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
);
371 static u64
rq_clock(struct rq
*rq
)
378 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
379 * See detach_destroy_domains: synchronize_sched for details.
381 * The domain tree of any CPU may only be accessed from within
382 * preempt-disabled sections.
384 #define for_each_domain(cpu, __sd) \
385 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
387 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
388 #define this_rq() (&__get_cpu_var(runqueues))
389 #define task_rq(p) cpu_rq(task_cpu(p))
390 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
393 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
394 * clock constructed from sched_clock():
396 unsigned long long cpu_clock(int cpu
)
398 unsigned long long now
;
402 local_irq_save(flags
);
406 local_irq_restore(flags
);
411 #ifdef CONFIG_FAIR_GROUP_SCHED
412 /* Change a task's ->cfs_rq if it moves across CPUs */
413 static inline void set_task_cfs_rq(struct task_struct
*p
)
415 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
418 static inline void set_task_cfs_rq(struct task_struct
*p
)
423 #ifndef prepare_arch_switch
424 # define prepare_arch_switch(next) do { } while (0)
426 #ifndef finish_arch_switch
427 # define finish_arch_switch(prev) do { } while (0)
430 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
431 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
433 return rq
->curr
== p
;
436 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
440 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
442 #ifdef CONFIG_DEBUG_SPINLOCK
443 /* this is a valid case when another task releases the spinlock */
444 rq
->lock
.owner
= current
;
447 * If we are tracking spinlock dependencies then we have to
448 * fix up the runqueue lock - which gets 'carried over' from
451 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
453 spin_unlock_irq(&rq
->lock
);
456 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
457 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
462 return rq
->curr
== p
;
466 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
470 * We can optimise this out completely for !SMP, because the
471 * SMP rebalancing from interrupt is the only thing that cares
476 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
477 spin_unlock_irq(&rq
->lock
);
479 spin_unlock(&rq
->lock
);
483 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
487 * After ->oncpu is cleared, the task can be moved to a different CPU.
488 * We must ensure this doesn't happen until the switch is completely
494 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
498 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
501 * __task_rq_lock - lock the runqueue a given task resides on.
502 * Must be called interrupts disabled.
504 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
511 spin_lock(&rq
->lock
);
512 if (unlikely(rq
!= task_rq(p
))) {
513 spin_unlock(&rq
->lock
);
514 goto repeat_lock_task
;
520 * task_rq_lock - lock the runqueue a given task resides on and disable
521 * interrupts. Note the ordering: we can safely lookup the task_rq without
522 * explicitly disabling preemption.
524 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
530 local_irq_save(*flags
);
532 spin_lock(&rq
->lock
);
533 if (unlikely(rq
!= task_rq(p
))) {
534 spin_unlock_irqrestore(&rq
->lock
, *flags
);
535 goto repeat_lock_task
;
540 static inline void __task_rq_unlock(struct rq
*rq
)
543 spin_unlock(&rq
->lock
);
546 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
549 spin_unlock_irqrestore(&rq
->lock
, *flags
);
553 * this_rq_lock - lock this runqueue and disable interrupts.
555 static inline struct rq
*this_rq_lock(void)
562 spin_lock(&rq
->lock
);
568 * CPU frequency is/was unstable - start new by setting prev_clock_raw:
570 void sched_clock_unstable_event(void)
575 rq
= task_rq_lock(current
, &flags
);
576 rq
->prev_clock_raw
= sched_clock();
577 rq
->clock_unstable_events
++;
578 task_rq_unlock(rq
, &flags
);
582 * resched_task - mark a task 'to be rescheduled now'.
584 * On UP this means the setting of the need_resched flag, on SMP it
585 * might also involve a cross-CPU call to trigger the scheduler on
590 #ifndef tsk_is_polling
591 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
594 static void resched_task(struct task_struct
*p
)
598 assert_spin_locked(&task_rq(p
)->lock
);
600 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
603 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
606 if (cpu
== smp_processor_id())
609 /* NEED_RESCHED must be visible before we test polling */
611 if (!tsk_is_polling(p
))
612 smp_send_reschedule(cpu
);
615 static void resched_cpu(int cpu
)
617 struct rq
*rq
= cpu_rq(cpu
);
620 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
622 resched_task(cpu_curr(cpu
));
623 spin_unlock_irqrestore(&rq
->lock
, flags
);
626 static inline void resched_task(struct task_struct
*p
)
628 assert_spin_locked(&task_rq(p
)->lock
);
629 set_tsk_need_resched(p
);
633 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
635 #if BITS_PER_LONG == 32
636 if (likely(divident
<= 0xffffffffULL
))
637 return (u32
)divident
/ divisor
;
638 do_div(divident
, divisor
);
642 return divident
/ divisor
;
646 #if BITS_PER_LONG == 32
647 # define WMULT_CONST (~0UL)
649 # define WMULT_CONST (1UL << 32)
652 #define WMULT_SHIFT 32
655 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
656 struct load_weight
*lw
)
660 if (unlikely(!lw
->inv_weight
))
661 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
663 tmp
= (u64
)delta_exec
* weight
;
665 * Check whether we'd overflow the 64-bit multiplication:
667 if (unlikely(tmp
> WMULT_CONST
)) {
668 tmp
= ((tmp
>> WMULT_SHIFT
/2) * lw
->inv_weight
)
671 tmp
= (tmp
* lw
->inv_weight
) >> WMULT_SHIFT
;
674 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
677 static inline unsigned long
678 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
680 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
683 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
689 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
696 * To aid in avoiding the subversion of "niceness" due to uneven distribution
697 * of tasks with abnormal "nice" values across CPUs the contribution that
698 * each task makes to its run queue's load is weighted according to its
699 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
700 * scaled version of the new time slice allocation that they receive on time
704 #define WEIGHT_IDLEPRIO 2
705 #define WMULT_IDLEPRIO (1 << 31)
708 * Nice levels are multiplicative, with a gentle 10% change for every
709 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
710 * nice 1, it will get ~10% less CPU time than another CPU-bound task
711 * that remained on nice 0.
713 * The "10% effect" is relative and cumulative: from _any_ nice level,
714 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
715 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
716 * If a task goes up by ~10% and another task goes down by ~10% then
717 * the relative distance between them is ~25%.)
719 static const int prio_to_weight
[40] = {
720 /* -20 */ 88818, 71054, 56843, 45475, 36380, 29104, 23283, 18626, 14901, 11921,
721 /* -10 */ 9537, 7629, 6103, 4883, 3906, 3125, 2500, 2000, 1600, 1280,
722 /* 0 */ NICE_0_LOAD
/* 1024 */,
723 /* 1 */ 819, 655, 524, 419, 336, 268, 215, 172, 137,
724 /* 10 */ 110, 87, 70, 56, 45, 36, 29, 23, 18, 15,
728 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
730 * In cases where the weight does not change often, we can use the
731 * precalculated inverse to speed up arithmetics by turning divisions
732 * into multiplications:
734 static const u32 prio_to_wmult
[40] = {
735 /* -20 */ 48356, 60446, 75558, 94446, 118058,
736 /* -15 */ 147573, 184467, 230589, 288233, 360285,
737 /* -10 */ 450347, 562979, 703746, 879575, 1099582,
738 /* -5 */ 1374389, 1717986, 2147483, 2684354, 3355443,
739 /* 0 */ 4194304, 5244160, 6557201, 8196502, 10250518,
740 /* 5 */ 12782640, 16025997, 19976592, 24970740, 31350126,
741 /* 10 */ 39045157, 49367440, 61356675, 76695844, 95443717,
742 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
745 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
748 * runqueue iterator, to support SMP load-balancing between different
749 * scheduling classes, without having to expose their internal data
750 * structures to the load-balancing proper:
754 struct task_struct
*(*start
)(void *);
755 struct task_struct
*(*next
)(void *);
758 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
759 unsigned long max_nr_move
, unsigned long max_load_move
,
760 struct sched_domain
*sd
, enum cpu_idle_type idle
,
761 int *all_pinned
, unsigned long *load_moved
,
762 int *this_best_prio
, struct rq_iterator
*iterator
);
764 #include "sched_stats.h"
765 #include "sched_rt.c"
766 #include "sched_fair.c"
767 #include "sched_idletask.c"
768 #ifdef CONFIG_SCHED_DEBUG
769 # include "sched_debug.c"
772 #define sched_class_highest (&rt_sched_class)
774 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
776 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
777 ls
->delta_exec
+= ls
->delta_stat
;
778 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
784 * Update delta_exec, delta_fair fields for rq.
786 * delta_fair clock advances at a rate inversely proportional to
787 * total load (rq->ls.load.weight) on the runqueue, while
788 * delta_exec advances at the same rate as wall-clock (provided
791 * delta_exec / delta_fair is a measure of the (smoothened) load on this
792 * runqueue over any given interval. This (smoothened) load is used
793 * during load balance.
795 * This function is called /before/ updating rq->ls.load
796 * and when switching tasks.
798 static void update_curr_load(struct rq
*rq
, u64 now
)
800 struct load_stat
*ls
= &rq
->ls
;
803 start
= ls
->load_update_start
;
804 ls
->load_update_start
= now
;
805 ls
->delta_stat
+= now
- start
;
807 * Stagger updates to ls->delta_fair. Very frequent updates
810 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
811 __update_curr_load(rq
, ls
);
815 inc_load(struct rq
*rq
, const struct task_struct
*p
, u64 now
)
817 update_curr_load(rq
, now
);
818 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
822 dec_load(struct rq
*rq
, const struct task_struct
*p
, u64 now
)
824 update_curr_load(rq
, now
);
825 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
828 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
, u64 now
)
831 inc_load(rq
, p
, now
);
834 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
, u64 now
)
837 dec_load(rq
, p
, now
);
840 static void set_load_weight(struct task_struct
*p
)
842 task_rq(p
)->cfs
.wait_runtime
-= p
->se
.wait_runtime
;
843 p
->se
.wait_runtime
= 0;
845 if (task_has_rt_policy(p
)) {
846 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
847 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
852 * SCHED_IDLE tasks get minimal weight:
854 if (p
->policy
== SCHED_IDLE
) {
855 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
856 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
860 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
861 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
865 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, u64 now
)
867 sched_info_queued(p
);
868 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, now
);
873 dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
875 p
->sched_class
->dequeue_task(rq
, p
, sleep
, now
);
880 * __normal_prio - return the priority that is based on the static prio
882 static inline int __normal_prio(struct task_struct
*p
)
884 return p
->static_prio
;
888 * Calculate the expected normal priority: i.e. priority
889 * without taking RT-inheritance into account. Might be
890 * boosted by interactivity modifiers. Changes upon fork,
891 * setprio syscalls, and whenever the interactivity
892 * estimator recalculates.
894 static inline int normal_prio(struct task_struct
*p
)
898 if (task_has_rt_policy(p
))
899 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
901 prio
= __normal_prio(p
);
906 * Calculate the current priority, i.e. the priority
907 * taken into account by the scheduler. This value might
908 * be boosted by RT tasks, or might be boosted by
909 * interactivity modifiers. Will be RT if the task got
910 * RT-boosted. If not then it returns p->normal_prio.
912 static int effective_prio(struct task_struct
*p
)
914 p
->normal_prio
= normal_prio(p
);
916 * If we are RT tasks or we were boosted to RT priority,
917 * keep the priority unchanged. Otherwise, update priority
918 * to the normal priority:
920 if (!rt_prio(p
->prio
))
921 return p
->normal_prio
;
926 * activate_task - move a task to the runqueue.
928 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
935 if (p
->state
== TASK_UNINTERRUPTIBLE
)
936 rq
->nr_uninterruptible
--;
938 enqueue_task(rq
, p
, wakeup
, now
);
939 inc_nr_running(p
, rq
, now
);
943 * activate_idle_task - move idle task to the _front_ of runqueue.
945 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
952 if (p
->state
== TASK_UNINTERRUPTIBLE
)
953 rq
->nr_uninterruptible
--;
955 enqueue_task(rq
, p
, 0, now
);
956 inc_nr_running(p
, rq
, now
);
960 * deactivate_task - remove a task from the runqueue.
963 deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
965 if (p
->state
== TASK_UNINTERRUPTIBLE
)
966 rq
->nr_uninterruptible
++;
968 dequeue_task(rq
, p
, sleep
, now
);
969 dec_nr_running(p
, rq
, now
);
973 * task_curr - is this task currently executing on a CPU?
974 * @p: the task in question.
976 inline int task_curr(const struct task_struct
*p
)
978 return cpu_curr(task_cpu(p
)) == p
;
981 /* Used instead of source_load when we know the type == 0 */
982 unsigned long weighted_cpuload(const int cpu
)
984 return cpu_rq(cpu
)->ls
.load
.weight
;
987 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
990 task_thread_info(p
)->cpu
= cpu
;
997 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
999 int old_cpu
= task_cpu(p
);
1000 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1001 u64 clock_offset
, fair_clock_offset
;
1003 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1004 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
1006 if (p
->se
.wait_start_fair
)
1007 p
->se
.wait_start_fair
-= fair_clock_offset
;
1008 if (p
->se
.sleep_start_fair
)
1009 p
->se
.sleep_start_fair
-= fair_clock_offset
;
1011 #ifdef CONFIG_SCHEDSTATS
1012 if (p
->se
.wait_start
)
1013 p
->se
.wait_start
-= clock_offset
;
1014 if (p
->se
.sleep_start
)
1015 p
->se
.sleep_start
-= clock_offset
;
1016 if (p
->se
.block_start
)
1017 p
->se
.block_start
-= clock_offset
;
1020 __set_task_cpu(p
, new_cpu
);
1023 struct migration_req
{
1024 struct list_head list
;
1026 struct task_struct
*task
;
1029 struct completion done
;
1033 * The task's runqueue lock must be held.
1034 * Returns true if you have to wait for migration thread.
1037 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1039 struct rq
*rq
= task_rq(p
);
1042 * If the task is not on a runqueue (and not running), then
1043 * it is sufficient to simply update the task's cpu field.
1045 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1046 set_task_cpu(p
, dest_cpu
);
1050 init_completion(&req
->done
);
1052 req
->dest_cpu
= dest_cpu
;
1053 list_add(&req
->list
, &rq
->migration_queue
);
1059 * wait_task_inactive - wait for a thread to unschedule.
1061 * The caller must ensure that the task *will* unschedule sometime soon,
1062 * else this function might spin for a *long* time. This function can't
1063 * be called with interrupts off, or it may introduce deadlock with
1064 * smp_call_function() if an IPI is sent by the same process we are
1065 * waiting to become inactive.
1067 void wait_task_inactive(struct task_struct
*p
)
1069 unsigned long flags
;
1075 * We do the initial early heuristics without holding
1076 * any task-queue locks at all. We'll only try to get
1077 * the runqueue lock when things look like they will
1083 * If the task is actively running on another CPU
1084 * still, just relax and busy-wait without holding
1087 * NOTE! Since we don't hold any locks, it's not
1088 * even sure that "rq" stays as the right runqueue!
1089 * But we don't care, since "task_running()" will
1090 * return false if the runqueue has changed and p
1091 * is actually now running somewhere else!
1093 while (task_running(rq
, p
))
1097 * Ok, time to look more closely! We need the rq
1098 * lock now, to be *sure*. If we're wrong, we'll
1099 * just go back and repeat.
1101 rq
= task_rq_lock(p
, &flags
);
1102 running
= task_running(rq
, p
);
1103 on_rq
= p
->se
.on_rq
;
1104 task_rq_unlock(rq
, &flags
);
1107 * Was it really running after all now that we
1108 * checked with the proper locks actually held?
1110 * Oops. Go back and try again..
1112 if (unlikely(running
)) {
1118 * It's not enough that it's not actively running,
1119 * it must be off the runqueue _entirely_, and not
1122 * So if it wa still runnable (but just not actively
1123 * running right now), it's preempted, and we should
1124 * yield - it could be a while.
1126 if (unlikely(on_rq
)) {
1132 * Ahh, all good. It wasn't running, and it wasn't
1133 * runnable, which means that it will never become
1134 * running in the future either. We're all done!
1139 * kick_process - kick a running thread to enter/exit the kernel
1140 * @p: the to-be-kicked thread
1142 * Cause a process which is running on another CPU to enter
1143 * kernel-mode, without any delay. (to get signals handled.)
1145 * NOTE: this function doesnt have to take the runqueue lock,
1146 * because all it wants to ensure is that the remote task enters
1147 * the kernel. If the IPI races and the task has been migrated
1148 * to another CPU then no harm is done and the purpose has been
1151 void kick_process(struct task_struct
*p
)
1157 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1158 smp_send_reschedule(cpu
);
1163 * Return a low guess at the load of a migration-source cpu weighted
1164 * according to the scheduling class and "nice" value.
1166 * We want to under-estimate the load of migration sources, to
1167 * balance conservatively.
1169 static inline unsigned long source_load(int cpu
, int type
)
1171 struct rq
*rq
= cpu_rq(cpu
);
1172 unsigned long total
= weighted_cpuload(cpu
);
1177 return min(rq
->cpu_load
[type
-1], total
);
1181 * Return a high guess at the load of a migration-target cpu weighted
1182 * according to the scheduling class and "nice" value.
