4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
68 * Scheduler clock - returns current time in nanosec units.
69 * This is default implementation.
70 * Architectures and sub-architectures can override this.
72 unsigned long long __attribute__((weak
)) sched_clock(void)
74 return (unsigned long long)jiffies
* (1000000000 / HZ
);
78 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
82 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
91 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96 * Some helpers for converting nanosecond timing to jiffy resolution
98 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
99 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
108 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
109 * Timeslices get refilled after they expire.
111 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
112 #define DEF_TIMESLICE (100 * HZ / 1000)
116 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
117 * Since cpu_power is a 'constant', we can use a reciprocal divide.
119 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
121 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
125 * Each time a sched group cpu_power is changed,
126 * we must compute its reciprocal value
128 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
130 sg
->__cpu_power
+= val
;
131 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
135 #define SCALE_PRIO(x, prio) \
136 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
139 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
140 * to time slice values: [800ms ... 100ms ... 5ms]
142 static unsigned int static_prio_timeslice(int static_prio
)
144 if (static_prio
== NICE_TO_PRIO(19))
147 if (static_prio
< NICE_TO_PRIO(0))
148 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
150 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
153 static inline int rt_policy(int policy
)
155 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
160 static inline int task_has_rt_policy(struct task_struct
*p
)
162 return rt_policy(p
->policy
);
166 * This is the priority-queue data structure of the RT scheduling class:
168 struct rt_prio_array
{
169 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
170 struct list_head queue
[MAX_RT_PRIO
];
174 struct load_weight load
;
175 u64 load_update_start
, load_update_last
;
176 unsigned long delta_fair
, delta_exec
, delta_stat
;
179 /* CFS-related fields in a runqueue */
181 struct load_weight load
;
182 unsigned long nr_running
;
188 unsigned long wait_runtime_overruns
, wait_runtime_underruns
;
190 struct rb_root tasks_timeline
;
191 struct rb_node
*rb_leftmost
;
192 struct rb_node
*rb_load_balance_curr
;
193 #ifdef CONFIG_FAIR_GROUP_SCHED
194 /* 'curr' points to currently running entity on this cfs_rq.
195 * It is set to NULL otherwise (i.e when none are currently running).
197 struct sched_entity
*curr
;
198 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
207 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
211 /* Real-Time classes' related field in a runqueue: */
213 struct rt_prio_array active
;
214 int rt_load_balance_idx
;
215 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
219 * This is the main, per-CPU runqueue data structure.
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
226 spinlock_t lock
; /* runqueue lock */
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
232 unsigned long nr_running
;
233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
235 unsigned char idle_at_tick
;
237 unsigned char in_nohz_recently
;
239 struct load_stat ls
; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates
;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
255 unsigned long nr_uninterruptible
;
257 struct task_struct
*curr
, *idle
;
258 unsigned long next_balance
;
259 struct mm_struct
*prev_mm
;
261 u64 clock
, prev_clock_raw
;
264 unsigned int clock_warps
, clock_overflows
;
265 unsigned int clock_unstable_events
;
270 struct sched_domain
*sd
;
272 /* For active balancing */
275 int cpu
; /* cpu of this runqueue */
277 struct task_struct
*migration_thread
;
278 struct list_head migration_queue
;
281 #ifdef CONFIG_SCHEDSTATS
283 struct sched_info rq_sched_info
;
285 /* sys_sched_yield() stats */
286 unsigned long yld_exp_empty
;
287 unsigned long yld_act_empty
;
288 unsigned long yld_both_empty
;
289 unsigned long yld_cnt
;
291 /* schedule() stats */
292 unsigned long sched_switch
;
293 unsigned long sched_cnt
;
294 unsigned long sched_goidle
;
296 /* try_to_wake_up() stats */
297 unsigned long ttwu_cnt
;
298 unsigned long ttwu_local
;
300 struct lock_class_key rq_lock_key
;
303 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
304 static DEFINE_MUTEX(sched_hotcpu_mutex
);
306 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
308 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
311 static inline int cpu_of(struct rq
*rq
)
321 * Update the per-runqueue clock, as finegrained as the platform can give
322 * us, but without assuming monotonicity, etc.:
324 static void __update_rq_clock(struct rq
*rq
)
326 u64 prev_raw
= rq
->prev_clock_raw
;
327 u64 now
= sched_clock();
328 s64 delta
= now
- prev_raw
;
329 u64 clock
= rq
->clock
;
331 #ifdef CONFIG_SCHED_DEBUG
332 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
335 * Protect against sched_clock() occasionally going backwards:
337 if (unlikely(delta
< 0)) {
342 * Catch too large forward jumps too:
344 if (unlikely(delta
> 2*TICK_NSEC
)) {
346 rq
->clock_overflows
++;
348 if (unlikely(delta
> rq
->clock_max_delta
))
349 rq
->clock_max_delta
= delta
;
354 rq
->prev_clock_raw
= now
;
358 static void update_rq_clock(struct rq
*rq
)
360 if (likely(smp_processor_id() == cpu_of(rq
)))
361 __update_rq_clock(rq
);
364 static u64
__rq_clock(struct rq
*rq
)
366 __update_rq_clock(rq
);
372 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
373 * See detach_destroy_domains: synchronize_sched for details.
375 * The domain tree of any CPU may only be accessed from within
376 * preempt-disabled sections.
378 #define for_each_domain(cpu, __sd) \
379 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
381 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
382 #define this_rq() (&__get_cpu_var(runqueues))
383 #define task_rq(p) cpu_rq(task_cpu(p))
384 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
387 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
388 * clock constructed from sched_clock():
390 unsigned long long cpu_clock(int cpu
)
392 unsigned long long now
;
396 local_irq_save(flags
);
400 local_irq_restore(flags
);
405 #ifdef CONFIG_FAIR_GROUP_SCHED
406 /* Change a task's ->cfs_rq if it moves across CPUs */
407 static inline void set_task_cfs_rq(struct task_struct
*p
)
409 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
412 static inline void set_task_cfs_rq(struct task_struct
*p
)
417 #ifndef prepare_arch_switch
418 # define prepare_arch_switch(next) do { } while (0)
420 #ifndef finish_arch_switch
421 # define finish_arch_switch(prev) do { } while (0)
424 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
425 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
427 return rq
->curr
== p
;
430 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
434 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
436 #ifdef CONFIG_DEBUG_SPINLOCK
437 /* this is a valid case when another task releases the spinlock */
438 rq
->lock
.owner
= current
;
441 * If we are tracking spinlock dependencies then we have to
442 * fix up the runqueue lock - which gets 'carried over' from
445 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
447 spin_unlock_irq(&rq
->lock
);
450 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
451 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
456 return rq
->curr
== p
;
460 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
464 * We can optimise this out completely for !SMP, because the
465 * SMP rebalancing from interrupt is the only thing that cares
470 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
471 spin_unlock_irq(&rq
->lock
);
473 spin_unlock(&rq
->lock
);
477 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
481 * After ->oncpu is cleared, the task can be moved to a different CPU.
482 * We must ensure this doesn't happen until the switch is completely
488 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
492 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
495 * __task_rq_lock - lock the runqueue a given task resides on.
496 * Must be called interrupts disabled.
498 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
505 spin_lock(&rq
->lock
);
506 if (unlikely(rq
!= task_rq(p
))) {
507 spin_unlock(&rq
->lock
);
508 goto repeat_lock_task
;
514 * task_rq_lock - lock the runqueue a given task resides on and disable
515 * interrupts. Note the ordering: we can safely lookup the task_rq without
516 * explicitly disabling preemption.
518 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
524 local_irq_save(*flags
);
526 spin_lock(&rq
->lock
);
527 if (unlikely(rq
!= task_rq(p
))) {
528 spin_unlock_irqrestore(&rq
->lock
, *flags
);
529 goto repeat_lock_task
;
534 static inline void __task_rq_unlock(struct rq
*rq
)
537 spin_unlock(&rq
->lock
);
540 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
543 spin_unlock_irqrestore(&rq
->lock
, *flags
);
547 * this_rq_lock - lock this runqueue and disable interrupts.
549 static inline struct rq
*this_rq_lock(void)
556 spin_lock(&rq
->lock
);
562 * CPU frequency is/was unstable - start new by setting prev_clock_raw:
564 void sched_clock_unstable_event(void)
569 rq
= task_rq_lock(current
, &flags
);
570 rq
->prev_clock_raw
= sched_clock();
571 rq
->clock_unstable_events
++;
572 task_rq_unlock(rq
, &flags
);
576 * resched_task - mark a task 'to be rescheduled now'.
578 * On UP this means the setting of the need_resched flag, on SMP it
579 * might also involve a cross-CPU call to trigger the scheduler on
584 #ifndef tsk_is_polling
585 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
588 static void resched_task(struct task_struct
*p
)
592 assert_spin_locked(&task_rq(p
)->lock
);
594 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
597 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
600 if (cpu
== smp_processor_id())
603 /* NEED_RESCHED must be visible before we test polling */
605 if (!tsk_is_polling(p
))
606 smp_send_reschedule(cpu
);
609 static void resched_cpu(int cpu
)
611 struct rq
*rq
= cpu_rq(cpu
);
614 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
616 resched_task(cpu_curr(cpu
));
617 spin_unlock_irqrestore(&rq
->lock
, flags
);
620 static inline void resched_task(struct task_struct
*p
)
622 assert_spin_locked(&task_rq(p
)->lock
);
623 set_tsk_need_resched(p
);
627 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
629 #if BITS_PER_LONG == 32
630 if (likely(divident
<= 0xffffffffULL
))
631 return (u32
)divident
/ divisor
;
632 do_div(divident
, divisor
);
636 return divident
/ divisor
;
640 #if BITS_PER_LONG == 32
641 # define WMULT_CONST (~0UL)
643 # define WMULT_CONST (1UL << 32)
646 #define WMULT_SHIFT 32
649 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
650 struct load_weight
*lw
)
654 if (unlikely(!lw
->inv_weight
))
655 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
657 tmp
= (u64
)delta_exec
* weight
;
659 * Check whether we'd overflow the 64-bit multiplication:
661 if (unlikely(tmp
> WMULT_CONST
)) {
662 tmp
= ((tmp
>> WMULT_SHIFT
/2) * lw
->inv_weight
)
665 tmp
= (tmp
* lw
->inv_weight
) >> WMULT_SHIFT
;
668 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
671 static inline unsigned long
672 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
674 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
677 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
683 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
690 * To aid in avoiding the subversion of "niceness" due to uneven distribution
691 * of tasks with abnormal "nice" values across CPUs the contribution that
692 * each task makes to its run queue's load is weighted according to its
693 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
694 * scaled version of the new time slice allocation that they receive on time
698 #define WEIGHT_IDLEPRIO 2
699 #define WMULT_IDLEPRIO (1 << 31)
702 * Nice levels are multiplicative, with a gentle 10% change for every
703 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
704 * nice 1, it will get ~10% less CPU time than another CPU-bound task
705 * that remained on nice 0.
707 * The "10% effect" is relative and cumulative: from _any_ nice level,
708 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
709 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
710 * If a task goes up by ~10% and another task goes down by ~10% then
711 * the relative distance between them is ~25%.)
713 static const int prio_to_weight
[40] = {
714 /* -20 */ 88818, 71054, 56843, 45475, 36380, 29104, 23283, 18626, 14901, 11921,
715 /* -10 */ 9537, 7629, 6103, 4883, 3906, 3125, 2500, 2000, 1600, 1280,
716 /* 0 */ NICE_0_LOAD
/* 1024 */,
717 /* 1 */ 819, 655, 524, 419, 336, 268, 215, 172, 137,
718 /* 10 */ 110, 87, 70, 56, 45, 36, 29, 23, 18, 15,
722 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
724 * In cases where the weight does not change often, we can use the
725 * precalculated inverse to speed up arithmetics by turning divisions
726 * into multiplications:
728 static const u32 prio_to_wmult
[40] = {
729 /* -20 */ 48356, 60446, 75558, 94446, 118058,
730 /* -15 */ 147573, 184467, 230589, 288233, 360285,
731 /* -10 */ 450347, 562979, 703746, 879575, 1099582,
732 /* -5 */ 1374389, 1717986, 2147483, 2684354, 3355443,
733 /* 0 */ 4194304, 5244160, 6557201, 8196502, 10250518,
734 /* 5 */ 12782640, 16025997, 19976592, 24970740, 31350126,
735 /* 10 */ 39045157, 49367440, 61356675, 76695844, 95443717,
736 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
739 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
742 * runqueue iterator, to support SMP load-balancing between different
743 * scheduling classes, without having to expose their internal data
744 * structures to the load-balancing proper:
748 struct task_struct
*(*start
)(void *);
749 struct task_struct
*(*next
)(void *);
752 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
753 unsigned long max_nr_move
, unsigned long max_load_move
,
754 struct sched_domain
*sd
, enum cpu_idle_type idle
,
755 int *all_pinned
, unsigned long *load_moved
,
756 int *this_best_prio
, struct rq_iterator
*iterator
);
758 #include "sched_stats.h"
759 #include "sched_rt.c"
760 #include "sched_fair.c"
761 #include "sched_idletask.c"
762 #ifdef CONFIG_SCHED_DEBUG
763 # include "sched_debug.c"
766 #define sched_class_highest (&rt_sched_class)
768 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
770 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
771 ls
->delta_exec
+= ls
->delta_stat
;
772 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
778 * Update delta_exec, delta_fair fields for rq.
780 * delta_fair clock advances at a rate inversely proportional to
781 * total load (rq->ls.load.weight) on the runqueue, while
782 * delta_exec advances at the same rate as wall-clock (provided
785 * delta_exec / delta_fair is a measure of the (smoothened) load on this
786 * runqueue over any given interval. This (smoothened) load is used
787 * during load balance.
789 * This function is called /before/ updating rq->ls.load
790 * and when switching tasks.
792 static void update_curr_load(struct rq
*rq
, u64 now
)
794 struct load_stat
*ls
= &rq
->ls
;
797 start
= ls
->load_update_start
;
798 ls
->load_update_start
= now
;
799 ls
->delta_stat
+= now
- start
;
801 * Stagger updates to ls->delta_fair. Very frequent updates
804 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
805 __update_curr_load(rq
, ls
);
809 inc_load(struct rq
*rq
, const struct task_struct
*p
, u64 now
)
811 update_curr_load(rq
, now
);
812 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
816 dec_load(struct rq
*rq
, const struct task_struct
*p
, u64 now
)
818 update_curr_load(rq
, now
);
819 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
822 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
, u64 now
)
825 inc_load(rq
, p
, now
);
828 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
, u64 now
)
831 dec_load(rq
, p
, now
);
834 static void set_load_weight(struct task_struct
*p
)
836 task_rq(p
)->cfs
.wait_runtime
-= p
->se
.wait_runtime
;
837 p
->se
.wait_runtime
= 0;
839 if (task_has_rt_policy(p
)) {
840 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
841 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
846 * SCHED_IDLE tasks get minimal weight:
848 if (p
->policy
== SCHED_IDLE
) {
849 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
850 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
854 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
855 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
859 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, u64 now
)
861 sched_info_queued(p
);
862 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, now
);
867 dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
869 p
->sched_class
->dequeue_task(rq
, p
, sleep
, now
);
874 * __normal_prio - return the priority that is based on the static prio
876 static inline int __normal_prio(struct task_struct
*p
)
878 return p
->static_prio
;
882 * Calculate the expected normal priority: i.e. priority
883 * without taking RT-inheritance into account. Might be
884 * boosted by interactivity modifiers. Changes upon fork,
885 * setprio syscalls, and whenever the interactivity
886 * estimator recalculates.
888 static inline int normal_prio(struct task_struct
*p
)
892 if (task_has_rt_policy(p
))
893 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
895 prio
= __normal_prio(p
);
900 * Calculate the current priority, i.e. the priority
901 * taken into account by the scheduler. This value might
902 * be boosted by RT tasks, or might be boosted by
903 * interactivity modifiers. Will be RT if the task got
904 * RT-boosted. If not then it returns p->normal_prio.
906 static int effective_prio(struct task_struct
*p
)
908 p
->normal_prio
= normal_prio(p
);
910 * If we are RT tasks or we were boosted to RT priority,
911 * keep the priority unchanged. Otherwise, update priority
912 * to the normal priority:
914 if (!rt_prio(p
->prio
))
915 return p
->normal_prio
;
920 * activate_task - move a task to the runqueue.
922 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
929 if (p
->state
== TASK_UNINTERRUPTIBLE
)
930 rq
->nr_uninterruptible
--;
932 enqueue_task(rq
, p
, wakeup
, now
);
933 inc_nr_running(p
, rq
, now
);
937 * activate_idle_task - move idle task to the _front_ of runqueue.
939 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
946 if (p
->state
== TASK_UNINTERRUPTIBLE
)
947 rq
->nr_uninterruptible
--;
949 enqueue_task(rq
, p
, 0, now
);
950 inc_nr_running(p
, rq
, now
);
954 * deactivate_task - remove a task from the runqueue.
957 deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
959 if (p
->state
== TASK_UNINTERRUPTIBLE
)
960 rq
->nr_uninterruptible
++;
962 dequeue_task(rq
, p
, sleep
, now
);
963 dec_nr_running(p
, rq
, now
);
967 * task_curr - is this task currently executing on a CPU?
968 * @p: the task in question.
970 inline int task_curr(const struct task_struct
*p
)
972 return cpu_curr(task_cpu(p
)) == p
;
975 /* Used instead of source_load when we know the type == 0 */
976 unsigned long weighted_cpuload(const int cpu
)
978 return cpu_rq(cpu
)->ls
.load
.weight
;
981 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
984 task_thread_info(p
)->cpu
= cpu
;
991 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
993 int old_cpu
= task_cpu(p
);
994 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
995 u64 clock_offset
, fair_clock_offset
;
997 clock_offset
= old_rq
->clock
- new_rq
->clock
;
998 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
1000 if (p
->se
.wait_start_fair
)
1001 p
->se
.wait_start_fair
-= fair_clock_offset
;
1002 if (p
->se
.sleep_start_fair
)
1003 p
->se
.sleep_start_fair
-= fair_clock_offset
;
1005 #ifdef CONFIG_SCHEDSTATS
1006 if (p
->se
.wait_start
)
1007 p
->se
.wait_start
-= clock_offset
;
1008 if (p
->se
.sleep_start
)
1009 p
->se
.sleep_start
-= clock_offset
;
1010 if (p
->se
.block_start
)
1011 p
->se
.block_start
-= clock_offset
;
1014 __set_task_cpu(p
, new_cpu
);
1017 struct migration_req
{
1018 struct list_head list
;
1020 struct task_struct
*task
;
1023 struct completion done
;
1027 * The task's runqueue lock must be held.
