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 * Shift right and round:
644 #define RSR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
647 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
648 struct load_weight
*lw
)
652 if (unlikely(!lw
->inv_weight
))
653 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
655 tmp
= (u64
)delta_exec
* weight
;
657 * Check whether we'd overflow the 64-bit multiplication:
659 if (unlikely(tmp
> WMULT_CONST
))
660 tmp
= RSR(RSR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
663 tmp
= RSR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
665 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
668 static inline unsigned long
669 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
671 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
674 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
680 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
687 * To aid in avoiding the subversion of "niceness" due to uneven distribution
688 * of tasks with abnormal "nice" values across CPUs the contribution that
689 * each task makes to its run queue's load is weighted according to its
690 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
691 * scaled version of the new time slice allocation that they receive on time
695 #define WEIGHT_IDLEPRIO 2
696 #define WMULT_IDLEPRIO (1 << 31)
699 * Nice levels are multiplicative, with a gentle 10% change for every
700 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
701 * nice 1, it will get ~10% less CPU time than another CPU-bound task
702 * that remained on nice 0.
704 * The "10% effect" is relative and cumulative: from _any_ nice level,
705 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
706 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
707 * If a task goes up by ~10% and another task goes down by ~10% then
708 * the relative distance between them is ~25%.)
710 static const int prio_to_weight
[40] = {
711 /* -20 */ 88761, 71755, 56483, 46273, 36291,
712 /* -15 */ 29154, 23254, 18705, 14949, 11916,
713 /* -10 */ 9548, 7620, 6100, 4904, 3906,
714 /* -5 */ 3121, 2501, 1991, 1586, 1277,
715 /* 0 */ 1024, 820, 655, 526, 423,
716 /* 5 */ 335, 272, 215, 172, 137,
717 /* 10 */ 110, 87, 70, 56, 45,
718 /* 15 */ 36, 29, 23, 18, 15,
722 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
724 * In cases where the weight does not change often, we can use the
725 * precalculated inverse to speed up arithmetics by turning divisions
726 * into multiplications:
728 static const u32 prio_to_wmult
[40] = {
729 /* -20 */ 48388, 59856, 76040, 92818, 118348,
730 /* -15 */ 147320, 184698, 229616, 287308, 360437,
731 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
732 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
733 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
734 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
735 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
736 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
739 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
742 * runqueue iterator, to support SMP load-balancing between different
743 * scheduling classes, without having to expose their internal data
744 * structures to the load-balancing proper:
748 struct task_struct
*(*start
)(void *);
749 struct task_struct
*(*next
)(void *);
752 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
753 unsigned long max_nr_move
, unsigned long max_load_move
,
754 struct sched_domain
*sd
, enum cpu_idle_type idle
,
755 int *all_pinned
, unsigned long *load_moved
,
756 int *this_best_prio
, struct rq_iterator
*iterator
);
758 #include "sched_stats.h"
759 #include "sched_rt.c"
760 #include "sched_fair.c"
761 #include "sched_idletask.c"
762 #ifdef CONFIG_SCHED_DEBUG
763 # include "sched_debug.c"
766 #define sched_class_highest (&rt_sched_class)
768 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
770 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
771 ls
->delta_exec
+= ls
->delta_stat
;
772 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
778 * Update delta_exec, delta_fair fields for rq.
780 * delta_fair clock advances at a rate inversely proportional to
781 * total load (rq->ls.load.weight) on the runqueue, while
782 * delta_exec advances at the same rate as wall-clock (provided
785 * delta_exec / delta_fair is a measure of the (smoothened) load on this
786 * runqueue over any given interval. This (smoothened) load is used
787 * during load balance.
789 * This function is called /before/ updating rq->ls.load
790 * and when switching tasks.
792 static void update_curr_load(struct rq
*rq
)
794 struct load_stat
*ls
= &rq
->ls
;
797 start
= ls
->load_update_start
;
798 ls
->load_update_start
= rq
->clock
;
799 ls
->delta_stat
+= rq
->clock
- start
;
801 * Stagger updates to ls->delta_fair. Very frequent updates
804 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
805 __update_curr_load(rq
, ls
);
808 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
810 update_curr_load(rq
);
811 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
814 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
816 update_curr_load(rq
);
817 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
820 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
826 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
832 static void set_load_weight(struct task_struct
*p
)
834 task_rq(p
)->cfs
.wait_runtime
-= p
->se
.wait_runtime
;
835 p
->se
.wait_runtime
= 0;
837 if (task_has_rt_policy(p
)) {
838 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
839 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
844 * SCHED_IDLE tasks get minimal weight:
846 if (p
->policy
== SCHED_IDLE
) {
847 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
848 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
852 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
853 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
856 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
858 sched_info_queued(p
);
859 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
863 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
865 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
870 * __normal_prio - return the priority that is based on the static prio
872 static inline int __normal_prio(struct task_struct
*p
)
874 return p
->static_prio
;
878 * Calculate the expected normal priority: i.e. priority
879 * without taking RT-inheritance into account. Might be
880 * boosted by interactivity modifiers. Changes upon fork,
881 * setprio syscalls, and whenever the interactivity
882 * estimator recalculates.
884 static inline int normal_prio(struct task_struct
*p
)
888 if (task_has_rt_policy(p
))
889 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
891 prio
= __normal_prio(p
);
896 * Calculate the current priority, i.e. the priority
897 * taken into account by the scheduler. This value might
898 * be boosted by RT tasks, or might be boosted by
899 * interactivity modifiers. Will be RT if the task got
900 * RT-boosted. If not then it returns p->normal_prio.
902 static int effective_prio(struct task_struct
*p
)
904 p
->normal_prio
= normal_prio(p
);
906 * If we are RT tasks or we were boosted to RT priority,
907 * keep the priority unchanged. Otherwise, update priority
908 * to the normal priority:
910 if (!rt_prio(p
->prio
))
911 return p
->normal_prio
;
916 * activate_task - move a task to the runqueue.
918 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
920 if (p
->state
== TASK_UNINTERRUPTIBLE
)
921 rq
->nr_uninterruptible
--;
923 enqueue_task(rq
, p
, wakeup
);
924 inc_nr_running(p
, rq
);
928 * activate_idle_task - move idle task to the _front_ of runqueue.
930 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
934 if (p
->state
== TASK_UNINTERRUPTIBLE
)
935 rq
->nr_uninterruptible
--;
937 enqueue_task(rq
, p
, 0);
938 inc_nr_running(p
, rq
);
942 * deactivate_task - remove a task from the runqueue.
944 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
946 if (p
->state
== TASK_UNINTERRUPTIBLE
)
947 rq
->nr_uninterruptible
++;
949 dequeue_task(rq
, p
, sleep
);
950 dec_nr_running(p
, rq
);
954 * task_curr - is this task currently executing on a CPU?
955 * @p: the task in question.
957 inline int task_curr(const struct task_struct
*p
)
959 return cpu_curr(task_cpu(p
)) == p
;
962 /* Used instead of source_load when we know the type == 0 */
963 unsigned long weighted_cpuload(const int cpu
)
965 return cpu_rq(cpu
)->ls
.load
.weight
;
968 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
971 task_thread_info(p
)->cpu
= cpu
;
978 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
980 int old_cpu
= task_cpu(p
);
981 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
982 u64 clock_offset
, fair_clock_offset
;
984 clock_offset
= old_rq
->clock
- new_rq
->clock
;
985 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
987 if (p
->se
.wait_start_fair
)
988 p
->se
.wait_start_fair
-= fair_clock_offset
;
989 if (p
->se
.sleep_start_fair
)
990 p
->se
.sleep_start_fair
-= fair_clock_offset
;
992 #ifdef CONFIG_SCHEDSTATS
993 if (p
->se
.wait_start
)
994 p
->se
.wait_start
-= clock_offset
;
995 if (p
->se
.sleep_start
)
996 p
->se
.sleep_start
-= clock_offset
;
997 if (p
->se
.block_start
)
998 p
->se
.block_start
-= clock_offset
;
1001 __set_task_cpu(p
, new_cpu
);
1004 struct migration_req
{
1005 struct list_head list
;
1007 struct task_struct
*task
;
1010 struct completion done
;
1014 * The task's runqueue lock must be held.
1015 * Returns true if you have to wait for migration thread.
1018 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1020 struct rq
*rq
= task_rq(p
);
1023 * If the task is not on a runqueue (and not running), then
1024 * it is sufficient to simply update the task's cpu field.
1026 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1027 set_task_cpu(p
, dest_cpu
);
1031 init_completion(&req
->done
);
1033 req
->dest_cpu
= dest_cpu
;
1034 list_add(&req
->list
, &rq
->migration_queue
);
1040 * wait_task_inactive - wait for a thread to unschedule.
1042 * The caller must ensure that the task *will* unschedule sometime soon,
1043 * else this function might spin for a *long* time. This function can't
1044 * be called with interrupts off, or it may introduce deadlock with
1045 * smp_call_function() if an IPI is sent by the same process we are
1046 * waiting to become inactive.
1048 void wait_task_inactive(struct task_struct
*p
)
1050 unsigned long flags
;
1056 * We do the initial early heuristics without holding
1057 * any task-queue locks at all. We'll only try to get
1058 * the runqueue lock when things look like they will
1064 * If the task is actively running on another CPU
1065 * still, just relax and busy-wait without holding
1068 * NOTE! Since we don't hold any locks, it's not
1069 * even sure that "rq" stays as the right runqueue!
1070 * But we don't care, since "task_running()" will
1071 * return false if the runqueue has changed and p
1072 * is actually now running somewhere else!
1074 while (task_running(rq
, p
))
1078 * Ok, time to look more closely! We need the rq
1079 * lock now, to be *sure*. If we're wrong, we'll
1080 * just go back and repeat.
1082 rq
= task_rq_lock(p
, &flags
);
1083 running
= task_running(rq
, p
);
1084 on_rq
= p
->se
.on_rq
;
1085 task_rq_unlock(rq
, &flags
);
1088 * Was it really running after all now that we
1089 * checked with the proper locks actually held?
1091 * Oops. Go back and try again..
1093 if (unlikely(running
)) {
1099 * It's not enough that it's not actively running,
1100 * it must be off the runqueue _entirely_, and not
1103 * So if it wa still runnable (but just not actively
1104 * running right now), it's preempted, and we should
1105 * yield - it could be a while.
1107 if (unlikely(on_rq
)) {
1113 * Ahh, all good. It wasn't running, and it wasn't
1114 * runnable, which means that it will never become
1115 * running in the future either. We're all done!
1120 * kick_process - kick a running thread to enter/exit the kernel
1121 * @p: the to-be-kicked thread
1123 * Cause a process which is running on another CPU to enter
1124 * kernel-mode, without any delay. (to get signals handled.)
1126 * NOTE: this function doesnt have to take the runqueue lock,
1127 * because all it wants to ensure is that the remote task enters
1128 * the kernel. If the IPI races and the task has been migrated
1129 * to another CPU then no harm is done and the purpose has been
1132 void kick_process(struct task_struct
*p
)
1138 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1139 smp_send_reschedule(cpu
);
1144 * Return a low guess at the load of a migration-source cpu weighted
1145 * according to the scheduling class and "nice" value.
1147 * We want to under-estimate the load of migration sources, to
1148 * balance conservatively.
1150 static inline unsigned long source_load(int cpu
, int type
)
1152 struct rq
*rq
= cpu_rq(cpu
);
1153 unsigned long total
= weighted_cpuload(cpu
);
1158 return min(rq
->cpu_load
[type
-1], total
);
1162 * Return a high guess at the load of a migration-target cpu weighted
1163 * according to the scheduling class and "nice" value.
1165 static inline unsigned long target_load(int cpu
, int type
)
1167 struct rq
*rq
= cpu_rq(cpu
);
1168 unsigned long total
= weighted_cpuload(cpu
);
1173 return max(rq
->cpu_load
[type
-1], total
);
1177 * Return the average load per task on the cpu's run queue
1179 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1181 struct rq
*rq
= cpu_rq(cpu
);
1182 unsigned long total
= weighted_cpuload(cpu
);
1183 unsigned long n
= rq
->nr_running
;
1185 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1189 * find_idlest_group finds and returns the least busy CPU group within the
1192 static struct sched_group
*
1193 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1195 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1196 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1197 int load_idx
= sd
->forkexec_idx
;
1198 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1201 unsigned long load
, avg_load
;
1205 /* Skip over this group if it has no CPUs allowed */
1206 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1209 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1211 /* Tally up the load of all CPUs in the group */
1214 for_each_cpu_mask(i
, group
->cpumask
) {
1215 /* Bias balancing toward cpus of our domain */
1217 load
= source_load(i
, load_idx
);
1219 load
= target_load(i
, load_idx
);
1224 /* Adjust by relative CPU power of the group */
1225 avg_load
= sg_div_cpu_power(group
,
1226 avg_load
* SCHED_LOAD_SCALE
);
1229 this_load
= avg_load
;
1231 } else if (avg_load
< min_load
) {
1232 min_load
= avg_load
;
1236 group
= group
->next
;
1237 } while (group
!= sd
->groups
);
1239 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1245 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1248 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1251 unsigned long load
, min_load
= ULONG_MAX
;
1255 /* Traverse only the allowed CPUs */
1256 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1258 for_each_cpu_mask(i
, tmp
) {
1259 load
= weighted_cpuload(i
);
1261 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1271 * sched_balance_self: balance the current task (running on cpu) in domains
1272 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1275 * Balance, ie. select the least loaded group.
1277 * Returns the target CPU number, or the same CPU if no balancing is needed.
1279 * preempt must be disabled.
1281 static int sched_balance_self(int cpu
, int flag
)
1283 struct task_struct
*t
= current
;
1284 struct sched_domain
*tmp
, *sd
= NULL
;
1286 for_each_domain(cpu
, tmp
) {
1288 * If power savings logic is enabled for a domain, stop there.
1290 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1292 if (tmp
->flags
& flag
)
1298 struct sched_group
*group
;
1299 int new_cpu
, weight
;
1301 if (!(sd
->flags
& flag
)) {
1307 group
= find_idlest_group(sd
, t
, cpu
);
1313 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1314 if (new_cpu
== -1 || new_cpu
== cpu
) {
1315 /* Now try balancing at a lower domain level of cpu */
1320 /* Now try balancing at a lower domain level of new_cpu */
1323 weight
= cpus_weight(span
);
1324 for_each_domain(cpu
, tmp
) {
1325 if (weight
<= cpus_weight(tmp
->span
))
1327 if (tmp
->flags
& flag
)
1330 /* while loop will break here if sd == NULL */
1336 #endif /* CONFIG_SMP */
1339 * wake_idle() will wake a task on an idle cpu if task->cpu is
1340 * not idle and an idle cpu is available. The span of cpus to
1341 * search starts with cpus closest then further out as needed,
1342 * so we always favor a closer, idle cpu.
1344 * Returns the CPU we should wake onto.
1346 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1347 static int wake_idle(int cpu
, struct task_struct
*p
)
1350 struct sched_domain
*sd
;
1354 * If it is idle, then it is the best cpu to run this task.
1356 * This cpu is also the best, if it has more than one task already.
1357 * Siblings must be also busy(in most cases) as they didn't already
1358 * pickup the extra load from this cpu and hence we need not check
1359 * sibling runqueue info. This will avoid the checks and cache miss
1360 * penalities associated with that.
1362 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1365 for_each_domain(cpu
, sd
) {
1366 if (sd
->flags
& SD_WAKE_IDLE
) {
1367 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1368 for_each_cpu_mask(i
, tmp
) {
1379 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1386 * try_to_wake_up - wake up a thread
1387 * @p: the to-be-woken-up thread
1388 * @state: the mask of task states that can be woken
1389 * @sync: do a synchronous wakeup?
1391 * Put it on the run-queue if it's not already there. The "current"
1392 * thread is always on the run-queue (except when the actual
1393 * re-schedule is in progress), and as such you're allowed to do
1394 * the simpler "current->state = TASK_RUNNING" to mark yourself
1395 * runnable without the overhead of this.
1397 * returns failure only if the task is already active.
1399 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1401 int cpu
, this_cpu
, success
= 0;
1402 unsigned long flags
;
1406 struct sched_domain
*sd
, *this_sd
= NULL
;
1407 unsigned long load
, this_load
;
1411 rq
= task_rq_lock(p
, &flags
);
1412 old_state
= p
->state
;
1413 if (!(old_state
& state
))
1420 this_cpu
= smp_processor_id();
1423 if (unlikely(task_running(rq
, p
)))
1428 schedstat_inc(rq
, ttwu_cnt
);
1429 if (cpu
== this_cpu
) {
1430 schedstat_inc(rq
, ttwu_local
);
1434 for_each_domain(this_cpu
, sd
) {
1435 if (cpu_isset(cpu
, sd
->span
)) {
1436 schedstat_inc(sd
, ttwu_wake_remote
);
1442 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1446 * Check for affine wakeup and passive balancing possibilities.
