __set_current_state(TASK_INTERRUPTIBLE
);
3796 spin_unlock_irq(&x
->wait
.lock
);
3798 spin_lock_irq(&x
->wait
.lock
);
3800 __remove_wait_queue(&x
->wait
, &wait
);
3804 spin_unlock_irq(&x
->wait
.lock
);
3808 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3810 unsigned long fastcall __sched
3811 wait_for_completion_interruptible_timeout(struct completion
*x
,
3812 unsigned long timeout
)
3816 spin_lock_irq(&x
->wait
.lock
);
3818 DECLARE_WAITQUEUE(wait
, current
);
3820 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3821 __add_wait_queue_tail(&x
->wait
, &wait
);
3823 if (signal_pending(current
)) {
3824 timeout
= -ERESTARTSYS
;
3825 __remove_wait_queue(&x
->wait
, &wait
);
3828 __set_current_state(TASK_INTERRUPTIBLE
);
3829 spin_unlock_irq(&x
->wait
.lock
);
3830 timeout
= schedule_timeout(timeout
);
3831 spin_lock_irq(&x
->wait
.lock
);
3833 __remove_wait_queue(&x
->wait
, &wait
);
3837 __remove_wait_queue(&x
->wait
, &wait
);
3841 spin_unlock_irq(&x
->wait
.lock
);
3844 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3847 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3849 spin_lock_irqsave(&q
->lock
, *flags
);
3850 __add_wait_queue(q
, wait
);
3851 spin_unlock(&q
->lock
);
3855 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3857 spin_lock_irq(&q
->lock
);
3858 __remove_wait_queue(q
, wait
);
3859 spin_unlock_irqrestore(&q
->lock
, *flags
);
3862 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3864 unsigned long flags
;
3867 init_waitqueue_entry(&wait
, current
);
3869 current
->state
= TASK_INTERRUPTIBLE
;
3871 sleep_on_head(q
, &wait
, &flags
);
3873 sleep_on_tail(q
, &wait
, &flags
);
3875 EXPORT_SYMBOL(interruptible_sleep_on
);
3878 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3880 unsigned long flags
;
3883 init_waitqueue_entry(&wait
, current
);
3885 current
->state
= TASK_INTERRUPTIBLE
;
3887 sleep_on_head(q
, &wait
, &flags
);
3888 timeout
= schedule_timeout(timeout
);
3889 sleep_on_tail(q
, &wait
, &flags
);
3893 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3895 void __sched
sleep_on(wait_queue_head_t
*q
)
3897 unsigned long flags
;
3900 init_waitqueue_entry(&wait
, current
);
3902 current
->state
= TASK_UNINTERRUPTIBLE
;
3904 sleep_on_head(q
, &wait
, &flags
);
3906 sleep_on_tail(q
, &wait
, &flags
);
3908 EXPORT_SYMBOL(sleep_on
);
3910 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3912 unsigned long flags
;
3915 init_waitqueue_entry(&wait
, current
);
3917 current
->state
= TASK_UNINTERRUPTIBLE
;
3919 sleep_on_head(q
, &wait
, &flags
);
3920 timeout
= schedule_timeout(timeout
);
3921 sleep_on_tail(q
, &wait
, &flags
);
3925 EXPORT_SYMBOL(sleep_on_timeout
);
3927 #ifdef CONFIG_RT_MUTEXES
3930 * rt_mutex_setprio - set the current priority of a task
3932 * @prio: prio value (kernel-internal form)
3934 * This function changes the 'effective' priority of a task. It does
3935 * not touch ->normal_prio like __setscheduler().
3937 * Used by the rt_mutex code to implement priority inheritance logic.
3939 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3941 unsigned long flags
;
3942 int oldprio
, on_rq
, running
;
3945 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3947 rq
= task_rq_lock(p
, &flags
);
3948 update_rq_clock(rq
);
3951 on_rq
= p
->se
.on_rq
;
3952 running
= task_running(rq
, p
);
3954 dequeue_task(rq
, p
, 0);
3956 p
->sched_class
->put_prev_task(rq
, p
);
3960 p
->sched_class
= &rt_sched_class
;
3962 p
->sched_class
= &fair_sched_class
;
3968 p
->sched_class
->set_curr_task(rq
);
3969 enqueue_task(rq
, p
, 0);
3971 * Reschedule if we are currently running on this runqueue and
3972 * our priority decreased, or if we are not currently running on
3973 * this runqueue and our priority is higher than the current's
3976 if (p
->prio
> oldprio
)
3977 resched_task(rq
->curr
);
3979 check_preempt_curr(rq
, p
);
3982 task_rq_unlock(rq
, &flags
);
3987 void set_user_nice(struct task_struct
*p
, long nice
)
3989 int old_prio
, delta
, on_rq
;
3990 unsigned long flags
;
3993 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3996 * We have to be careful, if called from sys_setpriority(),
3997 * the task might be in the middle of scheduling on another CPU.
3999 rq
= task_rq_lock(p
, &flags
);
4000 update_rq_clock(rq
);
4002 * The RT priorities are set via sched_setscheduler(), but we still
4003 * allow the 'normal' nice value to be set - but as expected
4004 * it wont have any effect on scheduling until the task is
4005 * SCHED_FIFO/SCHED_RR:
4007 if (task_has_rt_policy(p
)) {
4008 p
->static_prio
= NICE_TO_PRIO(nice
);
4011 on_rq
= p
->se
.on_rq
;
4013 dequeue_task(rq
, p
, 0);
4017 p
->static_prio
= NICE_TO_PRIO(nice
);
4020 p
->prio
= effective_prio(p
);
4021 delta
= p
->prio
- old_prio
;
4024 enqueue_task(rq
, p
, 0);
4027 * If the task increased its priority or is running and
4028 * lowered its priority, then reschedule its CPU:
4030 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4031 resched_task(rq
->curr
);
4034 task_rq_unlock(rq
, &flags
);
4036 EXPORT_SYMBOL(set_user_nice
);
4039 * can_nice - check if a task can reduce its nice value
4043 int can_nice(const struct task_struct
*p
, const int nice
)
4045 /* convert nice value [19,-20] to rlimit style value [1,40] */
4046 int nice_rlim
= 20 - nice
;
4048 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4049 capable(CAP_SYS_NICE
));
4052 #ifdef __ARCH_WANT_SYS_NICE
4055 * sys_nice - change the priority of the current process.
4056 * @increment: priority increment
4058 * sys_setpriority is a more generic, but much slower function that
4059 * does similar things.
4061 asmlinkage
long sys_nice(int increment
)
4066 * Setpriority might change our priority at the same moment.
4067 * We don't have to worry. Conceptually one call occurs first
4068 * and we have a single winner.
4070 if (increment
< -40)
4075 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4081 if (increment
< 0 && !can_nice(current
, nice
))
4084 retval
= security_task_setnice(current
, nice
);
4088 set_user_nice(current
, nice
);
4095 * task_prio - return the priority value of a given task.
4096 * @p: the task in question.
4098 * This is the priority value as seen by users in /proc.
4099 * RT tasks are offset by -200. Normal tasks are centered
4100 * around 0, value goes from -16 to +15.
4102 int task_prio(const struct task_struct
*p
)
4104 return p
->prio
- MAX_RT_PRIO
;
4108 * task_nice - return the nice value of a given task.
4109 * @p: the task in question.
4111 int task_nice(const struct task_struct
*p
)
4113 return TASK_NICE(p
);
4115 EXPORT_SYMBOL_GPL(task_nice
);
4118 * idle_cpu - is a given cpu idle currently?
4119 * @cpu: the processor in question.
4121 int idle_cpu(int cpu
)
4123 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4127 * idle_task - return the idle task for a given cpu.
4128 * @cpu: the processor in question.
4130 struct task_struct
*idle_task(int cpu
)
4132 return cpu_rq(cpu
)->idle
;
4136 * find_process_by_pid - find a process with a matching PID value.
4137 * @pid: the pid in question.
4139 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4141 return pid
? find_task_by_pid(pid
) : current
;
4144 /* Actually do priority change: must hold rq lock. */
4146 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4148 BUG_ON(p
->se
.on_rq
);
4151 switch (p
->policy
) {
4155 p
->sched_class
= &fair_sched_class
;
4159 p
->sched_class
= &rt_sched_class
;
4163 p
->rt_priority
= prio
;
4164 p
->normal_prio
= normal_prio(p
);
4165 /* we are holding p->pi_lock already */
4166 p
->prio
= rt_mutex_getprio(p
);
4171 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4172 * @p: the task in question.
4173 * @policy: new policy.
4174 * @param: structure containing the new RT priority.
4176 * NOTE that the task may be already dead.