1028 * Returns true if you have to wait for migration thread.
1031 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1033 struct rq
*rq
= task_rq(p
);
1036 * If the task is not on a runqueue (and not running), then
1037 * it is sufficient to simply update the task's cpu field.
1039 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1040 set_task_cpu(p
, dest_cpu
);
1044 init_completion(&req
->done
);
1046 req
->dest_cpu
= dest_cpu
;
1047 list_add(&req
->list
, &rq
->migration_queue
);
1053 * wait_task_inactive - wait for a thread to unschedule.
1055 * The caller must ensure that the task *will* unschedule sometime soon,
1056 * else this function might spin for a *long* time. This function can't
1057 * be called with interrupts off, or it may introduce deadlock with
1058 * smp_call_function() if an IPI is sent by the same process we are
1059 * waiting to become inactive.
1061 void wait_task_inactive(struct task_struct
*p
)
1063 unsigned long flags
;
1069 * We do the initial early heuristics without holding
1070 * any task-queue locks at all. We'll only try to get
1071 * the runqueue lock when things look like they will
1077 * If the task is actively running on another CPU
1078 * still, just relax and busy-wait without holding
1081 * NOTE! Since we don't hold any locks, it's not
1082 * even sure that "rq" stays as the right runqueue!
1083 * But we don't care, since "task_running()" will
1084 * return false if the runqueue has changed and p
1085 * is actually now running somewhere else!
1087 while (task_running(rq
, p
))
1091 * Ok, time to look more closely! We need the rq
1092 * lock now, to be *sure*. If we're wrong, we'll
1093 * just go back and repeat.
1095 rq
= task_rq_lock(p
, &flags
);
1096 running
= task_running(rq
, p
);
1097 on_rq
= p
->se
.on_rq
;
1098 task_rq_unlock(rq
, &flags
);
1101 * Was it really running after all now that we
1102 * checked with the proper locks actually held?
1104 * Oops. Go back and try again..
1106 if (unlikely(running
)) {
1112 * It's not enough that it's not actively running,
1113 * it must be off the runqueue _entirely_, and not
1116 * So if it wa still runnable (but just not actively
1117 * running right now), it's preempted, and we should
1118 * yield - it could be a while.
1120 if (unlikely(on_rq
)) {
1126 * Ahh, all good. It wasn't running, and it wasn't
1127 * runnable, which means that it will never become
1128 * running in the future either. We're all done!
1133 * kick_process - kick a running thread to enter/exit the kernel
1134 * @p: the to-be-kicked thread
1136 * Cause a process which is running on another CPU to enter
1137 * kernel-mode, without any delay. (to get signals handled.)
1139 * NOTE: this function doesnt have to take the runqueue lock,
1140 * because all it wants to ensure is that the remote task enters
1141 * the kernel. If the IPI races and the task has been migrated
1142 * to another CPU then no harm is done and the purpose has been
1145 void kick_process(struct task_struct
*p
)
1151 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1152 smp_send_reschedule(cpu
);
1157 * Return a low guess at the load of a migration-source cpu weighted
1158 * according to the scheduling class and "nice" value.
1160 * We want to under-estimate the load of migration sources, to
1161 * balance conservatively.
1163 static inline unsigned long source_load(int cpu
, int type
)
1165 struct rq
*rq
= cpu_rq(cpu
);
1166 unsigned long total
= weighted_cpuload(cpu
);
1171 return min(rq
->cpu_load
[type
-1], total
);
1175 * Return a high guess at the load of a migration-target cpu weighted
1176 * according to the scheduling class and "nice" value.
1178 static inline unsigned long target_load(int cpu
, int type
)
1180 struct rq
*rq
= cpu_rq(cpu
);
1181 unsigned long total
= weighted_cpuload(cpu
);
1186 return max(rq
->cpu_load
[type
-1], total
);
1190 * Return the average load per task on the cpu's run queue
1192 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1194 struct rq
*rq
= cpu_rq(cpu
);
1195 unsigned long total
= weighted_cpuload(cpu
);
1196 unsigned long n
= rq
->nr_running
;
1198 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1202 * find_idlest_group finds and returns the least busy CPU group within the
1205 static struct sched_group
*
1206 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1208 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1209 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1210 int load_idx
= sd
->forkexec_idx
;
1211 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1214 unsigned long load
, avg_load
;
1218 /* Skip over this group if it has no CPUs allowed */
1219 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1222 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1224 /* Tally up the load of all CPUs in the group */
1227 for_each_cpu_mask(i
, group
->cpumask
) {
1228 /* Bias balancing toward cpus of our domain */
1230 load
= source_load(i
, load_idx
);
1232 load
= target_load(i
, load_idx
);
1237 /* Adjust by relative CPU power of the group */
1238 avg_load
= sg_div_cpu_power(group
,
1239 avg_load
* SCHED_LOAD_SCALE
);
1242 this_load
= avg_load
;
1244 } else if (avg_load
< min_load
) {
1245 min_load
= avg_load
;
1249 group
= group
->next
;
1250 } while (group
!= sd
->groups
);
1252 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1258 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1261 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1264 unsigned long load
, min_load
= ULONG_MAX
;
1268 /* Traverse only the allowed CPUs */
1269 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1271 for_each_cpu_mask(i
, tmp
) {
1272 load
= weighted_cpuload(i
);
1274 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1284 * sched_balance_self: balance the current task (running on cpu) in domains
1285 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1288 * Balance, ie. select the least loaded group.
1290 * Returns the target CPU number, or the same CPU if no balancing is needed.
1292 * preempt must be disabled.
1294 static int sched_balance_self(int cpu
, int flag
)
1296 struct task_struct
*t
= current
;
1297 struct sched_domain
*tmp
, *sd
= NULL
;
1299 for_each_domain(cpu
, tmp
) {
1301 * If power savings logic is enabled for a domain, stop there.
1303 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1305 if (tmp
->flags
& flag
)
1311 struct sched_group
*group
;
1312 int new_cpu
, weight
;
1314 if (!(sd
->flags
& flag
)) {
1320 group
= find_idlest_group(sd
, t
, cpu
);
1326 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1327 if (new_cpu
== -1 || new_cpu
== cpu
) {
1328 /* Now try balancing at a lower domain level of cpu */
1333 /* Now try balancing at a lower domain level of new_cpu */
1336 weight
= cpus_weight(span
);
1337 for_each_domain(cpu
, tmp
) {
1338 if (weight
<= cpus_weight(tmp
->span
))
1340 if (tmp
->flags
& flag
)
1343 /* while loop will break here if sd == NULL */
1349 #endif /* CONFIG_SMP */
1352 * wake_idle() will wake a task on an idle cpu if task->cpu is
1353 * not idle and an idle cpu is available. The span of cpus to
1354 * search starts with cpus closest then further out as needed,
1355 * so we always favor a closer, idle cpu.
1357 * Returns the CPU we should wake onto.
1359 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1360 static int wake_idle(int cpu
, struct task_struct
*p
)
1363 struct sched_domain
*sd
;
1367 * If it is idle, then it is the best cpu to run this task.
1369 * This cpu is also the best, if it has more than one task already.
1370 * Siblings must be also busy(in most cases) as they didn't already
1371 * pickup the extra load from this cpu and hence we need not check
1372 * sibling runqueue info. This will avoid the checks and cache miss
1373 * penalities associated with that.
1375 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1378 for_each_domain(cpu
, sd
) {
1379 if (sd
->flags
& SD_WAKE_IDLE
) {
1380 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1381 for_each_cpu_mask(i
, tmp
) {
1392 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1399 * try_to_wake_up - wake up a thread
1400 * @p: the to-be-woken-up thread
1401 * @state: the mask of task states that can be woken
1402 * @sync: do a synchronous wakeup?
1404 * Put it on the run-queue if it's not already there. The "current"
1405 * thread is always on the run-queue (except when the actual
1406 * re-schedule is in progress), and as such you're allowed to do
1407 * the simpler "current->state = TASK_RUNNING" to mark yourself
1408 * runnable without the overhead of this.
1410 * returns failure only if the task is already active.
1412 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1414 int cpu
, this_cpu
, success
= 0;
1415 unsigned long flags
;
1419 struct sched_domain
*sd
, *this_sd
= NULL
;
1420 unsigned long load
, this_load
;
1424 rq
= task_rq_lock(p
, &flags
);
1425 old_state
= p
->state
;
1426 if (!(old_state
& state
))
1433 this_cpu
= smp_processor_id();
1436 if (unlikely(task_running(rq
, p
)))
1441 schedstat_inc(rq
, ttwu_cnt
);
1442 if (cpu
== this_cpu
) {
1443 schedstat_inc(rq
, ttwu_local
);
1447 for_each_domain(this_cpu
, sd
) {
1448 if (cpu_isset(cpu
, sd
->span
)) {
1449 schedstat_inc(sd
, ttwu_wake_remote
);
1455 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1459 * Check for affine wakeup and passive balancing possibilities.
1462 int idx
= this_sd
->wake_idx
;
1463 unsigned int imbalance
;
1465 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1467 load
= source_load(cpu
, idx
);
1468 this_load
= target_load(this_cpu
, idx
);
1470 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1472 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1473 unsigned long tl
= this_load
;
1474 unsigned long tl_per_task
;
1476 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1479 * If sync wakeup then subtract the (maximum possible)
1480 * effect of the currently running task from the load
1481 * of the current CPU:
1484 tl
-= current
->se
.load
.weight
;
1487 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1488 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1490 * This domain has SD_WAKE_AFFINE and
1491 * p is cache cold in this domain, and
1492 * there is no bad imbalance.
1494 schedstat_inc(this_sd
, ttwu_move_affine
);
1500 * Start passive balancing when half the imbalance_pct
1503 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1504 if (imbalance
*this_load
<= 100*load
) {
1505 schedstat_inc(this_sd
, ttwu_move_balance
);
1511 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1513 new_cpu
= wake_idle(new_cpu
, p
);
1514 if (new_cpu
!= cpu
) {
1515 set_task_cpu(p
, new_cpu
);
1516 task_rq_unlock(rq
, &flags
);
1517 /* might preempt at this point */
1518 rq
= task_rq_lock(p
, &flags
);
1519 old_state
= p
->state
;
1520 if (!(old_state
& state
))
1525 this_cpu
= smp_processor_id();
1530 #endif /* CONFIG_SMP */
1531 activate_task(rq
, p
, 1);
1533 * Sync wakeups (i.e. those types of wakeups where the waker
1534 * has indicated that it will leave the CPU in short order)
1535 * don't trigger a preemption, if the woken up task will run on
1536 * this cpu. (in this case the 'I will reschedule' promise of
1537 * the waker guarantees that the freshly woken up task is going
1538 * to be considered on this CPU.)
1540 if (!sync
|| cpu
!= this_cpu
)
1541 check_preempt_curr(rq
, p
);
1545 p
->state
= TASK_RUNNING
;
1547 task_rq_unlock(rq
, &flags
);
1552 int fastcall
wake_up_process(struct task_struct
*p
)
1554 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1555 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1557 EXPORT_SYMBOL(wake_up_process
);
1559 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1561 return try_to_wake_up(p
, state
, 0);
1565 * Perform scheduler related setup for a newly forked process p.
1566 * p is forked by current.
1568 * __sched_fork() is basic setup used by init_idle() too:
1570 static void __sched_fork(struct task_struct
*p
)
1572 p
->se
.wait_start_fair
= 0;
1573 p
->se
.exec_start
= 0;
1574 p
->se
.sum_exec_runtime
= 0;
1575 p
->se
.delta_exec
= 0;
1576 p
->se
.delta_fair_run
= 0;
1577 p
->se
.delta_fair_sleep
= 0;
1578 p
->se
.wait_runtime
= 0;
1579 p
->se
.sleep_start_fair
= 0;
1581 #ifdef CONFIG_SCHEDSTATS
1582 p
->se
.wait_start
= 0;
1583 p
->se
.sum_wait_runtime
= 0;
1584 p
->se
.sum_sleep_runtime
= 0;
1585 p
->se
.sleep_start
= 0;
1586 p
->se
.block_start
= 0;
1587 p
->se
.sleep_max
= 0;
1588 p
->se
.block_max
= 0;
1591 p
->se
.wait_runtime_overruns
= 0;
1592 p
->se
.wait_runtime_underruns
= 0;
1595 INIT_LIST_HEAD(&p
->run_list
);
1598 #ifdef CONFIG_PREEMPT_NOTIFIERS
1599 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1603 * We mark the process as running here, but have not actually
1604 * inserted it onto the runqueue yet. This guarantees that
1605 * nobody will actually run it, and a signal or other external
1606 * event cannot wake it up and insert it on the runqueue either.
1608 p
->state
= TASK_RUNNING
;
1612 * fork()/clone()-time setup:
1614 void sched_fork(struct task_struct
*p
, int clone_flags
)
1616 int cpu
= get_cpu();
1621 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1623 __set_task_cpu(p
, cpu
);
1626 * Make sure we do not leak PI boosting priority to the child:
1628 p
->prio
= current
->normal_prio
;
1630 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1631 if (likely(sched_info_on()))
1632 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1634 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1637 #ifdef CONFIG_PREEMPT
1638 /* Want to start with kernel preemption disabled. */
1639 task_thread_info(p
)->preempt_count
= 1;
1645 * After fork, child runs first. (default) If set to 0 then
1646 * parent will (try to) run first.
1648 unsigned int __read_mostly sysctl_sched_child_runs_first
= 1;
1651 * wake_up_new_task - wake up a newly created task for the first time.
1653 * This function will do some initial scheduler statistics housekeeping
1654 * that must be done for every newly created context, then puts the task
1655 * on the runqueue and wakes it.
1657 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1659 unsigned long flags
;
1664 rq
= task_rq_lock(p
, &flags
);
1665 BUG_ON(p
->state
!= TASK_RUNNING
);
1666 this_cpu
= smp_processor_id(); /* parent's CPU */
1667 update_rq_clock(rq
);
1670 p
->prio
= effective_prio(p
);
1672 if (!p
->sched_class
->task_new
|| !sysctl_sched_child_runs_first
||
1673 (clone_flags
& CLONE_VM
) || task_cpu(p
) != this_cpu
||
1674 !current
->se
.on_rq
) {
1676 activate_task(rq
, p
, 0);
1679 * Let the scheduling class do new task startup
1680 * management (if any):
1682 p
->sched_class
->task_new(rq
, p
, now
);
1683 inc_nr_running(p
, rq
, now
);
1685 check_preempt_curr(rq
, p
);
1686 task_rq_unlock(rq
, &flags
);
1689 #ifdef CONFIG_PREEMPT_NOTIFIERS
1692 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1693 * @notifier: notifier struct to register
1695 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1697 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1699 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1702 * preempt_notifier_unregister - no longer interested in preemption notifications
1703 * @notifier: notifier struct to unregister
1705 * This is safe to call from within a preemption notifier.
1707 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1709 hlist_del(¬ifier
->link
);
1711 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1713 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1715 struct preempt_notifier
*notifier
;
1716 struct hlist_node
*node
;
1718 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1719 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1723 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1724 struct task_struct
*next
)
1726 struct preempt_notifier
*notifier
;
1727 struct hlist_node
*node
;
1729 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1730 notifier
->ops
->sched_out(notifier
, next
);
1735 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1740 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1741 struct task_struct
*next
)
1748 * prepare_task_switch - prepare to switch tasks
1749 * @rq: the runqueue preparing to switch
1750 * @prev: the current task that is being switched out
1751 * @next: the task we are going to switch to.
1753 * This is called with the rq lock held and interrupts off. It must
1754 * be paired with a subsequent finish_task_switch after the context
1757 * prepare_task_switch sets up locking and calls architecture specific
1761 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1762 struct task_struct
*next
)
1764 fire_sched_out_preempt_notifiers(prev
, next
);
1765 prepare_lock_switch(rq
, next
);
1766 prepare_arch_switch(next
);
1770 * finish_task_switch - clean up after a task-switch
1771 * @rq: runqueue associated with task-switch
1772 * @prev: the thread we just switched away from.
1774 * finish_task_switch must be called after the context switch, paired
1775 * with a prepare_task_switch call before the context switch.
1776 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1777 * and do any other architecture-specific cleanup actions.
1779 * Note that we may have delayed dropping an mm in context_switch(). If
1780 * so, we finish that here outside of the runqueue lock. (Doing it
1781 * with the lock held can cause deadlocks; see schedule() for
1784 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1785 __releases(rq
->lock
)
1787 struct mm_struct
*mm
= rq
->prev_mm
;
1793 * A task struct has one reference for the use as "current".
1794 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1795 * schedule one last time. The schedule call will never return, and
1796 * the scheduled task must drop that reference.
1797 * The test for TASK_DEAD must occur while the runqueue locks are
1798 * still held, otherwise prev could be scheduled on another cpu, die
1799 * there before we look at prev->state, and then the reference would
1801 * Manfred Spraul <manfred@colorfullife.com>
1803 prev_state
= prev
->state
;
1804 finish_arch_switch(prev
);
1805 finish_lock_switch(rq
, prev
);
1806 fire_sched_in_preempt_notifiers(current
);
1809 if (unlikely(prev_state
== TASK_DEAD
)) {
1811 * Remove function-return probe instances associated with this
1812 * task and put them back on the free list.
1814 kprobe_flush_task(prev
);
1815 put_task_struct(prev
);
1820 * schedule_tail - first thing a freshly forked thread must call.
1821 * @prev: the thread we just switched away from.
1823 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1824 __releases(rq
->lock
)
1826 struct rq
*rq
= this_rq();
1828 finish_task_switch(rq
, prev
);
1829 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1830 /* In this case, finish_task_switch does not reenable preemption */
1833 if (current
->set_child_tid
)
1834 put_user(current
->pid
, current
->set_child_tid
);
1838 * context_switch - switch to the new MM and the new
1839 * thread's register state.
