4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
68 * Scheduler clock - returns current time in nanosec units.
69 * This is default implementation.
70 * Architectures and sub-architectures can override this.
72 unsigned long long __attribute__((weak
)) sched_clock(void)
74 return (unsigned long long)jiffies
* (1000000000 / HZ
);
78 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
82 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
91 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96 * Some helpers for converting nanosecond timing to jiffy resolution
98 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
99 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
108 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
109 * Timeslices get refilled after they expire.
111 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
112 #define DEF_TIMESLICE (100 * HZ / 1000)
116 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
117 * Since cpu_power is a 'constant', we can use a reciprocal divide.
119 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
121 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
125 * Each time a sched group cpu_power is changed,
126 * we must compute its reciprocal value
128 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
130 sg
->__cpu_power
+= val
;
131 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
135 #define SCALE_PRIO(x, prio) \
136 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
139 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
140 * to time slice values: [800ms ... 100ms ... 5ms]
142 static unsigned int static_prio_timeslice(int static_prio
)
144 if (static_prio
== NICE_TO_PRIO(19))
147 if (static_prio
< NICE_TO_PRIO(0))
148 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
150 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
153 static inline int rt_policy(int policy
)
155 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
160 static inline int task_has_rt_policy(struct task_struct
*p
)
162 return rt_policy(p
->policy
);
166 * This is the priority-queue data structure of the RT scheduling class:
168 struct rt_prio_array
{
169 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
170 struct list_head queue
[MAX_RT_PRIO
];
174 struct load_weight load
;
175 u64 load_update_start
, load_update_last
;
176 unsigned long delta_fair
, delta_exec
, delta_stat
;
179 /* CFS-related fields in a runqueue */
181 struct load_weight load
;
182 unsigned long nr_running
;
188 unsigned long wait_runtime_overruns
, wait_runtime_underruns
;
190 struct rb_root tasks_timeline
;
191 struct rb_node
*rb_leftmost
;
192 struct rb_node
*rb_load_balance_curr
;
193 #ifdef CONFIG_FAIR_GROUP_SCHED
194 /* 'curr' points to currently running entity on this cfs_rq.
195 * It is set to NULL otherwise (i.e when none are currently running).
197 struct sched_entity
*curr
;
198 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
207 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
211 /* Real-Time classes' related field in a runqueue: */
213 struct rt_prio_array active
;
214 int rt_load_balance_idx
;
215 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
219 * This is the main, per-CPU runqueue data structure.
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
226 spinlock_t lock
; /* runqueue lock */
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
232 unsigned long nr_running
;
233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
235 unsigned char idle_at_tick
;
237 unsigned char in_nohz_recently
;
239 struct load_stat ls
; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates
;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
255 unsigned long nr_uninterruptible
;
257 struct task_struct
*curr
, *idle
;
258 unsigned long next_balance
;
259 struct mm_struct
*prev_mm
;
261 u64 clock
, prev_clock_raw
;
264 unsigned int clock_warps
, clock_overflows
;
265 unsigned int clock_unstable_events
;
270 struct sched_domain
*sd
;
272 /* For active balancing */
275 int cpu
; /* cpu of this runqueue */
277 struct task_struct
*migration_thread
;
278 struct list_head migration_queue
;
281 #ifdef CONFIG_SCHEDSTATS
283 struct sched_info rq_sched_info
;
285 /* sys_sched_yield() stats */
286 unsigned long yld_exp_empty
;
287 unsigned long yld_act_empty
;
288 unsigned long yld_both_empty
;
289 unsigned long yld_cnt
;
291 /* schedule() stats */
292 unsigned long sched_switch
;
293 unsigned long sched_cnt
;
294 unsigned long sched_goidle
;
296 /* try_to_wake_up() stats */
297 unsigned long ttwu_cnt
;
298 unsigned long ttwu_local
;
300 struct lock_class_key rq_lock_key
;
303 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
304 static DEFINE_MUTEX(sched_hotcpu_mutex
);
306 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
308 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
311 static inline int cpu_of(struct rq
*rq
)
321 * Update the per-runqueue clock, as finegrained as the platform can give
322 * us, but without assuming monotonicity, etc.:
324 static void __update_rq_clock(struct rq
*rq
)
326 u64 prev_raw
= rq
->prev_clock_raw
;
327 u64 now
= sched_clock();
328 s64 delta
= now
- prev_raw
;
329 u64 clock
= rq
->clock
;
331 #ifdef CONFIG_SCHED_DEBUG
332 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
335 * Protect against sched_clock() occasionally going backwards:
337 if (unlikely(delta
< 0)) {
342 * Catch too large forward jumps too:
344 if (unlikely(delta
> 2*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 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
381 * clock constructed from sched_clock():
383 unsigned long long cpu_clock(int cpu
)
385 unsigned long long now
;
389 local_irq_save(flags
);
393 local_irq_restore(flags
);
398 #ifdef CONFIG_FAIR_GROUP_SCHED
399 /* Change a task's ->cfs_rq if it moves across CPUs */
400 static inline void set_task_cfs_rq(struct task_struct
*p
)
402 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
405 static inline void set_task_cfs_rq(struct task_struct
*p
)
410 #ifndef prepare_arch_switch
411 # define prepare_arch_switch(next) do { } while (0)
413 #ifndef finish_arch_switch
414 # define finish_arch_switch(prev) do { } while (0)
417 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
418 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
420 return rq
->curr
== p
;
423 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
427 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
429 #ifdef CONFIG_DEBUG_SPINLOCK
430 /* this is a valid case when another task releases the spinlock */
431 rq
->lock
.owner
= current
;
434 * If we are tracking spinlock dependencies then we have to
435 * fix up the runqueue lock - which gets 'carried over' from
438 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
440 spin_unlock_irq(&rq
->lock
);
443 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
444 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
449 return rq
->curr
== p
;
453 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
457 * We can optimise this out completely for !SMP, because the
458 * SMP rebalancing from interrupt is the only thing that cares
463 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
464 spin_unlock_irq(&rq
->lock
);
466 spin_unlock(&rq
->lock
);
470 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
474 * After ->oncpu is cleared, the task can be moved to a different CPU.
475 * We must ensure this doesn't happen until the switch is completely
481 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
485 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
488 * __task_rq_lock - lock the runqueue a given task resides on.
489 * Must be called interrupts disabled.
491 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
498 spin_lock(&rq
->lock
);
499 if (unlikely(rq
!= task_rq(p
))) {
500 spin_unlock(&rq
->lock
);
501 goto repeat_lock_task
;
507 * task_rq_lock - lock the runqueue a given task resides on and disable
508 * interrupts. Note the ordering: we can safely lookup the task_rq without
509 * explicitly disabling preemption.
511 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
517 local_irq_save(*flags
);
519 spin_lock(&rq
->lock
);
520 if (unlikely(rq
!= task_rq(p
))) {
521 spin_unlock_irqrestore(&rq
->lock
, *flags
);
522 goto repeat_lock_task
;
527 static inline void __task_rq_unlock(struct rq
*rq
)
530 spin_unlock(&rq
->lock
);
533 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
536 spin_unlock_irqrestore(&rq
->lock
, *flags
);
540 * this_rq_lock - lock this runqueue and disable interrupts.
542 static inline struct rq
*this_rq_lock(void)
549 spin_lock(&rq
->lock
);
555 * CPU frequency is/was unstable - start new by setting prev_clock_raw:
557 void sched_clock_unstable_event(void)
562 rq
= task_rq_lock(current
, &flags
);
563 rq
->prev_clock_raw
= sched_clock();
564 rq
->clock_unstable_events
++;
565 task_rq_unlock(rq
, &flags
);
569 * resched_task - mark a task 'to be rescheduled now'.
571 * On UP this means the setting of the need_resched flag, on SMP it
572 * might also involve a cross-CPU call to trigger the scheduler on
577 #ifndef tsk_is_polling
578 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
581 static void resched_task(struct task_struct
*p
)
585 assert_spin_locked(&task_rq(p
)->lock
);
587 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
590 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
593 if (cpu
== smp_processor_id())
596 /* NEED_RESCHED must be visible before we test polling */
598 if (!tsk_is_polling(p
))
599 smp_send_reschedule(cpu
);
602 static void resched_cpu(int cpu
)
604 struct rq
*rq
= cpu_rq(cpu
);
607 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
609 resched_task(cpu_curr(cpu
));
610 spin_unlock_irqrestore(&rq
->lock
, flags
);
613 static inline void resched_task(struct task_struct
*p
)
615 assert_spin_locked(&task_rq(p
)->lock
);
616 set_tsk_need_resched(p
);
620 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
622 #if BITS_PER_LONG == 32
623 if (likely(divident
<= 0xffffffffULL
))
624 return (u32
)divident
/ divisor
;
625 do_div(divident
, divisor
);
629 return divident
/ divisor
;
633 #if BITS_PER_LONG == 32
634 # define WMULT_CONST (~0UL)
636 # define WMULT_CONST (1UL << 32)
639 #define WMULT_SHIFT 32
642 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
643 struct load_weight
*lw
)
647 if (unlikely(!lw
->inv_weight
))
648 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
650 tmp
= (u64
)delta_exec
* weight
;
652 * Check whether we'd overflow the 64-bit multiplication:
654 if (unlikely(tmp
> WMULT_CONST
)) {
655 tmp
= ((tmp
>> WMULT_SHIFT
/2) * lw
->inv_weight
)
658 tmp
= (tmp
* lw
->inv_weight
) >> WMULT_SHIFT
;
661 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
664 static inline unsigned long
665 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
667 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
670 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
676 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
683 * To aid in avoiding the subversion of "niceness" due to uneven distribution
684 * of tasks with abnormal "nice" values across CPUs the contribution that
685 * each task makes to its run queue's load is weighted according to its
686 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
687 * scaled version of the new time slice allocation that they receive on time
691 #define WEIGHT_IDLEPRIO 2
692 #define WMULT_IDLEPRIO (1 << 31)
695 * Nice levels are multiplicative, with a gentle 10% change for every
696 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
697 * nice 1, it will get ~10% less CPU time than another CPU-bound task
698 * that remained on nice 0.
700 * The "10% effect" is relative and cumulative: from _any_ nice level,
701 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
702 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
703 * If a task goes up by ~10% and another task goes down by ~10% then
704 * the relative distance between them is ~25%.)
706 static const int prio_to_weight
[40] = {
707 /* -20 */ 88818, 71054, 56843, 45475, 36380, 29104, 23283, 18626, 14901, 11921,
708 /* -10 */ 9537, 7629, 6103, 4883, 3906, 3125, 2500, 2000, 1600, 1280,
709 /* 0 */ NICE_0_LOAD
/* 1024 */,
710 /* 1 */ 819, 655, 524, 419, 336, 268, 215, 172, 137,
711 /* 10 */ 110, 87, 70, 56, 45, 36, 29, 23, 18, 15,
715 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
717 * In cases where the weight does not change often, we can use the
718 * precalculated inverse to speed up arithmetics by turning divisions
719 * into multiplications:
721 static const u32 prio_to_wmult
[40] = {
722 /* -20 */ 48356, 60446, 75558, 94446, 118058,
723 /* -15 */ 147573, 184467, 230589, 288233, 360285,
724 /* -10 */ 450347, 562979, 703746, 879575, 1099582,
725 /* -5 */ 1374389, 1717986, 2147483, 2684354, 3355443,
726 /* 0 */ 4194304, 5244160, 6557201, 8196502, 10250518,
727 /* 5 */ 12782640, 16025997, 19976592, 24970740, 31350126,
728 /* 10 */ 39045157, 49367440, 61356675, 76695844, 95443717,
729 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
732 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
735 * runqueue iterator, to support SMP load-balancing between different
736 * scheduling classes, without having to expose their internal data
737 * structures to the load-balancing proper:
741 struct task_struct
*(*start
)(void *);
742 struct task_struct
*(*next
)(void *);
745 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
746 unsigned long max_nr_move
, unsigned long max_load_move
,
747 struct sched_domain
*sd
, enum cpu_idle_type idle
,
748 int *all_pinned
, unsigned long *load_moved
,
749 int *this_best_prio
, struct rq_iterator
*iterator
);
751 #include "sched_stats.h"
752 #include "sched_rt.c"
753 #include "sched_fair.c"
754 #include "sched_idletask.c"
755 #ifdef CONFIG_SCHED_DEBUG
756 # include "sched_debug.c"
759 #define sched_class_highest (&rt_sched_class)
761 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
763 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
764 ls
->delta_exec
+= ls
->delta_stat
;
765 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
771 * Update delta_exec, delta_fair fields for rq.
773 * delta_fair clock advances at a rate inversely proportional to
774 * total load (rq->ls.load.weight) on the runqueue, while
775 * delta_exec advances at the same rate as wall-clock (provided
778 * delta_exec / delta_fair is a measure of the (smoothened) load on this
779 * runqueue over any given interval. This (smoothened) load is used
780 * during load balance.
782 * This function is called /before/ updating rq->ls.load
783 * and when switching tasks.
785 static void update_curr_load(struct rq
*rq
)
787 struct load_stat
*ls
= &rq
->ls
;
790 start
= ls
->load_update_start
;
791 ls
->load_update_start
= rq
->clock
;
792 ls
->delta_stat
+= rq
->clock
- start
;
794 * Stagger updates to ls->delta_fair. Very frequent updates
797 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
798 __update_curr_load(rq
, ls
);
801 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
803 update_curr_load(rq
);
804 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
807 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
809 update_curr_load(rq
);
810 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
813 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
819 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
825 static void set_load_weight(struct task_struct
*p
)
827 task_rq(p
)->cfs
.wait_runtime
-= p
->se
.wait_runtime
;
828 p
->se
.wait_runtime
= 0;
830 if (task_has_rt_policy(p
)) {
831 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
832 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
837 * SCHED_IDLE tasks get minimal weight:
839 if (p
->policy
== SCHED_IDLE
) {
840 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
841 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
845 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
846 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
849 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
851 sched_info_queued(p
);
852 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
857 dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
859 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
864 * __normal_prio - return the priority that is based on the static prio
866 static inline int __normal_prio(struct task_struct
*p
)
868 return p
->static_prio
;
872 * Calculate the expected normal priority: i.e. priority
873 * without taking RT-inheritance into account. Might be
874 * boosted by interactivity modifiers. Changes upon fork,
875 * setprio syscalls, and whenever the interactivity
876 * estimator recalculates.
878 static inline int normal_prio(struct task_struct
*p
)
882 if (task_has_rt_policy(p
))
883 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
885 prio
= __normal_prio(p
);
890 * Calculate the current priority, i.e. the priority
891 * taken into account by the scheduler. This value might
892 * be boosted by RT tasks, or might be boosted by
893 * interactivity modifiers. Will be RT if the task got
894 * RT-boosted. If not then it returns p->normal_prio.
896 static int effective_prio(struct task_struct
*p
)
898 p
->normal_prio
= normal_prio(p
);
900 * If we are RT tasks or we were boosted to RT priority,
901 * keep the priority unchanged. Otherwise, update priority
902 * to the normal priority:
904 if (!rt_prio(p
->prio
))
905 return p
->normal_prio
;
910 * activate_task - move a task to the runqueue.
912 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
919 if (p
->state
== TASK_UNINTERRUPTIBLE
)
920 rq
->nr_uninterruptible
--;
922 enqueue_task(rq
, p
, wakeup
);
923 inc_nr_running(p
, rq
);
927 * activate_idle_task - move idle task to the _front_ of runqueue.
929 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
936 if (p
->state
== TASK_UNINTERRUPTIBLE
)
937 rq
->nr_uninterruptible
--;
939 enqueue_task(rq
, p
, 0);
940 inc_nr_running(p
, rq
);
944 * deactivate_task - remove a task from the runqueue.
947 deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
949 if (p
->state
== TASK_UNINTERRUPTIBLE
)
950 rq
->nr_uninterruptible
++;
952 dequeue_task(rq
, p
, sleep
, now
);
953 dec_nr_running(p
, rq
);
957 * task_curr - is this task currently executing on a CPU?
958 * @p: the task in question.
960 inline int task_curr(const struct task_struct
*p
)
962 return cpu_curr(task_cpu(p
)) == p
;
965 /* Used instead of source_load when we know the type == 0 */
966 unsigned long weighted_cpuload(const int cpu
)
968 return cpu_rq(cpu
)->ls
.load
.weight
;
971 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
974 task_thread_info(p
)->cpu
= cpu
;
981 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
983 int old_cpu
= task_cpu(p
);
984 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
985 u64 clock_offset
, fair_clock_offset
;
987 clock_offset
= old_rq
->clock
- new_rq
->clock
;
988 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
990 if (p
->se
.wait_start_fair
)
991 p
->se
.wait_start_fair
-= fair_clock_offset
;
992 if (p
->se
.sleep_start_fair
)
993 p
->se
.sleep_start_fair
-= fair_clock_offset
;
995 #ifdef CONFIG_SCHEDSTATS
996 if (p
->se
.wait_start
)
997 p
->se
.wait_start
-= clock_offset
;
998 if (p
->se
.sleep_start
)
999 p
->se
.sleep_start
-= clock_offset
;
1000 if (p
->se
.block_start
)
1001 p
->se
.block_start
-= clock_offset
;
1004 __set_task_cpu(p
, new_cpu
);
1007 struct migration_req
{
1008 struct list_head list
;
1010 struct task_struct
*task
;
1013 struct completion done
;
1017 * The task's runqueue lock must be held.
1018 * Returns true if you have to wait for migration thread.
1021 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1023 struct rq
*rq
= task_rq(p
);
1026 * If the task is not on a runqueue (and not running), then
1027 * it is sufficient to simply update the task's cpu field.
1029 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1030 set_task_cpu(p
, dest_cpu
);
1034 init_completion(&req
->done
);
1036 req
->dest_cpu
= dest_cpu
;
1037 list_add(&req
->list
, &rq
->migration_queue
);
1043 * wait_task_inactive - wait for a thread to unschedule.
1045 * The caller must ensure that the task *will* unschedule sometime soon,
1046 * else this function might spin for a *long* time. This function can't
1047 * be called with interrupts off, or it may introduce deadlock with
1048 * smp_call_function() if an IPI is sent by the same process we are
1049 * waiting to become inactive.
