4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak
)) sched_clock(void)
75 return (unsigned long long)jiffies
* (1000000000 / HZ
);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
122 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
131 sg
->__cpu_power
+= val
;
132 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
136 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio
)
145 if (static_prio
== NICE_TO_PRIO(19))
148 if (static_prio
< NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
151 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
154 static inline int rt_policy(int policy
)
156 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
161 static inline int task_has_rt_policy(struct task_struct
*p
)
163 return rt_policy(p
->policy
);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array
{
170 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
171 struct list_head queue
[MAX_RT_PRIO
];
174 /* CFS-related fields in a runqueue */
176 struct load_weight load
;
177 unsigned long nr_running
;
183 unsigned long wait_runtime_overruns
, wait_runtime_underruns
;
185 struct rb_root tasks_timeline
;
186 struct rb_node
*rb_leftmost
;
187 struct rb_node
*rb_load_balance_curr
;
188 /* 'curr' points to currently running entity on this cfs_rq.
189 * It is set to NULL otherwise (i.e when none are currently running).
191 struct sched_entity
*curr
;
192 #ifdef CONFIG_FAIR_GROUP_SCHED
193 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
195 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
196 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
197 * (like users, containers etc.)
199 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
200 * list is used during load balance.
202 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
206 /* Real-Time classes' related field in a runqueue: */
208 struct rt_prio_array active
;
209 int rt_load_balance_idx
;
210 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
214 * This is the main, per-CPU runqueue data structure.
216 * Locking rule: those places that want to lock multiple runqueues
217 * (such as the load balancing or the thread migration code), lock
218 * acquire operations must be ordered by ascending &runqueue.
221 spinlock_t lock
; /* runqueue lock */
224 * nr_running and cpu_load should be in the same cacheline because
225 * remote CPUs use both these fields when doing load calculation.
227 unsigned long nr_running
;
228 #define CPU_LOAD_IDX_MAX 5
229 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
230 unsigned char idle_at_tick
;
232 unsigned char in_nohz_recently
;
234 struct load_weight load
; /* capture load from *all* tasks on this cpu */
235 unsigned long nr_load_updates
;
239 #ifdef CONFIG_FAIR_GROUP_SCHED
240 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
245 * This is part of a global counter where only the total sum
246 * over all CPUs matters. A task can increase this counter on
247 * one CPU and if it got migrated afterwards it may decrease
248 * it on another CPU. Always updated under the runqueue lock:
250 unsigned long nr_uninterruptible
;
252 struct task_struct
*curr
, *idle
;
253 unsigned long next_balance
;
254 struct mm_struct
*prev_mm
;
256 u64 clock
, prev_clock_raw
;
259 unsigned int clock_warps
, clock_overflows
;
261 unsigned int clock_deep_idle_events
;
267 struct sched_domain
*sd
;
269 /* For active balancing */
272 int cpu
; /* cpu of this runqueue */
274 struct task_struct
*migration_thread
;
275 struct list_head migration_queue
;
278 #ifdef CONFIG_SCHEDSTATS
280 struct sched_info rq_sched_info
;
282 /* sys_sched_yield() stats */
283 unsigned long yld_exp_empty
;
284 unsigned long yld_act_empty
;
285 unsigned long yld_both_empty
;
286 unsigned long yld_cnt
;
288 /* schedule() stats */
289 unsigned long sched_switch
;
290 unsigned long sched_cnt
;
291 unsigned long sched_goidle
;
293 /* try_to_wake_up() stats */
294 unsigned long ttwu_cnt
;
295 unsigned long ttwu_local
;
297 struct lock_class_key rq_lock_key
;
300 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
301 static DEFINE_MUTEX(sched_hotcpu_mutex
);
303 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
305 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
308 static inline int cpu_of(struct rq
*rq
)
318 * Update the per-runqueue clock, as finegrained as the platform can give
319 * us, but without assuming monotonicity, etc.:
321 static void __update_rq_clock(struct rq
*rq
)
323 u64 prev_raw
= rq
->prev_clock_raw
;
324 u64 now
= sched_clock();
325 s64 delta
= now
- prev_raw
;
326 u64 clock
= rq
->clock
;
328 #ifdef CONFIG_SCHED_DEBUG
329 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
332 * Protect against sched_clock() occasionally going backwards:
334 if (unlikely(delta
< 0)) {
339 * Catch too large forward jumps too:
341 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
342 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
343 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
346 rq
->clock_overflows
++;
348 if (unlikely(delta
> rq
->clock_max_delta
))
349 rq
->clock_max_delta
= delta
;
354 rq
->prev_clock_raw
= now
;
358 static void update_rq_clock(struct rq
*rq
)
360 if (likely(smp_processor_id() == cpu_of(rq
)))
361 __update_rq_clock(rq
);
365 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
366 * See detach_destroy_domains: synchronize_sched for details.
368 * The domain tree of any CPU may only be accessed from within
369 * preempt-disabled sections.
371 #define for_each_domain(cpu, __sd) \
372 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
374 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
375 #define this_rq() (&__get_cpu_var(runqueues))
376 #define task_rq(p) cpu_rq(task_cpu(p))
377 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
380 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
382 #ifdef CONFIG_SCHED_DEBUG
383 # define const_debug __read_mostly
385 # define const_debug static const
389 * Debugging: various feature bits
392 SCHED_FEAT_FAIR_SLEEPERS
= 1,
393 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 2,
394 SCHED_FEAT_SLEEPER_AVG
= 4,
395 SCHED_FEAT_SLEEPER_LOAD_AVG
= 8,
396 SCHED_FEAT_START_DEBIT
= 16,
397 SCHED_FEAT_USE_TREE_AVG
= 32,
398 SCHED_FEAT_APPROX_AVG
= 64,
401 const_debug
unsigned int sysctl_sched_features
=
402 SCHED_FEAT_FAIR_SLEEPERS
*0 |
403 SCHED_FEAT_NEW_FAIR_SLEEPERS
*1 |
404 SCHED_FEAT_SLEEPER_AVG
*0 |
405 SCHED_FEAT_SLEEPER_LOAD_AVG
*1 |
406 SCHED_FEAT_START_DEBIT
*1 |
407 SCHED_FEAT_USE_TREE_AVG
*0 |
408 SCHED_FEAT_APPROX_AVG
*0;
410 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
413 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
414 * clock constructed from sched_clock():
416 unsigned long long cpu_clock(int cpu
)
418 unsigned long long now
;
422 local_irq_save(flags
);
426 local_irq_restore(flags
);
431 #ifdef CONFIG_FAIR_GROUP_SCHED
432 /* Change a task's ->cfs_rq if it moves across CPUs */
433 static inline void set_task_cfs_rq(struct task_struct
*p
)
435 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
438 static inline void set_task_cfs_rq(struct task_struct
*p
)
443 #ifndef prepare_arch_switch
444 # define prepare_arch_switch(next) do { } while (0)
446 #ifndef finish_arch_switch
447 # define finish_arch_switch(prev) do { } while (0)
450 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
451 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
453 return rq
->curr
== p
;
456 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
460 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
462 #ifdef CONFIG_DEBUG_SPINLOCK
463 /* this is a valid case when another task releases the spinlock */
464 rq
->lock
.owner
= current
;
467 * If we are tracking spinlock dependencies then we have to
468 * fix up the runqueue lock - which gets 'carried over' from
471 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
473 spin_unlock_irq(&rq
->lock
);
476 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
477 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
482 return rq
->curr
== p
;
486 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
490 * We can optimise this out completely for !SMP, because the
491 * SMP rebalancing from interrupt is the only thing that cares
496 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
497 spin_unlock_irq(&rq
->lock
);
499 spin_unlock(&rq
->lock
);
503 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
507 * After ->oncpu is cleared, the task can be moved to a different CPU.
508 * We must ensure this doesn't happen until the switch is completely
514 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
518 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
521 * __task_rq_lock - lock the runqueue a given task resides on.
522 * Must be called interrupts disabled.
524 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
531 spin_lock(&rq
->lock
);
532 if (unlikely(rq
!= task_rq(p
))) {
533 spin_unlock(&rq
->lock
);
534 goto repeat_lock_task
;
540 * task_rq_lock - lock the runqueue a given task resides on and disable
541 * interrupts. Note the ordering: we can safely lookup the task_rq without
542 * explicitly disabling preemption.
544 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
550 local_irq_save(*flags
);
552 spin_lock(&rq
->lock
);
553 if (unlikely(rq
!= task_rq(p
))) {
554 spin_unlock_irqrestore(&rq
->lock
, *flags
);
555 goto repeat_lock_task
;
560 static inline void __task_rq_unlock(struct rq
*rq
)
563 spin_unlock(&rq
->lock
);
566 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
569 spin_unlock_irqrestore(&rq
->lock
, *flags
);
573 * this_rq_lock - lock this runqueue and disable interrupts.
575 static inline struct rq
*this_rq_lock(void)
582 spin_lock(&rq
->lock
);
588 * We are going deep-idle (irqs are disabled):
590 void sched_clock_idle_sleep_event(void)
592 struct rq
*rq
= cpu_rq(smp_processor_id());
594 spin_lock(&rq
->lock
);
595 __update_rq_clock(rq
);
596 spin_unlock(&rq
->lock
);
597 rq
->clock_deep_idle_events
++;
599 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
602 * We just idled delta nanoseconds (called with irqs disabled):
604 void sched_clock_idle_wakeup_event(u64 delta_ns
)
606 struct rq
*rq
= cpu_rq(smp_processor_id());
607 u64 now
= sched_clock();
609 rq
->idle_clock
+= delta_ns
;
611 * Override the previous timestamp and ignore all
612 * sched_clock() deltas that occured while we idled,
613 * and use the PM-provided delta_ns to advance the
616 spin_lock(&rq
->lock
);
617 rq
->prev_clock_raw
= now
;
618 rq
->clock
+= delta_ns
;
619 spin_unlock(&rq
->lock
);
621 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
624 * resched_task - mark a task 'to be rescheduled now'.
626 * On UP this means the setting of the need_resched flag, on SMP it
627 * might also involve a cross-CPU call to trigger the scheduler on
632 #ifndef tsk_is_polling
633 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
636 static void resched_task(struct task_struct
*p
)
640 assert_spin_locked(&task_rq(p
)->lock
);
642 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
645 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
648 if (cpu
== smp_processor_id())
651 /* NEED_RESCHED must be visible before we test polling */
653 if (!tsk_is_polling(p
))
654 smp_send_reschedule(cpu
);
657 static void resched_cpu(int cpu
)
659 struct rq
*rq
= cpu_rq(cpu
);
662 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
664 resched_task(cpu_curr(cpu
));
665 spin_unlock_irqrestore(&rq
->lock
, flags
);
668 static inline void resched_task(struct task_struct
*p
)
670 assert_spin_locked(&task_rq(p
)->lock
);
671 set_tsk_need_resched(p
);
675 #if BITS_PER_LONG == 32
676 # define WMULT_CONST (~0UL)
678 # define WMULT_CONST (1UL << 32)
681 #define WMULT_SHIFT 32
684 * Shift right and round:
686 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
689 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
690 struct load_weight
*lw
)
694 if (unlikely(!lw
->inv_weight
))
695 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
697 tmp
= (u64
)delta_exec
* weight
;
699 * Check whether we'd overflow the 64-bit multiplication:
701 if (unlikely(tmp
> WMULT_CONST
))
702 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
705 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
707 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
710 static inline unsigned long
711 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
713 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
716 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
719 if (sched_feat(FAIR_SLEEPERS
))
720 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
723 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
726 if (sched_feat(FAIR_SLEEPERS
) && likely(lw
->weight
))
727 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
731 * To aid in avoiding the subversion of "niceness" due to uneven distribution
732 * of tasks with abnormal "nice" values across CPUs the contribution that
733 * each task makes to its run queue's load is weighted according to its
734 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
735 * scaled version of the new time slice allocation that they receive on time
739 #define WEIGHT_IDLEPRIO 2
740 #define WMULT_IDLEPRIO (1 << 31)
743 * Nice levels are multiplicative, with a gentle 10% change for every
744 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
745 * nice 1, it will get ~10% less CPU time than another CPU-bound task
746 * that remained on nice 0.
748 * The "10% effect" is relative and cumulative: from _any_ nice level,
749 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
750 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
751 * If a task goes up by ~10% and another task goes down by ~10% then
752 * the relative distance between them is ~25%.)
754 static const int prio_to_weight
[40] = {
755 /* -20 */ 88761, 71755, 56483, 46273, 36291,
756 /* -15 */ 29154, 23254, 18705, 14949, 11916,
757 /* -10 */ 9548, 7620, 6100, 4904, 3906,
758 /* -5 */ 3121, 2501, 1991, 1586, 1277,
759 /* 0 */ 1024, 820, 655, 526, 423,
760 /* 5 */ 335, 272, 215, 172, 137,
761 /* 10 */ 110, 87, 70, 56, 45,
762 /* 15 */ 36, 29, 23, 18, 15,
766 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
768 * In cases where the weight does not change often, we can use the
769 * precalculated inverse to speed up arithmetics by turning divisions
770 * into multiplications:
772 static const u32 prio_to_wmult
[40] = {
773 /* -20 */ 48388, 59856, 76040, 92818, 118348,
774 /* -15 */ 147320, 184698, 229616, 287308, 360437,
775 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
776 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
777 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
778 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
779 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
780 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
783 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
786 * runqueue iterator, to support SMP load-balancing between different
787 * scheduling classes, without having to expose their internal data
788 * structures to the load-balancing proper:
792 struct task_struct
*(*start
)(void *);
793 struct task_struct
*(*next
)(void *);
796 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
797 unsigned long max_nr_move
, unsigned long max_load_move
,
798 struct sched_domain
*sd
, enum cpu_idle_type idle
,
799 int *all_pinned
, unsigned long *load_moved
,
800 int *this_best_prio
, struct rq_iterator
*iterator
);
802 #include "sched_stats.h"
803 #include "sched_rt.c"
804 #include "sched_fair.c"
805 #include "sched_idletask.c"
806 #ifdef CONFIG_SCHED_DEBUG
807 # include "sched_debug.c"
810 #define sched_class_highest (&rt_sched_class)
813 * Update delta_exec, delta_fair fields for rq.
815 * delta_fair clock advances at a rate inversely proportional to
816 * total load (rq->load.weight) on the runqueue, while
817 * delta_exec advances at the same rate as wall-clock (provided
820 * delta_exec / delta_fair is a measure of the (smoothened) load on this
821 * runqueue over any given interval. This (smoothened) load is used
822 * during load balance.
824 * This function is called /before/ updating rq->load
825 * and when switching tasks.
827 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
829 update_load_add(&rq
->load
, p
->se
.load
.weight
);
832 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
834 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
837 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
843 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
849 static void set_load_weight(struct task_struct
*p
)
851 p
->se
.wait_runtime
= 0;
853 if (task_has_rt_policy(p
)) {
854 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
855 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
860 * SCHED_IDLE tasks get minimal weight:
862 if (p
->policy
== SCHED_IDLE
) {
863 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
864 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
868 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
869 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
872 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
874 sched_info_queued(p
);
875 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
879 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
881 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
886 * __normal_prio - return the priority that is based on the static prio
888 static inline int __normal_prio(struct task_struct
*p
)
890 return p
->static_prio
;
894 * Calculate the expected normal priority: i.e. priority
895 * without taking RT-inheritance into account. Might be
896 * boosted by interactivity modifiers. Changes upon fork,
897 * setprio syscalls, and whenever the interactivity
898 * estimator recalculates.
900 static inline int normal_prio(struct task_struct
*p
)
904 if (task_has_rt_policy(p
))
905 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
907 prio
= __normal_prio(p
);
912 * Calculate the current priority, i.e. the priority
913 * taken into account by the scheduler. This value might
914 * be boosted by RT tasks, or might be boosted by
915 * interactivity modifiers. Will be RT if the task got
916 * RT-boosted. If not then it returns p->normal_prio.
918 static int effective_prio(struct task_struct
*p
)
920 p
->normal_prio
= normal_prio(p
);
922 * If we are RT tasks or we were boosted to RT priority,
923 * keep the priority unchanged. Otherwise, update priority
924 * to the normal priority:
926 if (!rt_prio(p
->prio
))
927 return p
->normal_prio
;
932 * activate_task - move a task to the runqueue.
934 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
936 if (p
->state
== TASK_UNINTERRUPTIBLE
)
937 rq
->nr_uninterruptible
--;
939 enqueue_task(rq
, p
, wakeup
);
940 inc_nr_running(p
, rq
);
944 * activate_idle_task - move idle task to the _front_ of runqueue.
946 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
950 if (p
->state
== TASK_UNINTERRUPTIBLE
)
951 rq
->nr_uninterruptible
--;
953 enqueue_task(rq
, p
, 0);
954 inc_nr_running(p
, rq
);
958 * deactivate_task - remove a task from the runqueue.
960 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
962 if (p
->state
== TASK_UNINTERRUPTIBLE
)
963 rq
->nr_uninterruptible
++;
965 dequeue_task(rq
, p
, sleep
);
966 dec_nr_running(p
, rq
);
970 * task_curr - is this task currently executing on a CPU?
971 * @p: the task in question.
973 inline int task_curr(const struct task_struct
*p
)
975 return cpu_curr(task_cpu(p
)) == p
;
978 /* Used instead of source_load when we know the type == 0 */
979 unsigned long weighted_cpuload(const int cpu
)
981 return cpu_rq(cpu
)->load
.weight
;
984 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
987 task_thread_info(p
)->cpu
= cpu
;
994 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
996 int old_cpu
= task_cpu(p
);
997 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
998 u64 clock_offset
, fair_clock_offset
;
1000 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1001 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
1003 if (p
->se
.wait_start_fair
)
1004 p
->se
.wait_start_fair
-= fair_clock_offset
;
1006 #ifdef CONFIG_SCHEDSTATS
1007 if (p
->se
.wait_start
)
1008 p
->se
.wait_start
-= clock_offset
;
1009 if (p
->se
.sleep_start
)
1010 p
->se
.sleep_start
-= clock_offset
;
1011 if (p
->se
.block_start
)
1012 p
->se
.block_start
-= clock_offset
;
1015 __set_task_cpu(p
, new_cpu
);
1018 struct migration_req
{
1019 struct list_head list
;
1021 struct task_struct
*task
;
1024 struct completion done
;
1028 * The task's runqueue lock must be held.