1842 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1843 struct task_struct
*next
)
1845 struct mm_struct
*mm
, *oldmm
;
1847 prepare_task_switch(rq
, prev
, next
);
1849 oldmm
= prev
->active_mm
;
1851 * For paravirt, this is coupled with an exit in switch_to to
1852 * combine the page table reload and the switch backend into
1855 arch_enter_lazy_cpu_mode();
1857 if (unlikely(!mm
)) {
1858 next
->active_mm
= oldmm
;
1859 atomic_inc(&oldmm
->mm_count
);
1860 enter_lazy_tlb(oldmm
, next
);
1862 switch_mm(oldmm
, mm
, next
);
1864 if (unlikely(!prev
->mm
)) {
1865 prev
->active_mm
= NULL
;
1866 rq
->prev_mm
= oldmm
;
1869 * Since the runqueue lock will be released by the next
1870 * task (which is an invalid locking op but in the case
1871 * of the scheduler it's an obvious special-case), so we
1872 * do an early lockdep release here:
1874 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1875 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1878 /* Here we just switch the register state and the stack. */
1879 switch_to(prev
, next
, prev
);
1883 * this_rq must be evaluated again because prev may have moved
1884 * CPUs since it called schedule(), thus the 'rq' on its stack
1885 * frame will be invalid.
1887 finish_task_switch(this_rq(), prev
);
1891 * nr_running, nr_uninterruptible and nr_context_switches:
1893 * externally visible scheduler statistics: current number of runnable
1894 * threads, current number of uninterruptible-sleeping threads, total
1895 * number of context switches performed since bootup.
1897 unsigned long nr_running(void)
1899 unsigned long i
, sum
= 0;
1901 for_each_online_cpu(i
)
1902 sum
+= cpu_rq(i
)->nr_running
;
1907 unsigned long nr_uninterruptible(void)
1909 unsigned long i
, sum
= 0;
1911 for_each_possible_cpu(i
)
1912 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1915 * Since we read the counters lockless, it might be slightly
1916 * inaccurate. Do not allow it to go below zero though:
1918 if (unlikely((long)sum
< 0))
1924 unsigned long long nr_context_switches(void)
1927 unsigned long long sum
= 0;
1929 for_each_possible_cpu(i
)
1930 sum
+= cpu_rq(i
)->nr_switches
;
1935 unsigned long nr_iowait(void)
1937 unsigned long i
, sum
= 0;
1939 for_each_possible_cpu(i
)
1940 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1945 unsigned long nr_active(void)
1947 unsigned long i
, running
= 0, uninterruptible
= 0;
1949 for_each_online_cpu(i
) {
1950 running
+= cpu_rq(i
)->nr_running
;
1951 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1954 if (unlikely((long)uninterruptible
< 0))
1955 uninterruptible
= 0;
1957 return running
+ uninterruptible
;
1961 * Update rq->cpu_load[] statistics. This function is usually called every
1962 * scheduler tick (TICK_NSEC).
1964 static void update_cpu_load(struct rq
*this_rq
)
1966 u64 fair_delta64
, exec_delta64
, idle_delta64
, sample_interval64
, tmp64
;
1967 unsigned long total_load
= this_rq
->ls
.load
.weight
;
1968 unsigned long this_load
= total_load
;
1969 struct load_stat
*ls
= &this_rq
->ls
;
1970 u64 now
= __rq_clock(this_rq
);
1973 this_rq
->nr_load_updates
++;
1974 if (unlikely(!(sysctl_sched_features
& SCHED_FEAT_PRECISE_CPU_LOAD
)))
1977 /* Update delta_fair/delta_exec fields first */
1978 update_curr_load(this_rq
, now
);
1980 fair_delta64
= ls
->delta_fair
+ 1;
1983 exec_delta64
= ls
->delta_exec
+ 1;
1986 sample_interval64
= now
- ls
->load_update_last
;
1987 ls
->load_update_last
= now
;
1989 if ((s64
)sample_interval64
< (s64
)TICK_NSEC
)
1990 sample_interval64
= TICK_NSEC
;
1992 if (exec_delta64
> sample_interval64
)
1993 exec_delta64
= sample_interval64
;
1995 idle_delta64
= sample_interval64
- exec_delta64
;
1997 tmp64
= div64_64(SCHED_LOAD_SCALE
* exec_delta64
, fair_delta64
);
1998 tmp64
= div64_64(tmp64
* exec_delta64
, sample_interval64
);
2000 this_load
= (unsigned long)tmp64
;
2004 /* Update our load: */
2005 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2006 unsigned long old_load
, new_load
;
2008 /* scale is effectively 1 << i now, and >> i divides by scale */
2010 old_load
= this_rq
->cpu_load
[i
];
2011 new_load
= this_load
;
2013 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2020 * double_rq_lock - safely lock two runqueues
2022 * Note this does not disable interrupts like task_rq_lock,
2023 * you need to do so manually before calling.
2025 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2026 __acquires(rq1
->lock
)
2027 __acquires(rq2
->lock
)
2029 BUG_ON(!irqs_disabled());
2031 spin_lock(&rq1
->lock
);
2032 __acquire(rq2
->lock
); /* Fake it out ;) */
2035 spin_lock(&rq1
->lock
);
2036 spin_lock(&rq2
->lock
);
2038 spin_lock(&rq2
->lock
);
2039 spin_lock(&rq1
->lock
);
2045 * double_rq_unlock - safely unlock two runqueues
2047 * Note this does not restore interrupts like task_rq_unlock,
2048 * you need to do so manually after calling.
2050 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2051 __releases(rq1
->lock
)
2052 __releases(rq2
->lock
)
2054 spin_unlock(&rq1
->lock
);
2056 spin_unlock(&rq2
->lock
);
2058 __release(rq2
->lock
);
2062 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2064 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2065 __releases(this_rq
->lock
)
2066 __acquires(busiest
->lock
)
2067 __acquires(this_rq
->lock
)
2069 if (unlikely(!irqs_disabled())) {
2070 /* printk() doesn't work good under rq->lock */
2071 spin_unlock(&this_rq
->lock
);
2074 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2075 if (busiest
< this_rq
) {
2076 spin_unlock(&this_rq
->lock
);
2077 spin_lock(&busiest
->lock
);
2078 spin_lock(&this_rq
->lock
);
2080 spin_lock(&busiest
->lock
);
2085 * If dest_cpu is allowed for this process, migrate the task to it.
2086 * This is accomplished by forcing the cpu_allowed mask to only
2087 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2088 * the cpu_allowed mask is restored.
2090 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2092 struct migration_req req
;
2093 unsigned long flags
;
2096 rq
= task_rq_lock(p
, &flags
);
2097 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2098 || unlikely(cpu_is_offline(dest_cpu
)))
2101 /* force the process onto the specified CPU */
2102 if (migrate_task(p
, dest_cpu
, &req
)) {
2103 /* Need to wait for migration thread (might exit: take ref). */
2104 struct task_struct
*mt
= rq
->migration_thread
;
2106 get_task_struct(mt
);
2107 task_rq_unlock(rq
, &flags
);
2108 wake_up_process(mt
);
2109 put_task_struct(mt
);
2110 wait_for_completion(&req
.done
);
2115 task_rq_unlock(rq
, &flags
);
2119 * sched_exec - execve() is a valuable balancing opportunity, because at
2120 * this point the task has the smallest effective memory and cache footprint.
2122 void sched_exec(void)
2124 int new_cpu
, this_cpu
= get_cpu();
2125 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2127 if (new_cpu
!= this_cpu
)
2128 sched_migrate_task(current
, new_cpu
);
2132 * pull_task - move a task from a remote runqueue to the local runqueue.
2133 * Both runqueues must be locked.
2135 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2136 struct rq
*this_rq
, int this_cpu
)
2138 update_rq_clock(src_rq
);
2139 deactivate_task(src_rq
, p
, 0, src_rq
->clock
);
2140 set_task_cpu(p
, this_cpu
);
2141 activate_task(this_rq
, p
, 0);
2143 * Note that idle threads have a prio of MAX_PRIO, for this test
2144 * to be always true for them.
2146 check_preempt_curr(this_rq
, p
);
2150 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2153 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2154 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2158 * We do not migrate tasks that are:
2159 * 1) running (obviously), or
2160 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2161 * 3) are cache-hot on their current CPU.
2163 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2167 if (task_running(rq
, p
))
2171 * Aggressive migration if too many balance attempts have failed:
2173 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2179 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2180 unsigned long max_nr_move
, unsigned long max_load_move
,
2181 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2182 int *all_pinned
, unsigned long *load_moved
,
2183 int *this_best_prio
, struct rq_iterator
*iterator
)
2185 int pulled
= 0, pinned
= 0, skip_for_load
;
2186 struct task_struct
*p
;
2187 long rem_load_move
= max_load_move
;
2189 if (max_nr_move
== 0 || max_load_move
== 0)
2195 * Start the load-balancing iterator:
2197 p
= iterator
->start(iterator
->arg
);
2202 * To help distribute high priority tasks accross CPUs we don't
2203 * skip a task if it will be the highest priority task (i.e. smallest
2204 * prio value) on its new queue regardless of its load weight
2206 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2207 SCHED_LOAD_SCALE_FUZZ
;
2208 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2209 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2210 p
= iterator
->next(iterator
->arg
);
2214 pull_task(busiest
, p
, this_rq
, this_cpu
);
2216 rem_load_move
-= p
->se
.load
.weight
;
2219 * We only want to steal up to the prescribed number of tasks
2220 * and the prescribed amount of weighted load.
2222 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2223 if (p
->prio
< *this_best_prio
)
2224 *this_best_prio
= p
->prio
;
2225 p
= iterator
->next(iterator
->arg
);
2230 * Right now, this is the only place pull_task() is called,
2231 * so we can safely collect pull_task() stats here rather than
2232 * inside pull_task().
2234 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2237 *all_pinned
= pinned
;
2238 *load_moved
= max_load_move
- rem_load_move
;
2243 * move_tasks tries to move up to max_load_move weighted load from busiest to
2244 * this_rq, as part of a balancing operation within domain "sd".
2245 * Returns 1 if successful and 0 otherwise.
2247 * Called with both runqueues locked.
2249 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2250 unsigned long max_load_move
,
2251 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2254 struct sched_class
*class = sched_class_highest
;
2255 unsigned long total_load_moved
= 0;
2256 int this_best_prio
= this_rq
->curr
->prio
;
2260 class->load_balance(this_rq
, this_cpu
, busiest
,
2261 ULONG_MAX
, max_load_move
- total_load_moved
,
2262 sd
, idle
, all_pinned
, &this_best_prio
);
2263 class = class->next
;
2264 } while (class && max_load_move
> total_load_moved
);
2266 return total_load_moved
> 0;
2270 * move_one_task tries to move exactly one task from busiest to this_rq, as
2271 * part of active balancing operations within "domain".
2272 * Returns 1 if successful and 0 otherwise.
2274 * Called with both runqueues locked.
2276 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2277 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2279 struct sched_class
*class;
2280 int this_best_prio
= MAX_PRIO
;
2282 for (class = sched_class_highest
; class; class = class->next
)
2283 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2284 1, ULONG_MAX
, sd
, idle
, NULL
,
2292 * find_busiest_group finds and returns the busiest CPU group within the
2293 * domain. It calculates and returns the amount of weighted load which
2294 * should be moved to restore balance via the imbalance parameter.
2296 static struct sched_group
*
2297 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2298 unsigned long *imbalance
, enum cpu_idle_type idle
,
2299 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2301 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2302 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2303 unsigned long max_pull
;
2304 unsigned long busiest_load_per_task
, busiest_nr_running
;
2305 unsigned long this_load_per_task
, this_nr_running
;
2307 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2308 int power_savings_balance
= 1;
2309 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2310 unsigned long min_nr_running
= ULONG_MAX
;
2311 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2314 max_load
= this_load
= total_load
= total_pwr
= 0;
2315 busiest_load_per_task
= busiest_nr_running
= 0;
2316 this_load_per_task
= this_nr_running
= 0;
2317 if (idle
== CPU_NOT_IDLE
)
2318 load_idx
= sd
->busy_idx
;
2319 else if (idle
== CPU_NEWLY_IDLE
)
2320 load_idx
= sd
->newidle_idx
;
2322 load_idx
= sd
->idle_idx
;
2325 unsigned long load
, group_capacity
;
2328 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2329 unsigned long sum_nr_running
, sum_weighted_load
;
2331 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2334 balance_cpu
= first_cpu(group
->cpumask
);
2336 /* Tally up the load of all CPUs in the group */
2337 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2339 for_each_cpu_mask(i
, group
->cpumask
) {
2342 if (!cpu_isset(i
, *cpus
))
2347 if (*sd_idle
&& rq
->nr_running
)
2350 /* Bias balancing toward cpus of our domain */
2352 if (idle_cpu(i
) && !first_idle_cpu
) {
2357 load
= target_load(i
, load_idx
);
2359 load
= source_load(i
, load_idx
);
2362 sum_nr_running
+= rq
->nr_running
;
2363 sum_weighted_load
+= weighted_cpuload(i
);
2367 * First idle cpu or the first cpu(busiest) in this sched group
2368 * is eligible for doing load balancing at this and above
2369 * domains. In the newly idle case, we will allow all the cpu's
2370 * to do the newly idle load balance.
2372 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2373 balance_cpu
!= this_cpu
&& balance
) {
2378 total_load
+= avg_load
;
2379 total_pwr
+= group
->__cpu_power
;
2381 /* Adjust by relative CPU power of the group */
2382 avg_load
= sg_div_cpu_power(group
,
2383 avg_load
* SCHED_LOAD_SCALE
);
2385 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2388 this_load
= avg_load
;
2390 this_nr_running
= sum_nr_running
;
2391 this_load_per_task
= sum_weighted_load
;
2392 } else if (avg_load
> max_load
&&
2393 sum_nr_running
> group_capacity
) {
2394 max_load
= avg_load
;
2396 busiest_nr_running
= sum_nr_running
;
2397 busiest_load_per_task
= sum_weighted_load
;
2400 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2402 * Busy processors will not participate in power savings
2405 if (idle
== CPU_NOT_IDLE
||
2406 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2410 * If the local group is idle or completely loaded
2411 * no need to do power savings balance at this domain
2413 if (local_group
&& (this_nr_running
>= group_capacity
||
2415 power_savings_balance
= 0;
2418 * If a group is already running at full capacity or idle,
2419 * don't include that group in power savings calculations
2421 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2426 * Calculate the group which has the least non-idle load.
2427 * This is the group from where we need to pick up the load
2430 if ((sum_nr_running
< min_nr_running
) ||
2431 (sum_nr_running
== min_nr_running
&&
2432 first_cpu(group
->cpumask
) <
2433 first_cpu(group_min
->cpumask
))) {
2435 min_nr_running
= sum_nr_running
;
2436 min_load_per_task
= sum_weighted_load
/
2441 * Calculate the group which is almost near its
2442 * capacity but still has some space to pick up some load
2443 * from other group and save more power
2445 if (sum_nr_running
<= group_capacity
- 1) {
2446 if (sum_nr_running
> leader_nr_running
||
2447 (sum_nr_running
== leader_nr_running
&&
2448 first_cpu(group
->cpumask
) >
2449 first_cpu(group_leader
->cpumask
))) {
2450 group_leader
= group
;
2451 leader_nr_running
= sum_nr_running
;
2456 group
= group
->next
;
2457 } while (group
!= sd
->groups
);
2459 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2462 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2464 if (this_load
>= avg_load
||
2465 100*max_load
<= sd
->imbalance_pct
*this_load
)
2468 busiest_load_per_task
/= busiest_nr_running
;
2470 * We're trying to get all the cpus to the average_load, so we don't
2471 * want to push ourselves above the average load, nor do we wish to
2472 * reduce the max loaded cpu below the average load, as either of these
2473 * actions would just result in more rebalancing later, and ping-pong
2474 * tasks around. Thus we look for the minimum possible imbalance.
2475 * Negative imbalances (*we* are more loaded than anyone else) will
2476 * be counted as no imbalance for these purposes -- we can't fix that
2477 * by pulling tasks to us. Be careful of negative numbers as they'll
2478 * appear as very large values with unsigned longs.
2480 if (max_load
<= busiest_load_per_task
)
2484 * In the presence of smp nice balancing, certain scenarios can have
2485 * max load less than avg load(as we skip the groups at or below
2486 * its cpu_power, while calculating max_load..)
2488 if (max_load
< avg_load
) {
2490 goto small_imbalance
;
2493 /* Don't want to pull so many tasks that a group would go idle */
2494 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2496 /* How much load to actually move to equalise the imbalance */
2497 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2498 (avg_load
- this_load
) * this->__cpu_power
)
2502 * if *imbalance is less than the average load per runnable task
2503 * there is no gaurantee that any tasks will be moved so we'll have
2504 * a think about bumping its value to force at least one task to be
2507 if (*imbalance
+ SCHED_LOAD_SCALE_FUZZ
< busiest_load_per_task
/2) {
2508 unsigned long tmp
, pwr_now
, pwr_move
;
2512 pwr_move
= pwr_now
= 0;
2514 if (this_nr_running
) {
2515 this_load_per_task
/= this_nr_running
;
2516 if (busiest_load_per_task
> this_load_per_task
)
2519 this_load_per_task
= SCHED_LOAD_SCALE
;
2521 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2522 busiest_load_per_task
* imbn
) {
2523 *imbalance
= busiest_load_per_task
;
2528 * OK, we don't have enough imbalance to justify moving tasks,
2529 * however we may be able to increase total CPU power used by
2533 pwr_now
+= busiest
->__cpu_power
*
2534 min(busiest_load_per_task
, max_load
);
2535 pwr_now
+= this->__cpu_power
*
2536 min(this_load_per_task
, this_load
);
2537 pwr_now
/= SCHED_LOAD_SCALE
;
2539 /* Amount of load we'd subtract */
2540 tmp
= sg_div_cpu_power(busiest
,
2541 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2543 pwr_move
+= busiest
->__cpu_power
*
2544 min(busiest_load_per_task
, max_load
- tmp
);
2546 /* Amount of load we'd add */
2547 if (max_load
* busiest
->__cpu_power
<
2548 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2549 tmp
= sg_div_cpu_power(this,
2550 max_load
* busiest
->__cpu_power
);
2552 tmp
= sg_div_cpu_power(this,
2553 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2554 pwr_move
+= this->__cpu_power
*
2555 min(this_load_per_task
, this_load
+ tmp
);
2556 pwr_move
/= SCHED_LOAD_SCALE
;
2558 /* Move if we gain throughput */
2559 if (pwr_move
<= pwr_now
)
2562 *imbalance
= busiest_load_per_task
;
2568 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2569 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2572 if (this == group_leader
&& group_leader
!= group_min
) {
2573 *imbalance
= min_load_per_task
;
2583 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2586 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2587 unsigned long imbalance
, cpumask_t
*cpus
)
2589 struct rq
*busiest
= NULL
, *rq
;
2590 unsigned long max_load
= 0;
2593 for_each_cpu_mask(i
, group
->cpumask
) {
2596 if (!cpu_isset(i
, *cpus
))
2600 wl
= weighted_cpuload(i
);
2602 if (rq
->nr_running
== 1 && wl
> imbalance
)
2605 if (wl
> max_load
) {
2615 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2616 * so long as it is large enough.