4178 int sched_setscheduler(struct task_struct
*p
, int policy
,
4179 struct sched_param
*param
)
4181 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4182 unsigned long flags
;
4185 /* may grab non-irq protected spin_locks */
4186 BUG_ON(in_interrupt());
4188 /* double check policy once rq lock held */
4190 policy
= oldpolicy
= p
->policy
;
4191 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4192 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4193 policy
!= SCHED_IDLE
)
4196 * Valid priorities for SCHED_FIFO and SCHED_RR are
4197 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4198 * SCHED_BATCH and SCHED_IDLE is 0.
4200 if (param
->sched_priority
< 0 ||
4201 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4202 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4204 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4208 * Allow unprivileged RT tasks to decrease priority:
4210 if (!capable(CAP_SYS_NICE
)) {
4211 if (rt_policy(policy
)) {
4212 unsigned long rlim_rtprio
;
4214 if (!lock_task_sighand(p
, &flags
))
4216 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4217 unlock_task_sighand(p
, &flags
);
4219 /* can't set/change the rt policy */
4220 if (policy
!= p
->policy
&& !rlim_rtprio
)
4223 /* can't increase priority */
4224 if (param
->sched_priority
> p
->rt_priority
&&
4225 param
->sched_priority
> rlim_rtprio
)
4229 * Like positive nice levels, dont allow tasks to
4230 * move out of SCHED_IDLE either:
4232 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4235 /* can't change other user's priorities */
4236 if ((current
->euid
!= p
->euid
) &&
4237 (current
->euid
!= p
->uid
))
4241 retval
= security_task_setscheduler(p
, policy
, param
);
4245 * make sure no PI-waiters arrive (or leave) while we are
4246 * changing the priority of the task:
4248 spin_lock_irqsave(&p
->pi_lock
, flags
);
4250 * To be able to change p->policy safely, the apropriate
4251 * runqueue lock must be held.
4253 rq
= __task_rq_lock(p
);
4254 /* recheck policy now with rq lock held */
4255 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4256 policy
= oldpolicy
= -1;
4257 __task_rq_unlock(rq
);
4258 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4261 update_rq_clock(rq
);
4262 on_rq
= p
->se
.on_rq
;
4263 running
= task_running(rq
, p
);
4265 deactivate_task(rq
, p
, 0);
4267 p
->sched_class
->put_prev_task(rq
, p
);
4271 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4275 p
->sched_class
->set_curr_task(rq
);
4276 activate_task(rq
, p
, 0);
4278 * Reschedule if we are currently running on this runqueue and
4279 * our priority decreased, or if we are not currently running on
4280 * this runqueue and our priority is higher than the current's
4283 if (p
->prio
> oldprio
)
4284 resched_task(rq
->curr
);
4286 check_preempt_curr(rq
, p
);
4289 __task_rq_unlock(rq
);
4290 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4292 rt_mutex_adjust_pi(p
);
4296 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4299 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4301 struct sched_param lparam
;
4302 struct task_struct
*p
;
4305 if (!param
|| pid
< 0)
4307 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4312 p
= find_process_by_pid(pid
);
4314 retval
= sched_setscheduler(p
, policy
, &lparam
);
4321 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4322 * @pid: the pid in question.
4323 * @policy: new policy.
4324 * @param: structure containing the new RT priority.
4326 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4327 struct sched_param __user
*param
)
4329 /* negative values for policy are not valid */
4333 return do_sched_setscheduler(pid
, policy
, param
);
4337 * sys_sched_setparam - set/change the RT priority of a thread
4338 * @pid: the pid in question.
4339 * @param: structure containing the new RT priority.
4341 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4343 return do_sched_setscheduler(pid
, -1, param
);
4347 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4348 * @pid: the pid in question.
4350 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4352 struct task_struct
*p
;
4353 int retval
= -EINVAL
;
4359 read_lock(&tasklist_lock
);
4360 p
= find_process_by_pid(pid
);
4362 retval
= security_task_getscheduler(p
);
4366 read_unlock(&tasklist_lock
);
4373 * sys_sched_getscheduler - get the RT priority of a thread
4374 * @pid: the pid in question.
4375 * @param: structure containing the RT priority.
4377 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4379 struct sched_param lp
;
4380 struct task_struct
*p
;
4381 int retval
= -EINVAL
;
4383 if (!param
|| pid
< 0)
4386 read_lock(&tasklist_lock
);
4387 p
= find_process_by_pid(pid
);
4392 retval
= security_task_getscheduler(p
);
4396 lp
.sched_priority
= p
->rt_priority
;
4397 read_unlock(&tasklist_lock
);
4400 * This one might sleep, we cannot do it with a spinlock held ...
4402 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4408 read_unlock(&tasklist_lock
);
4412 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4414 cpumask_t cpus_allowed
;
4415 struct task_struct
*p
;
4418 mutex_lock(&sched_hotcpu_mutex
);
4419 read_lock(&tasklist_lock
);
4421 p
= find_process_by_pid(pid
);
4423 read_unlock(&tasklist_lock
);
4424 mutex_unlock(&sched_hotcpu_mutex
);
4429 * It is not safe to call set_cpus_allowed with the
4430 * tasklist_lock held. We will bump the task_struct's
4431 * usage count and then drop tasklist_lock.
4434 read_unlock(&tasklist_lock
);
4437 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4438 !capable(CAP_SYS_NICE
))
4441 retval
= security_task_setscheduler(p
, 0, NULL
);
4445 cpus_allowed
= cpuset_cpus_allowed(p
);
4446 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4447 retval
= set_cpus_allowed(p
, new_mask
);
4451 mutex_unlock(&sched_hotcpu_mutex
);
4455 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4456 cpumask_t
*new_mask
)
4458 if (len
< sizeof(cpumask_t
)) {
4459 memset(new_mask
, 0, sizeof(cpumask_t
));
4460 } else if (len
> sizeof(cpumask_t
)) {
4461 len
= sizeof(cpumask_t
);
4463 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4467 * sys_sched_setaffinity - set the cpu affinity of a process
4468 * @pid: pid of the process
4469 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4470 * @user_mask_ptr: user-space pointer to the new cpu mask
4472 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4473 unsigned long __user
*user_mask_ptr
)
4478 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4482 return sched_setaffinity(pid
, new_mask
);
4486 * Represents all cpu's present in the system
4487 * In systems capable of hotplug, this map could dynamically grow
4488 * as new cpu's are detected in the system via any platform specific
4489 * method, such as ACPI for e.g.
4492 cpumask_t cpu_present_map __read_mostly
;
4493 EXPORT_SYMBOL(cpu_present_map
);
4496 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4497 EXPORT_SYMBOL(cpu_online_map
);
4499 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4500 EXPORT_SYMBOL(cpu_possible_map
);
4503 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4505 struct task_struct
*p
;
4508 mutex_lock(&sched_hotcpu_mutex
);
4509 read_lock(&tasklist_lock
);
4512 p
= find_process_by_pid(pid
);
4516 retval
= security_task_getscheduler(p
);
4520 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4523 read_unlock(&tasklist_lock
);
4524 mutex_unlock(&sched_hotcpu_mutex
);
4530 * sys_sched_getaffinity - get the cpu affinity of a process
4531 * @pid: pid of the process
4532 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4533 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4535 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4536 unsigned long __user
*user_mask_ptr
)
4541 if (len
< sizeof(cpumask_t
))
4544 ret
= sched_getaffinity(pid
, &mask
);
4548 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4551 return sizeof(cpumask_t
);
4555 * sys_sched_yield - yield the current processor to other threads.
4557 * This function yields the current CPU to other tasks. If there are no
4558 * other threads running on this CPU then this function will return.
4560 asmlinkage
long sys_sched_yield(void)
4562 struct rq
*rq
= this_rq_lock();
4564 schedstat_inc(rq
, yld_cnt
);
4565 current
->sched_class
->yield_task(rq
);
4568 * Since we are going to call schedule() anyway, there's
4569 * no need to preempt or enable interrupts:
4571 __release(rq
->lock
);
4572 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4573 _raw_spin_unlock(&rq
->lock
);
4574 preempt_enable_no_resched();
4581 static void __cond_resched(void)
4583 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4584 __might_sleep(__FILE__
, __LINE__
);
4587 * The BKS might be reacquired before we have dropped
4588 * PREEMPT_ACTIVE, which could trigger a second
4589 * cond_resched() call.
4592 add_preempt_count(PREEMPT_ACTIVE
);
4594 sub_preempt_count(PREEMPT_ACTIVE
);
4595 } while (need_resched());
4598 int __sched
cond_resched(void)
4600 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4601 system_state
== SYSTEM_RUNNING
) {
4607 EXPORT_SYMBOL(cond_resched
);
4610 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4611 * call schedule, and on return reacquire the lock.
4613 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4614 * operations here to prevent schedule() from being called twice (once via
4615 * spin_unlock(), once by hand).