1449 int idx
= this_sd
->wake_idx
;
1450 unsigned int imbalance
;
1452 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1454 load
= source_load(cpu
, idx
);
1455 this_load
= target_load(this_cpu
, idx
);
1457 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1459 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1460 unsigned long tl
= this_load
;
1461 unsigned long tl_per_task
;
1463 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1466 * If sync wakeup then subtract the (maximum possible)
1467 * effect of the currently running task from the load
1468 * of the current CPU:
1471 tl
-= current
->se
.load
.weight
;
1474 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1475 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1477 * This domain has SD_WAKE_AFFINE and
1478 * p is cache cold in this domain, and
1479 * there is no bad imbalance.
1481 schedstat_inc(this_sd
, ttwu_move_affine
);
1487 * Start passive balancing when half the imbalance_pct
1490 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1491 if (imbalance
*this_load
<= 100*load
) {
1492 schedstat_inc(this_sd
, ttwu_move_balance
);
1498 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1500 new_cpu
= wake_idle(new_cpu
, p
);
1501 if (new_cpu
!= cpu
) {
1502 set_task_cpu(p
, new_cpu
);
1503 task_rq_unlock(rq
, &flags
);
1504 /* might preempt at this point */
1505 rq
= task_rq_lock(p
, &flags
);
1506 old_state
= p
->state
;
1507 if (!(old_state
& state
))
1512 this_cpu
= smp_processor_id();
1517 #endif /* CONFIG_SMP */
1518 update_rq_clock(rq
);
1519 activate_task(rq
, p
, 1);
1521 * Sync wakeups (i.e. those types of wakeups where the waker
1522 * has indicated that it will leave the CPU in short order)
1523 * don't trigger a preemption, if the woken up task will run on
1524 * this cpu. (in this case the 'I will reschedule' promise of
1525 * the waker guarantees that the freshly woken up task is going
1526 * to be considered on this CPU.)
1528 if (!sync
|| cpu
!= this_cpu
)
1529 check_preempt_curr(rq
, p
);
1533 p
->state
= TASK_RUNNING
;
1535 task_rq_unlock(rq
, &flags
);
1540 int fastcall
wake_up_process(struct task_struct
*p
)
1542 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1543 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1545 EXPORT_SYMBOL(wake_up_process
);
1547 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1549 return try_to_wake_up(p
, state
, 0);
1553 * Perform scheduler related setup for a newly forked process p.
1554 * p is forked by current.
1556 * __sched_fork() is basic setup used by init_idle() too:
1558 static void __sched_fork(struct task_struct
*p
)
1560 p
->se
.wait_start_fair
= 0;
1561 p
->se
.exec_start
= 0;
1562 p
->se
.sum_exec_runtime
= 0;
1563 p
->se
.delta_exec
= 0;
1564 p
->se
.delta_fair_run
= 0;
1565 p
->se
.delta_fair_sleep
= 0;
1566 p
->se
.wait_runtime
= 0;
1567 p
->se
.sleep_start_fair
= 0;
1569 #ifdef CONFIG_SCHEDSTATS
1570 p
->se
.wait_start
= 0;
1571 p
->se
.sum_wait_runtime
= 0;
1572 p
->se
.sum_sleep_runtime
= 0;
1573 p
->se
.sleep_start
= 0;
1574 p
->se
.block_start
= 0;
1575 p
->se
.sleep_max
= 0;
1576 p
->se
.block_max
= 0;
1579 p
->se
.wait_runtime_overruns
= 0;
1580 p
->se
.wait_runtime_underruns
= 0;
1583 INIT_LIST_HEAD(&p
->run_list
);
1586 #ifdef CONFIG_PREEMPT_NOTIFIERS
1587 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1591 * We mark the process as running here, but have not actually
1592 * inserted it onto the runqueue yet. This guarantees that
1593 * nobody will actually run it, and a signal or other external
1594 * event cannot wake it up and insert it on the runqueue either.
1596 p
->state
= TASK_RUNNING
;
1600 * fork()/clone()-time setup:
1602 void sched_fork(struct task_struct
*p
, int clone_flags
)
1604 int cpu
= get_cpu();
1609 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1611 __set_task_cpu(p
, cpu
);
1614 * Make sure we do not leak PI boosting priority to the child:
1616 p
->prio
= current
->normal_prio
;
1618 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1619 if (likely(sched_info_on()))
1620 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1622 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1625 #ifdef CONFIG_PREEMPT
1626 /* Want to start with kernel preemption disabled. */
1627 task_thread_info(p
)->preempt_count
= 1;
1633 * After fork, child runs first. (default) If set to 0 then
1634 * parent will (try to) run first.
1636 unsigned int __read_mostly sysctl_sched_child_runs_first
= 1;
1639 * wake_up_new_task - wake up a newly created task for the first time.
1641 * This function will do some initial scheduler statistics housekeeping
1642 * that must be done for every newly created context, then puts the task
1643 * on the runqueue and wakes it.
1645 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1647 unsigned long flags
;
1651 rq
= task_rq_lock(p
, &flags
);
1652 BUG_ON(p
->state
!= TASK_RUNNING
);
1653 this_cpu
= smp_processor_id(); /* parent's CPU */
1654 update_rq_clock(rq
);
1656 p
->prio
= effective_prio(p
);
1658 if (!p
->sched_class
->task_new
|| !sysctl_sched_child_runs_first
||
1659 (clone_flags
& CLONE_VM
) || task_cpu(p
) != this_cpu
||
1660 !current
->se
.on_rq
) {
1662 activate_task(rq
, p
, 0);
1665 * Let the scheduling class do new task startup
1666 * management (if any):
1668 p
->sched_class
->task_new(rq
, p
);
1669 inc_nr_running(p
, rq
);
1671 check_preempt_curr(rq
, p
);
1672 task_rq_unlock(rq
, &flags
);
1675 #ifdef CONFIG_PREEMPT_NOTIFIERS
1678 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1679 * @notifier: notifier struct to register
1681 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1683 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1685 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1688 * preempt_notifier_unregister - no longer interested in preemption notifications
1689 * @notifier: notifier struct to unregister
1691 * This is safe to call from within a preemption notifier.
1693 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1695 hlist_del(¬ifier
->link
);
1697 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1699 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1701 struct preempt_notifier
*notifier
;
1702 struct hlist_node
*node
;
1704 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1705 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1709 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1710 struct task_struct
*next
)
1712 struct preempt_notifier
*notifier
;
1713 struct hlist_node
*node
;
1715 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1716 notifier
->ops
->sched_out(notifier
, next
);
1721 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1726 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1727 struct task_struct
*next
)
1734 * prepare_task_switch - prepare to switch tasks
1735 * @rq: the runqueue preparing to switch
1736 * @prev: the current task that is being switched out
1737 * @next: the task we are going to switch to.
1739 * This is called with the rq lock held and interrupts off. It must
1740 * be paired with a subsequent finish_task_switch after the context
1743 * prepare_task_switch sets up locking and calls architecture specific
1747 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1748 struct task_struct
*next
)
1750 fire_sched_out_preempt_notifiers(prev
, next
);
1751 prepare_lock_switch(rq
, next
);
1752 prepare_arch_switch(next
);
1756 * finish_task_switch - clean up after a task-switch
1757 * @rq: runqueue associated with task-switch
1758 * @prev: the thread we just switched away from.
1760 * finish_task_switch must be called after the context switch, paired
1761 * with a prepare_task_switch call before the context switch.
1762 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1763 * and do any other architecture-specific cleanup actions.
1765 * Note that we may have delayed dropping an mm in context_switch(). If
1766 * so, we finish that here outside of the runqueue lock. (Doing it
1767 * with the lock held can cause deadlocks; see schedule() for
1770 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1771 __releases(rq
->lock
)
1773 struct mm_struct
*mm
= rq
->prev_mm
;
1779 * A task struct has one reference for the use as "current".
1780 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1781 * schedule one last time. The schedule call will never return, and
1782 * the scheduled task must drop that reference.
1783 * The test for TASK_DEAD must occur while the runqueue locks are
1784 * still held, otherwise prev could be scheduled on another cpu, die
1785 * there before we look at prev->state, and then the reference would
1787 * Manfred Spraul <manfred@colorfullife.com>
1789 prev_state
= prev
->state
;
1790 finish_arch_switch(prev
);
1791 finish_lock_switch(rq
, prev
);
1792 fire_sched_in_preempt_notifiers(current
);
1795 if (unlikely(prev_state
== TASK_DEAD
)) {
1797 * Remove function-return probe instances associated with this
1798 * task and put them back on the free list.
1800 kprobe_flush_task(prev
);
1801 put_task_struct(prev
);
1806 * schedule_tail - first thing a freshly forked thread must call.
1807 * @prev: the thread we just switched away from.
1809 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1810 __releases(rq
->lock
)
1812 struct rq
*rq
= this_rq();
1814 finish_task_switch(rq
, prev
);
1815 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1816 /* In this case, finish_task_switch does not reenable preemption */
1819 if (current
->set_child_tid
)
1820 put_user(current
->pid
, current
->set_child_tid
);
1824 * context_switch - switch to the new MM and the new
1825 * thread's register state.
1828 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1829 struct task_struct
*next
)
1831 struct mm_struct
*mm
, *oldmm
;
1833 prepare_task_switch(rq
, prev
, next
);
1835 oldmm
= prev
->active_mm
;
1837 * For paravirt, this is coupled with an exit in switch_to to
1838 * combine the page table reload and the switch backend into
1841 arch_enter_lazy_cpu_mode();
1843 if (unlikely(!mm
)) {
1844 next
->active_mm
= oldmm
;
1845 atomic_inc(&oldmm
->mm_count
);
1846 enter_lazy_tlb(oldmm
, next
);
1848 switch_mm(oldmm
, mm
, next
);
1850 if (unlikely(!prev
->mm
)) {
1851 prev
->active_mm
= NULL
;
1852 rq
->prev_mm
= oldmm
;
1855 * Since the runqueue lock will be released by the next
1856 * task (which is an invalid locking op but in the case
1857 * of the scheduler it's an obvious special-case), so we
1858 * do an early lockdep release here:
1860 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1861 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1864 /* Here we just switch the register state and the stack. */
1865 switch_to(prev
, next
, prev
);
1869 * this_rq must be evaluated again because prev may have moved
1870 * CPUs since it called schedule(), thus the 'rq' on its stack
1871 * frame will be invalid.
1873 finish_task_switch(this_rq(), prev
);
1877 * nr_running, nr_uninterruptible and nr_context_switches:
1879 * externally visible scheduler statistics: current number of runnable
1880 * threads, current number of uninterruptible-sleeping threads, total
1881 * number of context switches performed since bootup.
1883 unsigned long nr_running(void)
1885 unsigned long i
, sum
= 0;
1887 for_each_online_cpu(i
)
1888 sum
+= cpu_rq(i
)->nr_running
;
1893 unsigned long nr_uninterruptible(void)
1895 unsigned long i
, sum
= 0;
1897 for_each_possible_cpu(i
)
1898 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1901 * Since we read the counters lockless, it might be slightly
1902 * inaccurate. Do not allow it to go below zero though:
1904 if (unlikely((long)sum
< 0))
1910 unsigned long long nr_context_switches(void)
1913 unsigned long long sum
= 0;
1915 for_each_possible_cpu(i
)
1916 sum
+= cpu_rq(i
)->nr_switches
;
1921 unsigned long nr_iowait(void)
1923 unsigned long i
, sum
= 0;
1925 for_each_possible_cpu(i
)
1926 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1931 unsigned long nr_active(void)
1933 unsigned long i
, running
= 0, uninterruptible
= 0;
1935 for_each_online_cpu(i
) {
1936 running
+= cpu_rq(i
)->nr_running
;
1937 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1940 if (unlikely((long)uninterruptible
< 0))
1941 uninterruptible
= 0;
1943 return running
+ uninterruptible
;
1947 * Update rq->cpu_load[] statistics. This function is usually called every
1948 * scheduler tick (TICK_NSEC).
1950 static void update_cpu_load(struct rq
*this_rq
)
1952 u64 fair_delta64
, exec_delta64
, idle_delta64
, sample_interval64
, tmp64
;
1953 unsigned long total_load
= this_rq
->ls
.load
.weight
;
1954 unsigned long this_load
= total_load
;
1955 struct load_stat
*ls
= &this_rq
->ls
;
1958 this_rq
->nr_load_updates
++;
1959 if (unlikely(!(sysctl_sched_features
& SCHED_FEAT_PRECISE_CPU_LOAD
)))
1962 /* Update delta_fair/delta_exec fields first */
1963 update_curr_load(this_rq
);
1965 fair_delta64
= ls
->delta_fair
+ 1;
1968 exec_delta64
= ls
->delta_exec
+ 1;
1971 sample_interval64
= this_rq
->clock
- ls
->load_update_last
;
1972 ls
->load_update_last
= this_rq
->clock
;
1974 if ((s64
)sample_interval64
< (s64
)TICK_NSEC
)
1975 sample_interval64
= TICK_NSEC
;
1977 if (exec_delta64
> sample_interval64
)
1978 exec_delta64
= sample_interval64
;
1980 idle_delta64
= sample_interval64
- exec_delta64
;
1982 tmp64
= div64_64(SCHED_LOAD_SCALE
* exec_delta64
, fair_delta64
);
1983 tmp64
= div64_64(tmp64
* exec_delta64
, sample_interval64
);
1985 this_load
= (unsigned long)tmp64
;
1989 /* Update our load: */
1990 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1991 unsigned long old_load
, new_load
;
1993 /* scale is effectively 1 << i now, and >> i divides by scale */
1995 old_load
= this_rq
->cpu_load
[i
];
1996 new_load
= this_load
;
1998 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2005 * double_rq_lock - safely lock two runqueues
2007 * Note this does not disable interrupts like task_rq_lock,
2008 * you need to do so manually before calling.
2010 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2011 __acquires(rq1
->lock
)
2012 __acquires(rq2
->lock
)
2014 BUG_ON(!irqs_disabled());
2016 spin_lock(&rq1
->lock
);
2017 __acquire(rq2
->lock
); /* Fake it out ;) */
2020 spin_lock(&rq1
->lock
);
2021 spin_lock(&rq2
->lock
);
2023 spin_lock(&rq2
->lock
);
2024 spin_lock(&rq1
->lock
);
2027 update_rq_clock(rq1
);
2028 update_rq_clock(rq2
);
2032 * double_rq_unlock - safely unlock two runqueues
2034 * Note this does not restore interrupts like task_rq_unlock,
2035 * you need to do so manually after calling.
2037 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2038 __releases(rq1
->lock
)
2039 __releases(rq2
->lock
)
2041 spin_unlock(&rq1
->lock
);
2043 spin_unlock(&rq2
->lock
);
2045 __release(rq2
->lock
);
2049 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2051 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2052 __releases(this_rq
->lock
)
2053 __acquires(busiest
->lock
)
2054 __acquires(this_rq
->lock
)
2056 if (unlikely(!irqs_disabled())) {
2057 /* printk() doesn't work good under rq->lock */
2058 spin_unlock(&this_rq
->lock
);
2061 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2062 if (busiest
< this_rq
) {
2063 spin_unlock(&this_rq
->lock
);
2064 spin_lock(&busiest
->lock
);
2065 spin_lock(&this_rq
->lock
);
2067 spin_lock(&busiest
->lock
);
2072 * If dest_cpu is allowed for this process, migrate the task to it.
2073 * This is accomplished by forcing the cpu_allowed mask to only
2074 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2075 * the cpu_allowed mask is restored.
2077 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2079 struct migration_req req
;
2080 unsigned long flags
;
2083 rq
= task_rq_lock(p
, &flags
);
2084 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2085 || unlikely(cpu_is_offline(dest_cpu
)))
2088 /* force the process onto the specified CPU */
2089 if (migrate_task(p
, dest_cpu
, &req
)) {
2090 /* Need to wait for migration thread (might exit: take ref). */
2091 struct task_struct
*mt
= rq
->migration_thread
;
2093 get_task_struct(mt
);
2094 task_rq_unlock(rq
, &flags
);
2095 wake_up_process(mt
);
2096 put_task_struct(mt
);
2097 wait_for_completion(&req
.done
);
2102 task_rq_unlock(rq
, &flags
);
2106 * sched_exec - execve() is a valuable balancing opportunity, because at
2107 * this point the task has the smallest effective memory and cache footprint.
2109 void sched_exec(void)
2111 int new_cpu
, this_cpu
= get_cpu();
2112 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2114 if (new_cpu
!= this_cpu
)
2115 sched_migrate_task(current
, new_cpu
);
2119 * pull_task - move a task from a remote runqueue to the local runqueue.
2120 * Both runqueues must be locked.
2122 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2123 struct rq
*this_rq
, int this_cpu
)
2125 deactivate_task(src_rq
, p
, 0);
2126 set_task_cpu(p
, this_cpu
);
2127 activate_task(this_rq
, p
, 0);
2129 * Note that idle threads have a prio of MAX_PRIO, for this test
2130 * to be always true for them.
2132 check_preempt_curr(this_rq
, p
);
2136 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2139 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2140 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2144 * We do not migrate tasks that are:
2145 * 1) running (obviously), or
2146 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2147 * 3) are cache-hot on their current CPU.
