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 * Per-runqueue clock, as finegrained as the platform can give us:
323 static unsigned long long __rq_clock(struct rq
*rq
)
325 u64 prev_raw
= rq
->prev_clock_raw
;
326 u64 now
= sched_clock();
327 s64 delta
= now
- prev_raw
;
328 u64 clock
= rq
->clock
;
331 * Protect against sched_clock() occasionally going backwards:
333 if (unlikely(delta
< 0)) {
338 * Catch too large forward jumps too:
340 if (unlikely(delta
> 2*TICK_NSEC
)) {
342 rq
->clock_overflows
++;
344 if (unlikely(delta
> rq
->clock_max_delta
))
345 rq
->clock_max_delta
= delta
;
350 rq
->prev_clock_raw
= now
;
356 static inline unsigned long long rq_clock(struct rq
*rq
)
358 int this_cpu
= smp_processor_id();
360 if (this_cpu
== cpu_of(rq
))
361 return __rq_clock(rq
);
367 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
368 * See detach_destroy_domains: synchronize_sched for details.
370 * The domain tree of any CPU may only be accessed from within
371 * preempt-disabled sections.
373 #define for_each_domain(cpu, __sd) \
374 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
376 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
377 #define this_rq() (&__get_cpu_var(runqueues))
378 #define task_rq(p) cpu_rq(task_cpu(p))
379 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
382 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
383 * clock constructed from sched_clock():
385 unsigned long long cpu_clock(int cpu
)
387 unsigned long long now
;
390 local_irq_save(flags
);
391 now
= rq_clock(cpu_rq(cpu
));
392 local_irq_restore(flags
);
397 #ifdef CONFIG_FAIR_GROUP_SCHED
398 /* Change a task's ->cfs_rq if it moves across CPUs */
399 static inline void set_task_cfs_rq(struct task_struct
*p
)
401 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
404 static inline void set_task_cfs_rq(struct task_struct
*p
)
409 #ifndef prepare_arch_switch
410 # define prepare_arch_switch(next) do { } while (0)
412 #ifndef finish_arch_switch
413 # define finish_arch_switch(prev) do { } while (0)
416 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
417 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
419 return rq
->curr
== p
;
422 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
426 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
428 #ifdef CONFIG_DEBUG_SPINLOCK
429 /* this is a valid case when another task releases the spinlock */
430 rq
->lock
.owner
= current
;
433 * If we are tracking spinlock dependencies then we have to
434 * fix up the runqueue lock - which gets 'carried over' from
437 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
439 spin_unlock_irq(&rq
->lock
);
442 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
443 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
448 return rq
->curr
== p
;
452 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
456 * We can optimise this out completely for !SMP, because the
457 * SMP rebalancing from interrupt is the only thing that cares
462 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
463 spin_unlock_irq(&rq
->lock
);
465 spin_unlock(&rq
->lock
);
469 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
473 * After ->oncpu is cleared, the task can be moved to a different CPU.
474 * We must ensure this doesn't happen until the switch is completely
480 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
484 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
487 * __task_rq_lock - lock the runqueue a given task resides on.
488 * Must be called interrupts disabled.
490 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
497 spin_lock(&rq
->lock
);
498 if (unlikely(rq
!= task_rq(p
))) {
499 spin_unlock(&rq
->lock
);
500 goto repeat_lock_task
;
506 * task_rq_lock - lock the runqueue a given task resides on and disable
507 * interrupts. Note the ordering: we can safely lookup the task_rq without
508 * explicitly disabling preemption.
510 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
516 local_irq_save(*flags
);
518 spin_lock(&rq
->lock
);
519 if (unlikely(rq
!= task_rq(p
))) {
520 spin_unlock_irqrestore(&rq
->lock
, *flags
);
521 goto repeat_lock_task
;
526 static inline void __task_rq_unlock(struct rq
*rq
)
529 spin_unlock(&rq
->lock
);
532 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
535 spin_unlock_irqrestore(&rq
->lock
, *flags
);
539 * this_rq_lock - lock this runqueue and disable interrupts.
541 static inline struct rq
*this_rq_lock(void)
548 spin_lock(&rq
->lock
);
554 * CPU frequency is/was unstable - start new by setting prev_clock_raw:
556 void sched_clock_unstable_event(void)
561 rq
= task_rq_lock(current
, &flags
);
562 rq
->prev_clock_raw
= sched_clock();
563 rq
->clock_unstable_events
++;
564 task_rq_unlock(rq
, &flags
);
568 * resched_task - mark a task 'to be rescheduled now'.
570 * On UP this means the setting of the need_resched flag, on SMP it
571 * might also involve a cross-CPU call to trigger the scheduler on
576 #ifndef tsk_is_polling
577 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
580 static void resched_task(struct task_struct
*p
)
584 assert_spin_locked(&task_rq(p
)->lock
);
586 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
589 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
592 if (cpu
== smp_processor_id())
595 /* NEED_RESCHED must be visible before we test polling */
597 if (!tsk_is_polling(p
))
598 smp_send_reschedule(cpu
);
601 static void resched_cpu(int cpu
)
603 struct rq
*rq
= cpu_rq(cpu
);
606 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
608 resched_task(cpu_curr(cpu
));
609 spin_unlock_irqrestore(&rq
->lock
, flags
);
612 static inline void resched_task(struct task_struct
*p
)
614 assert_spin_locked(&task_rq(p
)->lock
);
615 set_tsk_need_resched(p
);
619 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
621 #if BITS_PER_LONG == 32
622 if (likely(divident
<= 0xffffffffULL
))
623 return (u32
)divident
/ divisor
;
624 do_div(divident
, divisor
);
628 return divident
/ divisor
;
632 #if BITS_PER_LONG == 32
633 # define WMULT_CONST (~0UL)
635 # define WMULT_CONST (1UL << 32)
638 #define WMULT_SHIFT 32
641 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
642 struct load_weight
*lw
)
646 if (unlikely(!lw
->inv_weight
))
647 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
649 tmp
= (u64
)delta_exec
* weight
;
651 * Check whether we'd overflow the 64-bit multiplication:
653 if (unlikely(tmp
> WMULT_CONST
)) {
654 tmp
= ((tmp
>> WMULT_SHIFT
/2) * lw
->inv_weight
)
657 tmp
= (tmp
* lw
->inv_weight
) >> WMULT_SHIFT
;
660 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
663 static inline unsigned long
664 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
666 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
669 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
675 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
682 * To aid in avoiding the subversion of "niceness" due to uneven distribution
683 * of tasks with abnormal "nice" values across CPUs the contribution that
684 * each task makes to its run queue's load is weighted according to its
685 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
686 * scaled version of the new time slice allocation that they receive on time
690 #define WEIGHT_IDLEPRIO 2
691 #define WMULT_IDLEPRIO (1 << 31)
694 * Nice levels are multiplicative, with a gentle 10% change for every
695 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
696 * nice 1, it will get ~10% less CPU time than another CPU-bound task
697 * that remained on nice 0.
699 * The "10% effect" is relative and cumulative: from _any_ nice level,
700 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
701 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
702 * If a task goes up by ~10% and another task goes down by ~10% then
703 * the relative distance between them is ~25%.)
705 static const int prio_to_weight
[40] = {
706 /* -20 */ 88818, 71054, 56843, 45475, 36380, 29104, 23283, 18626, 14901, 11921,
707 /* -10 */ 9537, 7629, 6103, 4883, 3906, 3125, 2500, 2000, 1600, 1280,
708 /* 0 */ NICE_0_LOAD
/* 1024 */,
709 /* 1 */ 819, 655, 524, 419, 336, 268, 215, 172, 137,
710 /* 10 */ 110, 87, 70, 56, 45, 36, 29, 23, 18, 15,
714 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
716 * In cases where the weight does not change often, we can use the
717 * precalculated inverse to speed up arithmetics by turning divisions
718 * into multiplications:
720 static const u32 prio_to_wmult
[40] = {
721 /* -20 */ 48356, 60446, 75558, 94446, 118058,
722 /* -15 */ 147573, 184467, 230589, 288233, 360285,
723 /* -10 */ 450347, 562979, 703746, 879575, 1099582,
724 /* -5 */ 1374389, 1717986, 2147483, 2684354, 3355443,
725 /* 0 */ 4194304, 5244160, 6557201, 8196502, 10250518,
726 /* 5 */ 12782640, 16025997, 19976592, 24970740, 31350126,
727 /* 10 */ 39045157, 49367440, 61356675, 76695844, 95443717,
728 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
731 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
734 * runqueue iterator, to support SMP load-balancing between different
735 * scheduling classes, without having to expose their internal data
736 * structures to the load-balancing proper:
740 struct task_struct
*(*start
)(void *);
741 struct task_struct
*(*next
)(void *);
744 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
745 unsigned long max_nr_move
, unsigned long max_load_move
,
746 struct sched_domain
*sd
, enum cpu_idle_type idle
,
747 int *all_pinned
, unsigned long *load_moved
,
748 int this_best_prio
, int best_prio
, int best_prio_seen
,
749 struct rq_iterator
*iterator
);
751 #include "sched_stats.h"
752 #include "sched_rt.c"
753 #include "sched_fair.c"
754 #include "sched_idletask.c"
755 #ifdef CONFIG_SCHED_DEBUG
756 # include "sched_debug.c"
759 #define sched_class_highest (&rt_sched_class)
761 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
763 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
764 ls
->delta_exec
+= ls
->delta_stat
;
765 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
771 * Update delta_exec, delta_fair fields for rq.
773 * delta_fair clock advances at a rate inversely proportional to
774 * total load (rq->ls.load.weight) on the runqueue, while
775 * delta_exec advances at the same rate as wall-clock (provided
778 * delta_exec / delta_fair is a measure of the (smoothened) load on this
779 * runqueue over any given interval. This (smoothened) load is used
780 * during load balance.
782 * This function is called /before/ updating rq->ls.load
783 * and when switching tasks.
785 static void update_curr_load(struct rq
*rq
, u64 now
)
787 struct load_stat
*ls
= &rq
->ls
;
790 start
= ls
->load_update_start
;
791 ls
->load_update_start
= now
;
792 ls
->delta_stat
+= now
- start
;
794 * Stagger updates to ls->delta_fair. Very frequent updates
797 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
798 __update_curr_load(rq
, ls
);
802 inc_load(struct rq
*rq
, const struct task_struct
*p
, u64 now
)
804 update_curr_load(rq
, now
);
805 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
809 dec_load(struct rq
*rq
, const struct task_struct
*p
, u64 now
)
811 update_curr_load(rq
, now
);
812 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
815 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
, u64 now
)
818 inc_load(rq
, p
, now
);
821 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
, u64 now
)
824 dec_load(rq
, p
, now
);
827 static void set_load_weight(struct task_struct
*p
)
829 task_rq(p
)->cfs
.wait_runtime
-= p
->se
.wait_runtime
;
830 p
->se
.wait_runtime
= 0;
832 if (task_has_rt_policy(p
)) {
833 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
834 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
839 * SCHED_IDLE tasks get minimal weight:
841 if (p
->policy
== SCHED_IDLE
) {
842 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
843 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
847 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
848 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
852 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, u64 now
)
854 sched_info_queued(p
);
855 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, now
);
860 dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
862 p
->sched_class
->dequeue_task(rq
, p
, sleep
, now
);
867 * __normal_prio - return the priority that is based on the static prio
869 static inline int __normal_prio(struct task_struct
*p
)
871 return p
->static_prio
;
875 * Calculate the expected normal priority: i.e. priority
876 * without taking RT-inheritance into account. Might be
877 * boosted by interactivity modifiers. Changes upon fork,
878 * setprio syscalls, and whenever the interactivity
879 * estimator recalculates.
881 static inline int normal_prio(struct task_struct
*p
)
885 if (task_has_rt_policy(p
))
886 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
888 prio
= __normal_prio(p
);
893 * Calculate the current priority, i.e. the priority
894 * taken into account by the scheduler. This value might
895 * be boosted by RT tasks, or might be boosted by
896 * interactivity modifiers. Will be RT if the task got
897 * RT-boosted. If not then it returns p->normal_prio.
899 static int effective_prio(struct task_struct
*p
)
901 p
->normal_prio
= normal_prio(p
);
903 * If we are RT tasks or we were boosted to RT priority,
904 * keep the priority unchanged. Otherwise, update priority
905 * to the normal priority:
907 if (!rt_prio(p
->prio
))
908 return p
->normal_prio
;
913 * activate_task - move a task to the runqueue.
915 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
917 u64 now
= rq_clock(rq
);
919 if (p
->state
== TASK_UNINTERRUPTIBLE
)
920 rq
->nr_uninterruptible
--;
922 enqueue_task(rq
, p
, wakeup
, now
);
923 inc_nr_running(p
, rq
, now
);
927 * activate_idle_task - move idle task to the _front_ of runqueue.
929 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
931 u64 now
= rq_clock(rq
);
933 if (p
->state
== TASK_UNINTERRUPTIBLE
)
934 rq
->nr_uninterruptible
--;
936 enqueue_task(rq
, p
, 0, now
);
937 inc_nr_running(p
, rq
, now
);
941 * deactivate_task - remove a task from the runqueue.
943 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
945 u64 now
= rq_clock(rq
);
947 if (p
->state
== TASK_UNINTERRUPTIBLE
)
948 rq
->nr_uninterruptible
++;
950 dequeue_task(rq
, p
, sleep
, now
);
951 dec_nr_running(p
, rq
, now
);
955 * task_curr - is this task currently executing on a CPU?
956 * @p: the task in question.
958 inline int task_curr(const struct task_struct
*p
)
960 return cpu_curr(task_cpu(p
)) == p
;
963 /* Used instead of source_load when we know the type == 0 */
964 unsigned long weighted_cpuload(const int cpu
)
966 return cpu_rq(cpu
)->ls
.load
.weight
;
969 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
972 task_thread_info(p
)->cpu
= cpu
;
979 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
981 int old_cpu
= task_cpu(p
);
982 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
983 u64 clock_offset
, fair_clock_offset
;
985 clock_offset
= old_rq
->clock
- new_rq
->clock
;
986 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
988 if (p
->se
.wait_start_fair
)
989 p
->se
.wait_start_fair
-= fair_clock_offset
;
990 if (p
->se
.sleep_start_fair
)
991 p
->se
.sleep_start_fair
-= fair_clock_offset
;
993 #ifdef CONFIG_SCHEDSTATS
994 if (p
->se
.wait_start
)
995 p
->se
.wait_start
-= clock_offset
;
996 if (p
->se
.sleep_start
)
997 p
->se
.sleep_start
-= clock_offset
;
998 if (p
->se
.block_start
)
999 p
->se
.block_start
-= clock_offset
;
1002 __set_task_cpu(p
, new_cpu
);
1005 struct migration_req
{
1006 struct list_head list
;
1008 struct task_struct
*task
;
1011 struct completion done
;
1015 * The task's runqueue lock must be held.
1016 * Returns true if you have to wait for migration thread.
1019 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1021 struct rq
*rq
= task_rq(p
);
1024 * If the task is not on a runqueue (and not running), then
1025 * it is sufficient to simply update the task's cpu field.
1027 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1028 set_task_cpu(p
, dest_cpu
);
1032 init_completion(&req
->done
);
1034 req
->dest_cpu
= dest_cpu
;
1035 list_add(&req
->list
, &rq
->migration_queue
);
1041 * wait_task_inactive - wait for a thread to unschedule.
1043 * The caller must ensure that the task *will* unschedule sometime soon,
1044 * else this function might spin for a *long* time. This function can't
1045 * be called with interrupts off, or it may introduce deadlock with
1046 * smp_call_function() if an IPI is sent by the same process we are
1047 * waiting to become inactive.
1049 void wait_task_inactive(struct task_struct
*p
)
1051 unsigned long flags
;
1057 * We do the initial early heuristics without holding
1058 * any task-queue locks at all. We'll only try to get
1059 * the runqueue lock when things look like they will
1065 * If the task is actively running on another CPU
1066 * still, just relax and busy-wait without holding
1069 * NOTE! Since we don't hold any locks, it's not
1070 * even sure that "rq" stays as the right runqueue!
1071 * But we don't care, since "task_running()" will
1072 * return false if the runqueue has changed and p
1073 * is actually now running somewhere else!
1075 while (task_running(rq
, p
))
1079 * Ok, time to look more closely! We need the rq
1080 * lock now, to be *sure*. If we're wrong, we'll
1081 * just go back and repeat.
1083 rq
= task_rq_lock(p
, &flags
);
1084 running
= task_running(rq
, p
);
1085 on_rq
= p
->se
.on_rq
;
1086 task_rq_unlock(rq
, &flags
);
1089 * Was it really running after all now that we
1090 * checked with the proper locks actually held?
1092 * Oops. Go back and try again..
1094 if (unlikely(running
)) {
1100 * It's not enough that it's not actively running,
1101 * it must be off the runqueue _entirely_, and not
1104 * So if it wa still runnable (but just not actively
1105 * running right now), it's preempted, and we should
1106 * yield - it could be a while.
1108 if (unlikely(on_rq
)) {
1114 * Ahh, all good. It wasn't running, and it wasn't
1115 * runnable, which means that it will never become
1116 * running in the future either. We're all done!