1184 static inline unsigned long target_load(int cpu
, int type
)
1186 struct rq
*rq
= cpu_rq(cpu
);
1187 unsigned long total
= weighted_cpuload(cpu
);
1192 return max(rq
->cpu_load
[type
-1], total
);
1196 * Return the average load per task on the cpu's run queue
1198 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1200 struct rq
*rq
= cpu_rq(cpu
);
1201 unsigned long total
= weighted_cpuload(cpu
);
1202 unsigned long n
= rq
->nr_running
;
1204 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1208 * find_idlest_group finds and returns the least busy CPU group within the
1211 static struct sched_group
*
1212 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1214 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1215 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1216 int load_idx
= sd
->forkexec_idx
;
1217 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1220 unsigned long load
, avg_load
;
1224 /* Skip over this group if it has no CPUs allowed */
1225 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1228 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1230 /* Tally up the load of all CPUs in the group */
1233 for_each_cpu_mask(i
, group
->cpumask
) {
1234 /* Bias balancing toward cpus of our domain */
1236 load
= source_load(i
, load_idx
);
1238 load
= target_load(i
, load_idx
);
1243 /* Adjust by relative CPU power of the group */
1244 avg_load
= sg_div_cpu_power(group
,
1245 avg_load
* SCHED_LOAD_SCALE
);
1248 this_load
= avg_load
;
1250 } else if (avg_load
< min_load
) {
1251 min_load
= avg_load
;
1255 group
= group
->next
;
1256 } while (group
!= sd
->groups
);
1258 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1264 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1267 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1270 unsigned long load
, min_load
= ULONG_MAX
;
1274 /* Traverse only the allowed CPUs */
1275 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1277 for_each_cpu_mask(i
, tmp
) {
1278 load
= weighted_cpuload(i
);
1280 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1290 * sched_balance_self: balance the current task (running on cpu) in domains
1291 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1294 * Balance, ie. select the least loaded group.
1296 * Returns the target CPU number, or the same CPU if no balancing is needed.
1298 * preempt must be disabled.
1300 static int sched_balance_self(int cpu
, int flag
)
1302 struct task_struct
*t
= current
;
1303 struct sched_domain
*tmp
, *sd
= NULL
;
1305 for_each_domain(cpu
, tmp
) {
1307 * If power savings logic is enabled for a domain, stop there.
1309 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1311 if (tmp
->flags
& flag
)
1317 struct sched_group
*group
;
1318 int new_cpu
, weight
;
1320 if (!(sd
->flags
& flag
)) {
1326 group
= find_idlest_group(sd
, t
, cpu
);
1332 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1333 if (new_cpu
== -1 || new_cpu
== cpu
) {
1334 /* Now try balancing at a lower domain level of cpu */
1339 /* Now try balancing at a lower domain level of new_cpu */
1342 weight
= cpus_weight(span
);
1343 for_each_domain(cpu
, tmp
) {
1344 if (weight
<= cpus_weight(tmp
->span
))
1346 if (tmp
->flags
& flag
)
1349 /* while loop will break here if sd == NULL */
1355 #endif /* CONFIG_SMP */
1358 * wake_idle() will wake a task on an idle cpu if task->cpu is
1359 * not idle and an idle cpu is available. The span of cpus to
1360 * search starts with cpus closest then further out as needed,
1361 * so we always favor a closer, idle cpu.
1363 * Returns the CPU we should wake onto.
1365 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1366 static int wake_idle(int cpu
, struct task_struct
*p
)
1369 struct sched_domain
*sd
;
1373 * If it is idle, then it is the best cpu to run this task.
1375 * This cpu is also the best, if it has more than one task already.
1376 * Siblings must be also busy(in most cases) as they didn't already
1377 * pickup the extra load from this cpu and hence we need not check
1378 * sibling runqueue info. This will avoid the checks and cache miss
1379 * penalities associated with that.
1381 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1384 for_each_domain(cpu
, sd
) {
1385 if (sd
->flags
& SD_WAKE_IDLE
) {
1386 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1387 for_each_cpu_mask(i
, tmp
) {
1398 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1405 * try_to_wake_up - wake up a thread
1406 * @p: the to-be-woken-up thread
1407 * @state: the mask of task states that can be woken
1408 * @sync: do a synchronous wakeup?
1410 * Put it on the run-queue if it's not already there. The "current"
1411 * thread is always on the run-queue (except when the actual
1412 * re-schedule is in progress), and as such you're allowed to do
1413 * the simpler "current->state = TASK_RUNNING" to mark yourself
1414 * runnable without the overhead of this.
1416 * returns failure only if the task is already active.
1418 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1420 int cpu
, this_cpu
, success
= 0;
1421 unsigned long flags
;
1425 struct sched_domain
*sd
, *this_sd
= NULL
;
1426 unsigned long load
, this_load
;
1430 rq
= task_rq_lock(p
, &flags
);
1431 old_state
= p
->state
;
1432 if (!(old_state
& state
))
1439 this_cpu
= smp_processor_id();
1442 if (unlikely(task_running(rq
, p
)))
1447 schedstat_inc(rq
, ttwu_cnt
);
1448 if (cpu
== this_cpu
) {
1449 schedstat_inc(rq
, ttwu_local
);
1453 for_each_domain(this_cpu
, sd
) {
1454 if (cpu_isset(cpu
, sd
->span
)) {
1455 schedstat_inc(sd
, ttwu_wake_remote
);
1461 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1465 * Check for affine wakeup and passive balancing possibilities.
1468 int idx
= this_sd
->wake_idx
;
1469 unsigned int imbalance
;
1471 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1473 load
= source_load(cpu
, idx
);
1474 this_load
= target_load(this_cpu
, idx
);
1476 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1478 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1479 unsigned long tl
= this_load
;
1480 unsigned long tl_per_task
;
1482 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1485 * If sync wakeup then subtract the (maximum possible)
1486 * effect of the currently running task from the load
1487 * of the current CPU:
1490 tl
-= current
->se
.load
.weight
;
1493 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1494 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1496 * This domain has SD_WAKE_AFFINE and
1497 * p is cache cold in this domain, and
1498 * there is no bad imbalance.
1500 schedstat_inc(this_sd
, ttwu_move_affine
);
1506 * Start passive balancing when half the imbalance_pct
1509 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1510 if (imbalance
*this_load
<= 100*load
) {
1511 schedstat_inc(this_sd
, ttwu_move_balance
);
1517 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1519 new_cpu
= wake_idle(new_cpu
, p
);
1520 if (new_cpu
!= cpu
) {
1521 set_task_cpu(p
, new_cpu
);
1522 task_rq_unlock(rq
, &flags
);
1523 /* might preempt at this point */
1524 rq
= task_rq_lock(p
, &flags
);
1525 old_state
= p
->state
;
1526 if (!(old_state
& state
))
1531 this_cpu
= smp_processor_id();
1536 #endif /* CONFIG_SMP */
1537 activate_task(rq
, p
, 1);
1539 * Sync wakeups (i.e. those types of wakeups where the waker
1540 * has indicated that it will leave the CPU in short order)
1541 * don't trigger a preemption, if the woken up task will run on
1542 * this cpu. (in this case the 'I will reschedule' promise of
1543 * the waker guarantees that the freshly woken up task is going
1544 * to be considered on this CPU.)
1546 if (!sync
|| cpu
!= this_cpu
)
1547 check_preempt_curr(rq
, p
);
1551 p
->state
= TASK_RUNNING
;
1553 task_rq_unlock(rq
, &flags
);
1558 int fastcall
wake_up_process(struct task_struct
*p
)
1560 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1561 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1563 EXPORT_SYMBOL(wake_up_process
);
1565 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1567 return try_to_wake_up(p
, state
, 0);
1571 * Perform scheduler related setup for a newly forked process p.
1572 * p is forked by current.
1574 * __sched_fork() is basic setup used by init_idle() too:
1576 static void __sched_fork(struct task_struct
*p
)
1578 p
->se
.wait_start_fair
= 0;
1579 p
->se
.exec_start
= 0;
1580 p
->se
.sum_exec_runtime
= 0;
1581 p
->se
.delta_exec
= 0;
1582 p
->se
.delta_fair_run
= 0;
1583 p
->se
.delta_fair_sleep
= 0;
1584 p
->se
.wait_runtime
= 0;
1585 p
->se
.sleep_start_fair
= 0;
1587 #ifdef CONFIG_SCHEDSTATS
1588 p
->se
.wait_start
= 0;
1589 p
->se
.sum_wait_runtime
= 0;
1590 p
->se
.sum_sleep_runtime
= 0;
1591 p
->se
.sleep_start
= 0;
1592 p
->se
.block_start
= 0;
1593 p
->se
.sleep_max
= 0;
1594 p
->se
.block_max
= 0;
1597 p
->se
.wait_runtime_overruns
= 0;
1598 p
->se
.wait_runtime_underruns
= 0;
1601 INIT_LIST_HEAD(&p
->run_list
);
1604 #ifdef CONFIG_PREEMPT_NOTIFIERS
1605 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1609 * We mark the process as running here, but have not actually
1610 * inserted it onto the runqueue yet. This guarantees that
1611 * nobody will actually run it, and a signal or other external
1612 * event cannot wake it up and insert it on the runqueue either.
1614 p
->state
= TASK_RUNNING
;
1618 * fork()/clone()-time setup:
1620 void sched_fork(struct task_struct
*p
, int clone_flags
)
1622 int cpu
= get_cpu();
1627 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1629 __set_task_cpu(p
, cpu
);
1632 * Make sure we do not leak PI boosting priority to the child:
1634 p
->prio
= current
->normal_prio
;
1636 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1637 if (likely(sched_info_on()))
1638 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1640 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1643 #ifdef CONFIG_PREEMPT
1644 /* Want to start with kernel preemption disabled. */
1645 task_thread_info(p
)->preempt_count
= 1;
1651 * After fork, child runs first. (default) If set to 0 then
1652 * parent will (try to) run first.
1654 unsigned int __read_mostly sysctl_sched_child_runs_first
= 1;
1657 * wake_up_new_task - wake up a newly created task for the first time.
1659 * This function will do some initial scheduler statistics housekeeping
1660 * that must be done for every newly created context, then puts the task
1661 * on the runqueue and wakes it.
1663 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1665 unsigned long flags
;
1670 rq
= task_rq_lock(p
, &flags
);
1671 BUG_ON(p
->state
!= TASK_RUNNING
);
1672 this_cpu
= smp_processor_id(); /* parent's CPU */
1673 update_rq_clock(rq
);
1676 p
->prio
= effective_prio(p
);
1678 if (!p
->sched_class
->task_new
|| !sysctl_sched_child_runs_first
||
1679 (clone_flags
& CLONE_VM
) || task_cpu(p
) != this_cpu
||
1680 !current
->se
.on_rq
) {
1682 activate_task(rq
, p
, 0);
1685 * Let the scheduling class do new task startup
1686 * management (if any):
1688 p
->sched_class
->task_new(rq
, p
, now
);
1689 inc_nr_running(p
, rq
, now
);
1691 check_preempt_curr(rq
, p
);
1692 task_rq_unlock(rq
, &flags
);
1695 #ifdef CONFIG_PREEMPT_NOTIFIERS
1698 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1699 * @notifier: notifier struct to register
1701 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1703 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1705 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1708 * preempt_notifier_unregister - no longer interested in preemption notifications
1709 * @notifier: notifier struct to unregister
1711 * This is safe to call from within a preemption notifier.
1713 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1715 hlist_del(¬ifier
->link
);
1717 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1719 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1721 struct preempt_notifier
*notifier
;
1722 struct hlist_node
*node
;
1724 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1725 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1729 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1730 struct task_struct
*next
)
1732 struct preempt_notifier
*notifier
;
1733 struct hlist_node
*node
;
1735 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1736 notifier
->ops
->sched_out(notifier
, next
);
1741 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1746 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1747 struct task_struct
*next
)
1754 * prepare_task_switch - prepare to switch tasks
1755 * @rq: the runqueue preparing to switch
1756 * @prev: the current task that is being switched out
1757 * @next: the task we are going to switch to.
1759 * This is called with the rq lock held and interrupts off. It must
1760 * be paired with a subsequent finish_task_switch after the context
1763 * prepare_task_switch sets up locking and calls architecture specific
1767 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1768 struct task_struct
*next
)
1770 fire_sched_out_preempt_notifiers(prev
, next
);
1771 prepare_lock_switch(rq
, next
);
1772 prepare_arch_switch(next
);
1776 * finish_task_switch - clean up after a task-switch
1777 * @rq: runqueue associated with task-switch
1778 * @prev: the thread we just switched away from.
1780 * finish_task_switch must be called after the context switch, paired
1781 * with a prepare_task_switch call before the context switch.
1782 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1783 * and do any other architecture-specific cleanup actions.
1785 * Note that we may have delayed dropping an mm in context_switch(). If
1786 * so, we finish that here outside of the runqueue lock. (Doing it
1787 * with the lock held can cause deadlocks; see schedule() for
1790 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1791 __releases(rq
->lock
)
1793 struct mm_struct
*mm
= rq
->prev_mm
;
1799 * A task struct has one reference for the use as "current".
1800 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1801 * schedule one last time. The schedule call will never return, and
1802 * the scheduled task must drop that reference.
1803 * The test for TASK_DEAD must occur while the runqueue locks are
1804 * still held, otherwise prev could be scheduled on another cpu, die
1805 * there before we look at prev->state, and then the reference would
1807 * Manfred Spraul <manfred@colorfullife.com>
1809 prev_state
= prev
->state
;
1810 finish_arch_switch(prev
);
1811 finish_lock_switch(rq
, prev
);
1812 fire_sched_in_preempt_notifiers(current
);
1815 if (unlikely(prev_state
== TASK_DEAD
)) {
1817 * Remove function-return probe instances associated with this
1818 * task and put them back on the free list.
1820 kprobe_flush_task(prev
);
1821 put_task_struct(prev
);
1826 * schedule_tail - first thing a freshly forked thread must call.
1827 * @prev: the thread we just switched away from.
1829 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1830 __releases(rq
->lock
)
1832 struct rq
*rq
= this_rq();
1834 finish_task_switch(rq
, prev
);
1835 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1836 /* In this case, finish_task_switch does not reenable preemption */
1839 if (current
->set_child_tid
)
1840 put_user(current
->pid
, current
->set_child_tid
);
1844 * context_switch - switch to the new MM and the new
1845 * thread's register state.
1848 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1849 struct task_struct
*next
)
1851 struct mm_struct
*mm
, *oldmm
;
1853 prepare_task_switch(rq
, prev
, next
);
1855 oldmm
= prev
->active_mm
;
1857 * For paravirt, this is coupled with an exit in switch_to to
1858 * combine the page table reload and the switch backend into
1861 arch_enter_lazy_cpu_mode();
1863 if (unlikely(!mm
)) {
1864 next
->active_mm
= oldmm
;
1865 atomic_inc(&oldmm
->mm_count
);
1866 enter_lazy_tlb(oldmm
, next
);
1868 switch_mm(oldmm
, mm
, next
);
1870 if (unlikely(!prev
->mm
)) {
1871 prev
->active_mm
= NULL
;
1872 rq
->prev_mm
= oldmm
;
1875 * Since the runqueue lock will be released by the next
1876 * task (which is an invalid locking op but in the case
1877 * of the scheduler it's an obvious special-case), so we
1878 * do an early lockdep release here:
1880 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1881 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1884 /* Here we just switch the register state and the stack. */
1885 switch_to(prev
, next
, prev
);
1889 * this_rq must be evaluated again because prev may have moved
1890 * CPUs since it called schedule(), thus the 'rq' on its stack
1891 * frame will be invalid.
1893 finish_task_switch(this_rq(), prev
);
1897 * nr_running, nr_uninterruptible and nr_context_switches:
1899 * externally visible scheduler statistics: current number of runnable
1900 * threads, current number of uninterruptible-sleeping threads, total
1901 * number of context switches performed since bootup.
1903 unsigned long nr_running(void)
1905 unsigned long i
, sum
= 0;
1907 for_each_online_cpu(i
)
1908 sum
+= cpu_rq(i
)->nr_running
;
1913 unsigned long nr_uninterruptible(void)
1915 unsigned long i
, sum
= 0;
1917 for_each_possible_cpu(i
)
1918 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1921 * Since we read the counters lockless, it might be slightly
1922 * inaccurate. Do not allow it to go below zero though:
1924 if (unlikely((long)sum
< 0))
1930 unsigned long long nr_context_switches(void)
1933 unsigned long long sum
= 0;
1935 for_each_possible_cpu(i
)
1936 sum
+= cpu_rq(i
)->nr_switches
;
1941 unsigned long nr_iowait(void)
1943 unsigned long i
, sum
= 0;
1945 for_each_possible_cpu(i
)
1946 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1951 unsigned long nr_active(void)
1953 unsigned long i
, running
= 0, uninterruptible
= 0;
1955 for_each_online_cpu(i
) {
1956 running
+= cpu_rq(i
)->nr_running
;
1957 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1960 if (unlikely((long)uninterruptible
< 0))
1961 uninterruptible
= 0;
1963 return running
+ uninterruptible
;
1967 * Update rq->cpu_load[] statistics. This function is usually called every
1968 * scheduler tick (TICK_NSEC).
1970 static void update_cpu_load(struct rq
*this_rq
)
1972 u64 fair_delta64
, exec_delta64
, idle_delta64
, sample_interval64
, tmp64
;
1973 unsigned long total_load
= this_rq
->ls
.load
.weight
;
1974 unsigned long this_load
= total_load
;
1975 struct load_stat
*ls
= &this_rq
->ls
;
1976 u64 now
= __rq_clock(this_rq
);
1979 this_rq
->nr_load_updates
++;
1980 if (unlikely(!(sysctl_sched_features
& SCHED_FEAT_PRECISE_CPU_LOAD
)))
1983 /* Update delta_fair/delta_exec fields first */
1984 update_curr_load(this_rq
, now
);
1986 fair_delta64
= ls
->delta_fair
+ 1;
1989 exec_delta64
= ls
->delta_exec
+ 1;
1992 sample_interval64
= now
- ls
->load_update_last
;
1993 ls
->load_update_last
= now
;
1995 if ((s64
)sample_interval64
< (s64
)TICK_NSEC
)
1996 sample_interval64
= TICK_NSEC
;
1998 if (exec_delta64
> sample_interval64
)
1999 exec_delta64
= sample_interval64
;
2001 idle_delta64
= sample_interval64
- exec_delta64
;
2003 tmp64
= div64_64(SCHED_LOAD_SCALE
* exec_delta64
, fair_delta64
);
2004 tmp64
= div64_64(tmp64
* exec_delta64
, sample_interval64
);
2006 this_load
= (unsigned long)tmp64
;
2010 /* Update our load: */
2011 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2012 unsigned long old_load
, new_load
;
2014 /* scale is effectively 1 << i now, and >> i divides by scale */
2016 old_load
= this_rq
->cpu_load
[i
];
2017 new_load
= this_load
;
2019 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2026 * double_rq_lock - safely lock two runqueues
2028 * Note this does not disable interrupts like task_rq_lock,
2029 * you need to do so manually before calling.
