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>
68 * Scheduler clock - returns current time in nanosec units.
69 * This is default implementation.
70 * Architectures and sub-architectures can override this.
72 unsigned long long __attribute__((weak
)) sched_clock(void)
74 return (unsigned long long)jiffies
* (1000000000 / HZ
);
78 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
82 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
91 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96 * Some helpers for converting nanosecond timing to jiffy resolution
98 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
99 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
108 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
109 * Timeslices get refilled after they expire.
111 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
112 #define DEF_TIMESLICE (100 * HZ / 1000)
116 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
117 * Since cpu_power is a 'constant', we can use a reciprocal divide.
119 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
121 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
125 * Each time a sched group cpu_power is changed,
126 * we must compute its reciprocal value
128 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
130 sg
->__cpu_power
+= val
;
131 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
135 #define SCALE_PRIO(x, prio) \
136 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
139 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
140 * to time slice values: [800ms ... 100ms ... 5ms]
142 static unsigned int static_prio_timeslice(int static_prio
)
144 if (static_prio
== NICE_TO_PRIO(19))
147 if (static_prio
< NICE_TO_PRIO(0))
148 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
150 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
153 static inline int rt_policy(int policy
)
155 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
160 static inline int task_has_rt_policy(struct task_struct
*p
)
162 return rt_policy(p
->policy
);
166 * This is the priority-queue data structure of the RT scheduling class:
168 struct rt_prio_array
{
169 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
170 struct list_head queue
[MAX_RT_PRIO
];
174 struct load_weight load
;
175 u64 load_update_start
, load_update_last
;
176 unsigned long delta_fair
, delta_exec
, delta_stat
;
179 /* CFS-related fields in a runqueue */
181 struct load_weight load
;
182 unsigned long nr_running
;
188 unsigned long wait_runtime_overruns
, wait_runtime_underruns
;
190 struct rb_root tasks_timeline
;
191 struct rb_node
*rb_leftmost
;
192 struct rb_node
*rb_load_balance_curr
;
193 #ifdef CONFIG_FAIR_GROUP_SCHED
194 /* 'curr' points to currently running entity on this cfs_rq.
195 * It is set to NULL otherwise (i.e when none are currently running).
197 struct sched_entity
*curr
;
198 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
207 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
211 /* Real-Time classes' related field in a runqueue: */
213 struct rt_prio_array active
;
214 int rt_load_balance_idx
;
215 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
219 * This is the main, per-CPU runqueue data structure.
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
226 spinlock_t lock
; /* runqueue lock */
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
232 unsigned long nr_running
;
233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
235 unsigned char idle_at_tick
;
237 unsigned char in_nohz_recently
;
239 struct load_stat ls
; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates
;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
255 unsigned long nr_uninterruptible
;
257 struct task_struct
*curr
, *idle
;
258 unsigned long next_balance
;
259 struct mm_struct
*prev_mm
;
261 u64 clock
, prev_clock_raw
;
264 unsigned int clock_warps
, clock_overflows
;
266 unsigned int clock_deep_idle_events
;
272 struct sched_domain
*sd
;
274 /* For active balancing */
277 int cpu
; /* cpu of this runqueue */
279 struct task_struct
*migration_thread
;
280 struct list_head migration_queue
;
283 #ifdef CONFIG_SCHEDSTATS
285 struct sched_info rq_sched_info
;
287 /* sys_sched_yield() stats */
288 unsigned long yld_exp_empty
;
289 unsigned long yld_act_empty
;
290 unsigned long yld_both_empty
;
291 unsigned long yld_cnt
;
293 /* schedule() stats */
294 unsigned long sched_switch
;
295 unsigned long sched_cnt
;
296 unsigned long sched_goidle
;
298 /* try_to_wake_up() stats */
299 unsigned long ttwu_cnt
;
300 unsigned long ttwu_local
;
302 struct lock_class_key rq_lock_key
;
305 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
306 static DEFINE_MUTEX(sched_hotcpu_mutex
);
308 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
310 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
313 static inline int cpu_of(struct rq
*rq
)
323 * Update the per-runqueue clock, as finegrained as the platform can give
324 * us, but without assuming monotonicity, etc.:
326 static void __update_rq_clock(struct rq
*rq
)
328 u64 prev_raw
= rq
->prev_clock_raw
;
329 u64 now
= sched_clock();
330 s64 delta
= now
- prev_raw
;
331 u64 clock
= rq
->clock
;
333 #ifdef CONFIG_SCHED_DEBUG
334 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
337 * Protect against sched_clock() occasionally going backwards:
339 if (unlikely(delta
< 0)) {
344 * Catch too large forward jumps too:
346 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
347 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
348 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
351 rq
->clock_overflows
++;
353 if (unlikely(delta
> rq
->clock_max_delta
))
354 rq
->clock_max_delta
= delta
;
359 rq
->prev_clock_raw
= now
;
363 static void update_rq_clock(struct rq
*rq
)
365 if (likely(smp_processor_id() == cpu_of(rq
)))
366 __update_rq_clock(rq
);
370 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
371 * See detach_destroy_domains: synchronize_sched for details.
373 * The domain tree of any CPU may only be accessed from within
374 * preempt-disabled sections.
376 #define for_each_domain(cpu, __sd) \
377 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
379 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
380 #define this_rq() (&__get_cpu_var(runqueues))
381 #define task_rq(p) cpu_rq(task_cpu(p))
382 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
385 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
386 * clock constructed from sched_clock():
388 unsigned long long cpu_clock(int cpu
)
390 unsigned long long now
;
394 local_irq_save(flags
);
398 local_irq_restore(flags
);
403 #ifdef CONFIG_FAIR_GROUP_SCHED
404 /* Change a task's ->cfs_rq if it moves across CPUs */
405 static inline void set_task_cfs_rq(struct task_struct
*p
)
407 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
410 static inline void set_task_cfs_rq(struct task_struct
*p
)
415 #ifndef prepare_arch_switch
416 # define prepare_arch_switch(next) do { } while (0)
418 #ifndef finish_arch_switch
419 # define finish_arch_switch(prev) do { } while (0)
422 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
423 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
425 return rq
->curr
== p
;
428 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
432 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
434 #ifdef CONFIG_DEBUG_SPINLOCK
435 /* this is a valid case when another task releases the spinlock */
436 rq
->lock
.owner
= current
;
439 * If we are tracking spinlock dependencies then we have to
440 * fix up the runqueue lock - which gets 'carried over' from
443 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
445 spin_unlock_irq(&rq
->lock
);
448 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
449 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
454 return rq
->curr
== p
;
458 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
462 * We can optimise this out completely for !SMP, because the
463 * SMP rebalancing from interrupt is the only thing that cares
468 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
469 spin_unlock_irq(&rq
->lock
);
471 spin_unlock(&rq
->lock
);
475 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
479 * After ->oncpu is cleared, the task can be moved to a different CPU.
480 * We must ensure this doesn't happen until the switch is completely
486 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
490 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
493 * __task_rq_lock - lock the runqueue a given task resides on.
494 * Must be called interrupts disabled.
496 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
503 spin_lock(&rq
->lock
);
504 if (unlikely(rq
!= task_rq(p
))) {
505 spin_unlock(&rq
->lock
);
506 goto repeat_lock_task
;
512 * task_rq_lock - lock the runqueue a given task resides on and disable
513 * interrupts. Note the ordering: we can safely lookup the task_rq without
514 * explicitly disabling preemption.
516 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
522 local_irq_save(*flags
);
524 spin_lock(&rq
->lock
);
525 if (unlikely(rq
!= task_rq(p
))) {
526 spin_unlock_irqrestore(&rq
->lock
, *flags
);
527 goto repeat_lock_task
;
532 static inline void __task_rq_unlock(struct rq
*rq
)
535 spin_unlock(&rq
->lock
);
538 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
541 spin_unlock_irqrestore(&rq
->lock
, *flags
);
545 * this_rq_lock - lock this runqueue and disable interrupts.
547 static inline struct rq
*this_rq_lock(void)
554 spin_lock(&rq
->lock
);
560 * We are going deep-idle (irqs are disabled):
562 void sched_clock_idle_sleep_event(void)
564 struct rq
*rq
= cpu_rq(smp_processor_id());
566 spin_lock(&rq
->lock
);
567 __update_rq_clock(rq
);
568 spin_unlock(&rq
->lock
);
569 rq
->clock_deep_idle_events
++;
571 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
574 * We just idled delta nanoseconds (called with irqs disabled):
576 void sched_clock_idle_wakeup_event(u64 delta_ns
)
578 struct rq
*rq
= cpu_rq(smp_processor_id());
579 u64 now
= sched_clock();
581 rq
->idle_clock
+= delta_ns
;
583 * Override the previous timestamp and ignore all
584 * sched_clock() deltas that occured while we idled,
585 * and use the PM-provided delta_ns to advance the
588 spin_lock(&rq
->lock
);
589 rq
->prev_clock_raw
= now
;
590 rq
->clock
+= delta_ns
;
591 spin_unlock(&rq
->lock
);
593 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
596 * resched_task - mark a task 'to be rescheduled now'.
598 * On UP this means the setting of the need_resched flag, on SMP it
599 * might also involve a cross-CPU call to trigger the scheduler on
604 #ifndef tsk_is_polling
605 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
608 static void resched_task(struct task_struct
*p
)
612 assert_spin_locked(&task_rq(p
)->lock
);
614 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
617 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
620 if (cpu
== smp_processor_id())
623 /* NEED_RESCHED must be visible before we test polling */
625 if (!tsk_is_polling(p
))
626 smp_send_reschedule(cpu
);
629 static void resched_cpu(int cpu
)
631 struct rq
*rq
= cpu_rq(cpu
);
634 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
636 resched_task(cpu_curr(cpu
));
637 spin_unlock_irqrestore(&rq
->lock
, flags
);
640 static inline void resched_task(struct task_struct
*p
)
642 assert_spin_locked(&task_rq(p
)->lock
);
643 set_tsk_need_resched(p
);
647 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
649 #if BITS_PER_LONG == 32
650 if (likely(divident
<= 0xffffffffULL
))
651 return (u32
)divident
/ divisor
;
652 do_div(divident
, divisor
);
656 return divident
/ divisor
;
660 #if BITS_PER_LONG == 32
661 # define WMULT_CONST (~0UL)
663 # define WMULT_CONST (1UL << 32)
666 #define WMULT_SHIFT 32
669 * Shift right and round:
671 #define RSR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
674 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
675 struct load_weight
*lw
)
679 if (unlikely(!lw
->inv_weight
))
680 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
682 tmp
= (u64
)delta_exec
* weight
;
684 * Check whether we'd overflow the 64-bit multiplication:
686 if (unlikely(tmp
> WMULT_CONST
))
687 tmp
= RSR(RSR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
690 tmp
= RSR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
692 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
695 static inline unsigned long
696 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
698 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
701 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
707 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
714 * To aid in avoiding the subversion of "niceness" due to uneven distribution
715 * of tasks with abnormal "nice" values across CPUs the contribution that
716 * each task makes to its run queue's load is weighted according to its
717 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
718 * scaled version of the new time slice allocation that they receive on time
722 #define WEIGHT_IDLEPRIO 2
723 #define WMULT_IDLEPRIO (1 << 31)
726 * Nice levels are multiplicative, with a gentle 10% change for every
727 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
728 * nice 1, it will get ~10% less CPU time than another CPU-bound task
729 * that remained on nice 0.
731 * The "10% effect" is relative and cumulative: from _any_ nice level,
732 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
733 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
734 * If a task goes up by ~10% and another task goes down by ~10% then
735 * the relative distance between them is ~25%.)
737 static const int prio_to_weight
[40] = {
738 /* -20 */ 88761, 71755, 56483, 46273, 36291,
739 /* -15 */ 29154, 23254, 18705, 14949, 11916,
740 /* -10 */ 9548, 7620, 6100, 4904, 3906,
741 /* -5 */ 3121, 2501, 1991, 1586, 1277,
742 /* 0 */ 1024, 820, 655, 526, 423,
743 /* 5 */ 335, 272, 215, 172, 137,
744 /* 10 */ 110, 87, 70, 56, 45,
745 /* 15 */ 36, 29, 23, 18, 15,
749 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
751 * In cases where the weight does not change often, we can use the
752 * precalculated inverse to speed up arithmetics by turning divisions
753 * into multiplications:
755 static const u32 prio_to_wmult
[40] = {
756 /* -20 */ 48388, 59856, 76040, 92818, 118348,
757 /* -15 */ 147320, 184698, 229616, 287308, 360437,
758 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
759 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
760 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
761 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
762 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
763 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
766 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
769 * runqueue iterator, to support SMP load-balancing between different
770 * scheduling classes, without having to expose their internal data
771 * structures to the load-balancing proper:
775 struct task_struct
*(*start
)(void *);
776 struct task_struct
*(*next
)(void *);
779 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
780 unsigned long max_nr_move
, unsigned long max_load_move
,
781 struct sched_domain
*sd
, enum cpu_idle_type idle
,
782 int *all_pinned
, unsigned long *load_moved
,
783 int *this_best_prio
, struct rq_iterator
*iterator
);
785 #include "sched_stats.h"
786 #include "sched_rt.c"
787 #include "sched_fair.c"
788 #include "sched_idletask.c"
789 #ifdef CONFIG_SCHED_DEBUG
790 # include "sched_debug.c"
793 #define sched_class_highest (&rt_sched_class)
795 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
797 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
798 ls
->delta_exec
+= ls
->delta_stat
;
799 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
805 * Update delta_exec, delta_fair fields for rq.
807 * delta_fair clock advances at a rate inversely proportional to
808 * total load (rq->ls.load.weight) on the runqueue, while
809 * delta_exec advances at the same rate as wall-clock (provided
812 * delta_exec / delta_fair is a measure of the (smoothened) load on this
813 * runqueue over any given interval. This (smoothened) load is used
814 * during load balance.
816 * This function is called /before/ updating rq->ls.load
817 * and when switching tasks.
819 static void update_curr_load(struct rq
*rq
)
821 struct load_stat
*ls
= &rq
->ls
;
824 start
= ls
->load_update_start
;
825 ls
->load_update_start
= rq
->clock
;
826 ls
->delta_stat
+= rq
->clock
- start
;
828 * Stagger updates to ls->delta_fair. Very frequent updates
831 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
832 __update_curr_load(rq
, ls
);
835 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
837 update_curr_load(rq
);
838 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
841 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
843 update_curr_load(rq
);
844 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
847 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
853 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
859 static void set_load_weight(struct task_struct
*p
)
861 task_rq(p
)->cfs
.wait_runtime
-= p
->se
.wait_runtime
;
862 p
->se
.wait_runtime
= 0;
864 if (task_has_rt_policy(p
)) {
865 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
866 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
871 * SCHED_IDLE tasks get minimal weight:
873 if (p
->policy
== SCHED_IDLE
) {
874 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
875 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
879 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
880 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
883 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
885 sched_info_queued(p
);
886 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
890 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
892 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
897 * __normal_prio - return the priority that is based on the static prio
899 static inline int __normal_prio(struct task_struct
*p
)
901 return p
->static_prio
;
905 * Calculate the expected normal priority: i.e. priority
906 * without taking RT-inheritance into account. Might be
907 * boosted by interactivity modifiers. Changes upon fork,
908 * setprio syscalls, and whenever the interactivity
909 * estimator recalculates.
911 static inline int normal_prio(struct task_struct
*p
)
915 if (task_has_rt_policy(p
))
916 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
918 prio
= __normal_prio(p
);
923 * Calculate the current priority, i.e. the priority
924 * taken into account by the scheduler. This value might
925 * be boosted by RT tasks, or might be boosted by
926 * interactivity modifiers. Will be RT if the task got
927 * RT-boosted. If not then it returns p->normal_prio.
929 static int effective_prio(struct task_struct
*p
)
931 p
->normal_prio
= normal_prio(p
);
933 * If we are RT tasks or we were boosted to RT priority,
934 * keep the priority unchanged. Otherwise, update priority
935 * to the normal priority:
937 if (!rt_prio(p
->prio
))
938 return p
->normal_prio
;
943 * activate_task - move a task to the runqueue.
945 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
947 if (p
->state
== TASK_UNINTERRUPTIBLE
)
948 rq
->nr_uninterruptible
--;
950 enqueue_task(rq
, p
, wakeup
);
951 inc_nr_running(p
, rq
);
955 * activate_idle_task - move idle task to the _front_ of runqueue.
957 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
961 if (p
->state
== TASK_UNINTERRUPTIBLE
)
962 rq
->nr_uninterruptible
--;
964 enqueue_task(rq
, p
, 0);
965 inc_nr_running(p
, rq
);
969 * deactivate_task - remove a task from the runqueue.
971 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
973 if (p
->state
== TASK_UNINTERRUPTIBLE
)
974 rq
->nr_uninterruptible
++;
976 dequeue_task(rq
, p
, sleep
);
977 dec_nr_running(p
, rq
);
981 * task_curr - is this task currently executing on a CPU?
982 * @p: the task in question.