1051 void wait_task_inactive(struct task_struct
*p
)
1053 unsigned long flags
;
1059 * We do the initial early heuristics without holding
1060 * any task-queue locks at all. We'll only try to get
1061 * the runqueue lock when things look like they will
1067 * If the task is actively running on another CPU
1068 * still, just relax and busy-wait without holding
1071 * NOTE! Since we don't hold any locks, it's not
1072 * even sure that "rq" stays as the right runqueue!
1073 * But we don't care, since "task_running()" will
1074 * return false if the runqueue has changed and p
1075 * is actually now running somewhere else!
1077 while (task_running(rq
, p
))
1081 * Ok, time to look more closely! We need the rq
1082 * lock now, to be *sure*. If we're wrong, we'll
1083 * just go back and repeat.
1085 rq
= task_rq_lock(p
, &flags
);
1086 running
= task_running(rq
, p
);
1087 on_rq
= p
->se
.on_rq
;
1088 task_rq_unlock(rq
, &flags
);
1091 * Was it really running after all now that we
1092 * checked with the proper locks actually held?
1094 * Oops. Go back and try again..
1096 if (unlikely(running
)) {
1102 * It's not enough that it's not actively running,
1103 * it must be off the runqueue _entirely_, and not
1106 * So if it wa still runnable (but just not actively
1107 * running right now), it's preempted, and we should
1108 * yield - it could be a while.
1110 if (unlikely(on_rq
)) {
1116 * Ahh, all good. It wasn't running, and it wasn't
1117 * runnable, which means that it will never become
1118 * running in the future either. We're all done!
1123 * kick_process - kick a running thread to enter/exit the kernel
1124 * @p: the to-be-kicked thread
1126 * Cause a process which is running on another CPU to enter
1127 * kernel-mode, without any delay. (to get signals handled.)
1129 * NOTE: this function doesnt have to take the runqueue lock,
1130 * because all it wants to ensure is that the remote task enters
1131 * the kernel. If the IPI races and the task has been migrated
1132 * to another CPU then no harm is done and the purpose has been
1135 void kick_process(struct task_struct
*p
)
1141 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1142 smp_send_reschedule(cpu
);
1147 * Return a low guess at the load of a migration-source cpu weighted
1148 * according to the scheduling class and "nice" value.
1150 * We want to under-estimate the load of migration sources, to
1151 * balance conservatively.
1153 static inline unsigned long source_load(int cpu
, int type
)
1155 struct rq
*rq
= cpu_rq(cpu
);
1156 unsigned long total
= weighted_cpuload(cpu
);
1161 return min(rq
->cpu_load
[type
-1], total
);
1165 * Return a high guess at the load of a migration-target cpu weighted
1166 * according to the scheduling class and "nice" value.
1168 static inline unsigned long target_load(int cpu
, int type
)
1170 struct rq
*rq
= cpu_rq(cpu
);
1171 unsigned long total
= weighted_cpuload(cpu
);
1176 return max(rq
->cpu_load
[type
-1], total
);
1180 * Return the average load per task on the cpu's run queue
1182 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1184 struct rq
*rq
= cpu_rq(cpu
);
1185 unsigned long total
= weighted_cpuload(cpu
);
1186 unsigned long n
= rq
->nr_running
;
1188 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1192 * find_idlest_group finds and returns the least busy CPU group within the
1195 static struct sched_group
*
1196 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1198 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1199 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1200 int load_idx
= sd
->forkexec_idx
;
1201 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1204 unsigned long load
, avg_load
;
1208 /* Skip over this group if it has no CPUs allowed */
1209 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1212 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1214 /* Tally up the load of all CPUs in the group */
1217 for_each_cpu_mask(i
, group
->cpumask
) {
1218 /* Bias balancing toward cpus of our domain */
1220 load
= source_load(i
, load_idx
);
1222 load
= target_load(i
, load_idx
);
1227 /* Adjust by relative CPU power of the group */
1228 avg_load
= sg_div_cpu_power(group
,
1229 avg_load
* SCHED_LOAD_SCALE
);
1232 this_load
= avg_load
;
1234 } else if (avg_load
< min_load
) {
1235 min_load
= avg_load
;
1239 group
= group
->next
;
1240 } while (group
!= sd
->groups
);
1242 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1248 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1251 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1254 unsigned long load
, min_load
= ULONG_MAX
;
1258 /* Traverse only the allowed CPUs */
1259 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1261 for_each_cpu_mask(i
, tmp
) {
1262 load
= weighted_cpuload(i
);
1264 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1274 * sched_balance_self: balance the current task (running on cpu) in domains
1275 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1278 * Balance, ie. select the least loaded group.
1280 * Returns the target CPU number, or the same CPU if no balancing is needed.
1282 * preempt must be disabled.
1284 static int sched_balance_self(int cpu
, int flag
)
1286 struct task_struct
*t
= current
;
1287 struct sched_domain
*tmp
, *sd
= NULL
;
1289 for_each_domain(cpu
, tmp
) {
1291 * If power savings logic is enabled for a domain, stop there.
1293 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1295 if (tmp
->flags
& flag
)
1301 struct sched_group
*group
;
1302 int new_cpu
, weight
;
1304 if (!(sd
->flags
& flag
)) {
1310 group
= find_idlest_group(sd
, t
, cpu
);
1316 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1317 if (new_cpu
== -1 || new_cpu
== cpu
) {
1318 /* Now try balancing at a lower domain level of cpu */
1323 /* Now try balancing at a lower domain level of new_cpu */
1326 weight
= cpus_weight(span
);
1327 for_each_domain(cpu
, tmp
) {
1328 if (weight
<= cpus_weight(tmp
->span
))
1330 if (tmp
->flags
& flag
)
1333 /* while loop will break here if sd == NULL */
1339 #endif /* CONFIG_SMP */
1342 * wake_idle() will wake a task on an idle cpu if task->cpu is
1343 * not idle and an idle cpu is available. The span of cpus to
1344 * search starts with cpus closest then further out as needed,
1345 * so we always favor a closer, idle cpu.
1347 * Returns the CPU we should wake onto.
1349 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1350 static int wake_idle(int cpu
, struct task_struct
*p
)
1353 struct sched_domain
*sd
;
1357 * If it is idle, then it is the best cpu to run this task.
1359 * This cpu is also the best, if it has more than one task already.
1360 * Siblings must be also busy(in most cases) as they didn't already
1361 * pickup the extra load from this cpu and hence we need not check
1362 * sibling runqueue info. This will avoid the checks and cache miss
1363 * penalities associated with that.
1365 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1368 for_each_domain(cpu
, sd
) {
1369 if (sd
->flags
& SD_WAKE_IDLE
) {
1370 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1371 for_each_cpu_mask(i
, tmp
) {
1382 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1389 * try_to_wake_up - wake up a thread
1390 * @p: the to-be-woken-up thread
1391 * @state: the mask of task states that can be woken
1392 * @sync: do a synchronous wakeup?
1394 * Put it on the run-queue if it's not already there. The "current"
1395 * thread is always on the run-queue (except when the actual
1396 * re-schedule is in progress), and as such you're allowed to do
1397 * the simpler "current->state = TASK_RUNNING" to mark yourself
1398 * runnable without the overhead of this.
1400 * returns failure only if the task is already active.
1402 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1404 int cpu
, this_cpu
, success
= 0;
1405 unsigned long flags
;
1409 struct sched_domain
*sd
, *this_sd
= NULL
;
1410 unsigned long load
, this_load
;
1414 rq
= task_rq_lock(p
, &flags
);
1415 old_state
= p
->state
;
1416 if (!(old_state
& state
))
1423 this_cpu
= smp_processor_id();
1426 if (unlikely(task_running(rq
, p
)))
1431 schedstat_inc(rq
, ttwu_cnt
);
1432 if (cpu
== this_cpu
) {
1433 schedstat_inc(rq
, ttwu_local
);
1437 for_each_domain(this_cpu
, sd
) {
1438 if (cpu_isset(cpu
, sd
->span
)) {
1439 schedstat_inc(sd
, ttwu_wake_remote
);
1445 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1449 * Check for affine wakeup and passive balancing possibilities.
1452 int idx
= this_sd
->wake_idx
;
1453 unsigned int imbalance
;
1455 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1457 load
= source_load(cpu
, idx
);
1458 this_load
= target_load(this_cpu
, idx
);
1460 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1462 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1463 unsigned long tl
= this_load
;
1464 unsigned long tl_per_task
;
1466 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1469 * If sync wakeup then subtract the (maximum possible)
1470 * effect of the currently running task from the load
1471 * of the current CPU:
1474 tl
-= current
->se
.load
.weight
;
1477 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1478 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1480 * This domain has SD_WAKE_AFFINE and
1481 * p is cache cold in this domain, and
1482 * there is no bad imbalance.
1484 schedstat_inc(this_sd
, ttwu_move_affine
);
1490 * Start passive balancing when half the imbalance_pct
1493 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1494 if (imbalance
*this_load
<= 100*load
) {
1495 schedstat_inc(this_sd
, ttwu_move_balance
);
1501 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1503 new_cpu
= wake_idle(new_cpu
, p
);
1504 if (new_cpu
!= cpu
) {
1505 set_task_cpu(p
, new_cpu
);
1506 task_rq_unlock(rq
, &flags
);
1507 /* might preempt at this point */
1508 rq
= task_rq_lock(p
, &flags
);
1509 old_state
= p
->state
;
1510 if (!(old_state
& state
))
1515 this_cpu
= smp_processor_id();
1520 #endif /* CONFIG_SMP */
1521 activate_task(rq
, p
, 1);
1523 * Sync wakeups (i.e. those types of wakeups where the waker
1524 * has indicated that it will leave the CPU in short order)
1525 * don't trigger a preemption, if the woken up task will run on
1526 * this cpu. (in this case the 'I will reschedule' promise of
1527 * the waker guarantees that the freshly woken up task is going
1528 * to be considered on this CPU.)
1530 if (!sync
|| cpu
!= this_cpu
)
1531 check_preempt_curr(rq
, p
);
1535 p
->state
= TASK_RUNNING
;
1537 task_rq_unlock(rq
, &flags
);
1542 int fastcall
wake_up_process(struct task_struct
*p
)
1544 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1545 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1547 EXPORT_SYMBOL(wake_up_process
);
1549 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1551 return try_to_wake_up(p
, state
, 0);
1555 * Perform scheduler related setup for a newly forked process p.
1556 * p is forked by current.
1558 * __sched_fork() is basic setup used by init_idle() too:
1560 static void __sched_fork(struct task_struct
*p
)
1562 p
->se
.wait_start_fair
= 0;
1563 p
->se
.exec_start
= 0;
1564 p
->se
.sum_exec_runtime
= 0;
1565 p
->se
.delta_exec
= 0;
1566 p
->se
.delta_fair_run
= 0;
1567 p
->se
.delta_fair_sleep
= 0;
1568 p
->se
.wait_runtime
= 0;
1569 p
->se
.sleep_start_fair
= 0;
1571 #ifdef CONFIG_SCHEDSTATS
1572 p
->se
.wait_start
= 0;
1573 p
->se
.sum_wait_runtime
= 0;
1574 p
->se
.sum_sleep_runtime
= 0;
1575 p
->se
.sleep_start
= 0;
1576 p
->se
.block_start
= 0;
1577 p
->se
.sleep_max
= 0;
1578 p
->se
.block_max
= 0;
1581 p
->se
.wait_runtime_overruns
= 0;
1582 p
->se
.wait_runtime_underruns
= 0;
1585 INIT_LIST_HEAD(&p
->run_list
);
1588 #ifdef CONFIG_PREEMPT_NOTIFIERS
1589 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1593 * We mark the process as running here, but have not actually
1594 * inserted it onto the runqueue yet. This guarantees that
1595 * nobody will actually run it, and a signal or other external
1596 * event cannot wake it up and insert it on the runqueue either.
1598 p
->state
= TASK_RUNNING
;
1602 * fork()/clone()-time setup:
1604 void sched_fork(struct task_struct
*p
, int clone_flags
)
1606 int cpu
= get_cpu();
1611 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1613 __set_task_cpu(p
, cpu
);
1616 * Make sure we do not leak PI boosting priority to the child:
1618 p
->prio
= current
->normal_prio
;
1620 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1621 if (likely(sched_info_on()))
1622 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1624 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1627 #ifdef CONFIG_PREEMPT
1628 /* Want to start with kernel preemption disabled. */
1629 task_thread_info(p
)->preempt_count
= 1;
1635 * After fork, child runs first. (default) If set to 0 then
1636 * parent will (try to) run first.
1638 unsigned int __read_mostly sysctl_sched_child_runs_first
= 1;
1641 * wake_up_new_task - wake up a newly created task for the first time.
1643 * This function will do some initial scheduler statistics housekeeping
1644 * that must be done for every newly created context, then puts the task
1645 * on the runqueue and wakes it.
1647 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1649 unsigned long flags
;
1654 rq
= task_rq_lock(p
, &flags
);
1655 BUG_ON(p
->state
!= TASK_RUNNING
);
1656 this_cpu
= smp_processor_id(); /* parent's CPU */
1657 update_rq_clock(rq
);
1660 p
->prio
= effective_prio(p
);
1662 if (!p
->sched_class
->task_new
|| !sysctl_sched_child_runs_first
||
1663 (clone_flags
& CLONE_VM
) || task_cpu(p
) != this_cpu
||
1664 !current
->se
.on_rq
) {
1666 activate_task(rq
, p
, 0);
1669 * Let the scheduling class do new task startup
1670 * management (if any):
1672 p
->sched_class
->task_new(rq
, p
);
1673 inc_nr_running(p
, rq
);
1675 check_preempt_curr(rq
, p
);
1676 task_rq_unlock(rq
, &flags
);
1679 #ifdef CONFIG_PREEMPT_NOTIFIERS
1682 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1683 * @notifier: notifier struct to register
1685 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1687 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1689 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1692 * preempt_notifier_unregister - no longer interested in preemption notifications
1693 * @notifier: notifier struct to unregister
1695 * This is safe to call from within a preemption notifier.
1697 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1699 hlist_del(¬ifier
->link
);
1701 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1703 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1705 struct preempt_notifier
*notifier
;
1706 struct hlist_node
*node
;
1708 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1709 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1713 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1714 struct task_struct
*next
)
1716 struct preempt_notifier
*notifier
;
1717 struct hlist_node
*node
;
1719 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1720 notifier
->ops
->sched_out(notifier
, next
);
1725 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1730 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1731 struct task_struct
*next
)
1738 * prepare_task_switch - prepare to switch tasks
1739 * @rq: the runqueue preparing to switch
1740 * @prev: the current task that is being switched out
1741 * @next: the task we are going to switch to.
1743 * This is called with the rq lock held and interrupts off. It must
1744 * be paired with a subsequent finish_task_switch after the context
1747 * prepare_task_switch sets up locking and calls architecture specific
1751 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1752 struct task_struct
*next
)
1754 fire_sched_out_preempt_notifiers(prev
, next
);
1755 prepare_lock_switch(rq
, next
);
1756 prepare_arch_switch(next
);
1760 * finish_task_switch - clean up after a task-switch
1761 * @rq: runqueue associated with task-switch
1762 * @prev: the thread we just switched away from.
1764 * finish_task_switch must be called after the context switch, paired
1765 * with a prepare_task_switch call before the context switch.
1766 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1767 * and do any other architecture-specific cleanup actions.
1769 * Note that we may have delayed dropping an mm in context_switch(). If
1770 * so, we finish that here outside of the runqueue lock. (Doing it
1771 * with the lock held can cause deadlocks; see schedule() for
1774 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1775 __releases(rq
->lock
)
1777 struct mm_struct
*mm
= rq
->prev_mm
;
1783 * A task struct has one reference for the use as "current".
1784 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1785 * schedule one last time. The schedule call will never return, and
1786 * the scheduled task must drop that reference.
1787 * The test for TASK_DEAD must occur while the runqueue locks are
1788 * still held, otherwise prev could be scheduled on another cpu, die
1789 * there before we look at prev->state, and then the reference would
1791 * Manfred Spraul <manfred@colorfullife.com>
1793 prev_state
= prev
->state
;
1794 finish_arch_switch(prev
);
1795 finish_lock_switch(rq
, prev
);
1796 fire_sched_in_preempt_notifiers(current
);
1799 if (unlikely(prev_state
== TASK_DEAD
)) {
1801 * Remove function-return probe instances associated with this
1802 * task and put them back on the free list.
1804 kprobe_flush_task(prev
);
1805 put_task_struct(prev
);
1810 * schedule_tail - first thing a freshly forked thread must call.
1811 * @prev: the thread we just switched away from.
1813 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1814 __releases(rq
->lock
)
1816 struct rq
*rq
= this_rq();
1818 finish_task_switch(rq
, prev
);
1819 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1820 /* In this case, finish_task_switch does not reenable preemption */
1823 if (current
->set_child_tid
)
1824 put_user(current
->pid
, current
->set_child_tid
);
1828 * context_switch - switch to the new MM and the new
1829 * thread's register state.
1832 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1833 struct task_struct
*next
)
1835 struct mm_struct
*mm
, *oldmm
;
1837 prepare_task_switch(rq
, prev
, next
);
1839 oldmm
= prev
->active_mm
;
1841 * For paravirt, this is coupled with an exit in switch_to to
1842 * combine the page table reload and the switch backend into
1845 arch_enter_lazy_cpu_mode();
1847 if (unlikely(!mm
)) {
1848 next
->active_mm
= oldmm
;
1849 atomic_inc(&oldmm
->mm_count
);
1850 enter_lazy_tlb(oldmm
, next
);
1852 switch_mm(oldmm
, mm
, next
);
1854 if (unlikely(!prev
->mm
)) {
1855 prev
->active_mm
= NULL
;
1856 rq
->prev_mm
= oldmm
;
1859 * Since the runqueue lock will be released by the next
1860 * task (which is an invalid locking op but in the case
1861 * of the scheduler it's an obvious special-case), so we
1862 * do an early lockdep release here:
1864 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1865 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1868 /* Here we just switch the register state and the stack. */
1869 switch_to(prev
, next
, prev
);
1873 * this_rq must be evaluated again because prev may have moved
1874 * CPUs since it called schedule(), thus the 'rq' on its stack
1875 * frame will be invalid.