2618 #define MAX_PINNED_INTERVAL 512
2621 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2622 * tasks if there is an imbalance.
2624 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2625 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2628 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2629 struct sched_group
*group
;
2630 unsigned long imbalance
;
2632 cpumask_t cpus
= CPU_MASK_ALL
;
2633 unsigned long flags
;
2636 * When power savings policy is enabled for the parent domain, idle
2637 * sibling can pick up load irrespective of busy siblings. In this case,
2638 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2639 * portraying it as CPU_NOT_IDLE.
2641 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2642 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2645 schedstat_inc(sd
, lb_cnt
[idle
]);
2648 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2655 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2659 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2661 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2665 BUG_ON(busiest
== this_rq
);
2667 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2670 if (busiest
->nr_running
> 1) {
2672 * Attempt to move tasks. If find_busiest_group has found
2673 * an imbalance but busiest->nr_running <= 1, the group is
2674 * still unbalanced. ld_moved simply stays zero, so it is
2675 * correctly treated as an imbalance.
2677 local_irq_save(flags
);
2678 double_rq_lock(this_rq
, busiest
);
2679 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2680 imbalance
, sd
, idle
, &all_pinned
);
2681 double_rq_unlock(this_rq
, busiest
);
2682 local_irq_restore(flags
);
2685 * some other cpu did the load balance for us.
2687 if (ld_moved
&& this_cpu
!= smp_processor_id())
2688 resched_cpu(this_cpu
);
2690 /* All tasks on this runqueue were pinned by CPU affinity */
2691 if (unlikely(all_pinned
)) {
2692 cpu_clear(cpu_of(busiest
), cpus
);
2693 if (!cpus_empty(cpus
))
2700 schedstat_inc(sd
, lb_failed
[idle
]);
2701 sd
->nr_balance_failed
++;
2703 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2705 spin_lock_irqsave(&busiest
->lock
, flags
);
2707 /* don't kick the migration_thread, if the curr
2708 * task on busiest cpu can't be moved to this_cpu
2710 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2711 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2713 goto out_one_pinned
;
2716 if (!busiest
->active_balance
) {
2717 busiest
->active_balance
= 1;
2718 busiest
->push_cpu
= this_cpu
;
2721 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2723 wake_up_process(busiest
->migration_thread
);
2726 * We've kicked active balancing, reset the failure
2729 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2732 sd
->nr_balance_failed
= 0;
2734 if (likely(!active_balance
)) {
2735 /* We were unbalanced, so reset the balancing interval */
2736 sd
->balance_interval
= sd
->min_interval
;
2739 * If we've begun active balancing, start to back off. This
2740 * case may not be covered by the all_pinned logic if there
2741 * is only 1 task on the busy runqueue (because we don't call
2744 if (sd
->balance_interval
< sd
->max_interval
)
2745 sd
->balance_interval
*= 2;
2748 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2749 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2754 schedstat_inc(sd
, lb_balanced
[idle
]);
2756 sd
->nr_balance_failed
= 0;
2759 /* tune up the balancing interval */
2760 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2761 (sd
->balance_interval
< sd
->max_interval
))
2762 sd
->balance_interval
*= 2;
2764 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2765 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2771 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2772 * tasks if there is an imbalance.
2774 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2775 * this_rq is locked.
2778 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2780 struct sched_group
*group
;
2781 struct rq
*busiest
= NULL
;
2782 unsigned long imbalance
;
2786 cpumask_t cpus
= CPU_MASK_ALL
;
2789 * When power savings policy is enabled for the parent domain, idle
2790 * sibling can pick up load irrespective of busy siblings. In this case,
2791 * let the state of idle sibling percolate up as IDLE, instead of
2792 * portraying it as CPU_NOT_IDLE.
2794 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2795 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2798 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2800 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2801 &sd_idle
, &cpus
, NULL
);
2803 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2807 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2810 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2814 BUG_ON(busiest
== this_rq
);
2816 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2819 if (busiest
->nr_running
> 1) {
2820 /* Attempt to move tasks */
2821 double_lock_balance(this_rq
, busiest
);
2822 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2823 imbalance
, sd
, CPU_NEWLY_IDLE
,
2825 spin_unlock(&busiest
->lock
);
2827 if (unlikely(all_pinned
)) {
2828 cpu_clear(cpu_of(busiest
), cpus
);
2829 if (!cpus_empty(cpus
))
2835 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2836 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2837 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2840 sd
->nr_balance_failed
= 0;
2845 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2846 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2847 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2849 sd
->nr_balance_failed
= 0;
2855 * idle_balance is called by schedule() if this_cpu is about to become
2856 * idle. Attempts to pull tasks from other CPUs.
2858 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2860 struct sched_domain
*sd
;
2861 int pulled_task
= -1;
2862 unsigned long next_balance
= jiffies
+ HZ
;
2864 for_each_domain(this_cpu
, sd
) {
2865 unsigned long interval
;
2867 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2870 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2871 /* If we've pulled tasks over stop searching: */
2872 pulled_task
= load_balance_newidle(this_cpu
,
2875 interval
= msecs_to_jiffies(sd
->balance_interval
);
2876 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2877 next_balance
= sd
->last_balance
+ interval
;
2881 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2883 * We are going idle. next_balance may be set based on
2884 * a busy processor. So reset next_balance.
2886 this_rq
->next_balance
= next_balance
;
2891 * active_load_balance is run by migration threads. It pushes running tasks
2892 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2893 * running on each physical CPU where possible, and avoids physical /
2894 * logical imbalances.
2896 * Called with busiest_rq locked.
2898 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2900 int target_cpu
= busiest_rq
->push_cpu
;
2901 struct sched_domain
*sd
;
2902 struct rq
*target_rq
;
2904 /* Is there any task to move? */
2905 if (busiest_rq
->nr_running
<= 1)
2908 target_rq
= cpu_rq(target_cpu
);
2911 * This condition is "impossible", if it occurs
2912 * we need to fix it. Originally reported by
2913 * Bjorn Helgaas on a 128-cpu setup.
2915 BUG_ON(busiest_rq
== target_rq
);
2917 /* move a task from busiest_rq to target_rq */
2918 double_lock_balance(busiest_rq
, target_rq
);
2920 /* Search for an sd spanning us and the target CPU. */
2921 for_each_domain(target_cpu
, sd
) {
2922 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2923 cpu_isset(busiest_cpu
, sd
->span
))
2928 schedstat_inc(sd
, alb_cnt
);
2930 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2932 schedstat_inc(sd
, alb_pushed
);
2934 schedstat_inc(sd
, alb_failed
);
2936 spin_unlock(&target_rq
->lock
);
2941 atomic_t load_balancer
;
2943 } nohz ____cacheline_aligned
= {
2944 .load_balancer
= ATOMIC_INIT(-1),
2945 .cpu_mask
= CPU_MASK_NONE
,
2949 * This routine will try to nominate the ilb (idle load balancing)
2950 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2951 * load balancing on behalf of all those cpus. If all the cpus in the system
2952 * go into this tickless mode, then there will be no ilb owner (as there is
2953 * no need for one) and all the cpus will sleep till the next wakeup event
2956 * For the ilb owner, tick is not stopped. And this tick will be used
2957 * for idle load balancing. ilb owner will still be part of
2960 * While stopping the tick, this cpu will become the ilb owner if there
2961 * is no other owner. And will be the owner till that cpu becomes busy
2962 * or if all cpus in the system stop their ticks at which point
2963 * there is no need for ilb owner.
2965 * When the ilb owner becomes busy, it nominates another owner, during the
2966 * next busy scheduler_tick()
2968 int select_nohz_load_balancer(int stop_tick
)
2970 int cpu
= smp_processor_id();
2973 cpu_set(cpu
, nohz
.cpu_mask
);
2974 cpu_rq(cpu
)->in_nohz_recently
= 1;
2977 * If we are going offline and still the leader, give up!
2979 if (cpu_is_offline(cpu
) &&
2980 atomic_read(&nohz
.load_balancer
) == cpu
) {
2981 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2986 /* time for ilb owner also to sleep */
2987 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2988 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2989 atomic_set(&nohz
.load_balancer
, -1);
2993 if (atomic_read(&nohz
.load_balancer
) == -1) {
2994 /* make me the ilb owner */
2995 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2997 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3000 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3003 cpu_clear(cpu
, nohz
.cpu_mask
);
3005 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3006 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3013 static DEFINE_SPINLOCK(balancing
);
3016 * It checks each scheduling domain to see if it is due to be balanced,
3017 * and initiates a balancing operation if so.
3019 * Balancing parameters are set up in arch_init_sched_domains.
3021 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3024 struct rq
*rq
= cpu_rq(cpu
);
3025 unsigned long interval
;
3026 struct sched_domain
*sd
;
3027 /* Earliest time when we have to do rebalance again */
3028 unsigned long next_balance
= jiffies
+ 60*HZ
;
3030 for_each_domain(cpu
, sd
) {
3031 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3034 interval
= sd
->balance_interval
;
3035 if (idle
!= CPU_IDLE
)
3036 interval
*= sd
->busy_factor
;
3038 /* scale ms to jiffies */
3039 interval
= msecs_to_jiffies(interval
);
3040 if (unlikely(!interval
))
3042 if (interval
> HZ
*NR_CPUS
/10)
3043 interval
= HZ
*NR_CPUS
/10;
3046 if (sd
->flags
& SD_SERIALIZE
) {
3047 if (!spin_trylock(&balancing
))
3051 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3052 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3054 * We've pulled tasks over so either we're no
3055 * longer idle, or one of our SMT siblings is
3058 idle
= CPU_NOT_IDLE
;
3060 sd
->last_balance
= jiffies
;
3062 if (sd
->flags
& SD_SERIALIZE
)
3063 spin_unlock(&balancing
);
3065 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3066 next_balance
= sd
->last_balance
+ interval
;
3069 * Stop the load balance at this level. There is another
3070 * CPU in our sched group which is doing load balancing more
3076 rq
->next_balance
= next_balance
;
3080 * run_rebalance_domains is triggered when needed from the scheduler tick.
3081 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3082 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3084 static void run_rebalance_domains(struct softirq_action
*h
)
3086 int this_cpu
= smp_processor_id();
3087 struct rq
*this_rq
= cpu_rq(this_cpu
);
3088 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3089 CPU_IDLE
: CPU_NOT_IDLE
;
3091 rebalance_domains(this_cpu
, idle
);
3095 * If this cpu is the owner for idle load balancing, then do the
3096 * balancing on behalf of the other idle cpus whose ticks are
3099 if (this_rq
->idle_at_tick
&&
3100 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3101 cpumask_t cpus
= nohz
.cpu_mask
;
3105 cpu_clear(this_cpu
, cpus
);
3106 for_each_cpu_mask(balance_cpu
, cpus
) {
3108 * If this cpu gets work to do, stop the load balancing
3109 * work being done for other cpus. Next load
3110 * balancing owner will pick it up.
3115 rebalance_domains(balance_cpu
, SCHED_IDLE
);
3117 rq
= cpu_rq(balance_cpu
);
3118 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3119 this_rq
->next_balance
= rq
->next_balance
;
3126 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3128 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3129 * idle load balancing owner or decide to stop the periodic load balancing,
3130 * if the whole system is idle.
3132 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3136 * If we were in the nohz mode recently and busy at the current
3137 * scheduler tick, then check if we need to nominate new idle
3140 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3141 rq
->in_nohz_recently
= 0;
3143 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3144 cpu_clear(cpu
, nohz
.cpu_mask
);
3145 atomic_set(&nohz
.load_balancer
, -1);
3148 if (atomic_read(&nohz
.load_balancer
) == -1) {
3150 * simple selection for now: Nominate the
3151 * first cpu in the nohz list to be the next
3154 * TBD: Traverse the sched domains and nominate
3155 * the nearest cpu in the nohz.cpu_mask.
3157 int ilb
= first_cpu(nohz
.cpu_mask
);
3165 * If this cpu is idle and doing idle load balancing for all the
3166 * cpus with ticks stopped, is it time for that to stop?
3168 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3169 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3175 * If this cpu is idle and the idle load balancing is done by
3176 * someone else, then no need raise the SCHED_SOFTIRQ
3178 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3179 cpu_isset(cpu
, nohz
.cpu_mask
))
3182 if (time_after_eq(jiffies
, rq
->next_balance
))
3183 raise_softirq(SCHED_SOFTIRQ
);
3186 #else /* CONFIG_SMP */
3189 * on UP we do not need to balance between CPUs:
3191 static inline void idle_balance(int cpu
, struct rq
*rq
)
3195 /* Avoid "used but not defined" warning on UP */
3196 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3197 unsigned long max_nr_move
, unsigned long max_load_move
,
3198 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3199 int *all_pinned
, unsigned long *load_moved
,
3200 int *this_best_prio
, struct rq_iterator
*iterator
)
3209 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3211 EXPORT_PER_CPU_SYMBOL(kstat
);
3214 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3215 * that have not yet been banked in case the task is currently running.
3217 unsigned long long task_sched_runtime(struct task_struct
*p
)
3219 unsigned long flags
;
3223 rq
= task_rq_lock(p
, &flags
);
3224 ns
= p
->se
.sum_exec_runtime
;
3225 if (rq
->curr
== p
) {
3226 update_rq_clock(rq
);
3227 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3228 if ((s64
)delta_exec
> 0)
3231 task_rq_unlock(rq
, &flags
);
3237 * Account user cpu time to a process.
3238 * @p: the process that the cpu time gets accounted to
3239 * @hardirq_offset: the offset to subtract from hardirq_count()
3240 * @cputime: the cpu time spent in user space since the last update
3242 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3244 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3247 p
->utime
= cputime_add(p
->utime
, cputime
);
3249 /* Add user time to cpustat. */
3250 tmp
= cputime_to_cputime64(cputime
);
3251 if (TASK_NICE(p
) > 0)
3252 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3254 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3258 * Account system cpu time to a process.
3259 * @p: the process that the cpu time gets accounted to
3260 * @hardirq_offset: the offset to subtract from hardirq_count()
3261 * @cputime: the cpu time spent in kernel space since the last update
3263 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3266 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3267 struct rq
*rq
= this_rq();
3270 p
->stime
= cputime_add(p
->stime
, cputime
);
3272 /* Add system time to cpustat. */
3273 tmp
= cputime_to_cputime64(cputime
);
3274 if (hardirq_count() - hardirq_offset
)
3275 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3276 else if (softirq_count())
3277 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3278 else if (p
!= rq
->idle
)
3279 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3280 else if (atomic_read(&rq
->nr_iowait
) > 0)
3281 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3283 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3284 /* Account for system time used */
3285 acct_update_integrals(p
);
3289 * Account for involuntary wait time.
3290 * @p: the process from which the cpu time has been stolen
3291 * @steal: the cpu time spent in involuntary wait
3293 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3295 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3296 cputime64_t tmp
= cputime_to_cputime64(steal
);
3297 struct rq
*rq
= this_rq();
3299 if (p
== rq
->idle
) {
3300 p
->stime
= cputime_add(p
->stime
, steal
);
3301 if (atomic_read(&rq
->nr_iowait
) > 0)
3302 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3304 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3306 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3310 * This function gets called by the timer code, with HZ frequency.
3311 * We call it with interrupts disabled.
3313 * It also gets called by the fork code, when changing the parent's
3316 void scheduler_tick(void)
3318 int cpu
= smp_processor_id();
3319 struct rq
*rq
= cpu_rq(cpu
);
3320 struct task_struct
*curr
= rq
->curr
;
3322 spin_lock(&rq
->lock
);
3323 update_cpu_load(rq
);
3324 if (curr
!= rq
->idle
) /* FIXME: needed? */
3325 curr
->sched_class
->task_tick(rq
, curr
);
3326 spin_unlock(&rq
->lock
);
3329 rq
->idle_at_tick
= idle_cpu(cpu
);
3330 trigger_load_balance(rq
, cpu
);
3334 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3336 void fastcall
add_preempt_count(int val
)
3341 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3343 preempt_count() += val
;
3345 * Spinlock count overflowing soon?
3347 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3350 EXPORT_SYMBOL(add_preempt_count
);
3352 void fastcall
sub_preempt_count(int val
)
3357 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3360 * Is the spinlock portion underflowing?
3362 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3363 !(preempt_count() & PREEMPT_MASK
)))
3366 preempt_count() -= val
;
3368 EXPORT_SYMBOL(sub_preempt_count
);
3373 * Print scheduling while atomic bug:
3375 static noinline
void __schedule_bug(struct task_struct
*prev
)
3377 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3378 prev
->comm
, preempt_count(), prev
->pid
);
3379 debug_show_held_locks(prev
);
3380 if (irqs_disabled())
3381 print_irqtrace_events(prev
);
3386 * Various schedule()-time debugging checks and statistics:
3388 static inline void schedule_debug(struct task_struct
*prev
)
3391 * Test if we are atomic. Since do_exit() needs to call into
3392 * schedule() atomically, we ignore that path for now.
3393 * Otherwise, whine if we are scheduling when we should not be.
3395 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3396 __schedule_bug(prev
);
3398 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3400 schedstat_inc(this_rq(), sched_cnt
);
3404 * Pick up the highest-prio task:
3406 static inline struct task_struct
*
3407 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, u64 now
)
3409 struct sched_class
*class;
3410 struct task_struct
*p
;
3413 * Optimization: we know that if all tasks are in
3414 * the fair class we can call that function directly:
3416 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3417 p
= fair_sched_class
.pick_next_task(rq
, now
);
3422 class = sched_class_highest
;
3424 p
= class->pick_next_task(rq
, now
);
3428 * Will never be NULL as the idle class always
3429 * returns a non-NULL p:
3431 class = class->next
;
3436 * schedule() is the main scheduler function.