1121 * kick_process - kick a running thread to enter/exit the kernel
1122 * @p: the to-be-kicked thread
1124 * Cause a process which is running on another CPU to enter
1125 * kernel-mode, without any delay. (to get signals handled.)
1127 * NOTE: this function doesnt have to take the runqueue lock,
1128 * because all it wants to ensure is that the remote task enters
1129 * the kernel. If the IPI races and the task has been migrated
1130 * to another CPU then no harm is done and the purpose has been
1133 void kick_process(struct task_struct
*p
)
1139 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1140 smp_send_reschedule(cpu
);
1145 * Return a low guess at the load of a migration-source cpu weighted
1146 * according to the scheduling class and "nice" value.
1148 * We want to under-estimate the load of migration sources, to
1149 * balance conservatively.
1151 static inline unsigned long source_load(int cpu
, int type
)
1153 struct rq
*rq
= cpu_rq(cpu
);
1154 unsigned long total
= weighted_cpuload(cpu
);
1159 return min(rq
->cpu_load
[type
-1], total
);
1163 * Return a high guess at the load of a migration-target cpu weighted
1164 * according to the scheduling class and "nice" value.
1166 static inline unsigned long target_load(int cpu
, int type
)
1168 struct rq
*rq
= cpu_rq(cpu
);
1169 unsigned long total
= weighted_cpuload(cpu
);
1174 return max(rq
->cpu_load
[type
-1], total
);
1178 * Return the average load per task on the cpu's run queue
1180 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1182 struct rq
*rq
= cpu_rq(cpu
);
1183 unsigned long total
= weighted_cpuload(cpu
);
1184 unsigned long n
= rq
->nr_running
;
1186 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1190 * find_idlest_group finds and returns the least busy CPU group within the
1193 static struct sched_group
*
1194 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1196 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1197 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1198 int load_idx
= sd
->forkexec_idx
;
1199 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1202 unsigned long load
, avg_load
;
1206 /* Skip over this group if it has no CPUs allowed */
1207 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1210 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1212 /* Tally up the load of all CPUs in the group */
1215 for_each_cpu_mask(i
, group
->cpumask
) {
1216 /* Bias balancing toward cpus of our domain */
1218 load
= source_load(i
, load_idx
);
1220 load
= target_load(i
, load_idx
);
1225 /* Adjust by relative CPU power of the group */
1226 avg_load
= sg_div_cpu_power(group
,
1227 avg_load
* SCHED_LOAD_SCALE
);
1230 this_load
= avg_load
;
1232 } else if (avg_load
< min_load
) {
1233 min_load
= avg_load
;
1237 group
= group
->next
;
1238 } while (group
!= sd
->groups
);
1240 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1246 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1249 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1252 unsigned long load
, min_load
= ULONG_MAX
;
1256 /* Traverse only the allowed CPUs */
1257 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1259 for_each_cpu_mask(i
, tmp
) {
1260 load
= weighted_cpuload(i
);
1262 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1272 * sched_balance_self: balance the current task (running on cpu) in domains
1273 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1276 * Balance, ie. select the least loaded group.
1278 * Returns the target CPU number, or the same CPU if no balancing is needed.
1280 * preempt must be disabled.
1282 static int sched_balance_self(int cpu
, int flag
)
1284 struct task_struct
*t
= current
;
1285 struct sched_domain
*tmp
, *sd
= NULL
;
1287 for_each_domain(cpu
, tmp
) {
1289 * If power savings logic is enabled for a domain, stop there.
1291 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1293 if (tmp
->flags
& flag
)
1299 struct sched_group
*group
;
1300 int new_cpu
, weight
;
1302 if (!(sd
->flags
& flag
)) {
1308 group
= find_idlest_group(sd
, t
, cpu
);
1314 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1315 if (new_cpu
== -1 || new_cpu
== cpu
) {
1316 /* Now try balancing at a lower domain level of cpu */
1321 /* Now try balancing at a lower domain level of new_cpu */
1324 weight
= cpus_weight(span
);
1325 for_each_domain(cpu
, tmp
) {
1326 if (weight
<= cpus_weight(tmp
->span
))
1328 if (tmp
->flags
& flag
)
1331 /* while loop will break here if sd == NULL */
1337 #endif /* CONFIG_SMP */
1340 * wake_idle() will wake a task on an idle cpu if task->cpu is
1341 * not idle and an idle cpu is available. The span of cpus to
1342 * search starts with cpus closest then further out as needed,
1343 * so we always favor a closer, idle cpu.
1345 * Returns the CPU we should wake onto.
1347 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1348 static int wake_idle(int cpu
, struct task_struct
*p
)
1351 struct sched_domain
*sd
;
1355 * If it is idle, then it is the best cpu to run this task.
1357 * This cpu is also the best, if it has more than one task already.
1358 * Siblings must be also busy(in most cases) as they didn't already
1359 * pickup the extra load from this cpu and hence we need not check
1360 * sibling runqueue info. This will avoid the checks and cache miss
1361 * penalities associated with that.
1363 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1366 for_each_domain(cpu
, sd
) {
1367 if (sd
->flags
& SD_WAKE_IDLE
) {
1368 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1369 for_each_cpu_mask(i
, tmp
) {
1380 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1387 * try_to_wake_up - wake up a thread
1388 * @p: the to-be-woken-up thread
1389 * @state: the mask of task states that can be woken
1390 * @sync: do a synchronous wakeup?
1392 * Put it on the run-queue if it's not already there. The "current"
1393 * thread is always on the run-queue (except when the actual
1394 * re-schedule is in progress), and as such you're allowed to do
1395 * the simpler "current->state = TASK_RUNNING" to mark yourself
1396 * runnable without the overhead of this.
1398 * returns failure only if the task is already active.
1400 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1402 int cpu
, this_cpu
, success
= 0;
1403 unsigned long flags
;
1407 struct sched_domain
*sd
, *this_sd
= NULL
;
1408 unsigned long load
, this_load
;
1412 rq
= task_rq_lock(p
, &flags
);
1413 old_state
= p
->state
;
1414 if (!(old_state
& state
))
1421 this_cpu
= smp_processor_id();
1424 if (unlikely(task_running(rq
, p
)))
1429 schedstat_inc(rq
, ttwu_cnt
);
1430 if (cpu
== this_cpu
) {
1431 schedstat_inc(rq
, ttwu_local
);
1435 for_each_domain(this_cpu
, sd
) {
1436 if (cpu_isset(cpu
, sd
->span
)) {
1437 schedstat_inc(sd
, ttwu_wake_remote
);
1443 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1447 * Check for affine wakeup and passive balancing possibilities.
1450 int idx
= this_sd
->wake_idx
;
1451 unsigned int imbalance
;
1453 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1455 load
= source_load(cpu
, idx
);
1456 this_load
= target_load(this_cpu
, idx
);
1458 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1460 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1461 unsigned long tl
= this_load
;
1462 unsigned long tl_per_task
;
1464 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1467 * If sync wakeup then subtract the (maximum possible)
1468 * effect of the currently running task from the load
1469 * of the current CPU:
1472 tl
-= current
->se
.load
.weight
;
1475 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1476 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1478 * This domain has SD_WAKE_AFFINE and
1479 * p is cache cold in this domain, and
1480 * there is no bad imbalance.
1482 schedstat_inc(this_sd
, ttwu_move_affine
);
1488 * Start passive balancing when half the imbalance_pct
1491 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1492 if (imbalance
*this_load
<= 100*load
) {
1493 schedstat_inc(this_sd
, ttwu_move_balance
);
1499 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1501 new_cpu
= wake_idle(new_cpu
, p
);
1502 if (new_cpu
!= cpu
) {
1503 set_task_cpu(p
, new_cpu
);
1504 task_rq_unlock(rq
, &flags
);
1505 /* might preempt at this point */
1506 rq
= task_rq_lock(p
, &flags
);
1507 old_state
= p
->state
;
1508 if (!(old_state
& state
))
1513 this_cpu
= smp_processor_id();
1518 #endif /* CONFIG_SMP */
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
;
1652 rq
= task_rq_lock(p
, &flags
);
1653 BUG_ON(p
->state
!= TASK_RUNNING
);
1654 this_cpu
= smp_processor_id(); /* parent's CPU */
1657 p
->prio
= effective_prio(p
);
1659 if (!p
->sched_class
->task_new
|| !sysctl_sched_child_runs_first
||
1660 (clone_flags
& CLONE_VM
) || task_cpu(p
) != this_cpu
||
1661 !current
->se
.on_rq
) {
1663 activate_task(rq
, p
, 0);
1666 * Let the scheduling class do new task startup
1667 * management (if any):
1669 p
->sched_class
->task_new(rq
, p
, now
);
1670 inc_nr_running(p
, rq
, now
);
1672 check_preempt_curr(rq
, p
);
1673 task_rq_unlock(rq
, &flags
);
1676 #ifdef CONFIG_PREEMPT_NOTIFIERS
1679 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1680 * @notifier: notifier struct to register
1682 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1684 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1686 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1689 * preempt_notifier_unregister - no longer interested in preemption notifications
1690 * @notifier: notifier struct to unregister
1692 * This is safe to call from within a preemption notifier.
1694 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1696 hlist_del(¬ifier
->link
);
1698 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1700 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1702 struct preempt_notifier
*notifier
;
1703 struct hlist_node
*node
;
1705 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1706 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1710 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1711 struct task_struct
*next
)
1713 struct preempt_notifier
*notifier
;
1714 struct hlist_node
*node
;
1716 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1717 notifier
->ops
->sched_out(notifier
, next
);
1722 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1727 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1728 struct task_struct
*next
)
1735 * prepare_task_switch - prepare to switch tasks
1736 * @rq: the runqueue preparing to switch
1737 * @prev: the current task that is being switched out
1738 * @next: the task we are going to switch to.
1740 * This is called with the rq lock held and interrupts off. It must
1741 * be paired with a subsequent finish_task_switch after the context
1744 * prepare_task_switch sets up locking and calls architecture specific
1748 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1749 struct task_struct
*next
)
1751 fire_sched_out_preempt_notifiers(prev
, next
);
1752 prepare_lock_switch(rq
, next
);
1753 prepare_arch_switch(next
);
1757 * finish_task_switch - clean up after a task-switch
1758 * @rq: runqueue associated with task-switch
1759 * @prev: the thread we just switched away from.
1761 * finish_task_switch must be called after the context switch, paired
1762 * with a prepare_task_switch call before the context switch.
1763 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1764 * and do any other architecture-specific cleanup actions.
1766 * Note that we may have delayed dropping an mm in context_switch(). If
1767 * so, we finish that here outside of the runqueue lock. (Doing it
1768 * with the lock held can cause deadlocks; see schedule() for
1771 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1772 __releases(rq
->lock
)
1774 struct mm_struct
*mm
= rq
->prev_mm
;
1780 * A task struct has one reference for the use as "current".
1781 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1782 * schedule one last time. The schedule call will never return, and
1783 * the scheduled task must drop that reference.
1784 * The test for TASK_DEAD must occur while the runqueue locks are
1785 * still held, otherwise prev could be scheduled on another cpu, die
1786 * there before we look at prev->state, and then the reference would
1788 * Manfred Spraul <manfred@colorfullife.com>
1790 prev_state
= prev
->state
;
1791 finish_arch_switch(prev
);
1792 finish_lock_switch(rq
, prev
);
1793 fire_sched_in_preempt_notifiers(current
);
1796 if (unlikely(prev_state
== TASK_DEAD
)) {
1798 * Remove function-return probe instances associated with this
1799 * task and put them back on the free list.
1801 kprobe_flush_task(prev
);
1802 put_task_struct(prev
);
1807 * schedule_tail - first thing a freshly forked thread must call.
1808 * @prev: the thread we just switched away from.
1810 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1811 __releases(rq
->lock
)
1813 struct rq
*rq
= this_rq();
1815 finish_task_switch(rq
, prev
);
1816 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1817 /* In this case, finish_task_switch does not reenable preemption */
1820 if (current
->set_child_tid
)
1821 put_user(current
->pid
, current
->set_child_tid
);
1825 * context_switch - switch to the new MM and the new
1826 * thread's register state.
1829 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1830 struct task_struct
*next
)
1832 struct mm_struct
*mm
, *oldmm
;
1834 prepare_task_switch(rq
, prev
, next
);
1836 oldmm
= prev
->active_mm
;
1838 * For paravirt, this is coupled with an exit in switch_to to
1839 * combine the page table reload and the switch backend into
1842 arch_enter_lazy_cpu_mode();
1844 if (unlikely(!mm
)) {
1845 next
->active_mm
= oldmm
;
1846 atomic_inc(&oldmm
->mm_count
);
1847 enter_lazy_tlb(oldmm
, next
);
1849 switch_mm(oldmm
, mm
, next
);
1851 if (unlikely(!prev
->mm
)) {
1852 prev
->active_mm
= NULL
;
1853 rq
->prev_mm
= oldmm
;
1856 * Since the runqueue lock will be released by the next
1857 * task (which is an invalid locking op but in the case
1858 * of the scheduler it's an obvious special-case), so we
1859 * do an early lockdep release here:
1861 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1862 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1865 /* Here we just switch the register state and the stack. */
1866 switch_to(prev
, next
, prev
);
1870 * this_rq must be evaluated again because prev may have moved
1871 * CPUs since it called schedule(), thus the 'rq' on its stack
1872 * frame will be invalid.
1874 finish_task_switch(this_rq(), prev
);
1878 * nr_running, nr_uninterruptible and nr_context_switches:
1880 * externally visible scheduler statistics: current number of runnable
1881 * threads, current number of uninterruptible-sleeping threads, total
1882 * number of context switches performed since bootup.
1884 unsigned long nr_running(void)
1886 unsigned long i
, sum
= 0;
1888 for_each_online_cpu(i
)
1889 sum
+= cpu_rq(i
)->nr_running
;
1894 unsigned long nr_uninterruptible(void)
1896 unsigned long i
, sum
= 0;
1898 for_each_possible_cpu(i
)
1899 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1902 * Since we read the counters lockless, it might be slightly
1903 * inaccurate. Do not allow it to go below zero though:
1905 if (unlikely((long)sum
< 0))
1911 unsigned long long nr_context_switches(void)
1914 unsigned long long sum
= 0;
1916 for_each_possible_cpu(i
)
1917 sum
+= cpu_rq(i
)->nr_switches
;
1922 unsigned long nr_iowait(void)
1924 unsigned long i
, sum
= 0;
1926 for_each_possible_cpu(i
)
1927 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1932 unsigned long nr_active(void)
1934 unsigned long i
, running
= 0, uninterruptible
= 0;
1936 for_each_online_cpu(i
) {
1937 running
+= cpu_rq(i
)->nr_running
;
1938 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1941 if (unlikely((long)uninterruptible
< 0))
1942 uninterruptible
= 0;
1944 return running
+ uninterruptible
;
1948 * Update rq->cpu_load[] statistics. This function is usually called every
1949 * scheduler tick (TICK_NSEC).
1951 static void update_cpu_load(struct rq
*this_rq
)
1953 u64 fair_delta64
, exec_delta64
, idle_delta64
, sample_interval64
, tmp64
;
1954 unsigned long total_load
= this_rq
->ls
.load
.weight
;
1955 unsigned long this_load
= total_load
;
1956 struct load_stat
*ls
= &this_rq
->ls
;
1957 u64 now
= __rq_clock(this_rq
);
1960 this_rq
->nr_load_updates
++;
1961 if (unlikely(!(sysctl_sched_features
& SCHED_FEAT_PRECISE_CPU_LOAD
)))
1964 /* Update delta_fair/delta_exec fields first */
1965 update_curr_load(this_rq
, now
);
1967 fair_delta64
= ls
->delta_fair
+ 1;
1970 exec_delta64
= ls
->delta_exec
+ 1;
1973 sample_interval64
= now
- ls
->load_update_last
;
1974 ls
->load_update_last
= now
;
1976 if ((s64
)sample_interval64
< (s64
)TICK_NSEC
)
1977 sample_interval64
= TICK_NSEC
;
1979 if (exec_delta64
> sample_interval64
)
1980 exec_delta64
= sample_interval64
;
1982 idle_delta64
= sample_interval64
- exec_delta64
;
1984 tmp64
= div64_64(SCHED_LOAD_SCALE
* exec_delta64
, fair_delta64
);
1985 tmp64
= div64_64(tmp64
* exec_delta64
, sample_interval64
);
1987 this_load
= (unsigned long)tmp64
;
1991 /* Update our load: */
1992 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1993 unsigned long old_load
, new_load
;
1995 /* scale is effectively 1 << i now, and >> i divides by scale */
1997 old_load
= this_rq
->cpu_load
[i
];
1998 new_load
= this_load
;
2000 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2007 * double_rq_lock - safely lock two runqueues
2009 * Note this does not disable interrupts like task_rq_lock,
2010 * you need to do so manually before calling.