4617 int cond_resched_lock(spinlock_t
*lock
)
4621 if (need_lockbreak(lock
)) {
4627 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4628 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4629 _raw_spin_unlock(lock
);
4630 preempt_enable_no_resched();
4637 EXPORT_SYMBOL(cond_resched_lock
);
4639 int __sched
cond_resched_softirq(void)
4641 BUG_ON(!in_softirq());
4643 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4651 EXPORT_SYMBOL(cond_resched_softirq
);
4654 * yield - yield the current processor to other threads.
4656 * This is a shortcut for kernel-space yielding - it marks the
4657 * thread runnable and calls sys_sched_yield().
4659 void __sched
yield(void)
4661 set_current_state(TASK_RUNNING
);
4664 EXPORT_SYMBOL(yield
);
4667 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4668 * that process accounting knows that this is a task in IO wait state.
4670 * But don't do that if it is a deliberate, throttling IO wait (this task
4671 * has set its backing_dev_info: the queue against which it should throttle)
4673 void __sched
io_schedule(void)
4675 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4677 delayacct_blkio_start();
4678 atomic_inc(&rq
->nr_iowait
);
4680 atomic_dec(&rq
->nr_iowait
);
4681 delayacct_blkio_end();
4683 EXPORT_SYMBOL(io_schedule
);
4685 long __sched
io_schedule_timeout(long timeout
)
4687 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4690 delayacct_blkio_start();
4691 atomic_inc(&rq
->nr_iowait
);
4692 ret
= schedule_timeout(timeout
);
4693 atomic_dec(&rq
->nr_iowait
);
4694 delayacct_blkio_end();
4699 * sys_sched_get_priority_max - return maximum RT priority.
4700 * @policy: scheduling class.
4702 * this syscall returns the maximum rt_priority that can be used
4703 * by a given scheduling class.
4705 asmlinkage
long sys_sched_get_priority_max(int policy
)
4712 ret
= MAX_USER_RT_PRIO
-1;
4724 * sys_sched_get_priority_min - return minimum RT priority.
4725 * @policy: scheduling class.
4727 * this syscall returns the minimum rt_priority that can be used
4728 * by a given scheduling class.
4730 asmlinkage
long sys_sched_get_priority_min(int policy
)
4748 * sys_sched_rr_get_interval - return the default timeslice of a process.
4749 * @pid: pid of the process.
4750 * @interval: userspace pointer to the timeslice value.
4752 * this syscall writes the default timeslice value of a given process
4753 * into the user-space timespec buffer. A value of '0' means infinity.
4756 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4758 struct task_struct
*p
;
4759 int retval
= -EINVAL
;
4766 read_lock(&tasklist_lock
);
4767 p
= find_process_by_pid(pid
);
4771 retval
= security_task_getscheduler(p
);
4775 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4776 0 : static_prio_timeslice(p
->static_prio
), &t
);
4777 read_unlock(&tasklist_lock
);
4778 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4782 read_unlock(&tasklist_lock
);
4786 static const char stat_nam
[] = "RSDTtZX";
4788 static void show_task(struct task_struct
*p
)
4790 unsigned long free
= 0;
4793 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4794 printk("%-13.13s %c", p
->comm
,
4795 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4796 #if BITS_PER_LONG == 32
4797 if (state
== TASK_RUNNING
)
4798 printk(" running ");
4800 printk(" %08lx ", thread_saved_pc(p
));
4802 if (state
== TASK_RUNNING
)
4803 printk(" running task ");
4805 printk(" %016lx ", thread_saved_pc(p
));
4807 #ifdef CONFIG_DEBUG_STACK_USAGE
4809 unsigned long *n
= end_of_stack(p
);
4812 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4815 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4817 if (state
!= TASK_RUNNING
)
4818 show_stack(p
, NULL
);
4821 void show_state_filter(unsigned long state_filter
)
4823 struct task_struct
*g
, *p
;
4825 #if BITS_PER_LONG == 32
4827 " task PC stack pid father\n");
4830 " task PC stack pid father\n");
4832 read_lock(&tasklist_lock
);
4833 do_each_thread(g
, p
) {
4835 * reset the NMI-timeout, listing all files on a slow
4836 * console might take alot of time:
4838 touch_nmi_watchdog();
4839 if (!state_filter
|| (p
->state
& state_filter
))
4841 } while_each_thread(g
, p
);
4843 touch_all_softlockup_watchdogs();
4845 #ifdef CONFIG_SCHED_DEBUG
4846 sysrq_sched_debug_show();
4848 read_unlock(&tasklist_lock
);
4850 * Only show locks if all tasks are dumped:
4852 if (state_filter
== -1)
4853 debug_show_all_locks();
4856 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4858 idle
->sched_class
= &idle_sched_class
;
4862 * init_idle - set up an idle thread for a given CPU
4863 * @idle: task in question
4864 * @cpu: cpu the idle task belongs to
4866 * NOTE: this function does not set the idle thread's NEED_RESCHED
4867 * flag, to make booting more robust.
4869 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4871 struct rq
*rq
= cpu_rq(cpu
);
4872 unsigned long flags
;
4875 idle
->se
.exec_start
= sched_clock();
4877 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4878 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4879 __set_task_cpu(idle
, cpu
);
4881 spin_lock_irqsave(&rq
->lock
, flags
);
4882 rq
->curr
= rq
->idle
= idle
;
4883 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4886 spin_unlock_irqrestore(&rq
->lock
, flags
);
4888 /* Set the preempt count _outside_ the spinlocks! */
4889 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4890 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4892 task_thread_info(idle
)->preempt_count
= 0;
4895 * The idle tasks have their own, simple scheduling class:
4897 idle
->sched_class
= &idle_sched_class
;
4901 * In a system that switches off the HZ timer nohz_cpu_mask
4902 * indicates which cpus entered this state. This is used
4903 * in the rcu update to wait only for active cpus. For system
4904 * which do not switch off the HZ timer nohz_cpu_mask should
4905 * always be CPU_MASK_NONE.
4907 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4911 * This is how migration works:
4913 * 1) we queue a struct migration_req structure in the source CPU's
4914 * runqueue and wake up that CPU's migration thread.
4915 * 2) we down() the locked semaphore => thread blocks.
4916 * 3) migration thread wakes up (implicitly it forces the migrated
4917 * thread off the CPU)
4918 * 4) it gets the migration request and checks whether the migrated
4919 * task is still in the wrong runqueue.
4920 * 5) if it's in the wrong runqueue then the migration thread removes
4921 * it and puts it into the right queue.
4922 * 6) migration thread up()s the semaphore.
4923 * 7) we wake up and the migration is done.
4927 * Change a given task's CPU affinity. Migrate the thread to a
4928 * proper CPU and schedule it away if the CPU it's executing on
4929 * is removed from the allowed bitmask.
4931 * NOTE: the caller must have a valid reference to the task, the
4932 * task must not exit() & deallocate itself prematurely. The
4933 * call is not atomic; no spinlocks may be held.
4935 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4937 struct migration_req req
;
4938 unsigned long flags
;
4942 rq
= task_rq_lock(p
, &flags
);
4943 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4948 p
->cpus_allowed
= new_mask
;
4949 /* Can the task run on the task's current CPU? If so, we're done */
4950 if (cpu_isset(task_cpu(p
), new_mask
))
4953 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4954 /* Need help from migration thread: drop lock and wait. */
4955 task_rq_unlock(rq
, &flags
);
4956 wake_up_process(rq
->migration_thread
);
4957 wait_for_completion(&req
.done
);
4958 tlb_migrate_finish(p
->mm
);
4962 task_rq_unlock(rq
, &flags
);
4966 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4969 * Move (not current) task off this cpu, onto dest cpu. We're doing
4970 * this because either it can't run here any more (set_cpus_allowed()
4971 * away from this CPU, or CPU going down), or because we're
4972 * attempting to rebalance this task on exec (sched_exec).
4974 * So we race with normal scheduler movements, but that's OK, as long
4975 * as the task is no longer on this CPU.
4977 * Returns non-zero if task was successfully migrated.