1029 * Returns true if you have to wait for migration thread.
1032 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1034 struct rq
*rq
= task_rq(p
);
1037 * If the task is not on a runqueue (and not running), then
1038 * it is sufficient to simply update the task's cpu field.
1040 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1041 set_task_cpu(p
, dest_cpu
);
1045 init_completion(&req
->done
);
1047 req
->dest_cpu
= dest_cpu
;
1048 list_add(&req
->list
, &rq
->migration_queue
);
1054 * wait_task_inactive - wait for a thread to unschedule.
1056 * The caller must ensure that the task *will* unschedule sometime soon,
1057 * else this function might spin for a *long* time. This function can't
1058 * be called with interrupts off, or it may introduce deadlock with
1059 * smp_call_function() if an IPI is sent by the same process we are
1060 * waiting to become inactive.
1062 void wait_task_inactive(struct task_struct
*p
)
1064 unsigned long flags
;
1070 * We do the initial early heuristics without holding
1071 * any task-queue locks at all. We'll only try to get
1072 * the runqueue lock when things look like they will
1078 * If the task is actively running on another CPU
1079 * still, just relax and busy-wait without holding
1082 * NOTE! Since we don't hold any locks, it's not
1083 * even sure that "rq" stays as the right runqueue!
1084 * But we don't care, since "task_running()" will
1085 * return false if the runqueue has changed and p
1086 * is actually now running somewhere else!
1088 while (task_running(rq
, p
))
1092 * Ok, time to look more closely! We need the rq
1093 * lock now, to be *sure*. If we're wrong, we'll
1094 * just go back and repeat.
1096 rq
= task_rq_lock(p
, &flags
);
1097 running
= task_running(rq
, p
);
1098 on_rq
= p
->se
.on_rq
;
1099 task_rq_unlock(rq
, &flags
);
1102 * Was it really running after all now that we
1103 * checked with the proper locks actually held?
1105 * Oops. Go back and try again..
1107 if (unlikely(running
)) {
1113 * It's not enough that it's not actively running,
1114 * it must be off the runqueue _entirely_, and not
1117 * So if it wa still runnable (but just not actively
1118 * running right now), it's preempted, and we should
1119 * yield - it could be a while.
1121 if (unlikely(on_rq
)) {
1127 * Ahh, all good. It wasn't running, and it wasn't
1128 * runnable, which means that it will never become
1129 * running in the future either. We're all done!
1134 * kick_process - kick a running thread to enter/exit the kernel
1135 * @p: the to-be-kicked thread
1137 * Cause a process which is running on another CPU to enter
1138 * kernel-mode, without any delay. (to get signals handled.)
1140 * NOTE: this function doesnt have to take the runqueue lock,
1141 * because all it wants to ensure is that the remote task enters
1142 * the kernel. If the IPI races and the task has been migrated
1143 * to another CPU then no harm is done and the purpose has been
1146 void kick_process(struct task_struct
*p
)
1152 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1153 smp_send_reschedule(cpu
);
1158 * Return a low guess at the load of a migration-source cpu weighted
1159 * according to the scheduling class and "nice" value.
1161 * We want to under-estimate the load of migration sources, to
1162 * balance conservatively.
1164 static inline unsigned long source_load(int cpu
, int type
)
1166 struct rq
*rq
= cpu_rq(cpu
);
1167 unsigned long total
= weighted_cpuload(cpu
);
1172 return min(rq
->cpu_load
[type
-1], total
);
1176 * Return a high guess at the load of a migration-target cpu weighted
1177 * according to the scheduling class and "nice" value.
1179 static inline unsigned long target_load(int cpu
, int type
)
1181 struct rq
*rq
= cpu_rq(cpu
);
1182 unsigned long total
= weighted_cpuload(cpu
);
1187 return max(rq
->cpu_load
[type
-1], total
);
1191 * Return the average load per task on the cpu's run queue
1193 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1195 struct rq
*rq
= cpu_rq(cpu
);
1196 unsigned long total
= weighted_cpuload(cpu
);
1197 unsigned long n
= rq
->nr_running
;
1199 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1203 * find_idlest_group finds and returns the least busy CPU group within the
1206 static struct sched_group
*
1207 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1209 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1210 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1211 int load_idx
= sd
->forkexec_idx
;
1212 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1215 unsigned long load
, avg_load
;
1219 /* Skip over this group if it has no CPUs allowed */
1220 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1223 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1225 /* Tally up the load of all CPUs in the group */
1228 for_each_cpu_mask(i
, group
->cpumask
) {
1229 /* Bias balancing toward cpus of our domain */
1231 load
= source_load(i
, load_idx
);
1233 load
= target_load(i
, load_idx
);
1238 /* Adjust by relative CPU power of the group */
1239 avg_load
= sg_div_cpu_power(group
,
1240 avg_load
* SCHED_LOAD_SCALE
);
1243 this_load
= avg_load
;
1245 } else if (avg_load
< min_load
) {
1246 min_load
= avg_load
;
1250 group
= group
->next
;
1251 } while (group
!= sd
->groups
);
1253 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1259 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1262 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1265 unsigned long load
, min_load
= ULONG_MAX
;
1269 /* Traverse only the allowed CPUs */
1270 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1272 for_each_cpu_mask(i
, tmp
) {
1273 load
= weighted_cpuload(i
);
1275 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1285 * sched_balance_self: balance the current task (running on cpu) in domains
1286 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1289 * Balance, ie. select the least loaded group.
1291 * Returns the target CPU number, or the same CPU if no balancing is needed.
1293 * preempt must be disabled.
1295 static int sched_balance_self(int cpu
, int flag
)
1297 struct task_struct
*t
= current
;
1298 struct sched_domain
*tmp
, *sd
= NULL
;
1300 for_each_domain(cpu
, tmp
) {
1302 * If power savings logic is enabled for a domain, stop there.
1304 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1306 if (tmp
->flags
& flag
)
1312 struct sched_group
*group
;
1313 int new_cpu
, weight
;
1315 if (!(sd
->flags
& flag
)) {
1321 group
= find_idlest_group(sd
, t
, cpu
);
1327 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1328 if (new_cpu
== -1 || new_cpu
== cpu
) {
1329 /* Now try balancing at a lower domain level of cpu */
1334 /* Now try balancing at a lower domain level of new_cpu */
1337 weight
= cpus_weight(span
);
1338 for_each_domain(cpu
, tmp
) {
1339 if (weight
<= cpus_weight(tmp
->span
))
1341 if (tmp
->flags
& flag
)
1344 /* while loop will break here if sd == NULL */
1350 #endif /* CONFIG_SMP */
1353 * wake_idle() will wake a task on an idle cpu if task->cpu is
1354 * not idle and an idle cpu is available. The span of cpus to
1355 * search starts with cpus closest then further out as needed,
1356 * so we always favor a closer, idle cpu.
1358 * Returns the CPU we should wake onto.
1360 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1361 static int wake_idle(int cpu
, struct task_struct
*p
)
1364 struct sched_domain
*sd
;
1368 * If it is idle, then it is the best cpu to run this task.
1370 * This cpu is also the best, if it has more than one task already.
1371 * Siblings must be also busy(in most cases) as they didn't already
1372 * pickup the extra load from this cpu and hence we need not check
1373 * sibling runqueue info. This will avoid the checks and cache miss
1374 * penalities associated with that.
1376 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1379 for_each_domain(cpu
, sd
) {
1380 if (sd
->flags
& SD_WAKE_IDLE
) {
1381 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1382 for_each_cpu_mask(i
, tmp
) {
1393 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1400 * try_to_wake_up - wake up a thread
1401 * @p: the to-be-woken-up thread
1402 * @state: the mask of task states that can be woken
1403 * @sync: do a synchronous wakeup?
1405 * Put it on the run-queue if it's not already there. The "current"
1406 * thread is always on the run-queue (except when the actual
1407 * re-schedule is in progress), and as such you're allowed to do
1408 * the simpler "current->state = TASK_RUNNING" to mark yourself
1409 * runnable without the overhead of this.
1411 * returns failure only if the task is already active.
1413 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1415 int cpu
, this_cpu
, success
= 0;
1416 unsigned long flags
;
1420 struct sched_domain
*sd
, *this_sd
= NULL
;
1421 unsigned long load
, this_load
;
1425 rq
= task_rq_lock(p
, &flags
);
1426 old_state
= p
->state
;
1427 if (!(old_state
& state
))
1434 this_cpu
= smp_processor_id();
1437 if (unlikely(task_running(rq
, p
)))
1442 schedstat_inc(rq
, ttwu_cnt
);
1443 if (cpu
== this_cpu
) {
1444 schedstat_inc(rq
, ttwu_local
);
1448 for_each_domain(this_cpu
, sd
) {
1449 if (cpu_isset(cpu
, sd
->span
)) {
1450 schedstat_inc(sd
, ttwu_wake_remote
);
1456 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1460 * Check for affine wakeup and passive balancing possibilities.
1463 int idx
= this_sd
->wake_idx
;
1464 unsigned int imbalance
;
1466 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1468 load
= source_load(cpu
, idx
);
1469 this_load
= target_load(this_cpu
, idx
);
1471 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1473 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1474 unsigned long tl
= this_load
;
1475 unsigned long tl_per_task
;
1477 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1480 * If sync wakeup then subtract the (maximum possible)
1481 * effect of the currently running task from the load
1482 * of the current CPU:
1485 tl
-= current
->se
.load
.weight
;
1488 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1489 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1491 * This domain has SD_WAKE_AFFINE and
1492 * p is cache cold in this domain, and
1493 * there is no bad imbalance.
1495 schedstat_inc(this_sd
, ttwu_move_affine
);
1501 * Start passive balancing when half the imbalance_pct
1504 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1505 if (imbalance
*this_load
<= 100*load
) {
1506 schedstat_inc(this_sd
, ttwu_move_balance
);
1512 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1514 new_cpu
= wake_idle(new_cpu
, p
);
1515 if (new_cpu
!= cpu
) {
1516 set_task_cpu(p
, new_cpu
);
1517 task_rq_unlock(rq
, &flags
);
1518 /* might preempt at this point */
1519 rq
= task_rq_lock(p
, &flags
);
1520 old_state
= p
->state
;
1521 if (!(old_state
& state
))
1526 this_cpu
= smp_processor_id();
1531 #endif /* CONFIG_SMP */
1532 update_rq_clock(rq
);
1533 activate_task(rq
, p
, 1);
1535 * Sync wakeups (i.e. those types of wakeups where the waker
1536 * has indicated that it will leave the CPU in short order)
1537 * don't trigger a preemption, if the woken up task will run on
1538 * this cpu. (in this case the 'I will reschedule' promise of
1539 * the waker guarantees that the freshly woken up task is going
1540 * to be considered on this CPU.)
1542 if (!sync
|| cpu
!= this_cpu
)
1543 check_preempt_curr(rq
, p
);
1547 p
->state
= TASK_RUNNING
;
1549 task_rq_unlock(rq
, &flags
);
1554 int fastcall
wake_up_process(struct task_struct
*p
)
1556 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1557 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1559 EXPORT_SYMBOL(wake_up_process
);
1561 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1563 return try_to_wake_up(p
, state
, 0);
1567 * Perform scheduler related setup for a newly forked process p.
1568 * p is forked by current.
1570 * __sched_fork() is basic setup used by init_idle() too:
1572 static void __sched_fork(struct task_struct
*p
)
1574 p
->se
.wait_start_fair
= 0;
1575 p
->se
.exec_start
= 0;
1576 p
->se
.sum_exec_runtime
= 0;
1577 p
->se
.prev_sum_exec_runtime
= 0;
1578 p
->se
.wait_runtime
= 0;
1580 #ifdef CONFIG_SCHEDSTATS
1581 p
->se
.wait_start
= 0;
1582 p
->se
.sum_wait_runtime
= 0;
1583 p
->se
.sum_sleep_runtime
= 0;
1584 p
->se
.sleep_start
= 0;
1585 p
->se
.block_start
= 0;
1586 p
->se
.sleep_max
= 0;
1587 p
->se
.block_max
= 0;
1589 p
->se
.slice_max
= 0;
1591 p
->se
.wait_runtime_overruns
= 0;
1592 p
->se
.wait_runtime_underruns
= 0;
1595 INIT_LIST_HEAD(&p
->run_list
);
1598 #ifdef CONFIG_PREEMPT_NOTIFIERS
1599 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1603 * We mark the process as running here, but have not actually
1604 * inserted it onto the runqueue yet. This guarantees that
1605 * nobody will actually run it, and a signal or other external
1606 * event cannot wake it up and insert it on the runqueue either.
1608 p
->state
= TASK_RUNNING
;
1612 * fork()/clone()-time setup:
1614 void sched_fork(struct task_struct
*p
, int clone_flags
)
1616 int cpu
= get_cpu();
1621 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1623 __set_task_cpu(p
, cpu
);
1626 * Make sure we do not leak PI boosting priority to the child:
1628 p
->prio
= current
->normal_prio
;
1630 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1631 if (likely(sched_info_on()))
1632 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1634 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1637 #ifdef CONFIG_PREEMPT
1638 /* Want to start with kernel preemption disabled. */
1639 task_thread_info(p
)->preempt_count
= 1;
1645 * wake_up_new_task - wake up a newly created task for the first time.
1647 * This function will do some initial scheduler statistics housekeeping
1648 * that must be done for every newly created context, then puts the task
1649 * on the runqueue and wakes it.
1651 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1653 unsigned long flags
;
1657 rq
= task_rq_lock(p
, &flags
);
1658 BUG_ON(p
->state
!= TASK_RUNNING
);
1659 this_cpu
= smp_processor_id(); /* parent's CPU */
1660 update_rq_clock(rq
);
1662 p
->prio
= effective_prio(p
);
1664 if (rt_prio(p
->prio
))
1665 p
->sched_class
= &rt_sched_class
;
1667 p
->sched_class
= &fair_sched_class
;
1669 if (task_cpu(p
) != this_cpu
|| !p
->sched_class
->task_new
||
1670 !current
->se
.on_rq
) {
1671 activate_task(rq
, p
, 0);
1674 * Let the scheduling class do new task startup
1675 * management (if any):
1677 p
->sched_class
->task_new(rq
, p
);
1678 inc_nr_running(p
, rq
);
1680 check_preempt_curr(rq
, p
);
1681 task_rq_unlock(rq
, &flags
);
1684 #ifdef CONFIG_PREEMPT_NOTIFIERS
1687 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1688 * @notifier: notifier struct to register
1690 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1692 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1694 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1697 * preempt_notifier_unregister - no longer interested in preemption notifications
1698 * @notifier: notifier struct to unregister
1700 * This is safe to call from within a preemption notifier.
1702 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1704 hlist_del(¬ifier
->link
);
1706 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1708 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1710 struct preempt_notifier
*notifier
;
1711 struct hlist_node
*node
;
1713 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1714 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1718 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1719 struct task_struct
*next
)
1721 struct preempt_notifier
*notifier
;
1722 struct hlist_node
*node
;
1724 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1725 notifier
->ops
->sched_out(notifier
, next
);
1730 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1735 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1736 struct task_struct
*next
)
1743 * prepare_task_switch - prepare to switch tasks
1744 * @rq: the runqueue preparing to switch
1745 * @prev: the current task that is being switched out
1746 * @next: the task we are going to switch to.
1748 * This is called with the rq lock held and interrupts off. It must
1749 * be paired with a subsequent finish_task_switch after the context
1752 * prepare_task_switch sets up locking and calls architecture specific
1756 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1757 struct task_struct
*next
)
1759 fire_sched_out_preempt_notifiers(prev
, next
);
1760 prepare_lock_switch(rq
, next
);
1761 prepare_arch_switch(next
);
1765 * finish_task_switch - clean up after a task-switch
1766 * @rq: runqueue associated with task-switch
1767 * @prev: the thread we just switched away from.
1769 * finish_task_switch must be called after the context switch, paired
1770 * with a prepare_task_switch call before the context switch.
1771 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1772 * and do any other architecture-specific cleanup actions.
1774 * Note that we may have delayed dropping an mm in context_switch(). If
1775 * so, we finish that here outside of the runqueue lock. (Doing it
1776 * with the lock held can cause deadlocks; see schedule() for
1779 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1780 __releases(rq
->lock
)
1782 struct mm_struct
*mm
= rq
->prev_mm
;
1788 * A task struct has one reference for the use as "current".
1789 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1790 * schedule one last time. The schedule call will never return, and
1791 * the scheduled task must drop that reference.
1792 * The test for TASK_DEAD must occur while the runqueue locks are
1793 * still held, otherwise prev could be scheduled on another cpu, die
1794 * there before we look at prev->state, and then the reference would
1796 * Manfred Spraul <manfred@colorfullife.com>
1798 prev_state
= prev
->state
;
1799 finish_arch_switch(prev
);
1800 finish_lock_switch(rq
, prev
);
1801 fire_sched_in_preempt_notifiers(current
);
1804 if (unlikely(prev_state
== TASK_DEAD
)) {
1806 * Remove function-return probe instances associated with this
1807 * task and put them back on the free list.
1809 kprobe_flush_task(prev
);
1810 put_task_struct(prev
);
1815 * schedule_tail - first thing a freshly forked thread must call.
1816 * @prev: the thread we just switched away from.
1818 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1819 __releases(rq
->lock
)
1821 struct rq
*rq
= this_rq();
1823 finish_task_switch(rq
, prev
);
1824 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1825 /* In this case, finish_task_switch does not reenable preemption */
1828 if (current
->set_child_tid
)
1829 put_user(current
->pid
, current
->set_child_tid
);
1833 * context_switch - switch to the new MM and the new
1834 * thread's register state.