2031 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2032 __acquires(rq1
->lock
)
2033 __acquires(rq2
->lock
)
2035 BUG_ON(!irqs_disabled());
2037 spin_lock(&rq1
->lock
);
2038 __acquire(rq2
->lock
); /* Fake it out ;) */
2041 spin_lock(&rq1
->lock
);
2042 spin_lock(&rq2
->lock
);
2044 spin_lock(&rq2
->lock
);
2045 spin_lock(&rq1
->lock
);
2051 * double_rq_unlock - safely unlock two runqueues
2053 * Note this does not restore interrupts like task_rq_unlock,
2054 * you need to do so manually after calling.
2056 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2057 __releases(rq1
->lock
)
2058 __releases(rq2
->lock
)
2060 spin_unlock(&rq1
->lock
);
2062 spin_unlock(&rq2
->lock
);
2064 __release(rq2
->lock
);
2068 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2070 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2071 __releases(this_rq
->lock
)
2072 __acquires(busiest
->lock
)
2073 __acquires(this_rq
->lock
)
2075 if (unlikely(!irqs_disabled())) {
2076 /* printk() doesn't work good under rq->lock */
2077 spin_unlock(&this_rq
->lock
);
2080 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2081 if (busiest
< this_rq
) {
2082 spin_unlock(&this_rq
->lock
);
2083 spin_lock(&busiest
->lock
);
2084 spin_lock(&this_rq
->lock
);
2086 spin_lock(&busiest
->lock
);
2091 * If dest_cpu is allowed for this process, migrate the task to it.
2092 * This is accomplished by forcing the cpu_allowed mask to only
2093 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2094 * the cpu_allowed mask is restored.
2096 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2098 struct migration_req req
;
2099 unsigned long flags
;
2102 rq
= task_rq_lock(p
, &flags
);
2103 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2104 || unlikely(cpu_is_offline(dest_cpu
)))
2107 /* force the process onto the specified CPU */
2108 if (migrate_task(p
, dest_cpu
, &req
)) {
2109 /* Need to wait for migration thread (might exit: take ref). */
2110 struct task_struct
*mt
= rq
->migration_thread
;
2112 get_task_struct(mt
);
2113 task_rq_unlock(rq
, &flags
);
2114 wake_up_process(mt
);
2115 put_task_struct(mt
);
2116 wait_for_completion(&req
.done
);
2121 task_rq_unlock(rq
, &flags
);
2125 * sched_exec - execve() is a valuable balancing opportunity, because at
2126 * this point the task has the smallest effective memory and cache footprint.
2128 void sched_exec(void)
2130 int new_cpu
, this_cpu
= get_cpu();
2131 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2133 if (new_cpu
!= this_cpu
)
2134 sched_migrate_task(current
, new_cpu
);
2138 * pull_task - move a task from a remote runqueue to the local runqueue.
2139 * Both runqueues must be locked.
2141 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2142 struct rq
*this_rq
, int this_cpu
)
2144 update_rq_clock(src_rq
);
2145 deactivate_task(src_rq
, p
, 0, src_rq
->clock
);
2146 set_task_cpu(p
, this_cpu
);
2147 activate_task(this_rq
, p
, 0);
2149 * Note that idle threads have a prio of MAX_PRIO, for this test
2150 * to be always true for them.
2152 check_preempt_curr(this_rq
, p
);
2156 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2159 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2160 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2164 * We do not migrate tasks that are:
2165 * 1) running (obviously), or
2166 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2167 * 3) are cache-hot on their current CPU.
2169 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2173 if (task_running(rq
, p
))
2177 * Aggressive migration if too many balance attempts have failed:
2179 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2185 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2186 unsigned long max_nr_move
, unsigned long max_load_move
,
2187 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2188 int *all_pinned
, unsigned long *load_moved
,
2189 int *this_best_prio
, struct rq_iterator
*iterator
)
2191 int pulled
= 0, pinned
= 0, skip_for_load
;
2192 struct task_struct
*p
;
2193 long rem_load_move
= max_load_move
;
2195 if (max_nr_move
== 0 || max_load_move
== 0)
2201 * Start the load-balancing iterator:
2203 p
= iterator
->start(iterator
->arg
);
2208 * To help distribute high priority tasks accross CPUs we don't
2209 * skip a task if it will be the highest priority task (i.e. smallest
2210 * prio value) on its new queue regardless of its load weight
2212 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2213 SCHED_LOAD_SCALE_FUZZ
;
2214 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2215 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2216 p
= iterator
->next(iterator
->arg
);
2220 pull_task(busiest
, p
, this_rq
, this_cpu
);
2222 rem_load_move
-= p
->se
.load
.weight
;
2225 * We only want to steal up to the prescribed number of tasks
2226 * and the prescribed amount of weighted load.
2228 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2229 if (p
->prio
< *this_best_prio
)
2230 *this_best_prio
= p
->prio
;
2231 p
= iterator
->next(iterator
->arg
);
2236 * Right now, this is the only place pull_task() is called,
2237 * so we can safely collect pull_task() stats here rather than
2238 * inside pull_task().
2240 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2243 *all_pinned
= pinned
;
2244 *load_moved
= max_load_move
- rem_load_move
;
2249 * move_tasks tries to move up to max_load_move weighted load from busiest to
2250 * this_rq, as part of a balancing operation within domain "sd".
2251 * Returns 1 if successful and 0 otherwise.
2253 * Called with both runqueues locked.
2255 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2256 unsigned long max_load_move
,
2257 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2260 struct sched_class
*class = sched_class_highest
;
2261 unsigned long total_load_moved
= 0;
2262 int this_best_prio
= this_rq
->curr
->prio
;
2266 class->load_balance(this_rq
, this_cpu
, busiest
,
2267 ULONG_MAX
, max_load_move
- total_load_moved
,
2268 sd
, idle
, all_pinned
, &this_best_prio
);
2269 class = class->next
;
2270 } while (class && max_load_move
> total_load_moved
);
2272 return total_load_moved
> 0;
2276 * move_one_task tries to move exactly one task from busiest to this_rq, as
2277 * part of active balancing operations within "domain".
2278 * Returns 1 if successful and 0 otherwise.
2280 * Called with both runqueues locked.
2282 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2283 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2285 struct sched_class
*class;
2286 int this_best_prio
= MAX_PRIO
;
2288 for (class = sched_class_highest
; class; class = class->next
)
2289 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2290 1, ULONG_MAX
, sd
, idle
, NULL
,
2298 * find_busiest_group finds and returns the busiest CPU group within the
2299 * domain. It calculates and returns the amount of weighted load which
2300 * should be moved to restore balance via the imbalance parameter.
2302 static struct sched_group
*
2303 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2304 unsigned long *imbalance
, enum cpu_idle_type idle
,
2305 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2307 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2308 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2309 unsigned long max_pull
;
2310 unsigned long busiest_load_per_task
, busiest_nr_running
;
2311 unsigned long this_load_per_task
, this_nr_running
;
2313 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2314 int power_savings_balance
= 1;
2315 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2316 unsigned long min_nr_running
= ULONG_MAX
;
2317 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2320 max_load
= this_load
= total_load
= total_pwr
= 0;
2321 busiest_load_per_task
= busiest_nr_running
= 0;
2322 this_load_per_task
= this_nr_running
= 0;
2323 if (idle
== CPU_NOT_IDLE
)
2324 load_idx
= sd
->busy_idx
;
2325 else if (idle
== CPU_NEWLY_IDLE
)
2326 load_idx
= sd
->newidle_idx
;
2328 load_idx
= sd
->idle_idx
;
2331 unsigned long load
, group_capacity
;
2334 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2335 unsigned long sum_nr_running
, sum_weighted_load
;
2337 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2340 balance_cpu
= first_cpu(group
->cpumask
);
2342 /* Tally up the load of all CPUs in the group */
2343 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2345 for_each_cpu_mask(i
, group
->cpumask
) {
2348 if (!cpu_isset(i
, *cpus
))
2353 if (*sd_idle
&& rq
->nr_running
)
2356 /* Bias balancing toward cpus of our domain */
2358 if (idle_cpu(i
) && !first_idle_cpu
) {
2363 load
= target_load(i
, load_idx
);
2365 load
= source_load(i
, load_idx
);
2368 sum_nr_running
+= rq
->nr_running
;
2369 sum_weighted_load
+= weighted_cpuload(i
);
2373 * First idle cpu or the first cpu(busiest) in this sched group
2374 * is eligible for doing load balancing at this and above
2375 * domains. In the newly idle case, we will allow all the cpu's
2376 * to do the newly idle load balance.
2378 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2379 balance_cpu
!= this_cpu
&& balance
) {
2384 total_load
+= avg_load
;
2385 total_pwr
+= group
->__cpu_power
;
2387 /* Adjust by relative CPU power of the group */
2388 avg_load
= sg_div_cpu_power(group
,
2389 avg_load
* SCHED_LOAD_SCALE
);
2391 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2394 this_load
= avg_load
;
2396 this_nr_running
= sum_nr_running
;
2397 this_load_per_task
= sum_weighted_load
;
2398 } else if (avg_load
> max_load
&&
2399 sum_nr_running
> group_capacity
) {
2400 max_load
= avg_load
;
2402 busiest_nr_running
= sum_nr_running
;
2403 busiest_load_per_task
= sum_weighted_load
;
2406 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2408 * Busy processors will not participate in power savings
2411 if (idle
== CPU_NOT_IDLE
||
2412 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2416 * If the local group is idle or completely loaded
2417 * no need to do power savings balance at this domain
2419 if (local_group
&& (this_nr_running
>= group_capacity
||
2421 power_savings_balance
= 0;
2424 * If a group is already running at full capacity or idle,
2425 * don't include that group in power savings calculations
2427 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2432 * Calculate the group which has the least non-idle load.
2433 * This is the group from where we need to pick up the load
2436 if ((sum_nr_running
< min_nr_running
) ||
2437 (sum_nr_running
== min_nr_running
&&
2438 first_cpu(group
->cpumask
) <
2439 first_cpu(group_min
->cpumask
))) {
2441 min_nr_running
= sum_nr_running
;
2442 min_load_per_task
= sum_weighted_load
/
2447 * Calculate the group which is almost near its
2448 * capacity but still has some space to pick up some load
2449 * from other group and save more power
2451 if (sum_nr_running
<= group_capacity
- 1) {
2452 if (sum_nr_running
> leader_nr_running
||
2453 (sum_nr_running
== leader_nr_running
&&
2454 first_cpu(group
->cpumask
) >
2455 first_cpu(group_leader
->cpumask
))) {
2456 group_leader
= group
;
2457 leader_nr_running
= sum_nr_running
;
2462 group
= group
->next
;
2463 } while (group
!= sd
->groups
);
2465 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2468 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2470 if (this_load
>= avg_load
||
2471 100*max_load
<= sd
->imbalance_pct
*this_load
)
2474 busiest_load_per_task
/= busiest_nr_running
;
2476 * We're trying to get all the cpus to the average_load, so we don't
2477 * want to push ourselves above the average load, nor do we wish to
2478 * reduce the max loaded cpu below the average load, as either of these
2479 * actions would just result in more rebalancing later, and ping-pong
2480 * tasks around. Thus we look for the minimum possible imbalance.
2481 * Negative imbalances (*we* are more loaded than anyone else) will
2482 * be counted as no imbalance for these purposes -- we can't fix that
2483 * by pulling tasks to us. Be careful of negative numbers as they'll
2484 * appear as very large values with unsigned longs.
2486 if (max_load
<= busiest_load_per_task
)
2490 * In the presence of smp nice balancing, certain scenarios can have
2491 * max load less than avg load(as we skip the groups at or below
2492 * its cpu_power, while calculating max_load..)
2494 if (max_load
< avg_load
) {
2496 goto small_imbalance
;
2499 /* Don't want to pull so many tasks that a group would go idle */
2500 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2502 /* How much load to actually move to equalise the imbalance */
2503 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2504 (avg_load
- this_load
) * this->__cpu_power
)
2508 * if *imbalance is less than the average load per runnable task
2509 * there is no gaurantee that any tasks will be moved so we'll have
2510 * a think about bumping its value to force at least one task to be
2513 if (*imbalance
+ SCHED_LOAD_SCALE_FUZZ
< busiest_load_per_task
/2) {
2514 unsigned long tmp
, pwr_now
, pwr_move
;
2518 pwr_move
= pwr_now
= 0;
2520 if (this_nr_running
) {
2521 this_load_per_task
/= this_nr_running
;
2522 if (busiest_load_per_task
> this_load_per_task
)
2525 this_load_per_task
= SCHED_LOAD_SCALE
;
2527 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2528 busiest_load_per_task
* imbn
) {
2529 *imbalance
= busiest_load_per_task
;
2534 * OK, we don't have enough imbalance to justify moving tasks,
2535 * however we may be able to increase total CPU power used by
2539 pwr_now
+= busiest
->__cpu_power
*
2540 min(busiest_load_per_task
, max_load
);
2541 pwr_now
+= this->__cpu_power
*
2542 min(this_load_per_task
, this_load
);
2543 pwr_now
/= SCHED_LOAD_SCALE
;
2545 /* Amount of load we'd subtract */
2546 tmp
= sg_div_cpu_power(busiest
,
2547 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2549 pwr_move
+= busiest
->__cpu_power
*
2550 min(busiest_load_per_task
, max_load
- tmp
);
2552 /* Amount of load we'd add */
2553 if (max_load
* busiest
->__cpu_power
<
2554 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2555 tmp
= sg_div_cpu_power(this,
2556 max_load
* busiest
->__cpu_power
);
2558 tmp
= sg_div_cpu_power(this,
2559 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2560 pwr_move
+= this->__cpu_power
*
2561 min(this_load_per_task
, this_load
+ tmp
);
2562 pwr_move
/= SCHED_LOAD_SCALE
;
2564 /* Move if we gain throughput */
2565 if (pwr_move
<= pwr_now
)
2568 *imbalance
= busiest_load_per_task
;
2574 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2575 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2578 if (this == group_leader
&& group_leader
!= group_min
) {
2579 *imbalance
= min_load_per_task
;
2589 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2592 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2593 unsigned long imbalance
, cpumask_t
*cpus
)
2595 struct rq
*busiest
= NULL
, *rq
;
2596 unsigned long max_load
= 0;
2599 for_each_cpu_mask(i
, group
->cpumask
) {
2602 if (!cpu_isset(i
, *cpus
))
2606 wl
= weighted_cpuload(i
);
2608 if (rq
->nr_running
== 1 && wl
> imbalance
)
2611 if (wl
> max_load
) {
2621 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2622 * so long as it is large enough.
2624 #define MAX_PINNED_INTERVAL 512
2627 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2628 * tasks if there is an imbalance.
2630 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2631 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2634 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2635 struct sched_group
*group
;
2636 unsigned long imbalance
;
2638 cpumask_t cpus
= CPU_MASK_ALL
;
2639 unsigned long flags
;
2642 * When power savings policy is enabled for the parent domain, idle
2643 * sibling can pick up load irrespective of busy siblings. In this case,
2644 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2645 * portraying it as CPU_NOT_IDLE.
2647 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2648 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2651 schedstat_inc(sd
, lb_cnt
[idle
]);
2654 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2661 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2665 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2667 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2671 BUG_ON(busiest
== this_rq
);
2673 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2676 if (busiest
->nr_running
> 1) {
2678 * Attempt to move tasks. If find_busiest_group has found
2679 * an imbalance but busiest->nr_running <= 1, the group is
2680 * still unbalanced. ld_moved simply stays zero, so it is
2681 * correctly treated as an imbalance.
2683 local_irq_save(flags
);
2684 double_rq_lock(this_rq
, busiest
);
2685 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2686 imbalance
, sd
, idle
, &all_pinned
);
2687 double_rq_unlock(this_rq
, busiest
);
2688 local_irq_restore(flags
);
2691 * some other cpu did the load balance for us.
2693 if (ld_moved
&& this_cpu
!= smp_processor_id())
2694 resched_cpu(this_cpu
);
2696 /* All tasks on this runqueue were pinned by CPU affinity */
2697 if (unlikely(all_pinned
)) {
2698 cpu_clear(cpu_of(busiest
), cpus
);
2699 if (!cpus_empty(cpus
))
2706 schedstat_inc(sd
, lb_failed
[idle
]);
2707 sd
->nr_balance_failed
++;
2709 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2711 spin_lock_irqsave(&busiest
->lock
, flags
);
2713 /* don't kick the migration_thread, if the curr
2714 * task on busiest cpu can't be moved to this_cpu
2716 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2717 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2719 goto out_one_pinned
;
2722 if (!busiest
->active_balance
) {
2723 busiest
->active_balance
= 1;
2724 busiest
->push_cpu
= this_cpu
;
2727 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2729 wake_up_process(busiest
->migration_thread
);
2732 * We've kicked active balancing, reset the failure
2735 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2738 sd
->nr_balance_failed
= 0;
2740 if (likely(!active_balance
)) {
2741 /* We were unbalanced, so reset the balancing interval */
2742 sd
->balance_interval
= sd
->min_interval
;
2745 * If we've begun active balancing, start to back off. This
2746 * case may not be covered by the all_pinned logic if there
2747 * is only 1 task on the busy runqueue (because we don't call
2750 if (sd
->balance_interval
< sd
->max_interval
)
2751 sd
->balance_interval
*= 2;
2754 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2755 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2760 schedstat_inc(sd
, lb_balanced
[idle
]);
2762 sd
->nr_balance_failed
= 0;
2765 /* tune up the balancing interval */
2766 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2767 (sd
->balance_interval
< sd
->max_interval
))
2768 sd
->balance_interval
*= 2;
2770 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2771 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2777 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2778 * tasks if there is an imbalance.