4979 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4981 struct rq
*rq_dest
, *rq_src
;
4984 if (unlikely(cpu_is_offline(dest_cpu
)))
4987 rq_src
= cpu_rq(src_cpu
);
4988 rq_dest
= cpu_rq(dest_cpu
);
4990 double_rq_lock(rq_src
, rq_dest
);
4991 /* Already moved. */
4992 if (task_cpu(p
) != src_cpu
)
4994 /* Affinity changed (again). */
4995 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4998 on_rq
= p
->se
.on_rq
;
5000 deactivate_task(rq_src
, p
, 0);
5002 set_task_cpu(p
, dest_cpu
);
5004 activate_task(rq_dest
, p
, 0);
5005 check_preempt_curr(rq_dest
, p
);
5009 double_rq_unlock(rq_src
, rq_dest
);
5014 * migration_thread - this is a highprio system thread that performs
5015 * thread migration by bumping thread off CPU then 'pushing' onto
5018 static int migration_thread(void *data
)
5020 int cpu
= (long)data
;
5024 BUG_ON(rq
->migration_thread
!= current
);
5026 set_current_state(TASK_INTERRUPTIBLE
);
5027 while (!kthread_should_stop()) {
5028 struct migration_req
*req
;
5029 struct list_head
*head
;
5031 spin_lock_irq(&rq
->lock
);
5033 if (cpu_is_offline(cpu
)) {
5034 spin_unlock_irq(&rq
->lock
);
5038 if (rq
->active_balance
) {
5039 active_load_balance(rq
, cpu
);
5040 rq
->active_balance
= 0;
5043 head
= &rq
->migration_queue
;
5045 if (list_empty(head
)) {
5046 spin_unlock_irq(&rq
->lock
);
5048 set_current_state(TASK_INTERRUPTIBLE
);
5051 req
= list_entry(head
->next
, struct migration_req
, list
);
5052 list_del_init(head
->next
);
5054 spin_unlock(&rq
->lock
);
5055 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5058 complete(&req
->done
);
5060 __set_current_state(TASK_RUNNING
);
5064 /* Wait for kthread_stop */
5065 set_current_state(TASK_INTERRUPTIBLE
);
5066 while (!kthread_should_stop()) {
5068 set_current_state(TASK_INTERRUPTIBLE
);
5070 __set_current_state(TASK_RUNNING
);
5074 #ifdef CONFIG_HOTPLUG_CPU
5076 * Figure out where task on dead CPU should go, use force if neccessary.
5077 * NOTE: interrupts should be disabled by the caller
5079 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5081 unsigned long flags
;
5088 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5089 cpus_and(mask
, mask
, p
->cpus_allowed
);
5090 dest_cpu
= any_online_cpu(mask
);
5092 /* On any allowed CPU? */
5093 if (dest_cpu
== NR_CPUS
)
5094 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5096 /* No more Mr. Nice Guy. */
5097 if (dest_cpu
== NR_CPUS
) {
5098 rq
= task_rq_lock(p
, &flags
);
5099 cpus_setall(p
->cpus_allowed
);
5100 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5101 task_rq_unlock(rq
, &flags
);
5104 * Don't tell them about moving exiting tasks or
5105 * kernel threads (both mm NULL), since they never
5108 if (p
->mm
&& printk_ratelimit())
5109 printk(KERN_INFO
"process %d (%s) no "
5110 "longer affine to cpu%d\n",
5111 p
->pid
, p
->comm
, dead_cpu
);
5113 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5118 * While a dead CPU has no uninterruptible tasks queued at this point,
5119 * it might still have a nonzero ->nr_uninterruptible counter, because
5120 * for performance reasons the counter is not stricly tracking tasks to
5121 * their home CPUs. So we just add the counter to another CPU's counter,
5122 * to keep the global sum constant after CPU-down:
5124 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5126 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5127 unsigned long flags
;
5129 local_irq_save(flags
);
5130 double_rq_lock(rq_src
, rq_dest
);
5131 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5132 rq_src
->nr_uninterruptible
= 0;
5133 double_rq_unlock(rq_src
, rq_dest
);
5134 local_irq_restore(flags
);
5137 /* Run through task list and migrate tasks from the dead cpu. */
5138 static void migrate_live_tasks(int src_cpu
)
5140 struct task_struct
*p
, *t
;
5142 write_lock_irq(&tasklist_lock
);
5144 do_each_thread(t
, p
) {
5148 if (task_cpu(p
) == src_cpu
)
5149 move_task_off_dead_cpu(src_cpu
, p
);
5150 } while_each_thread(t
, p
);
5152 write_unlock_irq(&tasklist_lock
);
5156 * Schedules idle task to be the next runnable task on current CPU.
5157 * It does so by boosting its priority to highest possible and adding it to
5158 * the _front_ of the runqueue. Used by CPU offline code.
5160 void sched_idle_next(void)
5162 int this_cpu
= smp_processor_id();
5163 struct rq
*rq
= cpu_rq(this_cpu
);
5164 struct task_struct
*p
= rq
->idle
;
5165 unsigned long flags
;
5167 /* cpu has to be offline */
5168 BUG_ON(cpu_online(this_cpu
));
5171 * Strictly not necessary since rest of the CPUs are stopped by now
5172 * and interrupts disabled on the current cpu.
5174 spin_lock_irqsave(&rq
->lock
, flags
);
5176 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5178 /* Add idle task to the _front_ of its priority queue: */
5179 activate_idle_task(p
, rq
);
5181 spin_unlock_irqrestore(&rq
->lock
, flags
);
5185 * Ensures that the idle task is using init_mm right before its cpu goes
5188 void idle_task_exit(void)
5190 struct mm_struct
*mm
= current
->active_mm
;
5192 BUG_ON(cpu_online(smp_processor_id()));
5195 switch_mm(mm
, &init_mm
, current
);
5199 /* called under rq->lock with disabled interrupts */
5200 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5202 struct rq
*rq
= cpu_rq(dead_cpu
);
5204 /* Must be exiting, otherwise would be on tasklist. */
5205 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5207 /* Cannot have done final schedule yet: would have vanished. */
5208 BUG_ON(p
->state
== TASK_DEAD
);
5213 * Drop lock around migration; if someone else moves it,
5214 * that's OK. No task can be added to this CPU, so iteration is
5216 * NOTE: interrupts should be left disabled --dev@
5218 spin_unlock(&rq
->lock
);
5219 move_task_off_dead_cpu(dead_cpu
, p
);
5220 spin_lock(&rq
->lock
);
5225 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5226 static void migrate_dead_tasks(unsigned int dead_cpu
)
5228 struct rq
*rq
= cpu_rq(dead_cpu
);
5229 struct task_struct
*next
;
5232 if (!rq
->nr_running
)
5234 update_rq_clock(rq
);
5235 next
= pick_next_task(rq
, rq
->curr
);
5238 migrate_dead(dead_cpu
, next
);
5242 #endif /* CONFIG_HOTPLUG_CPU */
5244 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5246 static struct ctl_table sd_ctl_dir
[] = {
5248 .procname
= "sched_domain",
5254 static struct ctl_table sd_ctl_root
[] = {
5256 .ctl_name
= CTL_KERN
,
5257 .procname
= "kernel",
5259 .child
= sd_ctl_dir
,
5264 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5266 struct ctl_table
*entry
=
5267 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5270 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5276 set_table_entry(struct ctl_table
*entry
,
5277 const char *procname
, void *data
, int maxlen
,
5278 mode_t mode
, proc_handler
*proc_handler
)
5280 entry
->procname
= procname
;
5282 entry
->maxlen
= maxlen
;
5284 entry
->proc_handler
= proc_handler
;
5287 static struct ctl_table
*
5288 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5290 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5292 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5293 sizeof(long), 0644, proc_doulongvec_minmax
);
5294 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5295 sizeof(long), 0644, proc_doulongvec_minmax
);
5296 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5297 sizeof(int), 0644, proc_dointvec_minmax
);
5298 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5299 sizeof(int), 0644, proc_dointvec_minmax
);
5300 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5301 sizeof(int), 0644, proc_dointvec_minmax
);
5302 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5303 sizeof(int), 0644, proc_dointvec_minmax
);
5304 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5305 sizeof(int), 0644, proc_dointvec_minmax
);
5306 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5307 sizeof(int), 0644, proc_dointvec_minmax
);
5308 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5309 sizeof(int), 0644, proc_dointvec_minmax
);
5310 set_table_entry(&table
[10], "cache_nice_tries",
5311 &sd
->cache_nice_tries
,
5312 sizeof(int), 0644, proc_dointvec_minmax
);
5313 set_table_entry(&table
[12], "flags", &sd
->flags
,
5314 sizeof(int), 0644, proc_dointvec_minmax
);
5319 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5321 struct ctl_table
*entry
, *table
;
5322 struct sched_domain
*sd
;
5323 int domain_num
= 0, i
;
5326 for_each_domain(cpu
, sd
)
5328 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5331 for_each_domain(cpu
, sd
) {
5332 snprintf(buf
, 32, "domain%d", i
);
5333 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5335 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5342 static struct ctl_table_header
*sd_sysctl_header
;
5343 static void init_sched_domain_sysctl(void)
5345 int i
, cpu_num
= num_online_cpus();
5346 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5349 sd_ctl_dir
[0].child
= entry
;
5351 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5352 snprintf(buf
, 32, "cpu%d", i
);
5353 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5355 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5357 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5360 static void init_sched_domain_sysctl(void)
5366 * migration_call - callback that gets triggered when a CPU is added.
5367 * Here we can start up the necessary migration thread for the new CPU.