1877 finish_task_switch(this_rq(), prev
);
1881 * nr_running, nr_uninterruptible and nr_context_switches:
1883 * externally visible scheduler statistics: current number of runnable
1884 * threads, current number of uninterruptible-sleeping threads, total
1885 * number of context switches performed since bootup.
1887 unsigned long nr_running(void)
1889 unsigned long i
, sum
= 0;
1891 for_each_online_cpu(i
)
1892 sum
+= cpu_rq(i
)->nr_running
;
1897 unsigned long nr_uninterruptible(void)
1899 unsigned long i
, sum
= 0;
1901 for_each_possible_cpu(i
)
1902 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1905 * Since we read the counters lockless, it might be slightly
1906 * inaccurate. Do not allow it to go below zero though:
1908 if (unlikely((long)sum
< 0))
1914 unsigned long long nr_context_switches(void)
1917 unsigned long long sum
= 0;
1919 for_each_possible_cpu(i
)
1920 sum
+= cpu_rq(i
)->nr_switches
;
1925 unsigned long nr_iowait(void)
1927 unsigned long i
, sum
= 0;
1929 for_each_possible_cpu(i
)
1930 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1935 unsigned long nr_active(void)
1937 unsigned long i
, running
= 0, uninterruptible
= 0;
1939 for_each_online_cpu(i
) {
1940 running
+= cpu_rq(i
)->nr_running
;
1941 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1944 if (unlikely((long)uninterruptible
< 0))
1945 uninterruptible
= 0;
1947 return running
+ uninterruptible
;
1951 * Update rq->cpu_load[] statistics. This function is usually called every
1952 * scheduler tick (TICK_NSEC).
1954 static void update_cpu_load(struct rq
*this_rq
)
1956 u64 fair_delta64
, exec_delta64
, idle_delta64
, sample_interval64
, tmp64
;
1957 unsigned long total_load
= this_rq
->ls
.load
.weight
;
1958 unsigned long this_load
= total_load
;
1959 struct load_stat
*ls
= &this_rq
->ls
;
1963 __update_rq_clock(this_rq
);
1964 now
= this_rq
->clock
;
1966 this_rq
->nr_load_updates
++;
1967 if (unlikely(!(sysctl_sched_features
& SCHED_FEAT_PRECISE_CPU_LOAD
)))
1970 /* Update delta_fair/delta_exec fields first */
1971 update_curr_load(this_rq
);
1973 fair_delta64
= ls
->delta_fair
+ 1;
1976 exec_delta64
= ls
->delta_exec
+ 1;
1979 sample_interval64
= this_rq
->clock
- ls
->load_update_last
;
1980 ls
->load_update_last
= this_rq
->clock
;
1982 if ((s64
)sample_interval64
< (s64
)TICK_NSEC
)
1983 sample_interval64
= TICK_NSEC
;
1985 if (exec_delta64
> sample_interval64
)
1986 exec_delta64
= sample_interval64
;
1988 idle_delta64
= sample_interval64
- exec_delta64
;
1990 tmp64
= div64_64(SCHED_LOAD_SCALE
* exec_delta64
, fair_delta64
);
1991 tmp64
= div64_64(tmp64
* exec_delta64
, sample_interval64
);
1993 this_load
= (unsigned long)tmp64
;
1997 /* Update our load: */
1998 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1999 unsigned long old_load
, new_load
;
2001 /* scale is effectively 1 << i now, and >> i divides by scale */
2003 old_load
= this_rq
->cpu_load
[i
];
2004 new_load
= this_load
;
2006 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2013 * double_rq_lock - safely lock two runqueues
2015 * Note this does not disable interrupts like task_rq_lock,
2016 * you need to do so manually before calling.
2018 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2019 __acquires(rq1
->lock
)
2020 __acquires(rq2
->lock
)
2022 BUG_ON(!irqs_disabled());
2024 spin_lock(&rq1
->lock
);
2025 __acquire(rq2
->lock
); /* Fake it out ;) */
2028 spin_lock(&rq1
->lock
);
2029 spin_lock(&rq2
->lock
);
2031 spin_lock(&rq2
->lock
);
2032 spin_lock(&rq1
->lock
);
2038 * double_rq_unlock - safely unlock two runqueues
2040 * Note this does not restore interrupts like task_rq_unlock,
2041 * you need to do so manually after calling.
2043 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2044 __releases(rq1
->lock
)
2045 __releases(rq2
->lock
)
2047 spin_unlock(&rq1
->lock
);
2049 spin_unlock(&rq2
->lock
);
2051 __release(rq2
->lock
);
2055 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2057 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2058 __releases(this_rq
->lock
)
2059 __acquires(busiest
->lock
)
2060 __acquires(this_rq
->lock
)
2062 if (unlikely(!irqs_disabled())) {
2063 /* printk() doesn't work good under rq->lock */
2064 spin_unlock(&this_rq
->lock
);
2067 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2068 if (busiest
< this_rq
) {
2069 spin_unlock(&this_rq
->lock
);
2070 spin_lock(&busiest
->lock
);
2071 spin_lock(&this_rq
->lock
);
2073 spin_lock(&busiest
->lock
);
2078 * If dest_cpu is allowed for this process, migrate the task to it.
2079 * This is accomplished by forcing the cpu_allowed mask to only
2080 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2081 * the cpu_allowed mask is restored.
2083 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2085 struct migration_req req
;
2086 unsigned long flags
;
2089 rq
= task_rq_lock(p
, &flags
);
2090 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2091 || unlikely(cpu_is_offline(dest_cpu
)))
2094 /* force the process onto the specified CPU */
2095 if (migrate_task(p
, dest_cpu
, &req
)) {
2096 /* Need to wait for migration thread (might exit: take ref). */
2097 struct task_struct
*mt
= rq
->migration_thread
;
2099 get_task_struct(mt
);
2100 task_rq_unlock(rq
, &flags
);
2101 wake_up_process(mt
);
2102 put_task_struct(mt
);
2103 wait_for_completion(&req
.done
);
2108 task_rq_unlock(rq
, &flags
);
2112 * sched_exec - execve() is a valuable balancing opportunity, because at
2113 * this point the task has the smallest effective memory and cache footprint.
2115 void sched_exec(void)
2117 int new_cpu
, this_cpu
= get_cpu();
2118 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2120 if (new_cpu
!= this_cpu
)
2121 sched_migrate_task(current
, new_cpu
);
2125 * pull_task - move a task from a remote runqueue to the local runqueue.
2126 * Both runqueues must be locked.
2128 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2129 struct rq
*this_rq
, int this_cpu
)
2131 update_rq_clock(src_rq
);
2132 deactivate_task(src_rq
, p
, 0, src_rq
->clock
);
2133 set_task_cpu(p
, this_cpu
);
2134 activate_task(this_rq
, p
, 0);
2136 * Note that idle threads have a prio of MAX_PRIO, for this test
2137 * to be always true for them.
2139 check_preempt_curr(this_rq
, p
);
2143 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2146 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2147 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2151 * We do not migrate tasks that are:
2152 * 1) running (obviously), or
2153 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2154 * 3) are cache-hot on their current CPU.
2156 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2160 if (task_running(rq
, p
))
2164 * Aggressive migration if too many balance attempts have failed:
2166 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2172 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2173 unsigned long max_nr_move
, unsigned long max_load_move
,
2174 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2175 int *all_pinned
, unsigned long *load_moved
,
2176 int *this_best_prio
, struct rq_iterator
*iterator
)
2178 int pulled
= 0, pinned
= 0, skip_for_load
;
2179 struct task_struct
*p
;
2180 long rem_load_move
= max_load_move
;
2182 if (max_nr_move
== 0 || max_load_move
== 0)
2188 * Start the load-balancing iterator:
2190 p
= iterator
->start(iterator
->arg
);
2195 * To help distribute high priority tasks accross CPUs we don't
2196 * skip a task if it will be the highest priority task (i.e. smallest
2197 * prio value) on its new queue regardless of its load weight
2199 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2200 SCHED_LOAD_SCALE_FUZZ
;
2201 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2202 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2203 p
= iterator
->next(iterator
->arg
);
2207 pull_task(busiest
, p
, this_rq
, this_cpu
);
2209 rem_load_move
-= p
->se
.load
.weight
;
2212 * We only want to steal up to the prescribed number of tasks
2213 * and the prescribed amount of weighted load.
2215 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2216 if (p
->prio
< *this_best_prio
)
2217 *this_best_prio
= p
->prio
;
2218 p
= iterator
->next(iterator
->arg
);
2223 * Right now, this is the only place pull_task() is called,
2224 * so we can safely collect pull_task() stats here rather than
2225 * inside pull_task().
2227 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2230 *all_pinned
= pinned
;
2231 *load_moved
= max_load_move
- rem_load_move
;
2236 * move_tasks tries to move up to max_load_move weighted load from busiest to
2237 * this_rq, as part of a balancing operation within domain "sd".
2238 * Returns 1 if successful and 0 otherwise.
2240 * Called with both runqueues locked.
2242 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2243 unsigned long max_load_move
,
2244 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2247 struct sched_class
*class = sched_class_highest
;
2248 unsigned long total_load_moved
= 0;
2249 int this_best_prio
= this_rq
->curr
->prio
;
2253 class->load_balance(this_rq
, this_cpu
, busiest
,
2254 ULONG_MAX
, max_load_move
- total_load_moved
,
2255 sd
, idle
, all_pinned
, &this_best_prio
);
2256 class = class->next
;
2257 } while (class && max_load_move
> total_load_moved
);
2259 return total_load_moved
> 0;
2263 * move_one_task tries to move exactly one task from busiest to this_rq, as
2264 * part of active balancing operations within "domain".
2265 * Returns 1 if successful and 0 otherwise.
2267 * Called with both runqueues locked.
2269 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2270 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2272 struct sched_class
*class;
2273 int this_best_prio
= MAX_PRIO
;
2275 for (class = sched_class_highest
; class; class = class->next
)
2276 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2277 1, ULONG_MAX
, sd
, idle
, NULL
,
2285 * find_busiest_group finds and returns the busiest CPU group within the
2286 * domain. It calculates and returns the amount of weighted load which
2287 * should be moved to restore balance via the imbalance parameter.
2289 static struct sched_group
*
2290 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2291 unsigned long *imbalance
, enum cpu_idle_type idle
,
2292 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2294 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2295 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2296 unsigned long max_pull
;
2297 unsigned long busiest_load_per_task
, busiest_nr_running
;
2298 unsigned long this_load_per_task
, this_nr_running
;
2300 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2301 int power_savings_balance
= 1;
2302 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2303 unsigned long min_nr_running
= ULONG_MAX
;
2304 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2307 max_load
= this_load
= total_load
= total_pwr
= 0;
2308 busiest_load_per_task
= busiest_nr_running
= 0;
2309 this_load_per_task
= this_nr_running
= 0;
2310 if (idle
== CPU_NOT_IDLE
)
2311 load_idx
= sd
->busy_idx
;
2312 else if (idle
== CPU_NEWLY_IDLE
)
2313 load_idx
= sd
->newidle_idx
;
2315 load_idx
= sd
->idle_idx
;
2318 unsigned long load
, group_capacity
;
2321 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2322 unsigned long sum_nr_running
, sum_weighted_load
;
2324 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2327 balance_cpu
= first_cpu(group
->cpumask
);
2329 /* Tally up the load of all CPUs in the group */
2330 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2332 for_each_cpu_mask(i
, group
->cpumask
) {
2335 if (!cpu_isset(i
, *cpus
))
2340 if (*sd_idle
&& rq
->nr_running
)
2343 /* Bias balancing toward cpus of our domain */
2345 if (idle_cpu(i
) && !first_idle_cpu
) {
2350 load
= target_load(i
, load_idx
);
2352 load
= source_load(i
, load_idx
);
2355 sum_nr_running
+= rq
->nr_running
;
2356 sum_weighted_load
+= weighted_cpuload(i
);
2360 * First idle cpu or the first cpu(busiest) in this sched group
2361 * is eligible for doing load balancing at this and above
2362 * domains. In the newly idle case, we will allow all the cpu's
2363 * to do the newly idle load balance.
2365 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2366 balance_cpu
!= this_cpu
&& balance
) {
2371 total_load
+= avg_load
;
2372 total_pwr
+= group
->__cpu_power
;
2374 /* Adjust by relative CPU power of the group */
2375 avg_load
= sg_div_cpu_power(group
,
2376 avg_load
* SCHED_LOAD_SCALE
);
2378 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2381 this_load
= avg_load
;
2383 this_nr_running
= sum_nr_running
;
2384 this_load_per_task
= sum_weighted_load
;
2385 } else if (avg_load
> max_load
&&
2386 sum_nr_running
> group_capacity
) {
2387 max_load
= avg_load
;
2389 busiest_nr_running
= sum_nr_running
;
2390 busiest_load_per_task
= sum_weighted_load
;
2393 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2395 * Busy processors will not participate in power savings
2398 if (idle
== CPU_NOT_IDLE
||
2399 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2403 * If the local group is idle or completely loaded
2404 * no need to do power savings balance at this domain
2406 if (local_group
&& (this_nr_running
>= group_capacity
||
2408 power_savings_balance
= 0;
2411 * If a group is already running at full capacity or idle,
2412 * don't include that group in power savings calculations
2414 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2419 * Calculate the group which has the least non-idle load.
2420 * This is the group from where we need to pick up the load
2423 if ((sum_nr_running
< min_nr_running
) ||
2424 (sum_nr_running
== min_nr_running
&&
2425 first_cpu(group
->cpumask
) <
2426 first_cpu(group_min
->cpumask
))) {
2428 min_nr_running
= sum_nr_running
;
2429 min_load_per_task
= sum_weighted_load
/
2434 * Calculate the group which is almost near its
2435 * capacity but still has some space to pick up some load
2436 * from other group and save more power
2438 if (sum_nr_running
<= group_capacity
- 1) {
2439 if (sum_nr_running
> leader_nr_running
||
2440 (sum_nr_running
== leader_nr_running
&&
2441 first_cpu(group
->cpumask
) >
2442 first_cpu(group_leader
->cpumask
))) {
2443 group_leader
= group
;
2444 leader_nr_running
= sum_nr_running
;
2449 group
= group
->next
;
2450 } while (group
!= sd
->groups
);
2452 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2455 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2457 if (this_load
>= avg_load
||
2458 100*max_load
<= sd
->imbalance_pct
*this_load
)
2461 busiest_load_per_task
/= busiest_nr_running
;
2463 * We're trying to get all the cpus to the average_load, so we don't
2464 * want to push ourselves above the average load, nor do we wish to
2465 * reduce the max loaded cpu below the average load, as either of these
2466 * actions would just result in more rebalancing later, and ping-pong
2467 * tasks around. Thus we look for the minimum possible imbalance.
2468 * Negative imbalances (*we* are more loaded than anyone else) will
2469 * be counted as no imbalance for these purposes -- we can't fix that
2470 * by pulling tasks to us. Be careful of negative numbers as they'll
2471 * appear as very large values with unsigned longs.
2473 if (max_load
<= busiest_load_per_task
)
2477 * In the presence of smp nice balancing, certain scenarios can have
2478 * max load less than avg load(as we skip the groups at or below
2479 * its cpu_power, while calculating max_load..)
2481 if (max_load
< avg_load
) {
2483 goto small_imbalance
;
2486 /* Don't want to pull so many tasks that a group would go idle */
2487 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2489 /* How much load to actually move to equalise the imbalance */
2490 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2491 (avg_load
- this_load
) * this->__cpu_power
)
2495 * if *imbalance is less than the average load per runnable task
2496 * there is no gaurantee that any tasks will be moved so we'll have
2497 * a think about bumping its value to force at least one task to be
2500 if (*imbalance
+ SCHED_LOAD_SCALE_FUZZ
< busiest_load_per_task
/2) {
2501 unsigned long tmp
, pwr_now
, pwr_move
;
2505 pwr_move
= pwr_now
= 0;
2507 if (this_nr_running
) {
2508 this_load_per_task
/= this_nr_running
;
2509 if (busiest_load_per_task
> this_load_per_task
)
2512 this_load_per_task
= SCHED_LOAD_SCALE
;
2514 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2515 busiest_load_per_task
* imbn
) {
2516 *imbalance
= busiest_load_per_task
;
2521 * OK, we don't have enough imbalance to justify moving tasks,
2522 * however we may be able to increase total CPU power used by
2526 pwr_now
+= busiest
->__cpu_power
*
2527 min(busiest_load_per_task
, max_load
);
2528 pwr_now
+= this->__cpu_power
*
2529 min(this_load_per_task
, this_load
);
2530 pwr_now
/= SCHED_LOAD_SCALE
;
2532 /* Amount of load we'd subtract */
2533 tmp
= sg_div_cpu_power(busiest
,
2534 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2536 pwr_move
+= busiest
->__cpu_power
*
2537 min(busiest_load_per_task
, max_load
- tmp
);
2539 /* Amount of load we'd add */
2540 if (max_load
* busiest
->__cpu_power
<
2541 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2542 tmp
= sg_div_cpu_power(this,
2543 max_load
* busiest
->__cpu_power
);
2545 tmp
= sg_div_cpu_power(this,
2546 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2547 pwr_move
+= this->__cpu_power
*
2548 min(this_load_per_task
, this_load
+ tmp
);
2549 pwr_move
/= SCHED_LOAD_SCALE
;
2551 /* Move if we gain throughput */
2552 if (pwr_move
<= pwr_now
)
2555 *imbalance
= busiest_load_per_task
;
2561 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2562 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2565 if (this == group_leader
&& group_leader
!= group_min
) {
2566 *imbalance
= min_load_per_task
;
2576 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2579 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2580 unsigned long imbalance
, cpumask_t
*cpus
)
2582 struct rq
*busiest
= NULL
, *rq
;
2583 unsigned long max_load
= 0;
2586 for_each_cpu_mask(i
, group
->cpumask
) {
2589 if (!cpu_isset(i
, *cpus
))
2593 wl
= weighted_cpuload(i
);
2595 if (rq
->nr_running
== 1 && wl
> imbalance
)
2598 if (wl
> max_load
) {
2608 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2609 * so long as it is large enough.
2611 #define MAX_PINNED_INTERVAL 512
2614 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2615 * tasks if there is an imbalance.