2012 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2013 __acquires(rq1
->lock
)
2014 __acquires(rq2
->lock
)
2016 BUG_ON(!irqs_disabled());
2018 spin_lock(&rq1
->lock
);
2019 __acquire(rq2
->lock
); /* Fake it out ;) */
2022 spin_lock(&rq1
->lock
);
2023 spin_lock(&rq2
->lock
);
2025 spin_lock(&rq2
->lock
);
2026 spin_lock(&rq1
->lock
);
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
, int best_prio
, int best_prio_seen
,
2170 struct rq_iterator
*iterator
)
2172 int pulled
= 0, pinned
= 0, skip_for_load
;
2173 struct task_struct
*p
;
2174 long rem_load_move
= max_load_move
;
2176 if (max_nr_move
== 0 || max_load_move
== 0)
2182 * Start the load-balancing iterator:
2184 p
= iterator
->start(iterator
->arg
);
2189 * To help distribute high priority tasks accross CPUs we don't
2190 * skip a task if it will be the highest priority task (i.e. smallest
2191 * prio value) on its new queue regardless of its load weight
2193 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2194 SCHED_LOAD_SCALE_FUZZ
;
2195 if (skip_for_load
&& p
->prio
< this_best_prio
)
2196 skip_for_load
= !best_prio_seen
&& p
->prio
== best_prio
;
2197 if (skip_for_load
||
2198 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2200 best_prio_seen
|= p
->prio
== best_prio
;
2201 p
= iterator
->next(iterator
->arg
);
2205 pull_task(busiest
, p
, this_rq
, this_cpu
);
2207 rem_load_move
-= p
->se
.load
.weight
;
2210 * We only want to steal up to the prescribed number of tasks
2211 * and the prescribed amount of weighted load.
2213 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2214 if (p
->prio
< this_best_prio
)
2215 this_best_prio
= p
->prio
;
2216 p
= iterator
->next(iterator
->arg
);
2221 * Right now, this is the only place pull_task() is called,
2222 * so we can safely collect pull_task() stats here rather than
2223 * inside pull_task().
2225 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2228 *all_pinned
= pinned
;
2229 *load_moved
= max_load_move
- rem_load_move
;
2234 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2235 * load from busiest to this_rq, as part of a balancing operation within
2236 * "domain". Returns the number of tasks moved.
2238 * Called with both runqueues locked.
2240 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2241 unsigned long max_nr_move
, unsigned long max_load_move
,
2242 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2245 struct sched_class
*class = sched_class_highest
;
2246 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
2247 long rem_load_move
= max_load_move
;
2250 nr_moved
= class->load_balance(this_rq
, this_cpu
, busiest
,
2251 max_nr_move
, (unsigned long)rem_load_move
,
2252 sd
, idle
, all_pinned
, &load_moved
);
2253 total_nr_moved
+= nr_moved
;
2254 max_nr_move
-= nr_moved
;
2255 rem_load_move
-= load_moved
;
2256 class = class->next
;
2257 } while (class && max_nr_move
&& rem_load_move
> 0);
2259 return total_nr_moved
;
2263 * find_busiest_group finds and returns the busiest CPU group within the
2264 * domain. It calculates and returns the amount of weighted load which
2265 * should be moved to restore balance via the imbalance parameter.
2267 static struct sched_group
*
2268 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2269 unsigned long *imbalance
, enum cpu_idle_type idle
,
2270 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2272 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2273 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2274 unsigned long max_pull
;
2275 unsigned long busiest_load_per_task
, busiest_nr_running
;
2276 unsigned long this_load_per_task
, this_nr_running
;
2278 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2279 int power_savings_balance
= 1;
2280 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2281 unsigned long min_nr_running
= ULONG_MAX
;
2282 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2285 max_load
= this_load
= total_load
= total_pwr
= 0;
2286 busiest_load_per_task
= busiest_nr_running
= 0;
2287 this_load_per_task
= this_nr_running
= 0;
2288 if (idle
== CPU_NOT_IDLE
)
2289 load_idx
= sd
->busy_idx
;
2290 else if (idle
== CPU_NEWLY_IDLE
)
2291 load_idx
= sd
->newidle_idx
;
2293 load_idx
= sd
->idle_idx
;
2296 unsigned long load
, group_capacity
;
2299 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2300 unsigned long sum_nr_running
, sum_weighted_load
;
2302 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2305 balance_cpu
= first_cpu(group
->cpumask
);
2307 /* Tally up the load of all CPUs in the group */
2308 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2310 for_each_cpu_mask(i
, group
->cpumask
) {
2313 if (!cpu_isset(i
, *cpus
))
2318 if (*sd_idle
&& rq
->nr_running
)
2321 /* Bias balancing toward cpus of our domain */
2323 if (idle_cpu(i
) && !first_idle_cpu
) {
2328 load
= target_load(i
, load_idx
);
2330 load
= source_load(i
, load_idx
);
2333 sum_nr_running
+= rq
->nr_running
;
2334 sum_weighted_load
+= weighted_cpuload(i
);
2338 * First idle cpu or the first cpu(busiest) in this sched group
2339 * is eligible for doing load balancing at this and above
2340 * domains. In the newly idle case, we will allow all the cpu's
2341 * to do the newly idle load balance.
2343 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2344 balance_cpu
!= this_cpu
&& balance
) {
2349 total_load
+= avg_load
;
2350 total_pwr
+= group
->__cpu_power
;
2352 /* Adjust by relative CPU power of the group */
2353 avg_load
= sg_div_cpu_power(group
,
2354 avg_load
* SCHED_LOAD_SCALE
);
2356 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2359 this_load
= avg_load
;
2361 this_nr_running
= sum_nr_running
;
2362 this_load_per_task
= sum_weighted_load
;
2363 } else if (avg_load
> max_load
&&
2364 sum_nr_running
> group_capacity
) {
2365 max_load
= avg_load
;
2367 busiest_nr_running
= sum_nr_running
;
2368 busiest_load_per_task
= sum_weighted_load
;
2371 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2373 * Busy processors will not participate in power savings
2376 if (idle
== CPU_NOT_IDLE
||
2377 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2381 * If the local group is idle or completely loaded
2382 * no need to do power savings balance at this domain
2384 if (local_group
&& (this_nr_running
>= group_capacity
||
2386 power_savings_balance
= 0;
2389 * If a group is already running at full capacity or idle,
2390 * don't include that group in power savings calculations
2392 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2397 * Calculate the group which has the least non-idle load.
2398 * This is the group from where we need to pick up the load
2401 if ((sum_nr_running
< min_nr_running
) ||
2402 (sum_nr_running
== min_nr_running
&&
2403 first_cpu(group
->cpumask
) <
2404 first_cpu(group_min
->cpumask
))) {
2406 min_nr_running
= sum_nr_running
;
2407 min_load_per_task
= sum_weighted_load
/
2412 * Calculate the group which is almost near its
2413 * capacity but still has some space to pick up some load
2414 * from other group and save more power
2416 if (sum_nr_running
<= group_capacity
- 1) {
2417 if (sum_nr_running
> leader_nr_running
||
2418 (sum_nr_running
== leader_nr_running
&&
2419 first_cpu(group
->cpumask
) >
2420 first_cpu(group_leader
->cpumask
))) {
2421 group_leader
= group
;
2422 leader_nr_running
= sum_nr_running
;
2427 group
= group
->next
;
2428 } while (group
!= sd
->groups
);
2430 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2433 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2435 if (this_load
>= avg_load
||
2436 100*max_load
<= sd
->imbalance_pct
*this_load
)
2439 busiest_load_per_task
/= busiest_nr_running
;
2441 * We're trying to get all the cpus to the average_load, so we don't
2442 * want to push ourselves above the average load, nor do we wish to
2443 * reduce the max loaded cpu below the average load, as either of these
2444 * actions would just result in more rebalancing later, and ping-pong
2445 * tasks around. Thus we look for the minimum possible imbalance.
2446 * Negative imbalances (*we* are more loaded than anyone else) will
2447 * be counted as no imbalance for these purposes -- we can't fix that
2448 * by pulling tasks to us. Be careful of negative numbers as they'll
2449 * appear as very large values with unsigned longs.
2451 if (max_load
<= busiest_load_per_task
)
2455 * In the presence of smp nice balancing, certain scenarios can have
2456 * max load less than avg load(as we skip the groups at or below
2457 * its cpu_power, while calculating max_load..)
2459 if (max_load
< avg_load
) {
2461 goto small_imbalance
;
2464 /* Don't want to pull so many tasks that a group would go idle */
2465 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2467 /* How much load to actually move to equalise the imbalance */
2468 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2469 (avg_load
- this_load
) * this->__cpu_power
)
2473 * if *imbalance is less than the average load per runnable task
2474 * there is no gaurantee that any tasks will be moved so we'll have
2475 * a think about bumping its value to force at least one task to be
2478 if (*imbalance
+ SCHED_LOAD_SCALE_FUZZ
< busiest_load_per_task
/2) {
2479 unsigned long tmp
, pwr_now
, pwr_move
;
2483 pwr_move
= pwr_now
= 0;
2485 if (this_nr_running
) {
2486 this_load_per_task
/= this_nr_running
;
2487 if (busiest_load_per_task
> this_load_per_task
)
2490 this_load_per_task
= SCHED_LOAD_SCALE
;
2492 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2493 busiest_load_per_task
* imbn
) {
2494 *imbalance
= busiest_load_per_task
;
2499 * OK, we don't have enough imbalance to justify moving tasks,
2500 * however we may be able to increase total CPU power used by
2504 pwr_now
+= busiest
->__cpu_power
*
2505 min(busiest_load_per_task
, max_load
);
2506 pwr_now
+= this->__cpu_power
*
2507 min(this_load_per_task
, this_load
);
2508 pwr_now
/= SCHED_LOAD_SCALE
;
2510 /* Amount of load we'd subtract */
2511 tmp
= sg_div_cpu_power(busiest
,
2512 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2514 pwr_move
+= busiest
->__cpu_power
*
2515 min(busiest_load_per_task
, max_load
- tmp
);
2517 /* Amount of load we'd add */
2518 if (max_load
* busiest
->__cpu_power
<
2519 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2520 tmp
= sg_div_cpu_power(this,
2521 max_load
* busiest
->__cpu_power
);
2523 tmp
= sg_div_cpu_power(this,
2524 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2525 pwr_move
+= this->__cpu_power
*
2526 min(this_load_per_task
, this_load
+ tmp
);
2527 pwr_move
/= SCHED_LOAD_SCALE
;
2529 /* Move if we gain throughput */
2530 if (pwr_move
<= pwr_now
)
2533 *imbalance
= busiest_load_per_task
;
2539 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2540 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2543 if (this == group_leader
&& group_leader
!= group_min
) {
2544 *imbalance
= min_load_per_task
;
2554 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2557 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2558 unsigned long imbalance
, cpumask_t
*cpus
)
2560 struct rq
*busiest
= NULL
, *rq
;
2561 unsigned long max_load
= 0;
2564 for_each_cpu_mask(i
, group
->cpumask
) {
2567 if (!cpu_isset(i
, *cpus
))
2571 wl
= weighted_cpuload(i
);
2573 if (rq
->nr_running
== 1 && wl
> imbalance
)
2576 if (wl
> max_load
) {
2586 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2587 * so long as it is large enough.
2589 #define MAX_PINNED_INTERVAL 512
2591 static inline unsigned long minus_1_or_zero(unsigned long n
)
2593 return n
> 0 ? n
- 1 : 0;
2597 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2598 * tasks if there is an imbalance.
2600 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2601 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2604 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2605 struct sched_group
*group
;
2606 unsigned long imbalance
;
2608 cpumask_t cpus
= CPU_MASK_ALL
;
2609 unsigned long flags
;
2612 * When power savings policy is enabled for the parent domain, idle
2613 * sibling can pick up load irrespective of busy siblings. In this case,
2614 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2615 * portraying it as CPU_NOT_IDLE.
2617 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2618 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2621 schedstat_inc(sd
, lb_cnt
[idle
]);
2624 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2631 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2635 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2637 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2641 BUG_ON(busiest
== this_rq
);
2643 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2646 if (busiest
->nr_running
> 1) {
2648 * Attempt to move tasks. If find_busiest_group has found
2649 * an imbalance but busiest->nr_running <= 1, the group is
2650 * still unbalanced. nr_moved simply stays zero, so it is
2651 * correctly treated as an imbalance.
2653 local_irq_save(flags
);
2654 double_rq_lock(this_rq
, busiest
);
2655 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2656 minus_1_or_zero(busiest
->nr_running
),
2657 imbalance
, sd
, idle
, &all_pinned
);
2658 double_rq_unlock(this_rq
, busiest
);
2659 local_irq_restore(flags
);
2662 * some other cpu did the load balance for us.
2664 if (nr_moved
&& this_cpu
!= smp_processor_id())
2665 resched_cpu(this_cpu
);
2667 /* All tasks on this runqueue were pinned by CPU affinity */
2668 if (unlikely(all_pinned
)) {
2669 cpu_clear(cpu_of(busiest
), cpus
);
2670 if (!cpus_empty(cpus
))
2677 schedstat_inc(sd
, lb_failed
[idle
]);
2678 sd
->nr_balance_failed
++;
2680 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2682 spin_lock_irqsave(&busiest
->lock
, flags
);
2684 /* don't kick the migration_thread, if the curr
2685 * task on busiest cpu can't be moved to this_cpu
2687 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2688 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2690 goto out_one_pinned
;
2693 if (!busiest
->active_balance
) {
2694 busiest
->active_balance
= 1;
2695 busiest
->push_cpu
= this_cpu
;
2698 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2700 wake_up_process(busiest
->migration_thread
);
2703 * We've kicked active balancing, reset the failure
2706 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2709 sd
->nr_balance_failed
= 0;
2711 if (likely(!active_balance
)) {
2712 /* We were unbalanced, so reset the balancing interval */
2713 sd
->balance_interval
= sd
->min_interval
;
2716 * If we've begun active balancing, start to back off. This
2717 * case may not be covered by the all_pinned logic if there
2718 * is only 1 task on the busy runqueue (because we don't call
2721 if (sd
->balance_interval
< sd
->max_interval
)
2722 sd
->balance_interval
*= 2;
2725 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2726 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2731 schedstat_inc(sd
, lb_balanced
[idle
]);
2733 sd
->nr_balance_failed
= 0;
2736 /* tune up the balancing interval */
2737 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2738 (sd
->balance_interval
< sd
->max_interval
))
2739 sd
->balance_interval
*= 2;
2741 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2742 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2748 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2749 * tasks if there is an imbalance.
2751 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2752 * this_rq is locked.
2755 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2757 struct sched_group
*group
;
2758 struct rq
*busiest
= NULL
;
2759 unsigned long imbalance
;
2763 cpumask_t cpus
= CPU_MASK_ALL
;
2766 * When power savings policy is enabled for the parent domain, idle
2767 * sibling can pick up load irrespective of busy siblings. In this case,
2768 * let the state of idle sibling percolate up as IDLE, instead of
2769 * portraying it as CPU_NOT_IDLE.
2771 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2772 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2775 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2777 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2778 &sd_idle
, &cpus
, NULL
);
2780 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2784 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2787 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2791 BUG_ON(busiest
== this_rq
);
2793 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2796 if (busiest
->nr_running
> 1) {
2797 /* Attempt to move tasks */
2798 double_lock_balance(this_rq
, busiest
);
2799 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2800 minus_1_or_zero(busiest
->nr_running
),
2801 imbalance
, sd
, CPU_NEWLY_IDLE
,
2803 spin_unlock(&busiest
->lock
);
2805 if (unlikely(all_pinned
)) {
2806 cpu_clear(cpu_of(busiest
), cpus
);
2807 if (!cpus_empty(cpus
))
2813 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2814 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2815 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2818 sd
->nr_balance_failed
= 0;
2823 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2824 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2825 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2827 sd
->nr_balance_failed
= 0;
2833 * idle_balance is called by schedule() if this_cpu is about to become
2834 * idle. Attempts to pull tasks from other CPUs.
2836 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2838 struct sched_domain
*sd
;
2839 int pulled_task
= -1;
2840 unsigned long next_balance
= jiffies
+ HZ
;
2842 for_each_domain(this_cpu
, sd
) {
2843 unsigned long interval
;
2845 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2848 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2849 /* If we've pulled tasks over stop searching: */
2850 pulled_task
= load_balance_newidle(this_cpu
,
2853 interval
= msecs_to_jiffies(sd
->balance_interval
);
2854 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2855 next_balance
= sd
->last_balance
+ interval
;
2859 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2861 * We are going idle. next_balance may be set based on
2862 * a busy processor. So reset next_balance.
2864 this_rq
->next_balance
= next_balance
;
2869 * active_load_balance is run by migration threads. It pushes running tasks
2870 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2871 * running on each physical CPU where possible, and avoids physical /
2872 * logical imbalances.
2874 * Called with busiest_rq locked.
2876 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2878 int target_cpu
= busiest_rq
->push_cpu
;
2879 struct sched_domain
*sd
;
2880 struct rq
*target_rq
;
2882 /* Is there any task to move? */
2883 if (busiest_rq
->nr_running
<= 1)
2886 target_rq
= cpu_rq(target_cpu
);
2889 * This condition is "impossible", if it occurs
2890 * we need to fix it. Originally reported by
2891 * Bjorn Helgaas on a 128-cpu setup.