2149 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2153 if (task_running(rq
, p
))
2157 * Aggressive migration if too many balance attempts have failed:
2159 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2165 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2166 unsigned long max_nr_move
, unsigned long max_load_move
,
2167 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2168 int *all_pinned
, unsigned long *load_moved
,
2169 int *this_best_prio
, struct rq_iterator
*iterator
)
2171 int pulled
= 0, pinned
= 0, skip_for_load
;
2172 struct task_struct
*p
;
2173 long rem_load_move
= max_load_move
;
2175 if (max_nr_move
== 0 || max_load_move
== 0)
2181 * Start the load-balancing iterator:
2183 p
= iterator
->start(iterator
->arg
);
2188 * To help distribute high priority tasks accross CPUs we don't
2189 * skip a task if it will be the highest priority task (i.e. smallest
2190 * prio value) on its new queue regardless of its load weight
2192 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2193 SCHED_LOAD_SCALE_FUZZ
;
2194 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2195 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2196 p
= iterator
->next(iterator
->arg
);
2200 pull_task(busiest
, p
, this_rq
, this_cpu
);
2202 rem_load_move
-= p
->se
.load
.weight
;
2205 * We only want to steal up to the prescribed number of tasks
2206 * and the prescribed amount of weighted load.
2208 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2209 if (p
->prio
< *this_best_prio
)
2210 *this_best_prio
= p
->prio
;
2211 p
= iterator
->next(iterator
->arg
);
2216 * Right now, this is the only place pull_task() is called,
2217 * so we can safely collect pull_task() stats here rather than
2218 * inside pull_task().
2220 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2223 *all_pinned
= pinned
;
2224 *load_moved
= max_load_move
- rem_load_move
;
2229 * move_tasks tries to move up to max_load_move weighted load from busiest to
2230 * this_rq, as part of a balancing operation within domain "sd".
2231 * Returns 1 if successful and 0 otherwise.
2233 * Called with both runqueues locked.
2235 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2236 unsigned long max_load_move
,
2237 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2240 struct sched_class
*class = sched_class_highest
;
2241 unsigned long total_load_moved
= 0;
2242 int this_best_prio
= this_rq
->curr
->prio
;
2246 class->load_balance(this_rq
, this_cpu
, busiest
,
2247 ULONG_MAX
, max_load_move
- total_load_moved
,
2248 sd
, idle
, all_pinned
, &this_best_prio
);
2249 class = class->next
;
2250 } while (class && max_load_move
> total_load_moved
);
2252 return total_load_moved
> 0;
2256 * move_one_task tries to move exactly one task from busiest to this_rq, as
2257 * part of active balancing operations within "domain".
2258 * Returns 1 if successful and 0 otherwise.
2260 * Called with both runqueues locked.
2262 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2263 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2265 struct sched_class
*class;
2266 int this_best_prio
= MAX_PRIO
;
2268 for (class = sched_class_highest
; class; class = class->next
)
2269 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2270 1, ULONG_MAX
, sd
, idle
, NULL
,
2278 * find_busiest_group finds and returns the busiest CPU group within the
2279 * domain. It calculates and returns the amount of weighted load which
2280 * should be moved to restore balance via the imbalance parameter.
2282 static struct sched_group
*
2283 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2284 unsigned long *imbalance
, enum cpu_idle_type idle
,
2285 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2287 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2288 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2289 unsigned long max_pull
;
2290 unsigned long busiest_load_per_task
, busiest_nr_running
;
2291 unsigned long this_load_per_task
, this_nr_running
;
2293 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2294 int power_savings_balance
= 1;
2295 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2296 unsigned long min_nr_running
= ULONG_MAX
;
2297 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2300 max_load
= this_load
= total_load
= total_pwr
= 0;
2301 busiest_load_per_task
= busiest_nr_running
= 0;
2302 this_load_per_task
= this_nr_running
= 0;
2303 if (idle
== CPU_NOT_IDLE
)
2304 load_idx
= sd
->busy_idx
;
2305 else if (idle
== CPU_NEWLY_IDLE
)
2306 load_idx
= sd
->newidle_idx
;
2308 load_idx
= sd
->idle_idx
;
2311 unsigned long load
, group_capacity
;
2314 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2315 unsigned long sum_nr_running
, sum_weighted_load
;
2317 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2320 balance_cpu
= first_cpu(group
->cpumask
);
2322 /* Tally up the load of all CPUs in the group */
2323 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2325 for_each_cpu_mask(i
, group
->cpumask
) {
2328 if (!cpu_isset(i
, *cpus
))
2333 if (*sd_idle
&& rq
->nr_running
)
2336 /* Bias balancing toward cpus of our domain */
2338 if (idle_cpu(i
) && !first_idle_cpu
) {
2343 load
= target_load(i
, load_idx
);
2345 load
= source_load(i
, load_idx
);
2348 sum_nr_running
+= rq
->nr_running
;
2349 sum_weighted_load
+= weighted_cpuload(i
);
2353 * First idle cpu or the first cpu(busiest) in this sched group
2354 * is eligible for doing load balancing at this and above
2355 * domains. In the newly idle case, we will allow all the cpu's
2356 * to do the newly idle load balance.
2358 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2359 balance_cpu
!= this_cpu
&& balance
) {
2364 total_load
+= avg_load
;
2365 total_pwr
+= group
->__cpu_power
;
2367 /* Adjust by relative CPU power of the group */
2368 avg_load
= sg_div_cpu_power(group
,
2369 avg_load
* SCHED_LOAD_SCALE
);
2371 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2374 this_load
= avg_load
;
2376 this_nr_running
= sum_nr_running
;
2377 this_load_per_task
= sum_weighted_load
;
2378 } else if (avg_load
> max_load
&&
2379 sum_nr_running
> group_capacity
) {
2380 max_load
= avg_load
;
2382 busiest_nr_running
= sum_nr_running
;
2383 busiest_load_per_task
= sum_weighted_load
;
2386 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2388 * Busy processors will not participate in power savings
2391 if (idle
== CPU_NOT_IDLE
||
2392 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2396 * If the local group is idle or completely loaded
2397 * no need to do power savings balance at this domain
2399 if (local_group
&& (this_nr_running
>= group_capacity
||
2401 power_savings_balance
= 0;
2404 * If a group is already running at full capacity or idle,
2405 * don't include that group in power savings calculations
2407 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2412 * Calculate the group which has the least non-idle load.
2413 * This is the group from where we need to pick up the load
2416 if ((sum_nr_running
< min_nr_running
) ||
2417 (sum_nr_running
== min_nr_running
&&
2418 first_cpu(group
->cpumask
) <
2419 first_cpu(group_min
->cpumask
))) {
2421 min_nr_running
= sum_nr_running
;
2422 min_load_per_task
= sum_weighted_load
/
2427 * Calculate the group which is almost near its
2428 * capacity but still has some space to pick up some load
2429 * from other group and save more power
2431 if (sum_nr_running
<= group_capacity
- 1) {
2432 if (sum_nr_running
> leader_nr_running
||
2433 (sum_nr_running
== leader_nr_running
&&
2434 first_cpu(group
->cpumask
) >
2435 first_cpu(group_leader
->cpumask
))) {
2436 group_leader
= group
;
2437 leader_nr_running
= sum_nr_running
;
2442 group
= group
->next
;
2443 } while (group
!= sd
->groups
);
2445 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2448 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2450 if (this_load
>= avg_load
||
2451 100*max_load
<= sd
->imbalance_pct
*this_load
)
2454 busiest_load_per_task
/= busiest_nr_running
;
2456 * We're trying to get all the cpus to the average_load, so we don't
2457 * want to push ourselves above the average load, nor do we wish to
2458 * reduce the max loaded cpu below the average load, as either of these
2459 * actions would just result in more rebalancing later, and ping-pong
2460 * tasks around. Thus we look for the minimum possible imbalance.
2461 * Negative imbalances (*we* are more loaded than anyone else) will
2462 * be counted as no imbalance for these purposes -- we can't fix that
2463 * by pulling tasks to us. Be careful of negative numbers as they'll
2464 * appear as very large values with unsigned longs.
2466 if (max_load
<= busiest_load_per_task
)
2470 * In the presence of smp nice balancing, certain scenarios can have
2471 * max load less than avg load(as we skip the groups at or below
2472 * its cpu_power, while calculating max_load..)
2474 if (max_load
< avg_load
) {
2476 goto small_imbalance
;
2479 /* Don't want to pull so many tasks that a group would go idle */
2480 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2482 /* How much load to actually move to equalise the imbalance */
2483 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2484 (avg_load
- this_load
) * this->__cpu_power
)
2488 * if *imbalance is less than the average load per runnable task
2489 * there is no gaurantee that any tasks will be moved so we'll have
2490 * a think about bumping its value to force at least one task to be
2493 if (*imbalance
+ SCHED_LOAD_SCALE_FUZZ
< busiest_load_per_task
/2) {
2494 unsigned long tmp
, pwr_now
, pwr_move
;
2498 pwr_move
= pwr_now
= 0;
2500 if (this_nr_running
) {
2501 this_load_per_task
/= this_nr_running
;
2502 if (busiest_load_per_task
> this_load_per_task
)
2505 this_load_per_task
= SCHED_LOAD_SCALE
;
2507 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2508 busiest_load_per_task
* imbn
) {
2509 *imbalance
= busiest_load_per_task
;
2514 * OK, we don't have enough imbalance to justify moving tasks,
2515 * however we may be able to increase total CPU power used by
2519 pwr_now
+= busiest
->__cpu_power
*
2520 min(busiest_load_per_task
, max_load
);
2521 pwr_now
+= this->__cpu_power
*
2522 min(this_load_per_task
, this_load
);
2523 pwr_now
/= SCHED_LOAD_SCALE
;
2525 /* Amount of load we'd subtract */
2526 tmp
= sg_div_cpu_power(busiest
,
2527 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2529 pwr_move
+= busiest
->__cpu_power
*
2530 min(busiest_load_per_task
, max_load
- tmp
);
2532 /* Amount of load we'd add */
2533 if (max_load
* busiest
->__cpu_power
<
2534 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2535 tmp
= sg_div_cpu_power(this,
2536 max_load
* busiest
->__cpu_power
);
2538 tmp
= sg_div_cpu_power(this,
2539 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2540 pwr_move
+= this->__cpu_power
*
2541 min(this_load_per_task
, this_load
+ tmp
);
2542 pwr_move
/= SCHED_LOAD_SCALE
;
2544 /* Move if we gain throughput */
2545 if (pwr_move
<= pwr_now
)
2548 *imbalance
= busiest_load_per_task
;
2554 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2555 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2558 if (this == group_leader
&& group_leader
!= group_min
) {
2559 *imbalance
= min_load_per_task
;
2569 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2572 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2573 unsigned long imbalance
, cpumask_t
*cpus
)
2575 struct rq
*busiest
= NULL
, *rq
;
2576 unsigned long max_load
= 0;
2579 for_each_cpu_mask(i
, group
->cpumask
) {
2582 if (!cpu_isset(i
, *cpus
))
2586 wl
= weighted_cpuload(i
);
2588 if (rq
->nr_running
== 1 && wl
> imbalance
)
2591 if (wl
> max_load
) {
2601 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2602 * so long as it is large enough.
2604 #define MAX_PINNED_INTERVAL 512
2607 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2608 * tasks if there is an imbalance.
2610 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2611 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2614 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2615 struct sched_group
*group
;
2616 unsigned long imbalance
;
2618 cpumask_t cpus
= CPU_MASK_ALL
;
2619 unsigned long flags
;
2622 * When power savings policy is enabled for the parent domain, idle
2623 * sibling can pick up load irrespective of busy siblings. In this case,
2624 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2625 * portraying it as CPU_NOT_IDLE.
2627 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2628 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2631 schedstat_inc(sd
, lb_cnt
[idle
]);
2634 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2641 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2645 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2647 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2651 BUG_ON(busiest
== this_rq
);
2653 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2656 if (busiest
->nr_running
> 1) {
2658 * Attempt to move tasks. If find_busiest_group has found
2659 * an imbalance but busiest->nr_running <= 1, the group is
2660 * still unbalanced. ld_moved simply stays zero, so it is
2661 * correctly treated as an imbalance.
2663 local_irq_save(flags
);
2664 double_rq_lock(this_rq
, busiest
);
2665 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2666 imbalance
, sd
, idle
, &all_pinned
);
2667 double_rq_unlock(this_rq
, busiest
);
2668 local_irq_restore(flags
);
2671 * some other cpu did the load balance for us.
2673 if (ld_moved
&& this_cpu
!= smp_processor_id())
2674 resched_cpu(this_cpu
);
2676 /* All tasks on this runqueue were pinned by CPU affinity */
2677 if (unlikely(all_pinned
)) {
2678 cpu_clear(cpu_of(busiest
), cpus
);
2679 if (!cpus_empty(cpus
))
2686 schedstat_inc(sd
, lb_failed
[idle
]);
2687 sd
->nr_balance_failed
++;
2689 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2691 spin_lock_irqsave(&busiest
->lock
, flags
);
2693 /* don't kick the migration_thread, if the curr
2694 * task on busiest cpu can't be moved to this_cpu
2696 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2697 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2699 goto out_one_pinned
;
2702 if (!busiest
->active_balance
) {
2703 busiest
->active_balance
= 1;
2704 busiest
->push_cpu
= this_cpu
;
2707 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2709 wake_up_process(busiest
->migration_thread
);
2712 * We've kicked active balancing, reset the failure
2715 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2718 sd
->nr_balance_failed
= 0;
2720 if (likely(!active_balance
)) {
2721 /* We were unbalanced, so reset the balancing interval */
2722 sd
->balance_interval
= sd
->min_interval
;
2725 * If we've begun active balancing, start to back off. This
2726 * case may not be covered by the all_pinned logic if there
2727 * is only 1 task on the busy runqueue (because we don't call
2730 if (sd
->balance_interval
< sd
->max_interval
)
2731 sd
->balance_interval
*= 2;
2734 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2735 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2740 schedstat_inc(sd
, lb_balanced
[idle
]);
2742 sd
->nr_balance_failed
= 0;
2745 /* tune up the balancing interval */
2746 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2747 (sd
->balance_interval
< sd
->max_interval
))
2748 sd
->balance_interval
*= 2;
2750 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2751 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2757 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2758 * tasks if there is an imbalance.
2760 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2761 * this_rq is locked.
2764 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2766 struct sched_group
*group
;
2767 struct rq
*busiest
= NULL
;
2768 unsigned long imbalance
;
2772 cpumask_t cpus
= CPU_MASK_ALL
;
2775 * When power savings policy is enabled for the parent domain, idle
2776 * sibling can pick up load irrespective of busy siblings. In this case,
2777 * let the state of idle sibling percolate up as IDLE, instead of
2778 * portraying it as CPU_NOT_IDLE.
2780 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2781 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2784 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2786 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2787 &sd_idle
, &cpus
, NULL
);
2789 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2793 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2796 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2800 BUG_ON(busiest
== this_rq
);
2802 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2805 if (busiest
->nr_running
> 1) {
2806 /* Attempt to move tasks */
2807 double_lock_balance(this_rq
, busiest
);
2808 /* this_rq->clock is already updated */
2809 update_rq_clock(busiest
);
2810 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2811 imbalance
, sd
, CPU_NEWLY_IDLE
,
2813 spin_unlock(&busiest
->lock
);
2815 if (unlikely(all_pinned
)) {
2816 cpu_clear(cpu_of(busiest
), cpus
);
2817 if (!cpus_empty(cpus
))
2823 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2824 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2825 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2828 sd
->nr_balance_failed
= 0;
2833 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2834 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2835 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2837 sd
->nr_balance_failed
= 0;
2843 * idle_balance is called by schedule() if this_cpu is about to become
2844 * idle. Attempts to pull tasks from other CPUs.
2846 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2848 struct sched_domain
*sd
;
2849 int pulled_task
= -1;
2850 unsigned long next_balance
= jiffies
+ HZ
;
2852 for_each_domain(this_cpu
, sd
) {
2853 unsigned long interval
;
2855 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2858 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2859 /* If we've pulled tasks over stop searching: */
2860 pulled_task
= load_balance_newidle(this_cpu
,
2863 interval
= msecs_to_jiffies(sd
->balance_interval
);
2864 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2865 next_balance
= sd
->last_balance
+ interval
;
2869 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2871 * We are going idle. next_balance may be set based on
2872 * a busy processor. So reset next_balance.
2874 this_rq
->next_balance
= next_balance
;
2879 * active_load_balance is run by migration threads. It pushes running tasks
2880 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2881 * running on each physical CPU where possible, and avoids physical /
2882 * logical imbalances.
2884 * Called with busiest_rq locked.
2886 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2888 int target_cpu
= busiest_rq
->push_cpu
;
2889 struct sched_domain
*sd
;
2890 struct rq
*target_rq
;
2892 /* Is there any task to move? */
2893 if (busiest_rq
->nr_running
<= 1)
2896 target_rq
= cpu_rq(target_cpu
);
2899 * This condition is "impossible", if it occurs
2900 * we need to fix it. Originally reported by
2901 * Bjorn Helgaas on a 128-cpu setup.