2617 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2618 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2621 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2622 struct sched_group
*group
;
2623 unsigned long imbalance
;
2625 cpumask_t cpus
= CPU_MASK_ALL
;
2626 unsigned long flags
;
2629 * When power savings policy is enabled for the parent domain, idle
2630 * sibling can pick up load irrespective of busy siblings. In this case,
2631 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2632 * portraying it as CPU_NOT_IDLE.
2634 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2635 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2638 schedstat_inc(sd
, lb_cnt
[idle
]);
2641 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2648 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2652 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2654 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2658 BUG_ON(busiest
== this_rq
);
2660 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2663 if (busiest
->nr_running
> 1) {
2665 * Attempt to move tasks. If find_busiest_group has found
2666 * an imbalance but busiest->nr_running <= 1, the group is
2667 * still unbalanced. ld_moved simply stays zero, so it is
2668 * correctly treated as an imbalance.
2670 local_irq_save(flags
);
2671 double_rq_lock(this_rq
, busiest
);
2672 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2673 imbalance
, sd
, idle
, &all_pinned
);
2674 double_rq_unlock(this_rq
, busiest
);
2675 local_irq_restore(flags
);
2678 * some other cpu did the load balance for us.
2680 if (ld_moved
&& this_cpu
!= smp_processor_id())
2681 resched_cpu(this_cpu
);
2683 /* All tasks on this runqueue were pinned by CPU affinity */
2684 if (unlikely(all_pinned
)) {
2685 cpu_clear(cpu_of(busiest
), cpus
);
2686 if (!cpus_empty(cpus
))
2693 schedstat_inc(sd
, lb_failed
[idle
]);
2694 sd
->nr_balance_failed
++;
2696 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2698 spin_lock_irqsave(&busiest
->lock
, flags
);
2700 /* don't kick the migration_thread, if the curr
2701 * task on busiest cpu can't be moved to this_cpu
2703 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2704 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2706 goto out_one_pinned
;
2709 if (!busiest
->active_balance
) {
2710 busiest
->active_balance
= 1;
2711 busiest
->push_cpu
= this_cpu
;
2714 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2716 wake_up_process(busiest
->migration_thread
);
2719 * We've kicked active balancing, reset the failure
2722 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2725 sd
->nr_balance_failed
= 0;
2727 if (likely(!active_balance
)) {
2728 /* We were unbalanced, so reset the balancing interval */
2729 sd
->balance_interval
= sd
->min_interval
;
2732 * If we've begun active balancing, start to back off. This
2733 * case may not be covered by the all_pinned logic if there
2734 * is only 1 task on the busy runqueue (because we don't call
2737 if (sd
->balance_interval
< sd
->max_interval
)
2738 sd
->balance_interval
*= 2;
2741 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2742 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2747 schedstat_inc(sd
, lb_balanced
[idle
]);
2749 sd
->nr_balance_failed
= 0;
2752 /* tune up the balancing interval */
2753 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2754 (sd
->balance_interval
< sd
->max_interval
))
2755 sd
->balance_interval
*= 2;
2757 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2758 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2764 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2765 * tasks if there is an imbalance.
2767 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2768 * this_rq is locked.
2771 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2773 struct sched_group
*group
;
2774 struct rq
*busiest
= NULL
;
2775 unsigned long imbalance
;
2779 cpumask_t cpus
= CPU_MASK_ALL
;
2782 * When power savings policy is enabled for the parent domain, idle
2783 * sibling can pick up load irrespective of busy siblings. In this case,
2784 * let the state of idle sibling percolate up as IDLE, instead of
2785 * portraying it as CPU_NOT_IDLE.
2787 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2788 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2791 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2793 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2794 &sd_idle
, &cpus
, NULL
);
2796 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2800 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2803 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2807 BUG_ON(busiest
== this_rq
);
2809 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2812 if (busiest
->nr_running
> 1) {
2813 /* Attempt to move tasks */
2814 double_lock_balance(this_rq
, busiest
);
2815 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2816 imbalance
, sd
, CPU_NEWLY_IDLE
,
2818 spin_unlock(&busiest
->lock
);
2820 if (unlikely(all_pinned
)) {
2821 cpu_clear(cpu_of(busiest
), cpus
);
2822 if (!cpus_empty(cpus
))
2828 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2829 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2830 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2833 sd
->nr_balance_failed
= 0;
2838 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2839 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2840 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2842 sd
->nr_balance_failed
= 0;
2848 * idle_balance is called by schedule() if this_cpu is about to become
2849 * idle. Attempts to pull tasks from other CPUs.
2851 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2853 struct sched_domain
*sd
;
2854 int pulled_task
= -1;
2855 unsigned long next_balance
= jiffies
+ HZ
;
2857 for_each_domain(this_cpu
, sd
) {
2858 unsigned long interval
;
2860 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2863 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2864 /* If we've pulled tasks over stop searching: */
2865 pulled_task
= load_balance_newidle(this_cpu
,
2868 interval
= msecs_to_jiffies(sd
->balance_interval
);
2869 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2870 next_balance
= sd
->last_balance
+ interval
;
2874 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2876 * We are going idle. next_balance may be set based on
2877 * a busy processor. So reset next_balance.
2879 this_rq
->next_balance
= next_balance
;
2884 * active_load_balance is run by migration threads. It pushes running tasks
2885 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2886 * running on each physical CPU where possible, and avoids physical /
2887 * logical imbalances.
2889 * Called with busiest_rq locked.
2891 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2893 int target_cpu
= busiest_rq
->push_cpu
;
2894 struct sched_domain
*sd
;
2895 struct rq
*target_rq
;
2897 /* Is there any task to move? */
2898 if (busiest_rq
->nr_running
<= 1)
2901 target_rq
= cpu_rq(target_cpu
);
2904 * This condition is "impossible", if it occurs
2905 * we need to fix it. Originally reported by
2906 * Bjorn Helgaas on a 128-cpu setup.
2908 BUG_ON(busiest_rq
== target_rq
);
2910 /* move a task from busiest_rq to target_rq */
2911 double_lock_balance(busiest_rq
, target_rq
);
2913 /* Search for an sd spanning us and the target CPU. */
2914 for_each_domain(target_cpu
, sd
) {
2915 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2916 cpu_isset(busiest_cpu
, sd
->span
))
2921 schedstat_inc(sd
, alb_cnt
);
2923 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2925 schedstat_inc(sd
, alb_pushed
);
2927 schedstat_inc(sd
, alb_failed
);
2929 spin_unlock(&target_rq
->lock
);
2934 atomic_t load_balancer
;
2936 } nohz ____cacheline_aligned
= {
2937 .load_balancer
= ATOMIC_INIT(-1),
2938 .cpu_mask
= CPU_MASK_NONE
,
2942 * This routine will try to nominate the ilb (idle load balancing)
2943 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2944 * load balancing on behalf of all those cpus. If all the cpus in the system
2945 * go into this tickless mode, then there will be no ilb owner (as there is
2946 * no need for one) and all the cpus will sleep till the next wakeup event
2949 * For the ilb owner, tick is not stopped. And this tick will be used
2950 * for idle load balancing. ilb owner will still be part of
2953 * While stopping the tick, this cpu will become the ilb owner if there
2954 * is no other owner. And will be the owner till that cpu becomes busy
2955 * or if all cpus in the system stop their ticks at which point
2956 * there is no need for ilb owner.
2958 * When the ilb owner becomes busy, it nominates another owner, during the
2959 * next busy scheduler_tick()
2961 int select_nohz_load_balancer(int stop_tick
)
2963 int cpu
= smp_processor_id();
2966 cpu_set(cpu
, nohz
.cpu_mask
);
2967 cpu_rq(cpu
)->in_nohz_recently
= 1;
2970 * If we are going offline and still the leader, give up!
2972 if (cpu_is_offline(cpu
) &&
2973 atomic_read(&nohz
.load_balancer
) == cpu
) {
2974 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2979 /* time for ilb owner also to sleep */
2980 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2981 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2982 atomic_set(&nohz
.load_balancer
, -1);
2986 if (atomic_read(&nohz
.load_balancer
) == -1) {
2987 /* make me the ilb owner */
2988 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2990 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2993 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2996 cpu_clear(cpu
, nohz
.cpu_mask
);
2998 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2999 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3006 static DEFINE_SPINLOCK(balancing
);
3009 * It checks each scheduling domain to see if it is due to be balanced,
3010 * and initiates a balancing operation if so.
3012 * Balancing parameters are set up in arch_init_sched_domains.
3014 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3017 struct rq
*rq
= cpu_rq(cpu
);
3018 unsigned long interval
;
3019 struct sched_domain
*sd
;
3020 /* Earliest time when we have to do rebalance again */
3021 unsigned long next_balance
= jiffies
+ 60*HZ
;
3023 for_each_domain(cpu
, sd
) {
3024 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3027 interval
= sd
->balance_interval
;
3028 if (idle
!= CPU_IDLE
)
3029 interval
*= sd
->busy_factor
;
3031 /* scale ms to jiffies */
3032 interval
= msecs_to_jiffies(interval
);
3033 if (unlikely(!interval
))
3035 if (interval
> HZ
*NR_CPUS
/10)
3036 interval
= HZ
*NR_CPUS
/10;
3039 if (sd
->flags
& SD_SERIALIZE
) {
3040 if (!spin_trylock(&balancing
))
3044 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3045 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3047 * We've pulled tasks over so either we're no
3048 * longer idle, or one of our SMT siblings is
3051 idle
= CPU_NOT_IDLE
;
3053 sd
->last_balance
= jiffies
;
3055 if (sd
->flags
& SD_SERIALIZE
)
3056 spin_unlock(&balancing
);
3058 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3059 next_balance
= sd
->last_balance
+ interval
;
3062 * Stop the load balance at this level. There is another
3063 * CPU in our sched group which is doing load balancing more
3069 rq
->next_balance
= next_balance
;
3073 * run_rebalance_domains is triggered when needed from the scheduler tick.
3074 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3075 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3077 static void run_rebalance_domains(struct softirq_action
*h
)
3079 int this_cpu
= smp_processor_id();
3080 struct rq
*this_rq
= cpu_rq(this_cpu
);
3081 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3082 CPU_IDLE
: CPU_NOT_IDLE
;
3084 rebalance_domains(this_cpu
, idle
);
3088 * If this cpu is the owner for idle load balancing, then do the
3089 * balancing on behalf of the other idle cpus whose ticks are
3092 if (this_rq
->idle_at_tick
&&
3093 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3094 cpumask_t cpus
= nohz
.cpu_mask
;
3098 cpu_clear(this_cpu
, cpus
);
3099 for_each_cpu_mask(balance_cpu
, cpus
) {
3101 * If this cpu gets work to do, stop the load balancing
3102 * work being done for other cpus. Next load
3103 * balancing owner will pick it up.
3108 rebalance_domains(balance_cpu
, SCHED_IDLE
);
3110 rq
= cpu_rq(balance_cpu
);
3111 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3112 this_rq
->next_balance
= rq
->next_balance
;
3119 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3121 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3122 * idle load balancing owner or decide to stop the periodic load balancing,
3123 * if the whole system is idle.
3125 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3129 * If we were in the nohz mode recently and busy at the current
3130 * scheduler tick, then check if we need to nominate new idle
3133 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3134 rq
->in_nohz_recently
= 0;
3136 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3137 cpu_clear(cpu
, nohz
.cpu_mask
);
3138 atomic_set(&nohz
.load_balancer
, -1);
3141 if (atomic_read(&nohz
.load_balancer
) == -1) {
3143 * simple selection for now: Nominate the
3144 * first cpu in the nohz list to be the next
3147 * TBD: Traverse the sched domains and nominate
3148 * the nearest cpu in the nohz.cpu_mask.
3150 int ilb
= first_cpu(nohz
.cpu_mask
);
3158 * If this cpu is idle and doing idle load balancing for all the
3159 * cpus with ticks stopped, is it time for that to stop?
3161 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3162 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3168 * If this cpu is idle and the idle load balancing is done by
3169 * someone else, then no need raise the SCHED_SOFTIRQ
3171 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3172 cpu_isset(cpu
, nohz
.cpu_mask
))
3175 if (time_after_eq(jiffies
, rq
->next_balance
))
3176 raise_softirq(SCHED_SOFTIRQ
);
3179 #else /* CONFIG_SMP */
3182 * on UP we do not need to balance between CPUs:
3184 static inline void idle_balance(int cpu
, struct rq
*rq
)
3188 /* Avoid "used but not defined" warning on UP */
3189 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3190 unsigned long max_nr_move
, unsigned long max_load_move
,
3191 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3192 int *all_pinned
, unsigned long *load_moved
,
3193 int *this_best_prio
, struct rq_iterator
*iterator
)
3202 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3204 EXPORT_PER_CPU_SYMBOL(kstat
);
3207 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3208 * that have not yet been banked in case the task is currently running.
3210 unsigned long long task_sched_runtime(struct task_struct
*p
)
3212 unsigned long flags
;
3216 rq
= task_rq_lock(p
, &flags
);
3217 ns
= p
->se
.sum_exec_runtime
;
3218 if (rq
->curr
== p
) {
3219 update_rq_clock(rq
);
3220 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3221 if ((s64
)delta_exec
> 0)
3224 task_rq_unlock(rq
, &flags
);
3230 * Account user cpu time to a process.
3231 * @p: the process that the cpu time gets accounted to
3232 * @hardirq_offset: the offset to subtract from hardirq_count()
3233 * @cputime: the cpu time spent in user space since the last update
3235 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3237 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3240 p
->utime
= cputime_add(p
->utime
, cputime
);
3242 /* Add user time to cpustat. */
3243 tmp
= cputime_to_cputime64(cputime
);
3244 if (TASK_NICE(p
) > 0)
3245 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3247 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3251 * Account system cpu time to a process.
3252 * @p: the process that the cpu time gets accounted to
3253 * @hardirq_offset: the offset to subtract from hardirq_count()
3254 * @cputime: the cpu time spent in kernel space since the last update
3256 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3259 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3260 struct rq
*rq
= this_rq();
3263 p
->stime
= cputime_add(p
->stime
, cputime
);
3265 /* Add system time to cpustat. */
3266 tmp
= cputime_to_cputime64(cputime
);
3267 if (hardirq_count() - hardirq_offset
)
3268 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3269 else if (softirq_count())
3270 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3271 else if (p
!= rq
->idle
)
3272 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3273 else if (atomic_read(&rq
->nr_iowait
) > 0)
3274 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3276 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3277 /* Account for system time used */
3278 acct_update_integrals(p
);
3282 * Account for involuntary wait time.
3283 * @p: the process from which the cpu time has been stolen
3284 * @steal: the cpu time spent in involuntary wait
3286 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3288 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3289 cputime64_t tmp
= cputime_to_cputime64(steal
);
3290 struct rq
*rq
= this_rq();
3292 if (p
== rq
->idle
) {
3293 p
->stime
= cputime_add(p
->stime
, steal
);
3294 if (atomic_read(&rq
->nr_iowait
) > 0)
3295 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3297 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3299 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3303 * This function gets called by the timer code, with HZ frequency.
3304 * We call it with interrupts disabled.
3306 * It also gets called by the fork code, when changing the parent's
3309 void scheduler_tick(void)
3311 int cpu
= smp_processor_id();
3312 struct rq
*rq
= cpu_rq(cpu
);
3313 struct task_struct
*curr
= rq
->curr
;
3315 spin_lock(&rq
->lock
);
3316 update_cpu_load(rq
);
3317 if (curr
!= rq
->idle
) /* FIXME: needed? */
3318 curr
->sched_class
->task_tick(rq
, curr
);
3319 spin_unlock(&rq
->lock
);
3322 rq
->idle_at_tick
= idle_cpu(cpu
);
3323 trigger_load_balance(rq
, cpu
);
3327 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3329 void fastcall
add_preempt_count(int val
)
3334 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3336 preempt_count() += val
;
3338 * Spinlock count overflowing soon?
3340 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3343 EXPORT_SYMBOL(add_preempt_count
);
3345 void fastcall
sub_preempt_count(int val
)
3350 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3353 * Is the spinlock portion underflowing?
3355 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3356 !(preempt_count() & PREEMPT_MASK
)))
3359 preempt_count() -= val
;
3361 EXPORT_SYMBOL(sub_preempt_count
);
3366 * Print scheduling while atomic bug:
3368 static noinline
void __schedule_bug(struct task_struct
*prev
)
3370 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3371 prev
->comm
, preempt_count(), prev
->pid
);
3372 debug_show_held_locks(prev
);
3373 if (irqs_disabled())
3374 print_irqtrace_events(prev
);
3379 * Various schedule()-time debugging checks and statistics:
3381 static inline void schedule_debug(struct task_struct
*prev
)
3384 * Test if we are atomic. Since do_exit() needs to call into
3385 * schedule() atomically, we ignore that path for now.
3386 * Otherwise, whine if we are scheduling when we should not be.
3388 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3389 __schedule_bug(prev
);
3391 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3393 schedstat_inc(this_rq(), sched_cnt
);
3397 * Pick up the highest-prio task:
3399 static inline struct task_struct
*
3400 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3402 struct sched_class
*class;
3403 struct task_struct
*p
;
3406 * Optimization: we know that if all tasks are in
3407 * the fair class we can call that function directly:
3409 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3410 p
= fair_sched_class
.pick_next_task(rq
);
3415 class = sched_class_highest
;
3417 p
= class->pick_next_task(rq
);
3421 * Will never be NULL as the idle class always
3422 * returns a non-NULL p:
3424 class = class->next
;
3429 * schedule() is the main scheduler function.
3431 asmlinkage
void __sched
schedule(void)
3433 struct task_struct
*prev
, *next
;
3441 cpu
= smp_processor_id();
3445 switch_count
= &prev
->nivcsw
;
3447 release_kernel_lock(prev
);
3448 need_resched_nonpreemptible
:
3450 schedule_debug(prev
);
3452 spin_lock_irq(&rq
->lock
);
3453 clear_tsk_need_resched(prev
);
3454 __update_rq_clock(rq
);
3457 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3458 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3459 unlikely(signal_pending(prev
)))) {
3460 prev
->state
= TASK_RUNNING
;
3462 deactivate_task(rq
, prev
, 1, now
);
3464 switch_count
= &prev
->nvcsw
;
3467 if (unlikely(!rq
->nr_running
))
3468 idle_balance(cpu
, rq
);
3470 prev
->sched_class
->put_prev_task(rq
, prev
);
3471 next
= pick_next_task(rq
, prev
);
3473 sched_info_switch(prev
, next
);
3475 if (likely(prev
!= next
)) {
3480 context_switch(rq
, prev
, next
); /* unlocks the rq */
3482 spin_unlock_irq(&rq
->lock
);
3484 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3485 cpu
= smp_processor_id();
3487 goto need_resched_nonpreemptible
;
3489 preempt_enable_no_resched();
3490 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3493 EXPORT_SYMBOL(schedule
);
3495 #ifdef CONFIG_PREEMPT
3497 * this is the entry point to schedule() from in-kernel preemption
3498 * off of preempt_enable. Kernel preemptions off return from interrupt
3499 * occur there and call schedule directly.