2893 BUG_ON(busiest_rq
== target_rq
);
2895 /* move a task from busiest_rq to target_rq */
2896 double_lock_balance(busiest_rq
, target_rq
);
2898 /* Search for an sd spanning us and the target CPU. */
2899 for_each_domain(target_cpu
, sd
) {
2900 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2901 cpu_isset(busiest_cpu
, sd
->span
))
2906 schedstat_inc(sd
, alb_cnt
);
2908 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2909 ULONG_MAX
, sd
, CPU_IDLE
, NULL
))
2910 schedstat_inc(sd
, alb_pushed
);
2912 schedstat_inc(sd
, alb_failed
);
2914 spin_unlock(&target_rq
->lock
);
2919 atomic_t load_balancer
;
2921 } nohz ____cacheline_aligned
= {
2922 .load_balancer
= ATOMIC_INIT(-1),
2923 .cpu_mask
= CPU_MASK_NONE
,
2927 * This routine will try to nominate the ilb (idle load balancing)
2928 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2929 * load balancing on behalf of all those cpus. If all the cpus in the system
2930 * go into this tickless mode, then there will be no ilb owner (as there is
2931 * no need for one) and all the cpus will sleep till the next wakeup event
2934 * For the ilb owner, tick is not stopped. And this tick will be used
2935 * for idle load balancing. ilb owner will still be part of
2938 * While stopping the tick, this cpu will become the ilb owner if there
2939 * is no other owner. And will be the owner till that cpu becomes busy
2940 * or if all cpus in the system stop their ticks at which point
2941 * there is no need for ilb owner.
2943 * When the ilb owner becomes busy, it nominates another owner, during the
2944 * next busy scheduler_tick()
2946 int select_nohz_load_balancer(int stop_tick
)
2948 int cpu
= smp_processor_id();
2951 cpu_set(cpu
, nohz
.cpu_mask
);
2952 cpu_rq(cpu
)->in_nohz_recently
= 1;
2955 * If we are going offline and still the leader, give up!
2957 if (cpu_is_offline(cpu
) &&
2958 atomic_read(&nohz
.load_balancer
) == cpu
) {
2959 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2964 /* time for ilb owner also to sleep */
2965 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2966 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2967 atomic_set(&nohz
.load_balancer
, -1);
2971 if (atomic_read(&nohz
.load_balancer
) == -1) {
2972 /* make me the ilb owner */
2973 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2975 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2978 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2981 cpu_clear(cpu
, nohz
.cpu_mask
);
2983 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2984 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2991 static DEFINE_SPINLOCK(balancing
);
2994 * It checks each scheduling domain to see if it is due to be balanced,
2995 * and initiates a balancing operation if so.
2997 * Balancing parameters are set up in arch_init_sched_domains.
2999 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3002 struct rq
*rq
= cpu_rq(cpu
);
3003 unsigned long interval
;
3004 struct sched_domain
*sd
;
3005 /* Earliest time when we have to do rebalance again */
3006 unsigned long next_balance
= jiffies
+ 60*HZ
;
3008 for_each_domain(cpu
, sd
) {
3009 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3012 interval
= sd
->balance_interval
;
3013 if (idle
!= CPU_IDLE
)
3014 interval
*= sd
->busy_factor
;
3016 /* scale ms to jiffies */
3017 interval
= msecs_to_jiffies(interval
);
3018 if (unlikely(!interval
))
3020 if (interval
> HZ
*NR_CPUS
/10)
3021 interval
= HZ
*NR_CPUS
/10;
3024 if (sd
->flags
& SD_SERIALIZE
) {
3025 if (!spin_trylock(&balancing
))
3029 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3030 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3032 * We've pulled tasks over so either we're no
3033 * longer idle, or one of our SMT siblings is
3036 idle
= CPU_NOT_IDLE
;
3038 sd
->last_balance
= jiffies
;
3040 if (sd
->flags
& SD_SERIALIZE
)
3041 spin_unlock(&balancing
);
3043 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3044 next_balance
= sd
->last_balance
+ interval
;
3047 * Stop the load balance at this level. There is another
3048 * CPU in our sched group which is doing load balancing more
3054 rq
->next_balance
= next_balance
;
3058 * run_rebalance_domains is triggered when needed from the scheduler tick.
3059 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3060 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3062 static void run_rebalance_domains(struct softirq_action
*h
)
3064 int this_cpu
= smp_processor_id();
3065 struct rq
*this_rq
= cpu_rq(this_cpu
);
3066 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3067 CPU_IDLE
: CPU_NOT_IDLE
;
3069 rebalance_domains(this_cpu
, idle
);
3073 * If this cpu is the owner for idle load balancing, then do the
3074 * balancing on behalf of the other idle cpus whose ticks are
3077 if (this_rq
->idle_at_tick
&&
3078 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3079 cpumask_t cpus
= nohz
.cpu_mask
;
3083 cpu_clear(this_cpu
, cpus
);
3084 for_each_cpu_mask(balance_cpu
, cpus
) {
3086 * If this cpu gets work to do, stop the load balancing
3087 * work being done for other cpus. Next load
3088 * balancing owner will pick it up.
3093 rebalance_domains(balance_cpu
, SCHED_IDLE
);
3095 rq
= cpu_rq(balance_cpu
);
3096 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3097 this_rq
->next_balance
= rq
->next_balance
;
3104 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3106 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3107 * idle load balancing owner or decide to stop the periodic load balancing,
3108 * if the whole system is idle.
3110 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3114 * If we were in the nohz mode recently and busy at the current
3115 * scheduler tick, then check if we need to nominate new idle
3118 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3119 rq
->in_nohz_recently
= 0;
3121 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3122 cpu_clear(cpu
, nohz
.cpu_mask
);
3123 atomic_set(&nohz
.load_balancer
, -1);
3126 if (atomic_read(&nohz
.load_balancer
) == -1) {
3128 * simple selection for now: Nominate the
3129 * first cpu in the nohz list to be the next
3132 * TBD: Traverse the sched domains and nominate
3133 * the nearest cpu in the nohz.cpu_mask.
3135 int ilb
= first_cpu(nohz
.cpu_mask
);
3143 * If this cpu is idle and doing idle load balancing for all the
3144 * cpus with ticks stopped, is it time for that to stop?
3146 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3147 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3153 * If this cpu is idle and the idle load balancing is done by
3154 * someone else, then no need raise the SCHED_SOFTIRQ
3156 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3157 cpu_isset(cpu
, nohz
.cpu_mask
))
3160 if (time_after_eq(jiffies
, rq
->next_balance
))
3161 raise_softirq(SCHED_SOFTIRQ
);
3164 #else /* CONFIG_SMP */
3167 * on UP we do not need to balance between CPUs:
3169 static inline void idle_balance(int cpu
, struct rq
*rq
)
3173 /* Avoid "used but not defined" warning on UP */
3174 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3175 unsigned long max_nr_move
, unsigned long max_load_move
,
3176 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3177 int *all_pinned
, unsigned long *load_moved
,
3178 int this_best_prio
, int best_prio
, int best_prio_seen
,
3179 struct rq_iterator
*iterator
)
3188 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3190 EXPORT_PER_CPU_SYMBOL(kstat
);
3193 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3194 * that have not yet been banked in case the task is currently running.
3196 unsigned long long task_sched_runtime(struct task_struct
*p
)
3198 unsigned long flags
;
3202 rq
= task_rq_lock(p
, &flags
);
3203 ns
= p
->se
.sum_exec_runtime
;
3204 if (rq
->curr
== p
) {
3205 delta_exec
= rq_clock(rq
) - p
->se
.exec_start
;
3206 if ((s64
)delta_exec
> 0)
3209 task_rq_unlock(rq
, &flags
);
3215 * Account user cpu time to a process.
3216 * @p: the process that the cpu time gets accounted to
3217 * @hardirq_offset: the offset to subtract from hardirq_count()
3218 * @cputime: the cpu time spent in user space since the last update
3220 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3222 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3225 p
->utime
= cputime_add(p
->utime
, cputime
);
3227 /* Add user time to cpustat. */
3228 tmp
= cputime_to_cputime64(cputime
);
3229 if (TASK_NICE(p
) > 0)
3230 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3232 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3236 * Account system cpu time to a process.
3237 * @p: the process that the cpu time gets accounted to
3238 * @hardirq_offset: the offset to subtract from hardirq_count()
3239 * @cputime: the cpu time spent in kernel space since the last update
3241 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3244 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3245 struct rq
*rq
= this_rq();
3248 p
->stime
= cputime_add(p
->stime
, cputime
);
3250 /* Add system time to cpustat. */
3251 tmp
= cputime_to_cputime64(cputime
);
3252 if (hardirq_count() - hardirq_offset
)
3253 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3254 else if (softirq_count())
3255 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3256 else if (p
!= rq
->idle
)
3257 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3258 else if (atomic_read(&rq
->nr_iowait
) > 0)
3259 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3261 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3262 /* Account for system time used */
3263 acct_update_integrals(p
);
3267 * Account for involuntary wait time.
3268 * @p: the process from which the cpu time has been stolen
3269 * @steal: the cpu time spent in involuntary wait
3271 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3273 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3274 cputime64_t tmp
= cputime_to_cputime64(steal
);
3275 struct rq
*rq
= this_rq();
3277 if (p
== rq
->idle
) {
3278 p
->stime
= cputime_add(p
->stime
, steal
);
3279 if (atomic_read(&rq
->nr_iowait
) > 0)
3280 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3282 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3284 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3288 * This function gets called by the timer code, with HZ frequency.
3289 * We call it with interrupts disabled.
3291 * It also gets called by the fork code, when changing the parent's
3294 void scheduler_tick(void)
3296 int cpu
= smp_processor_id();
3297 struct rq
*rq
= cpu_rq(cpu
);
3298 struct task_struct
*curr
= rq
->curr
;
3300 spin_lock(&rq
->lock
);
3301 update_cpu_load(rq
);
3302 if (curr
!= rq
->idle
) /* FIXME: needed? */
3303 curr
->sched_class
->task_tick(rq
, curr
);
3304 spin_unlock(&rq
->lock
);
3307 rq
->idle_at_tick
= idle_cpu(cpu
);
3308 trigger_load_balance(rq
, cpu
);
3312 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3314 void fastcall
add_preempt_count(int val
)
3319 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3321 preempt_count() += val
;
3323 * Spinlock count overflowing soon?
3325 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3328 EXPORT_SYMBOL(add_preempt_count
);
3330 void fastcall
sub_preempt_count(int val
)
3335 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3338 * Is the spinlock portion underflowing?
3340 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3341 !(preempt_count() & PREEMPT_MASK
)))
3344 preempt_count() -= val
;
3346 EXPORT_SYMBOL(sub_preempt_count
);
3351 * Print scheduling while atomic bug:
3353 static noinline
void __schedule_bug(struct task_struct
*prev
)
3355 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3356 prev
->comm
, preempt_count(), prev
->pid
);
3357 debug_show_held_locks(prev
);
3358 if (irqs_disabled())
3359 print_irqtrace_events(prev
);
3364 * Various schedule()-time debugging checks and statistics:
3366 static inline void schedule_debug(struct task_struct
*prev
)
3369 * Test if we are atomic. Since do_exit() needs to call into
3370 * schedule() atomically, we ignore that path for now.
3371 * Otherwise, whine if we are scheduling when we should not be.
3373 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3374 __schedule_bug(prev
);
3376 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3378 schedstat_inc(this_rq(), sched_cnt
);
3382 * Pick up the highest-prio task:
3384 static inline struct task_struct
*
3385 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, u64 now
)
3387 struct sched_class
*class;
3388 struct task_struct
*p
;
3391 * Optimization: we know that if all tasks are in
3392 * the fair class we can call that function directly:
3394 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3395 p
= fair_sched_class
.pick_next_task(rq
, now
);
3400 class = sched_class_highest
;
3402 p
= class->pick_next_task(rq
, now
);
3406 * Will never be NULL as the idle class always
3407 * returns a non-NULL p:
3409 class = class->next
;
3414 * schedule() is the main scheduler function.
3416 asmlinkage
void __sched
schedule(void)
3418 struct task_struct
*prev
, *next
;
3426 cpu
= smp_processor_id();
3430 switch_count
= &prev
->nivcsw
;
3432 release_kernel_lock(prev
);
3433 need_resched_nonpreemptible
:
3435 schedule_debug(prev
);
3437 spin_lock_irq(&rq
->lock
);
3438 clear_tsk_need_resched(prev
);
3440 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3441 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3442 unlikely(signal_pending(prev
)))) {
3443 prev
->state
= TASK_RUNNING
;
3445 deactivate_task(rq
, prev
, 1);
3447 switch_count
= &prev
->nvcsw
;
3450 if (unlikely(!rq
->nr_running
))
3451 idle_balance(cpu
, rq
);
3453 now
= __rq_clock(rq
);
3454 prev
->sched_class
->put_prev_task(rq
, prev
, now
);
3455 next
= pick_next_task(rq
, prev
, now
);
3457 sched_info_switch(prev
, next
);
3459 if (likely(prev
!= next
)) {
3464 context_switch(rq
, prev
, next
); /* unlocks the rq */
3466 spin_unlock_irq(&rq
->lock
);
3468 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3469 cpu
= smp_processor_id();
3471 goto need_resched_nonpreemptible
;
3473 preempt_enable_no_resched();
3474 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3477 EXPORT_SYMBOL(schedule
);
3479 #ifdef CONFIG_PREEMPT
3481 * this is the entry point to schedule() from in-kernel preemption
3482 * off of preempt_enable. Kernel preemptions off return from interrupt
3483 * occur there and call schedule directly.
3485 asmlinkage
void __sched
preempt_schedule(void)
3487 struct thread_info
*ti
= current_thread_info();
3488 #ifdef CONFIG_PREEMPT_BKL
3489 struct task_struct
*task
= current
;
3490 int saved_lock_depth
;
3493 * If there is a non-zero preempt_count or interrupts are disabled,
3494 * we do not want to preempt the current task. Just return..
3496 if (likely(ti
->preempt_count
|| irqs_disabled()))
3500 add_preempt_count(PREEMPT_ACTIVE
);
3502 * We keep the big kernel semaphore locked, but we
3503 * clear ->lock_depth so that schedule() doesnt
3504 * auto-release the semaphore:
3506 #ifdef CONFIG_PREEMPT_BKL
3507 saved_lock_depth
= task
->lock_depth
;
3508 task
->lock_depth
= -1;
3511 #ifdef CONFIG_PREEMPT_BKL
3512 task
->lock_depth
= saved_lock_depth
;
3514 sub_preempt_count(PREEMPT_ACTIVE
);
3516 /* we could miss a preemption opportunity between schedule and now */
3518 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3521 EXPORT_SYMBOL(preempt_schedule
);
3524 * this is the entry point to schedule() from kernel preemption
3525 * off of irq context.
3526 * Note, that this is called and return with irqs disabled. This will
3527 * protect us against recursive calling from irq.
3529 asmlinkage
void __sched
preempt_schedule_irq(void)
3531 struct thread_info
*ti
= current_thread_info();
3532 #ifdef CONFIG_PREEMPT_BKL
3533 struct task_struct
*task
= current
;
3534 int saved_lock_depth
;
3536 /* Catch callers which need to be fixed */
3537 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3540 add_preempt_count(PREEMPT_ACTIVE
);
3542 * We keep the big kernel semaphore locked, but we
3543 * clear ->lock_depth so that schedule() doesnt
3544 * auto-release the semaphore:
3546 #ifdef CONFIG_PREEMPT_BKL
3547 saved_lock_depth
= task
->lock_depth
;
3548 task
->lock_depth
= -1;
3552 local_irq_disable();
3553 #ifdef CONFIG_PREEMPT_BKL
3554 task
->lock_depth
= saved_lock_depth
;
3556 sub_preempt_count(PREEMPT_ACTIVE
);
3558 /* we could miss a preemption opportunity between schedule and now */
3560 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3564 #endif /* CONFIG_PREEMPT */
3566 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3569 return try_to_wake_up(curr
->private, mode
, sync
);
3571 EXPORT_SYMBOL(default_wake_function
);
3574 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3575 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3576 * number) then we wake all the non-exclusive tasks and one exclusive task.
3578 * There are circumstances in which we can try to wake a task which has already
3579 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3580 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3582 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3583 int nr_exclusive
, int sync
, void *key
)
3585 struct list_head
*tmp
, *next
;
3587 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3588 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3589 unsigned flags
= curr
->flags
;
3591 if (curr
->func(curr
, mode
, sync
, key
) &&
3592 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3598 * __wake_up - wake up threads blocked on a waitqueue.
3600 * @mode: which threads
3601 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3602 * @key: is directly passed to the wakeup function
3604 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3605 int nr_exclusive
, void *key
)
3607 unsigned long flags
;
3609 spin_lock_irqsave(&q
->lock
, flags
);
3610 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3611 spin_unlock_irqrestore(&q
->lock
, flags
);
3613 EXPORT_SYMBOL(__wake_up
);
3616 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3618 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3620 __wake_up_common(q
, mode
, 1, 0, NULL
);
3624 * __wake_up_sync - wake up threads blocked on a waitqueue.
3626 * @mode: which threads
3627 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3629 * The sync wakeup differs that the waker knows that it will schedule
3630 * away soon, so while the target thread will be woken up, it will not
3631 * be migrated to another CPU - ie. the two threads are 'synchronized'
3632 * with each other. This can prevent needless bouncing between CPUs.
3634 * On UP it can prevent extra preemption.