2780 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2781 * this_rq is locked.
2784 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2786 struct sched_group
*group
;
2787 struct rq
*busiest
= NULL
;
2788 unsigned long imbalance
;
2792 cpumask_t cpus
= CPU_MASK_ALL
;
2795 * When power savings policy is enabled for the parent domain, idle
2796 * sibling can pick up load irrespective of busy siblings. In this case,
2797 * let the state of idle sibling percolate up as IDLE, instead of
2798 * portraying it as CPU_NOT_IDLE.
2800 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2801 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2804 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2806 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2807 &sd_idle
, &cpus
, NULL
);
2809 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2813 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2816 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2820 BUG_ON(busiest
== this_rq
);
2822 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2825 if (busiest
->nr_running
> 1) {
2826 /* Attempt to move tasks */
2827 double_lock_balance(this_rq
, busiest
);
2828 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2829 imbalance
, sd
, CPU_NEWLY_IDLE
,
2831 spin_unlock(&busiest
->lock
);
2833 if (unlikely(all_pinned
)) {
2834 cpu_clear(cpu_of(busiest
), cpus
);
2835 if (!cpus_empty(cpus
))
2841 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2842 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2843 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2846 sd
->nr_balance_failed
= 0;
2851 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2852 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2853 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2855 sd
->nr_balance_failed
= 0;
2861 * idle_balance is called by schedule() if this_cpu is about to become
2862 * idle. Attempts to pull tasks from other CPUs.
2864 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2866 struct sched_domain
*sd
;
2867 int pulled_task
= -1;
2868 unsigned long next_balance
= jiffies
+ HZ
;
2870 for_each_domain(this_cpu
, sd
) {
2871 unsigned long interval
;
2873 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2876 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2877 /* If we've pulled tasks over stop searching: */
2878 pulled_task
= load_balance_newidle(this_cpu
,
2881 interval
= msecs_to_jiffies(sd
->balance_interval
);
2882 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2883 next_balance
= sd
->last_balance
+ interval
;
2887 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2889 * We are going idle. next_balance may be set based on
2890 * a busy processor. So reset next_balance.
2892 this_rq
->next_balance
= next_balance
;
2897 * active_load_balance is run by migration threads. It pushes running tasks
2898 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2899 * running on each physical CPU where possible, and avoids physical /
2900 * logical imbalances.
2902 * Called with busiest_rq locked.
2904 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2906 int target_cpu
= busiest_rq
->push_cpu
;
2907 struct sched_domain
*sd
;
2908 struct rq
*target_rq
;
2910 /* Is there any task to move? */
2911 if (busiest_rq
->nr_running
<= 1)
2914 target_rq
= cpu_rq(target_cpu
);
2917 * This condition is "impossible", if it occurs
2918 * we need to fix it. Originally reported by
2919 * Bjorn Helgaas on a 128-cpu setup.
2921 BUG_ON(busiest_rq
== target_rq
);
2923 /* move a task from busiest_rq to target_rq */
2924 double_lock_balance(busiest_rq
, target_rq
);
2926 /* Search for an sd spanning us and the target CPU. */
2927 for_each_domain(target_cpu
, sd
) {
2928 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2929 cpu_isset(busiest_cpu
, sd
->span
))
2934 schedstat_inc(sd
, alb_cnt
);
2936 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2938 schedstat_inc(sd
, alb_pushed
);
2940 schedstat_inc(sd
, alb_failed
);
2942 spin_unlock(&target_rq
->lock
);
2947 atomic_t load_balancer
;
2949 } nohz ____cacheline_aligned
= {
2950 .load_balancer
= ATOMIC_INIT(-1),
2951 .cpu_mask
= CPU_MASK_NONE
,
2955 * This routine will try to nominate the ilb (idle load balancing)
2956 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2957 * load balancing on behalf of all those cpus. If all the cpus in the system
2958 * go into this tickless mode, then there will be no ilb owner (as there is
2959 * no need for one) and all the cpus will sleep till the next wakeup event
2962 * For the ilb owner, tick is not stopped. And this tick will be used
2963 * for idle load balancing. ilb owner will still be part of
2966 * While stopping the tick, this cpu will become the ilb owner if there
2967 * is no other owner. And will be the owner till that cpu becomes busy
2968 * or if all cpus in the system stop their ticks at which point
2969 * there is no need for ilb owner.
2971 * When the ilb owner becomes busy, it nominates another owner, during the
2972 * next busy scheduler_tick()
2974 int select_nohz_load_balancer(int stop_tick
)
2976 int cpu
= smp_processor_id();
2979 cpu_set(cpu
, nohz
.cpu_mask
);
2980 cpu_rq(cpu
)->in_nohz_recently
= 1;
2983 * If we are going offline and still the leader, give up!
2985 if (cpu_is_offline(cpu
) &&
2986 atomic_read(&nohz
.load_balancer
) == cpu
) {
2987 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2992 /* time for ilb owner also to sleep */
2993 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2994 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2995 atomic_set(&nohz
.load_balancer
, -1);
2999 if (atomic_read(&nohz
.load_balancer
) == -1) {
3000 /* make me the ilb owner */
3001 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3003 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3006 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3009 cpu_clear(cpu
, nohz
.cpu_mask
);
3011 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3012 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3019 static DEFINE_SPINLOCK(balancing
);
3022 * It checks each scheduling domain to see if it is due to be balanced,
3023 * and initiates a balancing operation if so.
3025 * Balancing parameters are set up in arch_init_sched_domains.
3027 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3030 struct rq
*rq
= cpu_rq(cpu
);
3031 unsigned long interval
;
3032 struct sched_domain
*sd
;
3033 /* Earliest time when we have to do rebalance again */
3034 unsigned long next_balance
= jiffies
+ 60*HZ
;
3036 for_each_domain(cpu
, sd
) {
3037 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3040 interval
= sd
->balance_interval
;
3041 if (idle
!= CPU_IDLE
)
3042 interval
*= sd
->busy_factor
;
3044 /* scale ms to jiffies */
3045 interval
= msecs_to_jiffies(interval
);
3046 if (unlikely(!interval
))
3048 if (interval
> HZ
*NR_CPUS
/10)
3049 interval
= HZ
*NR_CPUS
/10;
3052 if (sd
->flags
& SD_SERIALIZE
) {
3053 if (!spin_trylock(&balancing
))
3057 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3058 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3060 * We've pulled tasks over so either we're no
3061 * longer idle, or one of our SMT siblings is
3064 idle
= CPU_NOT_IDLE
;
3066 sd
->last_balance
= jiffies
;
3068 if (sd
->flags
& SD_SERIALIZE
)
3069 spin_unlock(&balancing
);
3071 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3072 next_balance
= sd
->last_balance
+ interval
;
3075 * Stop the load balance at this level. There is another
3076 * CPU in our sched group which is doing load balancing more
3082 rq
->next_balance
= next_balance
;
3086 * run_rebalance_domains is triggered when needed from the scheduler tick.
3087 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3088 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3090 static void run_rebalance_domains(struct softirq_action
*h
)
3092 int this_cpu
= smp_processor_id();
3093 struct rq
*this_rq
= cpu_rq(this_cpu
);
3094 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3095 CPU_IDLE
: CPU_NOT_IDLE
;
3097 rebalance_domains(this_cpu
, idle
);
3101 * If this cpu is the owner for idle load balancing, then do the
3102 * balancing on behalf of the other idle cpus whose ticks are
3105 if (this_rq
->idle_at_tick
&&
3106 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3107 cpumask_t cpus
= nohz
.cpu_mask
;
3111 cpu_clear(this_cpu
, cpus
);
3112 for_each_cpu_mask(balance_cpu
, cpus
) {
3114 * If this cpu gets work to do, stop the load balancing
3115 * work being done for other cpus. Next load
3116 * balancing owner will pick it up.
3121 rebalance_domains(balance_cpu
, SCHED_IDLE
);
3123 rq
= cpu_rq(balance_cpu
);
3124 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3125 this_rq
->next_balance
= rq
->next_balance
;
3132 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3134 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3135 * idle load balancing owner or decide to stop the periodic load balancing,
3136 * if the whole system is idle.
3138 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3142 * If we were in the nohz mode recently and busy at the current
3143 * scheduler tick, then check if we need to nominate new idle
3146 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3147 rq
->in_nohz_recently
= 0;
3149 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3150 cpu_clear(cpu
, nohz
.cpu_mask
);
3151 atomic_set(&nohz
.load_balancer
, -1);
3154 if (atomic_read(&nohz
.load_balancer
) == -1) {
3156 * simple selection for now: Nominate the
3157 * first cpu in the nohz list to be the next
3160 * TBD: Traverse the sched domains and nominate
3161 * the nearest cpu in the nohz.cpu_mask.
3163 int ilb
= first_cpu(nohz
.cpu_mask
);
3171 * If this cpu is idle and doing idle load balancing for all the
3172 * cpus with ticks stopped, is it time for that to stop?
3174 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3175 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3181 * If this cpu is idle and the idle load balancing is done by
3182 * someone else, then no need raise the SCHED_SOFTIRQ
3184 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3185 cpu_isset(cpu
, nohz
.cpu_mask
))
3188 if (time_after_eq(jiffies
, rq
->next_balance
))
3189 raise_softirq(SCHED_SOFTIRQ
);
3192 #else /* CONFIG_SMP */
3195 * on UP we do not need to balance between CPUs:
3197 static inline void idle_balance(int cpu
, struct rq
*rq
)
3201 /* Avoid "used but not defined" warning on UP */
3202 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3203 unsigned long max_nr_move
, unsigned long max_load_move
,
3204 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3205 int *all_pinned
, unsigned long *load_moved
,
3206 int *this_best_prio
, struct rq_iterator
*iterator
)
3215 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3217 EXPORT_PER_CPU_SYMBOL(kstat
);
3220 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3221 * that have not yet been banked in case the task is currently running.
3223 unsigned long long task_sched_runtime(struct task_struct
*p
)
3225 unsigned long flags
;
3229 rq
= task_rq_lock(p
, &flags
);
3230 ns
= p
->se
.sum_exec_runtime
;
3231 if (rq
->curr
== p
) {
3232 update_rq_clock(rq
);
3233 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3234 if ((s64
)delta_exec
> 0)
3237 task_rq_unlock(rq
, &flags
);
3243 * Account user cpu time to a process.
3244 * @p: the process that the cpu time gets accounted to
3245 * @hardirq_offset: the offset to subtract from hardirq_count()
3246 * @cputime: the cpu time spent in user space since the last update
3248 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3250 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3253 p
->utime
= cputime_add(p
->utime
, cputime
);
3255 /* Add user time to cpustat. */
3256 tmp
= cputime_to_cputime64(cputime
);
3257 if (TASK_NICE(p
) > 0)
3258 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3260 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3264 * Account system cpu time to a process.
3265 * @p: the process that the cpu time gets accounted to
3266 * @hardirq_offset: the offset to subtract from hardirq_count()
3267 * @cputime: the cpu time spent in kernel space since the last update
3269 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3272 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3273 struct rq
*rq
= this_rq();
3276 p
->stime
= cputime_add(p
->stime
, cputime
);
3278 /* Add system time to cpustat. */
3279 tmp
= cputime_to_cputime64(cputime
);
3280 if (hardirq_count() - hardirq_offset
)
3281 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3282 else if (softirq_count())
3283 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3284 else if (p
!= rq
->idle
)
3285 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3286 else if (atomic_read(&rq
->nr_iowait
) > 0)
3287 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3289 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3290 /* Account for system time used */
3291 acct_update_integrals(p
);
3295 * Account for involuntary wait time.
3296 * @p: the process from which the cpu time has been stolen
3297 * @steal: the cpu time spent in involuntary wait
3299 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3301 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3302 cputime64_t tmp
= cputime_to_cputime64(steal
);
3303 struct rq
*rq
= this_rq();
3305 if (p
== rq
->idle
) {
3306 p
->stime
= cputime_add(p
->stime
, steal
);
3307 if (atomic_read(&rq
->nr_iowait
) > 0)
3308 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3310 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3312 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3316 * This function gets called by the timer code, with HZ frequency.
3317 * We call it with interrupts disabled.
3319 * It also gets called by the fork code, when changing the parent's
3322 void scheduler_tick(void)
3324 int cpu
= smp_processor_id();
3325 struct rq
*rq
= cpu_rq(cpu
);
3326 struct task_struct
*curr
= rq
->curr
;
3328 spin_lock(&rq
->lock
);
3329 update_cpu_load(rq
);
3330 if (curr
!= rq
->idle
) /* FIXME: needed? */
3331 curr
->sched_class
->task_tick(rq
, curr
);
3332 spin_unlock(&rq
->lock
);
3335 rq
->idle_at_tick
= idle_cpu(cpu
);
3336 trigger_load_balance(rq
, cpu
);
3340 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3342 void fastcall
add_preempt_count(int val
)
3347 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3349 preempt_count() += val
;
3351 * Spinlock count overflowing soon?
3353 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3356 EXPORT_SYMBOL(add_preempt_count
);
3358 void fastcall
sub_preempt_count(int val
)
3363 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3366 * Is the spinlock portion underflowing?
3368 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3369 !(preempt_count() & PREEMPT_MASK
)))
3372 preempt_count() -= val
;
3374 EXPORT_SYMBOL(sub_preempt_count
);
3379 * Print scheduling while atomic bug:
3381 static noinline
void __schedule_bug(struct task_struct
*prev
)
3383 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3384 prev
->comm
, preempt_count(), prev
->pid
);
3385 debug_show_held_locks(prev
);
3386 if (irqs_disabled())
3387 print_irqtrace_events(prev
);
3392 * Various schedule()-time debugging checks and statistics:
3394 static inline void schedule_debug(struct task_struct
*prev
)
3397 * Test if we are atomic. Since do_exit() needs to call into
3398 * schedule() atomically, we ignore that path for now.
3399 * Otherwise, whine if we are scheduling when we should not be.
3401 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3402 __schedule_bug(prev
);
3404 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3406 schedstat_inc(this_rq(), sched_cnt
);
3410 * Pick up the highest-prio task:
3412 static inline struct task_struct
*
3413 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, u64 now
)
3415 struct sched_class
*class;
3416 struct task_struct
*p
;
3419 * Optimization: we know that if all tasks are in
3420 * the fair class we can call that function directly:
3422 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3423 p
= fair_sched_class
.pick_next_task(rq
, now
);
3428 class = sched_class_highest
;
3430 p
= class->pick_next_task(rq
, now
);
3434 * Will never be NULL as the idle class always
3435 * returns a non-NULL p:
3437 class = class->next
;
3442 * schedule() is the main scheduler function.
3444 asmlinkage
void __sched
schedule(void)
3446 struct task_struct
*prev
, *next
;
3454 cpu
= smp_processor_id();
3458 switch_count
= &prev
->nivcsw
;
3460 release_kernel_lock(prev
);
3461 need_resched_nonpreemptible
:
3463 schedule_debug(prev
);
3465 spin_lock_irq(&rq
->lock
);
3466 clear_tsk_need_resched(prev
);
3467 now
= __rq_clock(rq
);
3469 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3470 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3471 unlikely(signal_pending(prev
)))) {
3472 prev
->state
= TASK_RUNNING
;
3474 deactivate_task(rq
, prev
, 1, now
);
3476 switch_count
= &prev
->nvcsw
;
3479 if (unlikely(!rq
->nr_running
))
3480 idle_balance(cpu
, rq
);
3482 prev
->sched_class
->put_prev_task(rq
, prev
, now
);
3483 next
= pick_next_task(rq
, prev
, now
);
3485 sched_info_switch(prev
, next
);
3487 if (likely(prev
!= next
)) {
3492 context_switch(rq
, prev
, next
); /* unlocks the rq */
3494 spin_unlock_irq(&rq
->lock
);
3496 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3497 cpu
= smp_processor_id();
3499 goto need_resched_nonpreemptible
;
3501 preempt_enable_no_resched();
3502 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3505 EXPORT_SYMBOL(schedule
);
3507 #ifdef CONFIG_PREEMPT
3509 * this is the entry point to schedule() from in-kernel preemption
3510 * off of preempt_enable. Kernel preemptions off return from interrupt
3511 * occur there and call schedule directly.
3513 asmlinkage
void __sched
preempt_schedule(void)
3515 struct thread_info
*ti
= current_thread_info();
3516 #ifdef CONFIG_PREEMPT_BKL
3517 struct task_struct
*task
= current
;
3518 int saved_lock_depth
;
3521 * If there is a non-zero preempt_count or interrupts are disabled,
3522 * we do not want to preempt the current task. Just return..
3524 if (likely(ti
->preempt_count
|| irqs_disabled()))
3528 add_preempt_count(PREEMPT_ACTIVE
);
3530 * We keep the big kernel semaphore locked, but we
3531 * clear ->lock_depth so that schedule() doesnt
3532 * auto-release the semaphore:
3534 #ifdef CONFIG_PREEMPT_BKL
3535 saved_lock_depth
= task
->lock_depth
;
3536 task
->lock_depth
= -1;
3539 #ifdef CONFIG_PREEMPT_BKL
3540 task
->lock_depth
= saved_lock_depth
;
3542 sub_preempt_count(PREEMPT_ACTIVE
);
3544 /* we could miss a preemption opportunity between schedule and now */
3546 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3549 EXPORT_SYMBOL(preempt_schedule
);
3552 * this is the entry point to schedule() from kernel preemption
3553 * off of irq context.