1837 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1838 struct task_struct
*next
)
1840 struct mm_struct
*mm
, *oldmm
;
1842 prepare_task_switch(rq
, prev
, next
);
1844 oldmm
= prev
->active_mm
;
1846 * For paravirt, this is coupled with an exit in switch_to to
1847 * combine the page table reload and the switch backend into
1850 arch_enter_lazy_cpu_mode();
1852 if (unlikely(!mm
)) {
1853 next
->active_mm
= oldmm
;
1854 atomic_inc(&oldmm
->mm_count
);
1855 enter_lazy_tlb(oldmm
, next
);
1857 switch_mm(oldmm
, mm
, next
);
1859 if (unlikely(!prev
->mm
)) {
1860 prev
->active_mm
= NULL
;
1861 rq
->prev_mm
= oldmm
;
1864 * Since the runqueue lock will be released by the next
1865 * task (which is an invalid locking op but in the case
1866 * of the scheduler it's an obvious special-case), so we
1867 * do an early lockdep release here:
1869 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1870 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1873 /* Here we just switch the register state and the stack. */
1874 switch_to(prev
, next
, prev
);
1878 * this_rq must be evaluated again because prev may have moved
1879 * CPUs since it called schedule(), thus the 'rq' on its stack
1880 * frame will be invalid.
1882 finish_task_switch(this_rq(), prev
);
1886 * nr_running, nr_uninterruptible and nr_context_switches:
1888 * externally visible scheduler statistics: current number of runnable
1889 * threads, current number of uninterruptible-sleeping threads, total
1890 * number of context switches performed since bootup.
1892 unsigned long nr_running(void)
1894 unsigned long i
, sum
= 0;
1896 for_each_online_cpu(i
)
1897 sum
+= cpu_rq(i
)->nr_running
;
1902 unsigned long nr_uninterruptible(void)
1904 unsigned long i
, sum
= 0;
1906 for_each_possible_cpu(i
)
1907 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1910 * Since we read the counters lockless, it might be slightly
1911 * inaccurate. Do not allow it to go below zero though:
1913 if (unlikely((long)sum
< 0))
1919 unsigned long long nr_context_switches(void)
1922 unsigned long long sum
= 0;
1924 for_each_possible_cpu(i
)
1925 sum
+= cpu_rq(i
)->nr_switches
;
1930 unsigned long nr_iowait(void)
1932 unsigned long i
, sum
= 0;
1934 for_each_possible_cpu(i
)
1935 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1940 unsigned long nr_active(void)
1942 unsigned long i
, running
= 0, uninterruptible
= 0;
1944 for_each_online_cpu(i
) {
1945 running
+= cpu_rq(i
)->nr_running
;
1946 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1949 if (unlikely((long)uninterruptible
< 0))
1950 uninterruptible
= 0;
1952 return running
+ uninterruptible
;
1956 * Update rq->cpu_load[] statistics. This function is usually called every
1957 * scheduler tick (TICK_NSEC).
1959 static void update_cpu_load(struct rq
*this_rq
)
1961 unsigned long this_load
= this_rq
->load
.weight
;
1964 this_rq
->nr_load_updates
++;
1966 /* Update our load: */
1967 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1968 unsigned long old_load
, new_load
;
1970 /* scale is effectively 1 << i now, and >> i divides by scale */
1972 old_load
= this_rq
->cpu_load
[i
];
1973 new_load
= this_load
;
1975 * Round up the averaging division if load is increasing. This
1976 * prevents us from getting stuck on 9 if the load is 10, for
1979 if (new_load
> old_load
)
1980 new_load
+= scale
-1;
1981 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
1988 * double_rq_lock - safely lock two runqueues
1990 * Note this does not disable interrupts like task_rq_lock,
1991 * you need to do so manually before calling.
1993 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1994 __acquires(rq1
->lock
)
1995 __acquires(rq2
->lock
)
1997 BUG_ON(!irqs_disabled());
1999 spin_lock(&rq1
->lock
);
2000 __acquire(rq2
->lock
); /* Fake it out ;) */
2003 spin_lock(&rq1
->lock
);
2004 spin_lock(&rq2
->lock
);
2006 spin_lock(&rq2
->lock
);
2007 spin_lock(&rq1
->lock
);
2010 update_rq_clock(rq1
);
2011 update_rq_clock(rq2
);
2015 * double_rq_unlock - safely unlock two runqueues
2017 * Note this does not restore interrupts like task_rq_unlock,
2018 * you need to do so manually after calling.
2020 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2021 __releases(rq1
->lock
)
2022 __releases(rq2
->lock
)
2024 spin_unlock(&rq1
->lock
);
2026 spin_unlock(&rq2
->lock
);
2028 __release(rq2
->lock
);
2032 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2034 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2035 __releases(this_rq
->lock
)
2036 __acquires(busiest
->lock
)
2037 __acquires(this_rq
->lock
)
2039 if (unlikely(!irqs_disabled())) {
2040 /* printk() doesn't work good under rq->lock */
2041 spin_unlock(&this_rq
->lock
);
2044 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2045 if (busiest
< this_rq
) {
2046 spin_unlock(&this_rq
->lock
);
2047 spin_lock(&busiest
->lock
);
2048 spin_lock(&this_rq
->lock
);
2050 spin_lock(&busiest
->lock
);
2055 * If dest_cpu is allowed for this process, migrate the task to it.
2056 * This is accomplished by forcing the cpu_allowed mask to only
2057 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2058 * the cpu_allowed mask is restored.
2060 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2062 struct migration_req req
;
2063 unsigned long flags
;
2066 rq
= task_rq_lock(p
, &flags
);
2067 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2068 || unlikely(cpu_is_offline(dest_cpu
)))
2071 /* force the process onto the specified CPU */
2072 if (migrate_task(p
, dest_cpu
, &req
)) {
2073 /* Need to wait for migration thread (might exit: take ref). */
2074 struct task_struct
*mt
= rq
->migration_thread
;
2076 get_task_struct(mt
);
2077 task_rq_unlock(rq
, &flags
);
2078 wake_up_process(mt
);
2079 put_task_struct(mt
);
2080 wait_for_completion(&req
.done
);
2085 task_rq_unlock(rq
, &flags
);
2089 * sched_exec - execve() is a valuable balancing opportunity, because at
2090 * this point the task has the smallest effective memory and cache footprint.
2092 void sched_exec(void)
2094 int new_cpu
, this_cpu
= get_cpu();
2095 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2097 if (new_cpu
!= this_cpu
)
2098 sched_migrate_task(current
, new_cpu
);
2102 * pull_task - move a task from a remote runqueue to the local runqueue.
2103 * Both runqueues must be locked.
2105 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2106 struct rq
*this_rq
, int this_cpu
)
2108 deactivate_task(src_rq
, p
, 0);
2109 set_task_cpu(p
, this_cpu
);
2110 activate_task(this_rq
, p
, 0);
2112 * Note that idle threads have a prio of MAX_PRIO, for this test
2113 * to be always true for them.
2115 check_preempt_curr(this_rq
, p
);
2119 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2122 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2123 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2127 * We do not migrate tasks that are:
2128 * 1) running (obviously), or
2129 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2130 * 3) are cache-hot on their current CPU.
2132 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2136 if (task_running(rq
, p
))
2142 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2143 unsigned long max_nr_move
, unsigned long max_load_move
,
2144 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2145 int *all_pinned
, unsigned long *load_moved
,
2146 int *this_best_prio
, struct rq_iterator
*iterator
)
2148 int pulled
= 0, pinned
= 0, skip_for_load
;
2149 struct task_struct
*p
;
2150 long rem_load_move
= max_load_move
;
2152 if (max_nr_move
== 0 || max_load_move
== 0)
2158 * Start the load-balancing iterator:
2160 p
= iterator
->start(iterator
->arg
);
2165 * To help distribute high priority tasks accross CPUs we don't
2166 * skip a task if it will be the highest priority task (i.e. smallest
2167 * prio value) on its new queue regardless of its load weight
2169 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2170 SCHED_LOAD_SCALE_FUZZ
;
2171 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2172 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2173 p
= iterator
->next(iterator
->arg
);
2177 pull_task(busiest
, p
, this_rq
, this_cpu
);
2179 rem_load_move
-= p
->se
.load
.weight
;
2182 * We only want to steal up to the prescribed number of tasks
2183 * and the prescribed amount of weighted load.
2185 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2186 if (p
->prio
< *this_best_prio
)
2187 *this_best_prio
= p
->prio
;
2188 p
= iterator
->next(iterator
->arg
);
2193 * Right now, this is the only place pull_task() is called,
2194 * so we can safely collect pull_task() stats here rather than
2195 * inside pull_task().
2197 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2200 *all_pinned
= pinned
;
2201 *load_moved
= max_load_move
- rem_load_move
;
2206 * move_tasks tries to move up to max_load_move weighted load from busiest to
2207 * this_rq, as part of a balancing operation within domain "sd".
2208 * Returns 1 if successful and 0 otherwise.
2210 * Called with both runqueues locked.
2212 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2213 unsigned long max_load_move
,
2214 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2217 struct sched_class
*class = sched_class_highest
;
2218 unsigned long total_load_moved
= 0;
2219 int this_best_prio
= this_rq
->curr
->prio
;
2223 class->load_balance(this_rq
, this_cpu
, busiest
,
2224 ULONG_MAX
, max_load_move
- total_load_moved
,
2225 sd
, idle
, all_pinned
, &this_best_prio
);
2226 class = class->next
;
2227 } while (class && max_load_move
> total_load_moved
);
2229 return total_load_moved
> 0;
2233 * move_one_task tries to move exactly one task from busiest to this_rq, as
2234 * part of active balancing operations within "domain".
2235 * Returns 1 if successful and 0 otherwise.
2237 * Called with both runqueues locked.
2239 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2240 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2242 struct sched_class
*class;
2243 int this_best_prio
= MAX_PRIO
;
2245 for (class = sched_class_highest
; class; class = class->next
)
2246 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2247 1, ULONG_MAX
, sd
, idle
, NULL
,
2255 * find_busiest_group finds and returns the busiest CPU group within the
2256 * domain. It calculates and returns the amount of weighted load which
2257 * should be moved to restore balance via the imbalance parameter.
2259 static struct sched_group
*
2260 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2261 unsigned long *imbalance
, enum cpu_idle_type idle
,
2262 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2264 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2265 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2266 unsigned long max_pull
;
2267 unsigned long busiest_load_per_task
, busiest_nr_running
;
2268 unsigned long this_load_per_task
, this_nr_running
;
2270 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2271 int power_savings_balance
= 1;
2272 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2273 unsigned long min_nr_running
= ULONG_MAX
;
2274 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2277 max_load
= this_load
= total_load
= total_pwr
= 0;
2278 busiest_load_per_task
= busiest_nr_running
= 0;
2279 this_load_per_task
= this_nr_running
= 0;
2280 if (idle
== CPU_NOT_IDLE
)
2281 load_idx
= sd
->busy_idx
;
2282 else if (idle
== CPU_NEWLY_IDLE
)
2283 load_idx
= sd
->newidle_idx
;
2285 load_idx
= sd
->idle_idx
;
2288 unsigned long load
, group_capacity
;
2291 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2292 unsigned long sum_nr_running
, sum_weighted_load
;
2294 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2297 balance_cpu
= first_cpu(group
->cpumask
);
2299 /* Tally up the load of all CPUs in the group */
2300 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2302 for_each_cpu_mask(i
, group
->cpumask
) {
2305 if (!cpu_isset(i
, *cpus
))
2310 if (*sd_idle
&& rq
->nr_running
)
2313 /* Bias balancing toward cpus of our domain */
2315 if (idle_cpu(i
) && !first_idle_cpu
) {
2320 load
= target_load(i
, load_idx
);
2322 load
= source_load(i
, load_idx
);
2325 sum_nr_running
+= rq
->nr_running
;
2326 sum_weighted_load
+= weighted_cpuload(i
);
2330 * First idle cpu or the first cpu(busiest) in this sched group
2331 * is eligible for doing load balancing at this and above
2332 * domains. In the newly idle case, we will allow all the cpu's
2333 * to do the newly idle load balance.
2335 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2336 balance_cpu
!= this_cpu
&& balance
) {
2341 total_load
+= avg_load
;
2342 total_pwr
+= group
->__cpu_power
;
2344 /* Adjust by relative CPU power of the group */
2345 avg_load
= sg_div_cpu_power(group
,
2346 avg_load
* SCHED_LOAD_SCALE
);
2348 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2351 this_load
= avg_load
;
2353 this_nr_running
= sum_nr_running
;
2354 this_load_per_task
= sum_weighted_load
;
2355 } else if (avg_load
> max_load
&&
2356 sum_nr_running
> group_capacity
) {
2357 max_load
= avg_load
;
2359 busiest_nr_running
= sum_nr_running
;
2360 busiest_load_per_task
= sum_weighted_load
;
2363 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2365 * Busy processors will not participate in power savings
2368 if (idle
== CPU_NOT_IDLE
||
2369 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2373 * If the local group is idle or completely loaded
2374 * no need to do power savings balance at this domain
2376 if (local_group
&& (this_nr_running
>= group_capacity
||
2378 power_savings_balance
= 0;
2381 * If a group is already running at full capacity or idle,
2382 * don't include that group in power savings calculations
2384 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2389 * Calculate the group which has the least non-idle load.
2390 * This is the group from where we need to pick up the load
2393 if ((sum_nr_running
< min_nr_running
) ||
2394 (sum_nr_running
== min_nr_running
&&
2395 first_cpu(group
->cpumask
) <
2396 first_cpu(group_min
->cpumask
))) {
2398 min_nr_running
= sum_nr_running
;
2399 min_load_per_task
= sum_weighted_load
/
2404 * Calculate the group which is almost near its
2405 * capacity but still has some space to pick up some load
2406 * from other group and save more power
2408 if (sum_nr_running
<= group_capacity
- 1) {
2409 if (sum_nr_running
> leader_nr_running
||
2410 (sum_nr_running
== leader_nr_running
&&
2411 first_cpu(group
->cpumask
) >
2412 first_cpu(group_leader
->cpumask
))) {
2413 group_leader
= group
;
2414 leader_nr_running
= sum_nr_running
;
2419 group
= group
->next
;
2420 } while (group
!= sd
->groups
);
2422 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2425 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2427 if (this_load
>= avg_load
||
2428 100*max_load
<= sd
->imbalance_pct
*this_load
)
2431 busiest_load_per_task
/= busiest_nr_running
;
2433 * We're trying to get all the cpus to the average_load, so we don't
2434 * want to push ourselves above the average load, nor do we wish to
2435 * reduce the max loaded cpu below the average load, as either of these
2436 * actions would just result in more rebalancing later, and ping-pong
2437 * tasks around. Thus we look for the minimum possible imbalance.
2438 * Negative imbalances (*we* are more loaded than anyone else) will
2439 * be counted as no imbalance for these purposes -- we can't fix that
2440 * by pulling tasks to us. Be careful of negative numbers as they'll
2441 * appear as very large values with unsigned longs.
2443 if (max_load
<= busiest_load_per_task
)
2447 * In the presence of smp nice balancing, certain scenarios can have
2448 * max load less than avg load(as we skip the groups at or below
2449 * its cpu_power, while calculating max_load..)
2451 if (max_load
< avg_load
) {
2453 goto small_imbalance
;
2456 /* Don't want to pull so many tasks that a group would go idle */
2457 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2459 /* How much load to actually move to equalise the imbalance */
2460 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2461 (avg_load
- this_load
) * this->__cpu_power
)
2465 * if *imbalance is less than the average load per runnable task
2466 * there is no gaurantee that any tasks will be moved so we'll have
2467 * a think about bumping its value to force at least one task to be
2470 if (*imbalance
< busiest_load_per_task
) {
2471 unsigned long tmp
, pwr_now
, pwr_move
;
2475 pwr_move
= pwr_now
= 0;
2477 if (this_nr_running
) {
2478 this_load_per_task
/= this_nr_running
;
2479 if (busiest_load_per_task
> this_load_per_task
)
2482 this_load_per_task
= SCHED_LOAD_SCALE
;
2484 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2485 busiest_load_per_task
* imbn
) {
2486 *imbalance
= busiest_load_per_task
;
2491 * OK, we don't have enough imbalance to justify moving tasks,
2492 * however we may be able to increase total CPU power used by
2496 pwr_now
+= busiest
->__cpu_power
*
2497 min(busiest_load_per_task
, max_load
);
2498 pwr_now
+= this->__cpu_power
*
2499 min(this_load_per_task
, this_load
);
2500 pwr_now
/= SCHED_LOAD_SCALE
;
2502 /* Amount of load we'd subtract */
2503 tmp
= sg_div_cpu_power(busiest
,
2504 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2506 pwr_move
+= busiest
->__cpu_power
*
2507 min(busiest_load_per_task
, max_load
- tmp
);
2509 /* Amount of load we'd add */
2510 if (max_load
* busiest
->__cpu_power
<
2511 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2512 tmp
= sg_div_cpu_power(this,
2513 max_load
* busiest
->__cpu_power
);
2515 tmp
= sg_div_cpu_power(this,
2516 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2517 pwr_move
+= this->__cpu_power
*
2518 min(this_load_per_task
, this_load
+ tmp
);
2519 pwr_move
/= SCHED_LOAD_SCALE
;
2521 /* Move if we gain throughput */
2522 if (pwr_move
> pwr_now
)
2523 *imbalance
= busiest_load_per_task
;
2529 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2530 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2533 if (this == group_leader
&& group_leader
!= group_min
) {
2534 *imbalance
= min_load_per_task
;
2544 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2547 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2548 unsigned long imbalance
, cpumask_t
*cpus
)
2550 struct rq
*busiest
= NULL
, *rq
;
2551 unsigned long max_load
= 0;
2554 for_each_cpu_mask(i
, group
->cpumask
) {
2557 if (!cpu_isset(i
, *cpus
))
2561 wl
= weighted_cpuload(i
);
2563 if (rq
->nr_running
== 1 && wl
> imbalance
)
2566 if (wl
> max_load
) {
2576 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2577 * so long as it is large enough.
2579 #define MAX_PINNED_INTERVAL 512
2582 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2583 * tasks if there is an imbalance.