2903 BUG_ON(busiest_rq
== target_rq
);
2905 /* move a task from busiest_rq to target_rq */
2906 double_lock_balance(busiest_rq
, target_rq
);
2907 update_rq_clock(busiest_rq
);
2908 update_rq_clock(target_rq
);
2910 /* Search for an sd spanning us and the target CPU. */
2911 for_each_domain(target_cpu
, sd
) {
2912 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2913 cpu_isset(busiest_cpu
, sd
->span
))
2918 schedstat_inc(sd
, alb_cnt
);
2920 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2922 schedstat_inc(sd
, alb_pushed
);
2924 schedstat_inc(sd
, alb_failed
);
2926 spin_unlock(&target_rq
->lock
);
2931 atomic_t load_balancer
;
2933 } nohz ____cacheline_aligned
= {
2934 .load_balancer
= ATOMIC_INIT(-1),
2935 .cpu_mask
= CPU_MASK_NONE
,
2939 * This routine will try to nominate the ilb (idle load balancing)
2940 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2941 * load balancing on behalf of all those cpus. If all the cpus in the system
2942 * go into this tickless mode, then there will be no ilb owner (as there is
2943 * no need for one) and all the cpus will sleep till the next wakeup event
2946 * For the ilb owner, tick is not stopped. And this tick will be used
2947 * for idle load balancing. ilb owner will still be part of
2950 * While stopping the tick, this cpu will become the ilb owner if there
2951 * is no other owner. And will be the owner till that cpu becomes busy
2952 * or if all cpus in the system stop their ticks at which point
2953 * there is no need for ilb owner.
2955 * When the ilb owner becomes busy, it nominates another owner, during the
2956 * next busy scheduler_tick()
2958 int select_nohz_load_balancer(int stop_tick
)
2960 int cpu
= smp_processor_id();
2963 cpu_set(cpu
, nohz
.cpu_mask
);
2964 cpu_rq(cpu
)->in_nohz_recently
= 1;
2967 * If we are going offline and still the leader, give up!
2969 if (cpu_is_offline(cpu
) &&
2970 atomic_read(&nohz
.load_balancer
) == cpu
) {
2971 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2976 /* time for ilb owner also to sleep */
2977 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2978 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2979 atomic_set(&nohz
.load_balancer
, -1);
2983 if (atomic_read(&nohz
.load_balancer
) == -1) {
2984 /* make me the ilb owner */
2985 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2987 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2990 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2993 cpu_clear(cpu
, nohz
.cpu_mask
);
2995 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2996 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3003 static DEFINE_SPINLOCK(balancing
);
3006 * It checks each scheduling domain to see if it is due to be balanced,
3007 * and initiates a balancing operation if so.
3009 * Balancing parameters are set up in arch_init_sched_domains.
3011 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3014 struct rq
*rq
= cpu_rq(cpu
);
3015 unsigned long interval
;
3016 struct sched_domain
*sd
;
3017 /* Earliest time when we have to do rebalance again */
3018 unsigned long next_balance
= jiffies
+ 60*HZ
;
3020 for_each_domain(cpu
, sd
) {
3021 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3024 interval
= sd
->balance_interval
;
3025 if (idle
!= CPU_IDLE
)
3026 interval
*= sd
->busy_factor
;
3028 /* scale ms to jiffies */
3029 interval
= msecs_to_jiffies(interval
);
3030 if (unlikely(!interval
))
3032 if (interval
> HZ
*NR_CPUS
/10)
3033 interval
= HZ
*NR_CPUS
/10;
3036 if (sd
->flags
& SD_SERIALIZE
) {
3037 if (!spin_trylock(&balancing
))
3041 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3042 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3044 * We've pulled tasks over so either we're no
3045 * longer idle, or one of our SMT siblings is
3048 idle
= CPU_NOT_IDLE
;
3050 sd
->last_balance
= jiffies
;
3052 if (sd
->flags
& SD_SERIALIZE
)
3053 spin_unlock(&balancing
);
3055 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3056 next_balance
= sd
->last_balance
+ interval
;
3059 * Stop the load balance at this level. There is another
3060 * CPU in our sched group which is doing load balancing more
3066 rq
->next_balance
= next_balance
;
3070 * run_rebalance_domains is triggered when needed from the scheduler tick.
3071 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3072 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3074 static void run_rebalance_domains(struct softirq_action
*h
)
3076 int this_cpu
= smp_processor_id();
3077 struct rq
*this_rq
= cpu_rq(this_cpu
);
3078 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3079 CPU_IDLE
: CPU_NOT_IDLE
;
3081 rebalance_domains(this_cpu
, idle
);
3085 * If this cpu is the owner for idle load balancing, then do the
3086 * balancing on behalf of the other idle cpus whose ticks are
3089 if (this_rq
->idle_at_tick
&&
3090 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3091 cpumask_t cpus
= nohz
.cpu_mask
;
3095 cpu_clear(this_cpu
, cpus
);
3096 for_each_cpu_mask(balance_cpu
, cpus
) {
3098 * If this cpu gets work to do, stop the load balancing
3099 * work being done for other cpus. Next load
3100 * balancing owner will pick it up.
3105 rebalance_domains(balance_cpu
, SCHED_IDLE
);
3107 rq
= cpu_rq(balance_cpu
);
3108 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3109 this_rq
->next_balance
= rq
->next_balance
;
3116 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3118 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3119 * idle load balancing owner or decide to stop the periodic load balancing,
3120 * if the whole system is idle.
3122 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3126 * If we were in the nohz mode recently and busy at the current
3127 * scheduler tick, then check if we need to nominate new idle
3130 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3131 rq
->in_nohz_recently
= 0;
3133 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3134 cpu_clear(cpu
, nohz
.cpu_mask
);
3135 atomic_set(&nohz
.load_balancer
, -1);
3138 if (atomic_read(&nohz
.load_balancer
) == -1) {
3140 * simple selection for now: Nominate the
3141 * first cpu in the nohz list to be the next
3144 * TBD: Traverse the sched domains and nominate
3145 * the nearest cpu in the nohz.cpu_mask.
3147 int ilb
= first_cpu(nohz
.cpu_mask
);
3155 * If this cpu is idle and doing idle load balancing for all the
3156 * cpus with ticks stopped, is it time for that to stop?
3158 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3159 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3165 * If this cpu is idle and the idle load balancing is done by
3166 * someone else, then no need raise the SCHED_SOFTIRQ
3168 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3169 cpu_isset(cpu
, nohz
.cpu_mask
))
3172 if (time_after_eq(jiffies
, rq
->next_balance
))
3173 raise_softirq(SCHED_SOFTIRQ
);
3176 #else /* CONFIG_SMP */
3179 * on UP we do not need to balance between CPUs:
3181 static inline void idle_balance(int cpu
, struct rq
*rq
)
3185 /* Avoid "used but not defined" warning on UP */
3186 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3187 unsigned long max_nr_move
, unsigned long max_load_move
,
3188 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3189 int *all_pinned
, unsigned long *load_moved
,
3190 int *this_best_prio
, struct rq_iterator
*iterator
)
3199 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3201 EXPORT_PER_CPU_SYMBOL(kstat
);
3204 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3205 * that have not yet been banked in case the task is currently running.
3207 unsigned long long task_sched_runtime(struct task_struct
*p
)
3209 unsigned long flags
;
3213 rq
= task_rq_lock(p
, &flags
);
3214 ns
= p
->se
.sum_exec_runtime
;
3215 if (rq
->curr
== p
) {
3216 update_rq_clock(rq
);
3217 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3218 if ((s64
)delta_exec
> 0)
3221 task_rq_unlock(rq
, &flags
);
3227 * Account user cpu time to a process.
3228 * @p: the process that the cpu time gets accounted to
3229 * @hardirq_offset: the offset to subtract from hardirq_count()
3230 * @cputime: the cpu time spent in user space since the last update
3232 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3234 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3237 p
->utime
= cputime_add(p
->utime
, cputime
);
3239 /* Add user time to cpustat. */
3240 tmp
= cputime_to_cputime64(cputime
);
3241 if (TASK_NICE(p
) > 0)
3242 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3244 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3248 * Account system cpu time to a process.
3249 * @p: the process that the cpu time gets accounted to
3250 * @hardirq_offset: the offset to subtract from hardirq_count()
3251 * @cputime: the cpu time spent in kernel space since the last update
3253 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3256 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3257 struct rq
*rq
= this_rq();
3260 p
->stime
= cputime_add(p
->stime
, cputime
);
3262 /* Add system time to cpustat. */
3263 tmp
= cputime_to_cputime64(cputime
);
3264 if (hardirq_count() - hardirq_offset
)
3265 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3266 else if (softirq_count())
3267 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3268 else if (p
!= rq
->idle
)
3269 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3270 else if (atomic_read(&rq
->nr_iowait
) > 0)
3271 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3273 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3274 /* Account for system time used */
3275 acct_update_integrals(p
);
3279 * Account for involuntary wait time.
3280 * @p: the process from which the cpu time has been stolen
3281 * @steal: the cpu time spent in involuntary wait
3283 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3285 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3286 cputime64_t tmp
= cputime_to_cputime64(steal
);
3287 struct rq
*rq
= this_rq();
3289 if (p
== rq
->idle
) {
3290 p
->stime
= cputime_add(p
->stime
, steal
);
3291 if (atomic_read(&rq
->nr_iowait
) > 0)
3292 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3294 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3296 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3300 * This function gets called by the timer code, with HZ frequency.
3301 * We call it with interrupts disabled.
3303 * It also gets called by the fork code, when changing the parent's
3306 void scheduler_tick(void)
3308 int cpu
= smp_processor_id();
3309 struct rq
*rq
= cpu_rq(cpu
);
3310 struct task_struct
*curr
= rq
->curr
;
3312 spin_lock(&rq
->lock
);
3313 __update_rq_clock(rq
);
3314 update_cpu_load(rq
);
3315 if (curr
!= rq
->idle
) /* FIXME: needed? */
3316 curr
->sched_class
->task_tick(rq
, curr
);
3317 spin_unlock(&rq
->lock
);
3320 rq
->idle_at_tick
= idle_cpu(cpu
);
3321 trigger_load_balance(rq
, cpu
);
3325 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3327 void fastcall
add_preempt_count(int val
)
3332 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3334 preempt_count() += val
;
3336 * Spinlock count overflowing soon?
3338 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3341 EXPORT_SYMBOL(add_preempt_count
);
3343 void fastcall
sub_preempt_count(int val
)
3348 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3351 * Is the spinlock portion underflowing?
3353 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3354 !(preempt_count() & PREEMPT_MASK
)))
3357 preempt_count() -= val
;
3359 EXPORT_SYMBOL(sub_preempt_count
);
3364 * Print scheduling while atomic bug:
3366 static noinline
void __schedule_bug(struct task_struct
*prev
)
3368 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3369 prev
->comm
, preempt_count(), prev
->pid
);
3370 debug_show_held_locks(prev
);
3371 if (irqs_disabled())
3372 print_irqtrace_events(prev
);
3377 * Various schedule()-time debugging checks and statistics:
3379 static inline void schedule_debug(struct task_struct
*prev
)
3382 * Test if we are atomic. Since do_exit() needs to call into
3383 * schedule() atomically, we ignore that path for now.
3384 * Otherwise, whine if we are scheduling when we should not be.
3386 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3387 __schedule_bug(prev
);
3389 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3391 schedstat_inc(this_rq(), sched_cnt
);
3395 * Pick up the highest-prio task:
3397 static inline struct task_struct
*
3398 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3400 struct sched_class
*class;
3401 struct task_struct
*p
;
3404 * Optimization: we know that if all tasks are in
3405 * the fair class we can call that function directly:
3407 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3408 p
= fair_sched_class
.pick_next_task(rq
);
3413 class = sched_class_highest
;
3415 p
= class->pick_next_task(rq
);
3419 * Will never be NULL as the idle class always
3420 * returns a non-NULL p:
3422 class = class->next
;
3427 * schedule() is the main scheduler function.
3429 asmlinkage
void __sched
schedule(void)
3431 struct task_struct
*prev
, *next
;
3438 cpu
= smp_processor_id();
3442 switch_count
= &prev
->nivcsw
;
3444 release_kernel_lock(prev
);
3445 need_resched_nonpreemptible
:
3447 schedule_debug(prev
);
3449 spin_lock_irq(&rq
->lock
);
3450 clear_tsk_need_resched(prev
);
3451 __update_rq_clock(rq
);
3453 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3454 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3455 unlikely(signal_pending(prev
)))) {
3456 prev
->state
= TASK_RUNNING
;
3458 deactivate_task(rq
, prev
, 1);
3460 switch_count
= &prev
->nvcsw
;
3463 if (unlikely(!rq
->nr_running
))
3464 idle_balance(cpu
, rq
);
3466 prev
->sched_class
->put_prev_task(rq
, prev
);
3467 next
= pick_next_task(rq
, prev
);
3469 sched_info_switch(prev
, next
);
3471 if (likely(prev
!= next
)) {
3476 context_switch(rq
, prev
, next
); /* unlocks the rq */
3478 spin_unlock_irq(&rq
->lock
);
3480 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3481 cpu
= smp_processor_id();
3483 goto need_resched_nonpreemptible
;
3485 preempt_enable_no_resched();
3486 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3489 EXPORT_SYMBOL(schedule
);
3491 #ifdef CONFIG_PREEMPT
3493 * this is the entry point to schedule() from in-kernel preemption
3494 * off of preempt_enable. Kernel preemptions off return from interrupt
3495 * occur there and call schedule directly.
3497 asmlinkage
void __sched
preempt_schedule(void)
3499 struct thread_info
*ti
= current_thread_info();
3500 #ifdef CONFIG_PREEMPT_BKL
3501 struct task_struct
*task
= current
;
3502 int saved_lock_depth
;
3505 * If there is a non-zero preempt_count or interrupts are disabled,
3506 * we do not want to preempt the current task. Just return..
3508 if (likely(ti
->preempt_count
|| irqs_disabled()))
3512 add_preempt_count(PREEMPT_ACTIVE
);
3514 * We keep the big kernel semaphore locked, but we
3515 * clear ->lock_depth so that schedule() doesnt
3516 * auto-release the semaphore:
3518 #ifdef CONFIG_PREEMPT_BKL
3519 saved_lock_depth
= task
->lock_depth
;
3520 task
->lock_depth
= -1;
3523 #ifdef CONFIG_PREEMPT_BKL
3524 task
->lock_depth
= saved_lock_depth
;
3526 sub_preempt_count(PREEMPT_ACTIVE
);
3528 /* we could miss a preemption opportunity between schedule and now */
3530 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3533 EXPORT_SYMBOL(preempt_schedule
);
3536 * this is the entry point to schedule() from kernel preemption
3537 * off of irq context.
3538 * Note, that this is called and return with irqs disabled. This will
3539 * protect us against recursive calling from irq.
3541 asmlinkage
void __sched
preempt_schedule_irq(void)
3543 struct thread_info
*ti
= current_thread_info();
3544 #ifdef CONFIG_PREEMPT_BKL
3545 struct task_struct
*task
= current
;
3546 int saved_lock_depth
;
3548 /* Catch callers which need to be fixed */
3549 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3552 add_preempt_count(PREEMPT_ACTIVE
);
3554 * We keep the big kernel semaphore locked, but we
3555 * clear ->lock_depth so that schedule() doesnt
3556 * auto-release the semaphore:
3558 #ifdef CONFIG_PREEMPT_BKL
3559 saved_lock_depth
= task
->lock_depth
;
3560 task
->lock_depth
= -1;
3564 local_irq_disable();
3565 #ifdef CONFIG_PREEMPT_BKL
3566 task
->lock_depth
= saved_lock_depth
;
3568 sub_preempt_count(PREEMPT_ACTIVE
);
3570 /* we could miss a preemption opportunity between schedule and now */
3572 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3576 #endif /* CONFIG_PREEMPT */
3578 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3581 return try_to_wake_up(curr
->private, mode
, sync
);
3583 EXPORT_SYMBOL(default_wake_function
);
3586 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3587 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3588 * number) then we wake all the non-exclusive tasks and one exclusive task.
3590 * There are circumstances in which we can try to wake a task which has already
3591 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3592 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3594 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3595 int nr_exclusive
, int sync
, void *key
)
3597 struct list_head
*tmp
, *next
;
3599 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3600 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3601 unsigned flags
= curr
->flags
;
3603 if (curr
->func(curr
, mode
, sync
, key
) &&
3604 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3610 * __wake_up - wake up threads blocked on a waitqueue.
3612 * @mode: which threads
3613 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3614 * @key: is directly passed to the wakeup function
3616 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3617 int nr_exclusive
, void *key
)
3619 unsigned long flags
;
3621 spin_lock_irqsave(&q
->lock
, flags
);
3622 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3623 spin_unlock_irqrestore(&q
->lock
, flags
);
3625 EXPORT_SYMBOL(__wake_up
);
3628 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3630 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3632 __wake_up_common(q
, mode
, 1, 0, NULL
);
3636 * __wake_up_sync - wake up threads blocked on a waitqueue.
3638 * @mode: which threads
3639 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3641 * The sync wakeup differs that the waker knows that it will schedule
3642 * away soon, so while the target thread will be woken up, it will not
3643 * be migrated to another CPU - ie. the two threads are 'synchronized'
3644 * with each other. This can prevent needless bouncing between CPUs.