5369 static int __cpuinit
5370 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5372 struct task_struct
*p
;
5373 int cpu
= (long)hcpu
;
5374 unsigned long flags
;
5378 case CPU_LOCK_ACQUIRE
:
5379 mutex_lock(&sched_hotcpu_mutex
);
5382 case CPU_UP_PREPARE
:
5383 case CPU_UP_PREPARE_FROZEN
:
5384 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5387 kthread_bind(p
, cpu
);
5388 /* Must be high prio: stop_machine expects to yield to it. */
5389 rq
= task_rq_lock(p
, &flags
);
5390 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5391 task_rq_unlock(rq
, &flags
);
5392 cpu_rq(cpu
)->migration_thread
= p
;
5396 case CPU_ONLINE_FROZEN
:
5397 /* Strictly unneccessary, as first user will wake it. */
5398 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5401 #ifdef CONFIG_HOTPLUG_CPU
5402 case CPU_UP_CANCELED
:
5403 case CPU_UP_CANCELED_FROZEN
:
5404 if (!cpu_rq(cpu
)->migration_thread
)
5406 /* Unbind it from offline cpu so it can run. Fall thru. */
5407 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5408 any_online_cpu(cpu_online_map
));
5409 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5410 cpu_rq(cpu
)->migration_thread
= NULL
;
5414 case CPU_DEAD_FROZEN
:
5415 migrate_live_tasks(cpu
);
5417 kthread_stop(rq
->migration_thread
);
5418 rq
->migration_thread
= NULL
;
5419 /* Idle task back to normal (off runqueue, low prio) */
5420 rq
= task_rq_lock(rq
->idle
, &flags
);
5421 update_rq_clock(rq
);
5422 deactivate_task(rq
, rq
->idle
, 0);
5423 rq
->idle
->static_prio
= MAX_PRIO
;
5424 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5425 rq
->idle
->sched_class
= &idle_sched_class
;
5426 migrate_dead_tasks(cpu
);
5427 task_rq_unlock(rq
, &flags
);
5428 migrate_nr_uninterruptible(rq
);
5429 BUG_ON(rq
->nr_running
!= 0);
5431 /* No need to migrate the tasks: it was best-effort if
5432 * they didn't take sched_hotcpu_mutex. Just wake up
5433 * the requestors. */
5434 spin_lock_irq(&rq
->lock
);
5435 while (!list_empty(&rq
->migration_queue
)) {
5436 struct migration_req
*req
;
5438 req
= list_entry(rq
->migration_queue
.next
,
5439 struct migration_req
, list
);
5440 list_del_init(&req
->list
);
5441 complete(&req
->done
);
5443 spin_unlock_irq(&rq
->lock
);
5446 case CPU_LOCK_RELEASE
:
5447 mutex_unlock(&sched_hotcpu_mutex
);
5453 /* Register at highest priority so that task migration (migrate_all_tasks)
5454 * happens before everything else.
5456 static struct notifier_block __cpuinitdata migration_notifier
= {
5457 .notifier_call
= migration_call
,
5461 int __init
migration_init(void)
5463 void *cpu
= (void *)(long)smp_processor_id();
5466 /* Start one for the boot CPU: */
5467 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5468 BUG_ON(err
== NOTIFY_BAD
);
5469 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5470 register_cpu_notifier(&migration_notifier
);
5478 /* Number of possible processor ids */
5479 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5480 EXPORT_SYMBOL(nr_cpu_ids
);
5482 #undef SCHED_DOMAIN_DEBUG
5483 #ifdef SCHED_DOMAIN_DEBUG
5484 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5489 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5493 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5498 struct sched_group
*group
= sd
->groups
;
5499 cpumask_t groupmask
;
5501 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5502 cpus_clear(groupmask
);
5505 for (i
= 0; i
< level
+ 1; i
++)
5507 printk("domain %d: ", level
);
5509 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5510 printk("does not load-balance\n");
5512 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5517 printk("span %s\n", str
);
5519 if (!cpu_isset(cpu
, sd
->span
))
5520 printk(KERN_ERR
"ERROR: domain->span does not contain "
5522 if (!cpu_isset(cpu
, group
->cpumask
))
5523 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5527 for (i
= 0; i
< level
+ 2; i
++)
5533 printk(KERN_ERR
"ERROR: group is NULL\n");
5537 if (!group
->__cpu_power
) {
5539 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5543 if (!cpus_weight(group
->cpumask
)) {
5545 printk(KERN_ERR
"ERROR: empty group\n");
5548 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5550 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5553 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5555 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5558 group
= group
->next
;
5559 } while (group
!= sd
->groups
);
5562 if (!cpus_equal(sd
->span
, groupmask
))
5563 printk(KERN_ERR
"ERROR: groups don't span "
5571 if (!cpus_subset(groupmask
, sd
->span
))
5572 printk(KERN_ERR
"ERROR: parent span is not a superset "
5573 "of domain->span\n");
5578 # define sched_domain_debug(sd, cpu) do { } while (0)
5581 static int sd_degenerate(struct sched_domain
*sd
)
5583 if (cpus_weight(sd
->span
) == 1)
5586 /* Following flags need at least 2 groups */
5587 if (sd
->flags
& (SD_LOAD_BALANCE
|
5588 SD_BALANCE_NEWIDLE
|
5592 SD_SHARE_PKG_RESOURCES
)) {
5593 if (sd
->groups
!= sd
->groups
->next
)
5597 /* Following flags don't use groups */
5598 if (sd
->flags
& (SD_WAKE_IDLE
|
5607 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5609 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5611 if (sd_degenerate(parent
))
5614 if (!cpus_equal(sd
->span
, parent
->span
))
5617 /* Does parent contain flags not in child? */
5618 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5619 if (cflags
& SD_WAKE_AFFINE
)
5620 pflags
&= ~SD_WAKE_BALANCE
;
5621 /* Flags needing groups don't count if only 1 group in parent */
5622 if (parent
->groups
== parent
->groups
->next
) {
5623 pflags
&= ~(SD_LOAD_BALANCE
|
5624 SD_BALANCE_NEWIDLE
|
5628 SD_SHARE_PKG_RESOURCES
);
5630 if (~cflags
& pflags
)
5637 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5638 * hold the hotplug lock.
5640 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5642 struct rq
*rq
= cpu_rq(cpu
);
5643 struct sched_domain
*tmp
;
5645 /* Remove the sched domains which do not contribute to scheduling. */
5646 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5647 struct sched_domain
*parent
= tmp
->parent
;
5650 if (sd_parent_degenerate(tmp
, parent
)) {
5651 tmp
->parent
= parent
->parent
;
5653 parent
->parent
->child
= tmp
;
5657 if (sd
&& sd_degenerate(sd
)) {
5663 sched_domain_debug(sd
, cpu
);
5665 rcu_assign_pointer(rq
->sd
, sd
);
5668 /* cpus with isolated domains */
5669 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5671 /* Setup the mask of cpus configured for isolated domains */
5672 static int __init
isolated_cpu_setup(char *str
)
5674 int ints
[NR_CPUS
], i
;
5676 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5677 cpus_clear(cpu_isolated_map
);
5678 for (i
= 1; i
<= ints
[0]; i
++)
5679 if (ints
[i
] < NR_CPUS
)
5680 cpu_set(ints
[i
], cpu_isolated_map
);
5684 __setup ("isolcpus=", isolated_cpu_setup
);
5687 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5688 * to a function which identifies what group(along with sched group) a CPU
5689 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5690 * (due to the fact that we keep track of groups covered with a cpumask_t).
5692 * init_sched_build_groups will build a circular linked list of the groups
5693 * covered by the given span, and will set each group's ->cpumask correctly,
5694 * and ->cpu_power to 0.
5697 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5698 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5699 struct sched_group
**sg
))
5701 struct sched_group
*first
= NULL
, *last
= NULL
;
5702 cpumask_t covered
= CPU_MASK_NONE
;
5705 for_each_cpu_mask(i
, span
) {
5706 struct sched_group
*sg
;
5707 int group
= group_fn(i
, cpu_map
, &sg
);
5710 if (cpu_isset(i
, covered
))
5713 sg
->cpumask
= CPU_MASK_NONE
;
5714 sg
->__cpu_power
= 0;
5716 for_each_cpu_mask(j
, span
) {
5717 if (group_fn(j
, cpu_map
, NULL
) != group
)
5720 cpu_set(j
, covered
);
5721 cpu_set(j
, sg
->cpumask
);
5732 #define SD_NODES_PER_DOMAIN 16
5737 * find_next_best_node - find the next node to include in a sched_domain
5738 * @node: node whose sched_domain we're building
5739 * @used_nodes: nodes already in the sched_domain
5741 * Find the next node to include in a given scheduling domain. Simply
5742 * finds the closest node not already in the @used_nodes map.
5744 * Should use nodemask_t.