984 inline int task_curr(const struct task_struct
*p
)
986 return cpu_curr(task_cpu(p
)) == p
;
989 /* Used instead of source_load when we know the type == 0 */
990 unsigned long weighted_cpuload(const int cpu
)
992 return cpu_rq(cpu
)->ls
.load
.weight
;
995 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
998 task_thread_info(p
)->cpu
= cpu
;
1005 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1007 int old_cpu
= task_cpu(p
);
1008 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1009 u64 clock_offset
, fair_clock_offset
;
1011 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1012 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
1014 if (p
->se
.wait_start_fair
)
1015 p
->se
.wait_start_fair
-= fair_clock_offset
;
1016 if (p
->se
.sleep_start_fair
)
1017 p
->se
.sleep_start_fair
-= fair_clock_offset
;
1019 #ifdef CONFIG_SCHEDSTATS
1020 if (p
->se
.wait_start
)
1021 p
->se
.wait_start
-= clock_offset
;
1022 if (p
->se
.sleep_start
)
1023 p
->se
.sleep_start
-= clock_offset
;
1024 if (p
->se
.block_start
)
1025 p
->se
.block_start
-= clock_offset
;
1028 __set_task_cpu(p
, new_cpu
);
1031 struct migration_req
{
1032 struct list_head list
;
1034 struct task_struct
*task
;
1037 struct completion done
;
1041 * The task's runqueue lock must be held.
1042 * Returns true if you have to wait for migration thread.
1045 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1047 struct rq
*rq
= task_rq(p
);
1050 * If the task is not on a runqueue (and not running), then
1051 * it is sufficient to simply update the task's cpu field.
1053 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1054 set_task_cpu(p
, dest_cpu
);
1058 init_completion(&req
->done
);
1060 req
->dest_cpu
= dest_cpu
;
1061 list_add(&req
->list
, &rq
->migration_queue
);
1067 * wait_task_inactive - wait for a thread to unschedule.
1069 * The caller must ensure that the task *will* unschedule sometime soon,
1070 * else this function might spin for a *long* time. This function can't
1071 * be called with interrupts off, or it may introduce deadlock with
1072 * smp_call_function() if an IPI is sent by the same process we are
1073 * waiting to become inactive.
1075 void wait_task_inactive(struct task_struct
*p
)
1077 unsigned long flags
;
1083 * We do the initial early heuristics without holding
1084 * any task-queue locks at all. We'll only try to get
1085 * the runqueue lock when things look like they will
1091 * If the task is actively running on another CPU
1092 * still, just relax and busy-wait without holding
1095 * NOTE! Since we don't hold any locks, it's not
1096 * even sure that "rq" stays as the right runqueue!
1097 * But we don't care, since "task_running()" will
1098 * return false if the runqueue has changed and p
1099 * is actually now running somewhere else!
1101 while (task_running(rq
, p
))
1105 * Ok, time to look more closely! We need the rq
1106 * lock now, to be *sure*. If we're wrong, we'll
1107 * just go back and repeat.
1109 rq
= task_rq_lock(p
, &flags
);
1110 running
= task_running(rq
, p
);
1111 on_rq
= p
->se
.on_rq
;
1112 task_rq_unlock(rq
, &flags
);
1115 * Was it really running after all now that we
1116 * checked with the proper locks actually held?
1118 * Oops. Go back and try again..
1120 if (unlikely(running
)) {
1126 * It's not enough that it's not actively running,
1127 * it must be off the runqueue _entirely_, and not
1130 * So if it wa still runnable (but just not actively
1131 * running right now), it's preempted, and we should
1132 * yield - it could be a while.
1134 if (unlikely(on_rq
)) {
1140 * Ahh, all good. It wasn't running, and it wasn't
1141 * runnable, which means that it will never become
1142 * running in the future either. We're all done!
1147 * kick_process - kick a running thread to enter/exit the kernel
1148 * @p: the to-be-kicked thread
1150 * Cause a process which is running on another CPU to enter
1151 * kernel-mode, without any delay. (to get signals handled.)
1153 * NOTE: this function doesnt have to take the runqueue lock,
1154 * because all it wants to ensure is that the remote task enters
1155 * the kernel. If the IPI races and the task has been migrated
1156 * to another CPU then no harm is done and the purpose has been
1159 void kick_process(struct task_struct
*p
)
1165 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1166 smp_send_reschedule(cpu
);
1171 * Return a low guess at the load of a migration-source cpu weighted
1172 * according to the scheduling class and "nice" value.
1174 * We want to under-estimate the load of migration sources, to
1175 * balance conservatively.
1177 static inline unsigned long source_load(int cpu
, int type
)
1179 struct rq
*rq
= cpu_rq(cpu
);
1180 unsigned long total
= weighted_cpuload(cpu
);
1185 return min(rq
->cpu_load
[type
-1], total
);
1189 * Return a high guess at the load of a migration-target cpu weighted
1190 * according to the scheduling class and "nice" value.
1192 static inline unsigned long target_load(int cpu
, int type
)
1194 struct rq
*rq
= cpu_rq(cpu
);
1195 unsigned long total
= weighted_cpuload(cpu
);
1200 return max(rq
->cpu_load
[type
-1], total
);
1204 * Return the average load per task on the cpu's run queue
1206 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1208 struct rq
*rq
= cpu_rq(cpu
);
1209 unsigned long total
= weighted_cpuload(cpu
);
1210 unsigned long n
= rq
->nr_running
;
1212 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1216 * find_idlest_group finds and returns the least busy CPU group within the
1219 static struct sched_group
*
1220 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1222 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1223 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1224 int load_idx
= sd
->forkexec_idx
;
1225 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1228 unsigned long load
, avg_load
;
1232 /* Skip over this group if it has no CPUs allowed */
1233 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1236 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1238 /* Tally up the load of all CPUs in the group */
1241 for_each_cpu_mask(i
, group
->cpumask
) {
1242 /* Bias balancing toward cpus of our domain */
1244 load
= source_load(i
, load_idx
);
1246 load
= target_load(i
, load_idx
);
1251 /* Adjust by relative CPU power of the group */
1252 avg_load
= sg_div_cpu_power(group
,
1253 avg_load
* SCHED_LOAD_SCALE
);
1256 this_load
= avg_load
;
1258 } else if (avg_load
< min_load
) {
1259 min_load
= avg_load
;
1263 group
= group
->next
;
1264 } while (group
!= sd
->groups
);
1266 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1272 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1275 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1278 unsigned long load
, min_load
= ULONG_MAX
;
1282 /* Traverse only the allowed CPUs */
1283 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1285 for_each_cpu_mask(i
, tmp
) {
1286 load
= weighted_cpuload(i
);
1288 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1298 * sched_balance_self: balance the current task (running on cpu) in domains
1299 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1302 * Balance, ie. select the least loaded group.
1304 * Returns the target CPU number, or the same CPU if no balancing is needed.
1306 * preempt must be disabled.
1308 static int sched_balance_self(int cpu
, int flag
)
1310 struct task_struct
*t
= current
;
1311 struct sched_domain
*tmp
, *sd
= NULL
;
1313 for_each_domain(cpu
, tmp
) {
1315 * If power savings logic is enabled for a domain, stop there.
1317 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1319 if (tmp
->flags
& flag
)
1325 struct sched_group
*group
;
1326 int new_cpu
, weight
;
1328 if (!(sd
->flags
& flag
)) {
1334 group
= find_idlest_group(sd
, t
, cpu
);
1340 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1341 if (new_cpu
== -1 || new_cpu
== cpu
) {
1342 /* Now try balancing at a lower domain level of cpu */
1347 /* Now try balancing at a lower domain level of new_cpu */
1350 weight
= cpus_weight(span
);
1351 for_each_domain(cpu
, tmp
) {
1352 if (weight
<= cpus_weight(tmp
->span
))
1354 if (tmp
->flags
& flag
)
1357 /* while loop will break here if sd == NULL */
1363 #endif /* CONFIG_SMP */
1366 * wake_idle() will wake a task on an idle cpu if task->cpu is
1367 * not idle and an idle cpu is available. The span of cpus to
1368 * search starts with cpus closest then further out as needed,
1369 * so we always favor a closer, idle cpu.
1371 * Returns the CPU we should wake onto.
1373 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1374 static int wake_idle(int cpu
, struct task_struct
*p
)
1377 struct sched_domain
*sd
;
1381 * If it is idle, then it is the best cpu to run this task.
1383 * This cpu is also the best, if it has more than one task already.
1384 * Siblings must be also busy(in most cases) as they didn't already
1385 * pickup the extra load from this cpu and hence we need not check
1386 * sibling runqueue info. This will avoid the checks and cache miss
1387 * penalities associated with that.
1389 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1392 for_each_domain(cpu
, sd
) {
1393 if (sd
->flags
& SD_WAKE_IDLE
) {
1394 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1395 for_each_cpu_mask(i
, tmp
) {
1406 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1413 * try_to_wake_up - wake up a thread
1414 * @p: the to-be-woken-up thread
1415 * @state: the mask of task states that can be woken
1416 * @sync: do a synchronous wakeup?
1418 * Put it on the run-queue if it's not already there. The "current"
1419 * thread is always on the run-queue (except when the actual
1420 * re-schedule is in progress), and as such you're allowed to do
1421 * the simpler "current->state = TASK_RUNNING" to mark yourself
1422 * runnable without the overhead of this.
1424 * returns failure only if the task is already active.
1426 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1428 int cpu
, this_cpu
, success
= 0;
1429 unsigned long flags
;
1433 struct sched_domain
*sd
, *this_sd
= NULL
;
1434 unsigned long load
, this_load
;
1438 rq
= task_rq_lock(p
, &flags
);
1439 old_state
= p
->state
;
1440 if (!(old_state
& state
))
1447 this_cpu
= smp_processor_id();
1450 if (unlikely(task_running(rq
, p
)))
1455 schedstat_inc(rq
, ttwu_cnt
);
1456 if (cpu
== this_cpu
) {
1457 schedstat_inc(rq
, ttwu_local
);
1461 for_each_domain(this_cpu
, sd
) {
1462 if (cpu_isset(cpu
, sd
->span
)) {
1463 schedstat_inc(sd
, ttwu_wake_remote
);
1469 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1473 * Check for affine wakeup and passive balancing possibilities.
1476 int idx
= this_sd
->wake_idx
;
1477 unsigned int imbalance
;
1479 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1481 load
= source_load(cpu
, idx
);
1482 this_load
= target_load(this_cpu
, idx
);
1484 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1486 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1487 unsigned long tl
= this_load
;
1488 unsigned long tl_per_task
;
1490 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1493 * If sync wakeup then subtract the (maximum possible)
1494 * effect of the currently running task from the load
1495 * of the current CPU:
1498 tl
-= current
->se
.load
.weight
;
1501 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1502 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1504 * This domain has SD_WAKE_AFFINE and
1505 * p is cache cold in this domain, and
1506 * there is no bad imbalance.
1508 schedstat_inc(this_sd
, ttwu_move_affine
);
1514 * Start passive balancing when half the imbalance_pct
1517 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1518 if (imbalance
*this_load
<= 100*load
) {
1519 schedstat_inc(this_sd
, ttwu_move_balance
);
1525 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1527 new_cpu
= wake_idle(new_cpu
, p
);
1528 if (new_cpu
!= cpu
) {
1529 set_task_cpu(p
, new_cpu
);
1530 task_rq_unlock(rq
, &flags
);
1531 /* might preempt at this point */
1532 rq
= task_rq_lock(p
, &flags
);
1533 old_state
= p
->state
;
1534 if (!(old_state
& state
))
1539 this_cpu
= smp_processor_id();
1544 #endif /* CONFIG_SMP */
1545 update_rq_clock(rq
);
1546 activate_task(rq
, p
, 1);
1548 * Sync wakeups (i.e. those types of wakeups where the waker
1549 * has indicated that it will leave the CPU in short order)
1550 * don't trigger a preemption, if the woken up task will run on
1551 * this cpu. (in this case the 'I will reschedule' promise of
1552 * the waker guarantees that the freshly woken up task is going
1553 * to be considered on this CPU.)
1555 if (!sync
|| cpu
!= this_cpu
)
1556 check_preempt_curr(rq
, p
);
1560 p
->state
= TASK_RUNNING
;
1562 task_rq_unlock(rq
, &flags
);
1567 int fastcall
wake_up_process(struct task_struct
*p
)
1569 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1570 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1572 EXPORT_SYMBOL(wake_up_process
);
1574 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1576 return try_to_wake_up(p
, state
, 0);
1580 * Perform scheduler related setup for a newly forked process p.
1581 * p is forked by current.
1583 * __sched_fork() is basic setup used by init_idle() too:
1585 static void __sched_fork(struct task_struct
*p
)
1587 p
->se
.wait_start_fair
= 0;
1588 p
->se
.exec_start
= 0;
1589 p
->se
.sum_exec_runtime
= 0;
1590 p
->se
.delta_exec
= 0;
1591 p
->se
.delta_fair_run
= 0;
1592 p
->se
.delta_fair_sleep
= 0;
1593 p
->se
.wait_runtime
= 0;
1594 p
->se
.sleep_start_fair
= 0;
1596 #ifdef CONFIG_SCHEDSTATS
1597 p
->se
.wait_start
= 0;
1598 p
->se
.sum_wait_runtime
= 0;
1599 p
->se
.sum_sleep_runtime
= 0;
1600 p
->se
.sleep_start
= 0;
1601 p
->se
.block_start
= 0;
1602 p
->se
.sleep_max
= 0;
1603 p
->se
.block_max
= 0;
1606 p
->se
.wait_runtime_overruns
= 0;
1607 p
->se
.wait_runtime_underruns
= 0;
1610 INIT_LIST_HEAD(&p
->run_list
);
1613 #ifdef CONFIG_PREEMPT_NOTIFIERS
1614 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1618 * We mark the process as running here, but have not actually
1619 * inserted it onto the runqueue yet. This guarantees that
1620 * nobody will actually run it, and a signal or other external
1621 * event cannot wake it up and insert it on the runqueue either.
1623 p
->state
= TASK_RUNNING
;
1627 * fork()/clone()-time setup:
1629 void sched_fork(struct task_struct
*p
, int clone_flags
)
1631 int cpu
= get_cpu();
1636 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1638 __set_task_cpu(p
, cpu
);
1641 * Make sure we do not leak PI boosting priority to the child:
1643 p
->prio
= current
->normal_prio
;
1645 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1646 if (likely(sched_info_on()))
1647 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1649 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1652 #ifdef CONFIG_PREEMPT
1653 /* Want to start with kernel preemption disabled. */
1654 task_thread_info(p
)->preempt_count
= 1;
1660 * After fork, child runs first. (default) If set to 0 then
1661 * parent will (try to) run first.
1663 unsigned int __read_mostly sysctl_sched_child_runs_first
= 1;
1666 * wake_up_new_task - wake up a newly created task for the first time.
1668 * This function will do some initial scheduler statistics housekeeping
1669 * that must be done for every newly created context, then puts the task
1670 * on the runqueue and wakes it.
1672 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1674 unsigned long flags
;
1678 rq
= task_rq_lock(p
, &flags
);
1679 BUG_ON(p
->state
!= TASK_RUNNING
);
1680 this_cpu
= smp_processor_id(); /* parent's CPU */
1681 update_rq_clock(rq
);
1683 p
->prio
= effective_prio(p
);
1685 if (!p
->sched_class
->task_new
|| !sysctl_sched_child_runs_first
||
1686 (clone_flags
& CLONE_VM
) || task_cpu(p
) != this_cpu
||
1687 !current
->se
.on_rq
) {
1689 activate_task(rq
, p
, 0);
1692 * Let the scheduling class do new task startup
1693 * management (if any):
1695 p
->sched_class
->task_new(rq
, p
);
1696 inc_nr_running(p
, rq
);
1698 check_preempt_curr(rq
, p
);
1699 task_rq_unlock(rq
, &flags
);
1702 #ifdef CONFIG_PREEMPT_NOTIFIERS
1705 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1706 * @notifier: notifier struct to register
1708 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1710 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1712 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1715 * preempt_notifier_unregister - no longer interested in preemption notifications
1716 * @notifier: notifier struct to unregister
1718 * This is safe to call from within a preemption notifier.
1720 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1722 hlist_del(¬ifier
->link
);
1724 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1726 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1728 struct preempt_notifier
*notifier
;
1729 struct hlist_node
*node
;
1731 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1732 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1736 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1737 struct task_struct
*next
)
1739 struct preempt_notifier
*notifier
;
1740 struct hlist_node
*node
;
1742 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1743 notifier
->ops
->sched_out(notifier
, next
);
1748 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1753 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1754 struct task_struct
*next
)
1761 * prepare_task_switch - prepare to switch tasks
1762 * @rq: the runqueue preparing to switch
1763 * @prev: the current task that is being switched out
1764 * @next: the task we are going to switch to.
1766 * This is called with the rq lock held and interrupts off. It must
1767 * be paired with a subsequent finish_task_switch after the context
1770 * prepare_task_switch sets up locking and calls architecture specific
1774 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1775 struct task_struct
*next
)
1777 fire_sched_out_preempt_notifiers(prev
, next
);
1778 prepare_lock_switch(rq
, next
);
1779 prepare_arch_switch(next
);
1783 * finish_task_switch - clean up after a task-switch
1784 * @rq: runqueue associated with task-switch
1785 * @prev: the thread we just switched away from.
1787 * finish_task_switch must be called after the context switch, paired
1788 * with a prepare_task_switch call before the context switch.
1789 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1790 * and do any other architecture-specific cleanup actions.
1792 * Note that we may have delayed dropping an mm in context_switch(). If
1793 * so, we finish that here outside of the runqueue lock. (Doing it
1794 * with the lock held can cause deadlocks; see schedule() for
1797 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1798 __releases(rq
->lock
)
1800 struct mm_struct
*mm
= rq
->prev_mm
;
1806 * A task struct has one reference for the use as "current".
1807 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1808 * schedule one last time. The schedule call will never return, and
1809 * the scheduled task must drop that reference.