3646 * On UP it can prevent extra preemption.
3649 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3651 unsigned long flags
;
3657 if (unlikely(!nr_exclusive
))
3660 spin_lock_irqsave(&q
->lock
, flags
);
3661 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3662 spin_unlock_irqrestore(&q
->lock
, flags
);
3664 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3666 void fastcall
complete(struct completion
*x
)
3668 unsigned long flags
;
3670 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3672 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3674 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3676 EXPORT_SYMBOL(complete
);
3678 void fastcall
complete_all(struct completion
*x
)
3680 unsigned long flags
;
3682 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3683 x
->done
+= UINT_MAX
/2;
3684 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3686 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3688 EXPORT_SYMBOL(complete_all
);
3690 void fastcall __sched
wait_for_completion(struct completion
*x
)
3694 spin_lock_irq(&x
->wait
.lock
);
3696 DECLARE_WAITQUEUE(wait
, current
);
3698 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3699 __add_wait_queue_tail(&x
->wait
, &wait
);
3701 __set_current_state(TASK_UNINTERRUPTIBLE
);
3702 spin_unlock_irq(&x
->wait
.lock
);
3704 spin_lock_irq(&x
->wait
.lock
);
3706 __remove_wait_queue(&x
->wait
, &wait
);
3709 spin_unlock_irq(&x
->wait
.lock
);
3711 EXPORT_SYMBOL(wait_for_completion
);
3713 unsigned long fastcall __sched
3714 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3718 spin_lock_irq(&x
->wait
.lock
);
3720 DECLARE_WAITQUEUE(wait
, current
);
3722 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3723 __add_wait_queue_tail(&x
->wait
, &wait
);
3725 __set_current_state(TASK_UNINTERRUPTIBLE
);
3726 spin_unlock_irq(&x
->wait
.lock
);
3727 timeout
= schedule_timeout(timeout
);
3728 spin_lock_irq(&x
->wait
.lock
);
3730 __remove_wait_queue(&x
->wait
, &wait
);
3734 __remove_wait_queue(&x
->wait
, &wait
);
3738 spin_unlock_irq(&x
->wait
.lock
);
3741 EXPORT_SYMBOL(wait_for_completion_timeout
);
3743 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3749 spin_lock_irq(&x
->wait
.lock
);
3751 DECLARE_WAITQUEUE(wait
, current
);
3753 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3754 __add_wait_queue_tail(&x
->wait
, &wait
);
3756 if (signal_pending(current
)) {
3758 __remove_wait_queue(&x
->wait
, &wait
);
3761 __set_current_state(TASK_INTERRUPTIBLE
);
3762 spin_unlock_irq(&x
->wait
.lock
);
3764 spin_lock_irq(&x
->wait
.lock
);
3766 __remove_wait_queue(&x
->wait
, &wait
);
3770 spin_unlock_irq(&x
->wait
.lock
);
3774 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3776 unsigned long fastcall __sched
3777 wait_for_completion_interruptible_timeout(struct completion
*x
,
3778 unsigned long timeout
)
3782 spin_lock_irq(&x
->wait
.lock
);
3784 DECLARE_WAITQUEUE(wait
, current
);
3786 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3787 __add_wait_queue_tail(&x
->wait
, &wait
);
3789 if (signal_pending(current
)) {
3790 timeout
= -ERESTARTSYS
;
3791 __remove_wait_queue(&x
->wait
, &wait
);
3794 __set_current_state(TASK_INTERRUPTIBLE
);
3795 spin_unlock_irq(&x
->wait
.lock
);
3796 timeout
= schedule_timeout(timeout
);
3797 spin_lock_irq(&x
->wait
.lock
);
3799 __remove_wait_queue(&x
->wait
, &wait
);
3803 __remove_wait_queue(&x
->wait
, &wait
);
3807 spin_unlock_irq(&x
->wait
.lock
);
3810 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3813 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3815 spin_lock_irqsave(&q
->lock
, *flags
);
3816 __add_wait_queue(q
, wait
);
3817 spin_unlock(&q
->lock
);
3821 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3823 spin_lock_irq(&q
->lock
);
3824 __remove_wait_queue(q
, wait
);
3825 spin_unlock_irqrestore(&q
->lock
, *flags
);
3828 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3830 unsigned long flags
;
3833 init_waitqueue_entry(&wait
, current
);
3835 current
->state
= TASK_INTERRUPTIBLE
;
3837 sleep_on_head(q
, &wait
, &flags
);
3839 sleep_on_tail(q
, &wait
, &flags
);
3841 EXPORT_SYMBOL(interruptible_sleep_on
);
3844 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3846 unsigned long flags
;
3849 init_waitqueue_entry(&wait
, current
);
3851 current
->state
= TASK_INTERRUPTIBLE
;
3853 sleep_on_head(q
, &wait
, &flags
);
3854 timeout
= schedule_timeout(timeout
);
3855 sleep_on_tail(q
, &wait
, &flags
);
3859 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3861 void __sched
sleep_on(wait_queue_head_t
*q
)
3863 unsigned long flags
;
3866 init_waitqueue_entry(&wait
, current
);
3868 current
->state
= TASK_UNINTERRUPTIBLE
;
3870 sleep_on_head(q
, &wait
, &flags
);
3872 sleep_on_tail(q
, &wait
, &flags
);
3874 EXPORT_SYMBOL(sleep_on
);
3876 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3878 unsigned long flags
;
3881 init_waitqueue_entry(&wait
, current
);
3883 current
->state
= TASK_UNINTERRUPTIBLE
;
3885 sleep_on_head(q
, &wait
, &flags
);
3886 timeout
= schedule_timeout(timeout
);
3887 sleep_on_tail(q
, &wait
, &flags
);
3891 EXPORT_SYMBOL(sleep_on_timeout
);
3893 #ifdef CONFIG_RT_MUTEXES
3896 * rt_mutex_setprio - set the current priority of a task
3898 * @prio: prio value (kernel-internal form)
3900 * This function changes the 'effective' priority of a task. It does
3901 * not touch ->normal_prio like __setscheduler().
3903 * Used by the rt_mutex code to implement priority inheritance logic.
3905 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3907 unsigned long flags
;
3911 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3913 rq
= task_rq_lock(p
, &flags
);
3914 update_rq_clock(rq
);
3917 on_rq
= p
->se
.on_rq
;
3919 dequeue_task(rq
, p
, 0);
3922 p
->sched_class
= &rt_sched_class
;
3924 p
->sched_class
= &fair_sched_class
;
3929 enqueue_task(rq
, p
, 0);
3931 * Reschedule if we are currently running on this runqueue and
3932 * our priority decreased, or if we are not currently running on
3933 * this runqueue and our priority is higher than the current's
3935 if (task_running(rq
, p
)) {
3936 if (p
->prio
> oldprio
)
3937 resched_task(rq
->curr
);
3939 check_preempt_curr(rq
, p
);
3942 task_rq_unlock(rq
, &flags
);
3947 void set_user_nice(struct task_struct
*p
, long nice
)
3949 int old_prio
, delta
, on_rq
;
3950 unsigned long flags
;
3953 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3956 * We have to be careful, if called from sys_setpriority(),
3957 * the task might be in the middle of scheduling on another CPU.
3959 rq
= task_rq_lock(p
, &flags
);
3960 update_rq_clock(rq
);
3962 * The RT priorities are set via sched_setscheduler(), but we still
3963 * allow the 'normal' nice value to be set - but as expected
3964 * it wont have any effect on scheduling until the task is
3965 * SCHED_FIFO/SCHED_RR:
3967 if (task_has_rt_policy(p
)) {
3968 p
->static_prio
= NICE_TO_PRIO(nice
);
3971 on_rq
= p
->se
.on_rq
;
3973 dequeue_task(rq
, p
, 0);
3977 p
->static_prio
= NICE_TO_PRIO(nice
);
3980 p
->prio
= effective_prio(p
);
3981 delta
= p
->prio
- old_prio
;
3984 enqueue_task(rq
, p
, 0);
3987 * If the task increased its priority or is running and
3988 * lowered its priority, then reschedule its CPU:
3990 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3991 resched_task(rq
->curr
);
3994 task_rq_unlock(rq
, &flags
);
3996 EXPORT_SYMBOL(set_user_nice
);
3999 * can_nice - check if a task can reduce its nice value
4003 int can_nice(const struct task_struct
*p
, const int nice
)
4005 /* convert nice value [19,-20] to rlimit style value [1,40] */
4006 int nice_rlim
= 20 - nice
;
4008 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4009 capable(CAP_SYS_NICE
));
4012 #ifdef __ARCH_WANT_SYS_NICE
4015 * sys_nice - change the priority of the current process.
4016 * @increment: priority increment
4018 * sys_setpriority is a more generic, but much slower function that
4019 * does similar things.
4021 asmlinkage
long sys_nice(int increment
)
4026 * Setpriority might change our priority at the same moment.
4027 * We don't have to worry. Conceptually one call occurs first
4028 * and we have a single winner.
4030 if (increment
< -40)
4035 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4041 if (increment
< 0 && !can_nice(current
, nice
))
4044 retval
= security_task_setnice(current
, nice
);
4048 set_user_nice(current
, nice
);
4055 * task_prio - return the priority value of a given task.
4056 * @p: the task in question.
4058 * This is the priority value as seen by users in /proc.
4059 * RT tasks are offset by -200. Normal tasks are centered
4060 * around 0, value goes from -16 to +15.
4062 int task_prio(const struct task_struct
*p
)
4064 return p
->prio
- MAX_RT_PRIO
;
4068 * task_nice - return the nice value of a given task.
4069 * @p: the task in question.
4071 int task_nice(const struct task_struct
*p
)
4073 return TASK_NICE(p
);
4075 EXPORT_SYMBOL_GPL(task_nice
);
4078 * idle_cpu - is a given cpu idle currently?
4079 * @cpu: the processor in question.
4081 int idle_cpu(int cpu
)
4083 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4087 * idle_task - return the idle task for a given cpu.
4088 * @cpu: the processor in question.
4090 struct task_struct
*idle_task(int cpu
)
4092 return cpu_rq(cpu
)->idle
;
4096 * find_process_by_pid - find a process with a matching PID value.
4097 * @pid: the pid in question.
4099 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4101 return pid
? find_task_by_pid(pid
) : current
;
4104 /* Actually do priority change: must hold rq lock. */
4106 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4108 BUG_ON(p
->se
.on_rq
);
4111 switch (p
->policy
) {
4115 p
->sched_class
= &fair_sched_class
;
4119 p
->sched_class
= &rt_sched_class
;
4123 p
->rt_priority
= prio
;
4124 p
->normal_prio
= normal_prio(p
);
4125 /* we are holding p->pi_lock already */
4126 p
->prio
= rt_mutex_getprio(p
);
4131 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4132 * @p: the task in question.
4133 * @policy: new policy.
4134 * @param: structure containing the new RT priority.
4136 * NOTE that the task may be already dead.
4138 int sched_setscheduler(struct task_struct
*p
, int policy
,
4139 struct sched_param
*param
)
4141 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4142 unsigned long flags
;
4145 /* may grab non-irq protected spin_locks */
4146 BUG_ON(in_interrupt());
4148 /* double check policy once rq lock held */
4150 policy
= oldpolicy
= p
->policy
;
4151 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4152 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4153 policy
!= SCHED_IDLE
)
4156 * Valid priorities for SCHED_FIFO and SCHED_RR are
4157 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4158 * SCHED_BATCH and SCHED_IDLE is 0.
4160 if (param
->sched_priority
< 0 ||
4161 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4162 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4164 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4168 * Allow unprivileged RT tasks to decrease priority:
4170 if (!capable(CAP_SYS_NICE
)) {
4171 if (rt_policy(policy
)) {
4172 unsigned long rlim_rtprio
;
4174 if (!lock_task_sighand(p
, &flags
))
4176 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4177 unlock_task_sighand(p
, &flags
);
4179 /* can't set/change the rt policy */
4180 if (policy
!= p
->policy
&& !rlim_rtprio
)
4183 /* can't increase priority */
4184 if (param
->sched_priority
> p
->rt_priority
&&
4185 param
->sched_priority
> rlim_rtprio
)
4189 * Like positive nice levels, dont allow tasks to
4190 * move out of SCHED_IDLE either:
4192 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4195 /* can't change other user's priorities */
4196 if ((current
->euid
!= p
->euid
) &&
4197 (current
->euid
!= p
->uid
))
4201 retval
= security_task_setscheduler(p
, policy
, param
);
4205 * make sure no PI-waiters arrive (or leave) while we are
4206 * changing the priority of the task:
4208 spin_lock_irqsave(&p
->pi_lock
, flags
);
4210 * To be able to change p->policy safely, the apropriate
4211 * runqueue lock must be held.
4213 rq
= __task_rq_lock(p
);
4214 /* recheck policy now with rq lock held */
4215 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4216 policy
= oldpolicy
= -1;
4217 __task_rq_unlock(rq
);
4218 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4221 update_rq_clock(rq
);
4222 on_rq
= p
->se
.on_rq
;
4224 deactivate_task(rq
, p
, 0);
4226 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4228 activate_task(rq
, p
, 0);
4230 * Reschedule if we are currently running on this runqueue and
4231 * our priority decreased, or if we are not currently running on
4232 * this runqueue and our priority is higher than the current's
4234 if (task_running(rq
, p
)) {
4235 if (p
->prio
> oldprio
)
4236 resched_task(rq
->curr
);
4238 check_preempt_curr(rq
, p
);
4241 __task_rq_unlock(rq
);
4242 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4244 rt_mutex_adjust_pi(p
);
4248 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4251 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4253 struct sched_param lparam
;
4254 struct task_struct
*p
;
4257 if (!param
|| pid
< 0)
4259 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4264 p
= find_process_by_pid(pid
);
4266 retval
= sched_setscheduler(p
, policy
, &lparam
);
4273 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4274 * @pid: the pid in question.
4275 * @policy: new policy.
4276 * @param: structure containing the new RT priority.
4278 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4279 struct sched_param __user
*param
)
4281 /* negative values for policy are not valid */
4285 return do_sched_setscheduler(pid
, policy
, param
);
4289 * sys_sched_setparam - set/change the RT priority of a thread
4290 * @pid: the pid in question.
4291 * @param: structure containing the new RT priority.
4293 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4295 return do_sched_setscheduler(pid
, -1, param
);
4299 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4300 * @pid: the pid in question.
4302 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4304 struct task_struct
*p
;
4305 int retval
= -EINVAL
;
4311 read_lock(&tasklist_lock
);
4312 p
= find_process_by_pid(pid
);
4314 retval
= security_task_getscheduler(p
);
4318 read_unlock(&tasklist_lock
);
4325 * sys_sched_getscheduler - get the RT priority of a thread
4326 * @pid: the pid in question.
4327 * @param: structure containing the RT priority.
4329 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4331 struct sched_param lp
;
4332 struct task_struct
*p
;
4333 int retval
= -EINVAL
;
4335 if (!param
|| pid
< 0)
4338 read_lock(&tasklist_lock
);
4339 p
= find_process_by_pid(pid
);
4344 retval
= security_task_getscheduler(p
);
4348 lp
.sched_priority
= p
->rt_priority
;
4349 read_unlock(&tasklist_lock
);
4352 * This one might sleep, we cannot do it with a spinlock held ...
4354 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4360 read_unlock(&tasklist_lock
);
4364 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4366 cpumask_t cpus_allowed
;
4367 struct task_struct
*p
;
4370 mutex_lock(&sched_hotcpu_mutex
);
4371 read_lock(&tasklist_lock
);
4373 p
= find_process_by_pid(pid
);
4375 read_unlock(&tasklist_lock
);
4376 mutex_unlock(&sched_hotcpu_mutex
);
4381 * It is not safe to call set_cpus_allowed with the
4382 * tasklist_lock held. We will bump the task_struct's
4383 * usage count and then drop tasklist_lock.
4386 read_unlock(&tasklist_lock
);
4389 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4390 !capable(CAP_SYS_NICE
))
4393 retval
= security_task_setscheduler(p
, 0, NULL
);
4397 cpus_allowed
= cpuset_cpus_allowed(p
);
4398 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4399 retval
= set_cpus_allowed(p
, new_mask
);
4403 mutex_unlock(&sched_hotcpu_mutex
);
4407 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4408 cpumask_t
*new_mask
)
4410 if (len
< sizeof(cpumask_t
)) {
4411 memset(new_mask
, 0, sizeof(cpumask_t
));
4412 } else if (len
> sizeof(cpumask_t
)) {
4413 len
= sizeof(cpumask_t
);
4415 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4419 * sys_sched_setaffinity - set the cpu affinity of a process
4420 * @pid: pid of the process
4421 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4422 * @user_mask_ptr: user-space pointer to the new cpu mask
4424 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4425 unsigned long __user
*user_mask_ptr
)
4430 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4434 return sched_setaffinity(pid
, new_mask
);
4438 * Represents all cpu's present in the system
4439 * In systems capable of hotplug, this map could dynamically grow
4440 * as new cpu's are detected in the system via any platform specific
4441 * method, such as ACPI for e.g.