3501 asmlinkage
void __sched
preempt_schedule(void)
3503 struct thread_info
*ti
= current_thread_info();
3504 #ifdef CONFIG_PREEMPT_BKL
3505 struct task_struct
*task
= current
;
3506 int saved_lock_depth
;
3509 * If there is a non-zero preempt_count or interrupts are disabled,
3510 * we do not want to preempt the current task. Just return..
3512 if (likely(ti
->preempt_count
|| irqs_disabled()))
3516 add_preempt_count(PREEMPT_ACTIVE
);
3518 * We keep the big kernel semaphore locked, but we
3519 * clear ->lock_depth so that schedule() doesnt
3520 * auto-release the semaphore:
3522 #ifdef CONFIG_PREEMPT_BKL
3523 saved_lock_depth
= task
->lock_depth
;
3524 task
->lock_depth
= -1;
3527 #ifdef CONFIG_PREEMPT_BKL
3528 task
->lock_depth
= saved_lock_depth
;
3530 sub_preempt_count(PREEMPT_ACTIVE
);
3532 /* we could miss a preemption opportunity between schedule and now */
3534 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3537 EXPORT_SYMBOL(preempt_schedule
);
3540 * this is the entry point to schedule() from kernel preemption
3541 * off of irq context.
3542 * Note, that this is called and return with irqs disabled. This will
3543 * protect us against recursive calling from irq.
3545 asmlinkage
void __sched
preempt_schedule_irq(void)
3547 struct thread_info
*ti
= current_thread_info();
3548 #ifdef CONFIG_PREEMPT_BKL
3549 struct task_struct
*task
= current
;
3550 int saved_lock_depth
;
3552 /* Catch callers which need to be fixed */
3553 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3556 add_preempt_count(PREEMPT_ACTIVE
);
3558 * We keep the big kernel semaphore locked, but we
3559 * clear ->lock_depth so that schedule() doesnt
3560 * auto-release the semaphore:
3562 #ifdef CONFIG_PREEMPT_BKL
3563 saved_lock_depth
= task
->lock_depth
;
3564 task
->lock_depth
= -1;
3568 local_irq_disable();
3569 #ifdef CONFIG_PREEMPT_BKL
3570 task
->lock_depth
= saved_lock_depth
;
3572 sub_preempt_count(PREEMPT_ACTIVE
);
3574 /* we could miss a preemption opportunity between schedule and now */
3576 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3580 #endif /* CONFIG_PREEMPT */
3582 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3585 return try_to_wake_up(curr
->private, mode
, sync
);
3587 EXPORT_SYMBOL(default_wake_function
);
3590 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3591 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3592 * number) then we wake all the non-exclusive tasks and one exclusive task.
3594 * There are circumstances in which we can try to wake a task which has already
3595 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3596 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3598 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3599 int nr_exclusive
, int sync
, void *key
)
3601 struct list_head
*tmp
, *next
;
3603 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3604 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3605 unsigned flags
= curr
->flags
;
3607 if (curr
->func(curr
, mode
, sync
, key
) &&
3608 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3614 * __wake_up - wake up threads blocked on a waitqueue.
3616 * @mode: which threads
3617 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3618 * @key: is directly passed to the wakeup function
3620 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3621 int nr_exclusive
, void *key
)
3623 unsigned long flags
;
3625 spin_lock_irqsave(&q
->lock
, flags
);
3626 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3627 spin_unlock_irqrestore(&q
->lock
, flags
);
3629 EXPORT_SYMBOL(__wake_up
);
3632 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3634 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3636 __wake_up_common(q
, mode
, 1, 0, NULL
);
3640 * __wake_up_sync - wake up threads blocked on a waitqueue.
3642 * @mode: which threads
3643 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3645 * The sync wakeup differs that the waker knows that it will schedule
3646 * away soon, so while the target thread will be woken up, it will not
3647 * be migrated to another CPU - ie. the two threads are 'synchronized'
3648 * with each other. This can prevent needless bouncing between CPUs.
3650 * On UP it can prevent extra preemption.
3653 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3655 unsigned long flags
;
3661 if (unlikely(!nr_exclusive
))
3664 spin_lock_irqsave(&q
->lock
, flags
);
3665 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3666 spin_unlock_irqrestore(&q
->lock
, flags
);
3668 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3670 void fastcall
complete(struct completion
*x
)
3672 unsigned long flags
;
3674 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3676 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3678 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3680 EXPORT_SYMBOL(complete
);
3682 void fastcall
complete_all(struct completion
*x
)
3684 unsigned long flags
;
3686 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3687 x
->done
+= UINT_MAX
/2;
3688 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3690 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3692 EXPORT_SYMBOL(complete_all
);
3694 void fastcall __sched
wait_for_completion(struct completion
*x
)
3698 spin_lock_irq(&x
->wait
.lock
);
3700 DECLARE_WAITQUEUE(wait
, current
);
3702 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3703 __add_wait_queue_tail(&x
->wait
, &wait
);
3705 __set_current_state(TASK_UNINTERRUPTIBLE
);
3706 spin_unlock_irq(&x
->wait
.lock
);
3708 spin_lock_irq(&x
->wait
.lock
);
3710 __remove_wait_queue(&x
->wait
, &wait
);
3713 spin_unlock_irq(&x
->wait
.lock
);
3715 EXPORT_SYMBOL(wait_for_completion
);
3717 unsigned long fastcall __sched
3718 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3722 spin_lock_irq(&x
->wait
.lock
);
3724 DECLARE_WAITQUEUE(wait
, current
);
3726 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3727 __add_wait_queue_tail(&x
->wait
, &wait
);
3729 __set_current_state(TASK_UNINTERRUPTIBLE
);
3730 spin_unlock_irq(&x
->wait
.lock
);
3731 timeout
= schedule_timeout(timeout
);
3732 spin_lock_irq(&x
->wait
.lock
);
3734 __remove_wait_queue(&x
->wait
, &wait
);
3738 __remove_wait_queue(&x
->wait
, &wait
);
3742 spin_unlock_irq(&x
->wait
.lock
);
3745 EXPORT_SYMBOL(wait_for_completion_timeout
);
3747 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3753 spin_lock_irq(&x
->wait
.lock
);
3755 DECLARE_WAITQUEUE(wait
, current
);
3757 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3758 __add_wait_queue_tail(&x
->wait
, &wait
);
3760 if (signal_pending(current
)) {
3762 __remove_wait_queue(&x
->wait
, &wait
);
3765 __set_current_state(TASK_INTERRUPTIBLE
);
3766 spin_unlock_irq(&x
->wait
.lock
);
3768 spin_lock_irq(&x
->wait
.lock
);
3770 __remove_wait_queue(&x
->wait
, &wait
);
3774 spin_unlock_irq(&x
->wait
.lock
);
3778 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3780 unsigned long fastcall __sched
3781 wait_for_completion_interruptible_timeout(struct completion
*x
,
3782 unsigned long timeout
)
3786 spin_lock_irq(&x
->wait
.lock
);
3788 DECLARE_WAITQUEUE(wait
, current
);
3790 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3791 __add_wait_queue_tail(&x
->wait
, &wait
);
3793 if (signal_pending(current
)) {
3794 timeout
= -ERESTARTSYS
;
3795 __remove_wait_queue(&x
->wait
, &wait
);
3798 __set_current_state(TASK_INTERRUPTIBLE
);
3799 spin_unlock_irq(&x
->wait
.lock
);
3800 timeout
= schedule_timeout(timeout
);
3801 spin_lock_irq(&x
->wait
.lock
);
3803 __remove_wait_queue(&x
->wait
, &wait
);
3807 __remove_wait_queue(&x
->wait
, &wait
);
3811 spin_unlock_irq(&x
->wait
.lock
);
3814 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3817 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3819 spin_lock_irqsave(&q
->lock
, *flags
);
3820 __add_wait_queue(q
, wait
);
3821 spin_unlock(&q
->lock
);
3825 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3827 spin_lock_irq(&q
->lock
);
3828 __remove_wait_queue(q
, wait
);
3829 spin_unlock_irqrestore(&q
->lock
, *flags
);
3832 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3834 unsigned long flags
;
3837 init_waitqueue_entry(&wait
, current
);
3839 current
->state
= TASK_INTERRUPTIBLE
;
3841 sleep_on_head(q
, &wait
, &flags
);
3843 sleep_on_tail(q
, &wait
, &flags
);
3845 EXPORT_SYMBOL(interruptible_sleep_on
);
3848 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3850 unsigned long flags
;
3853 init_waitqueue_entry(&wait
, current
);
3855 current
->state
= TASK_INTERRUPTIBLE
;
3857 sleep_on_head(q
, &wait
, &flags
);
3858 timeout
= schedule_timeout(timeout
);
3859 sleep_on_tail(q
, &wait
, &flags
);
3863 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3865 void __sched
sleep_on(wait_queue_head_t
*q
)
3867 unsigned long flags
;
3870 init_waitqueue_entry(&wait
, current
);
3872 current
->state
= TASK_UNINTERRUPTIBLE
;
3874 sleep_on_head(q
, &wait
, &flags
);
3876 sleep_on_tail(q
, &wait
, &flags
);
3878 EXPORT_SYMBOL(sleep_on
);
3880 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3882 unsigned long flags
;
3885 init_waitqueue_entry(&wait
, current
);
3887 current
->state
= TASK_UNINTERRUPTIBLE
;
3889 sleep_on_head(q
, &wait
, &flags
);
3890 timeout
= schedule_timeout(timeout
);
3891 sleep_on_tail(q
, &wait
, &flags
);
3895 EXPORT_SYMBOL(sleep_on_timeout
);
3897 #ifdef CONFIG_RT_MUTEXES
3900 * rt_mutex_setprio - set the current priority of a task
3902 * @prio: prio value (kernel-internal form)
3904 * This function changes the 'effective' priority of a task. It does
3905 * not touch ->normal_prio like __setscheduler().
3907 * Used by the rt_mutex code to implement priority inheritance logic.
3909 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3911 unsigned long flags
;
3916 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3918 rq
= task_rq_lock(p
, &flags
);
3919 update_rq_clock(rq
);
3923 on_rq
= p
->se
.on_rq
;
3925 dequeue_task(rq
, p
, 0, now
);
3928 p
->sched_class
= &rt_sched_class
;
3930 p
->sched_class
= &fair_sched_class
;
3935 enqueue_task(rq
, p
, 0);
3937 * Reschedule if we are currently running on this runqueue and
3938 * our priority decreased, or if we are not currently running on
3939 * this runqueue and our priority is higher than the current's
3941 if (task_running(rq
, p
)) {
3942 if (p
->prio
> oldprio
)
3943 resched_task(rq
->curr
);
3945 check_preempt_curr(rq
, p
);
3948 task_rq_unlock(rq
, &flags
);
3953 void set_user_nice(struct task_struct
*p
, long nice
)
3955 int old_prio
, delta
, on_rq
;
3956 unsigned long flags
;
3960 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3963 * We have to be careful, if called from sys_setpriority(),
3964 * the task might be in the middle of scheduling on another CPU.
3966 rq
= task_rq_lock(p
, &flags
);
3967 update_rq_clock(rq
);
3970 * The RT priorities are set via sched_setscheduler(), but we still
3971 * allow the 'normal' nice value to be set - but as expected
3972 * it wont have any effect on scheduling until the task is
3973 * SCHED_FIFO/SCHED_RR:
3975 if (task_has_rt_policy(p
)) {
3976 p
->static_prio
= NICE_TO_PRIO(nice
);
3979 on_rq
= p
->se
.on_rq
;
3981 dequeue_task(rq
, p
, 0, now
);
3985 p
->static_prio
= NICE_TO_PRIO(nice
);
3988 p
->prio
= effective_prio(p
);
3989 delta
= p
->prio
- old_prio
;
3992 enqueue_task(rq
, p
, 0);
3995 * If the task increased its priority or is running and
3996 * lowered its priority, then reschedule its CPU:
3998 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3999 resched_task(rq
->curr
);
4002 task_rq_unlock(rq
, &flags
);
4004 EXPORT_SYMBOL(set_user_nice
);
4007 * can_nice - check if a task can reduce its nice value
4011 int can_nice(const struct task_struct
*p
, const int nice
)
4013 /* convert nice value [19,-20] to rlimit style value [1,40] */
4014 int nice_rlim
= 20 - nice
;
4016 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4017 capable(CAP_SYS_NICE
));
4020 #ifdef __ARCH_WANT_SYS_NICE
4023 * sys_nice - change the priority of the current process.
4024 * @increment: priority increment
4026 * sys_setpriority is a more generic, but much slower function that
4027 * does similar things.
4029 asmlinkage
long sys_nice(int increment
)
4034 * Setpriority might change our priority at the same moment.
4035 * We don't have to worry. Conceptually one call occurs first
4036 * and we have a single winner.
4038 if (increment
< -40)
4043 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4049 if (increment
< 0 && !can_nice(current
, nice
))
4052 retval
= security_task_setnice(current
, nice
);
4056 set_user_nice(current
, nice
);
4063 * task_prio - return the priority value of a given task.
4064 * @p: the task in question.
4066 * This is the priority value as seen by users in /proc.
4067 * RT tasks are offset by -200. Normal tasks are centered
4068 * around 0, value goes from -16 to +15.
4070 int task_prio(const struct task_struct
*p
)
4072 return p
->prio
- MAX_RT_PRIO
;
4076 * task_nice - return the nice value of a given task.
4077 * @p: the task in question.
4079 int task_nice(const struct task_struct
*p
)
4081 return TASK_NICE(p
);
4083 EXPORT_SYMBOL_GPL(task_nice
);
4086 * idle_cpu - is a given cpu idle currently?
4087 * @cpu: the processor in question.
4089 int idle_cpu(int cpu
)
4091 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4095 * idle_task - return the idle task for a given cpu.
4096 * @cpu: the processor in question.
4098 struct task_struct
*idle_task(int cpu
)
4100 return cpu_rq(cpu
)->idle
;
4104 * find_process_by_pid - find a process with a matching PID value.
4105 * @pid: the pid in question.
4107 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4109 return pid
? find_task_by_pid(pid
) : current
;
4112 /* Actually do priority change: must hold rq lock. */
4114 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4116 BUG_ON(p
->se
.on_rq
);
4119 switch (p
->policy
) {
4123 p
->sched_class
= &fair_sched_class
;
4127 p
->sched_class
= &rt_sched_class
;
4131 p
->rt_priority
= prio
;
4132 p
->normal_prio
= normal_prio(p
);
4133 /* we are holding p->pi_lock already */
4134 p
->prio
= rt_mutex_getprio(p
);
4139 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4140 * @p: the task in question.
4141 * @policy: new policy.
4142 * @param: structure containing the new RT priority.
4144 * NOTE that the task may be already dead.
4146 int sched_setscheduler(struct task_struct
*p
, int policy
,
4147 struct sched_param
*param
)
4149 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4150 unsigned long flags
;
4153 /* may grab non-irq protected spin_locks */
4154 BUG_ON(in_interrupt());
4156 /* double check policy once rq lock held */
4158 policy
= oldpolicy
= p
->policy
;
4159 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4160 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4161 policy
!= SCHED_IDLE
)
4164 * Valid priorities for SCHED_FIFO and SCHED_RR are
4165 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4166 * SCHED_BATCH and SCHED_IDLE is 0.
4168 if (param
->sched_priority
< 0 ||
4169 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4170 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4172 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4176 * Allow unprivileged RT tasks to decrease priority:
4178 if (!capable(CAP_SYS_NICE
)) {
4179 if (rt_policy(policy
)) {
4180 unsigned long rlim_rtprio
;
4182 if (!lock_task_sighand(p
, &flags
))
4184 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4185 unlock_task_sighand(p
, &flags
);
4187 /* can't set/change the rt policy */
4188 if (policy
!= p
->policy
&& !rlim_rtprio
)
4191 /* can't increase priority */
4192 if (param
->sched_priority
> p
->rt_priority
&&
4193 param
->sched_priority
> rlim_rtprio
)
4197 * Like positive nice levels, dont allow tasks to
4198 * move out of SCHED_IDLE either:
4200 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4203 /* can't change other user's priorities */
4204 if ((current
->euid
!= p
->euid
) &&
4205 (current
->euid
!= p
->uid
))
4209 retval
= security_task_setscheduler(p
, policy
, param
);
4213 * make sure no PI-waiters arrive (or leave) while we are
4214 * changing the priority of the task:
4216 spin_lock_irqsave(&p
->pi_lock
, flags
);
4218 * To be able to change p->policy safely, the apropriate
4219 * runqueue lock must be held.
4221 rq
= __task_rq_lock(p
);
4222 /* recheck policy now with rq lock held */
4223 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4224 policy
= oldpolicy
= -1;
4225 __task_rq_unlock(rq
);
4226 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4229 on_rq
= p
->se
.on_rq
;
4231 update_rq_clock(rq
);
4232 deactivate_task(rq
, p
, 0, rq
->clock
);
4235 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4237 activate_task(rq
, p
, 0);
4239 * Reschedule if we are currently running on this runqueue and
4240 * our priority decreased, or if we are not currently running on
4241 * this runqueue and our priority is higher than the current's
4243 if (task_running(rq
, p
)) {
4244 if (p
->prio
> oldprio
)
4245 resched_task(rq
->curr
);
4247 check_preempt_curr(rq
, p
);
4250 __task_rq_unlock(rq
);
4251 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4253 rt_mutex_adjust_pi(p
);
4257 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4260 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4262 struct sched_param lparam
;
4263 struct task_struct
*p
;
4266 if (!param
|| pid
< 0)
4268 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4273 p
= find_process_by_pid(pid
);
4275 retval
= sched_setscheduler(p
, policy
, &lparam
);
4282 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4283 * @pid: the pid in question.