1810 * The test for TASK_DEAD must occur while the runqueue locks are
1811 * still held, otherwise prev could be scheduled on another cpu, die
1812 * there before we look at prev->state, and then the reference would
1814 * Manfred Spraul <manfred@colorfullife.com>
1816 prev_state
= prev
->state
;
1817 finish_arch_switch(prev
);
1818 finish_lock_switch(rq
, prev
);
1819 fire_sched_in_preempt_notifiers(current
);
1822 if (unlikely(prev_state
== TASK_DEAD
)) {
1824 * Remove function-return probe instances associated with this
1825 * task and put them back on the free list.
1827 kprobe_flush_task(prev
);
1828 put_task_struct(prev
);
1833 * schedule_tail - first thing a freshly forked thread must call.
1834 * @prev: the thread we just switched away from.
1836 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1837 __releases(rq
->lock
)
1839 struct rq
*rq
= this_rq();
1841 finish_task_switch(rq
, prev
);
1842 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1843 /* In this case, finish_task_switch does not reenable preemption */
1846 if (current
->set_child_tid
)
1847 put_user(current
->pid
, current
->set_child_tid
);
1851 * context_switch - switch to the new MM and the new
1852 * thread's register state.
1855 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1856 struct task_struct
*next
)
1858 struct mm_struct
*mm
, *oldmm
;
1860 prepare_task_switch(rq
, prev
, next
);
1862 oldmm
= prev
->active_mm
;
1864 * For paravirt, this is coupled with an exit in switch_to to
1865 * combine the page table reload and the switch backend into
1868 arch_enter_lazy_cpu_mode();
1870 if (unlikely(!mm
)) {
1871 next
->active_mm
= oldmm
;
1872 atomic_inc(&oldmm
->mm_count
);
1873 enter_lazy_tlb(oldmm
, next
);
1875 switch_mm(oldmm
, mm
, next
);
1877 if (unlikely(!prev
->mm
)) {
1878 prev
->active_mm
= NULL
;
1879 rq
->prev_mm
= oldmm
;
1882 * Since the runqueue lock will be released by the next
1883 * task (which is an invalid locking op but in the case
1884 * of the scheduler it's an obvious special-case), so we
1885 * do an early lockdep release here:
1887 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1888 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1891 /* Here we just switch the register state and the stack. */
1892 switch_to(prev
, next
, prev
);
1896 * this_rq must be evaluated again because prev may have moved
1897 * CPUs since it called schedule(), thus the 'rq' on its stack
1898 * frame will be invalid.
1900 finish_task_switch(this_rq(), prev
);
1904 * nr_running, nr_uninterruptible and nr_context_switches:
1906 * externally visible scheduler statistics: current number of runnable
1907 * threads, current number of uninterruptible-sleeping threads, total
1908 * number of context switches performed since bootup.
1910 unsigned long nr_running(void)
1912 unsigned long i
, sum
= 0;
1914 for_each_online_cpu(i
)
1915 sum
+= cpu_rq(i
)->nr_running
;
1920 unsigned long nr_uninterruptible(void)
1922 unsigned long i
, sum
= 0;
1924 for_each_possible_cpu(i
)
1925 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1928 * Since we read the counters lockless, it might be slightly
1929 * inaccurate. Do not allow it to go below zero though:
1931 if (unlikely((long)sum
< 0))
1937 unsigned long long nr_context_switches(void)
1940 unsigned long long sum
= 0;
1942 for_each_possible_cpu(i
)
1943 sum
+= cpu_rq(i
)->nr_switches
;
1948 unsigned long nr_iowait(void)
1950 unsigned long i
, sum
= 0;
1952 for_each_possible_cpu(i
)
1953 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1958 unsigned long nr_active(void)
1960 unsigned long i
, running
= 0, uninterruptible
= 0;
1962 for_each_online_cpu(i
) {
1963 running
+= cpu_rq(i
)->nr_running
;
1964 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1967 if (unlikely((long)uninterruptible
< 0))
1968 uninterruptible
= 0;
1970 return running
+ uninterruptible
;
1974 * Update rq->cpu_load[] statistics. This function is usually called every
1975 * scheduler tick (TICK_NSEC).
1977 static void update_cpu_load(struct rq
*this_rq
)
1979 u64 fair_delta64
, exec_delta64
, idle_delta64
, sample_interval64
, tmp64
;
1980 unsigned long total_load
= this_rq
->ls
.load
.weight
;
1981 unsigned long this_load
= total_load
;
1982 struct load_stat
*ls
= &this_rq
->ls
;
1985 this_rq
->nr_load_updates
++;
1986 if (unlikely(!(sysctl_sched_features
& SCHED_FEAT_PRECISE_CPU_LOAD
)))
1989 /* Update delta_fair/delta_exec fields first */
1990 update_curr_load(this_rq
);
1992 fair_delta64
= ls
->delta_fair
+ 1;
1995 exec_delta64
= ls
->delta_exec
+ 1;
1998 sample_interval64
= this_rq
->clock
- ls
->load_update_last
;
1999 ls
->load_update_last
= this_rq
->clock
;
2001 if ((s64
)sample_interval64
< (s64
)TICK_NSEC
)
2002 sample_interval64
= TICK_NSEC
;
2004 if (exec_delta64
> sample_interval64
)
2005 exec_delta64
= sample_interval64
;
2007 idle_delta64
= sample_interval64
- exec_delta64
;
2009 tmp64
= div64_64(SCHED_LOAD_SCALE
* exec_delta64
, fair_delta64
);
2010 tmp64
= div64_64(tmp64
* exec_delta64
, sample_interval64
);
2012 this_load
= (unsigned long)tmp64
;
2016 /* Update our load: */
2017 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2018 unsigned long old_load
, new_load
;
2020 /* scale is effectively 1 << i now, and >> i divides by scale */
2022 old_load
= this_rq
->cpu_load
[i
];
2023 new_load
= this_load
;
2025 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2032 * double_rq_lock - safely lock two runqueues
2034 * Note this does not disable interrupts like task_rq_lock,
2035 * you need to do so manually before calling.
2037 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2038 __acquires(rq1
->lock
)
2039 __acquires(rq2
->lock
)
2041 BUG_ON(!irqs_disabled());
2043 spin_lock(&rq1
->lock
);
2044 __acquire(rq2
->lock
); /* Fake it out ;) */
2047 spin_lock(&rq1
->lock
);
2048 spin_lock(&rq2
->lock
);
2050 spin_lock(&rq2
->lock
);
2051 spin_lock(&rq1
->lock
);
2054 update_rq_clock(rq1
);
2055 update_rq_clock(rq2
);
2059 * double_rq_unlock - safely unlock two runqueues
2061 * Note this does not restore interrupts like task_rq_unlock,
2062 * you need to do so manually after calling.
2064 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2065 __releases(rq1
->lock
)
2066 __releases(rq2
->lock
)
2068 spin_unlock(&rq1
->lock
);
2070 spin_unlock(&rq2
->lock
);
2072 __release(rq2
->lock
);
2076 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2078 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2079 __releases(this_rq
->lock
)
2080 __acquires(busiest
->lock
)
2081 __acquires(this_rq
->lock
)
2083 if (unlikely(!irqs_disabled())) {
2084 /* printk() doesn't work good under rq->lock */
2085 spin_unlock(&this_rq
->lock
);
2088 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2089 if (busiest
< this_rq
) {
2090 spin_unlock(&this_rq
->lock
);
2091 spin_lock(&busiest
->lock
);
2092 spin_lock(&this_rq
->lock
);
2094 spin_lock(&busiest
->lock
);
2099 * If dest_cpu is allowed for this process, migrate the task to it.
2100 * This is accomplished by forcing the cpu_allowed mask to only
2101 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2102 * the cpu_allowed mask is restored.
2104 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2106 struct migration_req req
;
2107 unsigned long flags
;
2110 rq
= task_rq_lock(p
, &flags
);
2111 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2112 || unlikely(cpu_is_offline(dest_cpu
)))
2115 /* force the process onto the specified CPU */
2116 if (migrate_task(p
, dest_cpu
, &req
)) {
2117 /* Need to wait for migration thread (might exit: take ref). */
2118 struct task_struct
*mt
= rq
->migration_thread
;
2120 get_task_struct(mt
);
2121 task_rq_unlock(rq
, &flags
);
2122 wake_up_process(mt
);
2123 put_task_struct(mt
);
2124 wait_for_completion(&req
.done
);
2129 task_rq_unlock(rq
, &flags
);
2133 * sched_exec - execve() is a valuable balancing opportunity, because at
2134 * this point the task has the smallest effective memory and cache footprint.
2136 void sched_exec(void)
2138 int new_cpu
, this_cpu
= get_cpu();
2139 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2141 if (new_cpu
!= this_cpu
)
2142 sched_migrate_task(current
, new_cpu
);
2146 * pull_task - move a task from a remote runqueue to the local runqueue.
2147 * Both runqueues must be locked.
2149 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2150 struct rq
*this_rq
, int this_cpu
)
2152 deactivate_task(src_rq
, p
, 0);
2153 set_task_cpu(p
, this_cpu
);
2154 activate_task(this_rq
, p
, 0);
2156 * Note that idle threads have a prio of MAX_PRIO, for this test
2157 * to be always true for them.
2159 check_preempt_curr(this_rq
, p
);
2163 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2166 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2167 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2171 * We do not migrate tasks that are:
2172 * 1) running (obviously), or
2173 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2174 * 3) are cache-hot on their current CPU.
2176 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2180 if (task_running(rq
, p
))
2184 * Aggressive migration if too many balance attempts have failed:
2186 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2192 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2193 unsigned long max_nr_move
, unsigned long max_load_move
,
2194 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2195 int *all_pinned
, unsigned long *load_moved
,
2196 int *this_best_prio
, struct rq_iterator
*iterator
)
2198 int pulled
= 0, pinned
= 0, skip_for_load
;
2199 struct task_struct
*p
;
2200 long rem_load_move
= max_load_move
;
2202 if (max_nr_move
== 0 || max_load_move
== 0)
2208 * Start the load-balancing iterator:
2210 p
= iterator
->start(iterator
->arg
);
2215 * To help distribute high priority tasks accross CPUs we don't
2216 * skip a task if it will be the highest priority task (i.e. smallest
2217 * prio value) on its new queue regardless of its load weight
2219 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2220 SCHED_LOAD_SCALE_FUZZ
;
2221 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2222 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2223 p
= iterator
->next(iterator
->arg
);
2227 pull_task(busiest
, p
, this_rq
, this_cpu
);
2229 rem_load_move
-= p
->se
.load
.weight
;
2232 * We only want to steal up to the prescribed number of tasks
2233 * and the prescribed amount of weighted load.
2235 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2236 if (p
->prio
< *this_best_prio
)
2237 *this_best_prio
= p
->prio
;
2238 p
= iterator
->next(iterator
->arg
);
2243 * Right now, this is the only place pull_task() is called,
2244 * so we can safely collect pull_task() stats here rather than
2245 * inside pull_task().
2247 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2250 *all_pinned
= pinned
;
2251 *load_moved
= max_load_move
- rem_load_move
;
2256 * move_tasks tries to move up to max_load_move weighted load from busiest to
2257 * this_rq, as part of a balancing operation within domain "sd".
2258 * Returns 1 if successful and 0 otherwise.
2260 * Called with both runqueues locked.
2262 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2263 unsigned long max_load_move
,
2264 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2267 struct sched_class
*class = sched_class_highest
;
2268 unsigned long total_load_moved
= 0;
2269 int this_best_prio
= this_rq
->curr
->prio
;
2273 class->load_balance(this_rq
, this_cpu
, busiest
,
2274 ULONG_MAX
, max_load_move
- total_load_moved
,
2275 sd
, idle
, all_pinned
, &this_best_prio
);
2276 class = class->next
;
2277 } while (class && max_load_move
> total_load_moved
);
2279 return total_load_moved
> 0;
2283 * move_one_task tries to move exactly one task from busiest to this_rq, as
2284 * part of active balancing operations within "domain".
2285 * Returns 1 if successful and 0 otherwise.
2287 * Called with both runqueues locked.
2289 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2290 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2292 struct sched_class
*class;
2293 int this_best_prio
= MAX_PRIO
;
2295 for (class = sched_class_highest
; class; class = class->next
)
2296 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2297 1, ULONG_MAX
, sd
, idle
, NULL
,
2305 * find_busiest_group finds and returns the busiest CPU group within the
2306 * domain. It calculates and returns the amount of weighted load which
2307 * should be moved to restore balance via the imbalance parameter.
2309 static struct sched_group
*
2310 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2311 unsigned long *imbalance
, enum cpu_idle_type idle
,
2312 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2314 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2315 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2316 unsigned long max_pull
;
2317 unsigned long busiest_load_per_task
, busiest_nr_running
;
2318 unsigned long this_load_per_task
, this_nr_running
;
2320 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2321 int power_savings_balance
= 1;
2322 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2323 unsigned long min_nr_running
= ULONG_MAX
;
2324 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2327 max_load
= this_load
= total_load
= total_pwr
= 0;
2328 busiest_load_per_task
= busiest_nr_running
= 0;
2329 this_load_per_task
= this_nr_running
= 0;
2330 if (idle
== CPU_NOT_IDLE
)
2331 load_idx
= sd
->busy_idx
;
2332 else if (idle
== CPU_NEWLY_IDLE
)
2333 load_idx
= sd
->newidle_idx
;
2335 load_idx
= sd
->idle_idx
;
2338 unsigned long load
, group_capacity
;
2341 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2342 unsigned long sum_nr_running
, sum_weighted_load
;
2344 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2347 balance_cpu
= first_cpu(group
->cpumask
);
2349 /* Tally up the load of all CPUs in the group */
2350 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2352 for_each_cpu_mask(i
, group
->cpumask
) {
2355 if (!cpu_isset(i
, *cpus
))
2360 if (*sd_idle
&& rq
->nr_running
)
2363 /* Bias balancing toward cpus of our domain */
2365 if (idle_cpu(i
) && !first_idle_cpu
) {
2370 load
= target_load(i
, load_idx
);
2372 load
= source_load(i
, load_idx
);
2375 sum_nr_running
+= rq
->nr_running
;
2376 sum_weighted_load
+= weighted_cpuload(i
);
2380 * First idle cpu or the first cpu(busiest) in this sched group
2381 * is eligible for doing load balancing at this and above
2382 * domains. In the newly idle case, we will allow all the cpu's
2383 * to do the newly idle load balance.
2385 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2386 balance_cpu
!= this_cpu
&& balance
) {
2391 total_load
+= avg_load
;
2392 total_pwr
+= group
->__cpu_power
;
2394 /* Adjust by relative CPU power of the group */
2395 avg_load
= sg_div_cpu_power(group
,
2396 avg_load
* SCHED_LOAD_SCALE
);
2398 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2401 this_load
= avg_load
;
2403 this_nr_running
= sum_nr_running
;
2404 this_load_per_task
= sum_weighted_load
;
2405 } else if (avg_load
> max_load
&&
2406 sum_nr_running
> group_capacity
) {
2407 max_load
= avg_load
;
2409 busiest_nr_running
= sum_nr_running
;
2410 busiest_load_per_task
= sum_weighted_load
;
2413 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2415 * Busy processors will not participate in power savings
2418 if (idle
== CPU_NOT_IDLE
||
2419 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2423 * If the local group is idle or completely loaded
2424 * no need to do power savings balance at this domain
2426 if (local_group
&& (this_nr_running
>= group_capacity
||
2428 power_savings_balance
= 0;
2431 * If a group is already running at full capacity or idle,
2432 * don't include that group in power savings calculations
2434 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2439 * Calculate the group which has the least non-idle load.
2440 * This is the group from where we need to pick up the load
2443 if ((sum_nr_running
< min_nr_running
) ||
2444 (sum_nr_running
== min_nr_running
&&
2445 first_cpu(group
->cpumask
) <
2446 first_cpu(group_min
->cpumask
))) {
2448 min_nr_running
= sum_nr_running
;
2449 min_load_per_task
= sum_weighted_load
/
2454 * Calculate the group which is almost near its
2455 * capacity but still has some space to pick up some load
2456 * from other group and save more power
2458 if (sum_nr_running
<= group_capacity
- 1) {
2459 if (sum_nr_running
> leader_nr_running
||
2460 (sum_nr_running
== leader_nr_running
&&
2461 first_cpu(group
->cpumask
) >
2462 first_cpu(group_leader
->cpumask
))) {
2463 group_leader
= group
;
2464 leader_nr_running
= sum_nr_running
;
2469 group
= group
->next
;
2470 } while (group
!= sd
->groups
);
2472 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2475 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2477 if (this_load
>= avg_load
||
2478 100*max_load
<= sd
->imbalance_pct
*this_load
)
2481 busiest_load_per_task
/= busiest_nr_running
;
2483 * We're trying to get all the cpus to the average_load, so we don't
2484 * want to push ourselves above the average load, nor do we wish to
2485 * reduce the max loaded cpu below the average load, as either of these
2486 * actions would just result in more rebalancing later, and ping-pong
2487 * tasks around. Thus we look for the minimum possible imbalance.
2488 * Negative imbalances (*we* are more loaded than anyone else) will
2489 * be counted as no imbalance for these purposes -- we can't fix that
2490 * by pulling tasks to us. Be careful of negative numbers as they'll
2491 * appear as very large values with unsigned longs.
2493 if (max_load
<= busiest_load_per_task
)
2497 * In the presence of smp nice balancing, certain scenarios can have
2498 * max load less than avg load(as we skip the groups at or below
2499 * its cpu_power, while calculating max_load..)