4444 cpumask_t cpu_present_map __read_mostly
;
4445 EXPORT_SYMBOL(cpu_present_map
);
4448 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4449 EXPORT_SYMBOL(cpu_online_map
);
4451 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4452 EXPORT_SYMBOL(cpu_possible_map
);
4455 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4457 struct task_struct
*p
;
4460 mutex_lock(&sched_hotcpu_mutex
);
4461 read_lock(&tasklist_lock
);
4464 p
= find_process_by_pid(pid
);
4468 retval
= security_task_getscheduler(p
);
4472 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4475 read_unlock(&tasklist_lock
);
4476 mutex_unlock(&sched_hotcpu_mutex
);
4482 * sys_sched_getaffinity - get the cpu affinity of a process
4483 * @pid: pid of the process
4484 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4485 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4487 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4488 unsigned long __user
*user_mask_ptr
)
4493 if (len
< sizeof(cpumask_t
))
4496 ret
= sched_getaffinity(pid
, &mask
);
4500 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4503 return sizeof(cpumask_t
);
4507 * sys_sched_yield - yield the current processor to other threads.
4509 * This function yields the current CPU to other tasks. If there are no
4510 * other threads running on this CPU then this function will return.
4512 asmlinkage
long sys_sched_yield(void)
4514 struct rq
*rq
= this_rq_lock();
4516 schedstat_inc(rq
, yld_cnt
);
4517 if (unlikely(rq
->nr_running
== 1))
4518 schedstat_inc(rq
, yld_act_empty
);
4520 current
->sched_class
->yield_task(rq
, current
);
4523 * Since we are going to call schedule() anyway, there's
4524 * no need to preempt or enable interrupts:
4526 __release(rq
->lock
);
4527 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4528 _raw_spin_unlock(&rq
->lock
);
4529 preempt_enable_no_resched();
4536 static void __cond_resched(void)
4538 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4539 __might_sleep(__FILE__
, __LINE__
);
4542 * The BKS might be reacquired before we have dropped
4543 * PREEMPT_ACTIVE, which could trigger a second
4544 * cond_resched() call.
4547 add_preempt_count(PREEMPT_ACTIVE
);
4549 sub_preempt_count(PREEMPT_ACTIVE
);
4550 } while (need_resched());
4553 int __sched
cond_resched(void)
4555 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4556 system_state
== SYSTEM_RUNNING
) {
4562 EXPORT_SYMBOL(cond_resched
);
4565 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4566 * call schedule, and on return reacquire the lock.
4568 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4569 * operations here to prevent schedule() from being called twice (once via
4570 * spin_unlock(), once by hand).
4572 int cond_resched_lock(spinlock_t
*lock
)
4576 if (need_lockbreak(lock
)) {
4582 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4583 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4584 _raw_spin_unlock(lock
);
4585 preempt_enable_no_resched();
4592 EXPORT_SYMBOL(cond_resched_lock
);
4594 int __sched
cond_resched_softirq(void)
4596 BUG_ON(!in_softirq());
4598 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4606 EXPORT_SYMBOL(cond_resched_softirq
);
4609 * yield - yield the current processor to other threads.
4611 * This is a shortcut for kernel-space yielding - it marks the
4612 * thread runnable and calls sys_sched_yield().
4614 void __sched
yield(void)
4616 set_current_state(TASK_RUNNING
);
4619 EXPORT_SYMBOL(yield
);
4622 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4623 * that process accounting knows that this is a task in IO wait state.
4625 * But don't do that if it is a deliberate, throttling IO wait (this task
4626 * has set its backing_dev_info: the queue against which it should throttle)
4628 void __sched
io_schedule(void)
4630 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4632 delayacct_blkio_start();
4633 atomic_inc(&rq
->nr_iowait
);
4635 atomic_dec(&rq
->nr_iowait
);
4636 delayacct_blkio_end();
4638 EXPORT_SYMBOL(io_schedule
);
4640 long __sched
io_schedule_timeout(long timeout
)
4642 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4645 delayacct_blkio_start();
4646 atomic_inc(&rq
->nr_iowait
);
4647 ret
= schedule_timeout(timeout
);
4648 atomic_dec(&rq
->nr_iowait
);
4649 delayacct_blkio_end();
4654 * sys_sched_get_priority_max - return maximum RT priority.
4655 * @policy: scheduling class.
4657 * this syscall returns the maximum rt_priority that can be used
4658 * by a given scheduling class.
4660 asmlinkage
long sys_sched_get_priority_max(int policy
)
4667 ret
= MAX_USER_RT_PRIO
-1;
4679 * sys_sched_get_priority_min - return minimum RT priority.
4680 * @policy: scheduling class.
4682 * this syscall returns the minimum rt_priority that can be used
4683 * by a given scheduling class.
4685 asmlinkage
long sys_sched_get_priority_min(int policy
)
4703 * sys_sched_rr_get_interval - return the default timeslice of a process.
4704 * @pid: pid of the process.
4705 * @interval: userspace pointer to the timeslice value.
4707 * this syscall writes the default timeslice value of a given process
4708 * into the user-space timespec buffer. A value of '0' means infinity.
4711 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4713 struct task_struct
*p
;
4714 int retval
= -EINVAL
;
4721 read_lock(&tasklist_lock
);
4722 p
= find_process_by_pid(pid
);
4726 retval
= security_task_getscheduler(p
);
4730 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4731 0 : static_prio_timeslice(p
->static_prio
), &t
);
4732 read_unlock(&tasklist_lock
);
4733 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4737 read_unlock(&tasklist_lock
);
4741 static const char stat_nam
[] = "RSDTtZX";
4743 static void show_task(struct task_struct
*p
)
4745 unsigned long free
= 0;
4748 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4749 printk("%-13.13s %c", p
->comm
,
4750 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4751 #if BITS_PER_LONG == 32
4752 if (state
== TASK_RUNNING
)
4753 printk(" running ");
4755 printk(" %08lx ", thread_saved_pc(p
));
4757 if (state
== TASK_RUNNING
)
4758 printk(" running task ");
4760 printk(" %016lx ", thread_saved_pc(p
));
4762 #ifdef CONFIG_DEBUG_STACK_USAGE
4764 unsigned long *n
= end_of_stack(p
);
4767 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4770 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4772 if (state
!= TASK_RUNNING
)
4773 show_stack(p
, NULL
);
4776 void show_state_filter(unsigned long state_filter
)
4778 struct task_struct
*g
, *p
;
4780 #if BITS_PER_LONG == 32
4782 " task PC stack pid father\n");
4785 " task PC stack pid father\n");
4787 read_lock(&tasklist_lock
);
4788 do_each_thread(g
, p
) {
4790 * reset the NMI-timeout, listing all files on a slow
4791 * console might take alot of time:
4793 touch_nmi_watchdog();
4794 if (!state_filter
|| (p
->state
& state_filter
))
4796 } while_each_thread(g
, p
);
4798 touch_all_softlockup_watchdogs();
4800 #ifdef CONFIG_SCHED_DEBUG
4801 sysrq_sched_debug_show();
4803 read_unlock(&tasklist_lock
);
4805 * Only show locks if all tasks are dumped:
4807 if (state_filter
== -1)
4808 debug_show_all_locks();
4811 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4813 idle
->sched_class
= &idle_sched_class
;
4817 * init_idle - set up an idle thread for a given CPU
4818 * @idle: task in question
4819 * @cpu: cpu the idle task belongs to
4821 * NOTE: this function does not set the idle thread's NEED_RESCHED
4822 * flag, to make booting more robust.
4824 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4826 struct rq
*rq
= cpu_rq(cpu
);
4827 unsigned long flags
;
4830 idle
->se
.exec_start
= sched_clock();
4832 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4833 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4834 __set_task_cpu(idle
, cpu
);
4836 spin_lock_irqsave(&rq
->lock
, flags
);
4837 rq
->curr
= rq
->idle
= idle
;
4838 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4841 spin_unlock_irqrestore(&rq
->lock
, flags
);
4843 /* Set the preempt count _outside_ the spinlocks! */
4844 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4845 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4847 task_thread_info(idle
)->preempt_count
= 0;
4850 * The idle tasks have their own, simple scheduling class:
4852 idle
->sched_class
= &idle_sched_class
;
4856 * In a system that switches off the HZ timer nohz_cpu_mask
4857 * indicates which cpus entered this state. This is used
4858 * in the rcu update to wait only for active cpus. For system
4859 * which do not switch off the HZ timer nohz_cpu_mask should
4860 * always be CPU_MASK_NONE.
4862 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4865 * Increase the granularity value when there are more CPUs,
4866 * because with more CPUs the 'effective latency' as visible
4867 * to users decreases. But the relationship is not linear,
4868 * so pick a second-best guess by going with the log2 of the
4871 * This idea comes from the SD scheduler of Con Kolivas:
4873 static inline void sched_init_granularity(void)
4875 unsigned int factor
= 1 + ilog2(num_online_cpus());
4876 const unsigned long gran_limit
= 100000000;
4878 sysctl_sched_granularity
*= factor
;
4879 if (sysctl_sched_granularity
> gran_limit
)
4880 sysctl_sched_granularity
= gran_limit
;
4882 sysctl_sched_runtime_limit
= sysctl_sched_granularity
* 4;
4883 sysctl_sched_wakeup_granularity
= sysctl_sched_granularity
/ 2;
4888 * This is how migration works:
4890 * 1) we queue a struct migration_req structure in the source CPU's
4891 * runqueue and wake up that CPU's migration thread.
4892 * 2) we down() the locked semaphore => thread blocks.
4893 * 3) migration thread wakes up (implicitly it forces the migrated
4894 * thread off the CPU)
4895 * 4) it gets the migration request and checks whether the migrated
4896 * task is still in the wrong runqueue.
4897 * 5) if it's in the wrong runqueue then the migration thread removes
4898 * it and puts it into the right queue.
4899 * 6) migration thread up()s the semaphore.
4900 * 7) we wake up and the migration is done.
4904 * Change a given task's CPU affinity. Migrate the thread to a
4905 * proper CPU and schedule it away if the CPU it's executing on
4906 * is removed from the allowed bitmask.
4908 * NOTE: the caller must have a valid reference to the task, the
4909 * task must not exit() & deallocate itself prematurely. The
4910 * call is not atomic; no spinlocks may be held.
4912 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4914 struct migration_req req
;
4915 unsigned long flags
;
4919 rq
= task_rq_lock(p
, &flags
);
4920 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4925 p
->cpus_allowed
= new_mask
;
4926 /* Can the task run on the task's current CPU? If so, we're done */
4927 if (cpu_isset(task_cpu(p
), new_mask
))
4930 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4931 /* Need help from migration thread: drop lock and wait. */
4932 task_rq_unlock(rq
, &flags
);
4933 wake_up_process(rq
->migration_thread
);
4934 wait_for_completion(&req
.done
);
4935 tlb_migrate_finish(p
->mm
);
4939 task_rq_unlock(rq
, &flags
);
4943 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4946 * Move (not current) task off this cpu, onto dest cpu. We're doing
4947 * this because either it can't run here any more (set_cpus_allowed()
4948 * away from this CPU, or CPU going down), or because we're
4949 * attempting to rebalance this task on exec (sched_exec).
4951 * So we race with normal scheduler movements, but that's OK, as long
4952 * as the task is no longer on this CPU.
4954 * Returns non-zero if task was successfully migrated.
4956 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4958 struct rq
*rq_dest
, *rq_src
;
4961 if (unlikely(cpu_is_offline(dest_cpu
)))
4964 rq_src
= cpu_rq(src_cpu
);
4965 rq_dest
= cpu_rq(dest_cpu
);
4967 double_rq_lock(rq_src
, rq_dest
);
4968 /* Already moved. */
4969 if (task_cpu(p
) != src_cpu
)
4971 /* Affinity changed (again). */
4972 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4975 on_rq
= p
->se
.on_rq
;
4977 deactivate_task(rq_src
, p
, 0);
4979 set_task_cpu(p
, dest_cpu
);
4981 activate_task(rq_dest
, p
, 0);
4982 check_preempt_curr(rq_dest
, p
);
4986 double_rq_unlock(rq_src
, rq_dest
);
4991 * migration_thread - this is a highprio system thread that performs
4992 * thread migration by bumping thread off CPU then 'pushing' onto
4995 static int migration_thread(void *data
)
4997 int cpu
= (long)data
;
5001 BUG_ON(rq
->migration_thread
!= current
);
5003 set_current_state(TASK_INTERRUPTIBLE
);
5004 while (!kthread_should_stop()) {
5005 struct migration_req
*req
;
5006 struct list_head
*head
;
5008 spin_lock_irq(&rq
->lock
);
5010 if (cpu_is_offline(cpu
)) {
5011 spin_unlock_irq(&rq
->lock
);
5015 if (rq
->active_balance
) {
5016 active_load_balance(rq
, cpu
);
5017 rq
->active_balance
= 0;
5020 head
= &rq
->migration_queue
;
5022 if (list_empty(head
)) {
5023 spin_unlock_irq(&rq
->lock
);
5025 set_current_state(TASK_INTERRUPTIBLE
);
5028 req
= list_entry(head
->next
, struct migration_req
, list
);
5029 list_del_init(head
->next
);
5031 spin_unlock(&rq
->lock
);
5032 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5035 complete(&req
->done
);
5037 __set_current_state(TASK_RUNNING
);
5041 /* Wait for kthread_stop */
5042 set_current_state(TASK_INTERRUPTIBLE
);
5043 while (!kthread_should_stop()) {
5045 set_current_state(TASK_INTERRUPTIBLE
);
5047 __set_current_state(TASK_RUNNING
);
5051 #ifdef CONFIG_HOTPLUG_CPU
5053 * Figure out where task on dead CPU should go, use force if neccessary.
5054 * NOTE: interrupts should be disabled by the caller
5056 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5058 unsigned long flags
;
5065 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5066 cpus_and(mask
, mask
, p
->cpus_allowed
);
5067 dest_cpu
= any_online_cpu(mask
);
5069 /* On any allowed CPU? */
5070 if (dest_cpu
== NR_CPUS
)
5071 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5073 /* No more Mr. Nice Guy. */
5074 if (dest_cpu
== NR_CPUS
) {
5075 rq
= task_rq_lock(p
, &flags
);
5076 cpus_setall(p
->cpus_allowed
);
5077 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5078 task_rq_unlock(rq
, &flags
);
5081 * Don't tell them about moving exiting tasks or
5082 * kernel threads (both mm NULL), since they never
5085 if (p
->mm
&& printk_ratelimit())
5086 printk(KERN_INFO
"process %d (%s) no "
5087 "longer affine to cpu%d\n",
5088 p
->pid
, p
->comm
, dead_cpu
);
5090 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5095 * While a dead CPU has no uninterruptible tasks queued at this point,
5096 * it might still have a nonzero ->nr_uninterruptible counter, because
5097 * for performance reasons the counter is not stricly tracking tasks to
5098 * their home CPUs. So we just add the counter to another CPU's counter,
5099 * to keep the global sum constant after CPU-down:
5101 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5103 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5104 unsigned long flags
;
5106 local_irq_save(flags
);
5107 double_rq_lock(rq_src
, rq_dest
);
5108 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5109 rq_src
->nr_uninterruptible
= 0;
5110 double_rq_unlock(rq_src
, rq_dest
);
5111 local_irq_restore(flags
);
5114 /* Run through task list and migrate tasks from the dead cpu. */
5115 static void migrate_live_tasks(int src_cpu
)
5117 struct task_struct
*p
, *t
;
5119 write_lock_irq(&tasklist_lock
);
5121 do_each_thread(t
, p
) {
5125 if (task_cpu(p
) == src_cpu
)
5126 move_task_off_dead_cpu(src_cpu
, p
);
5127 } while_each_thread(t
, p
);
5129 write_unlock_irq(&tasklist_lock
);
5133 * Schedules idle task to be the next runnable task on current CPU.
5134 * It does so by boosting its priority to highest possible and adding it to
5135 * the _front_ of the runqueue. Used by CPU offline code.
5137 void sched_idle_next(void)
5139 int this_cpu
= smp_processor_id();
5140 struct rq
*rq
= cpu_rq(this_cpu
);
5141 struct task_struct
*p
= rq
->idle
;
5142 unsigned long flags
;
5144 /* cpu has to be offline */
5145 BUG_ON(cpu_online(this_cpu
));
5148 * Strictly not necessary since rest of the CPUs are stopped by now
5149 * and interrupts disabled on the current cpu.