3637 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3639 unsigned long flags
;
3645 if (unlikely(!nr_exclusive
))
3648 spin_lock_irqsave(&q
->lock
, flags
);
3649 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3650 spin_unlock_irqrestore(&q
->lock
, flags
);
3652 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3654 void fastcall
complete(struct completion
*x
)
3656 unsigned long flags
;
3658 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3660 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3662 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3664 EXPORT_SYMBOL(complete
);
3666 void fastcall
complete_all(struct completion
*x
)
3668 unsigned long flags
;
3670 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3671 x
->done
+= UINT_MAX
/2;
3672 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3674 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3676 EXPORT_SYMBOL(complete_all
);
3678 void fastcall __sched
wait_for_completion(struct completion
*x
)
3682 spin_lock_irq(&x
->wait
.lock
);
3684 DECLARE_WAITQUEUE(wait
, current
);
3686 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3687 __add_wait_queue_tail(&x
->wait
, &wait
);
3689 __set_current_state(TASK_UNINTERRUPTIBLE
);
3690 spin_unlock_irq(&x
->wait
.lock
);
3692 spin_lock_irq(&x
->wait
.lock
);
3694 __remove_wait_queue(&x
->wait
, &wait
);
3697 spin_unlock_irq(&x
->wait
.lock
);
3699 EXPORT_SYMBOL(wait_for_completion
);
3701 unsigned long fastcall __sched
3702 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3706 spin_lock_irq(&x
->wait
.lock
);
3708 DECLARE_WAITQUEUE(wait
, current
);
3710 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3711 __add_wait_queue_tail(&x
->wait
, &wait
);
3713 __set_current_state(TASK_UNINTERRUPTIBLE
);
3714 spin_unlock_irq(&x
->wait
.lock
);
3715 timeout
= schedule_timeout(timeout
);
3716 spin_lock_irq(&x
->wait
.lock
);
3718 __remove_wait_queue(&x
->wait
, &wait
);
3722 __remove_wait_queue(&x
->wait
, &wait
);
3726 spin_unlock_irq(&x
->wait
.lock
);
3729 EXPORT_SYMBOL(wait_for_completion_timeout
);
3731 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3737 spin_lock_irq(&x
->wait
.lock
);
3739 DECLARE_WAITQUEUE(wait
, current
);
3741 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3742 __add_wait_queue_tail(&x
->wait
, &wait
);
3744 if (signal_pending(current
)) {
3746 __remove_wait_queue(&x
->wait
, &wait
);
3749 __set_current_state(TASK_INTERRUPTIBLE
);
3750 spin_unlock_irq(&x
->wait
.lock
);
3752 spin_lock_irq(&x
->wait
.lock
);
3754 __remove_wait_queue(&x
->wait
, &wait
);
3758 spin_unlock_irq(&x
->wait
.lock
);
3762 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3764 unsigned long fastcall __sched
3765 wait_for_completion_interruptible_timeout(struct completion
*x
,
3766 unsigned long timeout
)
3770 spin_lock_irq(&x
->wait
.lock
);
3772 DECLARE_WAITQUEUE(wait
, current
);
3774 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3775 __add_wait_queue_tail(&x
->wait
, &wait
);
3777 if (signal_pending(current
)) {
3778 timeout
= -ERESTARTSYS
;
3779 __remove_wait_queue(&x
->wait
, &wait
);
3782 __set_current_state(TASK_INTERRUPTIBLE
);
3783 spin_unlock_irq(&x
->wait
.lock
);
3784 timeout
= schedule_timeout(timeout
);
3785 spin_lock_irq(&x
->wait
.lock
);
3787 __remove_wait_queue(&x
->wait
, &wait
);
3791 __remove_wait_queue(&x
->wait
, &wait
);
3795 spin_unlock_irq(&x
->wait
.lock
);
3798 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3801 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3803 spin_lock_irqsave(&q
->lock
, *flags
);
3804 __add_wait_queue(q
, wait
);
3805 spin_unlock(&q
->lock
);
3809 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3811 spin_lock_irq(&q
->lock
);
3812 __remove_wait_queue(q
, wait
);
3813 spin_unlock_irqrestore(&q
->lock
, *flags
);
3816 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3818 unsigned long flags
;
3821 init_waitqueue_entry(&wait
, current
);
3823 current
->state
= TASK_INTERRUPTIBLE
;
3825 sleep_on_head(q
, &wait
, &flags
);
3827 sleep_on_tail(q
, &wait
, &flags
);
3829 EXPORT_SYMBOL(interruptible_sleep_on
);
3832 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3834 unsigned long flags
;
3837 init_waitqueue_entry(&wait
, current
);
3839 current
->state
= TASK_INTERRUPTIBLE
;
3841 sleep_on_head(q
, &wait
, &flags
);
3842 timeout
= schedule_timeout(timeout
);
3843 sleep_on_tail(q
, &wait
, &flags
);
3847 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3849 void __sched
sleep_on(wait_queue_head_t
*q
)
3851 unsigned long flags
;
3854 init_waitqueue_entry(&wait
, current
);
3856 current
->state
= TASK_UNINTERRUPTIBLE
;
3858 sleep_on_head(q
, &wait
, &flags
);
3860 sleep_on_tail(q
, &wait
, &flags
);
3862 EXPORT_SYMBOL(sleep_on
);
3864 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3866 unsigned long flags
;
3869 init_waitqueue_entry(&wait
, current
);
3871 current
->state
= TASK_UNINTERRUPTIBLE
;
3873 sleep_on_head(q
, &wait
, &flags
);
3874 timeout
= schedule_timeout(timeout
);
3875 sleep_on_tail(q
, &wait
, &flags
);
3879 EXPORT_SYMBOL(sleep_on_timeout
);
3881 #ifdef CONFIG_RT_MUTEXES
3884 * rt_mutex_setprio - set the current priority of a task
3886 * @prio: prio value (kernel-internal form)
3888 * This function changes the 'effective' priority of a task. It does
3889 * not touch ->normal_prio like __setscheduler().
3891 * Used by the rt_mutex code to implement priority inheritance logic.
3893 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3895 unsigned long flags
;
3900 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3902 rq
= task_rq_lock(p
, &flags
);
3906 on_rq
= p
->se
.on_rq
;
3908 dequeue_task(rq
, p
, 0, now
);
3911 p
->sched_class
= &rt_sched_class
;
3913 p
->sched_class
= &fair_sched_class
;
3918 enqueue_task(rq
, p
, 0, now
);
3920 * Reschedule if we are currently running on this runqueue and
3921 * our priority decreased, or if we are not currently running on
3922 * this runqueue and our priority is higher than the current's
3924 if (task_running(rq
, p
)) {
3925 if (p
->prio
> oldprio
)
3926 resched_task(rq
->curr
);
3928 check_preempt_curr(rq
, p
);
3931 task_rq_unlock(rq
, &flags
);
3936 void set_user_nice(struct task_struct
*p
, long nice
)
3938 int old_prio
, delta
, on_rq
;
3939 unsigned long flags
;
3943 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3946 * We have to be careful, if called from sys_setpriority(),
3947 * the task might be in the middle of scheduling on another CPU.
3949 rq
= task_rq_lock(p
, &flags
);
3952 * The RT priorities are set via sched_setscheduler(), but we still
3953 * allow the 'normal' nice value to be set - but as expected
3954 * it wont have any effect on scheduling until the task is
3955 * SCHED_FIFO/SCHED_RR:
3957 if (task_has_rt_policy(p
)) {
3958 p
->static_prio
= NICE_TO_PRIO(nice
);
3961 on_rq
= p
->se
.on_rq
;
3963 dequeue_task(rq
, p
, 0, now
);
3964 dec_load(rq
, p
, now
);
3967 p
->static_prio
= NICE_TO_PRIO(nice
);
3970 p
->prio
= effective_prio(p
);
3971 delta
= p
->prio
- old_prio
;
3974 enqueue_task(rq
, p
, 0, now
);
3975 inc_load(rq
, p
, now
);
3977 * If the task increased its priority or is running and
3978 * lowered its priority, then reschedule its CPU:
3980 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3981 resched_task(rq
->curr
);
3984 task_rq_unlock(rq
, &flags
);
3986 EXPORT_SYMBOL(set_user_nice
);
3989 * can_nice - check if a task can reduce its nice value
3993 int can_nice(const struct task_struct
*p
, const int nice
)
3995 /* convert nice value [19,-20] to rlimit style value [1,40] */
3996 int nice_rlim
= 20 - nice
;
3998 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3999 capable(CAP_SYS_NICE
));
4002 #ifdef __ARCH_WANT_SYS_NICE
4005 * sys_nice - change the priority of the current process.
4006 * @increment: priority increment
4008 * sys_setpriority is a more generic, but much slower function that
4009 * does similar things.
4011 asmlinkage
long sys_nice(int increment
)
4016 * Setpriority might change our priority at the same moment.
4017 * We don't have to worry. Conceptually one call occurs first
4018 * and we have a single winner.
4020 if (increment
< -40)
4025 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4031 if (increment
< 0 && !can_nice(current
, nice
))
4034 retval
= security_task_setnice(current
, nice
);
4038 set_user_nice(current
, nice
);
4045 * task_prio - return the priority value of a given task.
4046 * @p: the task in question.
4048 * This is the priority value as seen by users in /proc.
4049 * RT tasks are offset by -200. Normal tasks are centered
4050 * around 0, value goes from -16 to +15.
4052 int task_prio(const struct task_struct
*p
)
4054 return p
->prio
- MAX_RT_PRIO
;
4058 * task_nice - return the nice value of a given task.
4059 * @p: the task in question.
4061 int task_nice(const struct task_struct
*p
)
4063 return TASK_NICE(p
);
4065 EXPORT_SYMBOL_GPL(task_nice
);
4068 * idle_cpu - is a given cpu idle currently?
4069 * @cpu: the processor in question.
4071 int idle_cpu(int cpu
)
4073 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4077 * idle_task - return the idle task for a given cpu.
4078 * @cpu: the processor in question.
4080 struct task_struct
*idle_task(int cpu
)
4082 return cpu_rq(cpu
)->idle
;
4086 * find_process_by_pid - find a process with a matching PID value.
4087 * @pid: the pid in question.
4089 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4091 return pid
? find_task_by_pid(pid
) : current
;
4094 /* Actually do priority change: must hold rq lock. */
4096 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4098 BUG_ON(p
->se
.on_rq
);
4101 switch (p
->policy
) {
4105 p
->sched_class
= &fair_sched_class
;
4109 p
->sched_class
= &rt_sched_class
;
4113 p
->rt_priority
= prio
;
4114 p
->normal_prio
= normal_prio(p
);
4115 /* we are holding p->pi_lock already */
4116 p
->prio
= rt_mutex_getprio(p
);
4121 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4122 * @p: the task in question.
4123 * @policy: new policy.
4124 * @param: structure containing the new RT priority.
4126 * NOTE that the task may be already dead.
4128 int sched_setscheduler(struct task_struct
*p
, int policy
,
4129 struct sched_param
*param
)
4131 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4132 unsigned long flags
;
4135 /* may grab non-irq protected spin_locks */
4136 BUG_ON(in_interrupt());
4138 /* double check policy once rq lock held */
4140 policy
= oldpolicy
= p
->policy
;
4141 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4142 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4143 policy
!= SCHED_IDLE
)
4146 * Valid priorities for SCHED_FIFO and SCHED_RR are
4147 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4148 * SCHED_BATCH and SCHED_IDLE is 0.
4150 if (param
->sched_priority
< 0 ||
4151 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4152 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4154 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4158 * Allow unprivileged RT tasks to decrease priority:
4160 if (!capable(CAP_SYS_NICE
)) {
4161 if (rt_policy(policy
)) {
4162 unsigned long rlim_rtprio
;
4164 if (!lock_task_sighand(p
, &flags
))
4166 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4167 unlock_task_sighand(p
, &flags
);
4169 /* can't set/change the rt policy */
4170 if (policy
!= p
->policy
&& !rlim_rtprio
)
4173 /* can't increase priority */
4174 if (param
->sched_priority
> p
->rt_priority
&&
4175 param
->sched_priority
> rlim_rtprio
)
4179 * Like positive nice levels, dont allow tasks to
4180 * move out of SCHED_IDLE either:
4182 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4185 /* can't change other user's priorities */
4186 if ((current
->euid
!= p
->euid
) &&
4187 (current
->euid
!= p
->uid
))
4191 retval
= security_task_setscheduler(p
, policy
, param
);
4195 * make sure no PI-waiters arrive (or leave) while we are
4196 * changing the priority of the task:
4198 spin_lock_irqsave(&p
->pi_lock
, flags
);
4200 * To be able to change p->policy safely, the apropriate
4201 * runqueue lock must be held.
4203 rq
= __task_rq_lock(p
);
4204 /* recheck policy now with rq lock held */
4205 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4206 policy
= oldpolicy
= -1;
4207 __task_rq_unlock(rq
);
4208 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4211 on_rq
= p
->se
.on_rq
;
4213 deactivate_task(rq
, p
, 0);
4215 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4217 activate_task(rq
, p
, 0);
4219 * Reschedule if we are currently running on this runqueue and
4220 * our priority decreased, or if we are not currently running on
4221 * this runqueue and our priority is higher than the current's
4223 if (task_running(rq
, p
)) {
4224 if (p
->prio
> oldprio
)
4225 resched_task(rq
->curr
);
4227 check_preempt_curr(rq
, p
);
4230 __task_rq_unlock(rq
);
4231 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4233 rt_mutex_adjust_pi(p
);
4237 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4240 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4242 struct sched_param lparam
;
4243 struct task_struct
*p
;
4246 if (!param
|| pid
< 0)
4248 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4253 p
= find_process_by_pid(pid
);
4255 retval
= sched_setscheduler(p
, policy
, &lparam
);
4262 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4263 * @pid: the pid in question.
4264 * @policy: new policy.
4265 * @param: structure containing the new RT priority.
4267 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4268 struct sched_param __user
*param
)
4270 /* negative values for policy are not valid */
4274 return do_sched_setscheduler(pid
, policy
, param
);
4278 * sys_sched_setparam - set/change the RT priority of a thread
4279 * @pid: the pid in question.
4280 * @param: structure containing the new RT priority.
4282 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4284 return do_sched_setscheduler(pid
, -1, param
);
4288 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4289 * @pid: the pid in question.
4291 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4293 struct task_struct
*p
;
4294 int retval
= -EINVAL
;
4300 read_lock(&tasklist_lock
);
4301 p
= find_process_by_pid(pid
);
4303 retval
= security_task_getscheduler(p
);
4307 read_unlock(&tasklist_lock
);
4314 * sys_sched_getscheduler - get the RT priority of a thread
4315 * @pid: the pid in question.
4316 * @param: structure containing the RT priority.
4318 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4320 struct sched_param lp
;
4321 struct task_struct
*p
;
4322 int retval
= -EINVAL
;
4324 if (!param
|| pid
< 0)
4327 read_lock(&tasklist_lock
);
4328 p
= find_process_by_pid(pid
);
4333 retval
= security_task_getscheduler(p
);
4337 lp
.sched_priority
= p
->rt_priority
;
4338 read_unlock(&tasklist_lock
);
4341 * This one might sleep, we cannot do it with a spinlock held ...
4343 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4349 read_unlock(&tasklist_lock
);
4353 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4355 cpumask_t cpus_allowed
;
4356 struct task_struct
*p
;
4359 mutex_lock(&sched_hotcpu_mutex
);
4360 read_lock(&tasklist_lock
);
4362 p
= find_process_by_pid(pid
);
4364 read_unlock(&tasklist_lock
);
4365 mutex_unlock(&sched_hotcpu_mutex
);
4370 * It is not safe to call set_cpus_allowed with the
4371 * tasklist_lock held. We will bump the task_struct's
4372 * usage count and then drop tasklist_lock.
4375 read_unlock(&tasklist_lock
);
4378 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4379 !capable(CAP_SYS_NICE
))
4382 retval
= security_task_setscheduler(p
, 0, NULL
);
4386 cpus_allowed
= cpuset_cpus_allowed(p
);
4387 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4388 retval
= set_cpus_allowed(p
, new_mask
);
4392 mutex_unlock(&sched_hotcpu_mutex
);
4396 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4397 cpumask_t
*new_mask
)
4399 if (len
< sizeof(cpumask_t
)) {
4400 memset(new_mask
, 0, sizeof(cpumask_t
));
4401 } else if (len
> sizeof(cpumask_t
)) {
4402 len
= sizeof(cpumask_t
);
4404 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4408 * sys_sched_setaffinity - set the cpu affinity of a process
4409 * @pid: pid of the process
4410 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4411 * @user_mask_ptr: user-space pointer to the new cpu mask
4413 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4414 unsigned long __user
*user_mask_ptr
)
4419 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4423 return sched_setaffinity(pid
, new_mask
);
4427 * Represents all cpu's present in the system
4428 * In systems capable of hotplug, this map could dynamically grow
4429 * as new cpu's are detected in the system via any platform specific
4430 * method, such as ACPI for e.g.
4433 cpumask_t cpu_present_map __read_mostly
;
4434 EXPORT_SYMBOL(cpu_present_map
);
4437 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4438 EXPORT_SYMBOL(cpu_online_map
);
4440 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4441 EXPORT_SYMBOL(cpu_possible_map
);
4444 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4446 struct task_struct
*p
;
4449 mutex_lock(&sched_hotcpu_mutex
);
4450 read_lock(&tasklist_lock
);
4453 p
= find_process_by_pid(pid
);
4457 retval
= security_task_getscheduler(p
);
4461 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4464 read_unlock(&tasklist_lock
);
4465 mutex_unlock(&sched_hotcpu_mutex
);
4473 * sys_sched_getaffinity - get the cpu affinity of a process
4474 * @pid: pid of the process
4475 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4476 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4478 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4479 unsigned long __user
*user_mask_ptr
)
4484 if (len
< sizeof(cpumask_t
))
4487 ret
= sched_getaffinity(pid
, &mask
);
4491 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4494 return sizeof(cpumask_t
);
4498 * sys_sched_yield - yield the current processor to other threads.