5746 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5748 int i
, n
, val
, min_val
, best_node
= 0;
5752 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5753 /* Start at @node */
5754 n
= (node
+ i
) % MAX_NUMNODES
;
5756 if (!nr_cpus_node(n
))
5759 /* Skip already used nodes */
5760 if (test_bit(n
, used_nodes
))
5763 /* Simple min distance search */
5764 val
= node_distance(node
, n
);
5766 if (val
< min_val
) {
5772 set_bit(best_node
, used_nodes
);
5777 * sched_domain_node_span - get a cpumask for a node's sched_domain
5778 * @node: node whose cpumask we're constructing
5779 * @size: number of nodes to include in this span
5781 * Given a node, construct a good cpumask for its sched_domain to span. It
5782 * should be one that prevents unnecessary balancing, but also spreads tasks
5785 static cpumask_t
sched_domain_node_span(int node
)
5787 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5788 cpumask_t span
, nodemask
;
5792 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5794 nodemask
= node_to_cpumask(node
);
5795 cpus_or(span
, span
, nodemask
);
5796 set_bit(node
, used_nodes
);
5798 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5799 int next_node
= find_next_best_node(node
, used_nodes
);
5801 nodemask
= node_to_cpumask(next_node
);
5802 cpus_or(span
, span
, nodemask
);
5809 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5812 * SMT sched-domains:
5814 #ifdef CONFIG_SCHED_SMT
5815 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5816 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5818 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5819 struct sched_group
**sg
)
5822 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5828 * multi-core sched-domains:
5830 #ifdef CONFIG_SCHED_MC
5831 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5832 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5835 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5836 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5837 struct sched_group
**sg
)
5840 cpumask_t mask
= cpu_sibling_map
[cpu
];
5841 cpus_and(mask
, mask
, *cpu_map
);
5842 group
= first_cpu(mask
);
5844 *sg
= &per_cpu(sched_group_core
, group
);
5847 #elif defined(CONFIG_SCHED_MC)
5848 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5849 struct sched_group
**sg
)
5852 *sg
= &per_cpu(sched_group_core
, cpu
);
5857 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5858 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5860 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5861 struct sched_group
**sg
)
5864 #ifdef CONFIG_SCHED_MC
5865 cpumask_t mask
= cpu_coregroup_map(cpu
);
5866 cpus_and(mask
, mask
, *cpu_map
);
5867 group
= first_cpu(mask
);
5868 #elif defined(CONFIG_SCHED_SMT)
5869 cpumask_t mask
= cpu_sibling_map
[cpu
];
5870 cpus_and(mask
, mask
, *cpu_map
);
5871 group
= first_cpu(mask
);
5876 *sg
= &per_cpu(sched_group_phys
, group
);
5882 * The init_sched_build_groups can't handle what we want to do with node
5883 * groups, so roll our own. Now each node has its own list of groups which
5884 * gets dynamically allocated.
5886 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5887 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5889 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5890 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5892 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5893 struct sched_group
**sg
)
5895 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5898 cpus_and(nodemask
, nodemask
, *cpu_map
);
5899 group
= first_cpu(nodemask
);
5902 *sg
= &per_cpu(sched_group_allnodes
, group
);
5906 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5908 struct sched_group
*sg
= group_head
;
5914 for_each_cpu_mask(j
, sg
->cpumask
) {
5915 struct sched_domain
*sd
;
5917 sd
= &per_cpu(phys_domains
, j
);
5918 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5920 * Only add "power" once for each
5926 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5929 if (sg
!= group_head
)
5935 /* Free memory allocated for various sched_group structures */
5936 static void free_sched_groups(const cpumask_t
*cpu_map
)
5940 for_each_cpu_mask(cpu
, *cpu_map
) {
5941 struct sched_group
**sched_group_nodes
5942 = sched_group_nodes_bycpu
[cpu
];
5944 if (!sched_group_nodes
)
5947 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5948 cpumask_t nodemask
= node_to_cpumask(i
);
5949 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5951 cpus_and(nodemask
, nodemask
, *cpu_map
);
5952 if (cpus_empty(nodemask
))
5962 if (oldsg
!= sched_group_nodes
[i
])
5965 kfree(sched_group_nodes
);
5966 sched_group_nodes_bycpu
[cpu
] = NULL
;
5970 static void free_sched_groups(const cpumask_t
*cpu_map
)
5976 * Initialize sched groups cpu_power.
5978 * cpu_power indicates the capacity of sched group, which is used while
5979 * distributing the load between different sched groups in a sched domain.
5980 * Typically cpu_power for all the groups in a sched domain will be same unless
5981 * there are asymmetries in the topology. If there are asymmetries, group
5982 * having more cpu_power will pickup more load compared to the group having
5985 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5986 * the maximum number of tasks a group can handle in the presence of other idle
5987 * or lightly loaded groups in the same sched domain.
5989 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5991 struct sched_domain
*child
;
5992 struct sched_group
*group
;
5994 WARN_ON(!sd
|| !sd
->groups
);
5996 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6001 sd
->groups
->__cpu_power
= 0;
6004 * For perf policy, if the groups in child domain share resources
6005 * (for example cores sharing some portions of the cache hierarchy
6006 * or SMT), then set this domain groups cpu_power such that each group
6007 * can handle only one task, when there are other idle groups in the
6008 * same sched domain.
6010 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6012 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6013 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6018 * add cpu_power of each child group to this groups cpu_power
6020 group
= child
->groups
;
6022 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6023 group
= group
->next
;
6024 } while (group
!= child
->groups
);
6028 * Build sched domains for a given set of cpus and attach the sched domains
6029 * to the individual cpus
6031 static int build_sched_domains(const cpumask_t
*cpu_map
)
6035 struct sched_group
**sched_group_nodes
= NULL
;
6036 int sd_allnodes
= 0;
6039 * Allocate the per-node list of sched groups
6041 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6043 if (!sched_group_nodes
) {
6044 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6047 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6051 * Set up domains for cpus specified by the cpu_map.
6053 for_each_cpu_mask(i
, *cpu_map
) {
6054 struct sched_domain
*sd
= NULL
, *p
;
6055 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6057 cpus_and(nodemask
, nodemask
, *cpu_map
);
6060 if (cpus_weight(*cpu_map
) >
6061 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6062 sd
= &per_cpu(allnodes_domains
, i
);
6063 *sd
= SD_ALLNODES_INIT
;
6064 sd
->span
= *cpu_map
;
6065 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6071 sd
= &per_cpu(node_domains
, i
);
6073 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6077 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6081 sd
= &per_cpu(phys_domains
, i
);
6083 sd
->span
= nodemask
;
6087 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6089 #ifdef CONFIG_SCHED_MC
6091 sd
= &per_cpu(core_domains
, i
);
6093 sd
->span
= cpu_coregroup_map(i
);
6094 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6097 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6100 #ifdef CONFIG_SCHED_SMT
6102 sd
= &per_cpu(cpu_domains
, i
);
6103 *sd
= SD_SIBLING_INIT
;
6104 sd
->span
= cpu_sibling_map
[i
];
6105 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6108 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6112 #ifdef CONFIG_SCHED_SMT
6113 /* Set up CPU (sibling) groups */
6114 for_each_cpu_mask(i
, *cpu_map
) {
6115 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6116 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6117 if (i
!= first_cpu(this_sibling_map
))
6120 init_sched_build_groups(this_sibling_map
, cpu_map
,
6125 #ifdef CONFIG_SCHED_MC
6126 /* Set up multi-core groups */
6127 for_each_cpu_mask(i
, *cpu_map
) {
6128 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6129 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6130 if (i
!= first_cpu(this_core_map
))
6132 init_sched_build_groups(this_core_map
, cpu_map
,
6133 &cpu_to_core_group
);
6137 /* Set up physical groups */
6138 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6139 cpumask_t nodemask
= node_to_cpumask(i
);
6141 cpus_and(nodemask
, nodemask
, *cpu_map
);
6142 if (cpus_empty(nodemask
))
6145 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6149 /* Set up node groups */
6151 init_sched_build_groups(*cpu_map
, cpu_map
,
6152 &cpu_to_allnodes_group
);
6154 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6155 /* Set up node groups */
6156 struct sched_group
*sg
, *prev
;
6157 cpumask_t nodemask
= node_to_cpumask(i
);
6158 cpumask_t domainspan
;
6159 cpumask_t covered
= CPU_MASK_NONE
;
6162 cpus_and(nodemask
, nodemask
, *cpu_map
);
6163 if (cpus_empty(nodemask
)) {
6164 sched_group_nodes
[i
] = NULL
;
6168 domainspan
= sched_domain_node_span(i
);
6169 cpus_and(domainspan
, domainspan
, *cpu_map
);
6171 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6173 printk(KERN_WARNING
"Can not alloc domain group for "
6177 sched_group_nodes
[i
] = sg
;
6178 for_each_cpu_mask(j
, nodemask
) {
6179 struct sched_domain
*sd
;
6181 sd
= &per_cpu(node_domains
, j
);
6184 sg
->__cpu_power
= 0;
6185 sg
->cpumask
= nodemask
;
6187 cpus_or(covered
, covered
, nodemask
);
6190 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6191 cpumask_t tmp
, notcovered
;
6192 int n
= (i
+ j
) % MAX_NUMNODES
;
6194 cpus_complement(notcovered
, covered
);
6195 cpus_and(tmp
, notcovered
, *cpu_map
);
6196 cpus_and(tmp
, tmp
, domainspan
);
6197 if (cpus_empty(tmp
))
6200 nodemask
= node_to_cpumask(n
);
6201 cpus_and(tmp
, tmp
, nodemask
);
6202 if (cpus_empty(tmp
))
6205 sg
= kmalloc_node(sizeof(struct sched_group
),
6209 "Can not alloc domain group for node %d\n", j
);
6212 sg
->__cpu_power
= 0;
6214 sg
->next
= prev
->next
;
6215 cpus_or(covered
, covered
, tmp
);
6222 /* Calculate CPU power for physical packages and nodes */
6223 #ifdef CONFIG_SCHED_SMT
6224 for_each_cpu_mask(i
, *cpu_map
) {
6225 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6227 init_sched_groups_power(i
, sd
);
6230 #ifdef CONFIG_SCHED_MC
6231 for_each_cpu_mask(i
, *cpu_map
) {
6232 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6234 init_sched_groups_power(i
, sd
);
6238 for_each_cpu_mask(i
, *cpu_map
) {
6239 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6241 init_sched_groups_power(i
, sd
);
6245 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6246 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6249 struct sched_group
*sg
;
6251 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6252 init_numa_sched_groups_power(sg
);
6256 /* Attach the domains */
6257 for_each_cpu_mask(i
, *cpu_map
) {
6258 struct sched_domain
*sd
;
6259 #ifdef CONFIG_SCHED_SMT
6260 sd
= &per_cpu(cpu_domains
, i
);
6261 #elif defined(CONFIG_SCHED_MC)
6262 sd
= &per_cpu(core_domains
, i
);
6264 sd
= &per_cpu(phys_domains
, i
);
6266 cpu_attach_domain(sd
, i
);
6273 free_sched_groups(cpu_map
);
6278 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6280 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6282 cpumask_t cpu_default_map
;
6286 * Setup mask for cpus without special case scheduling requirements.