5151 spin_lock_irqsave(&rq
->lock
, flags
);
5153 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5155 /* Add idle task to the _front_ of its priority queue: */
5156 activate_idle_task(p
, rq
);
5158 spin_unlock_irqrestore(&rq
->lock
, flags
);
5162 * Ensures that the idle task is using init_mm right before its cpu goes
5165 void idle_task_exit(void)
5167 struct mm_struct
*mm
= current
->active_mm
;
5169 BUG_ON(cpu_online(smp_processor_id()));
5172 switch_mm(mm
, &init_mm
, current
);
5176 /* called under rq->lock with disabled interrupts */
5177 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5179 struct rq
*rq
= cpu_rq(dead_cpu
);
5181 /* Must be exiting, otherwise would be on tasklist. */
5182 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5184 /* Cannot have done final schedule yet: would have vanished. */
5185 BUG_ON(p
->state
== TASK_DEAD
);
5190 * Drop lock around migration; if someone else moves it,
5191 * that's OK. No task can be added to this CPU, so iteration is
5193 * NOTE: interrupts should be left disabled --dev@
5195 spin_unlock(&rq
->lock
);
5196 move_task_off_dead_cpu(dead_cpu
, p
);
5197 spin_lock(&rq
->lock
);
5202 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5203 static void migrate_dead_tasks(unsigned int dead_cpu
)
5205 struct rq
*rq
= cpu_rq(dead_cpu
);
5206 struct task_struct
*next
;
5209 if (!rq
->nr_running
)
5211 update_rq_clock(rq
);
5212 next
= pick_next_task(rq
, rq
->curr
);
5215 migrate_dead(dead_cpu
, next
);
5219 #endif /* CONFIG_HOTPLUG_CPU */
5221 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5223 static struct ctl_table sd_ctl_dir
[] = {
5225 .procname
= "sched_domain",
5231 static struct ctl_table sd_ctl_root
[] = {
5233 .procname
= "kernel",
5235 .child
= sd_ctl_dir
,
5240 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5242 struct ctl_table
*entry
=
5243 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5246 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5252 set_table_entry(struct ctl_table
*entry
,
5253 const char *procname
, void *data
, int maxlen
,
5254 mode_t mode
, proc_handler
*proc_handler
)
5256 entry
->procname
= procname
;
5258 entry
->maxlen
= maxlen
;
5260 entry
->proc_handler
= proc_handler
;
5263 static struct ctl_table
*
5264 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5266 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5268 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5269 sizeof(long), 0644, proc_doulongvec_minmax
);
5270 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5271 sizeof(long), 0644, proc_doulongvec_minmax
);
5272 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5273 sizeof(int), 0644, proc_dointvec_minmax
);
5274 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5275 sizeof(int), 0644, proc_dointvec_minmax
);
5276 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5277 sizeof(int), 0644, proc_dointvec_minmax
);
5278 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5279 sizeof(int), 0644, proc_dointvec_minmax
);
5280 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5281 sizeof(int), 0644, proc_dointvec_minmax
);
5282 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5283 sizeof(int), 0644, proc_dointvec_minmax
);
5284 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5285 sizeof(int), 0644, proc_dointvec_minmax
);
5286 set_table_entry(&table
[10], "cache_nice_tries",
5287 &sd
->cache_nice_tries
,
5288 sizeof(int), 0644, proc_dointvec_minmax
);
5289 set_table_entry(&table
[12], "flags", &sd
->flags
,
5290 sizeof(int), 0644, proc_dointvec_minmax
);
5295 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5297 struct ctl_table
*entry
, *table
;
5298 struct sched_domain
*sd
;
5299 int domain_num
= 0, i
;
5302 for_each_domain(cpu
, sd
)
5304 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5307 for_each_domain(cpu
, sd
) {
5308 snprintf(buf
, 32, "domain%d", i
);
5309 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5311 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5318 static struct ctl_table_header
*sd_sysctl_header
;
5319 static void init_sched_domain_sysctl(void)
5321 int i
, cpu_num
= num_online_cpus();
5322 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5325 sd_ctl_dir
[0].child
= entry
;
5327 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5328 snprintf(buf
, 32, "cpu%d", i
);
5329 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5331 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5333 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5336 static void init_sched_domain_sysctl(void)
5342 * migration_call - callback that gets triggered when a CPU is added.
5343 * Here we can start up the necessary migration thread for the new CPU.
5345 static int __cpuinit
5346 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5348 struct task_struct
*p
;
5349 int cpu
= (long)hcpu
;
5350 unsigned long flags
;
5354 case CPU_LOCK_ACQUIRE
:
5355 mutex_lock(&sched_hotcpu_mutex
);
5358 case CPU_UP_PREPARE
:
5359 case CPU_UP_PREPARE_FROZEN
:
5360 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5363 kthread_bind(p
, cpu
);
5364 /* Must be high prio: stop_machine expects to yield to it. */
5365 rq
= task_rq_lock(p
, &flags
);
5366 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5367 task_rq_unlock(rq
, &flags
);
5368 cpu_rq(cpu
)->migration_thread
= p
;
5372 case CPU_ONLINE_FROZEN
:
5373 /* Strictly unneccessary, as first user will wake it. */
5374 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5377 #ifdef CONFIG_HOTPLUG_CPU
5378 case CPU_UP_CANCELED
:
5379 case CPU_UP_CANCELED_FROZEN
:
5380 if (!cpu_rq(cpu
)->migration_thread
)
5382 /* Unbind it from offline cpu so it can run. Fall thru. */
5383 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5384 any_online_cpu(cpu_online_map
));
5385 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5386 cpu_rq(cpu
)->migration_thread
= NULL
;
5390 case CPU_DEAD_FROZEN
:
5391 migrate_live_tasks(cpu
);
5393 kthread_stop(rq
->migration_thread
);
5394 rq
->migration_thread
= NULL
;
5395 /* Idle task back to normal (off runqueue, low prio) */
5396 rq
= task_rq_lock(rq
->idle
, &flags
);
5397 update_rq_clock(rq
);
5398 deactivate_task(rq
, rq
->idle
, 0);
5399 rq
->idle
->static_prio
= MAX_PRIO
;
5400 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5401 rq
->idle
->sched_class
= &idle_sched_class
;
5402 migrate_dead_tasks(cpu
);
5403 task_rq_unlock(rq
, &flags
);
5404 migrate_nr_uninterruptible(rq
);
5405 BUG_ON(rq
->nr_running
!= 0);
5407 /* No need to migrate the tasks: it was best-effort if
5408 * they didn't take sched_hotcpu_mutex. Just wake up
5409 * the requestors. */
5410 spin_lock_irq(&rq
->lock
);
5411 while (!list_empty(&rq
->migration_queue
)) {
5412 struct migration_req
*req
;
5414 req
= list_entry(rq
->migration_queue
.next
,
5415 struct migration_req
, list
);
5416 list_del_init(&req
->list
);
5417 complete(&req
->done
);
5419 spin_unlock_irq(&rq
->lock
);
5422 case CPU_LOCK_RELEASE
:
5423 mutex_unlock(&sched_hotcpu_mutex
);
5429 /* Register at highest priority so that task migration (migrate_all_tasks)
5430 * happens before everything else.
5432 static struct notifier_block __cpuinitdata migration_notifier
= {
5433 .notifier_call
= migration_call
,
5437 int __init
migration_init(void)
5439 void *cpu
= (void *)(long)smp_processor_id();
5442 /* Start one for the boot CPU: */
5443 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5444 BUG_ON(err
== NOTIFY_BAD
);
5445 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5446 register_cpu_notifier(&migration_notifier
);
5454 /* Number of possible processor ids */
5455 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5456 EXPORT_SYMBOL(nr_cpu_ids
);
5458 #undef SCHED_DOMAIN_DEBUG
5459 #ifdef SCHED_DOMAIN_DEBUG
5460 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5465 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5469 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5474 struct sched_group
*group
= sd
->groups
;
5475 cpumask_t groupmask
;
5477 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5478 cpus_clear(groupmask
);
5481 for (i
= 0; i
< level
+ 1; i
++)
5483 printk("domain %d: ", level
);
5485 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5486 printk("does not load-balance\n");
5488 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5493 printk("span %s\n", str
);
5495 if (!cpu_isset(cpu
, sd
->span
))
5496 printk(KERN_ERR
"ERROR: domain->span does not contain "
5498 if (!cpu_isset(cpu
, group
->cpumask
))
5499 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5503 for (i
= 0; i
< level
+ 2; i
++)
5509 printk(KERN_ERR
"ERROR: group is NULL\n");
5513 if (!group
->__cpu_power
) {
5515 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5519 if (!cpus_weight(group
->cpumask
)) {
5521 printk(KERN_ERR
"ERROR: empty group\n");
5524 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5526 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5529 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5531 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5534 group
= group
->next
;
5535 } while (group
!= sd
->groups
);
5538 if (!cpus_equal(sd
->span
, groupmask
))
5539 printk(KERN_ERR
"ERROR: groups don't span "
5547 if (!cpus_subset(groupmask
, sd
->span
))
5548 printk(KERN_ERR
"ERROR: parent span is not a superset "
5549 "of domain->span\n");
5554 # define sched_domain_debug(sd, cpu) do { } while (0)
5557 static int sd_degenerate(struct sched_domain
*sd
)
5559 if (cpus_weight(sd
->span
) == 1)
5562 /* Following flags need at least 2 groups */
5563 if (sd
->flags
& (SD_LOAD_BALANCE
|
5564 SD_BALANCE_NEWIDLE
|
5568 SD_SHARE_PKG_RESOURCES
)) {
5569 if (sd
->groups
!= sd
->groups
->next
)
5573 /* Following flags don't use groups */
5574 if (sd
->flags
& (SD_WAKE_IDLE
|
5583 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5585 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5587 if (sd_degenerate(parent
))
5590 if (!cpus_equal(sd
->span
, parent
->span
))
5593 /* Does parent contain flags not in child? */
5594 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5595 if (cflags
& SD_WAKE_AFFINE
)
5596 pflags
&= ~SD_WAKE_BALANCE
;
5597 /* Flags needing groups don't count if only 1 group in parent */
5598 if (parent
->groups
== parent
->groups
->next
) {
5599 pflags
&= ~(SD_LOAD_BALANCE
|
5600 SD_BALANCE_NEWIDLE
|
5604 SD_SHARE_PKG_RESOURCES
);
5606 if (~cflags
& pflags
)
5613 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5614 * hold the hotplug lock.
5616 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5618 struct rq
*rq
= cpu_rq(cpu
);
5619 struct sched_domain
*tmp
;
5621 /* Remove the sched domains which do not contribute to scheduling. */
5622 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5623 struct sched_domain
*parent
= tmp
->parent
;
5626 if (sd_parent_degenerate(tmp
, parent
)) {
5627 tmp
->parent
= parent
->parent
;
5629 parent
->parent
->child
= tmp
;
5633 if (sd
&& sd_degenerate(sd
)) {
5639 sched_domain_debug(sd
, cpu
);
5641 rcu_assign_pointer(rq
->sd
, sd
);
5644 /* cpus with isolated domains */
5645 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5647 /* Setup the mask of cpus configured for isolated domains */
5648 static int __init
isolated_cpu_setup(char *str
)
5650 int ints
[NR_CPUS
], i
;
5652 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5653 cpus_clear(cpu_isolated_map
);
5654 for (i
= 1; i
<= ints
[0]; i
++)
5655 if (ints
[i
] < NR_CPUS
)
5656 cpu_set(ints
[i
], cpu_isolated_map
);
5660 __setup ("isolcpus=", isolated_cpu_setup
);
5663 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5664 * to a function which identifies what group(along with sched group) a CPU
5665 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5666 * (due to the fact that we keep track of groups covered with a cpumask_t).
5668 * init_sched_build_groups will build a circular linked list of the groups
5669 * covered by the given span, and will set each group's ->cpumask correctly,
5670 * and ->cpu_power to 0.
5673 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5674 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5675 struct sched_group
**sg
))
5677 struct sched_group
*first
= NULL
, *last
= NULL
;
5678 cpumask_t covered
= CPU_MASK_NONE
;
5681 for_each_cpu_mask(i
, span
) {
5682 struct sched_group
*sg
;
5683 int group
= group_fn(i
, cpu_map
, &sg
);
5686 if (cpu_isset(i
, covered
))
5689 sg
->cpumask
= CPU_MASK_NONE
;
5690 sg
->__cpu_power
= 0;
5692 for_each_cpu_mask(j
, span
) {
5693 if (group_fn(j
, cpu_map
, NULL
) != group
)
5696 cpu_set(j
, covered
);
5697 cpu_set(j
, sg
->cpumask
);
5708 #define SD_NODES_PER_DOMAIN 16
5713 * find_next_best_node - find the next node to include in a sched_domain
5714 * @node: node whose sched_domain we're building
5715 * @used_nodes: nodes already in the sched_domain
5717 * Find the next node to include in a given scheduling domain. Simply
5718 * finds the closest node not already in the @used_nodes map.
5720 * Should use nodemask_t.
5722 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5724 int i
, n
, val
, min_val
, best_node
= 0;
5728 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5729 /* Start at @node */
5730 n
= (node
+ i
) % MAX_NUMNODES
;
5732 if (!nr_cpus_node(n
))
5735 /* Skip already used nodes */
5736 if (test_bit(n
, used_nodes
))
5739 /* Simple min distance search */
5740 val
= node_distance(node
, n
);
5742 if (val
< min_val
) {
5748 set_bit(best_node
, used_nodes
);
5753 * sched_domain_node_span - get a cpumask for a node's sched_domain
5754 * @node: node whose cpumask we're constructing
5755 * @size: number of nodes to include in this span
5757 * Given a node, construct a good cpumask for its sched_domain to span. It
5758 * should be one that prevents unnecessary balancing, but also spreads tasks
5761 static cpumask_t
sched_domain_node_span(int node
)
5763 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5764 cpumask_t span
, nodemask
;
5768 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5770 nodemask
= node_to_cpumask(node
);
5771 cpus_or(span
, span
, nodemask
);
5772 set_bit(node
, used_nodes
);
5774 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5775 int next_node
= find_next_best_node(node
, used_nodes
);
5777 nodemask
= node_to_cpumask(next_node
);
5778 cpus_or(span
, span
, nodemask
);
5785 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5788 * SMT sched-domains:
5790 #ifdef CONFIG_SCHED_SMT
5791 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5792 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5794 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5795 struct sched_group
**sg
)
5798 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5804 * multi-core sched-domains:
5806 #ifdef CONFIG_SCHED_MC
5807 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5808 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5811 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5812 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5813 struct sched_group
**sg
)
5816 cpumask_t mask
= cpu_sibling_map
[cpu
];
5817 cpus_and(mask
, mask
, *cpu_map
);
5818 group
= first_cpu(mask
);
5820 *sg
= &per_cpu(sched_group_core
, group
);
5823 #elif defined(CONFIG_SCHED_MC)
5824 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5825 struct sched_group
**sg
)
5828 *sg
= &per_cpu(sched_group_core
, cpu
);
5833 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5834 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5836 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5837 struct sched_group
**sg
)
5840 #ifdef CONFIG_SCHED_MC
5841 cpumask_t mask
= cpu_coregroup_map(cpu
);
5842 cpus_and(mask
, mask
, *cpu_map
);
5843 group
= first_cpu(mask
);
5844 #elif defined(CONFIG_SCHED_SMT)
5845 cpumask_t mask
= cpu_sibling_map
[cpu
];
5846 cpus_and(mask
, mask
, *cpu_map
);
5847 group
= first_cpu(mask
);
5852 *sg
= &per_cpu(sched_group_phys
, group
);
5858 * The init_sched_build_groups can't handle what we want to do with node
5859 * groups, so roll our own. Now each node has its own list of groups which
5860 * gets dynamically allocated.
5862 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5863 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5865 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5866 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5868 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5869 struct sched_group
**sg
)
5871 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5874 cpus_and(nodemask
, nodemask
, *cpu_map
);
5875 group
= first_cpu(nodemask
);
5878 *sg
= &per_cpu(sched_group_allnodes
, group
);
5882 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5884 struct sched_group
*sg
= group_head
;
5890 for_each_cpu_mask(j
, sg
->cpumask
) {
5891 struct sched_domain
*sd
;
5893 sd
= &per_cpu(phys_domains
, j
);
5894 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5896 * Only add "power" once for each
5902 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5905 if (sg
!= group_head
)
5911 /* Free memory allocated for various sched_group structures */
5912 static void free_sched_groups(const cpumask_t
*cpu_map
)
5916 for_each_cpu_mask(cpu
, *cpu_map
) {
5917 struct sched_group
**sched_group_nodes
5918 = sched_group_nodes_bycpu
[cpu
];
5920 if (!sched_group_nodes
)
5923 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5924 cpumask_t nodemask
= node_to_cpumask(i
);
5925 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5927 cpus_and(nodemask
, nodemask
, *cpu_map
);
5928 if (cpus_empty(nodemask
))
5938 if (oldsg
!= sched_group_nodes
[i
])
5941 kfree(sched_group_nodes
);
5942 sched_group_nodes_bycpu
[cpu
] = NULL
;
5946 static void free_sched_groups(const cpumask_t
*cpu_map
)
5952 * Initialize sched groups cpu_power.
5954 * cpu_power indicates the capacity of sched group, which is used while
5955 * distributing the load between different sched groups in a sched domain.
5956 * Typically cpu_power for all the groups in a sched domain will be same unless
5957 * there are asymmetries in the topology. If there are asymmetries, group
5958 * having more cpu_power will pickup more load compared to the group having
5961 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5962 * the maximum number of tasks a group can handle in the presence of other idle
5963 * or lightly loaded groups in the same sched domain.