2501 if (max_load
< avg_load
) {
2503 goto small_imbalance
;
2506 /* Don't want to pull so many tasks that a group would go idle */
2507 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2509 /* How much load to actually move to equalise the imbalance */
2510 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2511 (avg_load
- this_load
) * this->__cpu_power
)
2515 * if *imbalance is less than the average load per runnable task
2516 * there is no gaurantee that any tasks will be moved so we'll have
2517 * a think about bumping its value to force at least one task to be
2520 if (*imbalance
+ SCHED_LOAD_SCALE_FUZZ
< busiest_load_per_task
/2) {
2521 unsigned long tmp
, pwr_now
, pwr_move
;
2525 pwr_move
= pwr_now
= 0;
2527 if (this_nr_running
) {
2528 this_load_per_task
/= this_nr_running
;
2529 if (busiest_load_per_task
> this_load_per_task
)
2532 this_load_per_task
= SCHED_LOAD_SCALE
;
2534 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2535 busiest_load_per_task
* imbn
) {
2536 *imbalance
= busiest_load_per_task
;
2541 * OK, we don't have enough imbalance to justify moving tasks,
2542 * however we may be able to increase total CPU power used by
2546 pwr_now
+= busiest
->__cpu_power
*
2547 min(busiest_load_per_task
, max_load
);
2548 pwr_now
+= this->__cpu_power
*
2549 min(this_load_per_task
, this_load
);
2550 pwr_now
/= SCHED_LOAD_SCALE
;
2552 /* Amount of load we'd subtract */
2553 tmp
= sg_div_cpu_power(busiest
,
2554 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2556 pwr_move
+= busiest
->__cpu_power
*
2557 min(busiest_load_per_task
, max_load
- tmp
);
2559 /* Amount of load we'd add */
2560 if (max_load
* busiest
->__cpu_power
<
2561 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2562 tmp
= sg_div_cpu_power(this,
2563 max_load
* busiest
->__cpu_power
);
2565 tmp
= sg_div_cpu_power(this,
2566 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2567 pwr_move
+= this->__cpu_power
*
2568 min(this_load_per_task
, this_load
+ tmp
);
2569 pwr_move
/= SCHED_LOAD_SCALE
;
2571 /* Move if we gain throughput */
2572 if (pwr_move
<= pwr_now
)
2575 *imbalance
= busiest_load_per_task
;
2581 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2582 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2585 if (this == group_leader
&& group_leader
!= group_min
) {
2586 *imbalance
= min_load_per_task
;
2596 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2599 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2600 unsigned long imbalance
, cpumask_t
*cpus
)
2602 struct rq
*busiest
= NULL
, *rq
;
2603 unsigned long max_load
= 0;
2606 for_each_cpu_mask(i
, group
->cpumask
) {
2609 if (!cpu_isset(i
, *cpus
))
2613 wl
= weighted_cpuload(i
);
2615 if (rq
->nr_running
== 1 && wl
> imbalance
)
2618 if (wl
> max_load
) {
2628 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2629 * so long as it is large enough.
2631 #define MAX_PINNED_INTERVAL 512
2634 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2635 * tasks if there is an imbalance.
2637 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2638 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2641 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2642 struct sched_group
*group
;
2643 unsigned long imbalance
;
2645 cpumask_t cpus
= CPU_MASK_ALL
;
2646 unsigned long flags
;
2649 * When power savings policy is enabled for the parent domain, idle
2650 * sibling can pick up load irrespective of busy siblings. In this case,
2651 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2652 * portraying it as CPU_NOT_IDLE.
2654 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2655 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2658 schedstat_inc(sd
, lb_cnt
[idle
]);
2661 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2668 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2672 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2674 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2678 BUG_ON(busiest
== this_rq
);
2680 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2683 if (busiest
->nr_running
> 1) {
2685 * Attempt to move tasks. If find_busiest_group has found
2686 * an imbalance but busiest->nr_running <= 1, the group is
2687 * still unbalanced. ld_moved simply stays zero, so it is
2688 * correctly treated as an imbalance.
2690 local_irq_save(flags
);
2691 double_rq_lock(this_rq
, busiest
);
2692 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2693 imbalance
, sd
, idle
, &all_pinned
);
2694 double_rq_unlock(this_rq
, busiest
);
2695 local_irq_restore(flags
);
2698 * some other cpu did the load balance for us.
2700 if (ld_moved
&& this_cpu
!= smp_processor_id())
2701 resched_cpu(this_cpu
);
2703 /* All tasks on this runqueue were pinned by CPU affinity */
2704 if (unlikely(all_pinned
)) {
2705 cpu_clear(cpu_of(busiest
), cpus
);
2706 if (!cpus_empty(cpus
))
2713 schedstat_inc(sd
, lb_failed
[idle
]);
2714 sd
->nr_balance_failed
++;
2716 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2718 spin_lock_irqsave(&busiest
->lock
, flags
);
2720 /* don't kick the migration_thread, if the curr
2721 * task on busiest cpu can't be moved to this_cpu
2723 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2724 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2726 goto out_one_pinned
;
2729 if (!busiest
->active_balance
) {
2730 busiest
->active_balance
= 1;
2731 busiest
->push_cpu
= this_cpu
;
2734 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2736 wake_up_process(busiest
->migration_thread
);
2739 * We've kicked active balancing, reset the failure
2742 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2745 sd
->nr_balance_failed
= 0;
2747 if (likely(!active_balance
)) {
2748 /* We were unbalanced, so reset the balancing interval */
2749 sd
->balance_interval
= sd
->min_interval
;
2752 * If we've begun active balancing, start to back off. This
2753 * case may not be covered by the all_pinned logic if there
2754 * is only 1 task on the busy runqueue (because we don't call
2757 if (sd
->balance_interval
< sd
->max_interval
)
2758 sd
->balance_interval
*= 2;
2761 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2762 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2767 schedstat_inc(sd
, lb_balanced
[idle
]);
2769 sd
->nr_balance_failed
= 0;
2772 /* tune up the balancing interval */
2773 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2774 (sd
->balance_interval
< sd
->max_interval
))
2775 sd
->balance_interval
*= 2;
2777 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2778 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2784 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2785 * tasks if there is an imbalance.
2787 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2788 * this_rq is locked.
2791 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2793 struct sched_group
*group
;
2794 struct rq
*busiest
= NULL
;
2795 unsigned long imbalance
;
2799 cpumask_t cpus
= CPU_MASK_ALL
;
2802 * When power savings policy is enabled for the parent domain, idle
2803 * sibling can pick up load irrespective of busy siblings. In this case,
2804 * let the state of idle sibling percolate up as IDLE, instead of
2805 * portraying it as CPU_NOT_IDLE.
2807 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2808 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2811 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2813 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2814 &sd_idle
, &cpus
, NULL
);
2816 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2820 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2823 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2827 BUG_ON(busiest
== this_rq
);
2829 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2832 if (busiest
->nr_running
> 1) {
2833 /* Attempt to move tasks */
2834 double_lock_balance(this_rq
, busiest
);
2835 /* this_rq->clock is already updated */
2836 update_rq_clock(busiest
);
2837 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2838 imbalance
, sd
, CPU_NEWLY_IDLE
,
2840 spin_unlock(&busiest
->lock
);
2842 if (unlikely(all_pinned
)) {
2843 cpu_clear(cpu_of(busiest
), cpus
);
2844 if (!cpus_empty(cpus
))
2850 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2851 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2852 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2855 sd
->nr_balance_failed
= 0;
2860 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2861 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2862 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2864 sd
->nr_balance_failed
= 0;
2870 * idle_balance is called by schedule() if this_cpu is about to become
2871 * idle. Attempts to pull tasks from other CPUs.
2873 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2875 struct sched_domain
*sd
;
2876 int pulled_task
= -1;
2877 unsigned long next_balance
= jiffies
+ HZ
;
2879 for_each_domain(this_cpu
, sd
) {
2880 unsigned long interval
;
2882 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2885 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2886 /* If we've pulled tasks over stop searching: */
2887 pulled_task
= load_balance_newidle(this_cpu
,
2890 interval
= msecs_to_jiffies(sd
->balance_interval
);
2891 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2892 next_balance
= sd
->last_balance
+ interval
;
2896 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2898 * We are going idle. next_balance may be set based on
2899 * a busy processor. So reset next_balance.
2901 this_rq
->next_balance
= next_balance
;
2906 * active_load_balance is run by migration threads. It pushes running tasks
2907 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2908 * running on each physical CPU where possible, and avoids physical /
2909 * logical imbalances.
2911 * Called with busiest_rq locked.
2913 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2915 int target_cpu
= busiest_rq
->push_cpu
;
2916 struct sched_domain
*sd
;
2917 struct rq
*target_rq
;
2919 /* Is there any task to move? */
2920 if (busiest_rq
->nr_running
<= 1)
2923 target_rq
= cpu_rq(target_cpu
);
2926 * This condition is "impossible", if it occurs
2927 * we need to fix it. Originally reported by
2928 * Bjorn Helgaas on a 128-cpu setup.
2930 BUG_ON(busiest_rq
== target_rq
);
2932 /* move a task from busiest_rq to target_rq */
2933 double_lock_balance(busiest_rq
, target_rq
);
2934 update_rq_clock(busiest_rq
);
2935 update_rq_clock(target_rq
);
2937 /* Search for an sd spanning us and the target CPU. */
2938 for_each_domain(target_cpu
, sd
) {
2939 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2940 cpu_isset(busiest_cpu
, sd
->span
))
2945 schedstat_inc(sd
, alb_cnt
);
2947 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2949 schedstat_inc(sd
, alb_pushed
);
2951 schedstat_inc(sd
, alb_failed
);
2953 spin_unlock(&target_rq
->lock
);
2958 atomic_t load_balancer
;
2960 } nohz ____cacheline_aligned
= {
2961 .load_balancer
= ATOMIC_INIT(-1),
2962 .cpu_mask
= CPU_MASK_NONE
,
2966 * This routine will try to nominate the ilb (idle load balancing)
2967 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2968 * load balancing on behalf of all those cpus. If all the cpus in the system
2969 * go into this tickless mode, then there will be no ilb owner (as there is
2970 * no need for one) and all the cpus will sleep till the next wakeup event
2973 * For the ilb owner, tick is not stopped. And this tick will be used
2974 * for idle load balancing. ilb owner will still be part of
2977 * While stopping the tick, this cpu will become the ilb owner if there
2978 * is no other owner. And will be the owner till that cpu becomes busy
2979 * or if all cpus in the system stop their ticks at which point
2980 * there is no need for ilb owner.
2982 * When the ilb owner becomes busy, it nominates another owner, during the
2983 * next busy scheduler_tick()
2985 int select_nohz_load_balancer(int stop_tick
)
2987 int cpu
= smp_processor_id();
2990 cpu_set(cpu
, nohz
.cpu_mask
);
2991 cpu_rq(cpu
)->in_nohz_recently
= 1;
2994 * If we are going offline and still the leader, give up!
2996 if (cpu_is_offline(cpu
) &&
2997 atomic_read(&nohz
.load_balancer
) == cpu
) {
2998 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3003 /* time for ilb owner also to sleep */
3004 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3005 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3006 atomic_set(&nohz
.load_balancer
, -1);
3010 if (atomic_read(&nohz
.load_balancer
) == -1) {
3011 /* make me the ilb owner */
3012 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3014 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3017 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3020 cpu_clear(cpu
, nohz
.cpu_mask
);
3022 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3023 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3030 static DEFINE_SPINLOCK(balancing
);
3033 * It checks each scheduling domain to see if it is due to be balanced,
3034 * and initiates a balancing operation if so.
3036 * Balancing parameters are set up in arch_init_sched_domains.
3038 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3041 struct rq
*rq
= cpu_rq(cpu
);
3042 unsigned long interval
;
3043 struct sched_domain
*sd
;
3044 /* Earliest time when we have to do rebalance again */
3045 unsigned long next_balance
= jiffies
+ 60*HZ
;
3047 for_each_domain(cpu
, sd
) {
3048 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3051 interval
= sd
->balance_interval
;
3052 if (idle
!= CPU_IDLE
)
3053 interval
*= sd
->busy_factor
;
3055 /* scale ms to jiffies */
3056 interval
= msecs_to_jiffies(interval
);
3057 if (unlikely(!interval
))
3059 if (interval
> HZ
*NR_CPUS
/10)
3060 interval
= HZ
*NR_CPUS
/10;
3063 if (sd
->flags
& SD_SERIALIZE
) {
3064 if (!spin_trylock(&balancing
))
3068 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3069 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3071 * We've pulled tasks over so either we're no
3072 * longer idle, or one of our SMT siblings is
3075 idle
= CPU_NOT_IDLE
;
3077 sd
->last_balance
= jiffies
;
3079 if (sd
->flags
& SD_SERIALIZE
)
3080 spin_unlock(&balancing
);
3082 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3083 next_balance
= sd
->last_balance
+ interval
;
3086 * Stop the load balance at this level. There is another
3087 * CPU in our sched group which is doing load balancing more
3093 rq
->next_balance
= next_balance
;
3097 * run_rebalance_domains is triggered when needed from the scheduler tick.
3098 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3099 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3101 static void run_rebalance_domains(struct softirq_action
*h
)
3103 int this_cpu
= smp_processor_id();
3104 struct rq
*this_rq
= cpu_rq(this_cpu
);
3105 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3106 CPU_IDLE
: CPU_NOT_IDLE
;
3108 rebalance_domains(this_cpu
, idle
);
3112 * If this cpu is the owner for idle load balancing, then do the
3113 * balancing on behalf of the other idle cpus whose ticks are
3116 if (this_rq
->idle_at_tick
&&
3117 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3118 cpumask_t cpus
= nohz
.cpu_mask
;
3122 cpu_clear(this_cpu
, cpus
);
3123 for_each_cpu_mask(balance_cpu
, cpus
) {
3125 * If this cpu gets work to do, stop the load balancing
3126 * work being done for other cpus. Next load
3127 * balancing owner will pick it up.
3132 rebalance_domains(balance_cpu
, CPU_IDLE
);
3134 rq
= cpu_rq(balance_cpu
);
3135 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3136 this_rq
->next_balance
= rq
->next_balance
;
3143 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3145 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3146 * idle load balancing owner or decide to stop the periodic load balancing,
3147 * if the whole system is idle.
3149 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3153 * If we were in the nohz mode recently and busy at the current
3154 * scheduler tick, then check if we need to nominate new idle
3157 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3158 rq
->in_nohz_recently
= 0;
3160 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3161 cpu_clear(cpu
, nohz
.cpu_mask
);
3162 atomic_set(&nohz
.load_balancer
, -1);
3165 if (atomic_read(&nohz
.load_balancer
) == -1) {
3167 * simple selection for now: Nominate the
3168 * first cpu in the nohz list to be the next
3171 * TBD: Traverse the sched domains and nominate
3172 * the nearest cpu in the nohz.cpu_mask.
3174 int ilb
= first_cpu(nohz
.cpu_mask
);
3182 * If this cpu is idle and doing idle load balancing for all the
3183 * cpus with ticks stopped, is it time for that to stop?
3185 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3186 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3192 * If this cpu is idle and the idle load balancing is done by
3193 * someone else, then no need raise the SCHED_SOFTIRQ
3195 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3196 cpu_isset(cpu
, nohz
.cpu_mask
))
3199 if (time_after_eq(jiffies
, rq
->next_balance
))
3200 raise_softirq(SCHED_SOFTIRQ
);
3203 #else /* CONFIG_SMP */
3206 * on UP we do not need to balance between CPUs:
3208 static inline void idle_balance(int cpu
, struct rq
*rq
)
3212 /* Avoid "used but not defined" warning on UP */
3213 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3214 unsigned long max_nr_move
, unsigned long max_load_move
,
3215 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3216 int *all_pinned
, unsigned long *load_moved
,
3217 int *this_best_prio
, struct rq_iterator
*iterator
)
3226 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3228 EXPORT_PER_CPU_SYMBOL(kstat
);
3231 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3232 * that have not yet been banked in case the task is currently running.
3234 unsigned long long task_sched_runtime(struct task_struct
*p
)
3236 unsigned long flags
;
3240 rq
= task_rq_lock(p
, &flags
);
3241 ns
= p
->se
.sum_exec_runtime
;
3242 if (rq
->curr
== p
) {
3243 update_rq_clock(rq
);
3244 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3245 if ((s64
)delta_exec
> 0)
3248 task_rq_unlock(rq
, &flags
);
3254 * Account user cpu time to a process.
3255 * @p: the process that the cpu time gets accounted to
3256 * @hardirq_offset: the offset to subtract from hardirq_count()
3257 * @cputime: the cpu time spent in user space since the last update
3259 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3261 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3264 p
->utime
= cputime_add(p
->utime
, cputime
);
3266 /* Add user time to cpustat. */
3267 tmp
= cputime_to_cputime64(cputime
);
3268 if (TASK_NICE(p
) > 0)
3269 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3271 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3275 * Account system cpu time to a process.
3276 * @p: the process that the cpu time gets accounted to
3277 * @hardirq_offset: the offset to subtract from hardirq_count()
3278 * @cputime: the cpu time spent in kernel space since the last update
3280 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3283 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3284 struct rq
*rq
= this_rq();
3287 p
->stime
= cputime_add(p
->stime
, cputime
);
3289 /* Add system time to cpustat. */
3290 tmp
= cputime_to_cputime64(cputime
);
3291 if (hardirq_count() - hardirq_offset
)
3292 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3293 else if (softirq_count())
3294 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3295 else if (p
!= rq
->idle
)
3296 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3297 else if (atomic_read(&rq
->nr_iowait
) > 0)
3298 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3300 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3301 /* Account for system time used */
3302 acct_update_integrals(p
);
3306 * Account for involuntary wait time.