2585 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2586 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2589 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2590 struct sched_group
*group
;
2591 unsigned long imbalance
;
2593 cpumask_t cpus
= CPU_MASK_ALL
;
2594 unsigned long flags
;
2597 * When power savings policy is enabled for the parent domain, idle
2598 * sibling can pick up load irrespective of busy siblings. In this case,
2599 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2600 * portraying it as CPU_NOT_IDLE.
2602 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2603 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2606 schedstat_inc(sd
, lb_cnt
[idle
]);
2609 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2616 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2620 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2622 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2626 BUG_ON(busiest
== this_rq
);
2628 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2631 if (busiest
->nr_running
> 1) {
2633 * Attempt to move tasks. If find_busiest_group has found
2634 * an imbalance but busiest->nr_running <= 1, the group is
2635 * still unbalanced. ld_moved simply stays zero, so it is
2636 * correctly treated as an imbalance.
2638 local_irq_save(flags
);
2639 double_rq_lock(this_rq
, busiest
);
2640 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2641 imbalance
, sd
, idle
, &all_pinned
);
2642 double_rq_unlock(this_rq
, busiest
);
2643 local_irq_restore(flags
);
2646 * some other cpu did the load balance for us.
2648 if (ld_moved
&& this_cpu
!= smp_processor_id())
2649 resched_cpu(this_cpu
);
2651 /* All tasks on this runqueue were pinned by CPU affinity */
2652 if (unlikely(all_pinned
)) {
2653 cpu_clear(cpu_of(busiest
), cpus
);
2654 if (!cpus_empty(cpus
))
2661 schedstat_inc(sd
, lb_failed
[idle
]);
2662 sd
->nr_balance_failed
++;
2664 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2666 spin_lock_irqsave(&busiest
->lock
, flags
);
2668 /* don't kick the migration_thread, if the curr
2669 * task on busiest cpu can't be moved to this_cpu
2671 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2672 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2674 goto out_one_pinned
;
2677 if (!busiest
->active_balance
) {
2678 busiest
->active_balance
= 1;
2679 busiest
->push_cpu
= this_cpu
;
2682 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2684 wake_up_process(busiest
->migration_thread
);
2687 * We've kicked active balancing, reset the failure
2690 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2693 sd
->nr_balance_failed
= 0;
2695 if (likely(!active_balance
)) {
2696 /* We were unbalanced, so reset the balancing interval */
2697 sd
->balance_interval
= sd
->min_interval
;
2700 * If we've begun active balancing, start to back off. This
2701 * case may not be covered by the all_pinned logic if there
2702 * is only 1 task on the busy runqueue (because we don't call
2705 if (sd
->balance_interval
< sd
->max_interval
)
2706 sd
->balance_interval
*= 2;
2709 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2710 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2715 schedstat_inc(sd
, lb_balanced
[idle
]);
2717 sd
->nr_balance_failed
= 0;
2720 /* tune up the balancing interval */
2721 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2722 (sd
->balance_interval
< sd
->max_interval
))
2723 sd
->balance_interval
*= 2;
2725 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2726 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2732 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2733 * tasks if there is an imbalance.
2735 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2736 * this_rq is locked.
2739 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2741 struct sched_group
*group
;
2742 struct rq
*busiest
= NULL
;
2743 unsigned long imbalance
;
2747 cpumask_t cpus
= CPU_MASK_ALL
;
2750 * When power savings policy is enabled for the parent domain, idle
2751 * sibling can pick up load irrespective of busy siblings. In this case,
2752 * let the state of idle sibling percolate up as IDLE, instead of
2753 * portraying it as CPU_NOT_IDLE.
2755 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2756 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2759 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2761 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2762 &sd_idle
, &cpus
, NULL
);
2764 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2768 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2771 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2775 BUG_ON(busiest
== this_rq
);
2777 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2780 if (busiest
->nr_running
> 1) {
2781 /* Attempt to move tasks */
2782 double_lock_balance(this_rq
, busiest
);
2783 /* this_rq->clock is already updated */
2784 update_rq_clock(busiest
);
2785 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2786 imbalance
, sd
, CPU_NEWLY_IDLE
,
2788 spin_unlock(&busiest
->lock
);
2790 if (unlikely(all_pinned
)) {
2791 cpu_clear(cpu_of(busiest
), cpus
);
2792 if (!cpus_empty(cpus
))
2798 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2799 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2800 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2803 sd
->nr_balance_failed
= 0;
2808 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2809 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2810 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2812 sd
->nr_balance_failed
= 0;
2818 * idle_balance is called by schedule() if this_cpu is about to become
2819 * idle. Attempts to pull tasks from other CPUs.
2821 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2823 struct sched_domain
*sd
;
2824 int pulled_task
= -1;
2825 unsigned long next_balance
= jiffies
+ HZ
;
2827 for_each_domain(this_cpu
, sd
) {
2828 unsigned long interval
;
2830 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2833 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2834 /* If we've pulled tasks over stop searching: */
2835 pulled_task
= load_balance_newidle(this_cpu
,
2838 interval
= msecs_to_jiffies(sd
->balance_interval
);
2839 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2840 next_balance
= sd
->last_balance
+ interval
;
2844 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2846 * We are going idle. next_balance may be set based on
2847 * a busy processor. So reset next_balance.
2849 this_rq
->next_balance
= next_balance
;
2854 * active_load_balance is run by migration threads. It pushes running tasks
2855 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2856 * running on each physical CPU where possible, and avoids physical /
2857 * logical imbalances.
2859 * Called with busiest_rq locked.
2861 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2863 int target_cpu
= busiest_rq
->push_cpu
;
2864 struct sched_domain
*sd
;
2865 struct rq
*target_rq
;
2867 /* Is there any task to move? */
2868 if (busiest_rq
->nr_running
<= 1)
2871 target_rq
= cpu_rq(target_cpu
);
2874 * This condition is "impossible", if it occurs
2875 * we need to fix it. Originally reported by
2876 * Bjorn Helgaas on a 128-cpu setup.
2878 BUG_ON(busiest_rq
== target_rq
);
2880 /* move a task from busiest_rq to target_rq */
2881 double_lock_balance(busiest_rq
, target_rq
);
2882 update_rq_clock(busiest_rq
);
2883 update_rq_clock(target_rq
);
2885 /* Search for an sd spanning us and the target CPU. */
2886 for_each_domain(target_cpu
, sd
) {
2887 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2888 cpu_isset(busiest_cpu
, sd
->span
))
2893 schedstat_inc(sd
, alb_cnt
);
2895 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2897 schedstat_inc(sd
, alb_pushed
);
2899 schedstat_inc(sd
, alb_failed
);
2901 spin_unlock(&target_rq
->lock
);
2906 atomic_t load_balancer
;
2908 } nohz ____cacheline_aligned
= {
2909 .load_balancer
= ATOMIC_INIT(-1),
2910 .cpu_mask
= CPU_MASK_NONE
,
2914 * This routine will try to nominate the ilb (idle load balancing)
2915 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2916 * load balancing on behalf of all those cpus. If all the cpus in the system
2917 * go into this tickless mode, then there will be no ilb owner (as there is
2918 * no need for one) and all the cpus will sleep till the next wakeup event
2921 * For the ilb owner, tick is not stopped. And this tick will be used
2922 * for idle load balancing. ilb owner will still be part of
2925 * While stopping the tick, this cpu will become the ilb owner if there
2926 * is no other owner. And will be the owner till that cpu becomes busy
2927 * or if all cpus in the system stop their ticks at which point
2928 * there is no need for ilb owner.
2930 * When the ilb owner becomes busy, it nominates another owner, during the
2931 * next busy scheduler_tick()
2933 int select_nohz_load_balancer(int stop_tick
)
2935 int cpu
= smp_processor_id();
2938 cpu_set(cpu
, nohz
.cpu_mask
);
2939 cpu_rq(cpu
)->in_nohz_recently
= 1;
2942 * If we are going offline and still the leader, give up!
2944 if (cpu_is_offline(cpu
) &&
2945 atomic_read(&nohz
.load_balancer
) == cpu
) {
2946 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2951 /* time for ilb owner also to sleep */
2952 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2953 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2954 atomic_set(&nohz
.load_balancer
, -1);
2958 if (atomic_read(&nohz
.load_balancer
) == -1) {
2959 /* make me the ilb owner */
2960 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2962 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2965 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2968 cpu_clear(cpu
, nohz
.cpu_mask
);
2970 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2971 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2978 static DEFINE_SPINLOCK(balancing
);
2981 * It checks each scheduling domain to see if it is due to be balanced,
2982 * and initiates a balancing operation if so.
2984 * Balancing parameters are set up in arch_init_sched_domains.
2986 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
2989 struct rq
*rq
= cpu_rq(cpu
);
2990 unsigned long interval
;
2991 struct sched_domain
*sd
;
2992 /* Earliest time when we have to do rebalance again */
2993 unsigned long next_balance
= jiffies
+ 60*HZ
;
2994 int update_next_balance
= 0;
2996 for_each_domain(cpu
, sd
) {
2997 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3000 interval
= sd
->balance_interval
;
3001 if (idle
!= CPU_IDLE
)
3002 interval
*= sd
->busy_factor
;
3004 /* scale ms to jiffies */
3005 interval
= msecs_to_jiffies(interval
);
3006 if (unlikely(!interval
))
3008 if (interval
> HZ
*NR_CPUS
/10)
3009 interval
= HZ
*NR_CPUS
/10;
3012 if (sd
->flags
& SD_SERIALIZE
) {
3013 if (!spin_trylock(&balancing
))
3017 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3018 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3020 * We've pulled tasks over so either we're no
3021 * longer idle, or one of our SMT siblings is
3024 idle
= CPU_NOT_IDLE
;
3026 sd
->last_balance
= jiffies
;
3028 if (sd
->flags
& SD_SERIALIZE
)
3029 spin_unlock(&balancing
);
3031 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3032 next_balance
= sd
->last_balance
+ interval
;
3033 update_next_balance
= 1;
3037 * Stop the load balance at this level. There is another
3038 * CPU in our sched group which is doing load balancing more
3046 * next_balance will be updated only when there is a need.
3047 * When the cpu is attached to null domain for ex, it will not be
3050 if (likely(update_next_balance
))
3051 rq
->next_balance
= next_balance
;
3055 * run_rebalance_domains is triggered when needed from the scheduler tick.
3056 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3057 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3059 static void run_rebalance_domains(struct softirq_action
*h
)
3061 int this_cpu
= smp_processor_id();
3062 struct rq
*this_rq
= cpu_rq(this_cpu
);
3063 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3064 CPU_IDLE
: CPU_NOT_IDLE
;
3066 rebalance_domains(this_cpu
, idle
);
3070 * If this cpu is the owner for idle load balancing, then do the
3071 * balancing on behalf of the other idle cpus whose ticks are
3074 if (this_rq
->idle_at_tick
&&
3075 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3076 cpumask_t cpus
= nohz
.cpu_mask
;
3080 cpu_clear(this_cpu
, cpus
);
3081 for_each_cpu_mask(balance_cpu
, cpus
) {
3083 * If this cpu gets work to do, stop the load balancing
3084 * work being done for other cpus. Next load
3085 * balancing owner will pick it up.
3090 rebalance_domains(balance_cpu
, CPU_IDLE
);
3092 rq
= cpu_rq(balance_cpu
);
3093 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3094 this_rq
->next_balance
= rq
->next_balance
;
3101 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3103 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3104 * idle load balancing owner or decide to stop the periodic load balancing,
3105 * if the whole system is idle.
3107 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3111 * If we were in the nohz mode recently and busy at the current
3112 * scheduler tick, then check if we need to nominate new idle
3115 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3116 rq
->in_nohz_recently
= 0;
3118 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3119 cpu_clear(cpu
, nohz
.cpu_mask
);
3120 atomic_set(&nohz
.load_balancer
, -1);
3123 if (atomic_read(&nohz
.load_balancer
) == -1) {
3125 * simple selection for now: Nominate the
3126 * first cpu in the nohz list to be the next
3129 * TBD: Traverse the sched domains and nominate
3130 * the nearest cpu in the nohz.cpu_mask.
3132 int ilb
= first_cpu(nohz
.cpu_mask
);
3140 * If this cpu is idle and doing idle load balancing for all the
3141 * cpus with ticks stopped, is it time for that to stop?
3143 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3144 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3150 * If this cpu is idle and the idle load balancing is done by
3151 * someone else, then no need raise the SCHED_SOFTIRQ
3153 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3154 cpu_isset(cpu
, nohz
.cpu_mask
))
3157 if (time_after_eq(jiffies
, rq
->next_balance
))
3158 raise_softirq(SCHED_SOFTIRQ
);
3161 #else /* CONFIG_SMP */
3164 * on UP we do not need to balance between CPUs:
3166 static inline void idle_balance(int cpu
, struct rq
*rq
)
3170 /* Avoid "used but not defined" warning on UP */
3171 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3172 unsigned long max_nr_move
, unsigned long max_load_move
,
3173 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3174 int *all_pinned
, unsigned long *load_moved
,
3175 int *this_best_prio
, struct rq_iterator
*iterator
)
3184 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3186 EXPORT_PER_CPU_SYMBOL(kstat
);
3189 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3190 * that have not yet been banked in case the task is currently running.
3192 unsigned long long task_sched_runtime(struct task_struct
*p
)
3194 unsigned long flags
;
3198 rq
= task_rq_lock(p
, &flags
);
3199 ns
= p
->se
.sum_exec_runtime
;
3200 if (rq
->curr
== p
) {
3201 update_rq_clock(rq
);
3202 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3203 if ((s64
)delta_exec
> 0)
3206 task_rq_unlock(rq
, &flags
);
3212 * Account user cpu time to a process.
3213 * @p: the process that the cpu time gets accounted to
3214 * @hardirq_offset: the offset to subtract from hardirq_count()
3215 * @cputime: the cpu time spent in user space since the last update
3217 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3219 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3222 p
->utime
= cputime_add(p
->utime
, cputime
);
3224 /* Add user time to cpustat. */
3225 tmp
= cputime_to_cputime64(cputime
);
3226 if (TASK_NICE(p
) > 0)
3227 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3229 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3233 * Account system cpu time to a process.
3234 * @p: the process that the cpu time gets accounted to
3235 * @hardirq_offset: the offset to subtract from hardirq_count()
3236 * @cputime: the cpu time spent in kernel space since the last update
3238 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3241 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3242 struct rq
*rq
= this_rq();
3245 p
->stime
= cputime_add(p
->stime
, cputime
);
3247 /* Add system time to cpustat. */
3248 tmp
= cputime_to_cputime64(cputime
);
3249 if (hardirq_count() - hardirq_offset
)
3250 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3251 else if (softirq_count())
3252 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3253 else if (p
!= rq
->idle
)
3254 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3255 else if (atomic_read(&rq
->nr_iowait
) > 0)
3256 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3258 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3259 /* Account for system time used */
3260 acct_update_integrals(p
);
3264 * Account for involuntary wait time.
3265 * @p: the process from which the cpu time has been stolen
3266 * @steal: the cpu time spent in involuntary wait
3268 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3270 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3271 cputime64_t tmp
= cputime_to_cputime64(steal
);
3272 struct rq
*rq
= this_rq();
3274 if (p
== rq
->idle
) {
3275 p
->stime
= cputime_add(p
->stime
, steal
);
3276 if (atomic_read(&rq
->nr_iowait
) > 0)
3277 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3279 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3281 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3285 * This function gets called by the timer code, with HZ frequency.
3286 * We call it with interrupts disabled.
3288 * It also gets called by the fork code, when changing the parent's
3291 void scheduler_tick(void)
3293 int cpu
= smp_processor_id();
3294 struct rq
*rq
= cpu_rq(cpu
);
3295 struct task_struct
*curr
= rq
->curr
;
3296 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3298 spin_lock(&rq
->lock
);
3299 __update_rq_clock(rq
);
3301 * Let rq->clock advance by at least TICK_NSEC:
3303 if (unlikely(rq
->clock
< next_tick
))
3304 rq
->clock
= next_tick
;
3305 rq
->tick_timestamp
= rq
->clock
;
3306 update_cpu_load(rq
);
3307 if (curr
!= rq
->idle
) /* FIXME: needed? */
3308 curr
->sched_class
->task_tick(rq
, curr
);
3309 spin_unlock(&rq
->lock
);
3312 rq
->idle_at_tick
= idle_cpu(cpu
);
3313 trigger_load_balance(rq
, cpu
);
3317 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3319 void fastcall
add_preempt_count(int val
)
3324 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3326 preempt_count() += val
;
3328 * Spinlock count overflowing soon?
3330 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3333 EXPORT_SYMBOL(add_preempt_count
);
3335 void fastcall
sub_preempt_count(int val
)
3340 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3343 * Is the spinlock portion underflowing?
3345 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3346 !(preempt_count() & PREEMPT_MASK
)))
3349 preempt_count() -= val
;
3351 EXPORT_SYMBOL(sub_preempt_count
);
3356 * Print scheduling while atomic bug:
3358 static noinline
void __schedule_bug(struct task_struct
*prev
)
3360 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3361 prev
->comm
, preempt_count(), prev
->pid
);
3362 debug_show_held_locks(prev
);
3363 if (irqs_disabled())
3364 print_irqtrace_events(prev
);
3369 * Various schedule()-time debugging checks and statistics:
3371 static inline void schedule_debug(struct task_struct
*prev
)
3374 * Test if we are atomic. Since do_exit() needs to call into
3375 * schedule() atomically, we ignore that path for now.
3376 * Otherwise, whine if we are scheduling when we should not be.
3378 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3379 __schedule_bug(prev
);
3381 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3383 schedstat_inc(this_rq(), sched_cnt
);
3387 * Pick up the highest-prio task:
3389 static inline struct task_struct
*
3390 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3392 struct sched_class
*class;
3393 struct task_struct
*p
;
3396 * Optimization: we know that if all tasks are in
3397 * the fair class we can call that function directly:
3399 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3400 p
= fair_sched_class
.pick_next_task(rq
);
3405 class = sched_class_highest
;
3407 p
= class->pick_next_task(rq
);
3411 * Will never be NULL as the idle class always
3412 * returns a non-NULL p:
3414 class = class->next
;
3419 * schedule() is the main scheduler function.