5965 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5967 struct sched_domain
*child
;
5968 struct sched_group
*group
;
5970 WARN_ON(!sd
|| !sd
->groups
);
5972 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5977 sd
->groups
->__cpu_power
= 0;
5980 * For perf policy, if the groups in child domain share resources
5981 * (for example cores sharing some portions of the cache hierarchy
5982 * or SMT), then set this domain groups cpu_power such that each group
5983 * can handle only one task, when there are other idle groups in the
5984 * same sched domain.
5986 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5988 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5989 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5994 * add cpu_power of each child group to this groups cpu_power
5996 group
= child
->groups
;
5998 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5999 group
= group
->next
;
6000 } while (group
!= child
->groups
);
6004 * Build sched domains for a given set of cpus and attach the sched domains
6005 * to the individual cpus
6007 static int build_sched_domains(const cpumask_t
*cpu_map
)
6011 struct sched_group
**sched_group_nodes
= NULL
;
6012 int sd_allnodes
= 0;
6015 * Allocate the per-node list of sched groups
6017 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6019 if (!sched_group_nodes
) {
6020 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6023 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6027 * Set up domains for cpus specified by the cpu_map.
6029 for_each_cpu_mask(i
, *cpu_map
) {
6030 struct sched_domain
*sd
= NULL
, *p
;
6031 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6033 cpus_and(nodemask
, nodemask
, *cpu_map
);
6036 if (cpus_weight(*cpu_map
) >
6037 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6038 sd
= &per_cpu(allnodes_domains
, i
);
6039 *sd
= SD_ALLNODES_INIT
;
6040 sd
->span
= *cpu_map
;
6041 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6047 sd
= &per_cpu(node_domains
, i
);
6049 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6053 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6057 sd
= &per_cpu(phys_domains
, i
);
6059 sd
->span
= nodemask
;
6063 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6065 #ifdef CONFIG_SCHED_MC
6067 sd
= &per_cpu(core_domains
, i
);
6069 sd
->span
= cpu_coregroup_map(i
);
6070 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6073 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6076 #ifdef CONFIG_SCHED_SMT
6078 sd
= &per_cpu(cpu_domains
, i
);
6079 *sd
= SD_SIBLING_INIT
;
6080 sd
->span
= cpu_sibling_map
[i
];
6081 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6084 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6088 #ifdef CONFIG_SCHED_SMT
6089 /* Set up CPU (sibling) groups */
6090 for_each_cpu_mask(i
, *cpu_map
) {
6091 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6092 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6093 if (i
!= first_cpu(this_sibling_map
))
6096 init_sched_build_groups(this_sibling_map
, cpu_map
,
6101 #ifdef CONFIG_SCHED_MC
6102 /* Set up multi-core groups */
6103 for_each_cpu_mask(i
, *cpu_map
) {
6104 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6105 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6106 if (i
!= first_cpu(this_core_map
))
6108 init_sched_build_groups(this_core_map
, cpu_map
,
6109 &cpu_to_core_group
);
6113 /* Set up physical groups */
6114 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6115 cpumask_t nodemask
= node_to_cpumask(i
);
6117 cpus_and(nodemask
, nodemask
, *cpu_map
);
6118 if (cpus_empty(nodemask
))
6121 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6125 /* Set up node groups */
6127 init_sched_build_groups(*cpu_map
, cpu_map
,
6128 &cpu_to_allnodes_group
);
6130 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6131 /* Set up node groups */
6132 struct sched_group
*sg
, *prev
;
6133 cpumask_t nodemask
= node_to_cpumask(i
);
6134 cpumask_t domainspan
;
6135 cpumask_t covered
= CPU_MASK_NONE
;
6138 cpus_and(nodemask
, nodemask
, *cpu_map
);
6139 if (cpus_empty(nodemask
)) {
6140 sched_group_nodes
[i
] = NULL
;
6144 domainspan
= sched_domain_node_span(i
);
6145 cpus_and(domainspan
, domainspan
, *cpu_map
);
6147 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6149 printk(KERN_WARNING
"Can not alloc domain group for "
6153 sched_group_nodes
[i
] = sg
;
6154 for_each_cpu_mask(j
, nodemask
) {
6155 struct sched_domain
*sd
;
6157 sd
= &per_cpu(node_domains
, j
);
6160 sg
->__cpu_power
= 0;
6161 sg
->cpumask
= nodemask
;
6163 cpus_or(covered
, covered
, nodemask
);
6166 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6167 cpumask_t tmp
, notcovered
;
6168 int n
= (i
+ j
) % MAX_NUMNODES
;
6170 cpus_complement(notcovered
, covered
);
6171 cpus_and(tmp
, notcovered
, *cpu_map
);
6172 cpus_and(tmp
, tmp
, domainspan
);
6173 if (cpus_empty(tmp
))
6176 nodemask
= node_to_cpumask(n
);
6177 cpus_and(tmp
, tmp
, nodemask
);
6178 if (cpus_empty(tmp
))
6181 sg
= kmalloc_node(sizeof(struct sched_group
),
6185 "Can not alloc domain group for node %d\n", j
);
6188 sg
->__cpu_power
= 0;
6190 sg
->next
= prev
->next
;
6191 cpus_or(covered
, covered
, tmp
);
6198 /* Calculate CPU power for physical packages and nodes */
6199 #ifdef CONFIG_SCHED_SMT
6200 for_each_cpu_mask(i
, *cpu_map
) {
6201 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6203 init_sched_groups_power(i
, sd
);
6206 #ifdef CONFIG_SCHED_MC
6207 for_each_cpu_mask(i
, *cpu_map
) {
6208 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6210 init_sched_groups_power(i
, sd
);
6214 for_each_cpu_mask(i
, *cpu_map
) {
6215 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6217 init_sched_groups_power(i
, sd
);
6221 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6222 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6225 struct sched_group
*sg
;
6227 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6228 init_numa_sched_groups_power(sg
);
6232 /* Attach the domains */
6233 for_each_cpu_mask(i
, *cpu_map
) {
6234 struct sched_domain
*sd
;
6235 #ifdef CONFIG_SCHED_SMT
6236 sd
= &per_cpu(cpu_domains
, i
);
6237 #elif defined(CONFIG_SCHED_MC)
6238 sd
= &per_cpu(core_domains
, i
);
6240 sd
= &per_cpu(phys_domains
, i
);
6242 cpu_attach_domain(sd
, i
);
6249 free_sched_groups(cpu_map
);
6254 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6256 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6258 cpumask_t cpu_default_map
;
6262 * Setup mask for cpus without special case scheduling requirements.
6263 * For now this just excludes isolated cpus, but could be used to
6264 * exclude other special cases in the future.
6266 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6268 err
= build_sched_domains(&cpu_default_map
);
6273 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6275 free_sched_groups(cpu_map
);
6279 * Detach sched domains from a group of cpus specified in cpu_map
6280 * These cpus will now be attached to the NULL domain
6282 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6286 for_each_cpu_mask(i
, *cpu_map
)
6287 cpu_attach_domain(NULL
, i
);
6288 synchronize_sched();
6289 arch_destroy_sched_domains(cpu_map
);
6293 * Partition sched domains as specified by the cpumasks below.
6294 * This attaches all cpus from the cpumasks to the NULL domain,
6295 * waits for a RCU quiescent period, recalculates sched
6296 * domain information and then attaches them back to the
6297 * correct sched domains
6298 * Call with hotplug lock held
6300 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6302 cpumask_t change_map
;
6305 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6306 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6307 cpus_or(change_map
, *partition1
, *partition2
);
6309 /* Detach sched domains from all of the affected cpus */
6310 detach_destroy_domains(&change_map
);
6311 if (!cpus_empty(*partition1
))
6312 err
= build_sched_domains(partition1
);
6313 if (!err
&& !cpus_empty(*partition2
))
6314 err
= build_sched_domains(partition2
);
6319 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6320 int arch_reinit_sched_domains(void)
6324 mutex_lock(&sched_hotcpu_mutex
);
6325 detach_destroy_domains(&cpu_online_map
);
6326 err
= arch_init_sched_domains(&cpu_online_map
);
6327 mutex_unlock(&sched_hotcpu_mutex
);
6332 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6336 if (buf
[0] != '0' && buf
[0] != '1')
6340 sched_smt_power_savings
= (buf
[0] == '1');
6342 sched_mc_power_savings
= (buf
[0] == '1');
6344 ret
= arch_reinit_sched_domains();
6346 return ret
? ret
: count
;
6349 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6353 #ifdef CONFIG_SCHED_SMT
6355 err
= sysfs_create_file(&cls
->kset
.kobj
,
6356 &attr_sched_smt_power_savings
.attr
);
6358 #ifdef CONFIG_SCHED_MC
6359 if (!err
&& mc_capable())
6360 err
= sysfs_create_file(&cls
->kset
.kobj
,
6361 &attr_sched_mc_power_savings
.attr
);
6367 #ifdef CONFIG_SCHED_MC
6368 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6370 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6372 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6373 const char *buf
, size_t count
)
6375 return sched_power_savings_store(buf
, count
, 0);
6377 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6378 sched_mc_power_savings_store
);
6381 #ifdef CONFIG_SCHED_SMT
6382 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6384 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6386 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6387 const char *buf
, size_t count
)
6389 return sched_power_savings_store(buf
, count
, 1);
6391 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6392 sched_smt_power_savings_store
);
6396 * Force a reinitialization of the sched domains hierarchy. The domains
6397 * and groups cannot be updated in place without racing with the balancing
6398 * code, so we temporarily attach all running cpus to the NULL domain
6399 * which will prevent rebalancing while the sched domains are recalculated.
6401 static int update_sched_domains(struct notifier_block
*nfb
,
6402 unsigned long action
, void *hcpu
)
6405 case CPU_UP_PREPARE
:
6406 case CPU_UP_PREPARE_FROZEN
:
6407 case CPU_DOWN_PREPARE
:
6408 case CPU_DOWN_PREPARE_FROZEN
:
6409 detach_destroy_domains(&cpu_online_map
);
6412 case CPU_UP_CANCELED
:
6413 case CPU_UP_CANCELED_FROZEN
:
6414 case CPU_DOWN_FAILED
:
6415 case CPU_DOWN_FAILED_FROZEN
:
6417 case CPU_ONLINE_FROZEN
:
6419 case CPU_DEAD_FROZEN
:
6421 * Fall through and re-initialise the domains.
6428 /* The hotplug lock is already held by cpu_up/cpu_down */
6429 arch_init_sched_domains(&cpu_online_map
);
6434 void __init
sched_init_smp(void)
6436 cpumask_t non_isolated_cpus
;
6438 mutex_lock(&sched_hotcpu_mutex
);
6439 arch_init_sched_domains(&cpu_online_map
);
6440 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6441 if (cpus_empty(non_isolated_cpus
))
6442 cpu_set(smp_processor_id(), non_isolated_cpus
);
6443 mutex_unlock(&sched_hotcpu_mutex
);
6444 /* XXX: Theoretical race here - CPU may be hotplugged now */
6445 hotcpu_notifier(update_sched_domains
, 0);
6447 init_sched_domain_sysctl();
6449 /* Move init over to a non-isolated CPU */
6450 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6452 sched_init_granularity();
6455 void __init
sched_init_smp(void)
6457 sched_init_granularity();
6459 #endif /* CONFIG_SMP */
6461 int in_sched_functions(unsigned long addr
)
6463 /* Linker adds these: start and end of __sched functions */
6464 extern char __sched_text_start
[], __sched_text_end
[];
6466 return in_lock_functions(addr
) ||
6467 (addr
>= (unsigned long)__sched_text_start
6468 && addr
< (unsigned long)__sched_text_end
);
6471 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6473 cfs_rq
->tasks_timeline
= RB_ROOT
;
6474 cfs_rq
->fair_clock
= 1;
6475 #ifdef CONFIG_FAIR_GROUP_SCHED
6480 void __init
sched_init(void)
6482 u64 now
= sched_clock();
6483 int highest_cpu
= 0;
6487 * Link up the scheduling class hierarchy:
6489 rt_sched_class
.next
= &fair_sched_class
;
6490 fair_sched_class
.next
= &idle_sched_class
;
6491 idle_sched_class
.next
= NULL
;
6493 for_each_possible_cpu(i
) {
6494 struct rt_prio_array
*array
;
6498 spin_lock_init(&rq
->lock
);
6499 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6502 init_cfs_rq(&rq
->cfs
, rq
);
6503 #ifdef CONFIG_FAIR_GROUP_SCHED
6504 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6505 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6507 rq
->ls
.load_update_last
= now
;
6508 rq
->ls
.load_update_start
= now
;
6510 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6511 rq
->cpu_load
[j
] = 0;
6514 rq
->active_balance
= 0;
6515 rq
->next_balance
= jiffies
;
6518 rq
->migration_thread
= NULL
;
6519 INIT_LIST_HEAD(&rq
->migration_queue
);
6521 atomic_set(&rq
->nr_iowait
, 0);
6523 array
= &rq
->rt
.active
;
6524 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6525 INIT_LIST_HEAD(array
->queue
+ j
);
6526 __clear_bit(j
, array
->bitmap
);
6529 /* delimiter for bitsearch: */
6530 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6533 set_load_weight(&init_task
);
6535 #ifdef CONFIG_PREEMPT_NOTIFIERS
6536 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6540 nr_cpu_ids
= highest_cpu
+ 1;
6541 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6544 #ifdef CONFIG_RT_MUTEXES
6545 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6549 * The boot idle thread does lazy MMU switching as well:
6551 atomic_inc(&init_mm
.mm_count
);
6552 enter_lazy_tlb(&init_mm
, current
);
6555 * Make us the idle thread. Technically, schedule() should not be
6556 * called from this thread, however somewhere below it might be,
6557 * but because we are the idle thread, we just pick up running again
6558 * when this runqueue becomes "idle".
6560 init_idle(current
, smp_processor_id());
6562 * During early bootup we pretend to be a normal task:
6564 current
->sched_class
= &fair_sched_class
;
6567 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6568 void __might_sleep(char *file
, int line
)
6571 static unsigned long prev_jiffy
; /* ratelimiting */
6573 if ((in_atomic() || irqs_disabled()) &&
6574 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6575 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6577 prev_jiffy
= jiffies
;
6578 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6579 " context at %s:%d\n", file
, line
);
6580 printk("in_atomic():%d, irqs_disabled():%d\n",
6581 in_atomic(), irqs_disabled());
6582 debug_show_held_locks(current
);
6583 if (irqs_disabled())
6584 print_irqtrace_events(current
);
6589 EXPORT_SYMBOL(__might_sleep
);
6592 #ifdef CONFIG_MAGIC_SYSRQ
6593 void normalize_rt_tasks(void)
6595 struct task_struct
*g
, *p
;
6596 unsigned long flags
;
6600 read_lock_irq(&tasklist_lock
);
6601 do_each_thread(g
, p
) {
6603 p
->se
.wait_runtime
= 0;
6604 p
->se
.exec_start
= 0;
6605 p
->se
.wait_start_fair
= 0;
6606 p
->se
.sleep_start_fair
= 0;
6607 #ifdef CONFIG_SCHEDSTATS
6608 p
->se
.wait_start
= 0;
6609 p
->se
.sleep_start
= 0;
6610 p
->se
.block_start
= 0;
6612 task_rq(p
)->cfs
.fair_clock
= 0;
6613 task_rq(p
)->clock
= 0;
6617 * Renice negative nice level userspace
6620 if (TASK_NICE(p
) < 0 && p
->mm
)
6621 set_user_nice(p
, 0);
6625 spin_lock_irqsave(&p
->pi_lock
, flags
);
6626 rq
= __task_rq_lock(p
);
6629 * Do not touch the migration thread:
6631 if (p
== rq
->migration_thread
)
6635 update_rq_clock(rq
);
6636 on_rq
= p
->se
.on_rq
;
6638 deactivate_task(rq
, p
, 0);
6639 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6641 activate_task(rq
, p
, 0);
6642 resched_task(rq
->curr
);
6647 __task_rq_unlock(rq
);
6648 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6649 } while_each_thread(g
, p
);
6651 read_unlock_irq(&tasklist_lock
);
6654 #endif /* CONFIG_MAGIC_SYSRQ */
6658 * These functions are only useful for the IA64 MCA handling.
6660 * They can only be called when the whole system has been
6661 * stopped - every CPU needs to be quiescent, and no scheduling
6662 * activity can take place. Using them for anything else would
6663 * be a serious bug, and as a result, they aren't even visible
6664 * under any other configuration.
6668 * curr_task - return the current task for a given cpu.
6669 * @cpu: the processor in question.
6671 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6673 struct task_struct
*curr_task(int cpu
)
6675 return cpu_curr(cpu
);
6679 * set_curr_task - set the current task for a given cpu.
6680 * @cpu: the processor in question.
6681 * @p: the task pointer to set.
6683 * Description: This function must only be used when non-maskable interrupts
6684 * are serviced on a separate stack. It allows the architecture to switch the
6685 * notion of the current task on a cpu in a non-blocking manner. This function
6686 * must be called with all CPU's synchronized, and interrupts disabled, the
6687 * and caller must save the original value of the current task (see
6688 * curr_task() above) and restore that value before reenabling interrupts and
6689 * re-starting the system.
6691 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6693 void set_curr_task(int cpu
, struct task_struct
*p
)