4500 * This function yields the current CPU to other tasks. If there are no
4501 * other threads running on this CPU then this function will return.
4503 asmlinkage
long sys_sched_yield(void)
4505 struct rq
*rq
= this_rq_lock();
4507 schedstat_inc(rq
, yld_cnt
);
4508 if (unlikely(rq
->nr_running
== 1))
4509 schedstat_inc(rq
, yld_act_empty
);
4511 current
->sched_class
->yield_task(rq
, current
);
4514 * Since we are going to call schedule() anyway, there's
4515 * no need to preempt or enable interrupts:
4517 __release(rq
->lock
);
4518 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4519 _raw_spin_unlock(&rq
->lock
);
4520 preempt_enable_no_resched();
4527 static void __cond_resched(void)
4529 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4530 __might_sleep(__FILE__
, __LINE__
);
4533 * The BKS might be reacquired before we have dropped
4534 * PREEMPT_ACTIVE, which could trigger a second
4535 * cond_resched() call.
4538 add_preempt_count(PREEMPT_ACTIVE
);
4540 sub_preempt_count(PREEMPT_ACTIVE
);
4541 } while (need_resched());
4544 int __sched
cond_resched(void)
4546 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4547 system_state
== SYSTEM_RUNNING
) {
4553 EXPORT_SYMBOL(cond_resched
);
4556 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4557 * call schedule, and on return reacquire the lock.
4559 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4560 * operations here to prevent schedule() from being called twice (once via
4561 * spin_unlock(), once by hand).
4563 int cond_resched_lock(spinlock_t
*lock
)
4567 if (need_lockbreak(lock
)) {
4573 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4574 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4575 _raw_spin_unlock(lock
);
4576 preempt_enable_no_resched();
4583 EXPORT_SYMBOL(cond_resched_lock
);
4585 int __sched
cond_resched_softirq(void)
4587 BUG_ON(!in_softirq());
4589 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4597 EXPORT_SYMBOL(cond_resched_softirq
);
4600 * yield - yield the current processor to other threads.
4602 * This is a shortcut for kernel-space yielding - it marks the
4603 * thread runnable and calls sys_sched_yield().
4605 void __sched
yield(void)
4607 set_current_state(TASK_RUNNING
);
4610 EXPORT_SYMBOL(yield
);
4613 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4614 * that process accounting knows that this is a task in IO wait state.
4616 * But don't do that if it is a deliberate, throttling IO wait (this task
4617 * has set its backing_dev_info: the queue against which it should throttle)
4619 void __sched
io_schedule(void)
4621 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4623 delayacct_blkio_start();
4624 atomic_inc(&rq
->nr_iowait
);
4626 atomic_dec(&rq
->nr_iowait
);
4627 delayacct_blkio_end();
4629 EXPORT_SYMBOL(io_schedule
);
4631 long __sched
io_schedule_timeout(long timeout
)
4633 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4636 delayacct_blkio_start();
4637 atomic_inc(&rq
->nr_iowait
);
4638 ret
= schedule_timeout(timeout
);
4639 atomic_dec(&rq
->nr_iowait
);
4640 delayacct_blkio_end();
4645 * sys_sched_get_priority_max - return maximum RT priority.
4646 * @policy: scheduling class.
4648 * this syscall returns the maximum rt_priority that can be used
4649 * by a given scheduling class.
4651 asmlinkage
long sys_sched_get_priority_max(int policy
)
4658 ret
= MAX_USER_RT_PRIO
-1;
4670 * sys_sched_get_priority_min - return minimum RT priority.
4671 * @policy: scheduling class.
4673 * this syscall returns the minimum rt_priority that can be used
4674 * by a given scheduling class.
4676 asmlinkage
long sys_sched_get_priority_min(int policy
)
4694 * sys_sched_rr_get_interval - return the default timeslice of a process.
4695 * @pid: pid of the process.
4696 * @interval: userspace pointer to the timeslice value.
4698 * this syscall writes the default timeslice value of a given process
4699 * into the user-space timespec buffer. A value of '0' means infinity.
4702 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4704 struct task_struct
*p
;
4705 int retval
= -EINVAL
;
4712 read_lock(&tasklist_lock
);
4713 p
= find_process_by_pid(pid
);
4717 retval
= security_task_getscheduler(p
);
4721 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4722 0 : static_prio_timeslice(p
->static_prio
), &t
);
4723 read_unlock(&tasklist_lock
);
4724 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4728 read_unlock(&tasklist_lock
);
4732 static const char stat_nam
[] = "RSDTtZX";
4734 static void show_task(struct task_struct
*p
)
4736 unsigned long free
= 0;
4739 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4740 printk("%-13.13s %c", p
->comm
,
4741 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4742 #if BITS_PER_LONG == 32
4743 if (state
== TASK_RUNNING
)
4744 printk(" running ");
4746 printk(" %08lx ", thread_saved_pc(p
));
4748 if (state
== TASK_RUNNING
)
4749 printk(" running task ");
4751 printk(" %016lx ", thread_saved_pc(p
));
4753 #ifdef CONFIG_DEBUG_STACK_USAGE
4755 unsigned long *n
= end_of_stack(p
);
4758 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4761 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4763 if (state
!= TASK_RUNNING
)
4764 show_stack(p
, NULL
);
4767 void show_state_filter(unsigned long state_filter
)
4769 struct task_struct
*g
, *p
;
4771 #if BITS_PER_LONG == 32
4773 " task PC stack pid father\n");
4776 " task PC stack pid father\n");
4778 read_lock(&tasklist_lock
);
4779 do_each_thread(g
, p
) {
4781 * reset the NMI-timeout, listing all files on a slow
4782 * console might take alot of time:
4784 touch_nmi_watchdog();
4785 if (!state_filter
|| (p
->state
& state_filter
))
4787 } while_each_thread(g
, p
);
4789 touch_all_softlockup_watchdogs();
4791 #ifdef CONFIG_SCHED_DEBUG
4792 sysrq_sched_debug_show();
4794 read_unlock(&tasklist_lock
);
4796 * Only show locks if all tasks are dumped:
4798 if (state_filter
== -1)
4799 debug_show_all_locks();
4802 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4804 idle
->sched_class
= &idle_sched_class
;
4808 * init_idle - set up an idle thread for a given CPU
4809 * @idle: task in question
4810 * @cpu: cpu the idle task belongs to
4812 * NOTE: this function does not set the idle thread's NEED_RESCHED
4813 * flag, to make booting more robust.
4815 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4817 struct rq
*rq
= cpu_rq(cpu
);
4818 unsigned long flags
;
4821 idle
->se
.exec_start
= sched_clock();
4823 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4824 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4825 __set_task_cpu(idle
, cpu
);
4827 spin_lock_irqsave(&rq
->lock
, flags
);
4828 rq
->curr
= rq
->idle
= idle
;
4829 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4832 spin_unlock_irqrestore(&rq
->lock
, flags
);
4834 /* Set the preempt count _outside_ the spinlocks! */
4835 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4836 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4838 task_thread_info(idle
)->preempt_count
= 0;
4841 * The idle tasks have their own, simple scheduling class:
4843 idle
->sched_class
= &idle_sched_class
;
4847 * In a system that switches off the HZ timer nohz_cpu_mask
4848 * indicates which cpus entered this state. This is used
4849 * in the rcu update to wait only for active cpus. For system
4850 * which do not switch off the HZ timer nohz_cpu_mask should
4851 * always be CPU_MASK_NONE.
4853 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4856 * Increase the granularity value when there are more CPUs,
4857 * because with more CPUs the 'effective latency' as visible
4858 * to users decreases. But the relationship is not linear,
4859 * so pick a second-best guess by going with the log2 of the
4862 * This idea comes from the SD scheduler of Con Kolivas:
4864 static inline void sched_init_granularity(void)
4866 unsigned int factor
= 1 + ilog2(num_online_cpus());
4867 const unsigned long gran_limit
= 100000000;
4869 sysctl_sched_granularity
*= factor
;
4870 if (sysctl_sched_granularity
> gran_limit
)
4871 sysctl_sched_granularity
= gran_limit
;
4873 sysctl_sched_runtime_limit
= sysctl_sched_granularity
* 4;
4874 sysctl_sched_wakeup_granularity
= sysctl_sched_granularity
/ 2;
4879 * This is how migration works:
4881 * 1) we queue a struct migration_req structure in the source CPU's
4882 * runqueue and wake up that CPU's migration thread.
4883 * 2) we down() the locked semaphore => thread blocks.
4884 * 3) migration thread wakes up (implicitly it forces the migrated
4885 * thread off the CPU)
4886 * 4) it gets the migration request and checks whether the migrated
4887 * task is still in the wrong runqueue.
4888 * 5) if it's in the wrong runqueue then the migration thread removes
4889 * it and puts it into the right queue.
4890 * 6) migration thread up()s the semaphore.
4891 * 7) we wake up and the migration is done.
4895 * Change a given task's CPU affinity. Migrate the thread to a
4896 * proper CPU and schedule it away if the CPU it's executing on
4897 * is removed from the allowed bitmask.
4899 * NOTE: the caller must have a valid reference to the task, the
4900 * task must not exit() & deallocate itself prematurely. The
4901 * call is not atomic; no spinlocks may be held.
4903 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4905 struct migration_req req
;
4906 unsigned long flags
;
4910 rq
= task_rq_lock(p
, &flags
);
4911 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4916 p
->cpus_allowed
= new_mask
;
4917 /* Can the task run on the task's current CPU? If so, we're done */
4918 if (cpu_isset(task_cpu(p
), new_mask
))
4921 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4922 /* Need help from migration thread: drop lock and wait. */
4923 task_rq_unlock(rq
, &flags
);
4924 wake_up_process(rq
->migration_thread
);
4925 wait_for_completion(&req
.done
);
4926 tlb_migrate_finish(p
->mm
);
4930 task_rq_unlock(rq
, &flags
);
4934 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4937 * Move (not current) task off this cpu, onto dest cpu. We're doing
4938 * this because either it can't run here any more (set_cpus_allowed()
4939 * away from this CPU, or CPU going down), or because we're
4940 * attempting to rebalance this task on exec (sched_exec).
4942 * So we race with normal scheduler movements, but that's OK, as long
4943 * as the task is no longer on this CPU.
4945 * Returns non-zero if task was successfully migrated.
4947 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4949 struct rq
*rq_dest
, *rq_src
;
4952 if (unlikely(cpu_is_offline(dest_cpu
)))
4955 rq_src
= cpu_rq(src_cpu
);
4956 rq_dest
= cpu_rq(dest_cpu
);
4958 double_rq_lock(rq_src
, rq_dest
);
4959 /* Already moved. */
4960 if (task_cpu(p
) != src_cpu
)
4962 /* Affinity changed (again). */
4963 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4966 on_rq
= p
->se
.on_rq
;
4968 deactivate_task(rq_src
, p
, 0);
4969 set_task_cpu(p
, dest_cpu
);
4971 activate_task(rq_dest
, p
, 0);
4972 check_preempt_curr(rq_dest
, p
);
4976 double_rq_unlock(rq_src
, rq_dest
);
4981 * migration_thread - this is a highprio system thread that performs
4982 * thread migration by bumping thread off CPU then 'pushing' onto
4985 static int migration_thread(void *data
)
4987 int cpu
= (long)data
;
4991 BUG_ON(rq
->migration_thread
!= current
);
4993 set_current_state(TASK_INTERRUPTIBLE
);
4994 while (!kthread_should_stop()) {
4995 struct migration_req
*req
;
4996 struct list_head
*head
;
4998 spin_lock_irq(&rq
->lock
);
5000 if (cpu_is_offline(cpu
)) {
5001 spin_unlock_irq(&rq
->lock
);
5005 if (rq
->active_balance
) {
5006 active_load_balance(rq
, cpu
);
5007 rq
->active_balance
= 0;
5010 head
= &rq
->migration_queue
;
5012 if (list_empty(head
)) {
5013 spin_unlock_irq(&rq
->lock
);
5015 set_current_state(TASK_INTERRUPTIBLE
);
5018 req
= list_entry(head
->next
, struct migration_req
, list
);
5019 list_del_init(head
->next
);
5021 spin_unlock(&rq
->lock
);
5022 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5025 complete(&req
->done
);
5027 __set_current_state(TASK_RUNNING
);
5031 /* Wait for kthread_stop */
5032 set_current_state(TASK_INTERRUPTIBLE
);
5033 while (!kthread_should_stop()) {
5035 set_current_state(TASK_INTERRUPTIBLE
);
5037 __set_current_state(TASK_RUNNING
);
5041 #ifdef CONFIG_HOTPLUG_CPU
5043 * Figure out where task on dead CPU should go, use force if neccessary.
5044 * NOTE: interrupts should be disabled by the caller
5046 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5048 unsigned long flags
;
5055 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5056 cpus_and(mask
, mask
, p
->cpus_allowed
);
5057 dest_cpu
= any_online_cpu(mask
);
5059 /* On any allowed CPU? */
5060 if (dest_cpu
== NR_CPUS
)
5061 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5063 /* No more Mr. Nice Guy. */
5064 if (dest_cpu
== NR_CPUS
) {
5065 rq
= task_rq_lock(p
, &flags
);
5066 cpus_setall(p
->cpus_allowed
);
5067 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5068 task_rq_unlock(rq
, &flags
);
5071 * Don't tell them about moving exiting tasks or
5072 * kernel threads (both mm NULL), since they never
5075 if (p
->mm
&& printk_ratelimit())
5076 printk(KERN_INFO
"process %d (%s) no "
5077 "longer affine to cpu%d\n",
5078 p
->pid
, p
->comm
, dead_cpu
);
5080 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5085 * While a dead CPU has no uninterruptible tasks queued at this point,
5086 * it might still have a nonzero ->nr_uninterruptible counter, because
5087 * for performance reasons the counter is not stricly tracking tasks to
5088 * their home CPUs. So we just add the counter to another CPU's counter,
5089 * to keep the global sum constant after CPU-down:
5091 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5093 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5094 unsigned long flags
;
5096 local_irq_save(flags
);
5097 double_rq_lock(rq_src
, rq_dest
);
5098 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5099 rq_src
->nr_uninterruptible
= 0;
5100 double_rq_unlock(rq_src
, rq_dest
);
5101 local_irq_restore(flags
);
5104 /* Run through task list and migrate tasks from the dead cpu. */
5105 static void migrate_live_tasks(int src_cpu
)
5107 struct task_struct
*p
, *t
;
5109 write_lock_irq(&tasklist_lock
);
5111 do_each_thread(t
, p
) {
5115 if (task_cpu(p
) == src_cpu
)
5116 move_task_off_dead_cpu(src_cpu
, p
);
5117 } while_each_thread(t
, p
);
5119 write_unlock_irq(&tasklist_lock
);
5123 * Schedules idle task to be the next runnable task on current CPU.
5124 * It does so by boosting its priority to highest possible and adding it to
5125 * the _front_ of the runqueue. Used by CPU offline code.
5127 void sched_idle_next(void)
5129 int this_cpu
= smp_processor_id();
5130 struct rq
*rq
= cpu_rq(this_cpu
);
5131 struct task_struct
*p
= rq
->idle
;
5132 unsigned long flags
;
5134 /* cpu has to be offline */
5135 BUG_ON(cpu_online(this_cpu
));
5138 * Strictly not necessary since rest of the CPUs are stopped by now
5139 * and interrupts disabled on the current cpu.