6287 * For now this just excludes isolated cpus, but could be used to
6288 * exclude other special cases in the future.
6290 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6292 err
= build_sched_domains(&cpu_default_map
);
6297 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6299 free_sched_groups(cpu_map
);
6303 * Detach sched domains from a group of cpus specified in cpu_map
6304 * These cpus will now be attached to the NULL domain
6306 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6310 for_each_cpu_mask(i
, *cpu_map
)
6311 cpu_attach_domain(NULL
, i
);
6312 synchronize_sched();
6313 arch_destroy_sched_domains(cpu_map
);
6317 * Partition sched domains as specified by the cpumasks below.
6318 * This attaches all cpus from the cpumasks to the NULL domain,
6319 * waits for a RCU quiescent period, recalculates sched
6320 * domain information and then attaches them back to the
6321 * correct sched domains
6322 * Call with hotplug lock held
6324 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6326 cpumask_t change_map
;
6329 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6330 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6331 cpus_or(change_map
, *partition1
, *partition2
);
6333 /* Detach sched domains from all of the affected cpus */
6334 detach_destroy_domains(&change_map
);
6335 if (!cpus_empty(*partition1
))
6336 err
= build_sched_domains(partition1
);
6337 if (!err
&& !cpus_empty(*partition2
))
6338 err
= build_sched_domains(partition2
);
6343 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6344 static int arch_reinit_sched_domains(void)
6348 mutex_lock(&sched_hotcpu_mutex
);
6349 detach_destroy_domains(&cpu_online_map
);
6350 err
= arch_init_sched_domains(&cpu_online_map
);
6351 mutex_unlock(&sched_hotcpu_mutex
);
6356 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6360 if (buf
[0] != '0' && buf
[0] != '1')
6364 sched_smt_power_savings
= (buf
[0] == '1');
6366 sched_mc_power_savings
= (buf
[0] == '1');
6368 ret
= arch_reinit_sched_domains();
6370 return ret
? ret
: count
;
6373 #ifdef CONFIG_SCHED_MC
6374 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6376 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6378 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6379 const char *buf
, size_t count
)
6381 return sched_power_savings_store(buf
, count
, 0);
6383 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6384 sched_mc_power_savings_store
);
6387 #ifdef CONFIG_SCHED_SMT
6388 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6390 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6392 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6393 const char *buf
, size_t count
)
6395 return sched_power_savings_store(buf
, count
, 1);
6397 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6398 sched_smt_power_savings_store
);
6401 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6405 #ifdef CONFIG_SCHED_SMT
6407 err
= sysfs_create_file(&cls
->kset
.kobj
,
6408 &attr_sched_smt_power_savings
.attr
);
6410 #ifdef CONFIG_SCHED_MC
6411 if (!err
&& mc_capable())
6412 err
= sysfs_create_file(&cls
->kset
.kobj
,
6413 &attr_sched_mc_power_savings
.attr
);
6420 * Force a reinitialization of the sched domains hierarchy. The domains
6421 * and groups cannot be updated in place without racing with the balancing
6422 * code, so we temporarily attach all running cpus to the NULL domain
6423 * which will prevent rebalancing while the sched domains are recalculated.
6425 static int update_sched_domains(struct notifier_block
*nfb
,
6426 unsigned long action
, void *hcpu
)
6429 case CPU_UP_PREPARE
:
6430 case CPU_UP_PREPARE_FROZEN
:
6431 case CPU_DOWN_PREPARE
:
6432 case CPU_DOWN_PREPARE_FROZEN
:
6433 detach_destroy_domains(&cpu_online_map
);
6436 case CPU_UP_CANCELED
:
6437 case CPU_UP_CANCELED_FROZEN
:
6438 case CPU_DOWN_FAILED
:
6439 case CPU_DOWN_FAILED_FROZEN
:
6441 case CPU_ONLINE_FROZEN
:
6443 case CPU_DEAD_FROZEN
:
6445 * Fall through and re-initialise the domains.
6452 /* The hotplug lock is already held by cpu_up/cpu_down */
6453 arch_init_sched_domains(&cpu_online_map
);
6458 void __init
sched_init_smp(void)
6460 cpumask_t non_isolated_cpus
;
6462 mutex_lock(&sched_hotcpu_mutex
);
6463 arch_init_sched_domains(&cpu_online_map
);
6464 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6465 if (cpus_empty(non_isolated_cpus
))
6466 cpu_set(smp_processor_id(), non_isolated_cpus
);
6467 mutex_unlock(&sched_hotcpu_mutex
);
6468 /* XXX: Theoretical race here - CPU may be hotplugged now */
6469 hotcpu_notifier(update_sched_domains
, 0);
6471 init_sched_domain_sysctl();
6473 /* Move init over to a non-isolated CPU */
6474 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6478 void __init
sched_init_smp(void)
6481 #endif /* CONFIG_SMP */
6483 int in_sched_functions(unsigned long addr
)
6485 /* Linker adds these: start and end of __sched functions */
6486 extern char __sched_text_start
[], __sched_text_end
[];
6488 return in_lock_functions(addr
) ||
6489 (addr
>= (unsigned long)__sched_text_start
6490 && addr
< (unsigned long)__sched_text_end
);
6493 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6495 cfs_rq
->tasks_timeline
= RB_ROOT
;
6496 #ifdef CONFIG_FAIR_GROUP_SCHED
6499 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6502 void __init
sched_init(void)
6504 int highest_cpu
= 0;
6508 * Link up the scheduling class hierarchy:
6510 rt_sched_class
.next
= &fair_sched_class
;
6511 fair_sched_class
.next
= &idle_sched_class
;
6512 idle_sched_class
.next
= NULL
;
6514 for_each_possible_cpu(i
) {
6515 struct rt_prio_array
*array
;
6519 spin_lock_init(&rq
->lock
);
6520 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6523 init_cfs_rq(&rq
->cfs
, rq
);
6524 #ifdef CONFIG_FAIR_GROUP_SCHED
6525 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6527 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6528 struct sched_entity
*se
=
6529 &per_cpu(init_sched_entity
, i
);
6531 init_cfs_rq_p
[i
] = cfs_rq
;
6532 init_cfs_rq(cfs_rq
, rq
);
6533 cfs_rq
->tg
= &init_task_grp
;
6534 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6535 &rq
->leaf_cfs_rq_list
);
6537 init_sched_entity_p
[i
] = se
;
6538 se
->cfs_rq
= &rq
->cfs
;
6540 se
->load
.weight
= init_task_grp_load
;
6541 se
->load
.inv_weight
=
6542 div64_64(1ULL<<32, init_task_grp_load
);
6545 init_task_grp
.shares
= init_task_grp_load
;
6548 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6549 rq
->cpu_load
[j
] = 0;
6552 rq
->active_balance
= 0;
6553 rq
->next_balance
= jiffies
;
6556 rq
->migration_thread
= NULL
;
6557 INIT_LIST_HEAD(&rq
->migration_queue
);
6559 atomic_set(&rq
->nr_iowait
, 0);
6561 array
= &rq
->rt
.active
;
6562 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6563 INIT_LIST_HEAD(array
->queue
+ j
);
6564 __clear_bit(j
, array
->bitmap
);
6567 /* delimiter for bitsearch: */
6568 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6571 set_load_weight(&init_task
);
6573 #ifdef CONFIG_PREEMPT_NOTIFIERS
6574 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6578 nr_cpu_ids
= highest_cpu
+ 1;
6579 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6582 #ifdef CONFIG_RT_MUTEXES
6583 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6587 * The boot idle thread does lazy MMU switching as well:
6589 atomic_inc(&init_mm
.mm_count
);
6590 enter_lazy_tlb(&init_mm
, current
);
6593 * Make us the idle thread. Technically, schedule() should not be
6594 * called from this thread, however somewhere below it might be,
6595 * but because we are the idle thread, we just pick up running again
6596 * when this runqueue becomes "idle".