3438 asmlinkage
void __sched
schedule(void)
3440 struct task_struct
*prev
, *next
;
3448 cpu
= smp_processor_id();
3452 switch_count
= &prev
->nivcsw
;
3454 release_kernel_lock(prev
);
3455 need_resched_nonpreemptible
:
3457 schedule_debug(prev
);
3459 spin_lock_irq(&rq
->lock
);
3460 clear_tsk_need_resched(prev
);
3461 now
= __rq_clock(rq
);
3463 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3464 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3465 unlikely(signal_pending(prev
)))) {
3466 prev
->state
= TASK_RUNNING
;
3468 deactivate_task(rq
, prev
, 1, now
);
3470 switch_count
= &prev
->nvcsw
;
3473 if (unlikely(!rq
->nr_running
))
3474 idle_balance(cpu
, rq
);
3476 prev
->sched_class
->put_prev_task(rq
, prev
, now
);
3477 next
= pick_next_task(rq
, prev
, now
);
3479 sched_info_switch(prev
, next
);
3481 if (likely(prev
!= next
)) {
3486 context_switch(rq
, prev
, next
); /* unlocks the rq */
3488 spin_unlock_irq(&rq
->lock
);
3490 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3491 cpu
= smp_processor_id();
3493 goto need_resched_nonpreemptible
;
3495 preempt_enable_no_resched();
3496 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3499 EXPORT_SYMBOL(schedule
);
3501 #ifdef CONFIG_PREEMPT
3503 * this is the entry point to schedule() from in-kernel preemption
3504 * off of preempt_enable. Kernel preemptions off return from interrupt
3505 * occur there and call schedule directly.
3507 asmlinkage
void __sched
preempt_schedule(void)
3509 struct thread_info
*ti
= current_thread_info();
3510 #ifdef CONFIG_PREEMPT_BKL
3511 struct task_struct
*task
= current
;
3512 int saved_lock_depth
;
3515 * If there is a non-zero preempt_count or interrupts are disabled,
3516 * we do not want to preempt the current task. Just return..
3518 if (likely(ti
->preempt_count
|| irqs_disabled()))
3522 add_preempt_count(PREEMPT_ACTIVE
);
3524 * We keep the big kernel semaphore locked, but we
3525 * clear ->lock_depth so that schedule() doesnt
3526 * auto-release the semaphore:
3528 #ifdef CONFIG_PREEMPT_BKL
3529 saved_lock_depth
= task
->lock_depth
;
3530 task
->lock_depth
= -1;
3533 #ifdef CONFIG_PREEMPT_BKL
3534 task
->lock_depth
= saved_lock_depth
;
3536 sub_preempt_count(PREEMPT_ACTIVE
);
3538 /* we could miss a preemption opportunity between schedule and now */
3540 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3543 EXPORT_SYMBOL(preempt_schedule
);
3546 * this is the entry point to schedule() from kernel preemption
3547 * off of irq context.
3548 * Note, that this is called and return with irqs disabled. This will
3549 * protect us against recursive calling from irq.
3551 asmlinkage
void __sched
preempt_schedule_irq(void)
3553 struct thread_info
*ti
= current_thread_info();
3554 #ifdef CONFIG_PREEMPT_BKL
3555 struct task_struct
*task
= current
;
3556 int saved_lock_depth
;
3558 /* Catch callers which need to be fixed */
3559 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3562 add_preempt_count(PREEMPT_ACTIVE
);
3564 * We keep the big kernel semaphore locked, but we
3565 * clear ->lock_depth so that schedule() doesnt
3566 * auto-release the semaphore:
3568 #ifdef CONFIG_PREEMPT_BKL
3569 saved_lock_depth
= task
->lock_depth
;
3570 task
->lock_depth
= -1;
3574 local_irq_disable();
3575 #ifdef CONFIG_PREEMPT_BKL
3576 task
->lock_depth
= saved_lock_depth
;
3578 sub_preempt_count(PREEMPT_ACTIVE
);
3580 /* we could miss a preemption opportunity between schedule and now */
3582 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3586 #endif /* CONFIG_PREEMPT */
3588 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3591 return try_to_wake_up(curr
->private, mode
, sync
);
3593 EXPORT_SYMBOL(default_wake_function
);
3596 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3597 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3598 * number) then we wake all the non-exclusive tasks and one exclusive task.
3600 * There are circumstances in which we can try to wake a task which has already
3601 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3602 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3604 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3605 int nr_exclusive
, int sync
, void *key
)
3607 struct list_head
*tmp
, *next
;
3609 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3610 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3611 unsigned flags
= curr
->flags
;
3613 if (curr
->func(curr
, mode
, sync
, key
) &&
3614 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3620 * __wake_up - wake up threads blocked on a waitqueue.
3622 * @mode: which threads
3623 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3624 * @key: is directly passed to the wakeup function
3626 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3627 int nr_exclusive
, void *key
)
3629 unsigned long flags
;
3631 spin_lock_irqsave(&q
->lock
, flags
);
3632 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3633 spin_unlock_irqrestore(&q
->lock
, flags
);
3635 EXPORT_SYMBOL(__wake_up
);
3638 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3640 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3642 __wake_up_common(q
, mode
, 1, 0, NULL
);
3646 * __wake_up_sync - wake up threads blocked on a waitqueue.
3648 * @mode: which threads
3649 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3651 * The sync wakeup differs that the waker knows that it will schedule
3652 * away soon, so while the target thread will be woken up, it will not
3653 * be migrated to another CPU - ie. the two threads are 'synchronized'
3654 * with each other. This can prevent needless bouncing between CPUs.
3656 * On UP it can prevent extra preemption.
3659 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3661 unsigned long flags
;
3667 if (unlikely(!nr_exclusive
))
3670 spin_lock_irqsave(&q
->lock
, flags
);
3671 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3672 spin_unlock_irqrestore(&q
->lock
, flags
);
3674 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3676 void fastcall
complete(struct completion
*x
)
3678 unsigned long flags
;
3680 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3682 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3684 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3686 EXPORT_SYMBOL(complete
);
3688 void fastcall
complete_all(struct completion
*x
)
3690 unsigned long flags
;
3692 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3693 x
->done
+= UINT_MAX
/2;
3694 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3696 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3698 EXPORT_SYMBOL(complete_all
);
3700 void fastcall __sched
wait_for_completion(struct completion
*x
)
3704 spin_lock_irq(&x
->wait
.lock
);
3706 DECLARE_WAITQUEUE(wait
, current
);
3708 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3709 __add_wait_queue_tail(&x
->wait
, &wait
);
3711 __set_current_state(TASK_UNINTERRUPTIBLE
);
3712 spin_unlock_irq(&x
->wait
.lock
);
3714 spin_lock_irq(&x
->wait
.lock
);
3716 __remove_wait_queue(&x
->wait
, &wait
);
3719 spin_unlock_irq(&x
->wait
.lock
);
3721 EXPORT_SYMBOL(wait_for_completion
);
3723 unsigned long fastcall __sched
3724 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3728 spin_lock_irq(&x
->wait
.lock
);
3730 DECLARE_WAITQUEUE(wait
, current
);
3732 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3733 __add_wait_queue_tail(&x
->wait
, &wait
);
3735 __set_current_state(TASK_UNINTERRUPTIBLE
);
3736 spin_unlock_irq(&x
->wait
.lock
);
3737 timeout
= schedule_timeout(timeout
);
3738 spin_lock_irq(&x
->wait
.lock
);
3740 __remove_wait_queue(&x
->wait
, &wait
);
3744 __remove_wait_queue(&x
->wait
, &wait
);
3748 spin_unlock_irq(&x
->wait
.lock
);
3751 EXPORT_SYMBOL(wait_for_completion_timeout
);
3753 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3759 spin_lock_irq(&x
->wait
.lock
);
3761 DECLARE_WAITQUEUE(wait
, current
);
3763 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3764 __add_wait_queue_tail(&x
->wait
, &wait
);
3766 if (signal_pending(current
)) {
3768 __remove_wait_queue(&x
->wait
, &wait
);
3771 __set_current_state(TASK_INTERRUPTIBLE
);
3772 spin_unlock_irq(&x
->wait
.lock
);
3774 spin_lock_irq(&x
->wait
.lock
);
3776 __remove_wait_queue(&x
->wait
, &wait
);
3780 spin_unlock_irq(&x
->wait
.lock
);
3784 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3786 unsigned long fastcall __sched
3787 wait_for_completion_interruptible_timeout(struct completion
*x
,
3788 unsigned long timeout
)
3792 spin_lock_irq(&x
->wait
.lock
);
3794 DECLARE_WAITQUEUE(wait
, current
);
3796 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3797 __add_wait_queue_tail(&x
->wait
, &wait
);
3799 if (signal_pending(current
)) {
3800 timeout
= -ERESTARTSYS
;
3801 __remove_wait_queue(&x
->wait
, &wait
);
3804 __set_current_state(TASK_INTERRUPTIBLE
);
3805 spin_unlock_irq(&x
->wait
.lock
);
3806 timeout
= schedule_timeout(timeout
);
3807 spin_lock_irq(&x
->wait
.lock
);
3809 __remove_wait_queue(&x
->wait
, &wait
);
3813 __remove_wait_queue(&x
->wait
, &wait
);
3817 spin_unlock_irq(&x
->wait
.lock
);
3820 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3823 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3825 spin_lock_irqsave(&q
->lock
, *flags
);
3826 __add_wait_queue(q
, wait
);
3827 spin_unlock(&q
->lock
);
3831 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3833 spin_lock_irq(&q
->lock
);
3834 __remove_wait_queue(q
, wait
);
3835 spin_unlock_irqrestore(&q
->lock
, *flags
);
3838 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3840 unsigned long flags
;
3843 init_waitqueue_entry(&wait
, current
);
3845 current
->state
= TASK_INTERRUPTIBLE
;
3847 sleep_on_head(q
, &wait
, &flags
);
3849 sleep_on_tail(q
, &wait
, &flags
);
3851 EXPORT_SYMBOL(interruptible_sleep_on
);
3854 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3856 unsigned long flags
;
3859 init_waitqueue_entry(&wait
, current
);
3861 current
->state
= TASK_INTERRUPTIBLE
;
3863 sleep_on_head(q
, &wait
, &flags
);
3864 timeout
= schedule_timeout(timeout
);
3865 sleep_on_tail(q
, &wait
, &flags
);
3869 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3871 void __sched
sleep_on(wait_queue_head_t
*q
)
3873 unsigned long flags
;
3876 init_waitqueue_entry(&wait
, current
);
3878 current
->state
= TASK_UNINTERRUPTIBLE
;
3880 sleep_on_head(q
, &wait
, &flags
);
3882 sleep_on_tail(q
, &wait
, &flags
);
3884 EXPORT_SYMBOL(sleep_on
);
3886 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3888 unsigned long flags
;
3891 init_waitqueue_entry(&wait
, current
);
3893 current
->state
= TASK_UNINTERRUPTIBLE
;
3895 sleep_on_head(q
, &wait
, &flags
);
3896 timeout
= schedule_timeout(timeout
);
3897 sleep_on_tail(q
, &wait
, &flags
);
3901 EXPORT_SYMBOL(sleep_on_timeout
);
3903 #ifdef CONFIG_RT_MUTEXES
3906 * rt_mutex_setprio - set the current priority of a task
3908 * @prio: prio value (kernel-internal form)
3910 * This function changes the 'effective' priority of a task. It does
3911 * not touch ->normal_prio like __setscheduler().
3913 * Used by the rt_mutex code to implement priority inheritance logic.
3915 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3917 unsigned long flags
;
3922 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3924 rq
= task_rq_lock(p
, &flags
);
3925 update_rq_clock(rq
);
3929 on_rq
= p
->se
.on_rq
;
3931 dequeue_task(rq
, p
, 0, now
);
3934 p
->sched_class
= &rt_sched_class
;
3936 p
->sched_class
= &fair_sched_class
;
3941 enqueue_task(rq
, p
, 0, now
);
3943 * Reschedule if we are currently running on this runqueue and
3944 * our priority decreased, or if we are not currently running on
3945 * this runqueue and our priority is higher than the current's
3947 if (task_running(rq
, p
)) {
3948 if (p
->prio
> oldprio
)
3949 resched_task(rq
->curr
);
3951 check_preempt_curr(rq
, p
);
3954 task_rq_unlock(rq
, &flags
);
3959 void set_user_nice(struct task_struct
*p
, long nice
)
3961 int old_prio
, delta
, on_rq
;
3962 unsigned long flags
;
3966 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3969 * We have to be careful, if called from sys_setpriority(),
3970 * the task might be in the middle of scheduling on another CPU.
3972 rq
= task_rq_lock(p
, &flags
);
3973 update_rq_clock(rq
);
3976 * The RT priorities are set via sched_setscheduler(), but we still
3977 * allow the 'normal' nice value to be set - but as expected
3978 * it wont have any effect on scheduling until the task is
3979 * SCHED_FIFO/SCHED_RR:
3981 if (task_has_rt_policy(p
)) {
3982 p
->static_prio
= NICE_TO_PRIO(nice
);
3985 on_rq
= p
->se
.on_rq
;
3987 dequeue_task(rq
, p
, 0, now
);
3988 dec_load(rq
, p
, now
);
3991 p
->static_prio
= NICE_TO_PRIO(nice
);
3994 p
->prio
= effective_prio(p
);
3995 delta
= p
->prio
- old_prio
;
3998 enqueue_task(rq
, p
, 0, now
);
3999 inc_load(rq
, p
, now
);
4001 * If the task increased its priority or is running and
4002 * lowered its priority, then reschedule its CPU:
4004 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4005 resched_task(rq
->curr
);
4008 task_rq_unlock(rq
, &flags
);
4010 EXPORT_SYMBOL(set_user_nice
);
4013 * can_nice - check if a task can reduce its nice value
4017 int can_nice(const struct task_struct
*p
, const int nice
)
4019 /* convert nice value [19,-20] to rlimit style value [1,40] */
4020 int nice_rlim
= 20 - nice
;
4022 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4023 capable(CAP_SYS_NICE
));
4026 #ifdef __ARCH_WANT_SYS_NICE
4029 * sys_nice - change the priority of the current process.
4030 * @increment: priority increment
4032 * sys_setpriority is a more generic, but much slower function that
4033 * does similar things.
4035 asmlinkage
long sys_nice(int increment
)
4040 * Setpriority might change our priority at the same moment.
4041 * We don't have to worry. Conceptually one call occurs first
4042 * and we have a single winner.
4044 if (increment
< -40)
4049 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4055 if (increment
< 0 && !can_nice(current
, nice
))
4058 retval
= security_task_setnice(current
, nice
);
4062 set_user_nice(current
, nice
);
4069 * task_prio - return the priority value of a given task.
4070 * @p: the task in question.
4072 * This is the priority value as seen by users in /proc.
4073 * RT tasks are offset by -200. Normal tasks are centered
4074 * around 0, value goes from -16 to +15.
4076 int task_prio(const struct task_struct
*p
)
4078 return p
->prio
- MAX_RT_PRIO
;
4082 * task_nice - return the nice value of a given task.
4083 * @p: the task in question.
4085 int task_nice(const struct task_struct
*p
)
4087 return TASK_NICE(p
);
4089 EXPORT_SYMBOL_GPL(task_nice
);
4092 * idle_cpu - is a given cpu idle currently?
4093 * @cpu: the processor in question.
4095 int idle_cpu(int cpu
)
4097 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4101 * idle_task - return the idle task for a given cpu.
4102 * @cpu: the processor in question.
4104 struct task_struct
*idle_task(int cpu
)
4106 return cpu_rq(cpu
)->idle
;
4110 * find_process_by_pid - find a process with a matching PID value.
4111 * @pid: the pid in question.
4113 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4115 return pid
? find_task_by_pid(pid
) : current
;
4118 /* Actually do priority change: must hold rq lock. */
4120 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4122 BUG_ON(p
->se
.on_rq
);
4125 switch (p
->policy
) {
4129 p
->sched_class
= &fair_sched_class
;
4133 p
->sched_class
= &rt_sched_class
;
4137 p
->rt_priority
= prio
;
4138 p
->normal_prio
= normal_prio(p
);
4139 /* we are holding p->pi_lock already */
4140 p
->prio
= rt_mutex_getprio(p
);
4145 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4146 * @p: the task in question.
4147 * @policy: new policy.
4148 * @param: structure containing the new RT priority.
4150 * NOTE that the task may be already dead.
4152 int sched_setscheduler(struct task_struct
*p
, int policy
,
4153 struct sched_param
*param
)
4155 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4156 unsigned long flags
;
4159 /* may grab non-irq protected spin_locks */
4160 BUG_ON(in_interrupt());
4162 /* double check policy once rq lock held */
4164 policy
= oldpolicy
= p
->policy
;
4165 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4166 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4167 policy
!= SCHED_IDLE
)
4170 * Valid priorities for SCHED_FIFO and SCHED_RR are
4171 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4172 * SCHED_BATCH and SCHED_IDLE is 0.
4174 if (param
->sched_priority
< 0 ||
4175 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4176 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4178 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4182 * Allow unprivileged RT tasks to decrease priority:
4184 if (!capable(CAP_SYS_NICE
)) {
4185 if (rt_policy(policy
)) {
4186 unsigned long rlim_rtprio
;
4188 if (!lock_task_sighand(p
, &flags
))
4190 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4191 unlock_task_sighand(p
, &flags
);
4193 /* can't set/change the rt policy */
4194 if (policy
!= p
->policy
&& !rlim_rtprio
)
4197 /* can't increase priority */
4198 if (param
->sched_priority
> p
->rt_priority
&&
4199 param
->sched_priority
> rlim_rtprio
)
4203 * Like positive nice levels, dont allow tasks to
4204 * move out of SCHED_IDLE either:
4206 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4209 /* can't change other user's priorities */
4210 if ((current
->euid
!= p
->euid
) &&
4211 (current
->euid
!= p
->uid
))
4215 retval
= security_task_setscheduler(p
, policy
, param
);
4219 * make sure no PI-waiters arrive (or leave) while we are
4220 * changing the priority of the task:
4222 spin_lock_irqsave(&p
->pi_lock
, flags
);
4224 * To be able to change p->policy safely, the apropriate
4225 * runqueue lock must be held.