3307 * @p: the process from which the cpu time has been stolen
3308 * @steal: the cpu time spent in involuntary wait
3310 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3312 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3313 cputime64_t tmp
= cputime_to_cputime64(steal
);
3314 struct rq
*rq
= this_rq();
3316 if (p
== rq
->idle
) {
3317 p
->stime
= cputime_add(p
->stime
, steal
);
3318 if (atomic_read(&rq
->nr_iowait
) > 0)
3319 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3321 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3323 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3327 * This function gets called by the timer code, with HZ frequency.
3328 * We call it with interrupts disabled.
3330 * It also gets called by the fork code, when changing the parent's
3333 void scheduler_tick(void)
3335 int cpu
= smp_processor_id();
3336 struct rq
*rq
= cpu_rq(cpu
);
3337 struct task_struct
*curr
= rq
->curr
;
3338 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3340 spin_lock(&rq
->lock
);
3341 __update_rq_clock(rq
);
3343 * Let rq->clock advance by at least TICK_NSEC:
3345 if (unlikely(rq
->clock
< next_tick
))
3346 rq
->clock
= next_tick
;
3347 rq
->tick_timestamp
= rq
->clock
;
3348 update_cpu_load(rq
);
3349 if (curr
!= rq
->idle
) /* FIXME: needed? */
3350 curr
->sched_class
->task_tick(rq
, curr
);
3351 spin_unlock(&rq
->lock
);
3354 rq
->idle_at_tick
= idle_cpu(cpu
);
3355 trigger_load_balance(rq
, cpu
);
3359 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3361 void fastcall
add_preempt_count(int val
)
3366 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3368 preempt_count() += val
;
3370 * Spinlock count overflowing soon?
3372 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3375 EXPORT_SYMBOL(add_preempt_count
);
3377 void fastcall
sub_preempt_count(int val
)
3382 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3385 * Is the spinlock portion underflowing?
3387 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3388 !(preempt_count() & PREEMPT_MASK
)))
3391 preempt_count() -= val
;
3393 EXPORT_SYMBOL(sub_preempt_count
);
3398 * Print scheduling while atomic bug:
3400 static noinline
void __schedule_bug(struct task_struct
*prev
)
3402 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3403 prev
->comm
, preempt_count(), prev
->pid
);
3404 debug_show_held_locks(prev
);
3405 if (irqs_disabled())
3406 print_irqtrace_events(prev
);
3411 * Various schedule()-time debugging checks and statistics:
3413 static inline void schedule_debug(struct task_struct
*prev
)
3416 * Test if we are atomic. Since do_exit() needs to call into
3417 * schedule() atomically, we ignore that path for now.
3418 * Otherwise, whine if we are scheduling when we should not be.
3420 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3421 __schedule_bug(prev
);
3423 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3425 schedstat_inc(this_rq(), sched_cnt
);
3429 * Pick up the highest-prio task:
3431 static inline struct task_struct
*
3432 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3434 struct sched_class
*class;
3435 struct task_struct
*p
;
3438 * Optimization: we know that if all tasks are in
3439 * the fair class we can call that function directly:
3441 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3442 p
= fair_sched_class
.pick_next_task(rq
);
3447 class = sched_class_highest
;
3449 p
= class->pick_next_task(rq
);
3453 * Will never be NULL as the idle class always
3454 * returns a non-NULL p:
3456 class = class->next
;
3461 * schedule() is the main scheduler function.
3463 asmlinkage
void __sched
schedule(void)
3465 struct task_struct
*prev
, *next
;
3472 cpu
= smp_processor_id();
3476 switch_count
= &prev
->nivcsw
;
3478 release_kernel_lock(prev
);
3479 need_resched_nonpreemptible
:
3481 schedule_debug(prev
);
3483 spin_lock_irq(&rq
->lock
);
3484 clear_tsk_need_resched(prev
);
3485 __update_rq_clock(rq
);
3487 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3488 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3489 unlikely(signal_pending(prev
)))) {
3490 prev
->state
= TASK_RUNNING
;
3492 deactivate_task(rq
, prev
, 1);
3494 switch_count
= &prev
->nvcsw
;
3497 if (unlikely(!rq
->nr_running
))
3498 idle_balance(cpu
, rq
);
3500 prev
->sched_class
->put_prev_task(rq
, prev
);
3501 next
= pick_next_task(rq
, prev
);
3503 sched_info_switch(prev
, next
);
3505 if (likely(prev
!= next
)) {
3510 context_switch(rq
, prev
, next
); /* unlocks the rq */
3512 spin_unlock_irq(&rq
->lock
);
3514 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3515 cpu
= smp_processor_id();
3517 goto need_resched_nonpreemptible
;
3519 preempt_enable_no_resched();
3520 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3523 EXPORT_SYMBOL(schedule
);
3525 #ifdef CONFIG_PREEMPT
3527 * this is the entry point to schedule() from in-kernel preemption
3528 * off of preempt_enable. Kernel preemptions off return from interrupt
3529 * occur there and call schedule directly.
3531 asmlinkage
void __sched
preempt_schedule(void)
3533 struct thread_info
*ti
= current_thread_info();
3534 #ifdef CONFIG_PREEMPT_BKL
3535 struct task_struct
*task
= current
;
3536 int saved_lock_depth
;
3539 * If there is a non-zero preempt_count or interrupts are disabled,
3540 * we do not want to preempt the current task. Just return..
3542 if (likely(ti
->preempt_count
|| irqs_disabled()))
3546 add_preempt_count(PREEMPT_ACTIVE
);
3548 * We keep the big kernel semaphore locked, but we
3549 * clear ->lock_depth so that schedule() doesnt
3550 * auto-release the semaphore:
3552 #ifdef CONFIG_PREEMPT_BKL
3553 saved_lock_depth
= task
->lock_depth
;
3554 task
->lock_depth
= -1;
3557 #ifdef CONFIG_PREEMPT_BKL
3558 task
->lock_depth
= saved_lock_depth
;
3560 sub_preempt_count(PREEMPT_ACTIVE
);
3562 /* we could miss a preemption opportunity between schedule and now */
3564 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3567 EXPORT_SYMBOL(preempt_schedule
);
3570 * this is the entry point to schedule() from kernel preemption
3571 * off of irq context.
3572 * Note, that this is called and return with irqs disabled. This will
3573 * protect us against recursive calling from irq.
3575 asmlinkage
void __sched
preempt_schedule_irq(void)
3577 struct thread_info
*ti
= current_thread_info();
3578 #ifdef CONFIG_PREEMPT_BKL
3579 struct task_struct
*task
= current
;
3580 int saved_lock_depth
;
3582 /* Catch callers which need to be fixed */
3583 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3586 add_preempt_count(PREEMPT_ACTIVE
);
3588 * We keep the big kernel semaphore locked, but we
3589 * clear ->lock_depth so that schedule() doesnt
3590 * auto-release the semaphore:
3592 #ifdef CONFIG_PREEMPT_BKL
3593 saved_lock_depth
= task
->lock_depth
;
3594 task
->lock_depth
= -1;
3598 local_irq_disable();
3599 #ifdef CONFIG_PREEMPT_BKL
3600 task
->lock_depth
= saved_lock_depth
;
3602 sub_preempt_count(PREEMPT_ACTIVE
);
3604 /* we could miss a preemption opportunity between schedule and now */
3606 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3610 #endif /* CONFIG_PREEMPT */
3612 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3615 return try_to_wake_up(curr
->private, mode
, sync
);
3617 EXPORT_SYMBOL(default_wake_function
);
3620 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3621 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3622 * number) then we wake all the non-exclusive tasks and one exclusive task.
3624 * There are circumstances in which we can try to wake a task which has already
3625 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3626 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3628 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3629 int nr_exclusive
, int sync
, void *key
)
3631 struct list_head
*tmp
, *next
;
3633 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3634 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, 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
;
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
;
3953 dequeue_task(rq
, p
, 0);
3956 p
->sched_class
= &rt_sched_class
;
3958 p
->sched_class
= &fair_sched_class
;
3963 enqueue_task(rq
, p
, 0);
3965 * Reschedule if we are currently running on this runqueue and
3966 * our priority decreased, or if we are not currently running on
3967 * this runqueue and our priority is higher than the current's
3969 if (task_running(rq
, p
)) {
3970 if (p
->prio
> oldprio
)
3971 resched_task(rq
->curr
);
3973 check_preempt_curr(rq
, p
);
3976 task_rq_unlock(rq
, &flags
);
3981 void set_user_nice(struct task_struct
*p
, long nice
)
3983 int old_prio
, delta
, on_rq
;
3984 unsigned long flags
;
3987 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3990 * We have to be careful, if called from sys_setpriority(),
3991 * the task might be in the middle of scheduling on another CPU.
3993 rq
= task_rq_lock(p
, &flags
);
3994 update_rq_clock(rq
);
3996 * The RT priorities are set via sched_setscheduler(), but we still
3997 * allow the 'normal' nice value to be set - but as expected
3998 * it wont have any effect on scheduling until the task is
3999 * SCHED_FIFO/SCHED_RR:
4001 if (task_has_rt_policy(p
)) {
4002 p
->static_prio
= NICE_TO_PRIO(nice
);
4005 on_rq
= p
->se
.on_rq
;
4007 dequeue_task(rq
, p
, 0);
4011 p
->static_prio
= NICE_TO_PRIO(nice
);
4014 p
->prio
= effective_prio(p
);
4015 delta
= p
->prio
- old_prio
;
4018 enqueue_task(rq
, p
, 0);
4021 * If the task increased its priority or is running and
4022 * lowered its priority, then reschedule its CPU:
4024 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4025 resched_task(rq
->curr
);
4028 task_rq_unlock(rq
, &flags
);
4030 EXPORT_SYMBOL(set_user_nice
);
4033 * can_nice - check if a task can reduce its nice value
4037 int can_nice(const struct task_struct
*p
, const int nice
)
4039 /* convert nice value [19,-20] to rlimit style value [1,40] */
4040 int nice_rlim
= 20 - nice
;
4042 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4043 capable(CAP_SYS_NICE
));
4046 #ifdef __ARCH_WANT_SYS_NICE
4049 * sys_nice - change the priority of the current process.
4050 * @increment: priority increment
4052 * sys_setpriority is a more generic, but much slower function that
4053 * does similar things.
4055 asmlinkage
long sys_nice(int increment
)
4060 * Setpriority might change our priority at the same moment.
4061 * We don't have to worry. Conceptually one call occurs first
4062 * and we have a single winner.
4064 if (increment
< -40)
4069 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4075 if (increment
< 0 && !can_nice(current
, nice
))
4078 retval
= security_task_setnice(current
, nice
);
4082 set_user_nice(current
, nice
);
4089 * task_prio - return the priority value of a given task.
4090 * @p: the task in question.
4092 * This is the priority value as seen by users in /proc.
4093 * RT tasks are offset by -200. Normal tasks are centered
4094 * around 0, value goes from -16 to +15.
4096 int task_prio(const struct task_struct
*p
)
4098 return p
->prio
- MAX_RT_PRIO
;
4102 * task_nice - return the nice value of a given task.
4103 * @p: the task in question.
4105 int task_nice(const struct task_struct
*p
)
4107 return TASK_NICE(p
);
4109 EXPORT_SYMBOL_GPL(task_nice
);
4112 * idle_cpu - is a given cpu idle currently?
4113 * @cpu: the processor in question.
4115 int idle_cpu(int cpu
)
4117 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4121 * idle_task - return the idle task for a given cpu.
4122 * @cpu: the processor in question.
4124 struct task_struct
*idle_task(int cpu
)
4126 return cpu_rq(cpu
)->idle
;
4130 * find_process_by_pid - find a process with a matching PID value.
4131 * @pid: the pid in question.
4133 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4135 return pid
? find_task_by_pid(pid
) : current
;
4138 /* Actually do priority change: must hold rq lock. */
4140 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4142 BUG_ON(p
->se
.on_rq
);
4145 switch (p
->policy
) {
4149 p
->sched_class
= &fair_sched_class
;
4153 p
->sched_class
= &rt_sched_class
;
4157 p
->rt_priority
= prio
;
4158 p
->normal_prio
= normal_prio(p
);
4159 /* we are holding p->pi_lock already */
4160 p
->prio
= rt_mutex_getprio(p
);
4165 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4166 * @p: the task in question.
4167 * @policy: new policy.
4168 * @param: structure containing the new RT priority.
4170 * NOTE that the task may be already dead.
4172 int sched_setscheduler(struct task_struct
*p
, int policy
,
4173 struct sched_param
*param
)
4175 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4176 unsigned long flags
;
4179 /* may grab non-irq protected spin_locks */
4180 BUG_ON(in_interrupt());
4182 /* double check policy once rq lock held */
4184 policy
= oldpolicy
= p
->policy
;
4185 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4186 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4187 policy
!= SCHED_IDLE
)
4190 * Valid priorities for SCHED_FIFO and SCHED_RR are
4191 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4192 * SCHED_BATCH and SCHED_IDLE is 0.
4194 if (param
->sched_priority
< 0 ||
4195 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4196 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4198 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4202 * Allow unprivileged RT tasks to decrease priority:
4204 if (!capable(CAP_SYS_NICE
)) {
4205 if (rt_policy(policy
)) {
4206 unsigned long rlim_rtprio
;
4208 if (!lock_task_sighand(p
, &flags
))
4210 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4211 unlock_task_sighand(p
, &flags
);
4213 /* can't set/change the rt policy */
4214 if (policy
!= p
->policy
&& !rlim_rtprio
)
4217 /* can't increase priority */
4218 if (param
->sched_priority
> p
->rt_priority
&&
4219 param
->sched_priority
> rlim_rtprio
)
4223 * Like positive nice levels, dont allow tasks to
4224 * move out of SCHED_IDLE either:
4226 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4229 /* can't change other user's priorities */
4230 if ((current
->euid
!= p
->euid
) &&
4231 (current
->euid
!= p
->uid
))
4235 retval
= security_task_setscheduler(p
, policy
, param
);
4239 * make sure no PI-waiters arrive (or leave) while we are
4240 * changing the priority of the task:
4242 spin_lock_irqsave(&p
->pi_lock
, flags
);
4244 * To be able to change p->policy safely, the apropriate
4245 * runqueue lock must be held.
4247 rq
= __task_rq_lock(p
);
4248 /* recheck policy now with rq lock held */
4249 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4250 policy
= oldpolicy
= -1;
4251 __task_rq_unlock(rq
);
4252 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4255 update_rq_clock(rq
);
4256 on_rq
= p
->se
.on_rq
;
4258 deactivate_task(rq
, p
, 0);
4260 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4262 activate_task(rq
, p
, 0);
4264 * Reschedule if we are currently running on this runqueue and
4265 * our priority decreased, or if we are not currently running on
4266 * this runqueue and our priority is higher than the current's
4268 if (task_running(rq
, p
)) {
4269 if (p
->prio
> oldprio
)
4270 resched_task(rq
->curr
);
4272 check_preempt_curr(rq
, p
);
4275 __task_rq_unlock(rq
);
4276 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4278 rt_mutex_adjust_pi(p
);
4282 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4285 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4287 struct sched_param lparam
;
4288 struct task_struct
*p
;
4291 if (!param
|| pid
< 0)
4293 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4298 p
= find_process_by_pid(pid
);
4300 retval
= sched_setscheduler(p
, policy
, &lparam
);
4307 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4308 * @pid: the pid in question.
4309 * @policy: new policy.
4310 * @param: structure containing the new RT priority.
4312 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4313 struct sched_param __user
*param
)
4315 /* negative values for policy are not valid */
4319 return do_sched_setscheduler(pid
, policy
, param
);
4323 * sys_sched_setparam - set/change the RT priority of a thread
4324 * @pid: the pid in question.
4325 * @param: structure containing the new RT priority.
4327 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4329 return do_sched_setscheduler(pid
, -1, param
);
4333 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4334 * @pid: the pid in question.
4336 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4338 struct task_struct
*p
;
4339 int retval
= -EINVAL
;
4345 read_lock(&tasklist_lock
);
4346 p
= find_process_by_pid(pid
);
4348 retval
= security_task_getscheduler(p
);
4352 read_unlock(&tasklist_lock
);
4359 * sys_sched_getscheduler - get the RT priority of a thread
4360 * @pid: the pid in question.
4361 * @param: structure containing the RT priority.
4363 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4365 struct sched_param lp
;
4366 struct task_struct
*p
;
4367 int retval
= -EINVAL
;
4369 if (!param
|| pid
< 0)
4372 read_lock(&tasklist_lock
);
4373 p
= find_process_by_pid(pid
);
4378 retval
= security_task_getscheduler(p
);
4382 lp
.sched_priority
= p
->rt_priority
;
4383 read_unlock(&tasklist_lock
);
4386 * This one might sleep, we cannot do it with a spinlock held ...
4388 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4394 read_unlock(&tasklist_lock
);
4398 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4400 cpumask_t cpus_allowed
;
4401 struct task_struct
*p
;
4404 mutex_lock(&sched_hotcpu_mutex
);
4405 read_lock(&tasklist_lock
);
4407 p
= find_process_by_pid(pid
);
4409 read_unlock(&tasklist_lock
);
4410 mutex_unlock(&sched_hotcpu_mutex
);
4415 * It is not safe to call set_cpus_allowed with the
4416 * tasklist_lock held. We will bump the task_struct's
4417 * usage count and then drop tasklist_lock.