5141 spin_lock_irqsave(&rq
->lock
, flags
);
5143 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5145 /* Add idle task to the _front_ of its priority queue: */
5146 activate_idle_task(p
, rq
);
5148 spin_unlock_irqrestore(&rq
->lock
, flags
);
5152 * Ensures that the idle task is using init_mm right before its cpu goes
5155 void idle_task_exit(void)
5157 struct mm_struct
*mm
= current
->active_mm
;
5159 BUG_ON(cpu_online(smp_processor_id()));
5162 switch_mm(mm
, &init_mm
, current
);
5166 /* called under rq->lock with disabled interrupts */
5167 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5169 struct rq
*rq
= cpu_rq(dead_cpu
);
5171 /* Must be exiting, otherwise would be on tasklist. */
5172 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5174 /* Cannot have done final schedule yet: would have vanished. */
5175 BUG_ON(p
->state
== TASK_DEAD
);
5180 * Drop lock around migration; if someone else moves it,
5181 * that's OK. No task can be added to this CPU, so iteration is
5183 * NOTE: interrupts should be left disabled --dev@
5185 spin_unlock(&rq
->lock
);
5186 move_task_off_dead_cpu(dead_cpu
, p
);
5187 spin_lock(&rq
->lock
);
5192 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5193 static void migrate_dead_tasks(unsigned int dead_cpu
)
5195 struct rq
*rq
= cpu_rq(dead_cpu
);
5196 struct task_struct
*next
;
5199 if (!rq
->nr_running
)
5201 next
= pick_next_task(rq
, rq
->curr
, rq_clock(rq
));
5204 migrate_dead(dead_cpu
, next
);
5208 #endif /* CONFIG_HOTPLUG_CPU */
5210 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5212 static struct ctl_table sd_ctl_dir
[] = {
5213 {CTL_UNNUMBERED
, "sched_domain", NULL
, 0, 0755, NULL
, },
5217 static struct ctl_table sd_ctl_root
[] = {
5218 {CTL_UNNUMBERED
, "kernel", NULL
, 0, 0755, sd_ctl_dir
, },
5222 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5224 struct ctl_table
*entry
=
5225 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5228 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5234 set_table_entry(struct ctl_table
*entry
, int ctl_name
,
5235 const char *procname
, void *data
, int maxlen
,
5236 mode_t mode
, proc_handler
*proc_handler
)
5238 entry
->ctl_name
= ctl_name
;
5239 entry
->procname
= procname
;
5241 entry
->maxlen
= maxlen
;
5243 entry
->proc_handler
= proc_handler
;
5246 static struct ctl_table
*
5247 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5249 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5251 set_table_entry(&table
[0], 1, "min_interval", &sd
->min_interval
,
5252 sizeof(long), 0644, proc_doulongvec_minmax
);
5253 set_table_entry(&table
[1], 2, "max_interval", &sd
->max_interval
,
5254 sizeof(long), 0644, proc_doulongvec_minmax
);
5255 set_table_entry(&table
[2], 3, "busy_idx", &sd
->busy_idx
,
5256 sizeof(int), 0644, proc_dointvec_minmax
);
5257 set_table_entry(&table
[3], 4, "idle_idx", &sd
->idle_idx
,
5258 sizeof(int), 0644, proc_dointvec_minmax
);
5259 set_table_entry(&table
[4], 5, "newidle_idx", &sd
->newidle_idx
,
5260 sizeof(int), 0644, proc_dointvec_minmax
);
5261 set_table_entry(&table
[5], 6, "wake_idx", &sd
->wake_idx
,
5262 sizeof(int), 0644, proc_dointvec_minmax
);
5263 set_table_entry(&table
[6], 7, "forkexec_idx", &sd
->forkexec_idx
,
5264 sizeof(int), 0644, proc_dointvec_minmax
);
5265 set_table_entry(&table
[7], 8, "busy_factor", &sd
->busy_factor
,
5266 sizeof(int), 0644, proc_dointvec_minmax
);
5267 set_table_entry(&table
[8], 9, "imbalance_pct", &sd
->imbalance_pct
,
5268 sizeof(int), 0644, proc_dointvec_minmax
);
5269 set_table_entry(&table
[10], 11, "cache_nice_tries",
5270 &sd
->cache_nice_tries
,
5271 sizeof(int), 0644, proc_dointvec_minmax
);
5272 set_table_entry(&table
[12], 13, "flags", &sd
->flags
,
5273 sizeof(int), 0644, proc_dointvec_minmax
);
5278 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5280 struct ctl_table
*entry
, *table
;
5281 struct sched_domain
*sd
;
5282 int domain_num
= 0, i
;
5285 for_each_domain(cpu
, sd
)
5287 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5290 for_each_domain(cpu
, sd
) {
5291 snprintf(buf
, 32, "domain%d", i
);
5292 entry
->ctl_name
= i
+ 1;
5293 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5295 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5302 static struct ctl_table_header
*sd_sysctl_header
;
5303 static void init_sched_domain_sysctl(void)
5305 int i
, cpu_num
= num_online_cpus();
5306 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5309 sd_ctl_dir
[0].child
= entry
;
5311 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5312 snprintf(buf
, 32, "cpu%d", i
);
5313 entry
->ctl_name
= i
+ 1;
5314 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5316 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5318 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5321 static void init_sched_domain_sysctl(void)
5327 * migration_call - callback that gets triggered when a CPU is added.
5328 * Here we can start up the necessary migration thread for the new CPU.
5330 static int __cpuinit
5331 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5333 struct task_struct
*p
;
5334 int cpu
= (long)hcpu
;
5335 unsigned long flags
;
5339 case CPU_LOCK_ACQUIRE
:
5340 mutex_lock(&sched_hotcpu_mutex
);
5343 case CPU_UP_PREPARE
:
5344 case CPU_UP_PREPARE_FROZEN
:
5345 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5348 kthread_bind(p
, cpu
);
5349 /* Must be high prio: stop_machine expects to yield to it. */
5350 rq
= task_rq_lock(p
, &flags
);
5351 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5352 task_rq_unlock(rq
, &flags
);
5353 cpu_rq(cpu
)->migration_thread
= p
;
5357 case CPU_ONLINE_FROZEN
:
5358 /* Strictly unneccessary, as first user will wake it. */
5359 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5362 #ifdef CONFIG_HOTPLUG_CPU
5363 case CPU_UP_CANCELED
:
5364 case CPU_UP_CANCELED_FROZEN
:
5365 if (!cpu_rq(cpu
)->migration_thread
)
5367 /* Unbind it from offline cpu so it can run. Fall thru. */
5368 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5369 any_online_cpu(cpu_online_map
));
5370 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5371 cpu_rq(cpu
)->migration_thread
= NULL
;
5375 case CPU_DEAD_FROZEN
:
5376 migrate_live_tasks(cpu
);
5378 kthread_stop(rq
->migration_thread
);
5379 rq
->migration_thread
= NULL
;
5380 /* Idle task back to normal (off runqueue, low prio) */
5381 rq
= task_rq_lock(rq
->idle
, &flags
);
5382 deactivate_task(rq
, rq
->idle
, 0);
5383 rq
->idle
->static_prio
= MAX_PRIO
;
5384 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5385 rq
->idle
->sched_class
= &idle_sched_class
;
5386 migrate_dead_tasks(cpu
);
5387 task_rq_unlock(rq
, &flags
);
5388 migrate_nr_uninterruptible(rq
);
5389 BUG_ON(rq
->nr_running
!= 0);
5391 /* No need to migrate the tasks: it was best-effort if
5392 * they didn't take sched_hotcpu_mutex. Just wake up
5393 * the requestors. */
5394 spin_lock_irq(&rq
->lock
);
5395 while (!list_empty(&rq
->migration_queue
)) {
5396 struct migration_req
*req
;
5398 req
= list_entry(rq
->migration_queue
.next
,
5399 struct migration_req
, list
);
5400 list_del_init(&req
->list
);
5401 complete(&req
->done
);
5403 spin_unlock_irq(&rq
->lock
);
5406 case CPU_LOCK_RELEASE
:
5407 mutex_unlock(&sched_hotcpu_mutex
);
5413 /* Register at highest priority so that task migration (migrate_all_tasks)
5414 * happens before everything else.
5416 static struct notifier_block __cpuinitdata migration_notifier
= {
5417 .notifier_call
= migration_call
,
5421 int __init
migration_init(void)
5423 void *cpu
= (void *)(long)smp_processor_id();
5426 /* Start one for the boot CPU: */
5427 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5428 BUG_ON(err
== NOTIFY_BAD
);
5429 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5430 register_cpu_notifier(&migration_notifier
);
5438 /* Number of possible processor ids */
5439 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5440 EXPORT_SYMBOL(nr_cpu_ids
);
5442 #undef SCHED_DOMAIN_DEBUG
5443 #ifdef SCHED_DOMAIN_DEBUG
5444 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5449 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5453 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5458 struct sched_group
*group
= sd
->groups
;
5459 cpumask_t groupmask
;
5461 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5462 cpus_clear(groupmask
);
5465 for (i
= 0; i
< level
+ 1; i
++)
5467 printk("domain %d: ", level
);
5469 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5470 printk("does not load-balance\n");
5472 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5477 printk("span %s\n", str
);
5479 if (!cpu_isset(cpu
, sd
->span
))
5480 printk(KERN_ERR
"ERROR: domain->span does not contain "
5482 if (!cpu_isset(cpu
, group
->cpumask
))
5483 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5487 for (i
= 0; i
< level
+ 2; i
++)
5493 printk(KERN_ERR
"ERROR: group is NULL\n");
5497 if (!group
->__cpu_power
) {
5499 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5503 if (!cpus_weight(group
->cpumask
)) {
5505 printk(KERN_ERR
"ERROR: empty group\n");
5508 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5510 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5513 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5515 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5518 group
= group
->next
;
5519 } while (group
!= sd
->groups
);
5522 if (!cpus_equal(sd
->span
, groupmask
))
5523 printk(KERN_ERR
"ERROR: groups don't span "
5531 if (!cpus_subset(groupmask
, sd
->span
))
5532 printk(KERN_ERR
"ERROR: parent span is not a superset "
5533 "of domain->span\n");
5538 # define sched_domain_debug(sd, cpu) do { } while (0)
5541 static int sd_degenerate(struct sched_domain
*sd
)
5543 if (cpus_weight(sd
->span
) == 1)
5546 /* Following flags need at least 2 groups */
5547 if (sd
->flags
& (SD_LOAD_BALANCE
|
5548 SD_BALANCE_NEWIDLE
|
5552 SD_SHARE_PKG_RESOURCES
)) {
5553 if (sd
->groups
!= sd
->groups
->next
)
5557 /* Following flags don't use groups */
5558 if (sd
->flags
& (SD_WAKE_IDLE
|
5567 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5569 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5571 if (sd_degenerate(parent
))
5574 if (!cpus_equal(sd
->span
, parent
->span
))
5577 /* Does parent contain flags not in child? */
5578 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5579 if (cflags
& SD_WAKE_AFFINE
)
5580 pflags
&= ~SD_WAKE_BALANCE
;
5581 /* Flags needing groups don't count if only 1 group in parent */
5582 if (parent
->groups
== parent
->groups
->next
) {
5583 pflags
&= ~(SD_LOAD_BALANCE
|
5584 SD_BALANCE_NEWIDLE
|
5588 SD_SHARE_PKG_RESOURCES
);
5590 if (~cflags
& pflags
)
5597 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5598 * hold the hotplug lock.
5600 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5602 struct rq
*rq
= cpu_rq(cpu
);
5603 struct sched_domain
*tmp
;
5605 /* Remove the sched domains which do not contribute to scheduling. */
5606 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5607 struct sched_domain
*parent
= tmp
->parent
;
5610 if (sd_parent_degenerate(tmp
, parent
)) {
5611 tmp
->parent
= parent
->parent
;
5613 parent
->parent
->child
= tmp
;
5617 if (sd
&& sd_degenerate(sd
)) {
5623 sched_domain_debug(sd
, cpu
);
5625 rcu_assign_pointer(rq
->sd
, sd
);
5628 /* cpus with isolated domains */
5629 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5631 /* Setup the mask of cpus configured for isolated domains */
5632 static int __init
isolated_cpu_setup(char *str
)
5634 int ints
[NR_CPUS
], i
;
5636 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5637 cpus_clear(cpu_isolated_map
);
5638 for (i
= 1; i
<= ints
[0]; i
++)
5639 if (ints
[i
] < NR_CPUS
)
5640 cpu_set(ints
[i
], cpu_isolated_map
);
5644 __setup ("isolcpus=", isolated_cpu_setup
);
5647 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5648 * to a function which identifies what group(along with sched group) a CPU
5649 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5650 * (due to the fact that we keep track of groups covered with a cpumask_t).
5652 * init_sched_build_groups will build a circular linked list of the groups
5653 * covered by the given span, and will set each group's ->cpumask correctly,
5654 * and ->cpu_power to 0.
5657 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5658 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5659 struct sched_group
**sg
))
5661 struct sched_group
*first
= NULL
, *last
= NULL
;
5662 cpumask_t covered
= CPU_MASK_NONE
;
5665 for_each_cpu_mask(i
, span
) {
5666 struct sched_group
*sg
;
5667 int group
= group_fn(i
, cpu_map
, &sg
);
5670 if (cpu_isset(i
, covered
))
5673 sg
->cpumask
= CPU_MASK_NONE
;
5674 sg
->__cpu_power
= 0;
5676 for_each_cpu_mask(j
, span
) {
5677 if (group_fn(j
, cpu_map
, NULL
) != group
)
5680 cpu_set(j
, covered
);
5681 cpu_set(j
, sg
->cpumask
);
5692 #define SD_NODES_PER_DOMAIN 16
5697 * find_next_best_node - find the next node to include in a sched_domain
5698 * @node: node whose sched_domain we're building
5699 * @used_nodes: nodes already in the sched_domain
5701 * Find the next node to include in a given scheduling domain. Simply
5702 * finds the closest node not already in the @used_nodes map.
5704 * Should use nodemask_t.
5706 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5708 int i
, n
, val
, min_val
, best_node
= 0;
5712 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5713 /* Start at @node */
5714 n
= (node
+ i
) % MAX_NUMNODES
;
5716 if (!nr_cpus_node(n
))
5719 /* Skip already used nodes */
5720 if (test_bit(n
, used_nodes
))
5723 /* Simple min distance search */
5724 val
= node_distance(node
, n
);
5726 if (val
< min_val
) {
5732 set_bit(best_node
, used_nodes
);
5737 * sched_domain_node_span - get a cpumask for a node's sched_domain
5738 * @node: node whose cpumask we're constructing
5739 * @size: number of nodes to include in this span
5741 * Given a node, construct a good cpumask for its sched_domain to span. It
5742 * should be one that prevents unnecessary balancing, but also spreads tasks
5745 static cpumask_t
sched_domain_node_span(int node
)
5747 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5748 cpumask_t span
, nodemask
;
5752 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5754 nodemask
= node_to_cpumask(node
);
5755 cpus_or(span
, span
, nodemask
);
5756 set_bit(node
, used_nodes
);
5758 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5759 int next_node
= find_next_best_node(node
, used_nodes
);
5761 nodemask
= node_to_cpumask(next_node
);
5762 cpus_or(span
, span
, nodemask
);
5769 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5772 * SMT sched-domains:
5774 #ifdef CONFIG_SCHED_SMT
5775 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5776 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5778 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5779 struct sched_group
**sg
)
5782 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5788 * multi-core sched-domains:
5790 #ifdef CONFIG_SCHED_MC
5791 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5792 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5795 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5796 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5797 struct sched_group
**sg
)
5800 cpumask_t mask
= cpu_sibling_map
[cpu
];
5801 cpus_and(mask
, mask
, *cpu_map
);
5802 group
= first_cpu(mask
);
5804 *sg
= &per_cpu(sched_group_core
, group
);
5807 #elif defined(CONFIG_SCHED_MC)
5808 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5809 struct sched_group
**sg
)
5812 *sg
= &per_cpu(sched_group_core
, cpu
);
5817 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5818 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5820 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5821 struct sched_group
**sg
)
5824 #ifdef CONFIG_SCHED_MC
5825 cpumask_t mask
= cpu_coregroup_map(cpu
);
5826 cpus_and(mask
, mask
, *cpu_map
);
5827 group
= first_cpu(mask
);
5828 #elif defined(CONFIG_SCHED_SMT)
5829 cpumask_t mask
= cpu_sibling_map
[cpu
];
5830 cpus_and(mask
, mask
, *cpu_map
);
5831 group
= first_cpu(mask
);
5836 *sg
= &per_cpu(sched_group_phys
, group
);
5842 * The init_sched_build_groups can't handle what we want to do with node
5843 * groups, so roll our own. Now each node has its own list of groups which
5844 * gets dynamically allocated.
5846 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5847 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5849 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5850 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5852 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5853 struct sched_group
**sg
)
5855 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5858 cpus_and(nodemask
, nodemask
, *cpu_map
);
5859 group
= first_cpu(nodemask
);
5862 *sg
= &per_cpu(sched_group_allnodes
, group
);
5866 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5868 struct sched_group
*sg
= group_head
;
5874 for_each_cpu_mask(j
, sg
->cpumask
) {
5875 struct sched_domain
*sd
;
5877 sd
= &per_cpu(phys_domains
, j
);
5878 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5880 * Only add "power" once for each
5886 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5889 if (sg
!= group_head
)
5895 /* Free memory allocated for various sched_group structures */
5896 static void free_sched_groups(const cpumask_t
*cpu_map
)
5900 for_each_cpu_mask(cpu
, *cpu_map
) {
5901 struct sched_group
**sched_group_nodes
5902 = sched_group_nodes_bycpu
[cpu
];
5904 if (!sched_group_nodes
)
5907 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5908 cpumask_t nodemask
= node_to_cpumask(i
);
5909 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5911 cpus_and(nodemask
, nodemask
, *cpu_map
);
5912 if (cpus_empty(nodemask
))
5922 if (oldsg
!= sched_group_nodes
[i
])
5925 kfree(sched_group_nodes
);
5926 sched_group_nodes_bycpu
[cpu
] = NULL
;
5930 static void free_sched_groups(const cpumask_t
*cpu_map
)
5936 * Initialize sched groups cpu_power.
5938 * cpu_power indicates the capacity of sched group, which is used while
5939 * distributing the load between different sched groups in a sched domain.
5940 * Typically cpu_power for all the groups in a sched domain will be same unless
5941 * there are asymmetries in the topology. If there are asymmetries, group
5942 * having more cpu_power will pickup more load compared to the group having
5945 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5946 * the maximum number of tasks a group can handle in the presence of other idle
5947 * or lightly loaded groups in the same sched domain.