4227 rq
= __task_rq_lock(p
);
4228 /* recheck policy now with rq lock held */
4229 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4230 policy
= oldpolicy
= -1;
4231 __task_rq_unlock(rq
);
4232 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4235 on_rq
= p
->se
.on_rq
;
4237 update_rq_clock(rq
);
4238 deactivate_task(rq
, p
, 0, rq
->clock
);
4241 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4243 activate_task(rq
, p
, 0);
4245 * Reschedule if we are currently running on this runqueue and
4246 * our priority decreased, or if we are not currently running on
4247 * this runqueue and our priority is higher than the current's
4249 if (task_running(rq
, p
)) {
4250 if (p
->prio
> oldprio
)
4251 resched_task(rq
->curr
);
4253 check_preempt_curr(rq
, p
);
4256 __task_rq_unlock(rq
);
4257 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4259 rt_mutex_adjust_pi(p
);
4263 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4266 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4268 struct sched_param lparam
;
4269 struct task_struct
*p
;
4272 if (!param
|| pid
< 0)
4274 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4279 p
= find_process_by_pid(pid
);
4281 retval
= sched_setscheduler(p
, policy
, &lparam
);
4288 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4289 * @pid: the pid in question.
4290 * @policy: new policy.
4291 * @param: structure containing the new RT priority.
4293 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4294 struct sched_param __user
*param
)
4296 /* negative values for policy are not valid */
4300 return do_sched_setscheduler(pid
, policy
, param
);
4304 * sys_sched_setparam - set/change the RT priority of a thread
4305 * @pid: the pid in question.
4306 * @param: structure containing the new RT priority.
4308 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4310 return do_sched_setscheduler(pid
, -1, param
);
4314 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4315 * @pid: the pid in question.
4317 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4319 struct task_struct
*p
;
4320 int retval
= -EINVAL
;
4326 read_lock(&tasklist_lock
);
4327 p
= find_process_by_pid(pid
);
4329 retval
= security_task_getscheduler(p
);
4333 read_unlock(&tasklist_lock
);
4340 * sys_sched_getscheduler - get the RT priority of a thread
4341 * @pid: the pid in question.
4342 * @param: structure containing the RT priority.
4344 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4346 struct sched_param lp
;
4347 struct task_struct
*p
;
4348 int retval
= -EINVAL
;
4350 if (!param
|| pid
< 0)
4353 read_lock(&tasklist_lock
);
4354 p
= find_process_by_pid(pid
);
4359 retval
= security_task_getscheduler(p
);
4363 lp
.sched_priority
= p
->rt_priority
;
4364 read_unlock(&tasklist_lock
);
4367 * This one might sleep, we cannot do it with a spinlock held ...
4369 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4375 read_unlock(&tasklist_lock
);
4379 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4381 cpumask_t cpus_allowed
;
4382 struct task_struct
*p
;
4385 mutex_lock(&sched_hotcpu_mutex
);
4386 read_lock(&tasklist_lock
);
4388 p
= find_process_by_pid(pid
);
4390 read_unlock(&tasklist_lock
);
4391 mutex_unlock(&sched_hotcpu_mutex
);
4396 * It is not safe to call set_cpus_allowed with the
4397 * tasklist_lock held. We will bump the task_struct's
4398 * usage count and then drop tasklist_lock.
4401 read_unlock(&tasklist_lock
);
4404 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4405 !capable(CAP_SYS_NICE
))
4408 retval
= security_task_setscheduler(p
, 0, NULL
);
4412 cpus_allowed
= cpuset_cpus_allowed(p
);
4413 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4414 retval
= set_cpus_allowed(p
, new_mask
);
4418 mutex_unlock(&sched_hotcpu_mutex
);
4422 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4423 cpumask_t
*new_mask
)
4425 if (len
< sizeof(cpumask_t
)) {
4426 memset(new_mask
, 0, sizeof(cpumask_t
));
4427 } else if (len
> sizeof(cpumask_t
)) {
4428 len
= sizeof(cpumask_t
);
4430 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4434 * sys_sched_setaffinity - set the cpu affinity of a process
4435 * @pid: pid of the process
4436 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4437 * @user_mask_ptr: user-space pointer to the new cpu mask
4439 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4440 unsigned long __user
*user_mask_ptr
)
4445 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4449 return sched_setaffinity(pid
, new_mask
);
4453 * Represents all cpu's present in the system
4454 * In systems capable of hotplug, this map could dynamically grow
4455 * as new cpu's are detected in the system via any platform specific
4456 * method, such as ACPI for e.g.
4459 cpumask_t cpu_present_map __read_mostly
;
4460 EXPORT_SYMBOL(cpu_present_map
);
4463 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4464 EXPORT_SYMBOL(cpu_online_map
);
4466 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4467 EXPORT_SYMBOL(cpu_possible_map
);
4470 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4472 struct task_struct
*p
;
4475 mutex_lock(&sched_hotcpu_mutex
);
4476 read_lock(&tasklist_lock
);
4479 p
= find_process_by_pid(pid
);
4483 retval
= security_task_getscheduler(p
);
4487 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4490 read_unlock(&tasklist_lock
);
4491 mutex_unlock(&sched_hotcpu_mutex
);
4497 * sys_sched_getaffinity - get the cpu affinity of a process
4498 * @pid: pid of the process
4499 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4500 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4502 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4503 unsigned long __user
*user_mask_ptr
)
4508 if (len
< sizeof(cpumask_t
))
4511 ret
= sched_getaffinity(pid
, &mask
);
4515 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4518 return sizeof(cpumask_t
);
4522 * sys_sched_yield - yield the current processor to other threads.
4524 * This function yields the current CPU to other tasks. If there are no
4525 * other threads running on this CPU then this function will return.
4527 asmlinkage
long sys_sched_yield(void)
4529 struct rq
*rq
= this_rq_lock();
4531 schedstat_inc(rq
, yld_cnt
);
4532 if (unlikely(rq
->nr_running
== 1))
4533 schedstat_inc(rq
, yld_act_empty
);
4535 current
->sched_class
->yield_task(rq
, current
);
4538 * Since we are going to call schedule() anyway, there's
4539 * no need to preempt or enable interrupts:
4541 __release(rq
->lock
);
4542 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4543 _raw_spin_unlock(&rq
->lock
);
4544 preempt_enable_no_resched();
4551 static void __cond_resched(void)
4553 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4554 __might_sleep(__FILE__
, __LINE__
);
4557 * The BKS might be reacquired before we have dropped
4558 * PREEMPT_ACTIVE, which could trigger a second
4559 * cond_resched() call.
4562 add_preempt_count(PREEMPT_ACTIVE
);
4564 sub_preempt_count(PREEMPT_ACTIVE
);
4565 } while (need_resched());
4568 int __sched
cond_resched(void)
4570 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4571 system_state
== SYSTEM_RUNNING
) {
4577 EXPORT_SYMBOL(cond_resched
);
4580 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4581 * call schedule, and on return reacquire the lock.
4583 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4584 * operations here to prevent schedule() from being called twice (once via
4585 * spin_unlock(), once by hand).
4587 int cond_resched_lock(spinlock_t
*lock
)
4591 if (need_lockbreak(lock
)) {
4597 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4598 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4599 _raw_spin_unlock(lock
);
4600 preempt_enable_no_resched();
4607 EXPORT_SYMBOL(cond_resched_lock
);
4609 int __sched
cond_resched_softirq(void)
4611 BUG_ON(!in_softirq());
4613 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4621 EXPORT_SYMBOL(cond_resched_softirq
);
4624 * yield - yield the current processor to other threads.
4626 * This is a shortcut for kernel-space yielding - it marks the
4627 * thread runnable and calls sys_sched_yield().
4629 void __sched
yield(void)
4631 set_current_state(TASK_RUNNING
);
4634 EXPORT_SYMBOL(yield
);
4637 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4638 * that process accounting knows that this is a task in IO wait state.
4640 * But don't do that if it is a deliberate, throttling IO wait (this task
4641 * has set its backing_dev_info: the queue against which it should throttle)
4643 void __sched
io_schedule(void)
4645 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4647 delayacct_blkio_start();
4648 atomic_inc(&rq
->nr_iowait
);
4650 atomic_dec(&rq
->nr_iowait
);
4651 delayacct_blkio_end();
4653 EXPORT_SYMBOL(io_schedule
);
4655 long __sched
io_schedule_timeout(long timeout
)
4657 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4660 delayacct_blkio_start();
4661 atomic_inc(&rq
->nr_iowait
);
4662 ret
= schedule_timeout(timeout
);
4663 atomic_dec(&rq
->nr_iowait
);
4664 delayacct_blkio_end();
4669 * sys_sched_get_priority_max - return maximum RT priority.
4670 * @policy: scheduling class.
4672 * this syscall returns the maximum rt_priority that can be used
4673 * by a given scheduling class.
4675 asmlinkage
long sys_sched_get_priority_max(int policy
)
4682 ret
= MAX_USER_RT_PRIO
-1;
4694 * sys_sched_get_priority_min - return minimum RT priority.
4695 * @policy: scheduling class.
4697 * this syscall returns the minimum rt_priority that can be used
4698 * by a given scheduling class.
4700 asmlinkage
long sys_sched_get_priority_min(int policy
)
4718 * sys_sched_rr_get_interval - return the default timeslice of a process.
4719 * @pid: pid of the process.
4720 * @interval: userspace pointer to the timeslice value.
4722 * this syscall writes the default timeslice value of a given process
4723 * into the user-space timespec buffer. A value of '0' means infinity.
4726 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4728 struct task_struct
*p
;
4729 int retval
= -EINVAL
;
4736 read_lock(&tasklist_lock
);
4737 p
= find_process_by_pid(pid
);
4741 retval
= security_task_getscheduler(p
);
4745 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4746 0 : static_prio_timeslice(p
->static_prio
), &t
);
4747 read_unlock(&tasklist_lock
);
4748 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4752 read_unlock(&tasklist_lock
);
4756 static const char stat_nam
[] = "RSDTtZX";
4758 static void show_task(struct task_struct
*p
)
4760 unsigned long free
= 0;
4763 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4764 printk("%-13.13s %c", p
->comm
,
4765 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4766 #if BITS_PER_LONG == 32
4767 if (state
== TASK_RUNNING
)
4768 printk(" running ");
4770 printk(" %08lx ", thread_saved_pc(p
));
4772 if (state
== TASK_RUNNING
)
4773 printk(" running task ");
4775 printk(" %016lx ", thread_saved_pc(p
));
4777 #ifdef CONFIG_DEBUG_STACK_USAGE
4779 unsigned long *n
= end_of_stack(p
);
4782 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4785 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4787 if (state
!= TASK_RUNNING
)
4788 show_stack(p
, NULL
);
4791 void show_state_filter(unsigned long state_filter
)
4793 struct task_struct
*g
, *p
;
4795 #if BITS_PER_LONG == 32
4797 " task PC stack pid father\n");
4800 " task PC stack pid father\n");
4802 read_lock(&tasklist_lock
);
4803 do_each_thread(g
, p
) {
4805 * reset the NMI-timeout, listing all files on a slow
4806 * console might take alot of time:
4808 touch_nmi_watchdog();
4809 if (!state_filter
|| (p
->state
& state_filter
))
4811 } while_each_thread(g
, p
);
4813 touch_all_softlockup_watchdogs();
4815 #ifdef CONFIG_SCHED_DEBUG
4816 sysrq_sched_debug_show();
4818 read_unlock(&tasklist_lock
);
4820 * Only show locks if all tasks are dumped:
4822 if (state_filter
== -1)
4823 debug_show_all_locks();
4826 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4828 idle
->sched_class
= &idle_sched_class
;
4832 * init_idle - set up an idle thread for a given CPU
4833 * @idle: task in question
4834 * @cpu: cpu the idle task belongs to
4836 * NOTE: this function does not set the idle thread's NEED_RESCHED
4837 * flag, to make booting more robust.
4839 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4841 struct rq
*rq
= cpu_rq(cpu
);
4842 unsigned long flags
;
4845 idle
->se
.exec_start
= sched_clock();
4847 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4848 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4849 __set_task_cpu(idle
, cpu
);
4851 spin_lock_irqsave(&rq
->lock
, flags
);
4852 rq
->curr
= rq
->idle
= idle
;
4853 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4856 spin_unlock_irqrestore(&rq
->lock
, flags
);
4858 /* Set the preempt count _outside_ the spinlocks! */
4859 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4860 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4862 task_thread_info(idle
)->preempt_count
= 0;
4865 * The idle tasks have their own, simple scheduling class:
4867 idle
->sched_class
= &idle_sched_class
;
4871 * In a system that switches off the HZ timer nohz_cpu_mask
4872 * indicates which cpus entered this state. This is used
4873 * in the rcu update to wait only for active cpus. For system
4874 * which do not switch off the HZ timer nohz_cpu_mask should
4875 * always be CPU_MASK_NONE.
4877 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4880 * Increase the granularity value when there are more CPUs,
4881 * because with more CPUs the 'effective latency' as visible
4882 * to users decreases. But the relationship is not linear,
4883 * so pick a second-best guess by going with the log2 of the
4886 * This idea comes from the SD scheduler of Con Kolivas:
4888 static inline void sched_init_granularity(void)
4890 unsigned int factor
= 1 + ilog2(num_online_cpus());
4891 const unsigned long gran_limit
= 100000000;
4893 sysctl_sched_granularity
*= factor
;
4894 if (sysctl_sched_granularity
> gran_limit
)
4895 sysctl_sched_granularity
= gran_limit
;
4897 sysctl_sched_runtime_limit
= sysctl_sched_granularity
* 4;
4898 sysctl_sched_wakeup_granularity
= sysctl_sched_granularity
/ 2;
4903 * This is how migration works:
4905 * 1) we queue a struct migration_req structure in the source CPU's
4906 * runqueue and wake up that CPU's migration thread.
4907 * 2) we down() the locked semaphore => thread blocks.
4908 * 3) migration thread wakes up (implicitly it forces the migrated
4909 * thread off the CPU)
4910 * 4) it gets the migration request and checks whether the migrated
4911 * task is still in the wrong runqueue.
4912 * 5) if it's in the wrong runqueue then the migration thread removes
4913 * it and puts it into the right queue.
4914 * 6) migration thread up()s the semaphore.
4915 * 7) we wake up and the migration is done.
4919 * Change a given task's CPU affinity. Migrate the thread to a
4920 * proper CPU and schedule it away if the CPU it's executing on
4921 * is removed from the allowed bitmask.
4923 * NOTE: the caller must have a valid reference to the task, the
4924 * task must not exit() & deallocate itself prematurely. The
4925 * call is not atomic; no spinlocks may be held.
4927 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4929 struct migration_req req
;
4930 unsigned long flags
;
4934 rq
= task_rq_lock(p
, &flags
);
4935 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4940 p
->cpus_allowed
= new_mask
;
4941 /* Can the task run on the task's current CPU? If so, we're done */
4942 if (cpu_isset(task_cpu(p
), new_mask
))
4945 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4946 /* Need help from migration thread: drop lock and wait. */
4947 task_rq_unlock(rq
, &flags
);
4948 wake_up_process(rq
->migration_thread
);
4949 wait_for_completion(&req
.done
);
4950 tlb_migrate_finish(p
->mm
);
4954 task_rq_unlock(rq
, &flags
);
4958 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4961 * Move (not current) task off this cpu, onto dest cpu. We're doing
4962 * this because either it can't run here any more (set_cpus_allowed()
4963 * away from this CPU, or CPU going down), or because we're
4964 * attempting to rebalance this task on exec (sched_exec).
4966 * So we race with normal scheduler movements, but that's OK, as long
4967 * as the task is no longer on this CPU.
4969 * Returns non-zero if task was successfully migrated.
4971 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4973 struct rq
*rq_dest
, *rq_src
;
4976 if (unlikely(cpu_is_offline(dest_cpu
)))
4979 rq_src
= cpu_rq(src_cpu
);
4980 rq_dest
= cpu_rq(dest_cpu
);
4982 double_rq_lock(rq_src
, rq_dest
);
4983 /* Already moved. */
4984 if (task_cpu(p
) != src_cpu
)
4986 /* Affinity changed (again). */
4987 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4990 on_rq
= p
->se
.on_rq
;
4992 update_rq_clock(rq_src
);
4993 deactivate_task(rq_src
, p
, 0, rq_src
->clock
);
4995 set_task_cpu(p
, dest_cpu
);
4997 activate_task(rq_dest
, p
, 0);
4998 check_preempt_curr(rq_dest
, p
);
5002 double_rq_unlock(rq_src
, rq_dest
);
5007 * migration_thread - this is a highprio system thread that performs
5008 * thread migration by bumping thread off CPU then 'pushing' onto
5011 static int migration_thread(void *data
)
5013 int cpu
= (long)data
;
5017 BUG_ON(rq
->migration_thread
!= current
);
5019 set_current_state(TASK_INTERRUPTIBLE
);
5020 while (!kthread_should_stop()) {
5021 struct migration_req
*req
;
5022 struct list_head
*head
;
5024 spin_lock_irq(&rq
->lock
);
5026 if (cpu_is_offline(cpu
)) {
5027 spin_unlock_irq(&rq
->lock
);
5031 if (rq
->active_balance
) {
5032 active_load_balance(rq
, cpu
);
5033 rq
->active_balance
= 0;
5036 head
= &rq
->migration_queue
;
5038 if (list_empty(head
)) {
5039 spin_unlock_irq(&rq
->lock
);
5041 set_current_state(TASK_INTERRUPTIBLE
);
5044 req
= list_entry(head
->next
, struct migration_req
, list
);
5045 list_del_init(head
->next
);
5047 spin_unlock(&rq
->lock
);
5048 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5051 complete(&req
->done
);
5053 __set_current_state(TASK_RUNNING
);
5057 /* Wait for kthread_stop */
5058 set_current_state(TASK_INTERRUPTIBLE
);
5059 while (!kthread_should_stop()) {
5061 set_current_state(TASK_INTERRUPTIBLE
);
5063 __set_current_state(TASK_RUNNING
);
5067 #ifdef CONFIG_HOTPLUG_CPU
5069 * Figure out where task on dead CPU should go, use force if neccessary.