4420 read_unlock(&tasklist_lock
);
4423 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4424 !capable(CAP_SYS_NICE
))
4427 retval
= security_task_setscheduler(p
, 0, NULL
);
4431 cpus_allowed
= cpuset_cpus_allowed(p
);
4432 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4433 retval
= set_cpus_allowed(p
, new_mask
);
4437 mutex_unlock(&sched_hotcpu_mutex
);
4441 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4442 cpumask_t
*new_mask
)
4444 if (len
< sizeof(cpumask_t
)) {
4445 memset(new_mask
, 0, sizeof(cpumask_t
));
4446 } else if (len
> sizeof(cpumask_t
)) {
4447 len
= sizeof(cpumask_t
);
4449 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4453 * sys_sched_setaffinity - set the cpu affinity of a process
4454 * @pid: pid of the process
4455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4456 * @user_mask_ptr: user-space pointer to the new cpu mask
4458 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4459 unsigned long __user
*user_mask_ptr
)
4464 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4468 return sched_setaffinity(pid
, new_mask
);
4472 * Represents all cpu's present in the system
4473 * In systems capable of hotplug, this map could dynamically grow
4474 * as new cpu's are detected in the system via any platform specific
4475 * method, such as ACPI for e.g.
4478 cpumask_t cpu_present_map __read_mostly
;
4479 EXPORT_SYMBOL(cpu_present_map
);
4482 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4483 EXPORT_SYMBOL(cpu_online_map
);
4485 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4486 EXPORT_SYMBOL(cpu_possible_map
);
4489 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4491 struct task_struct
*p
;
4494 mutex_lock(&sched_hotcpu_mutex
);
4495 read_lock(&tasklist_lock
);
4498 p
= find_process_by_pid(pid
);
4502 retval
= security_task_getscheduler(p
);
4506 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4509 read_unlock(&tasklist_lock
);
4510 mutex_unlock(&sched_hotcpu_mutex
);
4516 * sys_sched_getaffinity - get the cpu affinity of a process
4517 * @pid: pid of the process
4518 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4519 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4521 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4522 unsigned long __user
*user_mask_ptr
)
4527 if (len
< sizeof(cpumask_t
))
4530 ret
= sched_getaffinity(pid
, &mask
);
4534 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4537 return sizeof(cpumask_t
);
4541 * sys_sched_yield - yield the current processor to other threads.
4543 * This function yields the current CPU to other tasks. If there are no
4544 * other threads running on this CPU then this function will return.
4546 asmlinkage
long sys_sched_yield(void)
4548 struct rq
*rq
= this_rq_lock();
4550 schedstat_inc(rq
, yld_cnt
);
4551 if (unlikely(rq
->nr_running
== 1))
4552 schedstat_inc(rq
, yld_act_empty
);
4554 current
->sched_class
->yield_task(rq
, current
);
4557 * Since we are going to call schedule() anyway, there's
4558 * no need to preempt or enable interrupts:
4560 __release(rq
->lock
);
4561 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4562 _raw_spin_unlock(&rq
->lock
);
4563 preempt_enable_no_resched();
4570 static void __cond_resched(void)
4572 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4573 __might_sleep(__FILE__
, __LINE__
);
4576 * The BKS might be reacquired before we have dropped
4577 * PREEMPT_ACTIVE, which could trigger a second
4578 * cond_resched() call.
4581 add_preempt_count(PREEMPT_ACTIVE
);
4583 sub_preempt_count(PREEMPT_ACTIVE
);
4584 } while (need_resched());
4587 int __sched
cond_resched(void)
4589 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4590 system_state
== SYSTEM_RUNNING
) {
4596 EXPORT_SYMBOL(cond_resched
);
4599 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4600 * call schedule, and on return reacquire the lock.
4602 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4603 * operations here to prevent schedule() from being called twice (once via
4604 * spin_unlock(), once by hand).
4606 int cond_resched_lock(spinlock_t
*lock
)
4610 if (need_lockbreak(lock
)) {
4616 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4617 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4618 _raw_spin_unlock(lock
);
4619 preempt_enable_no_resched();
4626 EXPORT_SYMBOL(cond_resched_lock
);
4628 int __sched
cond_resched_softirq(void)
4630 BUG_ON(!in_softirq());
4632 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4640 EXPORT_SYMBOL(cond_resched_softirq
);
4643 * yield - yield the current processor to other threads.
4645 * This is a shortcut for kernel-space yielding - it marks the
4646 * thread runnable and calls sys_sched_yield().
4648 void __sched
yield(void)
4650 set_current_state(TASK_RUNNING
);
4653 EXPORT_SYMBOL(yield
);
4656 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4657 * that process accounting knows that this is a task in IO wait state.
4659 * But don't do that if it is a deliberate, throttling IO wait (this task
4660 * has set its backing_dev_info: the queue against which it should throttle)
4662 void __sched
io_schedule(void)
4664 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4666 delayacct_blkio_start();
4667 atomic_inc(&rq
->nr_iowait
);
4669 atomic_dec(&rq
->nr_iowait
);
4670 delayacct_blkio_end();
4672 EXPORT_SYMBOL(io_schedule
);
4674 long __sched
io_schedule_timeout(long timeout
)
4676 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4679 delayacct_blkio_start();
4680 atomic_inc(&rq
->nr_iowait
);
4681 ret
= schedule_timeout(timeout
);
4682 atomic_dec(&rq
->nr_iowait
);
4683 delayacct_blkio_end();
4688 * sys_sched_get_priority_max - return maximum RT priority.
4689 * @policy: scheduling class.
4691 * this syscall returns the maximum rt_priority that can be used
4692 * by a given scheduling class.
4694 asmlinkage
long sys_sched_get_priority_max(int policy
)
4701 ret
= MAX_USER_RT_PRIO
-1;
4713 * sys_sched_get_priority_min - return minimum RT priority.
4714 * @policy: scheduling class.
4716 * this syscall returns the minimum rt_priority that can be used
4717 * by a given scheduling class.
4719 asmlinkage
long sys_sched_get_priority_min(int policy
)
4737 * sys_sched_rr_get_interval - return the default timeslice of a process.
4738 * @pid: pid of the process.
4739 * @interval: userspace pointer to the timeslice value.
4741 * this syscall writes the default timeslice value of a given process
4742 * into the user-space timespec buffer. A value of '0' means infinity.
4745 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4747 struct task_struct
*p
;
4748 int retval
= -EINVAL
;
4755 read_lock(&tasklist_lock
);
4756 p
= find_process_by_pid(pid
);
4760 retval
= security_task_getscheduler(p
);
4764 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4765 0 : static_prio_timeslice(p
->static_prio
), &t
);
4766 read_unlock(&tasklist_lock
);
4767 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4771 read_unlock(&tasklist_lock
);
4775 static const char stat_nam
[] = "RSDTtZX";
4777 static void show_task(struct task_struct
*p
)
4779 unsigned long free
= 0;
4782 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4783 printk("%-13.13s %c", p
->comm
,
4784 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4785 #if BITS_PER_LONG == 32
4786 if (state
== TASK_RUNNING
)
4787 printk(" running ");
4789 printk(" %08lx ", thread_saved_pc(p
));
4791 if (state
== TASK_RUNNING
)
4792 printk(" running task ");
4794 printk(" %016lx ", thread_saved_pc(p
));
4796 #ifdef CONFIG_DEBUG_STACK_USAGE
4798 unsigned long *n
= end_of_stack(p
);
4801 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4804 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4806 if (state
!= TASK_RUNNING
)
4807 show_stack(p
, NULL
);
4810 void show_state_filter(unsigned long state_filter
)
4812 struct task_struct
*g
, *p
;
4814 #if BITS_PER_LONG == 32
4816 " task PC stack pid father\n");
4819 " task PC stack pid father\n");
4821 read_lock(&tasklist_lock
);
4822 do_each_thread(g
, p
) {
4824 * reset the NMI-timeout, listing all files on a slow
4825 * console might take alot of time:
4827 touch_nmi_watchdog();
4828 if (!state_filter
|| (p
->state
& state_filter
))
4830 } while_each_thread(g
, p
);
4832 touch_all_softlockup_watchdogs();
4834 #ifdef CONFIG_SCHED_DEBUG
4835 sysrq_sched_debug_show();
4837 read_unlock(&tasklist_lock
);
4839 * Only show locks if all tasks are dumped:
4841 if (state_filter
== -1)
4842 debug_show_all_locks();
4845 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4847 idle
->sched_class
= &idle_sched_class
;
4851 * init_idle - set up an idle thread for a given CPU
4852 * @idle: task in question
4853 * @cpu: cpu the idle task belongs to
4855 * NOTE: this function does not set the idle thread's NEED_RESCHED
4856 * flag, to make booting more robust.
4858 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4860 struct rq
*rq
= cpu_rq(cpu
);
4861 unsigned long flags
;
4864 idle
->se
.exec_start
= sched_clock();
4866 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4867 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4868 __set_task_cpu(idle
, cpu
);
4870 spin_lock_irqsave(&rq
->lock
, flags
);
4871 rq
->curr
= rq
->idle
= idle
;
4872 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4875 spin_unlock_irqrestore(&rq
->lock
, flags
);
4877 /* Set the preempt count _outside_ the spinlocks! */
4878 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4879 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4881 task_thread_info(idle
)->preempt_count
= 0;
4884 * The idle tasks have their own, simple scheduling class:
4886 idle
->sched_class
= &idle_sched_class
;
4890 * In a system that switches off the HZ timer nohz_cpu_mask
4891 * indicates which cpus entered this state. This is used
4892 * in the rcu update to wait only for active cpus. For system
4893 * which do not switch off the HZ timer nohz_cpu_mask should
4894 * always be CPU_MASK_NONE.
4896 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4899 * Increase the granularity value when there are more CPUs,
4900 * because with more CPUs the 'effective latency' as visible
4901 * to users decreases. But the relationship is not linear,
4902 * so pick a second-best guess by going with the log2 of the
4905 * This idea comes from the SD scheduler of Con Kolivas:
4907 static inline void sched_init_granularity(void)
4909 unsigned int factor
= 1 + ilog2(num_online_cpus());
4910 const unsigned long gran_limit
= 100000000;
4912 sysctl_sched_granularity
*= factor
;
4913 if (sysctl_sched_granularity
> gran_limit
)
4914 sysctl_sched_granularity
= gran_limit
;
4916 sysctl_sched_runtime_limit
= sysctl_sched_granularity
* 4;
4917 sysctl_sched_wakeup_granularity
= sysctl_sched_granularity
/ 2;
4922 * This is how migration works:
4924 * 1) we queue a struct migration_req structure in the source CPU's
4925 * runqueue and wake up that CPU's migration thread.
4926 * 2) we down() the locked semaphore => thread blocks.
4927 * 3) migration thread wakes up (implicitly it forces the migrated
4928 * thread off the CPU)
4929 * 4) it gets the migration request and checks whether the migrated
4930 * task is still in the wrong runqueue.
4931 * 5) if it's in the wrong runqueue then the migration thread removes
4932 * it and puts it into the right queue.
4933 * 6) migration thread up()s the semaphore.
4934 * 7) we wake up and the migration is done.
4938 * Change a given task's CPU affinity. Migrate the thread to a
4939 * proper CPU and schedule it away if the CPU it's executing on
4940 * is removed from the allowed bitmask.
4942 * NOTE: the caller must have a valid reference to the task, the
4943 * task must not exit() & deallocate itself prematurely. The
4944 * call is not atomic; no spinlocks may be held.
4946 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4948 struct migration_req req
;
4949 unsigned long flags
;
4953 rq
= task_rq_lock(p
, &flags
);
4954 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4959 p
->cpus_allowed
= new_mask
;
4960 /* Can the task run on the task's current CPU? If so, we're done */
4961 if (cpu_isset(task_cpu(p
), new_mask
))
4964 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4965 /* Need help from migration thread: drop lock and wait. */
4966 task_rq_unlock(rq
, &flags
);
4967 wake_up_process(rq
->migration_thread
);
4968 wait_for_completion(&req
.done
);
4969 tlb_migrate_finish(p
->mm
);
4973 task_rq_unlock(rq
, &flags
);
4977 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4980 * Move (not current) task off this cpu, onto dest cpu. We're doing
4981 * this because either it can't run here any more (set_cpus_allowed()
4982 * away from this CPU, or CPU going down), or because we're
4983 * attempting to rebalance this task on exec (sched_exec).
4985 * So we race with normal scheduler movements, but that's OK, as long
4986 * as the task is no longer on this CPU.
4988 * Returns non-zero if task was successfully migrated.
4990 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4992 struct rq
*rq_dest
, *rq_src
;
4995 if (unlikely(cpu_is_offline(dest_cpu
)))
4998 rq_src
= cpu_rq(src_cpu
);
4999 rq_dest
= cpu_rq(dest_cpu
);
5001 double_rq_lock(rq_src
, rq_dest
);
5002 /* Already moved. */
5003 if (task_cpu(p
) != src_cpu
)
5005 /* Affinity changed (again). */
5006 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5009 on_rq
= p
->se
.on_rq
;
5011 deactivate_task(rq_src
, p
, 0);
5013 set_task_cpu(p
, dest_cpu
);
5015 activate_task(rq_dest
, p
, 0);
5016 check_preempt_curr(rq_dest
, p
);
5020 double_rq_unlock(rq_src
, rq_dest
);
5025 * migration_thread - this is a highprio system thread that performs
5026 * thread migration by bumping thread off CPU then 'pushing' onto
5029 static int migration_thread(void *data
)
5031 int cpu
= (long)data
;
5035 BUG_ON(rq
->migration_thread
!= current
);
5037 set_current_state(TASK_INTERRUPTIBLE
);
5038 while (!kthread_should_stop()) {
5039 struct migration_req
*req
;
5040 struct list_head
*head
;
5042 spin_lock_irq(&rq
->lock
);
5044 if (cpu_is_offline(cpu
)) {
5045 spin_unlock_irq(&rq
->lock
);
5049 if (rq
->active_balance
) {
5050 active_load_balance(rq
, cpu
);
5051 rq
->active_balance
= 0;
5054 head
= &rq
->migration_queue
;
5056 if (list_empty(head
)) {
5057 spin_unlock_irq(&rq
->lock
);
5059 set_current_state(TASK_INTERRUPTIBLE
);
5062 req
= list_entry(head
->next
, struct migration_req
, list
);
5063 list_del_init(head
->next
);
5065 spin_unlock(&rq
->lock
);
5066 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5069 complete(&req
->done
);
5071 __set_current_state(TASK_RUNNING
);
5075 /* Wait for kthread_stop */
5076 set_current_state(TASK_INTERRUPTIBLE
);
5077 while (!kthread_should_stop()) {
5079 set_current_state(TASK_INTERRUPTIBLE
);
5081 __set_current_state(TASK_RUNNING
);
5085 #ifdef CONFIG_HOTPLUG_CPU
5087 * Figure out where task on dead CPU should go, use force if neccessary.
5088 * NOTE: interrupts should be disabled by the caller
5090 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5092 unsigned long flags
;
5099 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5100 cpus_and(mask
, mask
, p
->cpus_allowed
);
5101 dest_cpu
= any_online_cpu(mask
);
5103 /* On any allowed CPU? */
5104 if (dest_cpu
== NR_CPUS
)
5105 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5107 /* No more Mr. Nice Guy. */
5108 if (dest_cpu
== NR_CPUS
) {
5109 rq
= task_rq_lock(p
, &flags
);
5110 cpus_setall(p
->cpus_allowed
);
5111 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5112 task_rq_unlock(rq
, &flags
);
5115 * Don't tell them about moving exiting tasks or
5116 * kernel threads (both mm NULL), since they never
5119 if (p
->mm
&& printk_ratelimit())
5120 printk(KERN_INFO
"process %d (%s) no "
5121 "longer affine to cpu%d\n",
5122 p
->pid
, p
->comm
, dead_cpu
);
5124 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5129 * While a dead CPU has no uninterruptible tasks queued at this point,
5130 * it might still have a nonzero ->nr_uninterruptible counter, because
5131 * for performance reasons the counter is not stricly tracking tasks to
5132 * their home CPUs. So we just add the counter to another CPU's counter,
5133 * to keep the global sum constant after CPU-down:
5135 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5137 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5138 unsigned long flags
;
5140 local_irq_save(flags
);
5141 double_rq_lock(rq_src
, rq_dest
);
5142 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5143 rq_src
->nr_uninterruptible
= 0;
5144 double_rq_unlock(rq_src
, rq_dest
);
5145 local_irq_restore(flags
);
5148 /* Run through task list and migrate tasks from the dead cpu. */
5149 static void migrate_live_tasks(int src_cpu
)
5151 struct task_struct
*p
, *t
;
5153 write_lock_irq(&tasklist_lock
);
5155 do_each_thread(t
, p
) {
5159 if (task_cpu(p
) == src_cpu
)
5160 move_task_off_dead_cpu(src_cpu
, p
);
5161 } while_each_thread(t
, p
);
5163 write_unlock_irq(&tasklist_lock
);
5167 * Schedules idle task to be the next runnable task on current CPU.
5168 * It does so by boosting its priority to highest possible and adding it to
5169 * the _front_ of the runqueue. Used by CPU offline code.