6598 init_idle(current
, smp_processor_id());
6600 * During early bootup we pretend to be a normal task:
6602 current
->sched_class
= &fair_sched_class
;
6605 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6606 void __might_sleep(char *file
, int line
)
6609 static unsigned long prev_jiffy
; /* ratelimiting */
6611 if ((in_atomic() || irqs_disabled()) &&
6612 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6613 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6615 prev_jiffy
= jiffies
;
6616 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6617 " context at %s:%d\n", file
, line
);
6618 printk("in_atomic():%d, irqs_disabled():%d\n",
6619 in_atomic(), irqs_disabled());
6620 debug_show_held_locks(current
);
6621 if (irqs_disabled())
6622 print_irqtrace_events(current
);
6627 EXPORT_SYMBOL(__might_sleep
);
6630 #ifdef CONFIG_MAGIC_SYSRQ
6631 void normalize_rt_tasks(void)
6633 struct task_struct
*g
, *p
;
6634 unsigned long flags
;
6638 read_lock_irq(&tasklist_lock
);
6639 do_each_thread(g
, p
) {
6640 p
->se
.exec_start
= 0;
6641 #ifdef CONFIG_SCHEDSTATS
6642 p
->se
.wait_start
= 0;
6643 p
->se
.sleep_start
= 0;
6644 p
->se
.block_start
= 0;
6646 task_rq(p
)->clock
= 0;
6650 * Renice negative nice level userspace
6653 if (TASK_NICE(p
) < 0 && p
->mm
)
6654 set_user_nice(p
, 0);
6658 spin_lock_irqsave(&p
->pi_lock
, flags
);
6659 rq
= __task_rq_lock(p
);
6662 * Do not touch the migration thread:
6664 if (p
== rq
->migration_thread
)
6668 update_rq_clock(rq
);
6669 on_rq
= p
->se
.on_rq
;
6671 deactivate_task(rq
, p
, 0);
6672 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6674 activate_task(rq
, p
, 0);
6675 resched_task(rq
->curr
);
6680 __task_rq_unlock(rq
);
6681 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6682 } while_each_thread(g
, p
);
6684 read_unlock_irq(&tasklist_lock
);
6687 #endif /* CONFIG_MAGIC_SYSRQ */
6691 * These functions are only useful for the IA64 MCA handling.
6693 * They can only be called when the whole system has been
6694 * stopped - every CPU needs to be quiescent, and no scheduling
6695 * activity can take place. Using them for anything else would
6696 * be a serious bug, and as a result, they aren't even visible
6697 * under any other configuration.
6701 * curr_task - return the current task for a given cpu.
6702 * @cpu: the processor in question.
6704 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6706 struct task_struct
*curr_task(int cpu
)
6708 return cpu_curr(cpu
);
6712 * set_curr_task - set the current task for a given cpu.
6713 * @cpu: the processor in question.
6714 * @p: the task pointer to set.
6716 * Description: This function must only be used when non-maskable interrupts
6717 * are serviced on a separate stack. It allows the architecture to switch the
6718 * notion of the current task on a cpu in a non-blocking manner. This function
6719 * must be called with all CPU's synchronized, and interrupts disabled, the
6720 * and caller must save the original value of the current task (see
6721 * curr_task() above) and restore that value before reenabling interrupts and
6722 * re-starting the system.
6724 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6726 void set_curr_task(int cpu
, struct task_struct
*p
)
6733 #ifdef CONFIG_FAIR_GROUP_SCHED
6735 /* allocate runqueue etc for a new task group */
6736 struct task_grp
*sched_create_group(void)
6738 struct task_grp
*tg
;
6739 struct cfs_rq
*cfs_rq
;
6740 struct sched_entity
*se
;
6744 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6746 return ERR_PTR(-ENOMEM
);
6748 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
6751 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
6755 for_each_possible_cpu(i
) {
6758 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
6763 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
6768 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
6769 memset(se
, 0, sizeof(struct sched_entity
));
6771 tg
->cfs_rq
[i
] = cfs_rq
;
6772 init_cfs_rq(cfs_rq
, rq
);
6776 se
->cfs_rq
= &rq
->cfs
;
6778 se
->load
.weight
= NICE_0_LOAD
;
6779 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
6783 for_each_possible_cpu(i
) {
6785 cfs_rq
= tg
->cfs_rq
[i
];
6786 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6789 tg
->shares
= NICE_0_LOAD
;
6794 for_each_possible_cpu(i
) {
6795 if (tg
->cfs_rq
&& tg
->cfs_rq
[i
])
6796 kfree(tg
->cfs_rq
[i
]);
6797 if (tg
->se
&& tg
->se
[i
])
6807 return ERR_PTR(-ENOMEM
);
6810 /* rcu callback to free various structures associated with a task group */
6811 static void free_sched_group(struct rcu_head
*rhp
)
6813 struct cfs_rq
*cfs_rq
= container_of(rhp
, struct cfs_rq
, rcu
);
6814 struct task_grp
*tg
= cfs_rq
->tg
;
6815 struct sched_entity
*se
;
6818 /* now it should be safe to free those cfs_rqs */
6819 for_each_possible_cpu(i
) {
6820 cfs_rq
= tg
->cfs_rq
[i
];
6832 /* Destroy runqueue etc associated with a task group */
6833 void sched_destroy_group(struct task_grp
*tg
)
6835 struct cfs_rq
*cfs_rq
;
6838 for_each_possible_cpu(i
) {
6839 cfs_rq
= tg
->cfs_rq
[i
];
6840 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
6843 cfs_rq
= tg
->cfs_rq
[0];
6845 /* wait for possible concurrent references to cfs_rqs complete */
6846 call_rcu(&cfs_rq
->rcu
, free_sched_group
);
6849 /* change task's runqueue when it moves between groups.
6850 * The caller of this function should have put the task in its new group
6851 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6852 * reflect its new group.
6854 void sched_move_task(struct task_struct
*tsk
)
6857 unsigned long flags
;
6860 rq
= task_rq_lock(tsk
, &flags
);
6862 if (tsk
->sched_class
!= &fair_sched_class
)
6865 update_rq_clock(rq
);
6867 running
= task_running(rq
, tsk
);
6868 on_rq
= tsk
->se
.on_rq
;
6871 dequeue_task(rq
, tsk
, 0);
6872 if (unlikely(running
))
6873 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6876 set_task_cfs_rq(tsk
);
6879 if (unlikely(running
))
6880 tsk
->sched_class
->set_curr_task(rq
);
6881 enqueue_task(rq
, tsk
, 0);
6885 task_rq_unlock(rq
, &flags
);
6888 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
6890 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
6891 struct rq
*rq
= cfs_rq
->rq
;
6894 spin_lock_irq(&rq
->lock
);
6898 dequeue_entity(cfs_rq
, se
, 0);
6900 se
->load
.weight
= shares
;
6901 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
6904 enqueue_entity(cfs_rq
, se
, 0);
6906 spin_unlock_irq(&rq
->lock
);
6909 int sched_group_set_shares(struct task_grp
*tg
, unsigned long shares
)
6913 if (tg
->shares
== shares
)
6916 /* return -EINVAL if the new value is not sane */
6918 tg
->shares
= shares
;
6919 for_each_possible_cpu(i
)
6920 set_se_shares(tg
->se
[i
], shares
);
6925 #endif /* CONFIG_FAIR_GROUP_SCHED */