5070 * NOTE: interrupts should be disabled by the caller
5072 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5074 unsigned long flags
;
5081 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5082 cpus_and(mask
, mask
, p
->cpus_allowed
);
5083 dest_cpu
= any_online_cpu(mask
);
5085 /* On any allowed CPU? */
5086 if (dest_cpu
== NR_CPUS
)
5087 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5089 /* No more Mr. Nice Guy. */
5090 if (dest_cpu
== NR_CPUS
) {
5091 rq
= task_rq_lock(p
, &flags
);
5092 cpus_setall(p
->cpus_allowed
);
5093 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5094 task_rq_unlock(rq
, &flags
);
5097 * Don't tell them about moving exiting tasks or
5098 * kernel threads (both mm NULL), since they never
5101 if (p
->mm
&& printk_ratelimit())
5102 printk(KERN_INFO
"process %d (%s) no "
5103 "longer affine to cpu%d\n",
5104 p
->pid
, p
->comm
, dead_cpu
);
5106 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5111 * While a dead CPU has no uninterruptible tasks queued at this point,
5112 * it might still have a nonzero ->nr_uninterruptible counter, because
5113 * for performance reasons the counter is not stricly tracking tasks to
5114 * their home CPUs. So we just add the counter to another CPU's counter,
5115 * to keep the global sum constant after CPU-down:
5117 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5119 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5120 unsigned long flags
;
5122 local_irq_save(flags
);
5123 double_rq_lock(rq_src
, rq_dest
);
5124 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5125 rq_src
->nr_uninterruptible
= 0;
5126 double_rq_unlock(rq_src
, rq_dest
);
5127 local_irq_restore(flags
);
5130 /* Run through task list and migrate tasks from the dead cpu. */
5131 static void migrate_live_tasks(int src_cpu
)
5133 struct task_struct
*p
, *t
;
5135 write_lock_irq(&tasklist_lock
);
5137 do_each_thread(t
, p
) {
5141 if (task_cpu(p
) == src_cpu
)
5142 move_task_off_dead_cpu(src_cpu
, p
);
5143 } while_each_thread(t
, p
);
5145 write_unlock_irq(&tasklist_lock
);
5149 * Schedules idle task to be the next runnable task on current CPU.
5150 * It does so by boosting its priority to highest possible and adding it to
5151 * the _front_ of the runqueue. Used by CPU offline code.
5153 void sched_idle_next(void)
5155 int this_cpu
= smp_processor_id();
5156 struct rq
*rq
= cpu_rq(this_cpu
);
5157 struct task_struct
*p
= rq
->idle
;
5158 unsigned long flags
;
5160 /* cpu has to be offline */
5161 BUG_ON(cpu_online(this_cpu
));
5164 * Strictly not necessary since rest of the CPUs are stopped by now
5165 * and interrupts disabled on the current cpu.
5167 spin_lock_irqsave(&rq
->lock
, flags
);
5169 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5171 /* Add idle task to the _front_ of its priority queue: */
5172 activate_idle_task(p
, rq
);
5174 spin_unlock_irqrestore(&rq
->lock
, flags
);
5178 * Ensures that the idle task is using init_mm right before its cpu goes
5181 void idle_task_exit(void)
5183 struct mm_struct
*mm
= current
->active_mm
;
5185 BUG_ON(cpu_online(smp_processor_id()));
5188 switch_mm(mm
, &init_mm
, current
);
5192 /* called under rq->lock with disabled interrupts */
5193 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5195 struct rq
*rq
= cpu_rq(dead_cpu
);
5197 /* Must be exiting, otherwise would be on tasklist. */
5198 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5200 /* Cannot have done final schedule yet: would have vanished. */
5201 BUG_ON(p
->state
== TASK_DEAD
);
5206 * Drop lock around migration; if someone else moves it,
5207 * that's OK. No task can be added to this CPU, so iteration is
5209 * NOTE: interrupts should be left disabled --dev@
5211 spin_unlock(&rq
->lock
);
5212 move_task_off_dead_cpu(dead_cpu
, p
);
5213 spin_lock(&rq
->lock
);
5218 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5219 static void migrate_dead_tasks(unsigned int dead_cpu
)
5221 struct rq
*rq
= cpu_rq(dead_cpu
);
5222 struct task_struct
*next
;
5225 if (!rq
->nr_running
)
5227 update_rq_clock(rq
);
5228 next
= pick_next_task(rq
, rq
->curr
, rq
->clock
);
5231 migrate_dead(dead_cpu
, next
);
5235 #endif /* CONFIG_HOTPLUG_CPU */
5237 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5239 static struct ctl_table sd_ctl_dir
[] = {
5241 .procname
= "sched_domain",
5247 static struct ctl_table sd_ctl_root
[] = {
5249 .procname
= "kernel",
5251 .child
= sd_ctl_dir
,
5256 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5258 struct ctl_table
*entry
=
5259 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5262 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5268 set_table_entry(struct ctl_table
*entry
,
5269 const char *procname
, void *data
, int maxlen
,
5270 mode_t mode
, proc_handler
*proc_handler
)
5272 entry
->procname
= procname
;
5274 entry
->maxlen
= maxlen
;
5276 entry
->proc_handler
= proc_handler
;
5279 static struct ctl_table
*
5280 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5282 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5284 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5285 sizeof(long), 0644, proc_doulongvec_minmax
);
5286 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5287 sizeof(long), 0644, proc_doulongvec_minmax
);
5288 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5289 sizeof(int), 0644, proc_dointvec_minmax
);
5290 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5291 sizeof(int), 0644, proc_dointvec_minmax
);
5292 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5293 sizeof(int), 0644, proc_dointvec_minmax
);
5294 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5295 sizeof(int), 0644, proc_dointvec_minmax
);
5296 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5297 sizeof(int), 0644, proc_dointvec_minmax
);
5298 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5299 sizeof(int), 0644, proc_dointvec_minmax
);
5300 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5301 sizeof(int), 0644, proc_dointvec_minmax
);
5302 set_table_entry(&table
[10], "cache_nice_tries",
5303 &sd
->cache_nice_tries
,
5304 sizeof(int), 0644, proc_dointvec_minmax
);
5305 set_table_entry(&table
[12], "flags", &sd
->flags
,
5306 sizeof(int), 0644, proc_dointvec_minmax
);
5311 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5313 struct ctl_table
*entry
, *table
;
5314 struct sched_domain
*sd
;
5315 int domain_num
= 0, i
;
5318 for_each_domain(cpu
, sd
)
5320 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5323 for_each_domain(cpu
, sd
) {
5324 snprintf(buf
, 32, "domain%d", i
);
5325 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5327 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5334 static struct ctl_table_header
*sd_sysctl_header
;
5335 static void init_sched_domain_sysctl(void)
5337 int i
, cpu_num
= num_online_cpus();
5338 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5341 sd_ctl_dir
[0].child
= entry
;
5343 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5344 snprintf(buf
, 32, "cpu%d", i
);
5345 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5347 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5349 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5352 static void init_sched_domain_sysctl(void)
5358 * migration_call - callback that gets triggered when a CPU is added.
5359 * Here we can start up the necessary migration thread for the new CPU.
5361 static int __cpuinit
5362 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5364 struct task_struct
*p
;
5365 int cpu
= (long)hcpu
;
5366 unsigned long flags
;
5370 case CPU_LOCK_ACQUIRE
:
5371 mutex_lock(&sched_hotcpu_mutex
);
5374 case CPU_UP_PREPARE
:
5375 case CPU_UP_PREPARE_FROZEN
:
5376 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5379 kthread_bind(p
, cpu
);
5380 /* Must be high prio: stop_machine expects to yield to it. */
5381 rq
= task_rq_lock(p
, &flags
);
5382 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5383 task_rq_unlock(rq
, &flags
);
5384 cpu_rq(cpu
)->migration_thread
= p
;
5388 case CPU_ONLINE_FROZEN
:
5389 /* Strictly unneccessary, as first user will wake it. */
5390 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5393 #ifdef CONFIG_HOTPLUG_CPU
5394 case CPU_UP_CANCELED
:
5395 case CPU_UP_CANCELED_FROZEN
:
5396 if (!cpu_rq(cpu
)->migration_thread
)
5398 /* Unbind it from offline cpu so it can run. Fall thru. */
5399 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5400 any_online_cpu(cpu_online_map
));
5401 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5402 cpu_rq(cpu
)->migration_thread
= NULL
;
5406 case CPU_DEAD_FROZEN
:
5407 migrate_live_tasks(cpu
);
5409 kthread_stop(rq
->migration_thread
);
5410 rq
->migration_thread
= NULL
;
5411 /* Idle task back to normal (off runqueue, low prio) */
5412 rq
= task_rq_lock(rq
->idle
, &flags
);
5413 update_rq_clock(rq
);
5414 deactivate_task(rq
, rq
->idle
, 0, rq
->clock
);
5415 rq
->idle
->static_prio
= MAX_PRIO
;
5416 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5417 rq
->idle
->sched_class
= &idle_sched_class
;
5418 migrate_dead_tasks(cpu
);
5419 task_rq_unlock(rq
, &flags
);
5420 migrate_nr_uninterruptible(rq
);
5421 BUG_ON(rq
->nr_running
!= 0);
5423 /* No need to migrate the tasks: it was best-effort if
5424 * they didn't take sched_hotcpu_mutex. Just wake up
5425 * the requestors. */
5426 spin_lock_irq(&rq
->lock
);
5427 while (!list_empty(&rq
->migration_queue
)) {
5428 struct migration_req
*req
;
5430 req
= list_entry(rq
->migration_queue
.next
,
5431 struct migration_req
, list
);
5432 list_del_init(&req
->list
);
5433 complete(&req
->done
);
5435 spin_unlock_irq(&rq
->lock
);
5438 case CPU_LOCK_RELEASE
:
5439 mutex_unlock(&sched_hotcpu_mutex
);
5445 /* Register at highest priority so that task migration (migrate_all_tasks)
5446 * happens before everything else.
5448 static struct notifier_block __cpuinitdata migration_notifier
= {
5449 .notifier_call
= migration_call
,
5453 int __init
migration_init(void)
5455 void *cpu
= (void *)(long)smp_processor_id();
5458 /* Start one for the boot CPU: */
5459 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5460 BUG_ON(err
== NOTIFY_BAD
);
5461 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5462 register_cpu_notifier(&migration_notifier
);
5470 /* Number of possible processor ids */
5471 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5472 EXPORT_SYMBOL(nr_cpu_ids
);
5474 #undef SCHED_DOMAIN_DEBUG
5475 #ifdef SCHED_DOMAIN_DEBUG
5476 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5481 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5485 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5490 struct sched_group
*group
= sd
->groups
;
5491 cpumask_t groupmask
;
5493 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5494 cpus_clear(groupmask
);
5497 for (i
= 0; i
< level
+ 1; i
++)
5499 printk("domain %d: ", level
);
5501 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5502 printk("does not load-balance\n");
5504 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5509 printk("span %s\n", str
);
5511 if (!cpu_isset(cpu
, sd
->span
))
5512 printk(KERN_ERR
"ERROR: domain->span does not contain "
5514 if (!cpu_isset(cpu
, group
->cpumask
))
5515 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5519 for (i
= 0; i
< level
+ 2; i
++)
5525 printk(KERN_ERR
"ERROR: group is NULL\n");
5529 if (!group
->__cpu_power
) {
5531 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5535 if (!cpus_weight(group
->cpumask
)) {
5537 printk(KERN_ERR
"ERROR: empty group\n");
5540 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5542 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5545 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5547 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5550 group
= group
->next
;
5551 } while (group
!= sd
->groups
);
5554 if (!cpus_equal(sd
->span
, groupmask
))
5555 printk(KERN_ERR
"ERROR: groups don't span "
5563 if (!cpus_subset(groupmask
, sd
->span
))
5564 printk(KERN_ERR
"ERROR: parent span is not a superset "
5565 "of domain->span\n");
5570 # define sched_domain_debug(sd, cpu) do { } while (0)
5573 static int sd_degenerate(struct sched_domain
*sd
)
5575 if (cpus_weight(sd
->span
) == 1)
5578 /* Following flags need at least 2 groups */
5579 if (sd
->flags
& (SD_LOAD_BALANCE
|
5580 SD_BALANCE_NEWIDLE
|
5584 SD_SHARE_PKG_RESOURCES
)) {
5585 if (sd
->groups
!= sd
->groups
->next
)
5589 /* Following flags don't use groups */
5590 if (sd
->flags
& (SD_WAKE_IDLE
|
5599 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5601 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5603 if (sd_degenerate(parent
))
5606 if (!cpus_equal(sd
->span
, parent
->span
))
5609 /* Does parent contain flags not in child? */
5610 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5611 if (cflags
& SD_WAKE_AFFINE
)
5612 pflags
&= ~SD_WAKE_BALANCE
;
5613 /* Flags needing groups don't count if only 1 group in parent */
5614 if (parent
->groups
== parent
->groups
->next
) {
5615 pflags
&= ~(SD_LOAD_BALANCE
|
5616 SD_BALANCE_NEWIDLE
|
5620 SD_SHARE_PKG_RESOURCES
);
5622 if (~cflags
& pflags
)
5629 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5630 * hold the hotplug lock.
5632 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5634 struct rq
*rq
= cpu_rq(cpu
);
5635 struct sched_domain
*tmp
;
5637 /* Remove the sched domains which do not contribute to scheduling. */
5638 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5639 struct sched_domain
*parent
= tmp
->parent
;
5642 if (sd_parent_degenerate(tmp
, parent
)) {
5643 tmp
->parent
= parent
->parent
;
5645 parent
->parent
->child
= tmp
;
5649 if (sd
&& sd_degenerate(sd
)) {
5655 sched_domain_debug(sd
, cpu
);
5657 rcu_assign_pointer(rq
->sd
, sd
);
5660 /* cpus with isolated domains */
5661 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5663 /* Setup the mask of cpus configured for isolated domains */
5664 static int __init
isolated_cpu_setup(char *str
)
5666 int ints
[NR_CPUS
], i
;
5668 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5669 cpus_clear(cpu_isolated_map
);
5670 for (i
= 1; i
<= ints
[0]; i
++)
5671 if (ints
[i
] < NR_CPUS
)
5672 cpu_set(ints
[i
], cpu_isolated_map
);
5676 __setup ("isolcpus=", isolated_cpu_setup
);
5679 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5680 * to a function which identifies what group(along with sched group) a CPU
5681 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5682 * (due to the fact that we keep track of groups covered with a cpumask_t).
5684 * init_sched_build_groups will build a circular linked list of the groups
5685 * covered by the given span, and will set each group's ->cpumask correctly,
5686 * and ->cpu_power to 0.
5689 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5690 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5691 struct sched_group
**sg
))
5693 struct sched_group
*first
= NULL
, *last
= NULL
;
5694 cpumask_t covered
= CPU_MASK_NONE
;
5697 for_each_cpu_mask(i
, span
) {
5698 struct sched_group
*sg
;
5699 int group
= group_fn(i
, cpu_map
, &sg
);
5702 if (cpu_isset(i
, covered
))
5705 sg
->cpumask
= CPU_MASK_NONE
;
5706 sg
->__cpu_power
= 0;
5708 for_each_cpu_mask(j
, span
) {
5709 if (group_fn(j
, cpu_map
, NULL
) != group
)
5712 cpu_set(j
, covered
);
5713 cpu_set(j
, sg
->cpumask
);
5724 #define SD_NODES_PER_DOMAIN 16
5729 * find_next_best_node - find the next node to include in a sched_domain
5730 * @node: node whose sched_domain we're building
5731 * @used_nodes: nodes already in the sched_domain
5733 * Find the next node to include in a given scheduling domain. Simply
5734 * finds the closest node not already in the @used_nodes map.
5736 * Should use nodemask_t.
5738 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5740 int i
, n
, val
, min_val
, best_node
= 0;
5744 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5745 /* Start at @node */
5746 n
= (node
+ i
) % MAX_NUMNODES
;
5748 if (!nr_cpus_node(n
))
5751 /* Skip already used nodes */
5752 if (test_bit(n
, used_nodes
))
5755 /* Simple min distance search */
5756 val
= node_distance(node
, n
);
5758 if (val
< min_val
) {
5764 set_bit(best_node
, used_nodes
);
5769 * sched_domain_node_span - get a cpumask for a node's sched_domain
5770 * @node: node whose cpumask we're constructing
5771 * @size: number of nodes to include in this span
5773 * Given a node, construct a good cpumask for its sched_domain to span. It
5774 * should be one that prevents unnecessary balancing, but also spreads tasks
5777 static cpumask_t
sched_domain_node_span(int node
)
5779 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5780 cpumask_t span
, nodemask
;
5784 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5786 nodemask
= node_to_cpumask(node
);
5787 cpus_or(span
, span
, nodemask
);
5788 set_bit(node
, used_nodes
);
5790 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5791 int next_node
= find_next_best_node(node
, used_nodes
);
5793 nodemask
= node_to_cpumask(next_node
);
5794 cpus_or(span
, span
, nodemask
);
5801 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5804 * SMT sched-domains:
5806 #ifdef CONFIG_SCHED_SMT
5807 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5808 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5810 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5811 struct sched_group
**sg
)
5814 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5820 * multi-core sched-domains:
5822 #ifdef CONFIG_SCHED_MC
5823 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5824 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5827 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5828 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5829 struct sched_group
**sg
)
5832 cpumask_t mask
= cpu_sibling_map
[cpu
];
5833 cpus_and(mask
, mask
, *cpu_map
);
5834 group
= first_cpu(mask
);
5836 *sg
= &per_cpu(sched_group_core
, group
);
5839 #elif defined(CONFIG_SCHED_MC)
5840 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5841 struct sched_group
**sg
)
5844 *sg
= &per_cpu(sched_group_core
, cpu
);
5849 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5850 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5852 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5853 struct sched_group
**sg
)
5856 #ifdef CONFIG_SCHED_MC
5857 cpumask_t mask
= cpu_coregroup_map(cpu
);
5858 cpus_and(mask
, mask
, *cpu_map
);
5859 group
= first_cpu(mask
);
5860 #elif defined(CONFIG_SCHED_SMT)
5861 cpumask_t mask
= cpu_sibling_map
[cpu
];
5862 cpus_and(mask
, mask
, *cpu_map
);
5863 group
= first_cpu(mask
);
5868 *sg
= &per_cpu(sched_group_phys
, group
);
5874 * The init_sched_build_groups can't handle what we want to do with node
5875 * groups, so roll our own. Now each node has its own list of groups which
5876 * gets dynamically allocated.
5878 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5879 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5881 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5882 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5884 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5885 struct sched_group
**sg
)
5887 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5890 cpus_and(nodemask
, nodemask
, *cpu_map
);
5891 group
= first_cpu(nodemask
);
5894 *sg
= &per_cpu(sched_group_allnodes
, group
);
5898 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5900 struct sched_group
*sg
= group_head
;
5906 for_each_cpu_mask(j
, sg
->cpumask
) {
5907 struct sched_domain
*sd
;
5909 sd
= &per_cpu(phys_domains
, j
);
5910 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5912 * Only add "power" once for each
5918 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5921 if (sg
!= group_head
)
5927 /* Free memory allocated for various sched_group structures */
5928 static void free_sched_groups(const cpumask_t
*cpu_map
)
5932 for_each_cpu_mask(cpu
, *cpu_map
) {
5933 struct sched_group
**sched_group_nodes
5934 = sched_group_nodes_bycpu
[cpu
];
5936 if (!sched_group_nodes
)
5939 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5940 cpumask_t nodemask
= node_to_cpumask(i
);
5941 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5943 cpus_and(nodemask
, nodemask
, *cpu_map
);
5944 if (cpus_empty(nodemask
))
5954 if (oldsg
!= sched_group_nodes
[i
])
5957 kfree(sched_group_nodes
);
5958 sched_group_nodes_bycpu
[cpu
] = NULL
;
5962 static void free_sched_groups(const cpumask_t
*cpu_map
)
5968 * Initialize sched groups cpu_power.