3421 asmlinkage
void __sched
schedule(void)
3423 struct task_struct
*prev
, *next
;
3430 cpu
= smp_processor_id();
3434 switch_count
= &prev
->nivcsw
;
3436 release_kernel_lock(prev
);
3437 need_resched_nonpreemptible
:
3439 schedule_debug(prev
);
3441 spin_lock_irq(&rq
->lock
);
3442 clear_tsk_need_resched(prev
);
3443 __update_rq_clock(rq
);
3445 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3446 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3447 unlikely(signal_pending(prev
)))) {
3448 prev
->state
= TASK_RUNNING
;
3450 deactivate_task(rq
, prev
, 1);
3452 switch_count
= &prev
->nvcsw
;
3455 if (unlikely(!rq
->nr_running
))
3456 idle_balance(cpu
, rq
);
3458 prev
->sched_class
->put_prev_task(rq
, prev
);
3459 next
= pick_next_task(rq
, prev
);
3461 sched_info_switch(prev
, next
);
3463 if (likely(prev
!= next
)) {
3468 context_switch(rq
, prev
, next
); /* unlocks the rq */
3470 spin_unlock_irq(&rq
->lock
);
3472 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3473 cpu
= smp_processor_id();
3475 goto need_resched_nonpreemptible
;
3477 preempt_enable_no_resched();
3478 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3481 EXPORT_SYMBOL(schedule
);
3483 #ifdef CONFIG_PREEMPT
3485 * this is the entry point to schedule() from in-kernel preemption
3486 * off of preempt_enable. Kernel preemptions off return from interrupt
3487 * occur there and call schedule directly.
3489 asmlinkage
void __sched
preempt_schedule(void)
3491 struct thread_info
*ti
= current_thread_info();
3492 #ifdef CONFIG_PREEMPT_BKL
3493 struct task_struct
*task
= current
;
3494 int saved_lock_depth
;
3497 * If there is a non-zero preempt_count or interrupts are disabled,
3498 * we do not want to preempt the current task. Just return..
3500 if (likely(ti
->preempt_count
|| irqs_disabled()))
3504 add_preempt_count(PREEMPT_ACTIVE
);
3506 * We keep the big kernel semaphore locked, but we
3507 * clear ->lock_depth so that schedule() doesnt
3508 * auto-release the semaphore:
3510 #ifdef CONFIG_PREEMPT_BKL
3511 saved_lock_depth
= task
->lock_depth
;
3512 task
->lock_depth
= -1;
3515 #ifdef CONFIG_PREEMPT_BKL
3516 task
->lock_depth
= saved_lock_depth
;
3518 sub_preempt_count(PREEMPT_ACTIVE
);
3520 /* we could miss a preemption opportunity between schedule and now */
3522 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3525 EXPORT_SYMBOL(preempt_schedule
);
3528 * this is the entry point to schedule() from kernel preemption
3529 * off of irq context.
3530 * Note, that this is called and return with irqs disabled. This will
3531 * protect us against recursive calling from irq.
3533 asmlinkage
void __sched
preempt_schedule_irq(void)
3535 struct thread_info
*ti
= current_thread_info();
3536 #ifdef CONFIG_PREEMPT_BKL
3537 struct task_struct
*task
= current
;
3538 int saved_lock_depth
;
3540 /* Catch callers which need to be fixed */
3541 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3544 add_preempt_count(PREEMPT_ACTIVE
);
3546 * We keep the big kernel semaphore locked, but we
3547 * clear ->lock_depth so that schedule() doesnt
3548 * auto-release the semaphore:
3550 #ifdef CONFIG_PREEMPT_BKL
3551 saved_lock_depth
= task
->lock_depth
;
3552 task
->lock_depth
= -1;
3556 local_irq_disable();
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
)))
3568 #endif /* CONFIG_PREEMPT */
3570 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3573 return try_to_wake_up(curr
->private, mode
, sync
);
3575 EXPORT_SYMBOL(default_wake_function
);
3578 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3579 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3580 * number) then we wake all the non-exclusive tasks and one exclusive task.
3582 * There are circumstances in which we can try to wake a task which has already
3583 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3584 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3586 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3587 int nr_exclusive
, int sync
, void *key
)
3589 wait_queue_t
*curr
, *next
;
3591 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3592 unsigned flags
= curr
->flags
;
3594 if (curr
->func(curr
, mode
, sync
, key
) &&
3595 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3601 * __wake_up - wake up threads blocked on a waitqueue.
3603 * @mode: which threads
3604 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3605 * @key: is directly passed to the wakeup function
3607 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3608 int nr_exclusive
, void *key
)
3610 unsigned long flags
;
3612 spin_lock_irqsave(&q
->lock
, flags
);
3613 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3614 spin_unlock_irqrestore(&q
->lock
, flags
);
3616 EXPORT_SYMBOL(__wake_up
);
3619 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3621 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3623 __wake_up_common(q
, mode
, 1, 0, NULL
);
3627 * __wake_up_sync - wake up threads blocked on a waitqueue.
3629 * @mode: which threads
3630 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3632 * The sync wakeup differs that the waker knows that it will schedule
3633 * away soon, so while the target thread will be woken up, it will not
3634 * be migrated to another CPU - ie. the two threads are 'synchronized'
3635 * with each other. This can prevent needless bouncing between CPUs.
3637 * On UP it can prevent extra preemption.
3640 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3642 unsigned long flags
;
3648 if (unlikely(!nr_exclusive
))
3651 spin_lock_irqsave(&q
->lock
, flags
);
3652 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3653 spin_unlock_irqrestore(&q
->lock
, flags
);
3655 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3657 void fastcall
complete(struct completion
*x
)
3659 unsigned long flags
;
3661 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3663 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3665 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3667 EXPORT_SYMBOL(complete
);
3669 void fastcall
complete_all(struct completion
*x
)
3671 unsigned long flags
;
3673 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3674 x
->done
+= UINT_MAX
/2;
3675 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3677 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3679 EXPORT_SYMBOL(complete_all
);
3681 void fastcall __sched
wait_for_completion(struct completion
*x
)
3685 spin_lock_irq(&x
->wait
.lock
);
3687 DECLARE_WAITQUEUE(wait
, current
);
3689 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3690 __add_wait_queue_tail(&x
->wait
, &wait
);
3692 __set_current_state(TASK_UNINTERRUPTIBLE
);
3693 spin_unlock_irq(&x
->wait
.lock
);
3695 spin_lock_irq(&x
->wait
.lock
);
3697 __remove_wait_queue(&x
->wait
, &wait
);
3700 spin_unlock_irq(&x
->wait
.lock
);
3702 EXPORT_SYMBOL(wait_for_completion
);
3704 unsigned long fastcall __sched
3705 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3709 spin_lock_irq(&x
->wait
.lock
);
3711 DECLARE_WAITQUEUE(wait
, current
);
3713 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3714 __add_wait_queue_tail(&x
->wait
, &wait
);
3716 __set_current_state(TASK_UNINTERRUPTIBLE
);
3717 spin_unlock_irq(&x
->wait
.lock
);
3718 timeout
= schedule_timeout(timeout
);
3719 spin_lock_irq(&x
->wait
.lock
);
3721 __remove_wait_queue(&x
->wait
, &wait
);
3725 __remove_wait_queue(&x
->wait
, &wait
);
3729 spin_unlock_irq(&x
->wait
.lock
);
3732 EXPORT_SYMBOL(wait_for_completion_timeout
);
3734 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3740 spin_lock_irq(&x
->wait
.lock
);
3742 DECLARE_WAITQUEUE(wait
, current
);
3744 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3745 __add_wait_queue_tail(&x
->wait
, &wait
);
3747 if (signal_pending(current
)) {
3749 __remove_wait_queue(&x
->wait
, &wait
);
3752 __set_current_state(TASK_INTERRUPTIBLE
);
3753 spin_unlock_irq(&x
->wait
.lock
);
3755 spin_lock_irq(&x
->wait
.lock
);
3757 __remove_wait_queue(&x
->wait
, &wait
);
3761 spin_unlock_irq(&x
->wait
.lock
);
3765 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3767 unsigned long fastcall __sched
3768 wait_for_completion_interruptible_timeout(struct completion
*x
,
3769 unsigned long timeout
)
3773 spin_lock_irq(&x
->wait
.lock
);
3775 DECLARE_WAITQUEUE(wait
, current
);
3777 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3778 __add_wait_queue_tail(&x
->wait
, &wait
);
3780 if (signal_pending(current
)) {
3781 timeout
= -ERESTARTSYS
;
3782 __remove_wait_queue(&x
->wait
, &wait
);
3785 __set_current_state(TASK_INTERRUPTIBLE
);
3786 spin_unlock_irq(&x
->wait
.lock
);
3787 timeout
= schedule_timeout(timeout
);
3788 spin_lock_irq(&x
->wait
.lock
);
3790 __remove_wait_queue(&x
->wait
, &wait
);
3794 __remove_wait_queue(&x
->wait
, &wait
);
3798 spin_unlock_irq(&x
->wait
.lock
);
3801 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3804 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3806 spin_lock_irqsave(&q
->lock
, *flags
);
3807 __add_wait_queue(q
, wait
);
3808 spin_unlock(&q
->lock
);
3812 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3814 spin_lock_irq(&q
->lock
);
3815 __remove_wait_queue(q
, wait
);
3816 spin_unlock_irqrestore(&q
->lock
, *flags
);
3819 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3821 unsigned long flags
;
3824 init_waitqueue_entry(&wait
, current
);
3826 current
->state
= TASK_INTERRUPTIBLE
;
3828 sleep_on_head(q
, &wait
, &flags
);
3830 sleep_on_tail(q
, &wait
, &flags
);
3832 EXPORT_SYMBOL(interruptible_sleep_on
);
3835 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3837 unsigned long flags
;
3840 init_waitqueue_entry(&wait
, current
);
3842 current
->state
= TASK_INTERRUPTIBLE
;
3844 sleep_on_head(q
, &wait
, &flags
);
3845 timeout
= schedule_timeout(timeout
);
3846 sleep_on_tail(q
, &wait
, &flags
);
3850 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3852 void __sched
sleep_on(wait_queue_head_t
*q
)
3854 unsigned long flags
;
3857 init_waitqueue_entry(&wait
, current
);
3859 current
->state
= TASK_UNINTERRUPTIBLE
;
3861 sleep_on_head(q
, &wait
, &flags
);
3863 sleep_on_tail(q
, &wait
, &flags
);
3865 EXPORT_SYMBOL(sleep_on
);
3867 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3869 unsigned long flags
;
3872 init_waitqueue_entry(&wait
, current
);
3874 current
->state
= TASK_UNINTERRUPTIBLE
;
3876 sleep_on_head(q
, &wait
, &flags
);
3877 timeout
= schedule_timeout(timeout
);
3878 sleep_on_tail(q
, &wait
, &flags
);
3882 EXPORT_SYMBOL(sleep_on_timeout
);
3884 #ifdef CONFIG_RT_MUTEXES
3887 * rt_mutex_setprio - set the current priority of a task
3889 * @prio: prio value (kernel-internal form)
3891 * This function changes the 'effective' priority of a task. It does
3892 * not touch ->normal_prio like __setscheduler().
3894 * Used by the rt_mutex code to implement priority inheritance logic.
3896 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3898 unsigned long flags
;
3902 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3904 rq
= task_rq_lock(p
, &flags
);
3905 update_rq_clock(rq
);
3908 on_rq
= p
->se
.on_rq
;
3910 dequeue_task(rq
, p
, 0);
3913 p
->sched_class
= &rt_sched_class
;
3915 p
->sched_class
= &fair_sched_class
;
3920 enqueue_task(rq
, p
, 0);
3922 * Reschedule if we are currently running on this runqueue and
3923 * our priority decreased, or if we are not currently running on
3924 * this runqueue and our priority is higher than the current's
3926 if (task_running(rq
, p
)) {
3927 if (p
->prio
> oldprio
)
3928 resched_task(rq
->curr
);
3930 check_preempt_curr(rq
, p
);
3933 task_rq_unlock(rq
, &flags
);
3938 void set_user_nice(struct task_struct
*p
, long nice
)
3940 int old_prio
, delta
, on_rq
;
3941 unsigned long flags
;
3944 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3947 * We have to be careful, if called from sys_setpriority(),
3948 * the task might be in the middle of scheduling on another CPU.
3950 rq
= task_rq_lock(p
, &flags
);
3951 update_rq_clock(rq
);
3953 * The RT priorities are set via sched_setscheduler(), but we still
3954 * allow the 'normal' nice value to be set - but as expected
3955 * it wont have any effect on scheduling until the task is
3956 * SCHED_FIFO/SCHED_RR:
3958 if (task_has_rt_policy(p
)) {
3959 p
->static_prio
= NICE_TO_PRIO(nice
);
3962 on_rq
= p
->se
.on_rq
;
3964 dequeue_task(rq
, p
, 0);
3968 p
->static_prio
= NICE_TO_PRIO(nice
);
3971 p
->prio
= effective_prio(p
);
3972 delta
= p
->prio
- old_prio
;
3975 enqueue_task(rq
, p
, 0);
3978 * If the task increased its priority or is running and
3979 * lowered its priority, then reschedule its CPU:
3981 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3982 resched_task(rq
->curr
);
3985 task_rq_unlock(rq
, &flags
);
3987 EXPORT_SYMBOL(set_user_nice
);
3990 * can_nice - check if a task can reduce its nice value
3994 int can_nice(const struct task_struct
*p
, const int nice
)
3996 /* convert nice value [19,-20] to rlimit style value [1,40] */
3997 int nice_rlim
= 20 - nice
;
3999 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4000 capable(CAP_SYS_NICE
));
4003 #ifdef __ARCH_WANT_SYS_NICE
4006 * sys_nice - change the priority of the current process.
4007 * @increment: priority increment
4009 * sys_setpriority is a more generic, but much slower function that
4010 * does similar things.
4012 asmlinkage
long sys_nice(int increment
)
4017 * Setpriority might change our priority at the same moment.
4018 * We don't have to worry. Conceptually one call occurs first
4019 * and we have a single winner.
4021 if (increment
< -40)
4026 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4032 if (increment
< 0 && !can_nice(current
, nice
))
4035 retval
= security_task_setnice(current
, nice
);
4039 set_user_nice(current
, nice
);
4046 * task_prio - return the priority value of a given task.
4047 * @p: the task in question.
4049 * This is the priority value as seen by users in /proc.
4050 * RT tasks are offset by -200. Normal tasks are centered
4051 * around 0, value goes from -16 to +15.
4053 int task_prio(const struct task_struct
*p
)
4055 return p
->prio
- MAX_RT_PRIO
;
4059 * task_nice - return the nice value of a given task.
4060 * @p: the task in question.
4062 int task_nice(const struct task_struct
*p
)
4064 return TASK_NICE(p
);
4066 EXPORT_SYMBOL_GPL(task_nice
);
4069 * idle_cpu - is a given cpu idle currently?
4070 * @cpu: the processor in question.
4072 int idle_cpu(int cpu
)
4074 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4078 * idle_task - return the idle task for a given cpu.
4079 * @cpu: the processor in question.
4081 struct task_struct
*idle_task(int cpu
)
4083 return cpu_rq(cpu
)->idle
;
4087 * find_process_by_pid - find a process with a matching PID value.
4088 * @pid: the pid in question.
4090 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4092 return pid
? find_task_by_pid(pid
) : current
;
4095 /* Actually do priority change: must hold rq lock. */
4097 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4099 BUG_ON(p
->se
.on_rq
);
4102 switch (p
->policy
) {
4106 p
->sched_class
= &fair_sched_class
;
4110 p
->sched_class
= &rt_sched_class
;
4114 p
->rt_priority
= prio
;
4115 p
->normal_prio
= normal_prio(p
);
4116 /* we are holding p->pi_lock already */
4117 p
->prio
= rt_mutex_getprio(p
);
4122 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4123 * @p: the task in question.
4124 * @policy: new policy.
4125 * @param: structure containing the new RT priority.
4127 * NOTE that the task may be already dead.
4129 int sched_setscheduler(struct task_struct
*p
, int policy
,
4130 struct sched_param
*param
)
4132 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4133 unsigned long flags
;
4136 /* may grab non-irq protected spin_locks */
4137 BUG_ON(in_interrupt());
4139 /* double check policy once rq lock held */
4141 policy
= oldpolicy
= p
->policy
;
4142 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4143 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4144 policy
!= SCHED_IDLE
)
4147 * Valid priorities for SCHED_FIFO and SCHED_RR are
4148 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4149 * SCHED_BATCH and SCHED_IDLE is 0.
4151 if (param
->sched_priority
< 0 ||
4152 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4153 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4155 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4159 * Allow unprivileged RT tasks to decrease priority:
4161 if (!capable(CAP_SYS_NICE
)) {
4162 if (rt_policy(policy
)) {
4163 unsigned long rlim_rtprio
;
4165 if (!lock_task_sighand(p
, &flags
))
4167 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4168 unlock_task_sighand(p
, &flags
);
4170 /* can't set/change the rt policy */
4171 if (policy
!= p
->policy
&& !rlim_rtprio
)
4174 /* can't increase priority */
4175 if (param
->sched_priority
> p
->rt_priority
&&
4176 param
->sched_priority
> rlim_rtprio
)
4180 * Like positive nice levels, dont allow tasks to
4181 * move out of SCHED_IDLE either:
4183 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4186 /* can't change other user's priorities */
4187 if ((current
->euid
!= p
->euid
) &&
4188 (current
->euid
!= p
->uid
))
4192 retval
= security_task_setscheduler(p
, policy
, param
);
4196 * make sure no PI-waiters arrive (or leave) while we are
4197 * changing the priority of the task:
4199 spin_lock_irqsave(&p
->pi_lock
, flags
);
4201 * To be able to change p->policy safely, the apropriate
4202 * runqueue lock must be held.