3554 * Note, that this is called and return with irqs disabled. This will
3555 * protect us against recursive calling from irq.
3557 asmlinkage
void __sched
preempt_schedule_irq(void)
3559 struct thread_info
*ti
= current_thread_info();
3560 #ifdef CONFIG_PREEMPT_BKL
3561 struct task_struct
*task
= current
;
3562 int saved_lock_depth
;
3564 /* Catch callers which need to be fixed */
3565 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3568 add_preempt_count(PREEMPT_ACTIVE
);
3570 * We keep the big kernel semaphore locked, but we
3571 * clear ->lock_depth so that schedule() doesnt
3572 * auto-release the semaphore:
3574 #ifdef CONFIG_PREEMPT_BKL
3575 saved_lock_depth
= task
->lock_depth
;
3576 task
->lock_depth
= -1;
3580 local_irq_disable();
3581 #ifdef CONFIG_PREEMPT_BKL
3582 task
->lock_depth
= saved_lock_depth
;
3584 sub_preempt_count(PREEMPT_ACTIVE
);
3586 /* we could miss a preemption opportunity between schedule and now */
3588 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3592 #endif /* CONFIG_PREEMPT */
3594 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3597 return try_to_wake_up(curr
->private, mode
, sync
);
3599 EXPORT_SYMBOL(default_wake_function
);
3602 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3603 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3604 * number) then we wake all the non-exclusive tasks and one exclusive task.
3606 * There are circumstances in which we can try to wake a task which has already
3607 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3608 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3610 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3611 int nr_exclusive
, int sync
, void *key
)
3613 struct list_head
*tmp
, *next
;
3615 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3616 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3617 unsigned flags
= curr
->flags
;
3619 if (curr
->func(curr
, mode
, sync
, key
) &&
3620 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3626 * __wake_up - wake up threads blocked on a waitqueue.
3628 * @mode: which threads
3629 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3630 * @key: is directly passed to the wakeup function
3632 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3633 int nr_exclusive
, void *key
)
3635 unsigned long flags
;
3637 spin_lock_irqsave(&q
->lock
, flags
);
3638 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3639 spin_unlock_irqrestore(&q
->lock
, flags
);
3641 EXPORT_SYMBOL(__wake_up
);
3644 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3646 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3648 __wake_up_common(q
, mode
, 1, 0, NULL
);
3652 * __wake_up_sync - wake up threads blocked on a waitqueue.
3654 * @mode: which threads
3655 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3657 * The sync wakeup differs that the waker knows that it will schedule
3658 * away soon, so while the target thread will be woken up, it will not
3659 * be migrated to another CPU - ie. the two threads are 'synchronized'
3660 * with each other. This can prevent needless bouncing between CPUs.
3662 * On UP it can prevent extra preemption.
3665 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3667 unsigned long flags
;
3673 if (unlikely(!nr_exclusive
))
3676 spin_lock_irqsave(&q
->lock
, flags
);
3677 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3678 spin_unlock_irqrestore(&q
->lock
, flags
);
3680 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3682 void fastcall
complete(struct completion
*x
)
3684 unsigned long flags
;
3686 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3688 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3690 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3692 EXPORT_SYMBOL(complete
);
3694 void fastcall
complete_all(struct completion
*x
)
3696 unsigned long flags
;
3698 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3699 x
->done
+= UINT_MAX
/2;
3700 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3702 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3704 EXPORT_SYMBOL(complete_all
);
3706 void fastcall __sched
wait_for_completion(struct completion
*x
)
3710 spin_lock_irq(&x
->wait
.lock
);
3712 DECLARE_WAITQUEUE(wait
, current
);
3714 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3715 __add_wait_queue_tail(&x
->wait
, &wait
);
3717 __set_current_state(TASK_UNINTERRUPTIBLE
);
3718 spin_unlock_irq(&x
->wait
.lock
);
3720 spin_lock_irq(&x
->wait
.lock
);
3722 __remove_wait_queue(&x
->wait
, &wait
);
3725 spin_unlock_irq(&x
->wait
.lock
);
3727 EXPORT_SYMBOL(wait_for_completion
);
3729 unsigned long fastcall __sched
3730 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3734 spin_lock_irq(&x
->wait
.lock
);
3736 DECLARE_WAITQUEUE(wait
, current
);
3738 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3739 __add_wait_queue_tail(&x
->wait
, &wait
);
3741 __set_current_state(TASK_UNINTERRUPTIBLE
);
3742 spin_unlock_irq(&x
->wait
.lock
);
3743 timeout
= schedule_timeout(timeout
);
3744 spin_lock_irq(&x
->wait
.lock
);
3746 __remove_wait_queue(&x
->wait
, &wait
);
3750 __remove_wait_queue(&x
->wait
, &wait
);
3754 spin_unlock_irq(&x
->wait
.lock
);
3757 EXPORT_SYMBOL(wait_for_completion_timeout
);
3759 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3765 spin_lock_irq(&x
->wait
.lock
);
3767 DECLARE_WAITQUEUE(wait
, current
);
3769 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3770 __add_wait_queue_tail(&x
->wait
, &wait
);
3772 if (signal_pending(current
)) {
3774 __remove_wait_queue(&x
->wait
, &wait
);
3777 __set_current_state(TASK_INTERRUPTIBLE
);
3778 spin_unlock_irq(&x
->wait
.lock
);
3780 spin_lock_irq(&x
->wait
.lock
);
3782 __remove_wait_queue(&x
->wait
, &wait
);
3786 spin_unlock_irq(&x
->wait
.lock
);
3790 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3792 unsigned long fastcall __sched
3793 wait_for_completion_interruptible_timeout(struct completion
*x
,
3794 unsigned long timeout
)
3798 spin_lock_irq(&x
->wait
.lock
);
3800 DECLARE_WAITQUEUE(wait
, current
);
3802 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3803 __add_wait_queue_tail(&x
->wait
, &wait
);
3805 if (signal_pending(current
)) {
3806 timeout
= -ERESTARTSYS
;
3807 __remove_wait_queue(&x
->wait
, &wait
);
3810 __set_current_state(TASK_INTERRUPTIBLE
);
3811 spin_unlock_irq(&x
->wait
.lock
);
3812 timeout
= schedule_timeout(timeout
);
3813 spin_lock_irq(&x
->wait
.lock
);
3815 __remove_wait_queue(&x
->wait
, &wait
);
3819 __remove_wait_queue(&x
->wait
, &wait
);
3823 spin_unlock_irq(&x
->wait
.lock
);
3826 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3829 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3831 spin_lock_irqsave(&q
->lock
, *flags
);
3832 __add_wait_queue(q
, wait
);
3833 spin_unlock(&q
->lock
);
3837 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3839 spin_lock_irq(&q
->lock
);
3840 __remove_wait_queue(q
, wait
);
3841 spin_unlock_irqrestore(&q
->lock
, *flags
);
3844 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3846 unsigned long flags
;
3849 init_waitqueue_entry(&wait
, current
);
3851 current
->state
= TASK_INTERRUPTIBLE
;
3853 sleep_on_head(q
, &wait
, &flags
);
3855 sleep_on_tail(q
, &wait
, &flags
);
3857 EXPORT_SYMBOL(interruptible_sleep_on
);
3860 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3862 unsigned long flags
;
3865 init_waitqueue_entry(&wait
, current
);
3867 current
->state
= TASK_INTERRUPTIBLE
;
3869 sleep_on_head(q
, &wait
, &flags
);
3870 timeout
= schedule_timeout(timeout
);
3871 sleep_on_tail(q
, &wait
, &flags
);
3875 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3877 void __sched
sleep_on(wait_queue_head_t
*q
)
3879 unsigned long flags
;
3882 init_waitqueue_entry(&wait
, current
);
3884 current
->state
= TASK_UNINTERRUPTIBLE
;
3886 sleep_on_head(q
, &wait
, &flags
);
3888 sleep_on_tail(q
, &wait
, &flags
);
3890 EXPORT_SYMBOL(sleep_on
);
3892 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3894 unsigned long flags
;
3897 init_waitqueue_entry(&wait
, current
);
3899 current
->state
= TASK_UNINTERRUPTIBLE
;
3901 sleep_on_head(q
, &wait
, &flags
);
3902 timeout
= schedule_timeout(timeout
);
3903 sleep_on_tail(q
, &wait
, &flags
);
3907 EXPORT_SYMBOL(sleep_on_timeout
);
3909 #ifdef CONFIG_RT_MUTEXES
3912 * rt_mutex_setprio - set the current priority of a task
3914 * @prio: prio value (kernel-internal form)
3916 * This function changes the 'effective' priority of a task. It does
3917 * not touch ->normal_prio like __setscheduler().
3919 * Used by the rt_mutex code to implement priority inheritance logic.
3921 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3923 unsigned long flags
;
3928 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3930 rq
= task_rq_lock(p
, &flags
);
3931 update_rq_clock(rq
);
3935 on_rq
= p
->se
.on_rq
;
3937 dequeue_task(rq
, p
, 0, now
);
3940 p
->sched_class
= &rt_sched_class
;
3942 p
->sched_class
= &fair_sched_class
;
3947 enqueue_task(rq
, p
, 0, now
);
3949 * Reschedule if we are currently running on this runqueue and
3950 * our priority decreased, or if we are not currently running on
3951 * this runqueue and our priority is higher than the current's
3953 if (task_running(rq
, p
)) {
3954 if (p
->prio
> oldprio
)
3955 resched_task(rq
->curr
);
3957 check_preempt_curr(rq
, p
);
3960 task_rq_unlock(rq
, &flags
);
3965 void set_user_nice(struct task_struct
*p
, long nice
)
3967 int old_prio
, delta
, on_rq
;
3968 unsigned long flags
;
3972 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3975 * We have to be careful, if called from sys_setpriority(),
3976 * the task might be in the middle of scheduling on another CPU.
3978 rq
= task_rq_lock(p
, &flags
);
3979 update_rq_clock(rq
);
3982 * The RT priorities are set via sched_setscheduler(), but we still
3983 * allow the 'normal' nice value to be set - but as expected
3984 * it wont have any effect on scheduling until the task is
3985 * SCHED_FIFO/SCHED_RR:
3987 if (task_has_rt_policy(p
)) {
3988 p
->static_prio
= NICE_TO_PRIO(nice
);
3991 on_rq
= p
->se
.on_rq
;
3993 dequeue_task(rq
, p
, 0, now
);
3994 dec_load(rq
, p
, now
);
3997 p
->static_prio
= NICE_TO_PRIO(nice
);
4000 p
->prio
= effective_prio(p
);
4001 delta
= p
->prio
- old_prio
;
4004 enqueue_task(rq
, p
, 0, now
);
4005 inc_load(rq
, p
, now
);
4007 * If the task increased its priority or is running and
4008 * lowered its priority, then reschedule its CPU:
4010 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4011 resched_task(rq
->curr
);
4014 task_rq_unlock(rq
, &flags
);
4016 EXPORT_SYMBOL(set_user_nice
);
4019 * can_nice - check if a task can reduce its nice value
4023 int can_nice(const struct task_struct
*p
, const int nice
)
4025 /* convert nice value [19,-20] to rlimit style value [1,40] */
4026 int nice_rlim
= 20 - nice
;
4028 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4029 capable(CAP_SYS_NICE
));
4032 #ifdef __ARCH_WANT_SYS_NICE
4035 * sys_nice - change the priority of the current process.
4036 * @increment: priority increment
4038 * sys_setpriority is a more generic, but much slower function that
4039 * does similar things.
4041 asmlinkage
long sys_nice(int increment
)
4046 * Setpriority might change our priority at the same moment.
4047 * We don't have to worry. Conceptually one call occurs first
4048 * and we have a single winner.
4050 if (increment
< -40)
4055 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4061 if (increment
< 0 && !can_nice(current
, nice
))
4064 retval
= security_task_setnice(current
, nice
);
4068 set_user_nice(current
, nice
);
4075 * task_prio - return the priority value of a given task.
4076 * @p: the task in question.
4078 * This is the priority value as seen by users in /proc.
4079 * RT tasks are offset by -200. Normal tasks are centered
4080 * around 0, value goes from -16 to +15.
4082 int task_prio(const struct task_struct
*p
)
4084 return p
->prio
- MAX_RT_PRIO
;
4088 * task_nice - return the nice value of a given task.
4089 * @p: the task in question.
4091 int task_nice(const struct task_struct
*p
)
4093 return TASK_NICE(p
);
4095 EXPORT_SYMBOL_GPL(task_nice
);
4098 * idle_cpu - is a given cpu idle currently?
4099 * @cpu: the processor in question.
4101 int idle_cpu(int cpu
)
4103 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4107 * idle_task - return the idle task for a given cpu.
4108 * @cpu: the processor in question.
4110 struct task_struct
*idle_task(int cpu
)
4112 return cpu_rq(cpu
)->idle
;
4116 * find_process_by_pid - find a process with a matching PID value.
4117 * @pid: the pid in question.
4119 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4121 return pid
? find_task_by_pid(pid
) : current
;
4124 /* Actually do priority change: must hold rq lock. */
4126 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4128 BUG_ON(p
->se
.on_rq
);
4131 switch (p
->policy
) {
4135 p
->sched_class
= &fair_sched_class
;
4139 p
->sched_class
= &rt_sched_class
;
4143 p
->rt_priority
= prio
;
4144 p
->normal_prio
= normal_prio(p
);
4145 /* we are holding p->pi_lock already */
4146 p
->prio
= rt_mutex_getprio(p
);
4151 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4152 * @p: the task in question.
4153 * @policy: new policy.
4154 * @param: structure containing the new RT priority.
4156 * NOTE that the task may be already dead.
4158 int sched_setscheduler(struct task_struct
*p
, int policy
,
4159 struct sched_param
*param
)
4161 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4162 unsigned long flags
;
4165 /* may grab non-irq protected spin_locks */
4166 BUG_ON(in_interrupt());
4168 /* double check policy once rq lock held */
4170 policy
= oldpolicy
= p
->policy
;
4171 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4172 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4173 policy
!= SCHED_IDLE
)
4176 * Valid priorities for SCHED_FIFO and SCHED_RR are
4177 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4178 * SCHED_BATCH and SCHED_IDLE is 0.
4180 if (param
->sched_priority
< 0 ||
4181 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4182 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4184 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4188 * Allow unprivileged RT tasks to decrease priority:
4190 if (!capable(CAP_SYS_NICE
)) {
4191 if (rt_policy(policy
)) {
4192 unsigned long rlim_rtprio
;
4194 if (!lock_task_sighand(p
, &flags
))
4196 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4197 unlock_task_sighand(p
, &flags
);
4199 /* can't set/change the rt policy */
4200 if (policy
!= p
->policy
&& !rlim_rtprio
)
4203 /* can't increase priority */
4204 if (param
->sched_priority
> p
->rt_priority
&&
4205 param
->sched_priority
> rlim_rtprio
)
4209 * Like positive nice levels, dont allow tasks to
4210 * move out of SCHED_IDLE either:
4212 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4215 /* can't change other user's priorities */
4216 if ((current
->euid
!= p
->euid
) &&
4217 (current
->euid
!= p
->uid
))
4221 retval
= security_task_setscheduler(p
, policy
, param
);
4225 * make sure no PI-waiters arrive (or leave) while we are
4226 * changing the priority of the task:
4228 spin_lock_irqsave(&p
->pi_lock
, flags
);
4230 * To be able to change p->policy safely, the apropriate
4231 * runqueue lock must be held.
4233 rq
= __task_rq_lock(p
);
4234 /* recheck policy now with rq lock held */
4235 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4236 policy
= oldpolicy
= -1;
4237 __task_rq_unlock(rq
);
4238 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4241 on_rq
= p
->se
.on_rq
;
4243 update_rq_clock(rq
);
4244 deactivate_task(rq
, p
, 0, rq
->clock
);
4247 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4249 activate_task(rq
, p
, 0);
4251 * Reschedule if we are currently running on this runqueue and
4252 * our priority decreased, or if we are not currently running on
4253 * this runqueue and our priority is higher than the current's
4255 if (task_running(rq
, p
)) {
4256 if (p
->prio
> oldprio
)
4257 resched_task(rq
->curr
);
4259 check_preempt_curr(rq
, p
);
4262 __task_rq_unlock(rq
);
4263 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4265 rt_mutex_adjust_pi(p
);
4269 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4272 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4274 struct sched_param lparam
;
4275 struct task_struct
*p
;
4278 if (!param
|| pid
< 0)
4280 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4285 p
= find_process_by_pid(pid
);
4287 retval
= sched_setscheduler(p
, policy
, &lparam
);
4294 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4295 * @pid: the pid in question.
4296 * @policy: new policy.
4297 * @param: structure containing the new RT priority.
4299 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4300 struct sched_param __user
*param
)
4302 /* negative values for policy are not valid */
4306 return do_sched_setscheduler(pid
, policy
, param
);
4310 * sys_sched_setparam - set/change the RT priority of a thread
4311 * @pid: the pid in question.
4312 * @param: structure containing the new RT priority.
4314 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4316 return do_sched_setscheduler(pid
, -1, param
);
4320 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4321 * @pid: the pid in question.
4323 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4325 struct task_struct
*p
;
4326 int retval
= -EINVAL
;
4332 read_lock(&tasklist_lock
);
4333 p
= find_process_by_pid(pid
);
4335 retval
= security_task_getscheduler(p
);
4339 read_unlock(&tasklist_lock
);
4346 * sys_sched_getscheduler - get the RT priority of a thread
4347 * @pid: the pid in question.
4348 * @param: structure containing the RT priority.
4350 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4352 struct sched_param lp
;
4353 struct task_struct
*p
;
4354 int retval
= -EINVAL
;
4356 if (!param
|| pid
< 0)
4359 read_lock(&tasklist_lock
);
4360 p
= find_process_by_pid(pid
);
4365 retval
= security_task_getscheduler(p
);
4369 lp
.sched_priority
= p
->rt_priority
;
4370 read_unlock(&tasklist_lock
);
4373 * This one might sleep, we cannot do it with a spinlock held ...