4284 * @policy: new policy.
4285 * @param: structure containing the new RT priority.
4287 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4288 struct sched_param __user
*param
)
4290 /* negative values for policy are not valid */
4294 return do_sched_setscheduler(pid
, policy
, param
);
4298 * sys_sched_setparam - set/change the RT priority of a thread
4299 * @pid: the pid in question.
4300 * @param: structure containing the new RT priority.
4302 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4304 return do_sched_setscheduler(pid
, -1, param
);
4308 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4309 * @pid: the pid in question.
4311 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4313 struct task_struct
*p
;
4314 int retval
= -EINVAL
;
4320 read_lock(&tasklist_lock
);
4321 p
= find_process_by_pid(pid
);
4323 retval
= security_task_getscheduler(p
);
4327 read_unlock(&tasklist_lock
);
4334 * sys_sched_getscheduler - get the RT priority of a thread
4335 * @pid: the pid in question.
4336 * @param: structure containing the RT priority.
4338 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4340 struct sched_param lp
;
4341 struct task_struct
*p
;
4342 int retval
= -EINVAL
;
4344 if (!param
|| pid
< 0)
4347 read_lock(&tasklist_lock
);
4348 p
= find_process_by_pid(pid
);
4353 retval
= security_task_getscheduler(p
);
4357 lp
.sched_priority
= p
->rt_priority
;
4358 read_unlock(&tasklist_lock
);
4361 * This one might sleep, we cannot do it with a spinlock held ...
4363 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4369 read_unlock(&tasklist_lock
);
4373 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4375 cpumask_t cpus_allowed
;
4376 struct task_struct
*p
;
4379 mutex_lock(&sched_hotcpu_mutex
);
4380 read_lock(&tasklist_lock
);
4382 p
= find_process_by_pid(pid
);
4384 read_unlock(&tasklist_lock
);
4385 mutex_unlock(&sched_hotcpu_mutex
);
4390 * It is not safe to call set_cpus_allowed with the
4391 * tasklist_lock held. We will bump the task_struct's
4392 * usage count and then drop tasklist_lock.
4395 read_unlock(&tasklist_lock
);
4398 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4399 !capable(CAP_SYS_NICE
))
4402 retval
= security_task_setscheduler(p
, 0, NULL
);
4406 cpus_allowed
= cpuset_cpus_allowed(p
);
4407 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4408 retval
= set_cpus_allowed(p
, new_mask
);
4412 mutex_unlock(&sched_hotcpu_mutex
);
4416 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4417 cpumask_t
*new_mask
)
4419 if (len
< sizeof(cpumask_t
)) {
4420 memset(new_mask
, 0, sizeof(cpumask_t
));
4421 } else if (len
> sizeof(cpumask_t
)) {
4422 len
= sizeof(cpumask_t
);
4424 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4428 * sys_sched_setaffinity - set the cpu affinity of a process
4429 * @pid: pid of the process
4430 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4431 * @user_mask_ptr: user-space pointer to the new cpu mask
4433 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4434 unsigned long __user
*user_mask_ptr
)
4439 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4443 return sched_setaffinity(pid
, new_mask
);
4447 * Represents all cpu's present in the system
4448 * In systems capable of hotplug, this map could dynamically grow
4449 * as new cpu's are detected in the system via any platform specific
4450 * method, such as ACPI for e.g.
4453 cpumask_t cpu_present_map __read_mostly
;
4454 EXPORT_SYMBOL(cpu_present_map
);
4457 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4458 EXPORT_SYMBOL(cpu_online_map
);
4460 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4461 EXPORT_SYMBOL(cpu_possible_map
);
4464 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4466 struct task_struct
*p
;
4469 mutex_lock(&sched_hotcpu_mutex
);
4470 read_lock(&tasklist_lock
);
4473 p
= find_process_by_pid(pid
);
4477 retval
= security_task_getscheduler(p
);
4481 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4484 read_unlock(&tasklist_lock
);
4485 mutex_unlock(&sched_hotcpu_mutex
);
4491 * sys_sched_getaffinity - get the cpu affinity of a process
4492 * @pid: pid of the process
4493 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4494 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4496 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4497 unsigned long __user
*user_mask_ptr
)
4502 if (len
< sizeof(cpumask_t
))
4505 ret
= sched_getaffinity(pid
, &mask
);
4509 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4512 return sizeof(cpumask_t
);
4516 * sys_sched_yield - yield the current processor to other threads.
4518 * This function yields the current CPU to other tasks. If there are no
4519 * other threads running on this CPU then this function will return.
4521 asmlinkage
long sys_sched_yield(void)
4523 struct rq
*rq
= this_rq_lock();
4525 schedstat_inc(rq
, yld_cnt
);
4526 if (unlikely(rq
->nr_running
== 1))
4527 schedstat_inc(rq
, yld_act_empty
);
4529 current
->sched_class
->yield_task(rq
, current
);
4532 * Since we are going to call schedule() anyway, there's
4533 * no need to preempt or enable interrupts:
4535 __release(rq
->lock
);
4536 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4537 _raw_spin_unlock(&rq
->lock
);
4538 preempt_enable_no_resched();
4545 static void __cond_resched(void)
4547 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4548 __might_sleep(__FILE__
, __LINE__
);
4551 * The BKS might be reacquired before we have dropped
4552 * PREEMPT_ACTIVE, which could trigger a second
4553 * cond_resched() call.
4556 add_preempt_count(PREEMPT_ACTIVE
);
4558 sub_preempt_count(PREEMPT_ACTIVE
);
4559 } while (need_resched());
4562 int __sched
cond_resched(void)
4564 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4565 system_state
== SYSTEM_RUNNING
) {
4571 EXPORT_SYMBOL(cond_resched
);
4574 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4575 * call schedule, and on return reacquire the lock.
4577 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4578 * operations here to prevent schedule() from being called twice (once via
4579 * spin_unlock(), once by hand).
4581 int cond_resched_lock(spinlock_t
*lock
)
4585 if (need_lockbreak(lock
)) {
4591 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4592 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4593 _raw_spin_unlock(lock
);
4594 preempt_enable_no_resched();
4601 EXPORT_SYMBOL(cond_resched_lock
);
4603 int __sched
cond_resched_softirq(void)
4605 BUG_ON(!in_softirq());
4607 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4615 EXPORT_SYMBOL(cond_resched_softirq
);
4618 * yield - yield the current processor to other threads.
4620 * This is a shortcut for kernel-space yielding - it marks the
4621 * thread runnable and calls sys_sched_yield().
4623 void __sched
yield(void)
4625 set_current_state(TASK_RUNNING
);
4628 EXPORT_SYMBOL(yield
);
4631 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4632 * that process accounting knows that this is a task in IO wait state.
4634 * But don't do that if it is a deliberate, throttling IO wait (this task
4635 * has set its backing_dev_info: the queue against which it should throttle)
4637 void __sched
io_schedule(void)
4639 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4641 delayacct_blkio_start();
4642 atomic_inc(&rq
->nr_iowait
);
4644 atomic_dec(&rq
->nr_iowait
);
4645 delayacct_blkio_end();
4647 EXPORT_SYMBOL(io_schedule
);
4649 long __sched
io_schedule_timeout(long timeout
)
4651 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4654 delayacct_blkio_start();
4655 atomic_inc(&rq
->nr_iowait
);
4656 ret
= schedule_timeout(timeout
);
4657 atomic_dec(&rq
->nr_iowait
);
4658 delayacct_blkio_end();
4663 * sys_sched_get_priority_max - return maximum RT priority.
4664 * @policy: scheduling class.
4666 * this syscall returns the maximum rt_priority that can be used
4667 * by a given scheduling class.
4669 asmlinkage
long sys_sched_get_priority_max(int policy
)
4676 ret
= MAX_USER_RT_PRIO
-1;
4688 * sys_sched_get_priority_min - return minimum RT priority.
4689 * @policy: scheduling class.
4691 * this syscall returns the minimum rt_priority that can be used
4692 * by a given scheduling class.
4694 asmlinkage
long sys_sched_get_priority_min(int policy
)
4712 * sys_sched_rr_get_interval - return the default timeslice of a process.
4713 * @pid: pid of the process.
4714 * @interval: userspace pointer to the timeslice value.
4716 * this syscall writes the default timeslice value of a given process
4717 * into the user-space timespec buffer. A value of '0' means infinity.
4720 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4722 struct task_struct
*p
;
4723 int retval
= -EINVAL
;
4730 read_lock(&tasklist_lock
);
4731 p
= find_process_by_pid(pid
);
4735 retval
= security_task_getscheduler(p
);
4739 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4740 0 : static_prio_timeslice(p
->static_prio
), &t
);
4741 read_unlock(&tasklist_lock
);
4742 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4746 read_unlock(&tasklist_lock
);
4750 static const char stat_nam
[] = "RSDTtZX";
4752 static void show_task(struct task_struct
*p
)
4754 unsigned long free
= 0;
4757 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4758 printk("%-13.13s %c", p
->comm
,
4759 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4760 #if BITS_PER_LONG == 32
4761 if (state
== TASK_RUNNING
)
4762 printk(" running ");
4764 printk(" %08lx ", thread_saved_pc(p
));
4766 if (state
== TASK_RUNNING
)
4767 printk(" running task ");
4769 printk(" %016lx ", thread_saved_pc(p
));
4771 #ifdef CONFIG_DEBUG_STACK_USAGE
4773 unsigned long *n
= end_of_stack(p
);
4776 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4779 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4781 if (state
!= TASK_RUNNING
)
4782 show_stack(p
, NULL
);
4785 void show_state_filter(unsigned long state_filter
)
4787 struct task_struct
*g
, *p
;
4789 #if BITS_PER_LONG == 32
4791 " task PC stack pid father\n");
4794 " task PC stack pid father\n");
4796 read_lock(&tasklist_lock
);
4797 do_each_thread(g
, p
) {
4799 * reset the NMI-timeout, listing all files on a slow
4800 * console might take alot of time:
4802 touch_nmi_watchdog();
4803 if (!state_filter
|| (p
->state
& state_filter
))
4805 } while_each_thread(g
, p
);
4807 touch_all_softlockup_watchdogs();
4809 #ifdef CONFIG_SCHED_DEBUG
4810 sysrq_sched_debug_show();
4812 read_unlock(&tasklist_lock
);
4814 * Only show locks if all tasks are dumped:
4816 if (state_filter
== -1)
4817 debug_show_all_locks();
4820 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4822 idle
->sched_class
= &idle_sched_class
;
4826 * init_idle - set up an idle thread for a given CPU
4827 * @idle: task in question
4828 * @cpu: cpu the idle task belongs to
4830 * NOTE: this function does not set the idle thread's NEED_RESCHED
4831 * flag, to make booting more robust.
4833 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4835 struct rq
*rq
= cpu_rq(cpu
);
4836 unsigned long flags
;
4839 idle
->se
.exec_start
= sched_clock();
4841 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4842 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4843 __set_task_cpu(idle
, cpu
);
4845 spin_lock_irqsave(&rq
->lock
, flags
);
4846 rq
->curr
= rq
->idle
= idle
;
4847 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4850 spin_unlock_irqrestore(&rq
->lock
, flags
);
4852 /* Set the preempt count _outside_ the spinlocks! */
4853 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4854 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4856 task_thread_info(idle
)->preempt_count
= 0;
4859 * The idle tasks have their own, simple scheduling class:
4861 idle
->sched_class
= &idle_sched_class
;
4865 * In a system that switches off the HZ timer nohz_cpu_mask
4866 * indicates which cpus entered this state. This is used
4867 * in the rcu update to wait only for active cpus. For system
4868 * which do not switch off the HZ timer nohz_cpu_mask should
4869 * always be CPU_MASK_NONE.
4871 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4874 * Increase the granularity value when there are more CPUs,
4875 * because with more CPUs the 'effective latency' as visible
4876 * to users decreases. But the relationship is not linear,
4877 * so pick a second-best guess by going with the log2 of the
4880 * This idea comes from the SD scheduler of Con Kolivas:
4882 static inline void sched_init_granularity(void)
4884 unsigned int factor
= 1 + ilog2(num_online_cpus());
4885 const unsigned long gran_limit
= 100000000;
4887 sysctl_sched_granularity
*= factor
;
4888 if (sysctl_sched_granularity
> gran_limit
)
4889 sysctl_sched_granularity
= gran_limit
;
4891 sysctl_sched_runtime_limit
= sysctl_sched_granularity
* 4;
4892 sysctl_sched_wakeup_granularity
= sysctl_sched_granularity
/ 2;
4897 * This is how migration works:
4899 * 1) we queue a struct migration_req structure in the source CPU's
4900 * runqueue and wake up that CPU's migration thread.
4901 * 2) we down() the locked semaphore => thread blocks.
4902 * 3) migration thread wakes up (implicitly it forces the migrated
4903 * thread off the CPU)
4904 * 4) it gets the migration request and checks whether the migrated
4905 * task is still in the wrong runqueue.
4906 * 5) if it's in the wrong runqueue then the migration thread removes
4907 * it and puts it into the right queue.
4908 * 6) migration thread up()s the semaphore.
4909 * 7) we wake up and the migration is done.
4913 * Change a given task's CPU affinity. Migrate the thread to a
4914 * proper CPU and schedule it away if the CPU it's executing on
4915 * is removed from the allowed bitmask.
4917 * NOTE: the caller must have a valid reference to the task, the
4918 * task must not exit() & deallocate itself prematurely. The
4919 * call is not atomic; no spinlocks may be held.
4921 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4923 struct migration_req req
;
4924 unsigned long flags
;
4928 rq
= task_rq_lock(p
, &flags
);
4929 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4934 p
->cpus_allowed
= new_mask
;
4935 /* Can the task run on the task's current CPU? If so, we're done */
4936 if (cpu_isset(task_cpu(p
), new_mask
))
4939 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4940 /* Need help from migration thread: drop lock and wait. */
4941 task_rq_unlock(rq
, &flags
);
4942 wake_up_process(rq
->migration_thread
);
4943 wait_for_completion(&req
.done
);
4944 tlb_migrate_finish(p
->mm
);
4948 task_rq_unlock(rq
, &flags
);
4952 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4955 * Move (not current) task off this cpu, onto dest cpu. We're doing
4956 * this because either it can't run here any more (set_cpus_allowed()
4957 * away from this CPU, or CPU going down), or because we're
4958 * attempting to rebalance this task on exec (sched_exec).
4960 * So we race with normal scheduler movements, but that's OK, as long
4961 * as the task is no longer on this CPU.
4963 * Returns non-zero if task was successfully migrated.
4965 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4967 struct rq
*rq_dest
, *rq_src
;
4970 if (unlikely(cpu_is_offline(dest_cpu
)))
4973 rq_src
= cpu_rq(src_cpu
);
4974 rq_dest
= cpu_rq(dest_cpu
);
4976 double_rq_lock(rq_src
, rq_dest
);
4977 /* Already moved. */
4978 if (task_cpu(p
) != src_cpu
)
4980 /* Affinity changed (again). */
4981 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4984 on_rq
= p
->se
.on_rq
;
4986 update_rq_clock(rq_src
);
4987 deactivate_task(rq_src
, p
, 0, rq_src
->clock
);
4989 set_task_cpu(p
, dest_cpu
);
4991 activate_task(rq_dest
, p
, 0);
4992 check_preempt_curr(rq_dest
, p
);
4996 double_rq_unlock(rq_src
, rq_dest
);
5001 * migration_thread - this is a highprio system thread that performs
5002 * thread migration by bumping thread off CPU then 'pushing' onto
5005 static int migration_thread(void *data
)
5007 int cpu
= (long)data
;
5011 BUG_ON(rq
->migration_thread
!= current
);
5013 set_current_state(TASK_INTERRUPTIBLE
);
5014 while (!kthread_should_stop()) {
5015 struct migration_req
*req
;
5016 struct list_head
*head
;
5018 spin_lock_irq(&rq
->lock
);
5020 if (cpu_is_offline(cpu
)) {
5021 spin_unlock_irq(&rq
->lock
);
5025 if (rq
->active_balance
) {
5026 active_load_balance(rq
, cpu
);
5027 rq
->active_balance
= 0;
5030 head
= &rq
->migration_queue
;
5032 if (list_empty(head
)) {
5033 spin_unlock_irq(&rq
->lock
);
5035 set_current_state(TASK_INTERRUPTIBLE
);
5038 req
= list_entry(head
->next
, struct migration_req
, list
);
5039 list_del_init(head
->next
);
5041 spin_unlock(&rq
->lock
);
5042 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5045 complete(&req
->done
);
5047 __set_current_state(TASK_RUNNING
);
5051 /* Wait for kthread_stop */
5052 set_current_state(TASK_INTERRUPTIBLE
);
5053 while (!kthread_should_stop()) {
5055 set_current_state(TASK_INTERRUPTIBLE
);
5057 __set_current_state(TASK_RUNNING
);
5061 #ifdef CONFIG_HOTPLUG_CPU
5063 * Figure out where task on dead CPU should go, use force if neccessary.
5064 * NOTE: interrupts should be disabled by the caller
5066 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5068 unsigned long flags
;
5075 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5076 cpus_and(mask
, mask
, p
->cpus_allowed
);
5077 dest_cpu
= any_online_cpu(mask
);
5079 /* On any allowed CPU? */
5080 if (dest_cpu
== NR_CPUS
)
5081 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5083 /* No more Mr. Nice Guy. */
5084 if (dest_cpu
== NR_CPUS
) {
5085 rq
= task_rq_lock(p
, &flags
);
5086 cpus_setall(p
->cpus_allowed
);
5087 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5088 task_rq_unlock(rq
, &flags
);
5091 * Don't tell them about moving exiting tasks or
5092 * kernel threads (both mm NULL), since they never
5095 if (p
->mm
&& printk_ratelimit())
5096 printk(KERN_INFO
"process %d (%s) no "
5097 "longer affine to cpu%d\n",
5098 p
->pid
, p
->comm
, dead_cpu
);
5100 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5105 * While a dead CPU has no uninterruptible tasks queued at this point,
5106 * it might still have a nonzero ->nr_uninterruptible counter, because
5107 * for performance reasons the counter is not stricly tracking tasks to
5108 * their home CPUs. So we just add the counter to another CPU's counter,
5109 * to keep the global sum constant after CPU-down:
5111 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5113 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5114 unsigned long flags
;
5116 local_irq_save(flags
);
5117 double_rq_lock(rq_src
, rq_dest
);
5118 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5119 rq_src
->nr_uninterruptible
= 0;
5120 double_rq_unlock(rq_src
, rq_dest
);
5121 local_irq_restore(flags
);
5124 /* Run through task list and migrate tasks from the dead cpu. */
5125 static void migrate_live_tasks(int src_cpu
)
5127 struct task_struct
*p
, *t
;
5129 write_lock_irq(&tasklist_lock
);
5131 do_each_thread(t
, p
) {
5135 if (task_cpu(p
) == src_cpu
)
5136 move_task_off_dead_cpu(src_cpu
, p
);
5137 } while_each_thread(t
, p
);
5139 write_unlock_irq(&tasklist_lock
);
5143 * Schedules idle task to be the next runnable task on current CPU.