4204 rq
= __task_rq_lock(p
);
4205 /* recheck policy now with rq lock held */
4206 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4207 policy
= oldpolicy
= -1;
4208 __task_rq_unlock(rq
);
4209 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4212 update_rq_clock(rq
);
4213 on_rq
= p
->se
.on_rq
;
4215 deactivate_task(rq
, p
, 0);
4217 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4219 activate_task(rq
, p
, 0);
4221 * Reschedule if we are currently running on this runqueue and
4222 * our priority decreased, or if we are not currently running on
4223 * this runqueue and our priority is higher than the current's
4225 if (task_running(rq
, p
)) {
4226 if (p
->prio
> oldprio
)
4227 resched_task(rq
->curr
);
4229 check_preempt_curr(rq
, p
);
4232 __task_rq_unlock(rq
);
4233 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4235 rt_mutex_adjust_pi(p
);
4239 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4242 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4244 struct sched_param lparam
;
4245 struct task_struct
*p
;
4248 if (!param
|| pid
< 0)
4250 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4255 p
= find_process_by_pid(pid
);
4257 retval
= sched_setscheduler(p
, policy
, &lparam
);
4264 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4265 * @pid: the pid in question.
4266 * @policy: new policy.
4267 * @param: structure containing the new RT priority.
4269 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4270 struct sched_param __user
*param
)
4272 /* negative values for policy are not valid */
4276 return do_sched_setscheduler(pid
, policy
, param
);
4280 * sys_sched_setparam - set/change the RT priority of a thread
4281 * @pid: the pid in question.
4282 * @param: structure containing the new RT priority.
4284 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4286 return do_sched_setscheduler(pid
, -1, param
);
4290 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4291 * @pid: the pid in question.
4293 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4295 struct task_struct
*p
;
4296 int retval
= -EINVAL
;
4302 read_lock(&tasklist_lock
);
4303 p
= find_process_by_pid(pid
);
4305 retval
= security_task_getscheduler(p
);
4309 read_unlock(&tasklist_lock
);
4316 * sys_sched_getscheduler - get the RT priority of a thread
4317 * @pid: the pid in question.
4318 * @param: structure containing the RT priority.
4320 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4322 struct sched_param lp
;
4323 struct task_struct
*p
;
4324 int retval
= -EINVAL
;
4326 if (!param
|| pid
< 0)
4329 read_lock(&tasklist_lock
);
4330 p
= find_process_by_pid(pid
);
4335 retval
= security_task_getscheduler(p
);
4339 lp
.sched_priority
= p
->rt_priority
;
4340 read_unlock(&tasklist_lock
);
4343 * This one might sleep, we cannot do it with a spinlock held ...
4345 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4351 read_unlock(&tasklist_lock
);
4355 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4357 cpumask_t cpus_allowed
;
4358 struct task_struct
*p
;
4361 mutex_lock(&sched_hotcpu_mutex
);
4362 read_lock(&tasklist_lock
);
4364 p
= find_process_by_pid(pid
);
4366 read_unlock(&tasklist_lock
);
4367 mutex_unlock(&sched_hotcpu_mutex
);
4372 * It is not safe to call set_cpus_allowed with the
4373 * tasklist_lock held. We will bump the task_struct's
4374 * usage count and then drop tasklist_lock.
4377 read_unlock(&tasklist_lock
);
4380 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4381 !capable(CAP_SYS_NICE
))
4384 retval
= security_task_setscheduler(p
, 0, NULL
);
4388 cpus_allowed
= cpuset_cpus_allowed(p
);
4389 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4390 retval
= set_cpus_allowed(p
, new_mask
);
4394 mutex_unlock(&sched_hotcpu_mutex
);
4398 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4399 cpumask_t
*new_mask
)
4401 if (len
< sizeof(cpumask_t
)) {
4402 memset(new_mask
, 0, sizeof(cpumask_t
));
4403 } else if (len
> sizeof(cpumask_t
)) {
4404 len
= sizeof(cpumask_t
);
4406 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4410 * sys_sched_setaffinity - set the cpu affinity of a process
4411 * @pid: pid of the process
4412 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4413 * @user_mask_ptr: user-space pointer to the new cpu mask
4415 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4416 unsigned long __user
*user_mask_ptr
)
4421 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4425 return sched_setaffinity(pid
, new_mask
);
4429 * Represents all cpu's present in the system
4430 * In systems capable of hotplug, this map could dynamically grow
4431 * as new cpu's are detected in the system via any platform specific
4432 * method, such as ACPI for e.g.
4435 cpumask_t cpu_present_map __read_mostly
;
4436 EXPORT_SYMBOL(cpu_present_map
);
4439 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4440 EXPORT_SYMBOL(cpu_online_map
);
4442 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4443 EXPORT_SYMBOL(cpu_possible_map
);
4446 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4448 struct task_struct
*p
;
4451 mutex_lock(&sched_hotcpu_mutex
);
4452 read_lock(&tasklist_lock
);
4455 p
= find_process_by_pid(pid
);
4459 retval
= security_task_getscheduler(p
);
4463 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4466 read_unlock(&tasklist_lock
);
4467 mutex_unlock(&sched_hotcpu_mutex
);
4473 * sys_sched_getaffinity - get the cpu affinity of a process
4474 * @pid: pid of the process
4475 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4476 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4478 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4479 unsigned long __user
*user_mask_ptr
)
4484 if (len
< sizeof(cpumask_t
))
4487 ret
= sched_getaffinity(pid
, &mask
);
4491 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4494 return sizeof(cpumask_t
);
4498 * sys_sched_yield - yield the current processor to other threads.
4500 * This function yields the current CPU to other tasks. If there are no
4501 * other threads running on this CPU then this function will return.
4503 asmlinkage
long sys_sched_yield(void)
4505 struct rq
*rq
= this_rq_lock();
4507 schedstat_inc(rq
, yld_cnt
);
4508 current
->sched_class
->yield_task(rq
, current
);
4511 * Since we are going to call schedule() anyway, there's
4512 * no need to preempt or enable interrupts:
4514 __release(rq
->lock
);
4515 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4516 _raw_spin_unlock(&rq
->lock
);
4517 preempt_enable_no_resched();
4524 static void __cond_resched(void)
4526 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4527 __might_sleep(__FILE__
, __LINE__
);
4530 * The BKS might be reacquired before we have dropped
4531 * PREEMPT_ACTIVE, which could trigger a second
4532 * cond_resched() call.
4535 add_preempt_count(PREEMPT_ACTIVE
);
4537 sub_preempt_count(PREEMPT_ACTIVE
);
4538 } while (need_resched());
4541 int __sched
cond_resched(void)
4543 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4544 system_state
== SYSTEM_RUNNING
) {
4550 EXPORT_SYMBOL(cond_resched
);
4553 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4554 * call schedule, and on return reacquire the lock.
4556 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4557 * operations here to prevent schedule() from being called twice (once via
4558 * spin_unlock(), once by hand).
4560 int cond_resched_lock(spinlock_t
*lock
)
4564 if (need_lockbreak(lock
)) {
4570 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4571 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4572 _raw_spin_unlock(lock
);
4573 preempt_enable_no_resched();
4580 EXPORT_SYMBOL(cond_resched_lock
);
4582 int __sched
cond_resched_softirq(void)
4584 BUG_ON(!in_softirq());
4586 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4594 EXPORT_SYMBOL(cond_resched_softirq
);
4597 * yield - yield the current processor to other threads.
4599 * This is a shortcut for kernel-space yielding - it marks the
4600 * thread runnable and calls sys_sched_yield().
4602 void __sched
yield(void)
4604 set_current_state(TASK_RUNNING
);
4607 EXPORT_SYMBOL(yield
);
4610 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4611 * that process accounting knows that this is a task in IO wait state.
4613 * But don't do that if it is a deliberate, throttling IO wait (this task
4614 * has set its backing_dev_info: the queue against which it should throttle)
4616 void __sched
io_schedule(void)
4618 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4620 delayacct_blkio_start();
4621 atomic_inc(&rq
->nr_iowait
);
4623 atomic_dec(&rq
->nr_iowait
);
4624 delayacct_blkio_end();
4626 EXPORT_SYMBOL(io_schedule
);
4628 long __sched
io_schedule_timeout(long timeout
)
4630 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4633 delayacct_blkio_start();
4634 atomic_inc(&rq
->nr_iowait
);
4635 ret
= schedule_timeout(timeout
);
4636 atomic_dec(&rq
->nr_iowait
);
4637 delayacct_blkio_end();
4642 * sys_sched_get_priority_max - return maximum RT priority.
4643 * @policy: scheduling class.
4645 * this syscall returns the maximum rt_priority that can be used
4646 * by a given scheduling class.
4648 asmlinkage
long sys_sched_get_priority_max(int policy
)
4655 ret
= MAX_USER_RT_PRIO
-1;
4667 * sys_sched_get_priority_min - return minimum RT priority.
4668 * @policy: scheduling class.
4670 * this syscall returns the minimum rt_priority that can be used
4671 * by a given scheduling class.
4673 asmlinkage
long sys_sched_get_priority_min(int policy
)
4691 * sys_sched_rr_get_interval - return the default timeslice of a process.
4692 * @pid: pid of the process.
4693 * @interval: userspace pointer to the timeslice value.
4695 * this syscall writes the default timeslice value of a given process
4696 * into the user-space timespec buffer. A value of '0' means infinity.
4699 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4701 struct task_struct
*p
;
4702 int retval
= -EINVAL
;
4709 read_lock(&tasklist_lock
);
4710 p
= find_process_by_pid(pid
);
4714 retval
= security_task_getscheduler(p
);
4718 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4719 0 : static_prio_timeslice(p
->static_prio
), &t
);
4720 read_unlock(&tasklist_lock
);
4721 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4725 read_unlock(&tasklist_lock
);
4729 static const char stat_nam
[] = "RSDTtZX";
4731 static void show_task(struct task_struct
*p
)
4733 unsigned long free
= 0;
4736 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4737 printk("%-13.13s %c", p
->comm
,
4738 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4739 #if BITS_PER_LONG == 32
4740 if (state
== TASK_RUNNING
)
4741 printk(" running ");
4743 printk(" %08lx ", thread_saved_pc(p
));
4745 if (state
== TASK_RUNNING
)
4746 printk(" running task ");
4748 printk(" %016lx ", thread_saved_pc(p
));
4750 #ifdef CONFIG_DEBUG_STACK_USAGE
4752 unsigned long *n
= end_of_stack(p
);
4755 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4758 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4760 if (state
!= TASK_RUNNING
)
4761 show_stack(p
, NULL
);
4764 void show_state_filter(unsigned long state_filter
)
4766 struct task_struct
*g
, *p
;
4768 #if BITS_PER_LONG == 32
4770 " task PC stack pid father\n");
4773 " task PC stack pid father\n");
4775 read_lock(&tasklist_lock
);
4776 do_each_thread(g
, p
) {
4778 * reset the NMI-timeout, listing all files on a slow
4779 * console might take alot of time:
4781 touch_nmi_watchdog();
4782 if (!state_filter
|| (p
->state
& state_filter
))
4784 } while_each_thread(g
, p
);
4786 touch_all_softlockup_watchdogs();
4788 #ifdef CONFIG_SCHED_DEBUG
4789 sysrq_sched_debug_show();
4791 read_unlock(&tasklist_lock
);
4793 * Only show locks if all tasks are dumped:
4795 if (state_filter
== -1)
4796 debug_show_all_locks();
4799 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4801 idle
->sched_class
= &idle_sched_class
;
4805 * init_idle - set up an idle thread for a given CPU
4806 * @idle: task in question
4807 * @cpu: cpu the idle task belongs to
4809 * NOTE: this function does not set the idle thread's NEED_RESCHED
4810 * flag, to make booting more robust.
4812 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4814 struct rq
*rq
= cpu_rq(cpu
);
4815 unsigned long flags
;
4818 idle
->se
.exec_start
= sched_clock();
4820 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4821 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4822 __set_task_cpu(idle
, cpu
);
4824 spin_lock_irqsave(&rq
->lock
, flags
);
4825 rq
->curr
= rq
->idle
= idle
;
4826 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4829 spin_unlock_irqrestore(&rq
->lock
, flags
);
4831 /* Set the preempt count _outside_ the spinlocks! */
4832 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4833 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4835 task_thread_info(idle
)->preempt_count
= 0;
4838 * The idle tasks have their own, simple scheduling class:
4840 idle
->sched_class
= &idle_sched_class
;
4844 * In a system that switches off the HZ timer nohz_cpu_mask
4845 * indicates which cpus entered this state. This is used
4846 * in the rcu update to wait only for active cpus. For system
4847 * which do not switch off the HZ timer nohz_cpu_mask should
4848 * always be CPU_MASK_NONE.
4850 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4854 * This is how migration works:
4856 * 1) we queue a struct migration_req structure in the source CPU's
4857 * runqueue and wake up that CPU's migration thread.
4858 * 2) we down() the locked semaphore => thread blocks.
4859 * 3) migration thread wakes up (implicitly it forces the migrated
4860 * thread off the CPU)
4861 * 4) it gets the migration request and checks whether the migrated
4862 * task is still in the wrong runqueue.
4863 * 5) if it's in the wrong runqueue then the migration thread removes
4864 * it and puts it into the right queue.
4865 * 6) migration thread up()s the semaphore.
4866 * 7) we wake up and the migration is done.
4870 * Change a given task's CPU affinity. Migrate the thread to a
4871 * proper CPU and schedule it away if the CPU it's executing on
4872 * is removed from the allowed bitmask.
4874 * NOTE: the caller must have a valid reference to the task, the
4875 * task must not exit() & deallocate itself prematurely. The
4876 * call is not atomic; no spinlocks may be held.
4878 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4880 struct migration_req req
;
4881 unsigned long flags
;
4885 rq
= task_rq_lock(p
, &flags
);
4886 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4891 p
->cpus_allowed
= new_mask
;
4892 /* Can the task run on the task's current CPU? If so, we're done */
4893 if (cpu_isset(task_cpu(p
), new_mask
))
4896 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4897 /* Need help from migration thread: drop lock and wait. */
4898 task_rq_unlock(rq
, &flags
);
4899 wake_up_process(rq
->migration_thread
);
4900 wait_for_completion(&req
.done
);
4901 tlb_migrate_finish(p
->mm
);
4905 task_rq_unlock(rq
, &flags
);
4909 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4912 * Move (not current) task off this cpu, onto dest cpu. We're doing
4913 * this because either it can't run here any more (set_cpus_allowed()
4914 * away from this CPU, or CPU going down), or because we're
4915 * attempting to rebalance this task on exec (sched_exec).
4917 * So we race with normal scheduler movements, but that's OK, as long
4918 * as the task is no longer on this CPU.
4920 * Returns non-zero if task was successfully migrated.
4922 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4924 struct rq
*rq_dest
, *rq_src
;
4927 if (unlikely(cpu_is_offline(dest_cpu
)))
4930 rq_src
= cpu_rq(src_cpu
);
4931 rq_dest
= cpu_rq(dest_cpu
);
4933 double_rq_lock(rq_src
, rq_dest
);
4934 /* Already moved. */
4935 if (task_cpu(p
) != src_cpu
)
4937 /* Affinity changed (again). */
4938 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4941 on_rq
= p
->se
.on_rq
;
4943 deactivate_task(rq_src
, p
, 0);
4945 set_task_cpu(p
, dest_cpu
);
4947 activate_task(rq_dest
, p
, 0);
4948 check_preempt_curr(rq_dest
, p
);
4952 double_rq_unlock(rq_src
, rq_dest
);
4957 * migration_thread - this is a highprio system thread that performs
4958 * thread migration by bumping thread off CPU then 'pushing' onto
4961 static int migration_thread(void *data
)
4963 int cpu
= (long)data
;
4967 BUG_ON(rq
->migration_thread
!= current
);
4969 set_current_state(TASK_INTERRUPTIBLE
);
4970 while (!kthread_should_stop()) {
4971 struct migration_req
*req
;
4972 struct list_head
*head
;
4974 spin_lock_irq(&rq
->lock
);
4976 if (cpu_is_offline(cpu
)) {
4977 spin_unlock_irq(&rq
->lock
);
4981 if (rq
->active_balance
) {
4982 active_load_balance(rq
, cpu
);
4983 rq
->active_balance
= 0;
4986 head
= &rq
->migration_queue
;
4988 if (list_empty(head
)) {
4989 spin_unlock_irq(&rq
->lock
);
4991 set_current_state(TASK_INTERRUPTIBLE
);
4994 req
= list_entry(head
->next
, struct migration_req
, list
);
4995 list_del_init(head
->next
);
4997 spin_unlock(&rq
->lock
);
4998 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5001 complete(&req
->done
);
5003 __set_current_state(TASK_RUNNING
);
5007 /* Wait for kthread_stop */
5008 set_current_state(TASK_INTERRUPTIBLE
);
5009 while (!kthread_should_stop()) {
5011 set_current_state(TASK_INTERRUPTIBLE
);
5013 __set_current_state(TASK_RUNNING
);
5017 #ifdef CONFIG_HOTPLUG_CPU
5019 * Figure out where task on dead CPU should go, use force if neccessary.
5020 * NOTE: interrupts should be disabled by the caller
5022 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5024 unsigned long flags
;
5031 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5032 cpus_and(mask
, mask
, p
->cpus_allowed
);
5033 dest_cpu
= any_online_cpu(mask
);
5035 /* On any allowed CPU? */
5036 if (dest_cpu
== NR_CPUS
)
5037 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5039 /* No more Mr. Nice Guy. */
5040 if (dest_cpu
== NR_CPUS
) {
5041 rq
= task_rq_lock(p
, &flags
);
5042 cpus_setall(p
->cpus_allowed
);
5043 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5044 task_rq_unlock(rq
, &flags
);
5047 * Don't tell them about moving exiting tasks or
5048 * kernel threads (both mm NULL), since they never
5051 if (p
->mm
&& printk_ratelimit())
5052 printk(KERN_INFO
"process %d (%s) no "
5053 "longer affine to cpu%d\n",
5054 p
->pid
, p
->comm
, dead_cpu
);
5056 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5061 * While a dead CPU has no uninterruptible tasks queued at this point,
5062 * it might still have a nonzero ->nr_uninterruptible counter, because
5063 * for performance reasons the counter is not stricly tracking tasks to
5064 * their home CPUs. So we just add the counter to another CPU's counter,
5065 * to keep the global sum constant after CPU-down:
5067 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5069 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5070 unsigned long flags
;
5072 local_irq_save(flags
);
5073 double_rq_lock(rq_src
, rq_dest
);
5074 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5075 rq_src
->nr_uninterruptible
= 0;
5076 double_rq_unlock(rq_src
, rq_dest
);
5077 local_irq_restore(flags
);
5080 /* Run through task list and migrate tasks from the dead cpu. */
5081 static void migrate_live_tasks(int src_cpu
)
5083 struct task_struct
*p
, *t
;
5085 write_lock_irq(&tasklist_lock
);
5087 do_each_thread(t
, p
) {
5091 if (task_cpu(p
) == src_cpu
)
5092 move_task_off_dead_cpu(src_cpu
, p
);
5093 } while_each_thread(t
, p
);
5095 write_unlock_irq(&tasklist_lock
);
5099 * Schedules idle task to be the next runnable task on current CPU.