4375 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4381 read_unlock(&tasklist_lock
);
4385 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4387 cpumask_t cpus_allowed
;
4388 struct task_struct
*p
;
4391 mutex_lock(&sched_hotcpu_mutex
);
4392 read_lock(&tasklist_lock
);
4394 p
= find_process_by_pid(pid
);
4396 read_unlock(&tasklist_lock
);
4397 mutex_unlock(&sched_hotcpu_mutex
);
4402 * It is not safe to call set_cpus_allowed with the
4403 * tasklist_lock held. We will bump the task_struct's
4404 * usage count and then drop tasklist_lock.
4407 read_unlock(&tasklist_lock
);
4410 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4411 !capable(CAP_SYS_NICE
))
4414 retval
= security_task_setscheduler(p
, 0, NULL
);
4418 cpus_allowed
= cpuset_cpus_allowed(p
);
4419 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4420 retval
= set_cpus_allowed(p
, new_mask
);
4424 mutex_unlock(&sched_hotcpu_mutex
);
4428 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4429 cpumask_t
*new_mask
)
4431 if (len
< sizeof(cpumask_t
)) {
4432 memset(new_mask
, 0, sizeof(cpumask_t
));
4433 } else if (len
> sizeof(cpumask_t
)) {
4434 len
= sizeof(cpumask_t
);
4436 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4440 * sys_sched_setaffinity - set the cpu affinity of a process
4441 * @pid: pid of the process
4442 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4443 * @user_mask_ptr: user-space pointer to the new cpu mask
4445 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4446 unsigned long __user
*user_mask_ptr
)
4451 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4455 return sched_setaffinity(pid
, new_mask
);
4459 * Represents all cpu's present in the system
4460 * In systems capable of hotplug, this map could dynamically grow
4461 * as new cpu's are detected in the system via any platform specific
4462 * method, such as ACPI for e.g.
4465 cpumask_t cpu_present_map __read_mostly
;
4466 EXPORT_SYMBOL(cpu_present_map
);
4469 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4470 EXPORT_SYMBOL(cpu_online_map
);
4472 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4473 EXPORT_SYMBOL(cpu_possible_map
);
4476 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4478 struct task_struct
*p
;
4481 mutex_lock(&sched_hotcpu_mutex
);
4482 read_lock(&tasklist_lock
);
4485 p
= find_process_by_pid(pid
);
4489 retval
= security_task_getscheduler(p
);
4493 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4496 read_unlock(&tasklist_lock
);
4497 mutex_unlock(&sched_hotcpu_mutex
);
4503 * sys_sched_getaffinity - get the cpu affinity of a process
4504 * @pid: pid of the process
4505 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4506 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4508 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4509 unsigned long __user
*user_mask_ptr
)
4514 if (len
< sizeof(cpumask_t
))
4517 ret
= sched_getaffinity(pid
, &mask
);
4521 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4524 return sizeof(cpumask_t
);
4528 * sys_sched_yield - yield the current processor to other threads.
4530 * This function yields the current CPU to other tasks. If there are no
4531 * other threads running on this CPU then this function will return.
4533 asmlinkage
long sys_sched_yield(void)
4535 struct rq
*rq
= this_rq_lock();
4537 schedstat_inc(rq
, yld_cnt
);
4538 if (unlikely(rq
->nr_running
== 1))
4539 schedstat_inc(rq
, yld_act_empty
);
4541 current
->sched_class
->yield_task(rq
, current
);
4544 * Since we are going to call schedule() anyway, there's
4545 * no need to preempt or enable interrupts:
4547 __release(rq
->lock
);
4548 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4549 _raw_spin_unlock(&rq
->lock
);
4550 preempt_enable_no_resched();
4557 static void __cond_resched(void)
4559 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4560 __might_sleep(__FILE__
, __LINE__
);
4563 * The BKS might be reacquired before we have dropped
4564 * PREEMPT_ACTIVE, which could trigger a second
4565 * cond_resched() call.
4568 add_preempt_count(PREEMPT_ACTIVE
);
4570 sub_preempt_count(PREEMPT_ACTIVE
);
4571 } while (need_resched());
4574 int __sched
cond_resched(void)
4576 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4577 system_state
== SYSTEM_RUNNING
) {
4583 EXPORT_SYMBOL(cond_resched
);
4586 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4587 * call schedule, and on return reacquire the lock.
4589 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4590 * operations here to prevent schedule() from being called twice (once via
4591 * spin_unlock(), once by hand).
4593 int cond_resched_lock(spinlock_t
*lock
)
4597 if (need_lockbreak(lock
)) {
4603 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4604 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4605 _raw_spin_unlock(lock
);
4606 preempt_enable_no_resched();
4613 EXPORT_SYMBOL(cond_resched_lock
);
4615 int __sched
cond_resched_softirq(void)
4617 BUG_ON(!in_softirq());
4619 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4627 EXPORT_SYMBOL(cond_resched_softirq
);
4630 * yield - yield the current processor to other threads.
4632 * This is a shortcut for kernel-space yielding - it marks the
4633 * thread runnable and calls sys_sched_yield().
4635 void __sched
yield(void)
4637 set_current_state(TASK_RUNNING
);
4640 EXPORT_SYMBOL(yield
);
4643 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4644 * that process accounting knows that this is a task in IO wait state.
4646 * But don't do that if it is a deliberate, throttling IO wait (this task
4647 * has set its backing_dev_info: the queue against which it should throttle)
4649 void __sched
io_schedule(void)
4651 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4653 delayacct_blkio_start();
4654 atomic_inc(&rq
->nr_iowait
);
4656 atomic_dec(&rq
->nr_iowait
);
4657 delayacct_blkio_end();
4659 EXPORT_SYMBOL(io_schedule
);
4661 long __sched
io_schedule_timeout(long timeout
)
4663 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4666 delayacct_blkio_start();
4667 atomic_inc(&rq
->nr_iowait
);
4668 ret
= schedule_timeout(timeout
);
4669 atomic_dec(&rq
->nr_iowait
);
4670 delayacct_blkio_end();
4675 * sys_sched_get_priority_max - return maximum RT priority.
4676 * @policy: scheduling class.
4678 * this syscall returns the maximum rt_priority that can be used
4679 * by a given scheduling class.
4681 asmlinkage
long sys_sched_get_priority_max(int policy
)
4688 ret
= MAX_USER_RT_PRIO
-1;
4700 * sys_sched_get_priority_min - return minimum RT priority.
4701 * @policy: scheduling class.
4703 * this syscall returns the minimum rt_priority that can be used
4704 * by a given scheduling class.
4706 asmlinkage
long sys_sched_get_priority_min(int policy
)
4724 * sys_sched_rr_get_interval - return the default timeslice of a process.
4725 * @pid: pid of the process.
4726 * @interval: userspace pointer to the timeslice value.
4728 * this syscall writes the default timeslice value of a given process
4729 * into the user-space timespec buffer. A value of '0' means infinity.
4732 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4734 struct task_struct
*p
;
4735 int retval
= -EINVAL
;
4742 read_lock(&tasklist_lock
);
4743 p
= find_process_by_pid(pid
);
4747 retval
= security_task_getscheduler(p
);
4751 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4752 0 : static_prio_timeslice(p
->static_prio
), &t
);
4753 read_unlock(&tasklist_lock
);
4754 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4758 read_unlock(&tasklist_lock
);
4762 static const char stat_nam
[] = "RSDTtZX";
4764 static void show_task(struct task_struct
*p
)
4766 unsigned long free
= 0;
4769 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4770 printk("%-13.13s %c", p
->comm
,
4771 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4772 #if BITS_PER_LONG == 32
4773 if (state
== TASK_RUNNING
)
4774 printk(" running ");
4776 printk(" %08lx ", thread_saved_pc(p
));
4778 if (state
== TASK_RUNNING
)
4779 printk(" running task ");
4781 printk(" %016lx ", thread_saved_pc(p
));
4783 #ifdef CONFIG_DEBUG_STACK_USAGE
4785 unsigned long *n
= end_of_stack(p
);
4788 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4791 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4793 if (state
!= TASK_RUNNING
)
4794 show_stack(p
, NULL
);
4797 void show_state_filter(unsigned long state_filter
)
4799 struct task_struct
*g
, *p
;
4801 #if BITS_PER_LONG == 32
4803 " task PC stack pid father\n");
4806 " task PC stack pid father\n");
4808 read_lock(&tasklist_lock
);
4809 do_each_thread(g
, p
) {
4811 * reset the NMI-timeout, listing all files on a slow
4812 * console might take alot of time:
4814 touch_nmi_watchdog();
4815 if (!state_filter
|| (p
->state
& state_filter
))
4817 } while_each_thread(g
, p
);
4819 touch_all_softlockup_watchdogs();
4821 #ifdef CONFIG_SCHED_DEBUG
4822 sysrq_sched_debug_show();
4824 read_unlock(&tasklist_lock
);
4826 * Only show locks if all tasks are dumped:
4828 if (state_filter
== -1)
4829 debug_show_all_locks();
4832 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4834 idle
->sched_class
= &idle_sched_class
;
4838 * init_idle - set up an idle thread for a given CPU
4839 * @idle: task in question
4840 * @cpu: cpu the idle task belongs to
4842 * NOTE: this function does not set the idle thread's NEED_RESCHED
4843 * flag, to make booting more robust.
4845 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4847 struct rq
*rq
= cpu_rq(cpu
);
4848 unsigned long flags
;
4851 idle
->se
.exec_start
= sched_clock();
4853 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4854 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4855 __set_task_cpu(idle
, cpu
);
4857 spin_lock_irqsave(&rq
->lock
, flags
);
4858 rq
->curr
= rq
->idle
= idle
;
4859 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4862 spin_unlock_irqrestore(&rq
->lock
, flags
);
4864 /* Set the preempt count _outside_ the spinlocks! */
4865 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4866 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4868 task_thread_info(idle
)->preempt_count
= 0;
4871 * The idle tasks have their own, simple scheduling class:
4873 idle
->sched_class
= &idle_sched_class
;
4877 * In a system that switches off the HZ timer nohz_cpu_mask
4878 * indicates which cpus entered this state. This is used
4879 * in the rcu update to wait only for active cpus. For system
4880 * which do not switch off the HZ timer nohz_cpu_mask should
4881 * always be CPU_MASK_NONE.
4883 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4886 * Increase the granularity value when there are more CPUs,
4887 * because with more CPUs the 'effective latency' as visible
4888 * to users decreases. But the relationship is not linear,
4889 * so pick a second-best guess by going with the log2 of the
4892 * This idea comes from the SD scheduler of Con Kolivas:
4894 static inline void sched_init_granularity(void)
4896 unsigned int factor
= 1 + ilog2(num_online_cpus());
4897 const unsigned long gran_limit
= 100000000;
4899 sysctl_sched_granularity
*= factor
;
4900 if (sysctl_sched_granularity
> gran_limit
)
4901 sysctl_sched_granularity
= gran_limit
;
4903 sysctl_sched_runtime_limit
= sysctl_sched_granularity
* 4;
4904 sysctl_sched_wakeup_granularity
= sysctl_sched_granularity
/ 2;
4909 * This is how migration works:
4911 * 1) we queue a struct migration_req structure in the source CPU's
4912 * runqueue and wake up that CPU's migration thread.
4913 * 2) we down() the locked semaphore => thread blocks.
4914 * 3) migration thread wakes up (implicitly it forces the migrated
4915 * thread off the CPU)
4916 * 4) it gets the migration request and checks whether the migrated
4917 * task is still in the wrong runqueue.
4918 * 5) if it's in the wrong runqueue then the migration thread removes
4919 * it and puts it into the right queue.
4920 * 6) migration thread up()s the semaphore.
4921 * 7) we wake up and the migration is done.
4925 * Change a given task's CPU affinity. Migrate the thread to a
4926 * proper CPU and schedule it away if the CPU it's executing on
4927 * is removed from the allowed bitmask.
4929 * NOTE: the caller must have a valid reference to the task, the
4930 * task must not exit() & deallocate itself prematurely. The
4931 * call is not atomic; no spinlocks may be held.
4933 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4935 struct migration_req req
;
4936 unsigned long flags
;
4940 rq
= task_rq_lock(p
, &flags
);
4941 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4946 p
->cpus_allowed
= new_mask
;
4947 /* Can the task run on the task's current CPU? If so, we're done */
4948 if (cpu_isset(task_cpu(p
), new_mask
))
4951 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4952 /* Need help from migration thread: drop lock and wait. */
4953 task_rq_unlock(rq
, &flags
);
4954 wake_up_process(rq
->migration_thread
);
4955 wait_for_completion(&req
.done
);
4956 tlb_migrate_finish(p
->mm
);
4960 task_rq_unlock(rq
, &flags
);
4964 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4967 * Move (not current) task off this cpu, onto dest cpu. We're doing
4968 * this because either it can't run here any more (set_cpus_allowed()
4969 * away from this CPU, or CPU going down), or because we're
4970 * attempting to rebalance this task on exec (sched_exec).
4972 * So we race with normal scheduler movements, but that's OK, as long
4973 * as the task is no longer on this CPU.
4975 * Returns non-zero if task was successfully migrated.
4977 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4979 struct rq
*rq_dest
, *rq_src
;
4982 if (unlikely(cpu_is_offline(dest_cpu
)))
4985 rq_src
= cpu_rq(src_cpu
);
4986 rq_dest
= cpu_rq(dest_cpu
);
4988 double_rq_lock(rq_src
, rq_dest
);
4989 /* Already moved. */
4990 if (task_cpu(p
) != src_cpu
)
4992 /* Affinity changed (again). */
4993 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4996 on_rq
= p
->se
.on_rq
;
4998 update_rq_clock(rq_src
);
4999 deactivate_task(rq_src
, p
, 0, rq_src
->clock
);
5001 set_task_cpu(p
, dest_cpu
);
5003 activate_task(rq_dest
, p
, 0);
5004 check_preempt_curr(rq_dest
, p
);
5008 double_rq_unlock(rq_src
, rq_dest
);
5013 * migration_thread - this is a highprio system thread that performs
5014 * thread migration by bumping thread off CPU then 'pushing' onto
5017 static int migration_thread(void *data
)
5019 int cpu
= (long)data
;
5023 BUG_ON(rq
->migration_thread
!= current
);
5025 set_current_state(TASK_INTERRUPTIBLE
);
5026 while (!kthread_should_stop()) {
5027 struct migration_req
*req
;
5028 struct list_head
*head
;
5030 spin_lock_irq(&rq
->lock
);
5032 if (cpu_is_offline(cpu
)) {
5033 spin_unlock_irq(&rq
->lock
);
5037 if (rq
->active_balance
) {
5038 active_load_balance(rq
, cpu
);
5039 rq
->active_balance
= 0;
5042 head
= &rq
->migration_queue
;
5044 if (list_empty(head
)) {
5045 spin_unlock_irq(&rq
->lock
);
5047 set_current_state(TASK_INTERRUPTIBLE
);
5050 req
= list_entry(head
->next
, struct migration_req
, list
);
5051 list_del_init(head
->next
);
5053 spin_unlock(&rq
->lock
);
5054 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5057 complete(&req
->done
);
5059 __set_current_state(TASK_RUNNING
);
5063 /* Wait for kthread_stop */
5064 set_current_state(TASK_INTERRUPTIBLE
);
5065 while (!kthread_should_stop()) {
5067 set_current_state(TASK_INTERRUPTIBLE
);
5069 __set_current_state(TASK_RUNNING
);
5073 #ifdef CONFIG_HOTPLUG_CPU
5075 * Figure out where task on dead CPU should go, use force if neccessary.
5076 * NOTE: interrupts should be disabled by the caller
5078 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5080 unsigned long flags
;
5087 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5088 cpus_and(mask
, mask
, p
->cpus_allowed
);
5089 dest_cpu
= any_online_cpu(mask
);
5091 /* On any allowed CPU? */
5092 if (dest_cpu
== NR_CPUS
)
5093 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5095 /* No more Mr. Nice Guy. */
5096 if (dest_cpu
== NR_CPUS
) {
5097 rq
= task_rq_lock(p
, &flags
);
5098 cpus_setall(p
->cpus_allowed
);
5099 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5100 task_rq_unlock(rq
, &flags
);
5103 * Don't tell them about moving exiting tasks or
5104 * kernel threads (both mm NULL), since they never
5107 if (p
->mm
&& printk_ratelimit())
5108 printk(KERN_INFO
"process %d (%s) no "
5109 "longer affine to cpu%d\n",
5110 p
->pid
, p
->comm
, dead_cpu
);
5112 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5117 * While a dead CPU has no uninterruptible tasks queued at this point,
5118 * it might still have a nonzero ->nr_uninterruptible counter, because
5119 * for performance reasons the counter is not stricly tracking tasks to
5120 * their home CPUs. So we just add the counter to another CPU's counter,
5121 * to keep the global sum constant after CPU-down:
5123 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5125 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5126 unsigned long flags
;
5128 local_irq_save(flags
);
5129 double_rq_lock(rq_src
, rq_dest
);
5130 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5131 rq_src
->nr_uninterruptible
= 0;
5132 double_rq_unlock(rq_src
, rq_dest
);
5133 local_irq_restore(flags
);
5136 /* Run through task list and migrate tasks from the dead cpu. */
5137 static void migrate_live_tasks(int src_cpu
)
5139 struct task_struct
*p
, *t
;
5141 write_lock_irq(&tasklist_lock
);
5143 do_each_thread(t
, p
) {
5147 if (task_cpu(p
) == src_cpu
)
5148 move_task_off_dead_cpu(src_cpu
, p
);
5149 } while_each_thread(t
, p
);
5151 write_unlock_irq(&tasklist_lock
);
5155 * Schedules idle task to be the next runnable task on current CPU.