5144 * It does so by boosting its priority to highest possible and adding it to
5145 * the _front_ of the runqueue. Used by CPU offline code.
5147 void sched_idle_next(void)
5149 int this_cpu
= smp_processor_id();
5150 struct rq
*rq
= cpu_rq(this_cpu
);
5151 struct task_struct
*p
= rq
->idle
;
5152 unsigned long flags
;
5154 /* cpu has to be offline */
5155 BUG_ON(cpu_online(this_cpu
));
5158 * Strictly not necessary since rest of the CPUs are stopped by now
5159 * and interrupts disabled on the current cpu.
5161 spin_lock_irqsave(&rq
->lock
, flags
);
5163 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5165 /* Add idle task to the _front_ of its priority queue: */
5166 activate_idle_task(p
, rq
);
5168 spin_unlock_irqrestore(&rq
->lock
, flags
);
5172 * Ensures that the idle task is using init_mm right before its cpu goes
5175 void idle_task_exit(void)
5177 struct mm_struct
*mm
= current
->active_mm
;
5179 BUG_ON(cpu_online(smp_processor_id()));
5182 switch_mm(mm
, &init_mm
, current
);
5186 /* called under rq->lock with disabled interrupts */
5187 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5189 struct rq
*rq
= cpu_rq(dead_cpu
);
5191 /* Must be exiting, otherwise would be on tasklist. */
5192 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5194 /* Cannot have done final schedule yet: would have vanished. */
5195 BUG_ON(p
->state
== TASK_DEAD
);
5200 * Drop lock around migration; if someone else moves it,
5201 * that's OK. No task can be added to this CPU, so iteration is
5203 * NOTE: interrupts should be left disabled --dev@
5205 spin_unlock(&rq
->lock
);
5206 move_task_off_dead_cpu(dead_cpu
, p
);
5207 spin_lock(&rq
->lock
);
5212 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5213 static void migrate_dead_tasks(unsigned int dead_cpu
)
5215 struct rq
*rq
= cpu_rq(dead_cpu
);
5216 struct task_struct
*next
;
5219 if (!rq
->nr_running
)
5221 update_rq_clock(rq
);
5222 next
= pick_next_task(rq
, rq
->curr
);
5225 migrate_dead(dead_cpu
, next
);
5229 #endif /* CONFIG_HOTPLUG_CPU */
5231 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5233 static struct ctl_table sd_ctl_dir
[] = {
5235 .procname
= "sched_domain",
5241 static struct ctl_table sd_ctl_root
[] = {
5243 .procname
= "kernel",
5245 .child
= sd_ctl_dir
,
5250 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5252 struct ctl_table
*entry
=
5253 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5256 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5262 set_table_entry(struct ctl_table
*entry
,
5263 const char *procname
, void *data
, int maxlen
,
5264 mode_t mode
, proc_handler
*proc_handler
)
5266 entry
->procname
= procname
;
5268 entry
->maxlen
= maxlen
;
5270 entry
->proc_handler
= proc_handler
;
5273 static struct ctl_table
*
5274 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5276 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5278 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5279 sizeof(long), 0644, proc_doulongvec_minmax
);
5280 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5281 sizeof(long), 0644, proc_doulongvec_minmax
);
5282 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5283 sizeof(int), 0644, proc_dointvec_minmax
);
5284 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5285 sizeof(int), 0644, proc_dointvec_minmax
);
5286 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5287 sizeof(int), 0644, proc_dointvec_minmax
);
5288 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5289 sizeof(int), 0644, proc_dointvec_minmax
);
5290 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5291 sizeof(int), 0644, proc_dointvec_minmax
);
5292 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5293 sizeof(int), 0644, proc_dointvec_minmax
);
5294 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5295 sizeof(int), 0644, proc_dointvec_minmax
);
5296 set_table_entry(&table
[10], "cache_nice_tries",
5297 &sd
->cache_nice_tries
,
5298 sizeof(int), 0644, proc_dointvec_minmax
);
5299 set_table_entry(&table
[12], "flags", &sd
->flags
,
5300 sizeof(int), 0644, proc_dointvec_minmax
);
5305 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5307 struct ctl_table
*entry
, *table
;
5308 struct sched_domain
*sd
;
5309 int domain_num
= 0, i
;
5312 for_each_domain(cpu
, sd
)
5314 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5317 for_each_domain(cpu
, sd
) {
5318 snprintf(buf
, 32, "domain%d", i
);
5319 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5321 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5328 static struct ctl_table_header
*sd_sysctl_header
;
5329 static void init_sched_domain_sysctl(void)
5331 int i
, cpu_num
= num_online_cpus();
5332 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5335 sd_ctl_dir
[0].child
= entry
;
5337 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5338 snprintf(buf
, 32, "cpu%d", i
);
5339 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5341 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5343 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5346 static void init_sched_domain_sysctl(void)
5352 * migration_call - callback that gets triggered when a CPU is added.
5353 * Here we can start up the necessary migration thread for the new CPU.
5355 static int __cpuinit
5356 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5358 struct task_struct
*p
;
5359 int cpu
= (long)hcpu
;
5360 unsigned long flags
;
5364 case CPU_LOCK_ACQUIRE
:
5365 mutex_lock(&sched_hotcpu_mutex
);
5368 case CPU_UP_PREPARE
:
5369 case CPU_UP_PREPARE_FROZEN
:
5370 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5373 kthread_bind(p
, cpu
);
5374 /* Must be high prio: stop_machine expects to yield to it. */
5375 rq
= task_rq_lock(p
, &flags
);
5376 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5377 task_rq_unlock(rq
, &flags
);
5378 cpu_rq(cpu
)->migration_thread
= p
;
5382 case CPU_ONLINE_FROZEN
:
5383 /* Strictly unneccessary, as first user will wake it. */
5384 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5387 #ifdef CONFIG_HOTPLUG_CPU
5388 case CPU_UP_CANCELED
:
5389 case CPU_UP_CANCELED_FROZEN
:
5390 if (!cpu_rq(cpu
)->migration_thread
)
5392 /* Unbind it from offline cpu so it can run. Fall thru. */
5393 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5394 any_online_cpu(cpu_online_map
));
5395 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5396 cpu_rq(cpu
)->migration_thread
= NULL
;
5400 case CPU_DEAD_FROZEN
:
5401 migrate_live_tasks(cpu
);
5403 kthread_stop(rq
->migration_thread
);
5404 rq
->migration_thread
= NULL
;
5405 /* Idle task back to normal (off runqueue, low prio) */
5406 rq
= task_rq_lock(rq
->idle
, &flags
);
5407 update_rq_clock(rq
);
5408 deactivate_task(rq
, rq
->idle
, 0, rq
->clock
);
5409 rq
->idle
->static_prio
= MAX_PRIO
;
5410 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5411 rq
->idle
->sched_class
= &idle_sched_class
;
5412 migrate_dead_tasks(cpu
);
5413 task_rq_unlock(rq
, &flags
);
5414 migrate_nr_uninterruptible(rq
);
5415 BUG_ON(rq
->nr_running
!= 0);
5417 /* No need to migrate the tasks: it was best-effort if
5418 * they didn't take sched_hotcpu_mutex. Just wake up
5419 * the requestors. */
5420 spin_lock_irq(&rq
->lock
);
5421 while (!list_empty(&rq
->migration_queue
)) {
5422 struct migration_req
*req
;
5424 req
= list_entry(rq
->migration_queue
.next
,
5425 struct migration_req
, list
);
5426 list_del_init(&req
->list
);
5427 complete(&req
->done
);
5429 spin_unlock_irq(&rq
->lock
);
5432 case CPU_LOCK_RELEASE
:
5433 mutex_unlock(&sched_hotcpu_mutex
);
5439 /* Register at highest priority so that task migration (migrate_all_tasks)
5440 * happens before everything else.
5442 static struct notifier_block __cpuinitdata migration_notifier
= {
5443 .notifier_call
= migration_call
,
5447 int __init
migration_init(void)
5449 void *cpu
= (void *)(long)smp_processor_id();
5452 /* Start one for the boot CPU: */
5453 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5454 BUG_ON(err
== NOTIFY_BAD
);
5455 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5456 register_cpu_notifier(&migration_notifier
);
5464 /* Number of possible processor ids */
5465 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5466 EXPORT_SYMBOL(nr_cpu_ids
);
5468 #undef SCHED_DOMAIN_DEBUG
5469 #ifdef SCHED_DOMAIN_DEBUG
5470 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5475 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5479 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5484 struct sched_group
*group
= sd
->groups
;
5485 cpumask_t groupmask
;
5487 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5488 cpus_clear(groupmask
);
5491 for (i
= 0; i
< level
+ 1; i
++)
5493 printk("domain %d: ", level
);
5495 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5496 printk("does not load-balance\n");
5498 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5503 printk("span %s\n", str
);
5505 if (!cpu_isset(cpu
, sd
->span
))
5506 printk(KERN_ERR
"ERROR: domain->span does not contain "
5508 if (!cpu_isset(cpu
, group
->cpumask
))
5509 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5513 for (i
= 0; i
< level
+ 2; i
++)
5519 printk(KERN_ERR
"ERROR: group is NULL\n");
5523 if (!group
->__cpu_power
) {
5525 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5529 if (!cpus_weight(group
->cpumask
)) {
5531 printk(KERN_ERR
"ERROR: empty group\n");
5534 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5536 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5539 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5541 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5544 group
= group
->next
;
5545 } while (group
!= sd
->groups
);
5548 if (!cpus_equal(sd
->span
, groupmask
))
5549 printk(KERN_ERR
"ERROR: groups don't span "
5557 if (!cpus_subset(groupmask
, sd
->span
))
5558 printk(KERN_ERR
"ERROR: parent span is not a superset "
5559 "of domain->span\n");
5564 # define sched_domain_debug(sd, cpu) do { } while (0)
5567 static int sd_degenerate(struct sched_domain
*sd
)
5569 if (cpus_weight(sd
->span
) == 1)
5572 /* Following flags need at least 2 groups */
5573 if (sd
->flags
& (SD_LOAD_BALANCE
|
5574 SD_BALANCE_NEWIDLE
|
5578 SD_SHARE_PKG_RESOURCES
)) {
5579 if (sd
->groups
!= sd
->groups
->next
)
5583 /* Following flags don't use groups */
5584 if (sd
->flags
& (SD_WAKE_IDLE
|
5593 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5595 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5597 if (sd_degenerate(parent
))
5600 if (!cpus_equal(sd
->span
, parent
->span
))
5603 /* Does parent contain flags not in child? */
5604 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5605 if (cflags
& SD_WAKE_AFFINE
)
5606 pflags
&= ~SD_WAKE_BALANCE
;
5607 /* Flags needing groups don't count if only 1 group in parent */
5608 if (parent
->groups
== parent
->groups
->next
) {
5609 pflags
&= ~(SD_LOAD_BALANCE
|
5610 SD_BALANCE_NEWIDLE
|
5614 SD_SHARE_PKG_RESOURCES
);
5616 if (~cflags
& pflags
)
5623 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5624 * hold the hotplug lock.
5626 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5628 struct rq
*rq
= cpu_rq(cpu
);
5629 struct sched_domain
*tmp
;
5631 /* Remove the sched domains which do not contribute to scheduling. */
5632 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5633 struct sched_domain
*parent
= tmp
->parent
;
5636 if (sd_parent_degenerate(tmp
, parent
)) {
5637 tmp
->parent
= parent
->parent
;
5639 parent
->parent
->child
= tmp
;
5643 if (sd
&& sd_degenerate(sd
)) {
5649 sched_domain_debug(sd
, cpu
);
5651 rcu_assign_pointer(rq
->sd
, sd
);
5654 /* cpus with isolated domains */
5655 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5657 /* Setup the mask of cpus configured for isolated domains */
5658 static int __init
isolated_cpu_setup(char *str
)
5660 int ints
[NR_CPUS
], i
;
5662 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5663 cpus_clear(cpu_isolated_map
);
5664 for (i
= 1; i
<= ints
[0]; i
++)
5665 if (ints
[i
] < NR_CPUS
)
5666 cpu_set(ints
[i
], cpu_isolated_map
);
5670 __setup ("isolcpus=", isolated_cpu_setup
);
5673 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5674 * to a function which identifies what group(along with sched group) a CPU
5675 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5676 * (due to the fact that we keep track of groups covered with a cpumask_t).
5678 * init_sched_build_groups will build a circular linked list of the groups
5679 * covered by the given span, and will set each group's ->cpumask correctly,
5680 * and ->cpu_power to 0.
5683 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5684 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5685 struct sched_group
**sg
))
5687 struct sched_group
*first
= NULL
, *last
= NULL
;
5688 cpumask_t covered
= CPU_MASK_NONE
;
5691 for_each_cpu_mask(i
, span
) {
5692 struct sched_group
*sg
;
5693 int group
= group_fn(i
, cpu_map
, &sg
);
5696 if (cpu_isset(i
, covered
))
5699 sg
->cpumask
= CPU_MASK_NONE
;
5700 sg
->__cpu_power
= 0;
5702 for_each_cpu_mask(j
, span
) {
5703 if (group_fn(j
, cpu_map
, NULL
) != group
)
5706 cpu_set(j
, covered
);
5707 cpu_set(j
, sg
->cpumask
);
5718 #define SD_NODES_PER_DOMAIN 16
5723 * find_next_best_node - find the next node to include in a sched_domain
5724 * @node: node whose sched_domain we're building
5725 * @used_nodes: nodes already in the sched_domain
5727 * Find the next node to include in a given scheduling domain. Simply
5728 * finds the closest node not already in the @used_nodes map.
5730 * Should use nodemask_t.
5732 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5734 int i
, n
, val
, min_val
, best_node
= 0;
5738 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5739 /* Start at @node */
5740 n
= (node
+ i
) % MAX_NUMNODES
;
5742 if (!nr_cpus_node(n
))
5745 /* Skip already used nodes */
5746 if (test_bit(n
, used_nodes
))
5749 /* Simple min distance search */
5750 val
= node_distance(node
, n
);
5752 if (val
< min_val
) {
5758 set_bit(best_node
, used_nodes
);
5763 * sched_domain_node_span - get a cpumask for a node's sched_domain
5764 * @node: node whose cpumask we're constructing
5765 * @size: number of nodes to include in this span
5767 * Given a node, construct a good cpumask for its sched_domain to span. It
5768 * should be one that prevents unnecessary balancing, but also spreads tasks
5771 static cpumask_t
sched_domain_node_span(int node
)
5773 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5774 cpumask_t span
, nodemask
;
5778 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5780 nodemask
= node_to_cpumask(node
);
5781 cpus_or(span
, span
, nodemask
);
5782 set_bit(node
, used_nodes
);
5784 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5785 int next_node
= find_next_best_node(node
, used_nodes
);
5787 nodemask
= node_to_cpumask(next_node
);
5788 cpus_or(span
, span
, nodemask
);
5795 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5798 * SMT sched-domains:
5800 #ifdef CONFIG_SCHED_SMT
5801 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5802 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5804 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5805 struct sched_group
**sg
)
5808 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5814 * multi-core sched-domains:
5816 #ifdef CONFIG_SCHED_MC
5817 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5818 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5821 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5822 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5823 struct sched_group
**sg
)
5826 cpumask_t mask
= cpu_sibling_map
[cpu
];
5827 cpus_and(mask
, mask
, *cpu_map
);
5828 group
= first_cpu(mask
);
5830 *sg
= &per_cpu(sched_group_core
, group
);
5833 #elif defined(CONFIG_SCHED_MC)
5834 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5835 struct sched_group
**sg
)
5838 *sg
= &per_cpu(sched_group_core
, cpu
);
5843 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5844 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5846 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5847 struct sched_group
**sg
)
5850 #ifdef CONFIG_SCHED_MC
5851 cpumask_t mask
= cpu_coregroup_map(cpu
);
5852 cpus_and(mask
, mask
, *cpu_map
);
5853 group
= first_cpu(mask
);
5854 #elif defined(CONFIG_SCHED_SMT)
5855 cpumask_t mask
= cpu_sibling_map
[cpu
];
5856 cpus_and(mask
, mask
, *cpu_map
);
5857 group
= first_cpu(mask
);
5862 *sg
= &per_cpu(sched_group_phys
, group
);
5868 * The init_sched_build_groups can't handle what we want to do with node
5869 * groups, so roll our own. Now each node has its own list of groups which
5870 * gets dynamically allocated.
5872 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5873 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5875 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5876 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5878 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5879 struct sched_group
**sg
)
5881 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5884 cpus_and(nodemask
, nodemask
, *cpu_map
);
5885 group
= first_cpu(nodemask
);
5888 *sg
= &per_cpu(sched_group_allnodes
, group
);
5892 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5894 struct sched_group
*sg
= group_head
;
5900 for_each_cpu_mask(j
, sg
->cpumask
) {
5901 struct sched_domain
*sd
;
5903 sd
= &per_cpu(phys_domains
, j
);
5904 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5906 * Only add "power" once for each
5912 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5915 if (sg
!= group_head
)
5921 /* Free memory allocated for various sched_group structures */
5922 static void free_sched_groups(const cpumask_t
*cpu_map
)
5926 for_each_cpu_mask(cpu
, *cpu_map
) {
5927 struct sched_group
**sched_group_nodes
5928 = sched_group_nodes_bycpu
[cpu
];
5930 if (!sched_group_nodes
)
5933 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5934 cpumask_t nodemask
= node_to_cpumask(i
);
5935 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5937 cpus_and(nodemask
, nodemask
, *cpu_map
);
5938 if (cpus_empty(nodemask
))
5948 if (oldsg
!= sched_group_nodes
[i
])
5951 kfree(sched_group_nodes
);
5952 sched_group_nodes_bycpu
[cpu
] = NULL
;
5956 static void free_sched_groups(const cpumask_t
*cpu_map
)
5962 * Initialize sched groups cpu_power.