5100 * It does so by boosting its priority to highest possible and adding it to
5101 * the _front_ of the runqueue. Used by CPU offline code.
5103 void sched_idle_next(void)
5105 int this_cpu
= smp_processor_id();
5106 struct rq
*rq
= cpu_rq(this_cpu
);
5107 struct task_struct
*p
= rq
->idle
;
5108 unsigned long flags
;
5110 /* cpu has to be offline */
5111 BUG_ON(cpu_online(this_cpu
));
5114 * Strictly not necessary since rest of the CPUs are stopped by now
5115 * and interrupts disabled on the current cpu.
5117 spin_lock_irqsave(&rq
->lock
, flags
);
5119 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5121 /* Add idle task to the _front_ of its priority queue: */
5122 activate_idle_task(p
, rq
);
5124 spin_unlock_irqrestore(&rq
->lock
, flags
);
5128 * Ensures that the idle task is using init_mm right before its cpu goes
5131 void idle_task_exit(void)
5133 struct mm_struct
*mm
= current
->active_mm
;
5135 BUG_ON(cpu_online(smp_processor_id()));
5138 switch_mm(mm
, &init_mm
, current
);
5142 /* called under rq->lock with disabled interrupts */
5143 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5145 struct rq
*rq
= cpu_rq(dead_cpu
);
5147 /* Must be exiting, otherwise would be on tasklist. */
5148 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5150 /* Cannot have done final schedule yet: would have vanished. */
5151 BUG_ON(p
->state
== TASK_DEAD
);
5156 * Drop lock around migration; if someone else moves it,
5157 * that's OK. No task can be added to this CPU, so iteration is
5159 * NOTE: interrupts should be left disabled --dev@
5161 spin_unlock(&rq
->lock
);
5162 move_task_off_dead_cpu(dead_cpu
, p
);
5163 spin_lock(&rq
->lock
);
5168 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5169 static void migrate_dead_tasks(unsigned int dead_cpu
)
5171 struct rq
*rq
= cpu_rq(dead_cpu
);
5172 struct task_struct
*next
;
5175 if (!rq
->nr_running
)
5177 update_rq_clock(rq
);
5178 next
= pick_next_task(rq
, rq
->curr
);
5181 migrate_dead(dead_cpu
, next
);
5185 #endif /* CONFIG_HOTPLUG_CPU */
5187 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5189 static struct ctl_table sd_ctl_dir
[] = {
5191 .procname
= "sched_domain",
5197 static struct ctl_table sd_ctl_root
[] = {
5199 .ctl_name
= CTL_KERN
,
5200 .procname
= "kernel",
5202 .child
= sd_ctl_dir
,
5207 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5209 struct ctl_table
*entry
=
5210 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5213 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5219 set_table_entry(struct ctl_table
*entry
,
5220 const char *procname
, void *data
, int maxlen
,
5221 mode_t mode
, proc_handler
*proc_handler
)
5223 entry
->procname
= procname
;
5225 entry
->maxlen
= maxlen
;
5227 entry
->proc_handler
= proc_handler
;
5230 static struct ctl_table
*
5231 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5233 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5235 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5236 sizeof(long), 0644, proc_doulongvec_minmax
);
5237 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5238 sizeof(long), 0644, proc_doulongvec_minmax
);
5239 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5240 sizeof(int), 0644, proc_dointvec_minmax
);
5241 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5242 sizeof(int), 0644, proc_dointvec_minmax
);
5243 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5244 sizeof(int), 0644, proc_dointvec_minmax
);
5245 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5246 sizeof(int), 0644, proc_dointvec_minmax
);
5247 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5248 sizeof(int), 0644, proc_dointvec_minmax
);
5249 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5250 sizeof(int), 0644, proc_dointvec_minmax
);
5251 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5252 sizeof(int), 0644, proc_dointvec_minmax
);
5253 set_table_entry(&table
[10], "cache_nice_tries",
5254 &sd
->cache_nice_tries
,
5255 sizeof(int), 0644, proc_dointvec_minmax
);
5256 set_table_entry(&table
[12], "flags", &sd
->flags
,
5257 sizeof(int), 0644, proc_dointvec_minmax
);
5262 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5264 struct ctl_table
*entry
, *table
;
5265 struct sched_domain
*sd
;
5266 int domain_num
= 0, i
;
5269 for_each_domain(cpu
, sd
)
5271 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5274 for_each_domain(cpu
, sd
) {
5275 snprintf(buf
, 32, "domain%d", i
);
5276 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5278 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5285 static struct ctl_table_header
*sd_sysctl_header
;
5286 static void init_sched_domain_sysctl(void)
5288 int i
, cpu_num
= num_online_cpus();
5289 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5292 sd_ctl_dir
[0].child
= entry
;
5294 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5295 snprintf(buf
, 32, "cpu%d", i
);
5296 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5298 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5300 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5303 static void init_sched_domain_sysctl(void)
5309 * migration_call - callback that gets triggered when a CPU is added.
5310 * Here we can start up the necessary migration thread for the new CPU.
5312 static int __cpuinit
5313 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5315 struct task_struct
*p
;
5316 int cpu
= (long)hcpu
;
5317 unsigned long flags
;
5321 case CPU_LOCK_ACQUIRE
:
5322 mutex_lock(&sched_hotcpu_mutex
);
5325 case CPU_UP_PREPARE
:
5326 case CPU_UP_PREPARE_FROZEN
:
5327 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5330 kthread_bind(p
, cpu
);
5331 /* Must be high prio: stop_machine expects to yield to it. */
5332 rq
= task_rq_lock(p
, &flags
);
5333 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5334 task_rq_unlock(rq
, &flags
);
5335 cpu_rq(cpu
)->migration_thread
= p
;
5339 case CPU_ONLINE_FROZEN
:
5340 /* Strictly unneccessary, as first user will wake it. */
5341 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5344 #ifdef CONFIG_HOTPLUG_CPU
5345 case CPU_UP_CANCELED
:
5346 case CPU_UP_CANCELED_FROZEN
:
5347 if (!cpu_rq(cpu
)->migration_thread
)
5349 /* Unbind it from offline cpu so it can run. Fall thru. */
5350 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5351 any_online_cpu(cpu_online_map
));
5352 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5353 cpu_rq(cpu
)->migration_thread
= NULL
;
5357 case CPU_DEAD_FROZEN
:
5358 migrate_live_tasks(cpu
);
5360 kthread_stop(rq
->migration_thread
);
5361 rq
->migration_thread
= NULL
;
5362 /* Idle task back to normal (off runqueue, low prio) */
5363 rq
= task_rq_lock(rq
->idle
, &flags
);
5364 update_rq_clock(rq
);
5365 deactivate_task(rq
, rq
->idle
, 0);
5366 rq
->idle
->static_prio
= MAX_PRIO
;
5367 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5368 rq
->idle
->sched_class
= &idle_sched_class
;
5369 migrate_dead_tasks(cpu
);
5370 task_rq_unlock(rq
, &flags
);
5371 migrate_nr_uninterruptible(rq
);
5372 BUG_ON(rq
->nr_running
!= 0);
5374 /* No need to migrate the tasks: it was best-effort if
5375 * they didn't take sched_hotcpu_mutex. Just wake up
5376 * the requestors. */
5377 spin_lock_irq(&rq
->lock
);
5378 while (!list_empty(&rq
->migration_queue
)) {
5379 struct migration_req
*req
;
5381 req
= list_entry(rq
->migration_queue
.next
,
5382 struct migration_req
, list
);
5383 list_del_init(&req
->list
);
5384 complete(&req
->done
);
5386 spin_unlock_irq(&rq
->lock
);
5389 case CPU_LOCK_RELEASE
:
5390 mutex_unlock(&sched_hotcpu_mutex
);
5396 /* Register at highest priority so that task migration (migrate_all_tasks)
5397 * happens before everything else.
5399 static struct notifier_block __cpuinitdata migration_notifier
= {
5400 .notifier_call
= migration_call
,
5404 int __init
migration_init(void)
5406 void *cpu
= (void *)(long)smp_processor_id();
5409 /* Start one for the boot CPU: */
5410 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5411 BUG_ON(err
== NOTIFY_BAD
);
5412 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5413 register_cpu_notifier(&migration_notifier
);
5421 /* Number of possible processor ids */
5422 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5423 EXPORT_SYMBOL(nr_cpu_ids
);
5425 #undef SCHED_DOMAIN_DEBUG
5426 #ifdef SCHED_DOMAIN_DEBUG
5427 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5432 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5436 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5441 struct sched_group
*group
= sd
->groups
;
5442 cpumask_t groupmask
;
5444 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5445 cpus_clear(groupmask
);
5448 for (i
= 0; i
< level
+ 1; i
++)
5450 printk("domain %d: ", level
);
5452 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5453 printk("does not load-balance\n");
5455 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5460 printk("span %s\n", str
);
5462 if (!cpu_isset(cpu
, sd
->span
))
5463 printk(KERN_ERR
"ERROR: domain->span does not contain "
5465 if (!cpu_isset(cpu
, group
->cpumask
))
5466 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5470 for (i
= 0; i
< level
+ 2; i
++)
5476 printk(KERN_ERR
"ERROR: group is NULL\n");
5480 if (!group
->__cpu_power
) {
5482 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5486 if (!cpus_weight(group
->cpumask
)) {
5488 printk(KERN_ERR
"ERROR: empty group\n");
5491 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5493 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5496 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5498 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5501 group
= group
->next
;
5502 } while (group
!= sd
->groups
);
5505 if (!cpus_equal(sd
->span
, groupmask
))
5506 printk(KERN_ERR
"ERROR: groups don't span "
5514 if (!cpus_subset(groupmask
, sd
->span
))
5515 printk(KERN_ERR
"ERROR: parent span is not a superset "
5516 "of domain->span\n");
5521 # define sched_domain_debug(sd, cpu) do { } while (0)
5524 static int sd_degenerate(struct sched_domain
*sd
)
5526 if (cpus_weight(sd
->span
) == 1)
5529 /* Following flags need at least 2 groups */
5530 if (sd
->flags
& (SD_LOAD_BALANCE
|
5531 SD_BALANCE_NEWIDLE
|
5535 SD_SHARE_PKG_RESOURCES
)) {
5536 if (sd
->groups
!= sd
->groups
->next
)
5540 /* Following flags don't use groups */
5541 if (sd
->flags
& (SD_WAKE_IDLE
|
5550 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5552 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5554 if (sd_degenerate(parent
))
5557 if (!cpus_equal(sd
->span
, parent
->span
))
5560 /* Does parent contain flags not in child? */
5561 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5562 if (cflags
& SD_WAKE_AFFINE
)
5563 pflags
&= ~SD_WAKE_BALANCE
;
5564 /* Flags needing groups don't count if only 1 group in parent */
5565 if (parent
->groups
== parent
->groups
->next
) {
5566 pflags
&= ~(SD_LOAD_BALANCE
|
5567 SD_BALANCE_NEWIDLE
|
5571 SD_SHARE_PKG_RESOURCES
);
5573 if (~cflags
& pflags
)
5580 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5581 * hold the hotplug lock.
5583 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5585 struct rq
*rq
= cpu_rq(cpu
);
5586 struct sched_domain
*tmp
;
5588 /* Remove the sched domains which do not contribute to scheduling. */
5589 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5590 struct sched_domain
*parent
= tmp
->parent
;
5593 if (sd_parent_degenerate(tmp
, parent
)) {
5594 tmp
->parent
= parent
->parent
;
5596 parent
->parent
->child
= tmp
;
5600 if (sd
&& sd_degenerate(sd
)) {
5606 sched_domain_debug(sd
, cpu
);
5608 rcu_assign_pointer(rq
->sd
, sd
);
5611 /* cpus with isolated domains */
5612 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5614 /* Setup the mask of cpus configured for isolated domains */
5615 static int __init
isolated_cpu_setup(char *str
)
5617 int ints
[NR_CPUS
], i
;
5619 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5620 cpus_clear(cpu_isolated_map
);
5621 for (i
= 1; i
<= ints
[0]; i
++)
5622 if (ints
[i
] < NR_CPUS
)
5623 cpu_set(ints
[i
], cpu_isolated_map
);
5627 __setup ("isolcpus=", isolated_cpu_setup
);
5630 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5631 * to a function which identifies what group(along with sched group) a CPU
5632 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5633 * (due to the fact that we keep track of groups covered with a cpumask_t).
5635 * init_sched_build_groups will build a circular linked list of the groups
5636 * covered by the given span, and will set each group's ->cpumask correctly,
5637 * and ->cpu_power to 0.
5640 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5641 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5642 struct sched_group
**sg
))
5644 struct sched_group
*first
= NULL
, *last
= NULL
;
5645 cpumask_t covered
= CPU_MASK_NONE
;
5648 for_each_cpu_mask(i
, span
) {
5649 struct sched_group
*sg
;
5650 int group
= group_fn(i
, cpu_map
, &sg
);
5653 if (cpu_isset(i
, covered
))
5656 sg
->cpumask
= CPU_MASK_NONE
;
5657 sg
->__cpu_power
= 0;
5659 for_each_cpu_mask(j
, span
) {
5660 if (group_fn(j
, cpu_map
, NULL
) != group
)
5663 cpu_set(j
, covered
);
5664 cpu_set(j
, sg
->cpumask
);
5675 #define SD_NODES_PER_DOMAIN 16
5680 * find_next_best_node - find the next node to include in a sched_domain
5681 * @node: node whose sched_domain we're building
5682 * @used_nodes: nodes already in the sched_domain
5684 * Find the next node to include in a given scheduling domain. Simply
5685 * finds the closest node not already in the @used_nodes map.
5687 * Should use nodemask_t.
5689 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5691 int i
, n
, val
, min_val
, best_node
= 0;
5695 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5696 /* Start at @node */
5697 n
= (node
+ i
) % MAX_NUMNODES
;
5699 if (!nr_cpus_node(n
))
5702 /* Skip already used nodes */
5703 if (test_bit(n
, used_nodes
))
5706 /* Simple min distance search */
5707 val
= node_distance(node
, n
);
5709 if (val
< min_val
) {
5715 set_bit(best_node
, used_nodes
);
5720 * sched_domain_node_span - get a cpumask for a node's sched_domain
5721 * @node: node whose cpumask we're constructing
5722 * @size: number of nodes to include in this span
5724 * Given a node, construct a good cpumask for its sched_domain to span. It
5725 * should be one that prevents unnecessary balancing, but also spreads tasks
5728 static cpumask_t
sched_domain_node_span(int node
)
5730 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5731 cpumask_t span
, nodemask
;
5735 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5737 nodemask
= node_to_cpumask(node
);
5738 cpus_or(span
, span
, nodemask
);
5739 set_bit(node
, used_nodes
);
5741 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5742 int next_node
= find_next_best_node(node
, used_nodes
);
5744 nodemask
= node_to_cpumask(next_node
);
5745 cpus_or(span
, span
, nodemask
);
5752 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5755 * SMT sched-domains:
5757 #ifdef CONFIG_SCHED_SMT
5758 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5759 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5761 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5762 struct sched_group
**sg
)
5765 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5771 * multi-core sched-domains:
5773 #ifdef CONFIG_SCHED_MC
5774 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5775 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5778 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5779 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5780 struct sched_group
**sg
)
5783 cpumask_t mask
= cpu_sibling_map
[cpu
];
5784 cpus_and(mask
, mask
, *cpu_map
);
5785 group
= first_cpu(mask
);
5787 *sg
= &per_cpu(sched_group_core
, group
);
5790 #elif defined(CONFIG_SCHED_MC)
5791 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5792 struct sched_group
**sg
)
5795 *sg
= &per_cpu(sched_group_core
, cpu
);
5800 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5801 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5803 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5804 struct sched_group
**sg
)
5807 #ifdef CONFIG_SCHED_MC
5808 cpumask_t mask
= cpu_coregroup_map(cpu
);
5809 cpus_and(mask
, mask
, *cpu_map
);
5810 group
= first_cpu(mask
);
5811 #elif defined(CONFIG_SCHED_SMT)
5812 cpumask_t mask
= cpu_sibling_map
[cpu
];
5813 cpus_and(mask
, mask
, *cpu_map
);
5814 group
= first_cpu(mask
);
5819 *sg
= &per_cpu(sched_group_phys
, group
);
5825 * The init_sched_build_groups can't handle what we want to do with node
5826 * groups, so roll our own. Now each node has its own list of groups which
5827 * gets dynamically allocated.
5829 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5830 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5832 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5833 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5835 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5836 struct sched_group
**sg
)
5838 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5841 cpus_and(nodemask
, nodemask
, *cpu_map
);
5842 group
= first_cpu(nodemask
);
5845 *sg
= &per_cpu(sched_group_allnodes
, group
);
5849 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5851 struct sched_group
*sg
= group_head
;
5857 for_each_cpu_mask(j
, sg
->cpumask
) {
5858 struct sched_domain
*sd
;
5860 sd
= &per_cpu(phys_domains
, j
);
5861 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5863 * Only add "power" once for each
5869 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5872 if (sg
!= group_head
)
5878 /* Free memory allocated for various sched_group structures */
5879 static void free_sched_groups(const cpumask_t
*cpu_map
)
5883 for_each_cpu_mask(cpu
, *cpu_map
) {
5884 struct sched_group
**sched_group_nodes
5885 = sched_group_nodes_bycpu
[cpu
];
5887 if (!sched_group_nodes
)
5890 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5891 cpumask_t nodemask
= node_to_cpumask(i
);
5892 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5894 cpus_and(nodemask
, nodemask
, *cpu_map
);
5895 if (cpus_empty(nodemask
))
5905 if (oldsg
!= sched_group_nodes
[i
])
5908 kfree(sched_group_nodes
);
5909 sched_group_nodes_bycpu
[cpu
] = NULL
;
5913 static void free_sched_groups(const cpumask_t
*cpu_map
)
5919 * Initialize sched groups cpu_power.