5949 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5951 struct sched_domain
*child
;
5952 struct sched_group
*group
;
5954 WARN_ON(!sd
|| !sd
->groups
);
5956 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5961 sd
->groups
->__cpu_power
= 0;
5964 * For perf policy, if the groups in child domain share resources
5965 * (for example cores sharing some portions of the cache hierarchy
5966 * or SMT), then set this domain groups cpu_power such that each group
5967 * can handle only one task, when there are other idle groups in the
5968 * same sched domain.
5970 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5972 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5973 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5978 * add cpu_power of each child group to this groups cpu_power
5980 group
= child
->groups
;
5982 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5983 group
= group
->next
;
5984 } while (group
!= child
->groups
);
5988 * Build sched domains for a given set of cpus and attach the sched domains
5989 * to the individual cpus
5991 static int build_sched_domains(const cpumask_t
*cpu_map
)
5995 struct sched_group
**sched_group_nodes
= NULL
;
5996 int sd_allnodes
= 0;
5999 * Allocate the per-node list of sched groups
6001 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6003 if (!sched_group_nodes
) {
6004 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6007 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6011 * Set up domains for cpus specified by the cpu_map.
6013 for_each_cpu_mask(i
, *cpu_map
) {
6014 struct sched_domain
*sd
= NULL
, *p
;
6015 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6017 cpus_and(nodemask
, nodemask
, *cpu_map
);
6020 if (cpus_weight(*cpu_map
) >
6021 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6022 sd
= &per_cpu(allnodes_domains
, i
);
6023 *sd
= SD_ALLNODES_INIT
;
6024 sd
->span
= *cpu_map
;
6025 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6031 sd
= &per_cpu(node_domains
, i
);
6033 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6037 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6041 sd
= &per_cpu(phys_domains
, i
);
6043 sd
->span
= nodemask
;
6047 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6049 #ifdef CONFIG_SCHED_MC
6051 sd
= &per_cpu(core_domains
, i
);
6053 sd
->span
= cpu_coregroup_map(i
);
6054 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6057 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6060 #ifdef CONFIG_SCHED_SMT
6062 sd
= &per_cpu(cpu_domains
, i
);
6063 *sd
= SD_SIBLING_INIT
;
6064 sd
->span
= cpu_sibling_map
[i
];
6065 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6068 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6072 #ifdef CONFIG_SCHED_SMT
6073 /* Set up CPU (sibling) groups */
6074 for_each_cpu_mask(i
, *cpu_map
) {
6075 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6076 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6077 if (i
!= first_cpu(this_sibling_map
))
6080 init_sched_build_groups(this_sibling_map
, cpu_map
,
6085 #ifdef CONFIG_SCHED_MC
6086 /* Set up multi-core groups */
6087 for_each_cpu_mask(i
, *cpu_map
) {
6088 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6089 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6090 if (i
!= first_cpu(this_core_map
))
6092 init_sched_build_groups(this_core_map
, cpu_map
,
6093 &cpu_to_core_group
);
6097 /* Set up physical groups */
6098 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6099 cpumask_t nodemask
= node_to_cpumask(i
);
6101 cpus_and(nodemask
, nodemask
, *cpu_map
);
6102 if (cpus_empty(nodemask
))
6105 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6109 /* Set up node groups */
6111 init_sched_build_groups(*cpu_map
, cpu_map
,
6112 &cpu_to_allnodes_group
);
6114 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6115 /* Set up node groups */
6116 struct sched_group
*sg
, *prev
;
6117 cpumask_t nodemask
= node_to_cpumask(i
);
6118 cpumask_t domainspan
;
6119 cpumask_t covered
= CPU_MASK_NONE
;
6122 cpus_and(nodemask
, nodemask
, *cpu_map
);
6123 if (cpus_empty(nodemask
)) {
6124 sched_group_nodes
[i
] = NULL
;
6128 domainspan
= sched_domain_node_span(i
);
6129 cpus_and(domainspan
, domainspan
, *cpu_map
);
6131 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6133 printk(KERN_WARNING
"Can not alloc domain group for "
6137 sched_group_nodes
[i
] = sg
;
6138 for_each_cpu_mask(j
, nodemask
) {
6139 struct sched_domain
*sd
;
6141 sd
= &per_cpu(node_domains
, j
);
6144 sg
->__cpu_power
= 0;
6145 sg
->cpumask
= nodemask
;
6147 cpus_or(covered
, covered
, nodemask
);
6150 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6151 cpumask_t tmp
, notcovered
;
6152 int n
= (i
+ j
) % MAX_NUMNODES
;
6154 cpus_complement(notcovered
, covered
);
6155 cpus_and(tmp
, notcovered
, *cpu_map
);
6156 cpus_and(tmp
, tmp
, domainspan
);
6157 if (cpus_empty(tmp
))
6160 nodemask
= node_to_cpumask(n
);
6161 cpus_and(tmp
, tmp
, nodemask
);
6162 if (cpus_empty(tmp
))
6165 sg
= kmalloc_node(sizeof(struct sched_group
),
6169 "Can not alloc domain group for node %d\n", j
);
6172 sg
->__cpu_power
= 0;
6174 sg
->next
= prev
->next
;
6175 cpus_or(covered
, covered
, tmp
);
6182 /* Calculate CPU power for physical packages and nodes */
6183 #ifdef CONFIG_SCHED_SMT
6184 for_each_cpu_mask(i
, *cpu_map
) {
6185 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6187 init_sched_groups_power(i
, sd
);
6190 #ifdef CONFIG_SCHED_MC
6191 for_each_cpu_mask(i
, *cpu_map
) {
6192 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6194 init_sched_groups_power(i
, sd
);
6198 for_each_cpu_mask(i
, *cpu_map
) {
6199 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6201 init_sched_groups_power(i
, sd
);
6205 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6206 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6209 struct sched_group
*sg
;
6211 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6212 init_numa_sched_groups_power(sg
);
6216 /* Attach the domains */
6217 for_each_cpu_mask(i
, *cpu_map
) {
6218 struct sched_domain
*sd
;
6219 #ifdef CONFIG_SCHED_SMT
6220 sd
= &per_cpu(cpu_domains
, i
);
6221 #elif defined(CONFIG_SCHED_MC)
6222 sd
= &per_cpu(core_domains
, i
);
6224 sd
= &per_cpu(phys_domains
, i
);
6226 cpu_attach_domain(sd
, i
);
6233 free_sched_groups(cpu_map
);
6238 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6240 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6242 cpumask_t cpu_default_map
;
6246 * Setup mask for cpus without special case scheduling requirements.
6247 * For now this just excludes isolated cpus, but could be used to
6248 * exclude other special cases in the future.
6250 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6252 err
= build_sched_domains(&cpu_default_map
);
6257 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6259 free_sched_groups(cpu_map
);
6263 * Detach sched domains from a group of cpus specified in cpu_map
6264 * These cpus will now be attached to the NULL domain
6266 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6270 for_each_cpu_mask(i
, *cpu_map
)
6271 cpu_attach_domain(NULL
, i
);
6272 synchronize_sched();
6273 arch_destroy_sched_domains(cpu_map
);
6277 * Partition sched domains as specified by the cpumasks below.
6278 * This attaches all cpus from the cpumasks to the NULL domain,
6279 * waits for a RCU quiescent period, recalculates sched
6280 * domain information and then attaches them back to the
6281 * correct sched domains
6282 * Call with hotplug lock held
6284 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6286 cpumask_t change_map
;
6289 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6290 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6291 cpus_or(change_map
, *partition1
, *partition2
);
6293 /* Detach sched domains from all of the affected cpus */
6294 detach_destroy_domains(&change_map
);
6295 if (!cpus_empty(*partition1
))
6296 err
= build_sched_domains(partition1
);
6297 if (!err
&& !cpus_empty(*partition2
))
6298 err
= build_sched_domains(partition2
);
6303 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6304 int arch_reinit_sched_domains(void)
6308 mutex_lock(&sched_hotcpu_mutex
);
6309 detach_destroy_domains(&cpu_online_map
);
6310 err
= arch_init_sched_domains(&cpu_online_map
);
6311 mutex_unlock(&sched_hotcpu_mutex
);
6316 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6320 if (buf
[0] != '0' && buf
[0] != '1')
6324 sched_smt_power_savings
= (buf
[0] == '1');
6326 sched_mc_power_savings
= (buf
[0] == '1');
6328 ret
= arch_reinit_sched_domains();
6330 return ret
? ret
: count
;
6333 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6337 #ifdef CONFIG_SCHED_SMT
6339 err
= sysfs_create_file(&cls
->kset
.kobj
,
6340 &attr_sched_smt_power_savings
.attr
);
6342 #ifdef CONFIG_SCHED_MC
6343 if (!err
&& mc_capable())
6344 err
= sysfs_create_file(&cls
->kset
.kobj
,
6345 &attr_sched_mc_power_savings
.attr
);
6351 #ifdef CONFIG_SCHED_MC
6352 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6354 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6356 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6357 const char *buf
, size_t count
)
6359 return sched_power_savings_store(buf
, count
, 0);
6361 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6362 sched_mc_power_savings_store
);
6365 #ifdef CONFIG_SCHED_SMT
6366 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6368 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6370 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6371 const char *buf
, size_t count
)
6373 return sched_power_savings_store(buf
, count
, 1);
6375 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6376 sched_smt_power_savings_store
);
6380 * Force a reinitialization of the sched domains hierarchy. The domains
6381 * and groups cannot be updated in place without racing with the balancing
6382 * code, so we temporarily attach all running cpus to the NULL domain
6383 * which will prevent rebalancing while the sched domains are recalculated.
6385 static int update_sched_domains(struct notifier_block
*nfb
,
6386 unsigned long action
, void *hcpu
)
6389 case CPU_UP_PREPARE
:
6390 case CPU_UP_PREPARE_FROZEN
:
6391 case CPU_DOWN_PREPARE
:
6392 case CPU_DOWN_PREPARE_FROZEN
:
6393 detach_destroy_domains(&cpu_online_map
);
6396 case CPU_UP_CANCELED
:
6397 case CPU_UP_CANCELED_FROZEN
:
6398 case CPU_DOWN_FAILED
:
6399 case CPU_DOWN_FAILED_FROZEN
:
6401 case CPU_ONLINE_FROZEN
:
6403 case CPU_DEAD_FROZEN
:
6405 * Fall through and re-initialise the domains.
6412 /* The hotplug lock is already held by cpu_up/cpu_down */
6413 arch_init_sched_domains(&cpu_online_map
);
6418 void __init
sched_init_smp(void)
6420 cpumask_t non_isolated_cpus
;
6422 mutex_lock(&sched_hotcpu_mutex
);
6423 arch_init_sched_domains(&cpu_online_map
);
6424 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6425 if (cpus_empty(non_isolated_cpus
))
6426 cpu_set(smp_processor_id(), non_isolated_cpus
);
6427 mutex_unlock(&sched_hotcpu_mutex
);
6428 /* XXX: Theoretical race here - CPU may be hotplugged now */
6429 hotcpu_notifier(update_sched_domains
, 0);
6431 init_sched_domain_sysctl();
6433 /* Move init over to a non-isolated CPU */
6434 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6436 sched_init_granularity();
6439 void __init
sched_init_smp(void)
6441 sched_init_granularity();
6443 #endif /* CONFIG_SMP */
6445 int in_sched_functions(unsigned long addr
)
6447 /* Linker adds these: start and end of __sched functions */
6448 extern char __sched_text_start
[], __sched_text_end
[];
6450 return in_lock_functions(addr
) ||
6451 (addr
>= (unsigned long)__sched_text_start
6452 && addr
< (unsigned long)__sched_text_end
);
6455 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6457 cfs_rq
->tasks_timeline
= RB_ROOT
;
6458 cfs_rq
->fair_clock
= 1;
6459 #ifdef CONFIG_FAIR_GROUP_SCHED
6464 void __init
sched_init(void)
6466 u64 now
= sched_clock();
6467 int highest_cpu
= 0;
6471 * Link up the scheduling class hierarchy:
6473 rt_sched_class
.next
= &fair_sched_class
;
6474 fair_sched_class
.next
= &idle_sched_class
;
6475 idle_sched_class
.next
= NULL
;
6477 for_each_possible_cpu(i
) {
6478 struct rt_prio_array
*array
;
6482 spin_lock_init(&rq
->lock
);
6483 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6486 init_cfs_rq(&rq
->cfs
, rq
);
6487 #ifdef CONFIG_FAIR_GROUP_SCHED
6488 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6489 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6491 rq
->ls
.load_update_last
= now
;
6492 rq
->ls
.load_update_start
= now
;
6494 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6495 rq
->cpu_load
[j
] = 0;
6498 rq
->active_balance
= 0;
6499 rq
->next_balance
= jiffies
;
6502 rq
->migration_thread
= NULL
;
6503 INIT_LIST_HEAD(&rq
->migration_queue
);
6505 atomic_set(&rq
->nr_iowait
, 0);
6507 array
= &rq
->rt
.active
;
6508 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6509 INIT_LIST_HEAD(array
->queue
+ j
);
6510 __clear_bit(j
, array
->bitmap
);
6513 /* delimiter for bitsearch: */
6514 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6517 set_load_weight(&init_task
);
6519 #ifdef CONFIG_PREEMPT_NOTIFIERS
6520 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6524 nr_cpu_ids
= highest_cpu
+ 1;
6525 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6528 #ifdef CONFIG_RT_MUTEXES
6529 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6533 * The boot idle thread does lazy MMU switching as well:
6535 atomic_inc(&init_mm
.mm_count
);
6536 enter_lazy_tlb(&init_mm
, current
);
6539 * Make us the idle thread. Technically, schedule() should not be
6540 * called from this thread, however somewhere below it might be,
6541 * but because we are the idle thread, we just pick up running again
6542 * when this runqueue becomes "idle".
6544 init_idle(current
, smp_processor_id());
6546 * During early bootup we pretend to be a normal task:
6548 current
->sched_class
= &fair_sched_class
;
6551 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6552 void __might_sleep(char *file
, int line
)
6555 static unsigned long prev_jiffy
; /* ratelimiting */
6557 if ((in_atomic() || irqs_disabled()) &&
6558 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6559 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6561 prev_jiffy
= jiffies
;
6562 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6563 " context at %s:%d\n", file
, line
);
6564 printk("in_atomic():%d, irqs_disabled():%d\n",
6565 in_atomic(), irqs_disabled());
6566 debug_show_held_locks(current
);
6567 if (irqs_disabled())
6568 print_irqtrace_events(current
);
6573 EXPORT_SYMBOL(__might_sleep
);
6576 #ifdef CONFIG_MAGIC_SYSRQ
6577 void normalize_rt_tasks(void)
6579 struct task_struct
*g
, *p
;
6580 unsigned long flags
;
6584 read_lock_irq(&tasklist_lock
);
6585 do_each_thread(g
, p
) {
6587 p
->se
.wait_runtime
= 0;
6588 p
->se
.exec_start
= 0;
6589 p
->se
.wait_start_fair
= 0;
6590 p
->se
.sleep_start_fair
= 0;
6591 #ifdef CONFIG_SCHEDSTATS
6592 p
->se
.wait_start
= 0;
6593 p
->se
.sleep_start
= 0;
6594 p
->se
.block_start
= 0;
6596 task_rq(p
)->cfs
.fair_clock
= 0;
6597 task_rq(p
)->clock
= 0;
6601 * Renice negative nice level userspace
6604 if (TASK_NICE(p
) < 0 && p
->mm
)
6605 set_user_nice(p
, 0);
6609 spin_lock_irqsave(&p
->pi_lock
, flags
);
6610 rq
= __task_rq_lock(p
);
6613 * Do not touch the migration thread:
6615 if (p
== rq
->migration_thread
)
6619 on_rq
= p
->se
.on_rq
;
6621 deactivate_task(task_rq(p
), p
, 0);
6622 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6624 activate_task(task_rq(p
), p
, 0);
6625 resched_task(rq
->curr
);
6630 __task_rq_unlock(rq
);
6631 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6632 } while_each_thread(g
, p
);
6634 read_unlock_irq(&tasklist_lock
);
6637 #endif /* CONFIG_MAGIC_SYSRQ */
6641 * These functions are only useful for the IA64 MCA handling.
6643 * They can only be called when the whole system has been
6644 * stopped - every CPU needs to be quiescent, and no scheduling
6645 * activity can take place. Using them for anything else would
6646 * be a serious bug, and as a result, they aren't even visible
6647 * under any other configuration.
6651 * curr_task - return the current task for a given cpu.
6652 * @cpu: the processor in question.
6654 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6656 struct task_struct
*curr_task(int cpu
)
6658 return cpu_curr(cpu
);
6662 * set_curr_task - set the current task for a given cpu.
6663 * @cpu: the processor in question.
6664 * @p: the task pointer to set.
6666 * Description: This function must only be used when non-maskable interrupts
6667 * are serviced on a separate stack. It allows the architecture to switch the
6668 * notion of the current task on a cpu in a non-blocking manner. This function
6669 * must be called with all CPU's synchronized, and interrupts disabled, the
6670 * and caller must save the original value of the current task (see
6671 * curr_task() above) and restore that value before reenabling interrupts and
6672 * re-starting the system.
6674 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6676 void set_curr_task(int cpu
, struct task_struct
*p
)