5171 void sched_idle_next(void)
5173 int this_cpu
= smp_processor_id();
5174 struct rq
*rq
= cpu_rq(this_cpu
);
5175 struct task_struct
*p
= rq
->idle
;
5176 unsigned long flags
;
5178 /* cpu has to be offline */
5179 BUG_ON(cpu_online(this_cpu
));
5182 * Strictly not necessary since rest of the CPUs are stopped by now
5183 * and interrupts disabled on the current cpu.
5185 spin_lock_irqsave(&rq
->lock
, flags
);
5187 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5189 /* Add idle task to the _front_ of its priority queue: */
5190 activate_idle_task(p
, rq
);
5192 spin_unlock_irqrestore(&rq
->lock
, flags
);
5196 * Ensures that the idle task is using init_mm right before its cpu goes
5199 void idle_task_exit(void)
5201 struct mm_struct
*mm
= current
->active_mm
;
5203 BUG_ON(cpu_online(smp_processor_id()));
5206 switch_mm(mm
, &init_mm
, current
);
5210 /* called under rq->lock with disabled interrupts */
5211 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5213 struct rq
*rq
= cpu_rq(dead_cpu
);
5215 /* Must be exiting, otherwise would be on tasklist. */
5216 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5218 /* Cannot have done final schedule yet: would have vanished. */
5219 BUG_ON(p
->state
== TASK_DEAD
);
5224 * Drop lock around migration; if someone else moves it,
5225 * that's OK. No task can be added to this CPU, so iteration is
5227 * NOTE: interrupts should be left disabled --dev@
5229 spin_unlock(&rq
->lock
);
5230 move_task_off_dead_cpu(dead_cpu
, p
);
5231 spin_lock(&rq
->lock
);
5236 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5237 static void migrate_dead_tasks(unsigned int dead_cpu
)
5239 struct rq
*rq
= cpu_rq(dead_cpu
);
5240 struct task_struct
*next
;
5243 if (!rq
->nr_running
)
5245 update_rq_clock(rq
);
5246 next
= pick_next_task(rq
, rq
->curr
);
5249 migrate_dead(dead_cpu
, next
);
5253 #endif /* CONFIG_HOTPLUG_CPU */
5255 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5257 static struct ctl_table sd_ctl_dir
[] = {
5259 .procname
= "sched_domain",
5265 static struct ctl_table sd_ctl_root
[] = {
5267 .procname
= "kernel",
5269 .child
= sd_ctl_dir
,
5274 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5276 struct ctl_table
*entry
=
5277 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5280 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5286 set_table_entry(struct ctl_table
*entry
,
5287 const char *procname
, void *data
, int maxlen
,
5288 mode_t mode
, proc_handler
*proc_handler
)
5290 entry
->procname
= procname
;
5292 entry
->maxlen
= maxlen
;
5294 entry
->proc_handler
= proc_handler
;
5297 static struct ctl_table
*
5298 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5300 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5302 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5303 sizeof(long), 0644, proc_doulongvec_minmax
);
5304 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5305 sizeof(long), 0644, proc_doulongvec_minmax
);
5306 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5307 sizeof(int), 0644, proc_dointvec_minmax
);
5308 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5309 sizeof(int), 0644, proc_dointvec_minmax
);
5310 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5311 sizeof(int), 0644, proc_dointvec_minmax
);
5312 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5313 sizeof(int), 0644, proc_dointvec_minmax
);
5314 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5315 sizeof(int), 0644, proc_dointvec_minmax
);
5316 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5317 sizeof(int), 0644, proc_dointvec_minmax
);
5318 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5319 sizeof(int), 0644, proc_dointvec_minmax
);
5320 set_table_entry(&table
[10], "cache_nice_tries",
5321 &sd
->cache_nice_tries
,
5322 sizeof(int), 0644, proc_dointvec_minmax
);
5323 set_table_entry(&table
[12], "flags", &sd
->flags
,
5324 sizeof(int), 0644, proc_dointvec_minmax
);
5329 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5331 struct ctl_table
*entry
, *table
;
5332 struct sched_domain
*sd
;
5333 int domain_num
= 0, i
;
5336 for_each_domain(cpu
, sd
)
5338 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5341 for_each_domain(cpu
, sd
) {
5342 snprintf(buf
, 32, "domain%d", i
);
5343 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5345 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5352 static struct ctl_table_header
*sd_sysctl_header
;
5353 static void init_sched_domain_sysctl(void)
5355 int i
, cpu_num
= num_online_cpus();
5356 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5359 sd_ctl_dir
[0].child
= entry
;
5361 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5362 snprintf(buf
, 32, "cpu%d", i
);
5363 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5365 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5367 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5370 static void init_sched_domain_sysctl(void)
5376 * migration_call - callback that gets triggered when a CPU is added.
5377 * Here we can start up the necessary migration thread for the new CPU.
5379 static int __cpuinit
5380 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5382 struct task_struct
*p
;
5383 int cpu
= (long)hcpu
;
5384 unsigned long flags
;
5388 case CPU_LOCK_ACQUIRE
:
5389 mutex_lock(&sched_hotcpu_mutex
);
5392 case CPU_UP_PREPARE
:
5393 case CPU_UP_PREPARE_FROZEN
:
5394 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5397 kthread_bind(p
, cpu
);
5398 /* Must be high prio: stop_machine expects to yield to it. */
5399 rq
= task_rq_lock(p
, &flags
);
5400 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5401 task_rq_unlock(rq
, &flags
);
5402 cpu_rq(cpu
)->migration_thread
= p
;
5406 case CPU_ONLINE_FROZEN
:
5407 /* Strictly unneccessary, as first user will wake it. */
5408 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5411 #ifdef CONFIG_HOTPLUG_CPU
5412 case CPU_UP_CANCELED
:
5413 case CPU_UP_CANCELED_FROZEN
:
5414 if (!cpu_rq(cpu
)->migration_thread
)
5416 /* Unbind it from offline cpu so it can run. Fall thru. */
5417 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5418 any_online_cpu(cpu_online_map
));
5419 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5420 cpu_rq(cpu
)->migration_thread
= NULL
;
5424 case CPU_DEAD_FROZEN
:
5425 migrate_live_tasks(cpu
);
5427 kthread_stop(rq
->migration_thread
);
5428 rq
->migration_thread
= NULL
;
5429 /* Idle task back to normal (off runqueue, low prio) */
5430 rq
= task_rq_lock(rq
->idle
, &flags
);
5431 update_rq_clock(rq
);
5432 deactivate_task(rq
, rq
->idle
, 0);
5433 rq
->idle
->static_prio
= MAX_PRIO
;
5434 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5435 rq
->idle
->sched_class
= &idle_sched_class
;
5436 migrate_dead_tasks(cpu
);
5437 task_rq_unlock(rq
, &flags
);
5438 migrate_nr_uninterruptible(rq
);
5439 BUG_ON(rq
->nr_running
!= 0);
5441 /* No need to migrate the tasks: it was best-effort if
5442 * they didn't take sched_hotcpu_mutex. Just wake up
5443 * the requestors. */
5444 spin_lock_irq(&rq
->lock
);
5445 while (!list_empty(&rq
->migration_queue
)) {
5446 struct migration_req
*req
;
5448 req
= list_entry(rq
->migration_queue
.next
,
5449 struct migration_req
, list
);
5450 list_del_init(&req
->list
);
5451 complete(&req
->done
);
5453 spin_unlock_irq(&rq
->lock
);
5456 case CPU_LOCK_RELEASE
:
5457 mutex_unlock(&sched_hotcpu_mutex
);
5463 /* Register at highest priority so that task migration (migrate_all_tasks)
5464 * happens before everything else.
5466 static struct notifier_block __cpuinitdata migration_notifier
= {
5467 .notifier_call
= migration_call
,
5471 int __init
migration_init(void)
5473 void *cpu
= (void *)(long)smp_processor_id();
5476 /* Start one for the boot CPU: */
5477 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5478 BUG_ON(err
== NOTIFY_BAD
);
5479 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5480 register_cpu_notifier(&migration_notifier
);
5488 /* Number of possible processor ids */
5489 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5490 EXPORT_SYMBOL(nr_cpu_ids
);
5492 #undef SCHED_DOMAIN_DEBUG
5493 #ifdef SCHED_DOMAIN_DEBUG
5494 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5499 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5503 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5508 struct sched_group
*group
= sd
->groups
;
5509 cpumask_t groupmask
;
5511 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5512 cpus_clear(groupmask
);
5515 for (i
= 0; i
< level
+ 1; i
++)
5517 printk("domain %d: ", level
);
5519 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5520 printk("does not load-balance\n");
5522 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5527 printk("span %s\n", str
);
5529 if (!cpu_isset(cpu
, sd
->span
))
5530 printk(KERN_ERR
"ERROR: domain->span does not contain "
5532 if (!cpu_isset(cpu
, group
->cpumask
))
5533 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5537 for (i
= 0; i
< level
+ 2; i
++)
5543 printk(KERN_ERR
"ERROR: group is NULL\n");
5547 if (!group
->__cpu_power
) {
5549 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5553 if (!cpus_weight(group
->cpumask
)) {
5555 printk(KERN_ERR
"ERROR: empty group\n");
5558 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5560 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5563 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5565 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5568 group
= group
->next
;
5569 } while (group
!= sd
->groups
);
5572 if (!cpus_equal(sd
->span
, groupmask
))
5573 printk(KERN_ERR
"ERROR: groups don't span "
5581 if (!cpus_subset(groupmask
, sd
->span
))
5582 printk(KERN_ERR
"ERROR: parent span is not a superset "
5583 "of domain->span\n");
5588 # define sched_domain_debug(sd, cpu) do { } while (0)
5591 static int sd_degenerate(struct sched_domain
*sd
)
5593 if (cpus_weight(sd
->span
) == 1)
5596 /* Following flags need at least 2 groups */
5597 if (sd
->flags
& (SD_LOAD_BALANCE
|
5598 SD_BALANCE_NEWIDLE
|
5602 SD_SHARE_PKG_RESOURCES
)) {
5603 if (sd
->groups
!= sd
->groups
->next
)
5607 /* Following flags don't use groups */
5608 if (sd
->flags
& (SD_WAKE_IDLE
|
5617 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5619 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5621 if (sd_degenerate(parent
))
5624 if (!cpus_equal(sd
->span
, parent
->span
))
5627 /* Does parent contain flags not in child? */
5628 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5629 if (cflags
& SD_WAKE_AFFINE
)
5630 pflags
&= ~SD_WAKE_BALANCE
;
5631 /* Flags needing groups don't count if only 1 group in parent */
5632 if (parent
->groups
== parent
->groups
->next
) {
5633 pflags
&= ~(SD_LOAD_BALANCE
|
5634 SD_BALANCE_NEWIDLE
|
5638 SD_SHARE_PKG_RESOURCES
);
5640 if (~cflags
& pflags
)
5647 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5648 * hold the hotplug lock.
5650 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5652 struct rq
*rq
= cpu_rq(cpu
);
5653 struct sched_domain
*tmp
;
5655 /* Remove the sched domains which do not contribute to scheduling. */
5656 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5657 struct sched_domain
*parent
= tmp
->parent
;
5660 if (sd_parent_degenerate(tmp
, parent
)) {
5661 tmp
->parent
= parent
->parent
;
5663 parent
->parent
->child
= tmp
;
5667 if (sd
&& sd_degenerate(sd
)) {
5673 sched_domain_debug(sd
, cpu
);
5675 rcu_assign_pointer(rq
->sd
, sd
);
5678 /* cpus with isolated domains */
5679 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5681 /* Setup the mask of cpus configured for isolated domains */
5682 static int __init
isolated_cpu_setup(char *str
)
5684 int ints
[NR_CPUS
], i
;
5686 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5687 cpus_clear(cpu_isolated_map
);
5688 for (i
= 1; i
<= ints
[0]; i
++)
5689 if (ints
[i
] < NR_CPUS
)
5690 cpu_set(ints
[i
], cpu_isolated_map
);
5694 __setup ("isolcpus=", isolated_cpu_setup
);
5697 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5698 * to a function which identifies what group(along with sched group) a CPU
5699 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5700 * (due to the fact that we keep track of groups covered with a cpumask_t).
5702 * init_sched_build_groups will build a circular linked list of the groups
5703 * covered by the given span, and will set each group's ->cpumask correctly,
5704 * and ->cpu_power to 0.
5707 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5708 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5709 struct sched_group
**sg
))
5711 struct sched_group
*first
= NULL
, *last
= NULL
;
5712 cpumask_t covered
= CPU_MASK_NONE
;
5715 for_each_cpu_mask(i
, span
) {
5716 struct sched_group
*sg
;
5717 int group
= group_fn(i
, cpu_map
, &sg
);
5720 if (cpu_isset(i
, covered
))
5723 sg
->cpumask
= CPU_MASK_NONE
;
5724 sg
->__cpu_power
= 0;
5726 for_each_cpu_mask(j
, span
) {
5727 if (group_fn(j
, cpu_map
, NULL
) != group
)
5730 cpu_set(j
, covered
);
5731 cpu_set(j
, sg
->cpumask
);
5742 #define SD_NODES_PER_DOMAIN 16
5747 * find_next_best_node - find the next node to include in a sched_domain
5748 * @node: node whose sched_domain we're building
5749 * @used_nodes: nodes already in the sched_domain
5751 * Find the next node to include in a given scheduling domain. Simply
5752 * finds the closest node not already in the @used_nodes map.
5754 * Should use nodemask_t.
5756 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5758 int i
, n
, val
, min_val
, best_node
= 0;
5762 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5763 /* Start at @node */
5764 n
= (node
+ i
) % MAX_NUMNODES
;
5766 if (!nr_cpus_node(n
))
5769 /* Skip already used nodes */
5770 if (test_bit(n
, used_nodes
))
5773 /* Simple min distance search */
5774 val
= node_distance(node
, n
);
5776 if (val
< min_val
) {
5782 set_bit(best_node
, used_nodes
);
5787 * sched_domain_node_span - get a cpumask for a node's sched_domain
5788 * @node: node whose cpumask we're constructing
5789 * @size: number of nodes to include in this span
5791 * Given a node, construct a good cpumask for its sched_domain to span. It
5792 * should be one that prevents unnecessary balancing, but also spreads tasks
5795 static cpumask_t
sched_domain_node_span(int node
)
5797 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5798 cpumask_t span
, nodemask
;
5802 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5804 nodemask
= node_to_cpumask(node
);
5805 cpus_or(span
, span
, nodemask
);
5806 set_bit(node
, used_nodes
);
5808 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5809 int next_node
= find_next_best_node(node
, used_nodes
);
5811 nodemask
= node_to_cpumask(next_node
);
5812 cpus_or(span
, span
, nodemask
);
5819 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5822 * SMT sched-domains:
5824 #ifdef CONFIG_SCHED_SMT
5825 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5826 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5828 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5829 struct sched_group
**sg
)
5832 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5838 * multi-core sched-domains:
5840 #ifdef CONFIG_SCHED_MC
5841 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5842 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5845 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5846 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5847 struct sched_group
**sg
)
5850 cpumask_t mask
= cpu_sibling_map
[cpu
];
5851 cpus_and(mask
, mask
, *cpu_map
);
5852 group
= first_cpu(mask
);
5854 *sg
= &per_cpu(sched_group_core
, group
);
5857 #elif defined(CONFIG_SCHED_MC)
5858 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5859 struct sched_group
**sg
)
5862 *sg
= &per_cpu(sched_group_core
, cpu
);
5867 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5868 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5870 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5871 struct sched_group
**sg
)
5874 #ifdef CONFIG_SCHED_MC
5875 cpumask_t mask
= cpu_coregroup_map(cpu
);
5876 cpus_and(mask
, mask
, *cpu_map
);
5877 group
= first_cpu(mask
);
5878 #elif defined(CONFIG_SCHED_SMT)
5879 cpumask_t mask
= cpu_sibling_map
[cpu
];
5880 cpus_and(mask
, mask
, *cpu_map
);
5881 group
= first_cpu(mask
);
5886 *sg
= &per_cpu(sched_group_phys
, group
);
5892 * The init_sched_build_groups can't handle what we want to do with node
5893 * groups, so roll our own. Now each node has its own list of groups which
5894 * gets dynamically allocated.
5896 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5897 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5899 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5900 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5902 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5903 struct sched_group
**sg
)
5905 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5908 cpus_and(nodemask
, nodemask
, *cpu_map
);
5909 group
= first_cpu(nodemask
);
5912 *sg
= &per_cpu(sched_group_allnodes
, group
);
5916 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5918 struct sched_group
*sg
= group_head
;
5924 for_each_cpu_mask(j
, sg
->cpumask
) {
5925 struct sched_domain
*sd
;
5927 sd
= &per_cpu(phys_domains
, j
);
5928 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5930 * Only add "power" once for each
5936 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5939 if (sg
!= group_head
)
5945 /* Free memory allocated for various sched_group structures */
5946 static void free_sched_groups(const cpumask_t
*cpu_map
)
5950 for_each_cpu_mask(cpu
, *cpu_map
) {
5951 struct sched_group
**sched_group_nodes
5952 = sched_group_nodes_bycpu
[cpu
];
5954 if (!sched_group_nodes
)
5957 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5958 cpumask_t nodemask
= node_to_cpumask(i
);
5959 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5961 cpus_and(nodemask
, nodemask
, *cpu_map
);
5962 if (cpus_empty(nodemask
))
5972 if (oldsg
!= sched_group_nodes
[i
])
5975 kfree(sched_group_nodes
);
5976 sched_group_nodes_bycpu
[cpu
] = NULL
;
5980 static void free_sched_groups(const cpumask_t
*cpu_map
)
5986 * Initialize sched groups cpu_power.