5156 * It does so by boosting its priority to highest possible and adding it to
5157 * the _front_ of the runqueue. Used by CPU offline code.
5159 void sched_idle_next(void)
5161 int this_cpu
= smp_processor_id();
5162 struct rq
*rq
= cpu_rq(this_cpu
);
5163 struct task_struct
*p
= rq
->idle
;
5164 unsigned long flags
;
5166 /* cpu has to be offline */
5167 BUG_ON(cpu_online(this_cpu
));
5170 * Strictly not necessary since rest of the CPUs are stopped by now
5171 * and interrupts disabled on the current cpu.
5173 spin_lock_irqsave(&rq
->lock
, flags
);
5175 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5177 /* Add idle task to the _front_ of its priority queue: */
5178 activate_idle_task(p
, rq
);
5180 spin_unlock_irqrestore(&rq
->lock
, flags
);
5184 * Ensures that the idle task is using init_mm right before its cpu goes
5187 void idle_task_exit(void)
5189 struct mm_struct
*mm
= current
->active_mm
;
5191 BUG_ON(cpu_online(smp_processor_id()));
5194 switch_mm(mm
, &init_mm
, current
);
5198 /* called under rq->lock with disabled interrupts */
5199 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5201 struct rq
*rq
= cpu_rq(dead_cpu
);
5203 /* Must be exiting, otherwise would be on tasklist. */
5204 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5206 /* Cannot have done final schedule yet: would have vanished. */
5207 BUG_ON(p
->state
== TASK_DEAD
);
5212 * Drop lock around migration; if someone else moves it,
5213 * that's OK. No task can be added to this CPU, so iteration is
5215 * NOTE: interrupts should be left disabled --dev@
5217 spin_unlock(&rq
->lock
);
5218 move_task_off_dead_cpu(dead_cpu
, p
);
5219 spin_lock(&rq
->lock
);
5224 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5225 static void migrate_dead_tasks(unsigned int dead_cpu
)
5227 struct rq
*rq
= cpu_rq(dead_cpu
);
5228 struct task_struct
*next
;
5231 if (!rq
->nr_running
)
5233 update_rq_clock(rq
);
5234 next
= pick_next_task(rq
, rq
->curr
, rq
->clock
);
5237 migrate_dead(dead_cpu
, next
);
5241 #endif /* CONFIG_HOTPLUG_CPU */
5243 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5245 static struct ctl_table sd_ctl_dir
[] = {
5247 .procname
= "sched_domain",
5253 static struct ctl_table sd_ctl_root
[] = {
5255 .procname
= "kernel",
5257 .child
= sd_ctl_dir
,
5262 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5264 struct ctl_table
*entry
=
5265 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5268 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5274 set_table_entry(struct ctl_table
*entry
,
5275 const char *procname
, void *data
, int maxlen
,
5276 mode_t mode
, proc_handler
*proc_handler
)
5278 entry
->procname
= procname
;
5280 entry
->maxlen
= maxlen
;
5282 entry
->proc_handler
= proc_handler
;
5285 static struct ctl_table
*
5286 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5288 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5290 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5291 sizeof(long), 0644, proc_doulongvec_minmax
);
5292 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5293 sizeof(long), 0644, proc_doulongvec_minmax
);
5294 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5295 sizeof(int), 0644, proc_dointvec_minmax
);
5296 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5297 sizeof(int), 0644, proc_dointvec_minmax
);
5298 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5299 sizeof(int), 0644, proc_dointvec_minmax
);
5300 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5301 sizeof(int), 0644, proc_dointvec_minmax
);
5302 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5303 sizeof(int), 0644, proc_dointvec_minmax
);
5304 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5305 sizeof(int), 0644, proc_dointvec_minmax
);
5306 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5307 sizeof(int), 0644, proc_dointvec_minmax
);
5308 set_table_entry(&table
[10], "cache_nice_tries",
5309 &sd
->cache_nice_tries
,
5310 sizeof(int), 0644, proc_dointvec_minmax
);
5311 set_table_entry(&table
[12], "flags", &sd
->flags
,
5312 sizeof(int), 0644, proc_dointvec_minmax
);
5317 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5319 struct ctl_table
*entry
, *table
;
5320 struct sched_domain
*sd
;
5321 int domain_num
= 0, i
;
5324 for_each_domain(cpu
, sd
)
5326 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5329 for_each_domain(cpu
, sd
) {
5330 snprintf(buf
, 32, "domain%d", i
);
5331 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5333 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5340 static struct ctl_table_header
*sd_sysctl_header
;
5341 static void init_sched_domain_sysctl(void)
5343 int i
, cpu_num
= num_online_cpus();
5344 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5347 sd_ctl_dir
[0].child
= entry
;
5349 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5350 snprintf(buf
, 32, "cpu%d", i
);
5351 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5353 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5355 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5358 static void init_sched_domain_sysctl(void)
5364 * migration_call - callback that gets triggered when a CPU is added.
5365 * Here we can start up the necessary migration thread for the new CPU.
5367 static int __cpuinit
5368 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5370 struct task_struct
*p
;
5371 int cpu
= (long)hcpu
;
5372 unsigned long flags
;
5376 case CPU_LOCK_ACQUIRE
:
5377 mutex_lock(&sched_hotcpu_mutex
);
5380 case CPU_UP_PREPARE
:
5381 case CPU_UP_PREPARE_FROZEN
:
5382 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5385 kthread_bind(p
, cpu
);
5386 /* Must be high prio: stop_machine expects to yield to it. */
5387 rq
= task_rq_lock(p
, &flags
);
5388 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5389 task_rq_unlock(rq
, &flags
);
5390 cpu_rq(cpu
)->migration_thread
= p
;
5394 case CPU_ONLINE_FROZEN
:
5395 /* Strictly unneccessary, as first user will wake it. */
5396 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5399 #ifdef CONFIG_HOTPLUG_CPU
5400 case CPU_UP_CANCELED
:
5401 case CPU_UP_CANCELED_FROZEN
:
5402 if (!cpu_rq(cpu
)->migration_thread
)
5404 /* Unbind it from offline cpu so it can run. Fall thru. */
5405 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5406 any_online_cpu(cpu_online_map
));
5407 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5408 cpu_rq(cpu
)->migration_thread
= NULL
;
5412 case CPU_DEAD_FROZEN
:
5413 migrate_live_tasks(cpu
);
5415 kthread_stop(rq
->migration_thread
);
5416 rq
->migration_thread
= NULL
;
5417 /* Idle task back to normal (off runqueue, low prio) */
5418 rq
= task_rq_lock(rq
->idle
, &flags
);
5419 update_rq_clock(rq
);
5420 deactivate_task(rq
, rq
->idle
, 0, rq
->clock
);
5421 rq
->idle
->static_prio
= MAX_PRIO
;
5422 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5423 rq
->idle
->sched_class
= &idle_sched_class
;
5424 migrate_dead_tasks(cpu
);
5425 task_rq_unlock(rq
, &flags
);
5426 migrate_nr_uninterruptible(rq
);
5427 BUG_ON(rq
->nr_running
!= 0);
5429 /* No need to migrate the tasks: it was best-effort if
5430 * they didn't take sched_hotcpu_mutex. Just wake up
5431 * the requestors. */
5432 spin_lock_irq(&rq
->lock
);
5433 while (!list_empty(&rq
->migration_queue
)) {
5434 struct migration_req
*req
;
5436 req
= list_entry(rq
->migration_queue
.next
,
5437 struct migration_req
, list
);
5438 list_del_init(&req
->list
);
5439 complete(&req
->done
);
5441 spin_unlock_irq(&rq
->lock
);
5444 case CPU_LOCK_RELEASE
:
5445 mutex_unlock(&sched_hotcpu_mutex
);
5451 /* Register at highest priority so that task migration (migrate_all_tasks)
5452 * happens before everything else.
5454 static struct notifier_block __cpuinitdata migration_notifier
= {
5455 .notifier_call
= migration_call
,
5459 int __init
migration_init(void)
5461 void *cpu
= (void *)(long)smp_processor_id();
5464 /* Start one for the boot CPU: */
5465 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5466 BUG_ON(err
== NOTIFY_BAD
);
5467 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5468 register_cpu_notifier(&migration_notifier
);
5476 /* Number of possible processor ids */
5477 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5478 EXPORT_SYMBOL(nr_cpu_ids
);
5480 #undef SCHED_DOMAIN_DEBUG
5481 #ifdef SCHED_DOMAIN_DEBUG
5482 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5487 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5491 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5496 struct sched_group
*group
= sd
->groups
;
5497 cpumask_t groupmask
;
5499 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5500 cpus_clear(groupmask
);
5503 for (i
= 0; i
< level
+ 1; i
++)
5505 printk("domain %d: ", level
);
5507 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5508 printk("does not load-balance\n");
5510 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5515 printk("span %s\n", str
);
5517 if (!cpu_isset(cpu
, sd
->span
))
5518 printk(KERN_ERR
"ERROR: domain->span does not contain "
5520 if (!cpu_isset(cpu
, group
->cpumask
))
5521 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5525 for (i
= 0; i
< level
+ 2; i
++)
5531 printk(KERN_ERR
"ERROR: group is NULL\n");
5535 if (!group
->__cpu_power
) {
5537 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5541 if (!cpus_weight(group
->cpumask
)) {
5543 printk(KERN_ERR
"ERROR: empty group\n");
5546 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5548 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5551 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5553 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5556 group
= group
->next
;
5557 } while (group
!= sd
->groups
);
5560 if (!cpus_equal(sd
->span
, groupmask
))
5561 printk(KERN_ERR
"ERROR: groups don't span "
5569 if (!cpus_subset(groupmask
, sd
->span
))
5570 printk(KERN_ERR
"ERROR: parent span is not a superset "
5571 "of domain->span\n");
5576 # define sched_domain_debug(sd, cpu) do { } while (0)
5579 static int sd_degenerate(struct sched_domain
*sd
)
5581 if (cpus_weight(sd
->span
) == 1)
5584 /* Following flags need at least 2 groups */
5585 if (sd
->flags
& (SD_LOAD_BALANCE
|
5586 SD_BALANCE_NEWIDLE
|
5590 SD_SHARE_PKG_RESOURCES
)) {
5591 if (sd
->groups
!= sd
->groups
->next
)
5595 /* Following flags don't use groups */
5596 if (sd
->flags
& (SD_WAKE_IDLE
|
5605 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5607 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5609 if (sd_degenerate(parent
))
5612 if (!cpus_equal(sd
->span
, parent
->span
))
5615 /* Does parent contain flags not in child? */
5616 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5617 if (cflags
& SD_WAKE_AFFINE
)
5618 pflags
&= ~SD_WAKE_BALANCE
;
5619 /* Flags needing groups don't count if only 1 group in parent */
5620 if (parent
->groups
== parent
->groups
->next
) {
5621 pflags
&= ~(SD_LOAD_BALANCE
|
5622 SD_BALANCE_NEWIDLE
|
5626 SD_SHARE_PKG_RESOURCES
);
5628 if (~cflags
& pflags
)
5635 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5636 * hold the hotplug lock.
5638 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5640 struct rq
*rq
= cpu_rq(cpu
);
5641 struct sched_domain
*tmp
;
5643 /* Remove the sched domains which do not contribute to scheduling. */
5644 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5645 struct sched_domain
*parent
= tmp
->parent
;
5648 if (sd_parent_degenerate(tmp
, parent
)) {
5649 tmp
->parent
= parent
->parent
;
5651 parent
->parent
->child
= tmp
;
5655 if (sd
&& sd_degenerate(sd
)) {
5661 sched_domain_debug(sd
, cpu
);
5663 rcu_assign_pointer(rq
->sd
, sd
);
5666 /* cpus with isolated domains */
5667 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5669 /* Setup the mask of cpus configured for isolated domains */
5670 static int __init
isolated_cpu_setup(char *str
)
5672 int ints
[NR_CPUS
], i
;
5674 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5675 cpus_clear(cpu_isolated_map
);
5676 for (i
= 1; i
<= ints
[0]; i
++)
5677 if (ints
[i
] < NR_CPUS
)
5678 cpu_set(ints
[i
], cpu_isolated_map
);
5682 __setup ("isolcpus=", isolated_cpu_setup
);
5685 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5686 * to a function which identifies what group(along with sched group) a CPU
5687 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5688 * (due to the fact that we keep track of groups covered with a cpumask_t).
5690 * init_sched_build_groups will build a circular linked list of the groups
5691 * covered by the given span, and will set each group's ->cpumask correctly,
5692 * and ->cpu_power to 0.
5695 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5696 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5697 struct sched_group
**sg
))
5699 struct sched_group
*first
= NULL
, *last
= NULL
;
5700 cpumask_t covered
= CPU_MASK_NONE
;
5703 for_each_cpu_mask(i
, span
) {
5704 struct sched_group
*sg
;
5705 int group
= group_fn(i
, cpu_map
, &sg
);
5708 if (cpu_isset(i
, covered
))
5711 sg
->cpumask
= CPU_MASK_NONE
;
5712 sg
->__cpu_power
= 0;
5714 for_each_cpu_mask(j
, span
) {
5715 if (group_fn(j
, cpu_map
, NULL
) != group
)
5718 cpu_set(j
, covered
);
5719 cpu_set(j
, sg
->cpumask
);
5730 #define SD_NODES_PER_DOMAIN 16
5735 * find_next_best_node - find the next node to include in a sched_domain
5736 * @node: node whose sched_domain we're building
5737 * @used_nodes: nodes already in the sched_domain
5739 * Find the next node to include in a given scheduling domain. Simply
5740 * finds the closest node not already in the @used_nodes map.
5742 * Should use nodemask_t.
5744 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5746 int i
, n
, val
, min_val
, best_node
= 0;
5750 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5751 /* Start at @node */
5752 n
= (node
+ i
) % MAX_NUMNODES
;
5754 if (!nr_cpus_node(n
))
5757 /* Skip already used nodes */
5758 if (test_bit(n
, used_nodes
))
5761 /* Simple min distance search */
5762 val
= node_distance(node
, n
);
5764 if (val
< min_val
) {
5770 set_bit(best_node
, used_nodes
);
5775 * sched_domain_node_span - get a cpumask for a node's sched_domain
5776 * @node: node whose cpumask we're constructing
5777 * @size: number of nodes to include in this span
5779 * Given a node, construct a good cpumask for its sched_domain to span. It
5780 * should be one that prevents unnecessary balancing, but also spreads tasks
5783 static cpumask_t
sched_domain_node_span(int node
)
5785 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5786 cpumask_t span
, nodemask
;
5790 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5792 nodemask
= node_to_cpumask(node
);
5793 cpus_or(span
, span
, nodemask
);
5794 set_bit(node
, used_nodes
);
5796 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5797 int next_node
= find_next_best_node(node
, used_nodes
);
5799 nodemask
= node_to_cpumask(next_node
);
5800 cpus_or(span
, span
, nodemask
);
5807 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5810 * SMT sched-domains:
5812 #ifdef CONFIG_SCHED_SMT
5813 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5814 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5816 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5817 struct sched_group
**sg
)
5820 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5826 * multi-core sched-domains:
5828 #ifdef CONFIG_SCHED_MC
5829 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5830 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5833 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5834 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5835 struct sched_group
**sg
)
5838 cpumask_t mask
= cpu_sibling_map
[cpu
];
5839 cpus_and(mask
, mask
, *cpu_map
);
5840 group
= first_cpu(mask
);
5842 *sg
= &per_cpu(sched_group_core
, group
);
5845 #elif defined(CONFIG_SCHED_MC)
5846 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5847 struct sched_group
**sg
)
5850 *sg
= &per_cpu(sched_group_core
, cpu
);
5855 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5856 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5858 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5859 struct sched_group
**sg
)
5862 #ifdef CONFIG_SCHED_MC
5863 cpumask_t mask
= cpu_coregroup_map(cpu
);
5864 cpus_and(mask
, mask
, *cpu_map
);
5865 group
= first_cpu(mask
);
5866 #elif defined(CONFIG_SCHED_SMT)
5867 cpumask_t mask
= cpu_sibling_map
[cpu
];
5868 cpus_and(mask
, mask
, *cpu_map
);
5869 group
= first_cpu(mask
);
5874 *sg
= &per_cpu(sched_group_phys
, group
);
5880 * The init_sched_build_groups can't handle what we want to do with node
5881 * groups, so roll our own. Now each node has its own list of groups which
5882 * gets dynamically allocated.
5884 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5885 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5887 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5888 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5890 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5891 struct sched_group
**sg
)
5893 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5896 cpus_and(nodemask
, nodemask
, *cpu_map
);
5897 group
= first_cpu(nodemask
);
5900 *sg
= &per_cpu(sched_group_allnodes
, group
);
5904 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5906 struct sched_group
*sg
= group_head
;
5912 for_each_cpu_mask(j
, sg
->cpumask
) {
5913 struct sched_domain
*sd
;
5915 sd
= &per_cpu(phys_domains
, j
);
5916 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5918 * Only add "power" once for each
5924 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5927 if (sg
!= group_head
)
5933 /* Free memory allocated for various sched_group structures */
5934 static void free_sched_groups(const cpumask_t
*cpu_map
)
5938 for_each_cpu_mask(cpu
, *cpu_map
) {
5939 struct sched_group
**sched_group_nodes
5940 = sched_group_nodes_bycpu
[cpu
];
5942 if (!sched_group_nodes
)
5945 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5946 cpumask_t nodemask
= node_to_cpumask(i
);
5947 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5949 cpus_and(nodemask
, nodemask
, *cpu_map
);
5950 if (cpus_empty(nodemask
))
5960 if (oldsg
!= sched_group_nodes
[i
])
5963 kfree(sched_group_nodes
);
5964 sched_group_nodes_bycpu
[cpu
] = NULL
;
5968 static void free_sched_groups(const cpumask_t
*cpu_map
)
5974 * Initialize sched groups cpu_power.