5964 * cpu_power indicates the capacity of sched group, which is used while
5965 * distributing the load between different sched groups in a sched domain.
5966 * Typically cpu_power for all the groups in a sched domain will be same unless
5967 * there are asymmetries in the topology. If there are asymmetries, group
5968 * having more cpu_power will pickup more load compared to the group having
5971 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5972 * the maximum number of tasks a group can handle in the presence of other idle
5973 * or lightly loaded groups in the same sched domain.
5975 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5977 struct sched_domain
*child
;
5978 struct sched_group
*group
;
5980 WARN_ON(!sd
|| !sd
->groups
);
5982 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5987 sd
->groups
->__cpu_power
= 0;
5990 * For perf policy, if the groups in child domain share resources
5991 * (for example cores sharing some portions of the cache hierarchy
5992 * or SMT), then set this domain groups cpu_power such that each group
5993 * can handle only one task, when there are other idle groups in the
5994 * same sched domain.
5996 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5998 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5999 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6004 * add cpu_power of each child group to this groups cpu_power
6006 group
= child
->groups
;
6008 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6009 group
= group
->next
;
6010 } while (group
!= child
->groups
);
6014 * Build sched domains for a given set of cpus and attach the sched domains
6015 * to the individual cpus
6017 static int build_sched_domains(const cpumask_t
*cpu_map
)
6021 struct sched_group
**sched_group_nodes
= NULL
;
6022 int sd_allnodes
= 0;
6025 * Allocate the per-node list of sched groups
6027 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6029 if (!sched_group_nodes
) {
6030 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6033 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6037 * Set up domains for cpus specified by the cpu_map.
6039 for_each_cpu_mask(i
, *cpu_map
) {
6040 struct sched_domain
*sd
= NULL
, *p
;
6041 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6043 cpus_and(nodemask
, nodemask
, *cpu_map
);
6046 if (cpus_weight(*cpu_map
) >
6047 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6048 sd
= &per_cpu(allnodes_domains
, i
);
6049 *sd
= SD_ALLNODES_INIT
;
6050 sd
->span
= *cpu_map
;
6051 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6057 sd
= &per_cpu(node_domains
, i
);
6059 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6063 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6067 sd
= &per_cpu(phys_domains
, i
);
6069 sd
->span
= nodemask
;
6073 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6075 #ifdef CONFIG_SCHED_MC
6077 sd
= &per_cpu(core_domains
, i
);
6079 sd
->span
= cpu_coregroup_map(i
);
6080 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6083 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6086 #ifdef CONFIG_SCHED_SMT
6088 sd
= &per_cpu(cpu_domains
, i
);
6089 *sd
= SD_SIBLING_INIT
;
6090 sd
->span
= cpu_sibling_map
[i
];
6091 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6094 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6098 #ifdef CONFIG_SCHED_SMT
6099 /* Set up CPU (sibling) groups */
6100 for_each_cpu_mask(i
, *cpu_map
) {
6101 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6102 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6103 if (i
!= first_cpu(this_sibling_map
))
6106 init_sched_build_groups(this_sibling_map
, cpu_map
,
6111 #ifdef CONFIG_SCHED_MC
6112 /* Set up multi-core groups */
6113 for_each_cpu_mask(i
, *cpu_map
) {
6114 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6115 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6116 if (i
!= first_cpu(this_core_map
))
6118 init_sched_build_groups(this_core_map
, cpu_map
,
6119 &cpu_to_core_group
);
6123 /* Set up physical groups */
6124 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6125 cpumask_t nodemask
= node_to_cpumask(i
);
6127 cpus_and(nodemask
, nodemask
, *cpu_map
);
6128 if (cpus_empty(nodemask
))
6131 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6135 /* Set up node groups */
6137 init_sched_build_groups(*cpu_map
, cpu_map
,
6138 &cpu_to_allnodes_group
);
6140 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6141 /* Set up node groups */
6142 struct sched_group
*sg
, *prev
;
6143 cpumask_t nodemask
= node_to_cpumask(i
);
6144 cpumask_t domainspan
;
6145 cpumask_t covered
= CPU_MASK_NONE
;
6148 cpus_and(nodemask
, nodemask
, *cpu_map
);
6149 if (cpus_empty(nodemask
)) {
6150 sched_group_nodes
[i
] = NULL
;
6154 domainspan
= sched_domain_node_span(i
);
6155 cpus_and(domainspan
, domainspan
, *cpu_map
);
6157 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6159 printk(KERN_WARNING
"Can not alloc domain group for "
6163 sched_group_nodes
[i
] = sg
;
6164 for_each_cpu_mask(j
, nodemask
) {
6165 struct sched_domain
*sd
;
6167 sd
= &per_cpu(node_domains
, j
);
6170 sg
->__cpu_power
= 0;
6171 sg
->cpumask
= nodemask
;
6173 cpus_or(covered
, covered
, nodemask
);
6176 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6177 cpumask_t tmp
, notcovered
;
6178 int n
= (i
+ j
) % MAX_NUMNODES
;
6180 cpus_complement(notcovered
, covered
);
6181 cpus_and(tmp
, notcovered
, *cpu_map
);
6182 cpus_and(tmp
, tmp
, domainspan
);
6183 if (cpus_empty(tmp
))
6186 nodemask
= node_to_cpumask(n
);
6187 cpus_and(tmp
, tmp
, nodemask
);
6188 if (cpus_empty(tmp
))
6191 sg
= kmalloc_node(sizeof(struct sched_group
),
6195 "Can not alloc domain group for node %d\n", j
);
6198 sg
->__cpu_power
= 0;
6200 sg
->next
= prev
->next
;
6201 cpus_or(covered
, covered
, tmp
);
6208 /* Calculate CPU power for physical packages and nodes */
6209 #ifdef CONFIG_SCHED_SMT
6210 for_each_cpu_mask(i
, *cpu_map
) {
6211 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6213 init_sched_groups_power(i
, sd
);
6216 #ifdef CONFIG_SCHED_MC
6217 for_each_cpu_mask(i
, *cpu_map
) {
6218 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6220 init_sched_groups_power(i
, sd
);
6224 for_each_cpu_mask(i
, *cpu_map
) {
6225 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6227 init_sched_groups_power(i
, sd
);
6231 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6232 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6235 struct sched_group
*sg
;
6237 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6238 init_numa_sched_groups_power(sg
);
6242 /* Attach the domains */
6243 for_each_cpu_mask(i
, *cpu_map
) {
6244 struct sched_domain
*sd
;
6245 #ifdef CONFIG_SCHED_SMT
6246 sd
= &per_cpu(cpu_domains
, i
);
6247 #elif defined(CONFIG_SCHED_MC)
6248 sd
= &per_cpu(core_domains
, i
);
6250 sd
= &per_cpu(phys_domains
, i
);
6252 cpu_attach_domain(sd
, i
);
6259 free_sched_groups(cpu_map
);
6264 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6266 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6268 cpumask_t cpu_default_map
;
6272 * Setup mask for cpus without special case scheduling requirements.
6273 * For now this just excludes isolated cpus, but could be used to
6274 * exclude other special cases in the future.
6276 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6278 err
= build_sched_domains(&cpu_default_map
);
6283 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6285 free_sched_groups(cpu_map
);
6289 * Detach sched domains from a group of cpus specified in cpu_map
6290 * These cpus will now be attached to the NULL domain
6292 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6296 for_each_cpu_mask(i
, *cpu_map
)
6297 cpu_attach_domain(NULL
, i
);
6298 synchronize_sched();
6299 arch_destroy_sched_domains(cpu_map
);
6303 * Partition sched domains as specified by the cpumasks below.
6304 * This attaches all cpus from the cpumasks to the NULL domain,
6305 * waits for a RCU quiescent period, recalculates sched
6306 * domain information and then attaches them back to the
6307 * correct sched domains
6308 * Call with hotplug lock held
6310 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6312 cpumask_t change_map
;
6315 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6316 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6317 cpus_or(change_map
, *partition1
, *partition2
);
6319 /* Detach sched domains from all of the affected cpus */
6320 detach_destroy_domains(&change_map
);
6321 if (!cpus_empty(*partition1
))
6322 err
= build_sched_domains(partition1
);
6323 if (!err
&& !cpus_empty(*partition2
))
6324 err
= build_sched_domains(partition2
);
6329 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6330 int arch_reinit_sched_domains(void)
6334 mutex_lock(&sched_hotcpu_mutex
);
6335 detach_destroy_domains(&cpu_online_map
);
6336 err
= arch_init_sched_domains(&cpu_online_map
);
6337 mutex_unlock(&sched_hotcpu_mutex
);
6342 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6346 if (buf
[0] != '0' && buf
[0] != '1')
6350 sched_smt_power_savings
= (buf
[0] == '1');
6352 sched_mc_power_savings
= (buf
[0] == '1');
6354 ret
= arch_reinit_sched_domains();
6356 return ret
? ret
: count
;
6359 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6363 #ifdef CONFIG_SCHED_SMT
6365 err
= sysfs_create_file(&cls
->kset
.kobj
,
6366 &attr_sched_smt_power_savings
.attr
);
6368 #ifdef CONFIG_SCHED_MC
6369 if (!err
&& mc_capable())
6370 err
= sysfs_create_file(&cls
->kset
.kobj
,
6371 &attr_sched_mc_power_savings
.attr
);
6377 #ifdef CONFIG_SCHED_MC
6378 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6380 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6382 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6383 const char *buf
, size_t count
)
6385 return sched_power_savings_store(buf
, count
, 0);
6387 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6388 sched_mc_power_savings_store
);
6391 #ifdef CONFIG_SCHED_SMT
6392 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6394 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6396 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6397 const char *buf
, size_t count
)
6399 return sched_power_savings_store(buf
, count
, 1);
6401 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6402 sched_smt_power_savings_store
);
6406 * Force a reinitialization of the sched domains hierarchy. The domains
6407 * and groups cannot be updated in place without racing with the balancing
6408 * code, so we temporarily attach all running cpus to the NULL domain
6409 * which will prevent rebalancing while the sched domains are recalculated.
6411 static int update_sched_domains(struct notifier_block
*nfb
,
6412 unsigned long action
, void *hcpu
)
6415 case CPU_UP_PREPARE
:
6416 case CPU_UP_PREPARE_FROZEN
:
6417 case CPU_DOWN_PREPARE
:
6418 case CPU_DOWN_PREPARE_FROZEN
:
6419 detach_destroy_domains(&cpu_online_map
);
6422 case CPU_UP_CANCELED
:
6423 case CPU_UP_CANCELED_FROZEN
:
6424 case CPU_DOWN_FAILED
:
6425 case CPU_DOWN_FAILED_FROZEN
:
6427 case CPU_ONLINE_FROZEN
:
6429 case CPU_DEAD_FROZEN
:
6431 * Fall through and re-initialise the domains.
6438 /* The hotplug lock is already held by cpu_up/cpu_down */
6439 arch_init_sched_domains(&cpu_online_map
);
6444 void __init
sched_init_smp(void)
6446 cpumask_t non_isolated_cpus
;
6448 mutex_lock(&sched_hotcpu_mutex
);
6449 arch_init_sched_domains(&cpu_online_map
);
6450 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6451 if (cpus_empty(non_isolated_cpus
))
6452 cpu_set(smp_processor_id(), non_isolated_cpus
);
6453 mutex_unlock(&sched_hotcpu_mutex
);
6454 /* XXX: Theoretical race here - CPU may be hotplugged now */
6455 hotcpu_notifier(update_sched_domains
, 0);
6457 init_sched_domain_sysctl();
6459 /* Move init over to a non-isolated CPU */
6460 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6462 sched_init_granularity();
6465 void __init
sched_init_smp(void)
6467 sched_init_granularity();
6469 #endif /* CONFIG_SMP */
6471 int in_sched_functions(unsigned long addr
)
6473 /* Linker adds these: start and end of __sched functions */
6474 extern char __sched_text_start
[], __sched_text_end
[];
6476 return in_lock_functions(addr
) ||
6477 (addr
>= (unsigned long)__sched_text_start
6478 && addr
< (unsigned long)__sched_text_end
);
6481 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6483 cfs_rq
->tasks_timeline
= RB_ROOT
;
6484 cfs_rq
->fair_clock
= 1;
6485 #ifdef CONFIG_FAIR_GROUP_SCHED
6490 void __init
sched_init(void)
6492 u64 now
= sched_clock();
6493 int highest_cpu
= 0;
6497 * Link up the scheduling class hierarchy:
6499 rt_sched_class
.next
= &fair_sched_class
;
6500 fair_sched_class
.next
= &idle_sched_class
;
6501 idle_sched_class
.next
= NULL
;
6503 for_each_possible_cpu(i
) {
6504 struct rt_prio_array
*array
;
6508 spin_lock_init(&rq
->lock
);
6509 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6512 init_cfs_rq(&rq
->cfs
, rq
);
6513 #ifdef CONFIG_FAIR_GROUP_SCHED
6514 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6515 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6517 rq
->ls
.load_update_last
= now
;
6518 rq
->ls
.load_update_start
= now
;
6520 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6521 rq
->cpu_load
[j
] = 0;
6524 rq
->active_balance
= 0;
6525 rq
->next_balance
= jiffies
;
6528 rq
->migration_thread
= NULL
;
6529 INIT_LIST_HEAD(&rq
->migration_queue
);
6531 atomic_set(&rq
->nr_iowait
, 0);
6533 array
= &rq
->rt
.active
;
6534 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6535 INIT_LIST_HEAD(array
->queue
+ j
);
6536 __clear_bit(j
, array
->bitmap
);
6539 /* delimiter for bitsearch: */
6540 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6543 set_load_weight(&init_task
);
6545 #ifdef CONFIG_PREEMPT_NOTIFIERS
6546 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6550 nr_cpu_ids
= highest_cpu
+ 1;
6551 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6554 #ifdef CONFIG_RT_MUTEXES
6555 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6559 * The boot idle thread does lazy MMU switching as well:
6561 atomic_inc(&init_mm
.mm_count
);
6562 enter_lazy_tlb(&init_mm
, current
);
6565 * Make us the idle thread. Technically, schedule() should not be
6566 * called from this thread, however somewhere below it might be,
6567 * but because we are the idle thread, we just pick up running again
6568 * when this runqueue becomes "idle".
6570 init_idle(current
, smp_processor_id());
6572 * During early bootup we pretend to be a normal task:
6574 current
->sched_class
= &fair_sched_class
;
6577 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6578 void __might_sleep(char *file
, int line
)
6581 static unsigned long prev_jiffy
; /* ratelimiting */
6583 if ((in_atomic() || irqs_disabled()) &&
6584 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6585 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6587 prev_jiffy
= jiffies
;
6588 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6589 " context at %s:%d\n", file
, line
);
6590 printk("in_atomic():%d, irqs_disabled():%d\n",
6591 in_atomic(), irqs_disabled());
6592 debug_show_held_locks(current
);
6593 if (irqs_disabled())
6594 print_irqtrace_events(current
);
6599 EXPORT_SYMBOL(__might_sleep
);
6602 #ifdef CONFIG_MAGIC_SYSRQ
6603 void normalize_rt_tasks(void)
6605 struct task_struct
*g
, *p
;
6606 unsigned long flags
;
6610 read_lock_irq(&tasklist_lock
);
6611 do_each_thread(g
, p
) {
6613 p
->se
.wait_runtime
= 0;
6614 p
->se
.exec_start
= 0;
6615 p
->se
.wait_start_fair
= 0;
6616 p
->se
.sleep_start_fair
= 0;
6617 #ifdef CONFIG_SCHEDSTATS
6618 p
->se
.wait_start
= 0;
6619 p
->se
.sleep_start
= 0;
6620 p
->se
.block_start
= 0;
6622 task_rq(p
)->cfs
.fair_clock
= 0;
6623 task_rq(p
)->clock
= 0;
6627 * Renice negative nice level userspace
6630 if (TASK_NICE(p
) < 0 && p
->mm
)
6631 set_user_nice(p
, 0);
6635 spin_lock_irqsave(&p
->pi_lock
, flags
);
6636 rq
= __task_rq_lock(p
);
6639 * Do not touch the migration thread:
6641 if (p
== rq
->migration_thread
)
6645 on_rq
= p
->se
.on_rq
;
6647 update_rq_clock(task_rq(p
));
6648 deactivate_task(task_rq(p
), p
, 0, task_rq(p
)->clock
);
6650 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6652 activate_task(task_rq(p
), p
, 0);
6653 resched_task(rq
->curr
);
6658 __task_rq_unlock(rq
);
6659 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6660 } while_each_thread(g
, p
);
6662 read_unlock_irq(&tasklist_lock
);
6665 #endif /* CONFIG_MAGIC_SYSRQ */
6669 * These functions are only useful for the IA64 MCA handling.
6671 * They can only be called when the whole system has been
6672 * stopped - every CPU needs to be quiescent, and no scheduling
6673 * activity can take place. Using them for anything else would
6674 * be a serious bug, and as a result, they aren't even visible
6675 * under any other configuration.
6679 * curr_task - return the current task for a given cpu.
6680 * @cpu: the processor in question.
6682 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6684 struct task_struct
*curr_task(int cpu
)
6686 return cpu_curr(cpu
);
6690 * set_curr_task - set the current task for a given cpu.
6691 * @cpu: the processor in question.
6692 * @p: the task pointer to set.
6694 * Description: This function must only be used when non-maskable interrupts
6695 * are serviced on a separate stack. It allows the architecture to switch the
6696 * notion of the current task on a cpu in a non-blocking manner. This function
6697 * must be called with all CPU's synchronized, and interrupts disabled, the
6698 * and caller must save the original value of the current task (see
6699 * curr_task() above) and restore that value before reenabling interrupts and
6700 * re-starting the system.
6702 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6704 void set_curr_task(int cpu
, struct task_struct
*p
)