5921 * cpu_power indicates the capacity of sched group, which is used while
5922 * distributing the load between different sched groups in a sched domain.
5923 * Typically cpu_power for all the groups in a sched domain will be same unless
5924 * there are asymmetries in the topology. If there are asymmetries, group
5925 * having more cpu_power will pickup more load compared to the group having
5928 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5929 * the maximum number of tasks a group can handle in the presence of other idle
5930 * or lightly loaded groups in the same sched domain.
5932 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5934 struct sched_domain
*child
;
5935 struct sched_group
*group
;
5937 WARN_ON(!sd
|| !sd
->groups
);
5939 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5944 sd
->groups
->__cpu_power
= 0;
5947 * For perf policy, if the groups in child domain share resources
5948 * (for example cores sharing some portions of the cache hierarchy
5949 * or SMT), then set this domain groups cpu_power such that each group
5950 * can handle only one task, when there are other idle groups in the
5951 * same sched domain.
5953 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5955 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5956 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5961 * add cpu_power of each child group to this groups cpu_power
5963 group
= child
->groups
;
5965 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5966 group
= group
->next
;
5967 } while (group
!= child
->groups
);
5971 * Build sched domains for a given set of cpus and attach the sched domains
5972 * to the individual cpus
5974 static int build_sched_domains(const cpumask_t
*cpu_map
)
5978 struct sched_group
**sched_group_nodes
= NULL
;
5979 int sd_allnodes
= 0;
5982 * Allocate the per-node list of sched groups
5984 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5986 if (!sched_group_nodes
) {
5987 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5990 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5994 * Set up domains for cpus specified by the cpu_map.
5996 for_each_cpu_mask(i
, *cpu_map
) {
5997 struct sched_domain
*sd
= NULL
, *p
;
5998 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6000 cpus_and(nodemask
, nodemask
, *cpu_map
);
6003 if (cpus_weight(*cpu_map
) >
6004 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6005 sd
= &per_cpu(allnodes_domains
, i
);
6006 *sd
= SD_ALLNODES_INIT
;
6007 sd
->span
= *cpu_map
;
6008 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6014 sd
= &per_cpu(node_domains
, i
);
6016 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6020 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6024 sd
= &per_cpu(phys_domains
, i
);
6026 sd
->span
= nodemask
;
6030 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6032 #ifdef CONFIG_SCHED_MC
6034 sd
= &per_cpu(core_domains
, i
);
6036 sd
->span
= cpu_coregroup_map(i
);
6037 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6040 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6043 #ifdef CONFIG_SCHED_SMT
6045 sd
= &per_cpu(cpu_domains
, i
);
6046 *sd
= SD_SIBLING_INIT
;
6047 sd
->span
= cpu_sibling_map
[i
];
6048 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6051 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6055 #ifdef CONFIG_SCHED_SMT
6056 /* Set up CPU (sibling) groups */
6057 for_each_cpu_mask(i
, *cpu_map
) {
6058 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6059 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6060 if (i
!= first_cpu(this_sibling_map
))
6063 init_sched_build_groups(this_sibling_map
, cpu_map
,
6068 #ifdef CONFIG_SCHED_MC
6069 /* Set up multi-core groups */
6070 for_each_cpu_mask(i
, *cpu_map
) {
6071 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6072 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6073 if (i
!= first_cpu(this_core_map
))
6075 init_sched_build_groups(this_core_map
, cpu_map
,
6076 &cpu_to_core_group
);
6080 /* Set up physical groups */
6081 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6082 cpumask_t nodemask
= node_to_cpumask(i
);
6084 cpus_and(nodemask
, nodemask
, *cpu_map
);
6085 if (cpus_empty(nodemask
))
6088 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6092 /* Set up node groups */
6094 init_sched_build_groups(*cpu_map
, cpu_map
,
6095 &cpu_to_allnodes_group
);
6097 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6098 /* Set up node groups */
6099 struct sched_group
*sg
, *prev
;
6100 cpumask_t nodemask
= node_to_cpumask(i
);
6101 cpumask_t domainspan
;
6102 cpumask_t covered
= CPU_MASK_NONE
;
6105 cpus_and(nodemask
, nodemask
, *cpu_map
);
6106 if (cpus_empty(nodemask
)) {
6107 sched_group_nodes
[i
] = NULL
;
6111 domainspan
= sched_domain_node_span(i
);
6112 cpus_and(domainspan
, domainspan
, *cpu_map
);
6114 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6116 printk(KERN_WARNING
"Can not alloc domain group for "
6120 sched_group_nodes
[i
] = sg
;
6121 for_each_cpu_mask(j
, nodemask
) {
6122 struct sched_domain
*sd
;
6124 sd
= &per_cpu(node_domains
, j
);
6127 sg
->__cpu_power
= 0;
6128 sg
->cpumask
= nodemask
;
6130 cpus_or(covered
, covered
, nodemask
);
6133 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6134 cpumask_t tmp
, notcovered
;
6135 int n
= (i
+ j
) % MAX_NUMNODES
;
6137 cpus_complement(notcovered
, covered
);
6138 cpus_and(tmp
, notcovered
, *cpu_map
);
6139 cpus_and(tmp
, tmp
, domainspan
);
6140 if (cpus_empty(tmp
))
6143 nodemask
= node_to_cpumask(n
);
6144 cpus_and(tmp
, tmp
, nodemask
);
6145 if (cpus_empty(tmp
))
6148 sg
= kmalloc_node(sizeof(struct sched_group
),
6152 "Can not alloc domain group for node %d\n", j
);
6155 sg
->__cpu_power
= 0;
6157 sg
->next
= prev
->next
;
6158 cpus_or(covered
, covered
, tmp
);
6165 /* Calculate CPU power for physical packages and nodes */
6166 #ifdef CONFIG_SCHED_SMT
6167 for_each_cpu_mask(i
, *cpu_map
) {
6168 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6170 init_sched_groups_power(i
, sd
);
6173 #ifdef CONFIG_SCHED_MC
6174 for_each_cpu_mask(i
, *cpu_map
) {
6175 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6177 init_sched_groups_power(i
, sd
);
6181 for_each_cpu_mask(i
, *cpu_map
) {
6182 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6184 init_sched_groups_power(i
, sd
);
6188 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6189 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6192 struct sched_group
*sg
;
6194 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6195 init_numa_sched_groups_power(sg
);
6199 /* Attach the domains */
6200 for_each_cpu_mask(i
, *cpu_map
) {
6201 struct sched_domain
*sd
;
6202 #ifdef CONFIG_SCHED_SMT
6203 sd
= &per_cpu(cpu_domains
, i
);
6204 #elif defined(CONFIG_SCHED_MC)
6205 sd
= &per_cpu(core_domains
, i
);
6207 sd
= &per_cpu(phys_domains
, i
);
6209 cpu_attach_domain(sd
, i
);
6216 free_sched_groups(cpu_map
);
6221 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6223 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6225 cpumask_t cpu_default_map
;
6229 * Setup mask for cpus without special case scheduling requirements.
6230 * For now this just excludes isolated cpus, but could be used to
6231 * exclude other special cases in the future.
6233 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6235 err
= build_sched_domains(&cpu_default_map
);
6240 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6242 free_sched_groups(cpu_map
);
6246 * Detach sched domains from a group of cpus specified in cpu_map
6247 * These cpus will now be attached to the NULL domain
6249 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6253 for_each_cpu_mask(i
, *cpu_map
)
6254 cpu_attach_domain(NULL
, i
);
6255 synchronize_sched();
6256 arch_destroy_sched_domains(cpu_map
);
6260 * Partition sched domains as specified by the cpumasks below.
6261 * This attaches all cpus from the cpumasks to the NULL domain,
6262 * waits for a RCU quiescent period, recalculates sched
6263 * domain information and then attaches them back to the
6264 * correct sched domains
6265 * Call with hotplug lock held
6267 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6269 cpumask_t change_map
;
6272 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6273 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6274 cpus_or(change_map
, *partition1
, *partition2
);
6276 /* Detach sched domains from all of the affected cpus */
6277 detach_destroy_domains(&change_map
);
6278 if (!cpus_empty(*partition1
))
6279 err
= build_sched_domains(partition1
);
6280 if (!err
&& !cpus_empty(*partition2
))
6281 err
= build_sched_domains(partition2
);
6286 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6287 static int arch_reinit_sched_domains(void)
6291 mutex_lock(&sched_hotcpu_mutex
);
6292 detach_destroy_domains(&cpu_online_map
);
6293 err
= arch_init_sched_domains(&cpu_online_map
);
6294 mutex_unlock(&sched_hotcpu_mutex
);
6299 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6303 if (buf
[0] != '0' && buf
[0] != '1')
6307 sched_smt_power_savings
= (buf
[0] == '1');
6309 sched_mc_power_savings
= (buf
[0] == '1');
6311 ret
= arch_reinit_sched_domains();
6313 return ret
? ret
: count
;
6316 #ifdef CONFIG_SCHED_MC
6317 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6319 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6321 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6322 const char *buf
, size_t count
)
6324 return sched_power_savings_store(buf
, count
, 0);
6326 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6327 sched_mc_power_savings_store
);
6330 #ifdef CONFIG_SCHED_SMT
6331 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6333 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6335 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6336 const char *buf
, size_t count
)
6338 return sched_power_savings_store(buf
, count
, 1);
6340 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6341 sched_smt_power_savings_store
);
6344 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6348 #ifdef CONFIG_SCHED_SMT
6350 err
= sysfs_create_file(&cls
->kset
.kobj
,
6351 &attr_sched_smt_power_savings
.attr
);
6353 #ifdef CONFIG_SCHED_MC
6354 if (!err
&& mc_capable())
6355 err
= sysfs_create_file(&cls
->kset
.kobj
,
6356 &attr_sched_mc_power_savings
.attr
);
6363 * Force a reinitialization of the sched domains hierarchy. The domains
6364 * and groups cannot be updated in place without racing with the balancing
6365 * code, so we temporarily attach all running cpus to the NULL domain
6366 * which will prevent rebalancing while the sched domains are recalculated.
6368 static int update_sched_domains(struct notifier_block
*nfb
,
6369 unsigned long action
, void *hcpu
)
6372 case CPU_UP_PREPARE
:
6373 case CPU_UP_PREPARE_FROZEN
:
6374 case CPU_DOWN_PREPARE
:
6375 case CPU_DOWN_PREPARE_FROZEN
:
6376 detach_destroy_domains(&cpu_online_map
);
6379 case CPU_UP_CANCELED
:
6380 case CPU_UP_CANCELED_FROZEN
:
6381 case CPU_DOWN_FAILED
:
6382 case CPU_DOWN_FAILED_FROZEN
:
6384 case CPU_ONLINE_FROZEN
:
6386 case CPU_DEAD_FROZEN
:
6388 * Fall through and re-initialise the domains.
6395 /* The hotplug lock is already held by cpu_up/cpu_down */
6396 arch_init_sched_domains(&cpu_online_map
);
6401 void __init
sched_init_smp(void)
6403 cpumask_t non_isolated_cpus
;
6405 mutex_lock(&sched_hotcpu_mutex
);
6406 arch_init_sched_domains(&cpu_online_map
);
6407 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6408 if (cpus_empty(non_isolated_cpus
))
6409 cpu_set(smp_processor_id(), non_isolated_cpus
);
6410 mutex_unlock(&sched_hotcpu_mutex
);
6411 /* XXX: Theoretical race here - CPU may be hotplugged now */
6412 hotcpu_notifier(update_sched_domains
, 0);
6414 init_sched_domain_sysctl();
6416 /* Move init over to a non-isolated CPU */
6417 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6421 void __init
sched_init_smp(void)
6424 #endif /* CONFIG_SMP */
6426 int in_sched_functions(unsigned long addr
)
6428 /* Linker adds these: start and end of __sched functions */
6429 extern char __sched_text_start
[], __sched_text_end
[];
6431 return in_lock_functions(addr
) ||
6432 (addr
>= (unsigned long)__sched_text_start
6433 && addr
< (unsigned long)__sched_text_end
);
6436 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6438 cfs_rq
->tasks_timeline
= RB_ROOT
;
6439 cfs_rq
->fair_clock
= 1;
6440 #ifdef CONFIG_FAIR_GROUP_SCHED
6445 void __init
sched_init(void)
6447 int highest_cpu
= 0;
6451 * Link up the scheduling class hierarchy:
6453 rt_sched_class
.next
= &fair_sched_class
;
6454 fair_sched_class
.next
= &idle_sched_class
;
6455 idle_sched_class
.next
= NULL
;
6457 for_each_possible_cpu(i
) {
6458 struct rt_prio_array
*array
;
6462 spin_lock_init(&rq
->lock
);
6463 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6466 init_cfs_rq(&rq
->cfs
, rq
);
6467 #ifdef CONFIG_FAIR_GROUP_SCHED
6468 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6469 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6472 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6473 rq
->cpu_load
[j
] = 0;
6476 rq
->active_balance
= 0;
6477 rq
->next_balance
= jiffies
;
6480 rq
->migration_thread
= NULL
;
6481 INIT_LIST_HEAD(&rq
->migration_queue
);
6483 atomic_set(&rq
->nr_iowait
, 0);
6485 array
= &rq
->rt
.active
;
6486 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6487 INIT_LIST_HEAD(array
->queue
+ j
);
6488 __clear_bit(j
, array
->bitmap
);
6491 /* delimiter for bitsearch: */
6492 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6495 set_load_weight(&init_task
);
6497 #ifdef CONFIG_PREEMPT_NOTIFIERS
6498 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6502 nr_cpu_ids
= highest_cpu
+ 1;
6503 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6506 #ifdef CONFIG_RT_MUTEXES
6507 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6511 * The boot idle thread does lazy MMU switching as well:
6513 atomic_inc(&init_mm
.mm_count
);
6514 enter_lazy_tlb(&init_mm
, current
);
6517 * Make us the idle thread. Technically, schedule() should not be
6518 * called from this thread, however somewhere below it might be,
6519 * but because we are the idle thread, we just pick up running again
6520 * when this runqueue becomes "idle".
6522 init_idle(current
, smp_processor_id());
6524 * During early bootup we pretend to be a normal task:
6526 current
->sched_class
= &fair_sched_class
;
6529 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6530 void __might_sleep(char *file
, int line
)
6533 static unsigned long prev_jiffy
; /* ratelimiting */
6535 if ((in_atomic() || irqs_disabled()) &&
6536 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6537 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6539 prev_jiffy
= jiffies
;
6540 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6541 " context at %s:%d\n", file
, line
);
6542 printk("in_atomic():%d, irqs_disabled():%d\n",
6543 in_atomic(), irqs_disabled());
6544 debug_show_held_locks(current
);
6545 if (irqs_disabled())
6546 print_irqtrace_events(current
);
6551 EXPORT_SYMBOL(__might_sleep
);
6554 #ifdef CONFIG_MAGIC_SYSRQ
6555 void normalize_rt_tasks(void)
6557 struct task_struct
*g
, *p
;
6558 unsigned long flags
;
6562 read_lock_irq(&tasklist_lock
);
6563 do_each_thread(g
, p
) {
6565 p
->se
.wait_runtime
= 0;
6566 p
->se
.exec_start
= 0;
6567 p
->se
.wait_start_fair
= 0;
6568 #ifdef CONFIG_SCHEDSTATS
6569 p
->se
.wait_start
= 0;
6570 p
->se
.sleep_start
= 0;
6571 p
->se
.block_start
= 0;
6573 task_rq(p
)->cfs
.fair_clock
= 0;
6574 task_rq(p
)->clock
= 0;
6578 * Renice negative nice level userspace
6581 if (TASK_NICE(p
) < 0 && p
->mm
)
6582 set_user_nice(p
, 0);
6586 spin_lock_irqsave(&p
->pi_lock
, flags
);
6587 rq
= __task_rq_lock(p
);
6590 * Do not touch the migration thread:
6592 if (p
== rq
->migration_thread
)
6596 update_rq_clock(rq
);
6597 on_rq
= p
->se
.on_rq
;
6599 deactivate_task(rq
, p
, 0);
6600 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6602 activate_task(rq
, p
, 0);
6603 resched_task(rq
->curr
);
6608 __task_rq_unlock(rq
);
6609 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6610 } while_each_thread(g
, p
);
6612 read_unlock_irq(&tasklist_lock
);
6615 #endif /* CONFIG_MAGIC_SYSRQ */
6619 * These functions are only useful for the IA64 MCA handling.
6621 * They can only be called when the whole system has been
6622 * stopped - every CPU needs to be quiescent, and no scheduling
6623 * activity can take place. Using them for anything else would
6624 * be a serious bug, and as a result, they aren't even visible
6625 * under any other configuration.
6629 * curr_task - return the current task for a given cpu.
6630 * @cpu: the processor in question.
6632 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6634 struct task_struct
*curr_task(int cpu
)
6636 return cpu_curr(cpu
);
6640 * set_curr_task - set the current task for a given cpu.
6641 * @cpu: the processor in question.
6642 * @p: the task pointer to set.
6644 * Description: This function must only be used when non-maskable interrupts
6645 * are serviced on a separate stack. It allows the architecture to switch the
6646 * notion of the current task on a cpu in a non-blocking manner. This function
6647 * must be called with all CPU's synchronized, and interrupts disabled, the
6648 * and caller must save the original value of the current task (see
6649 * curr_task() above) and restore that value before reenabling interrupts and
6650 * re-starting the system.
6652 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6654 void set_curr_task(int cpu
, struct task_struct
*p
)