5970 * cpu_power indicates the capacity of sched group, which is used while
5971 * distributing the load between different sched groups in a sched domain.
5972 * Typically cpu_power for all the groups in a sched domain will be same unless
5973 * there are asymmetries in the topology. If there are asymmetries, group
5974 * having more cpu_power will pickup more load compared to the group having
5977 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5978 * the maximum number of tasks a group can handle in the presence of other idle
5979 * or lightly loaded groups in the same sched domain.
5981 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5983 struct sched_domain
*child
;
5984 struct sched_group
*group
;
5986 WARN_ON(!sd
|| !sd
->groups
);
5988 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5993 sd
->groups
->__cpu_power
= 0;
5996 * For perf policy, if the groups in child domain share resources
5997 * (for example cores sharing some portions of the cache hierarchy
5998 * or SMT), then set this domain groups cpu_power such that each group
5999 * can handle only one task, when there are other idle groups in the
6000 * same sched domain.
6002 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6004 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6005 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6010 * add cpu_power of each child group to this groups cpu_power
6012 group
= child
->groups
;
6014 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6015 group
= group
->next
;
6016 } while (group
!= child
->groups
);
6020 * Build sched domains for a given set of cpus and attach the sched domains
6021 * to the individual cpus
6023 static int build_sched_domains(const cpumask_t
*cpu_map
)
6027 struct sched_group
**sched_group_nodes
= NULL
;
6028 int sd_allnodes
= 0;
6031 * Allocate the per-node list of sched groups
6033 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6035 if (!sched_group_nodes
) {
6036 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6039 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6043 * Set up domains for cpus specified by the cpu_map.
6045 for_each_cpu_mask(i
, *cpu_map
) {
6046 struct sched_domain
*sd
= NULL
, *p
;
6047 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6049 cpus_and(nodemask
, nodemask
, *cpu_map
);
6052 if (cpus_weight(*cpu_map
) >
6053 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6054 sd
= &per_cpu(allnodes_domains
, i
);
6055 *sd
= SD_ALLNODES_INIT
;
6056 sd
->span
= *cpu_map
;
6057 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6063 sd
= &per_cpu(node_domains
, i
);
6065 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6069 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6073 sd
= &per_cpu(phys_domains
, i
);
6075 sd
->span
= nodemask
;
6079 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6081 #ifdef CONFIG_SCHED_MC
6083 sd
= &per_cpu(core_domains
, i
);
6085 sd
->span
= cpu_coregroup_map(i
);
6086 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6089 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6092 #ifdef CONFIG_SCHED_SMT
6094 sd
= &per_cpu(cpu_domains
, i
);
6095 *sd
= SD_SIBLING_INIT
;
6096 sd
->span
= cpu_sibling_map
[i
];
6097 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6100 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6104 #ifdef CONFIG_SCHED_SMT
6105 /* Set up CPU (sibling) groups */
6106 for_each_cpu_mask(i
, *cpu_map
) {
6107 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6108 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6109 if (i
!= first_cpu(this_sibling_map
))
6112 init_sched_build_groups(this_sibling_map
, cpu_map
,
6117 #ifdef CONFIG_SCHED_MC
6118 /* Set up multi-core groups */
6119 for_each_cpu_mask(i
, *cpu_map
) {
6120 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6121 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6122 if (i
!= first_cpu(this_core_map
))
6124 init_sched_build_groups(this_core_map
, cpu_map
,
6125 &cpu_to_core_group
);
6129 /* Set up physical groups */
6130 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6131 cpumask_t nodemask
= node_to_cpumask(i
);
6133 cpus_and(nodemask
, nodemask
, *cpu_map
);
6134 if (cpus_empty(nodemask
))
6137 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6141 /* Set up node groups */
6143 init_sched_build_groups(*cpu_map
, cpu_map
,
6144 &cpu_to_allnodes_group
);
6146 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6147 /* Set up node groups */
6148 struct sched_group
*sg
, *prev
;
6149 cpumask_t nodemask
= node_to_cpumask(i
);
6150 cpumask_t domainspan
;
6151 cpumask_t covered
= CPU_MASK_NONE
;
6154 cpus_and(nodemask
, nodemask
, *cpu_map
);
6155 if (cpus_empty(nodemask
)) {
6156 sched_group_nodes
[i
] = NULL
;
6160 domainspan
= sched_domain_node_span(i
);
6161 cpus_and(domainspan
, domainspan
, *cpu_map
);
6163 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6165 printk(KERN_WARNING
"Can not alloc domain group for "
6169 sched_group_nodes
[i
] = sg
;
6170 for_each_cpu_mask(j
, nodemask
) {
6171 struct sched_domain
*sd
;
6173 sd
= &per_cpu(node_domains
, j
);
6176 sg
->__cpu_power
= 0;
6177 sg
->cpumask
= nodemask
;
6179 cpus_or(covered
, covered
, nodemask
);
6182 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6183 cpumask_t tmp
, notcovered
;
6184 int n
= (i
+ j
) % MAX_NUMNODES
;
6186 cpus_complement(notcovered
, covered
);
6187 cpus_and(tmp
, notcovered
, *cpu_map
);
6188 cpus_and(tmp
, tmp
, domainspan
);
6189 if (cpus_empty(tmp
))
6192 nodemask
= node_to_cpumask(n
);
6193 cpus_and(tmp
, tmp
, nodemask
);
6194 if (cpus_empty(tmp
))
6197 sg
= kmalloc_node(sizeof(struct sched_group
),
6201 "Can not alloc domain group for node %d\n", j
);
6204 sg
->__cpu_power
= 0;
6206 sg
->next
= prev
->next
;
6207 cpus_or(covered
, covered
, tmp
);
6214 /* Calculate CPU power for physical packages and nodes */
6215 #ifdef CONFIG_SCHED_SMT
6216 for_each_cpu_mask(i
, *cpu_map
) {
6217 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6219 init_sched_groups_power(i
, sd
);
6222 #ifdef CONFIG_SCHED_MC
6223 for_each_cpu_mask(i
, *cpu_map
) {
6224 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6226 init_sched_groups_power(i
, sd
);
6230 for_each_cpu_mask(i
, *cpu_map
) {
6231 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6233 init_sched_groups_power(i
, sd
);
6237 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6238 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6241 struct sched_group
*sg
;
6243 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6244 init_numa_sched_groups_power(sg
);
6248 /* Attach the domains */
6249 for_each_cpu_mask(i
, *cpu_map
) {
6250 struct sched_domain
*sd
;
6251 #ifdef CONFIG_SCHED_SMT
6252 sd
= &per_cpu(cpu_domains
, i
);
6253 #elif defined(CONFIG_SCHED_MC)
6254 sd
= &per_cpu(core_domains
, i
);
6256 sd
= &per_cpu(phys_domains
, i
);
6258 cpu_attach_domain(sd
, i
);
6265 free_sched_groups(cpu_map
);
6270 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6272 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6274 cpumask_t cpu_default_map
;
6278 * Setup mask for cpus without special case scheduling requirements.
6279 * For now this just excludes isolated cpus, but could be used to
6280 * exclude other special cases in the future.
6282 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6284 err
= build_sched_domains(&cpu_default_map
);
6289 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6291 free_sched_groups(cpu_map
);
6295 * Detach sched domains from a group of cpus specified in cpu_map
6296 * These cpus will now be attached to the NULL domain
6298 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6302 for_each_cpu_mask(i
, *cpu_map
)
6303 cpu_attach_domain(NULL
, i
);
6304 synchronize_sched();
6305 arch_destroy_sched_domains(cpu_map
);
6309 * Partition sched domains as specified by the cpumasks below.
6310 * This attaches all cpus from the cpumasks to the NULL domain,
6311 * waits for a RCU quiescent period, recalculates sched
6312 * domain information and then attaches them back to the
6313 * correct sched domains
6314 * Call with hotplug lock held
6316 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6318 cpumask_t change_map
;
6321 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6322 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6323 cpus_or(change_map
, *partition1
, *partition2
);
6325 /* Detach sched domains from all of the affected cpus */
6326 detach_destroy_domains(&change_map
);
6327 if (!cpus_empty(*partition1
))
6328 err
= build_sched_domains(partition1
);
6329 if (!err
&& !cpus_empty(*partition2
))
6330 err
= build_sched_domains(partition2
);
6335 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6336 int arch_reinit_sched_domains(void)
6340 mutex_lock(&sched_hotcpu_mutex
);
6341 detach_destroy_domains(&cpu_online_map
);
6342 err
= arch_init_sched_domains(&cpu_online_map
);
6343 mutex_unlock(&sched_hotcpu_mutex
);
6348 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6352 if (buf
[0] != '0' && buf
[0] != '1')
6356 sched_smt_power_savings
= (buf
[0] == '1');
6358 sched_mc_power_savings
= (buf
[0] == '1');
6360 ret
= arch_reinit_sched_domains();
6362 return ret
? ret
: count
;
6365 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6369 #ifdef CONFIG_SCHED_SMT
6371 err
= sysfs_create_file(&cls
->kset
.kobj
,
6372 &attr_sched_smt_power_savings
.attr
);
6374 #ifdef CONFIG_SCHED_MC
6375 if (!err
&& mc_capable())
6376 err
= sysfs_create_file(&cls
->kset
.kobj
,
6377 &attr_sched_mc_power_savings
.attr
);
6383 #ifdef CONFIG_SCHED_MC
6384 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6386 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6388 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6389 const char *buf
, size_t count
)
6391 return sched_power_savings_store(buf
, count
, 0);
6393 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6394 sched_mc_power_savings_store
);
6397 #ifdef CONFIG_SCHED_SMT
6398 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6400 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6402 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6403 const char *buf
, size_t count
)
6405 return sched_power_savings_store(buf
, count
, 1);
6407 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6408 sched_smt_power_savings_store
);
6412 * Force a reinitialization of the sched domains hierarchy. The domains
6413 * and groups cannot be updated in place without racing with the balancing
6414 * code, so we temporarily attach all running cpus to the NULL domain
6415 * which will prevent rebalancing while the sched domains are recalculated.
6417 static int update_sched_domains(struct notifier_block
*nfb
,
6418 unsigned long action
, void *hcpu
)
6421 case CPU_UP_PREPARE
:
6422 case CPU_UP_PREPARE_FROZEN
:
6423 case CPU_DOWN_PREPARE
:
6424 case CPU_DOWN_PREPARE_FROZEN
:
6425 detach_destroy_domains(&cpu_online_map
);
6428 case CPU_UP_CANCELED
:
6429 case CPU_UP_CANCELED_FROZEN
:
6430 case CPU_DOWN_FAILED
:
6431 case CPU_DOWN_FAILED_FROZEN
:
6433 case CPU_ONLINE_FROZEN
:
6435 case CPU_DEAD_FROZEN
:
6437 * Fall through and re-initialise the domains.
6444 /* The hotplug lock is already held by cpu_up/cpu_down */
6445 arch_init_sched_domains(&cpu_online_map
);
6450 void __init
sched_init_smp(void)
6452 cpumask_t non_isolated_cpus
;
6454 mutex_lock(&sched_hotcpu_mutex
);
6455 arch_init_sched_domains(&cpu_online_map
);
6456 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6457 if (cpus_empty(non_isolated_cpus
))
6458 cpu_set(smp_processor_id(), non_isolated_cpus
);
6459 mutex_unlock(&sched_hotcpu_mutex
);
6460 /* XXX: Theoretical race here - CPU may be hotplugged now */
6461 hotcpu_notifier(update_sched_domains
, 0);
6463 init_sched_domain_sysctl();
6465 /* Move init over to a non-isolated CPU */
6466 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6468 sched_init_granularity();
6471 void __init
sched_init_smp(void)
6473 sched_init_granularity();
6475 #endif /* CONFIG_SMP */
6477 int in_sched_functions(unsigned long addr
)
6479 /* Linker adds these: start and end of __sched functions */
6480 extern char __sched_text_start
[], __sched_text_end
[];
6482 return in_lock_functions(addr
) ||
6483 (addr
>= (unsigned long)__sched_text_start
6484 && addr
< (unsigned long)__sched_text_end
);
6487 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6489 cfs_rq
->tasks_timeline
= RB_ROOT
;
6490 cfs_rq
->fair_clock
= 1;
6491 #ifdef CONFIG_FAIR_GROUP_SCHED
6496 void __init
sched_init(void)
6498 u64 now
= sched_clock();
6499 int highest_cpu
= 0;
6503 * Link up the scheduling class hierarchy:
6505 rt_sched_class
.next
= &fair_sched_class
;
6506 fair_sched_class
.next
= &idle_sched_class
;
6507 idle_sched_class
.next
= NULL
;
6509 for_each_possible_cpu(i
) {
6510 struct rt_prio_array
*array
;
6514 spin_lock_init(&rq
->lock
);
6515 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6518 init_cfs_rq(&rq
->cfs
, rq
);
6519 #ifdef CONFIG_FAIR_GROUP_SCHED
6520 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6521 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6523 rq
->ls
.load_update_last
= now
;
6524 rq
->ls
.load_update_start
= now
;
6526 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6527 rq
->cpu_load
[j
] = 0;
6530 rq
->active_balance
= 0;
6531 rq
->next_balance
= jiffies
;
6534 rq
->migration_thread
= NULL
;
6535 INIT_LIST_HEAD(&rq
->migration_queue
);
6537 atomic_set(&rq
->nr_iowait
, 0);
6539 array
= &rq
->rt
.active
;
6540 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6541 INIT_LIST_HEAD(array
->queue
+ j
);
6542 __clear_bit(j
, array
->bitmap
);
6545 /* delimiter for bitsearch: */
6546 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6549 set_load_weight(&init_task
);
6551 #ifdef CONFIG_PREEMPT_NOTIFIERS
6552 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6556 nr_cpu_ids
= highest_cpu
+ 1;
6557 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6560 #ifdef CONFIG_RT_MUTEXES
6561 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6565 * The boot idle thread does lazy MMU switching as well:
6567 atomic_inc(&init_mm
.mm_count
);
6568 enter_lazy_tlb(&init_mm
, current
);
6571 * Make us the idle thread. Technically, schedule() should not be
6572 * called from this thread, however somewhere below it might be,
6573 * but because we are the idle thread, we just pick up running again
6574 * when this runqueue becomes "idle".
6576 init_idle(current
, smp_processor_id());
6578 * During early bootup we pretend to be a normal task:
6580 current
->sched_class
= &fair_sched_class
;
6583 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6584 void __might_sleep(char *file
, int line
)
6587 static unsigned long prev_jiffy
; /* ratelimiting */
6589 if ((in_atomic() || irqs_disabled()) &&
6590 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6591 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6593 prev_jiffy
= jiffies
;
6594 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6595 " context at %s:%d\n", file
, line
);
6596 printk("in_atomic():%d, irqs_disabled():%d\n",
6597 in_atomic(), irqs_disabled());
6598 debug_show_held_locks(current
);
6599 if (irqs_disabled())
6600 print_irqtrace_events(current
);
6605 EXPORT_SYMBOL(__might_sleep
);
6608 #ifdef CONFIG_MAGIC_SYSRQ
6609 void normalize_rt_tasks(void)
6611 struct task_struct
*g
, *p
;
6612 unsigned long flags
;
6616 read_lock_irq(&tasklist_lock
);
6617 do_each_thread(g
, p
) {
6619 p
->se
.wait_runtime
= 0;
6620 p
->se
.exec_start
= 0;
6621 p
->se
.wait_start_fair
= 0;
6622 p
->se
.sleep_start_fair
= 0;
6623 #ifdef CONFIG_SCHEDSTATS
6624 p
->se
.wait_start
= 0;
6625 p
->se
.sleep_start
= 0;
6626 p
->se
.block_start
= 0;
6628 task_rq(p
)->cfs
.fair_clock
= 0;
6629 task_rq(p
)->clock
= 0;
6633 * Renice negative nice level userspace
6636 if (TASK_NICE(p
) < 0 && p
->mm
)
6637 set_user_nice(p
, 0);
6641 spin_lock_irqsave(&p
->pi_lock
, flags
);
6642 rq
= __task_rq_lock(p
);
6645 * Do not touch the migration thread:
6647 if (p
== rq
->migration_thread
)
6651 on_rq
= p
->se
.on_rq
;
6653 update_rq_clock(task_rq(p
));
6654 deactivate_task(task_rq(p
), p
, 0, task_rq(p
)->clock
);
6656 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6658 activate_task(task_rq(p
), p
, 0);
6659 resched_task(rq
->curr
);
6664 __task_rq_unlock(rq
);
6665 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6666 } while_each_thread(g
, p
);
6668 read_unlock_irq(&tasklist_lock
);
6671 #endif /* CONFIG_MAGIC_SYSRQ */
6675 * These functions are only useful for the IA64 MCA handling.
6677 * They can only be called when the whole system has been
6678 * stopped - every CPU needs to be quiescent, and no scheduling
6679 * activity can take place. Using them for anything else would
6680 * be a serious bug, and as a result, they aren't even visible
6681 * under any other configuration.
6685 * curr_task - return the current task for a given cpu.
6686 * @cpu: the processor in question.
6688 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6690 struct task_struct
*curr_task(int cpu
)
6692 return cpu_curr(cpu
);
6696 * set_curr_task - set the current task for a given cpu.
6697 * @cpu: the processor in question.
6698 * @p: the task pointer to set.
6700 * Description: This function must only be used when non-maskable interrupts
6701 * are serviced on a separate stack. It allows the architecture to switch the
6702 * notion of the current task on a cpu in a non-blocking manner. This function
6703 * must be called with all CPU's synchronized, and interrupts disabled, the
6704 * and caller must save the original value of the current task (see
6705 * curr_task() above) and restore that value before reenabling interrupts and
6706 * re-starting the system.
6708 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6710 void set_curr_task(int cpu
, struct task_struct
*p
)