5988 * cpu_power indicates the capacity of sched group, which is used while
5989 * distributing the load between different sched groups in a sched domain.
5990 * Typically cpu_power for all the groups in a sched domain will be same unless
5991 * there are asymmetries in the topology. If there are asymmetries, group
5992 * having more cpu_power will pickup more load compared to the group having
5995 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5996 * the maximum number of tasks a group can handle in the presence of other idle
5997 * or lightly loaded groups in the same sched domain.
5999 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6001 struct sched_domain
*child
;
6002 struct sched_group
*group
;
6004 WARN_ON(!sd
|| !sd
->groups
);
6006 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6011 sd
->groups
->__cpu_power
= 0;
6014 * For perf policy, if the groups in child domain share resources
6015 * (for example cores sharing some portions of the cache hierarchy
6016 * or SMT), then set this domain groups cpu_power such that each group
6017 * can handle only one task, when there are other idle groups in the
6018 * same sched domain.
6020 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6022 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6023 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6028 * add cpu_power of each child group to this groups cpu_power
6030 group
= child
->groups
;
6032 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6033 group
= group
->next
;
6034 } while (group
!= child
->groups
);
6038 * Build sched domains for a given set of cpus and attach the sched domains
6039 * to the individual cpus
6041 static int build_sched_domains(const cpumask_t
*cpu_map
)
6045 struct sched_group
**sched_group_nodes
= NULL
;
6046 int sd_allnodes
= 0;
6049 * Allocate the per-node list of sched groups
6051 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6053 if (!sched_group_nodes
) {
6054 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6057 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6061 * Set up domains for cpus specified by the cpu_map.
6063 for_each_cpu_mask(i
, *cpu_map
) {
6064 struct sched_domain
*sd
= NULL
, *p
;
6065 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6067 cpus_and(nodemask
, nodemask
, *cpu_map
);
6070 if (cpus_weight(*cpu_map
) >
6071 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6072 sd
= &per_cpu(allnodes_domains
, i
);
6073 *sd
= SD_ALLNODES_INIT
;
6074 sd
->span
= *cpu_map
;
6075 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6081 sd
= &per_cpu(node_domains
, i
);
6083 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6087 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6091 sd
= &per_cpu(phys_domains
, i
);
6093 sd
->span
= nodemask
;
6097 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6099 #ifdef CONFIG_SCHED_MC
6101 sd
= &per_cpu(core_domains
, i
);
6103 sd
->span
= cpu_coregroup_map(i
);
6104 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6107 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6110 #ifdef CONFIG_SCHED_SMT
6112 sd
= &per_cpu(cpu_domains
, i
);
6113 *sd
= SD_SIBLING_INIT
;
6114 sd
->span
= cpu_sibling_map
[i
];
6115 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6118 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6122 #ifdef CONFIG_SCHED_SMT
6123 /* Set up CPU (sibling) groups */
6124 for_each_cpu_mask(i
, *cpu_map
) {
6125 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6126 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6127 if (i
!= first_cpu(this_sibling_map
))
6130 init_sched_build_groups(this_sibling_map
, cpu_map
,
6135 #ifdef CONFIG_SCHED_MC
6136 /* Set up multi-core groups */
6137 for_each_cpu_mask(i
, *cpu_map
) {
6138 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6139 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6140 if (i
!= first_cpu(this_core_map
))
6142 init_sched_build_groups(this_core_map
, cpu_map
,
6143 &cpu_to_core_group
);
6147 /* Set up physical groups */
6148 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6149 cpumask_t nodemask
= node_to_cpumask(i
);
6151 cpus_and(nodemask
, nodemask
, *cpu_map
);
6152 if (cpus_empty(nodemask
))
6155 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6159 /* Set up node groups */
6161 init_sched_build_groups(*cpu_map
, cpu_map
,
6162 &cpu_to_allnodes_group
);
6164 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6165 /* Set up node groups */
6166 struct sched_group
*sg
, *prev
;
6167 cpumask_t nodemask
= node_to_cpumask(i
);
6168 cpumask_t domainspan
;
6169 cpumask_t covered
= CPU_MASK_NONE
;
6172 cpus_and(nodemask
, nodemask
, *cpu_map
);
6173 if (cpus_empty(nodemask
)) {
6174 sched_group_nodes
[i
] = NULL
;
6178 domainspan
= sched_domain_node_span(i
);
6179 cpus_and(domainspan
, domainspan
, *cpu_map
);
6181 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6183 printk(KERN_WARNING
"Can not alloc domain group for "
6187 sched_group_nodes
[i
] = sg
;
6188 for_each_cpu_mask(j
, nodemask
) {
6189 struct sched_domain
*sd
;
6191 sd
= &per_cpu(node_domains
, j
);
6194 sg
->__cpu_power
= 0;
6195 sg
->cpumask
= nodemask
;
6197 cpus_or(covered
, covered
, nodemask
);
6200 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6201 cpumask_t tmp
, notcovered
;
6202 int n
= (i
+ j
) % MAX_NUMNODES
;
6204 cpus_complement(notcovered
, covered
);
6205 cpus_and(tmp
, notcovered
, *cpu_map
);
6206 cpus_and(tmp
, tmp
, domainspan
);
6207 if (cpus_empty(tmp
))
6210 nodemask
= node_to_cpumask(n
);
6211 cpus_and(tmp
, tmp
, nodemask
);
6212 if (cpus_empty(tmp
))
6215 sg
= kmalloc_node(sizeof(struct sched_group
),
6219 "Can not alloc domain group for node %d\n", j
);
6222 sg
->__cpu_power
= 0;
6224 sg
->next
= prev
->next
;
6225 cpus_or(covered
, covered
, tmp
);
6232 /* Calculate CPU power for physical packages and nodes */
6233 #ifdef CONFIG_SCHED_SMT
6234 for_each_cpu_mask(i
, *cpu_map
) {
6235 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6237 init_sched_groups_power(i
, sd
);
6240 #ifdef CONFIG_SCHED_MC
6241 for_each_cpu_mask(i
, *cpu_map
) {
6242 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6244 init_sched_groups_power(i
, sd
);
6248 for_each_cpu_mask(i
, *cpu_map
) {
6249 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6251 init_sched_groups_power(i
, sd
);
6255 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6256 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6259 struct sched_group
*sg
;
6261 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6262 init_numa_sched_groups_power(sg
);
6266 /* Attach the domains */
6267 for_each_cpu_mask(i
, *cpu_map
) {
6268 struct sched_domain
*sd
;
6269 #ifdef CONFIG_SCHED_SMT
6270 sd
= &per_cpu(cpu_domains
, i
);
6271 #elif defined(CONFIG_SCHED_MC)
6272 sd
= &per_cpu(core_domains
, i
);
6274 sd
= &per_cpu(phys_domains
, i
);
6276 cpu_attach_domain(sd
, i
);
6283 free_sched_groups(cpu_map
);
6288 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6290 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6292 cpumask_t cpu_default_map
;
6296 * Setup mask for cpus without special case scheduling requirements.
6297 * For now this just excludes isolated cpus, but could be used to
6298 * exclude other special cases in the future.
6300 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6302 err
= build_sched_domains(&cpu_default_map
);
6307 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6309 free_sched_groups(cpu_map
);
6313 * Detach sched domains from a group of cpus specified in cpu_map
6314 * These cpus will now be attached to the NULL domain
6316 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6320 for_each_cpu_mask(i
, *cpu_map
)
6321 cpu_attach_domain(NULL
, i
);
6322 synchronize_sched();
6323 arch_destroy_sched_domains(cpu_map
);
6327 * Partition sched domains as specified by the cpumasks below.
6328 * This attaches all cpus from the cpumasks to the NULL domain,
6329 * waits for a RCU quiescent period, recalculates sched
6330 * domain information and then attaches them back to the
6331 * correct sched domains
6332 * Call with hotplug lock held
6334 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6336 cpumask_t change_map
;
6339 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6340 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6341 cpus_or(change_map
, *partition1
, *partition2
);
6343 /* Detach sched domains from all of the affected cpus */
6344 detach_destroy_domains(&change_map
);
6345 if (!cpus_empty(*partition1
))
6346 err
= build_sched_domains(partition1
);
6347 if (!err
&& !cpus_empty(*partition2
))
6348 err
= build_sched_domains(partition2
);
6353 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6354 static int arch_reinit_sched_domains(void)
6358 mutex_lock(&sched_hotcpu_mutex
);
6359 detach_destroy_domains(&cpu_online_map
);
6360 err
= arch_init_sched_domains(&cpu_online_map
);
6361 mutex_unlock(&sched_hotcpu_mutex
);
6366 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6370 if (buf
[0] != '0' && buf
[0] != '1')
6374 sched_smt_power_savings
= (buf
[0] == '1');
6376 sched_mc_power_savings
= (buf
[0] == '1');
6378 ret
= arch_reinit_sched_domains();
6380 return ret
? ret
: count
;
6383 #ifdef CONFIG_SCHED_MC
6384 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6386 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6388 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6389 const char *buf
, size_t count
)
6391 return sched_power_savings_store(buf
, count
, 0);
6393 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6394 sched_mc_power_savings_store
);
6397 #ifdef CONFIG_SCHED_SMT
6398 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6400 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6402 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6403 const char *buf
, size_t count
)
6405 return sched_power_savings_store(buf
, count
, 1);
6407 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6408 sched_smt_power_savings_store
);
6411 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6415 #ifdef CONFIG_SCHED_SMT
6417 err
= sysfs_create_file(&cls
->kset
.kobj
,
6418 &attr_sched_smt_power_savings
.attr
);
6420 #ifdef CONFIG_SCHED_MC
6421 if (!err
&& mc_capable())
6422 err
= sysfs_create_file(&cls
->kset
.kobj
,
6423 &attr_sched_mc_power_savings
.attr
);
6430 * Force a reinitialization of the sched domains hierarchy. The domains
6431 * and groups cannot be updated in place without racing with the balancing
6432 * code, so we temporarily attach all running cpus to the NULL domain
6433 * which will prevent rebalancing while the sched domains are recalculated.
6435 static int update_sched_domains(struct notifier_block
*nfb
,
6436 unsigned long action
, void *hcpu
)
6439 case CPU_UP_PREPARE
:
6440 case CPU_UP_PREPARE_FROZEN
:
6441 case CPU_DOWN_PREPARE
:
6442 case CPU_DOWN_PREPARE_FROZEN
:
6443 detach_destroy_domains(&cpu_online_map
);
6446 case CPU_UP_CANCELED
:
6447 case CPU_UP_CANCELED_FROZEN
:
6448 case CPU_DOWN_FAILED
:
6449 case CPU_DOWN_FAILED_FROZEN
:
6451 case CPU_ONLINE_FROZEN
:
6453 case CPU_DEAD_FROZEN
:
6455 * Fall through and re-initialise the domains.
6462 /* The hotplug lock is already held by cpu_up/cpu_down */
6463 arch_init_sched_domains(&cpu_online_map
);
6468 void __init
sched_init_smp(void)
6470 cpumask_t non_isolated_cpus
;
6472 mutex_lock(&sched_hotcpu_mutex
);
6473 arch_init_sched_domains(&cpu_online_map
);
6474 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6475 if (cpus_empty(non_isolated_cpus
))
6476 cpu_set(smp_processor_id(), non_isolated_cpus
);
6477 mutex_unlock(&sched_hotcpu_mutex
);
6478 /* XXX: Theoretical race here - CPU may be hotplugged now */
6479 hotcpu_notifier(update_sched_domains
, 0);
6481 init_sched_domain_sysctl();
6483 /* Move init over to a non-isolated CPU */
6484 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6486 sched_init_granularity();
6489 void __init
sched_init_smp(void)
6491 sched_init_granularity();
6493 #endif /* CONFIG_SMP */
6495 int in_sched_functions(unsigned long addr
)
6497 /* Linker adds these: start and end of __sched functions */
6498 extern char __sched_text_start
[], __sched_text_end
[];
6500 return in_lock_functions(addr
) ||
6501 (addr
>= (unsigned long)__sched_text_start
6502 && addr
< (unsigned long)__sched_text_end
);
6505 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6507 cfs_rq
->tasks_timeline
= RB_ROOT
;
6508 cfs_rq
->fair_clock
= 1;
6509 #ifdef CONFIG_FAIR_GROUP_SCHED
6514 void __init
sched_init(void)
6516 u64 now
= sched_clock();
6517 int highest_cpu
= 0;
6521 * Link up the scheduling class hierarchy:
6523 rt_sched_class
.next
= &fair_sched_class
;
6524 fair_sched_class
.next
= &idle_sched_class
;
6525 idle_sched_class
.next
= NULL
;
6527 for_each_possible_cpu(i
) {
6528 struct rt_prio_array
*array
;
6532 spin_lock_init(&rq
->lock
);
6533 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6536 init_cfs_rq(&rq
->cfs
, rq
);
6537 #ifdef CONFIG_FAIR_GROUP_SCHED
6538 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6539 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6541 rq
->ls
.load_update_last
= now
;
6542 rq
->ls
.load_update_start
= now
;
6544 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6545 rq
->cpu_load
[j
] = 0;
6548 rq
->active_balance
= 0;
6549 rq
->next_balance
= jiffies
;
6552 rq
->migration_thread
= NULL
;
6553 INIT_LIST_HEAD(&rq
->migration_queue
);
6555 atomic_set(&rq
->nr_iowait
, 0);
6557 array
= &rq
->rt
.active
;
6558 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6559 INIT_LIST_HEAD(array
->queue
+ j
);
6560 __clear_bit(j
, array
->bitmap
);
6563 /* delimiter for bitsearch: */
6564 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6567 set_load_weight(&init_task
);
6569 #ifdef CONFIG_PREEMPT_NOTIFIERS
6570 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6574 nr_cpu_ids
= highest_cpu
+ 1;
6575 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6578 #ifdef CONFIG_RT_MUTEXES
6579 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6583 * The boot idle thread does lazy MMU switching as well:
6585 atomic_inc(&init_mm
.mm_count
);
6586 enter_lazy_tlb(&init_mm
, current
);
6589 * Make us the idle thread. Technically, schedule() should not be
6590 * called from this thread, however somewhere below it might be,
6591 * but because we are the idle thread, we just pick up running again
6592 * when this runqueue becomes "idle".
6594 init_idle(current
, smp_processor_id());
6596 * During early bootup we pretend to be a normal task:
6598 current
->sched_class
= &fair_sched_class
;
6601 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6602 void __might_sleep(char *file
, int line
)
6605 static unsigned long prev_jiffy
; /* ratelimiting */
6607 if ((in_atomic() || irqs_disabled()) &&
6608 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6609 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6611 prev_jiffy
= jiffies
;
6612 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6613 " context at %s:%d\n", file
, line
);
6614 printk("in_atomic():%d, irqs_disabled():%d\n",
6615 in_atomic(), irqs_disabled());
6616 debug_show_held_locks(current
);
6617 if (irqs_disabled())
6618 print_irqtrace_events(current
);
6623 EXPORT_SYMBOL(__might_sleep
);
6626 #ifdef CONFIG_MAGIC_SYSRQ
6627 void normalize_rt_tasks(void)
6629 struct task_struct
*g
, *p
;
6630 unsigned long flags
;
6634 read_lock_irq(&tasklist_lock
);
6635 do_each_thread(g
, p
) {
6637 p
->se
.wait_runtime
= 0;
6638 p
->se
.exec_start
= 0;
6639 p
->se
.wait_start_fair
= 0;
6640 p
->se
.sleep_start_fair
= 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
)->cfs
.fair_clock
= 0;
6647 task_rq(p
)->clock
= 0;
6651 * Renice negative nice level userspace
6654 if (TASK_NICE(p
) < 0 && p
->mm
)
6655 set_user_nice(p
, 0);
6659 spin_lock_irqsave(&p
->pi_lock
, flags
);
6660 rq
= __task_rq_lock(p
);
6663 * Do not touch the migration thread:
6665 if (p
== rq
->migration_thread
)
6669 update_rq_clock(rq
);
6670 on_rq
= p
->se
.on_rq
;
6672 deactivate_task(rq
, p
, 0);
6673 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6675 activate_task(rq
, p
, 0);
6676 resched_task(rq
->curr
);
6681 __task_rq_unlock(rq
);
6682 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6683 } while_each_thread(g
, p
);
6685 read_unlock_irq(&tasklist_lock
);
6688 #endif /* CONFIG_MAGIC_SYSRQ */
6692 * These functions are only useful for the IA64 MCA handling.
6694 * They can only be called when the whole system has been
6695 * stopped - every CPU needs to be quiescent, and no scheduling
6696 * activity can take place. Using them for anything else would
6697 * be a serious bug, and as a result, they aren't even visible
6698 * under any other configuration.
6702 * curr_task - return the current task for a given cpu.
6703 * @cpu: the processor in question.
6705 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6707 struct task_struct
*curr_task(int cpu
)
6709 return cpu_curr(cpu
);
6713 * set_curr_task - set the current task for a given cpu.
6714 * @cpu: the processor in question.
6715 * @p: the task pointer to set.
6717 * Description: This function must only be used when non-maskable interrupts
6718 * are serviced on a separate stack. It allows the architecture to switch the
6719 * notion of the current task on a cpu in a non-blocking manner. This function
6720 * must be called with all CPU's synchronized, and interrupts disabled, the
6721 * and caller must save the original value of the current task (see
6722 * curr_task() above) and restore that value before reenabling interrupts and
6723 * re-starting the system.
6725 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6727 void set_curr_task(int cpu
, struct task_struct
*p
)