5976 * cpu_power indicates the capacity of sched group, which is used while
5977 * distributing the load between different sched groups in a sched domain.
5978 * Typically cpu_power for all the groups in a sched domain will be same unless
5979 * there are asymmetries in the topology. If there are asymmetries, group
5980 * having more cpu_power will pickup more load compared to the group having
5983 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5984 * the maximum number of tasks a group can handle in the presence of other idle
5985 * or lightly loaded groups in the same sched domain.
5987 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5989 struct sched_domain
*child
;
5990 struct sched_group
*group
;
5992 WARN_ON(!sd
|| !sd
->groups
);
5994 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5999 sd
->groups
->__cpu_power
= 0;
6002 * For perf policy, if the groups in child domain share resources
6003 * (for example cores sharing some portions of the cache hierarchy
6004 * or SMT), then set this domain groups cpu_power such that each group
6005 * can handle only one task, when there are other idle groups in the
6006 * same sched domain.
6008 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6010 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6011 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6016 * add cpu_power of each child group to this groups cpu_power
6018 group
= child
->groups
;
6020 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6021 group
= group
->next
;
6022 } while (group
!= child
->groups
);
6026 * Build sched domains for a given set of cpus and attach the sched domains
6027 * to the individual cpus
6029 static int build_sched_domains(const cpumask_t
*cpu_map
)
6033 struct sched_group
**sched_group_nodes
= NULL
;
6034 int sd_allnodes
= 0;
6037 * Allocate the per-node list of sched groups
6039 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6041 if (!sched_group_nodes
) {
6042 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6045 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6049 * Set up domains for cpus specified by the cpu_map.
6051 for_each_cpu_mask(i
, *cpu_map
) {
6052 struct sched_domain
*sd
= NULL
, *p
;
6053 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6055 cpus_and(nodemask
, nodemask
, *cpu_map
);
6058 if (cpus_weight(*cpu_map
) >
6059 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6060 sd
= &per_cpu(allnodes_domains
, i
);
6061 *sd
= SD_ALLNODES_INIT
;
6062 sd
->span
= *cpu_map
;
6063 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6069 sd
= &per_cpu(node_domains
, i
);
6071 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6075 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6079 sd
= &per_cpu(phys_domains
, i
);
6081 sd
->span
= nodemask
;
6085 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6087 #ifdef CONFIG_SCHED_MC
6089 sd
= &per_cpu(core_domains
, i
);
6091 sd
->span
= cpu_coregroup_map(i
);
6092 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6095 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6098 #ifdef CONFIG_SCHED_SMT
6100 sd
= &per_cpu(cpu_domains
, i
);
6101 *sd
= SD_SIBLING_INIT
;
6102 sd
->span
= cpu_sibling_map
[i
];
6103 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6106 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6110 #ifdef CONFIG_SCHED_SMT
6111 /* Set up CPU (sibling) groups */
6112 for_each_cpu_mask(i
, *cpu_map
) {
6113 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6114 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6115 if (i
!= first_cpu(this_sibling_map
))
6118 init_sched_build_groups(this_sibling_map
, cpu_map
,
6123 #ifdef CONFIG_SCHED_MC
6124 /* Set up multi-core groups */
6125 for_each_cpu_mask(i
, *cpu_map
) {
6126 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6127 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6128 if (i
!= first_cpu(this_core_map
))
6130 init_sched_build_groups(this_core_map
, cpu_map
,
6131 &cpu_to_core_group
);
6135 /* Set up physical groups */
6136 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6137 cpumask_t nodemask
= node_to_cpumask(i
);
6139 cpus_and(nodemask
, nodemask
, *cpu_map
);
6140 if (cpus_empty(nodemask
))
6143 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6147 /* Set up node groups */
6149 init_sched_build_groups(*cpu_map
, cpu_map
,
6150 &cpu_to_allnodes_group
);
6152 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6153 /* Set up node groups */
6154 struct sched_group
*sg
, *prev
;
6155 cpumask_t nodemask
= node_to_cpumask(i
);
6156 cpumask_t domainspan
;
6157 cpumask_t covered
= CPU_MASK_NONE
;
6160 cpus_and(nodemask
, nodemask
, *cpu_map
);
6161 if (cpus_empty(nodemask
)) {
6162 sched_group_nodes
[i
] = NULL
;
6166 domainspan
= sched_domain_node_span(i
);
6167 cpus_and(domainspan
, domainspan
, *cpu_map
);
6169 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6171 printk(KERN_WARNING
"Can not alloc domain group for "
6175 sched_group_nodes
[i
] = sg
;
6176 for_each_cpu_mask(j
, nodemask
) {
6177 struct sched_domain
*sd
;
6179 sd
= &per_cpu(node_domains
, j
);
6182 sg
->__cpu_power
= 0;
6183 sg
->cpumask
= nodemask
;
6185 cpus_or(covered
, covered
, nodemask
);
6188 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6189 cpumask_t tmp
, notcovered
;
6190 int n
= (i
+ j
) % MAX_NUMNODES
;
6192 cpus_complement(notcovered
, covered
);
6193 cpus_and(tmp
, notcovered
, *cpu_map
);
6194 cpus_and(tmp
, tmp
, domainspan
);
6195 if (cpus_empty(tmp
))
6198 nodemask
= node_to_cpumask(n
);
6199 cpus_and(tmp
, tmp
, nodemask
);
6200 if (cpus_empty(tmp
))
6203 sg
= kmalloc_node(sizeof(struct sched_group
),
6207 "Can not alloc domain group for node %d\n", j
);
6210 sg
->__cpu_power
= 0;
6212 sg
->next
= prev
->next
;
6213 cpus_or(covered
, covered
, tmp
);
6220 /* Calculate CPU power for physical packages and nodes */
6221 #ifdef CONFIG_SCHED_SMT
6222 for_each_cpu_mask(i
, *cpu_map
) {
6223 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6225 init_sched_groups_power(i
, sd
);
6228 #ifdef CONFIG_SCHED_MC
6229 for_each_cpu_mask(i
, *cpu_map
) {
6230 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6232 init_sched_groups_power(i
, sd
);
6236 for_each_cpu_mask(i
, *cpu_map
) {
6237 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6239 init_sched_groups_power(i
, sd
);
6243 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6244 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6247 struct sched_group
*sg
;
6249 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6250 init_numa_sched_groups_power(sg
);
6254 /* Attach the domains */
6255 for_each_cpu_mask(i
, *cpu_map
) {
6256 struct sched_domain
*sd
;
6257 #ifdef CONFIG_SCHED_SMT
6258 sd
= &per_cpu(cpu_domains
, i
);
6259 #elif defined(CONFIG_SCHED_MC)
6260 sd
= &per_cpu(core_domains
, i
);
6262 sd
= &per_cpu(phys_domains
, i
);
6264 cpu_attach_domain(sd
, i
);
6271 free_sched_groups(cpu_map
);
6276 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6278 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6280 cpumask_t cpu_default_map
;
6284 * Setup mask for cpus without special case scheduling requirements.
6285 * For now this just excludes isolated cpus, but could be used to
6286 * exclude other special cases in the future.
6288 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6290 err
= build_sched_domains(&cpu_default_map
);
6295 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6297 free_sched_groups(cpu_map
);
6301 * Detach sched domains from a group of cpus specified in cpu_map
6302 * These cpus will now be attached to the NULL domain
6304 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6308 for_each_cpu_mask(i
, *cpu_map
)
6309 cpu_attach_domain(NULL
, i
);
6310 synchronize_sched();
6311 arch_destroy_sched_domains(cpu_map
);
6315 * Partition sched domains as specified by the cpumasks below.
6316 * This attaches all cpus from the cpumasks to the NULL domain,
6317 * waits for a RCU quiescent period, recalculates sched
6318 * domain information and then attaches them back to the
6319 * correct sched domains
6320 * Call with hotplug lock held
6322 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6324 cpumask_t change_map
;
6327 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6328 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6329 cpus_or(change_map
, *partition1
, *partition2
);
6331 /* Detach sched domains from all of the affected cpus */
6332 detach_destroy_domains(&change_map
);
6333 if (!cpus_empty(*partition1
))
6334 err
= build_sched_domains(partition1
);
6335 if (!err
&& !cpus_empty(*partition2
))
6336 err
= build_sched_domains(partition2
);
6341 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6342 int arch_reinit_sched_domains(void)
6346 mutex_lock(&sched_hotcpu_mutex
);
6347 detach_destroy_domains(&cpu_online_map
);
6348 err
= arch_init_sched_domains(&cpu_online_map
);
6349 mutex_unlock(&sched_hotcpu_mutex
);
6354 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6358 if (buf
[0] != '0' && buf
[0] != '1')
6362 sched_smt_power_savings
= (buf
[0] == '1');
6364 sched_mc_power_savings
= (buf
[0] == '1');
6366 ret
= arch_reinit_sched_domains();
6368 return ret
? ret
: count
;
6371 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6375 #ifdef CONFIG_SCHED_SMT
6377 err
= sysfs_create_file(&cls
->kset
.kobj
,
6378 &attr_sched_smt_power_savings
.attr
);
6380 #ifdef CONFIG_SCHED_MC
6381 if (!err
&& mc_capable())
6382 err
= sysfs_create_file(&cls
->kset
.kobj
,
6383 &attr_sched_mc_power_savings
.attr
);
6389 #ifdef CONFIG_SCHED_MC
6390 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6392 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6394 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6395 const char *buf
, size_t count
)
6397 return sched_power_savings_store(buf
, count
, 0);
6399 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6400 sched_mc_power_savings_store
);
6403 #ifdef CONFIG_SCHED_SMT
6404 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6406 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6408 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6409 const char *buf
, size_t count
)
6411 return sched_power_savings_store(buf
, count
, 1);
6413 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6414 sched_smt_power_savings_store
);
6418 * Force a reinitialization of the sched domains hierarchy. The domains
6419 * and groups cannot be updated in place without racing with the balancing
6420 * code, so we temporarily attach all running cpus to the NULL domain
6421 * which will prevent rebalancing while the sched domains are recalculated.
6423 static int update_sched_domains(struct notifier_block
*nfb
,
6424 unsigned long action
, void *hcpu
)
6427 case CPU_UP_PREPARE
:
6428 case CPU_UP_PREPARE_FROZEN
:
6429 case CPU_DOWN_PREPARE
:
6430 case CPU_DOWN_PREPARE_FROZEN
:
6431 detach_destroy_domains(&cpu_online_map
);
6434 case CPU_UP_CANCELED
:
6435 case CPU_UP_CANCELED_FROZEN
:
6436 case CPU_DOWN_FAILED
:
6437 case CPU_DOWN_FAILED_FROZEN
:
6439 case CPU_ONLINE_FROZEN
:
6441 case CPU_DEAD_FROZEN
:
6443 * Fall through and re-initialise the domains.
6450 /* The hotplug lock is already held by cpu_up/cpu_down */
6451 arch_init_sched_domains(&cpu_online_map
);
6456 void __init
sched_init_smp(void)
6458 cpumask_t non_isolated_cpus
;
6460 mutex_lock(&sched_hotcpu_mutex
);
6461 arch_init_sched_domains(&cpu_online_map
);
6462 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6463 if (cpus_empty(non_isolated_cpus
))
6464 cpu_set(smp_processor_id(), non_isolated_cpus
);
6465 mutex_unlock(&sched_hotcpu_mutex
);
6466 /* XXX: Theoretical race here - CPU may be hotplugged now */
6467 hotcpu_notifier(update_sched_domains
, 0);
6469 init_sched_domain_sysctl();
6471 /* Move init over to a non-isolated CPU */
6472 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6474 sched_init_granularity();
6477 void __init
sched_init_smp(void)
6479 sched_init_granularity();
6481 #endif /* CONFIG_SMP */
6483 int in_sched_functions(unsigned long addr
)
6485 /* Linker adds these: start and end of __sched functions */
6486 extern char __sched_text_start
[], __sched_text_end
[];
6488 return in_lock_functions(addr
) ||
6489 (addr
>= (unsigned long)__sched_text_start
6490 && addr
< (unsigned long)__sched_text_end
);
6493 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6495 cfs_rq
->tasks_timeline
= RB_ROOT
;
6496 cfs_rq
->fair_clock
= 1;
6497 #ifdef CONFIG_FAIR_GROUP_SCHED
6502 void __init
sched_init(void)
6504 u64 now
= sched_clock();
6505 int highest_cpu
= 0;
6509 * Link up the scheduling class hierarchy:
6511 rt_sched_class
.next
= &fair_sched_class
;
6512 fair_sched_class
.next
= &idle_sched_class
;
6513 idle_sched_class
.next
= NULL
;
6515 for_each_possible_cpu(i
) {
6516 struct rt_prio_array
*array
;
6520 spin_lock_init(&rq
->lock
);
6521 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6524 init_cfs_rq(&rq
->cfs
, rq
);
6525 #ifdef CONFIG_FAIR_GROUP_SCHED
6526 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6527 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6529 rq
->ls
.load_update_last
= now
;
6530 rq
->ls
.load_update_start
= now
;
6532 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6533 rq
->cpu_load
[j
] = 0;
6536 rq
->active_balance
= 0;
6537 rq
->next_balance
= jiffies
;
6540 rq
->migration_thread
= NULL
;
6541 INIT_LIST_HEAD(&rq
->migration_queue
);
6543 atomic_set(&rq
->nr_iowait
, 0);
6545 array
= &rq
->rt
.active
;
6546 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6547 INIT_LIST_HEAD(array
->queue
+ j
);
6548 __clear_bit(j
, array
->bitmap
);
6551 /* delimiter for bitsearch: */
6552 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6555 set_load_weight(&init_task
);
6557 #ifdef CONFIG_PREEMPT_NOTIFIERS
6558 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6562 nr_cpu_ids
= highest_cpu
+ 1;
6563 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6566 #ifdef CONFIG_RT_MUTEXES
6567 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6571 * The boot idle thread does lazy MMU switching as well:
6573 atomic_inc(&init_mm
.mm_count
);
6574 enter_lazy_tlb(&init_mm
, current
);
6577 * Make us the idle thread. Technically, schedule() should not be
6578 * called from this thread, however somewhere below it might be,
6579 * but because we are the idle thread, we just pick up running again
6580 * when this runqueue becomes "idle".
6582 init_idle(current
, smp_processor_id());
6584 * During early bootup we pretend to be a normal task:
6586 current
->sched_class
= &fair_sched_class
;
6589 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6590 void __might_sleep(char *file
, int line
)
6593 static unsigned long prev_jiffy
; /* ratelimiting */
6595 if ((in_atomic() || irqs_disabled()) &&
6596 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6597 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6599 prev_jiffy
= jiffies
;
6600 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6601 " context at %s:%d\n", file
, line
);
6602 printk("in_atomic():%d, irqs_disabled():%d\n",
6603 in_atomic(), irqs_disabled());
6604 debug_show_held_locks(current
);
6605 if (irqs_disabled())
6606 print_irqtrace_events(current
);
6611 EXPORT_SYMBOL(__might_sleep
);
6614 #ifdef CONFIG_MAGIC_SYSRQ
6615 void normalize_rt_tasks(void)
6617 struct task_struct
*g
, *p
;
6618 unsigned long flags
;
6622 read_lock_irq(&tasklist_lock
);
6623 do_each_thread(g
, p
) {
6625 p
->se
.wait_runtime
= 0;
6626 p
->se
.exec_start
= 0;
6627 p
->se
.wait_start_fair
= 0;
6628 p
->se
.sleep_start_fair
= 0;
6629 #ifdef CONFIG_SCHEDSTATS
6630 p
->se
.wait_start
= 0;
6631 p
->se
.sleep_start
= 0;
6632 p
->se
.block_start
= 0;
6634 task_rq(p
)->cfs
.fair_clock
= 0;
6635 task_rq(p
)->clock
= 0;
6639 * Renice negative nice level userspace
6642 if (TASK_NICE(p
) < 0 && p
->mm
)
6643 set_user_nice(p
, 0);
6647 spin_lock_irqsave(&p
->pi_lock
, flags
);
6648 rq
= __task_rq_lock(p
);
6651 * Do not touch the migration thread:
6653 if (p
== rq
->migration_thread
)
6657 on_rq
= p
->se
.on_rq
;
6659 update_rq_clock(task_rq(p
));
6660 deactivate_task(task_rq(p
), p
, 0, task_rq(p
)->clock
);
6662 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6664 activate_task(task_rq(p
), p
, 0);
6665 resched_task(rq
->curr
);
6670 __task_rq_unlock(rq
);
6671 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6672 } while_each_thread(g
, p
);
6674 read_unlock_irq(&tasklist_lock
);
6677 #endif /* CONFIG_MAGIC_SYSRQ */
6681 * These functions are only useful for the IA64 MCA handling.
6683 * They can only be called when the whole system has been
6684 * stopped - every CPU needs to be quiescent, and no scheduling
6685 * activity can take place. Using them for anything else would
6686 * be a serious bug, and as a result, they aren't even visible
6687 * under any other configuration.
6691 * curr_task - return the current task for a given cpu.
6692 * @cpu: the processor in question.
6694 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6696 struct task_struct
*curr_task(int cpu
)
6698 return cpu_curr(cpu
);
6702 * set_curr_task - set the current task for a given cpu.
6703 * @cpu: the processor in question.
6704 * @p: the task pointer to set.
6706 * Description: This function must only be used when non-maskable interrupts
6707 * are serviced on a separate stack. It allows the architecture to switch the
6708 * notion of the current task on a cpu in a non-blocking manner. This function
6709 * must be called with all CPU's synchronized, and interrupts disabled, the
6710 * and caller must save the original value of the current task (see
6711 * curr_task() above) and restore that value before reenabling interrupts and
6712 * re-starting the system.
6714 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6716 void set_curr_task(int cpu
, struct task_struct
*p
)