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>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak
)) sched_clock(void)
75 return (unsigned long long)jiffies
* (1000000000 / HZ
);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
122 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
131 sg
->__cpu_power
+= val
;
132 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
136 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio
)
145 if (static_prio
== NICE_TO_PRIO(19))
148 if (static_prio
< NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
151 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
154 static inline int rt_policy(int policy
)
156 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
161 static inline int task_has_rt_policy(struct task_struct
*p
)
163 return rt_policy(p
->policy
);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array
{
170 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
171 struct list_head queue
[MAX_RT_PRIO
];
175 struct load_weight load
;
176 u64 load_update_start
, load_update_last
;
177 unsigned long delta_fair
, delta_exec
, delta_stat
;
180 /* CFS-related fields in a runqueue */
182 struct load_weight load
;
183 unsigned long nr_running
;
189 unsigned long wait_runtime_overruns
, wait_runtime_underruns
;
191 struct rb_root tasks_timeline
;
192 struct rb_node
*rb_leftmost
;
193 struct rb_node
*rb_load_balance_curr
;
194 #ifdef CONFIG_FAIR_GROUP_SCHED
195 /* 'curr' points to currently running entity on this cfs_rq.
196 * It is set to NULL otherwise (i.e when none are currently running).
198 struct sched_entity
*curr
;
199 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
201 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
202 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
203 * (like users, containers etc.)
205 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
206 * list is used during load balance.
208 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
212 /* Real-Time classes' related field in a runqueue: */
214 struct rt_prio_array active
;
215 int rt_load_balance_idx
;
216 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
220 * This is the main, per-CPU runqueue data structure.
222 * Locking rule: those places that want to lock multiple runqueues
223 * (such as the load balancing or the thread migration code), lock
224 * acquire operations must be ordered by ascending &runqueue.
227 spinlock_t lock
; /* runqueue lock */
230 * nr_running and cpu_load should be in the same cacheline because
231 * remote CPUs use both these fields when doing load calculation.
233 unsigned long nr_running
;
234 #define CPU_LOAD_IDX_MAX 5
235 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
236 unsigned char idle_at_tick
;
238 unsigned char in_nohz_recently
;
240 struct load_stat ls
; /* capture load from *all* tasks on this cpu */
241 unsigned long nr_load_updates
;
245 #ifdef CONFIG_FAIR_GROUP_SCHED
246 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
251 * This is part of a global counter where only the total sum
252 * over all CPUs matters. A task can increase this counter on
253 * one CPU and if it got migrated afterwards it may decrease
254 * it on another CPU. Always updated under the runqueue lock:
256 unsigned long nr_uninterruptible
;
258 struct task_struct
*curr
, *idle
;
259 unsigned long next_balance
;
260 struct mm_struct
*prev_mm
;
262 u64 clock
, prev_clock_raw
;
265 unsigned int clock_warps
, clock_overflows
;
267 unsigned int clock_deep_idle_events
;
273 struct sched_domain
*sd
;
275 /* For active balancing */
278 int cpu
; /* cpu of this runqueue */
280 struct task_struct
*migration_thread
;
281 struct list_head migration_queue
;
284 #ifdef CONFIG_SCHEDSTATS
286 struct sched_info rq_sched_info
;
288 /* sys_sched_yield() stats */
289 unsigned long yld_exp_empty
;
290 unsigned long yld_act_empty
;
291 unsigned long yld_both_empty
;
292 unsigned long yld_cnt
;
294 /* schedule() stats */
295 unsigned long sched_switch
;
296 unsigned long sched_cnt
;
297 unsigned long sched_goidle
;
299 /* try_to_wake_up() stats */
300 unsigned long ttwu_cnt
;
301 unsigned long ttwu_local
;
303 struct lock_class_key rq_lock_key
;
306 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
307 static DEFINE_MUTEX(sched_hotcpu_mutex
);
309 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
311 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
314 static inline int cpu_of(struct rq
*rq
)
324 * Update the per-runqueue clock, as finegrained as the platform can give
325 * us, but without assuming monotonicity, etc.:
327 static void __update_rq_clock(struct rq
*rq
)
329 u64 prev_raw
= rq
->prev_clock_raw
;
330 u64 now
= sched_clock();
331 s64 delta
= now
- prev_raw
;
332 u64 clock
= rq
->clock
;
334 #ifdef CONFIG_SCHED_DEBUG
335 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
338 * Protect against sched_clock() occasionally going backwards:
340 if (unlikely(delta
< 0)) {
345 * Catch too large forward jumps too:
347 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
348 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
349 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
352 rq
->clock_overflows
++;
354 if (unlikely(delta
> rq
->clock_max_delta
))
355 rq
->clock_max_delta
= delta
;
360 rq
->prev_clock_raw
= now
;
364 static void update_rq_clock(struct rq
*rq
)
366 if (likely(smp_processor_id() == cpu_of(rq
)))
367 __update_rq_clock(rq
);
371 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
372 * See detach_destroy_domains: synchronize_sched for details.
374 * The domain tree of any CPU may only be accessed from within
375 * preempt-disabled sections.
377 #define for_each_domain(cpu, __sd) \
378 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
380 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
381 #define this_rq() (&__get_cpu_var(runqueues))
382 #define task_rq(p) cpu_rq(task_cpu(p))
383 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
386 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
387 * clock constructed from sched_clock():
389 unsigned long long cpu_clock(int cpu
)
391 unsigned long long now
;
395 local_irq_save(flags
);
399 local_irq_restore(flags
);
404 #ifdef CONFIG_FAIR_GROUP_SCHED
405 /* Change a task's ->cfs_rq if it moves across CPUs */
406 static inline void set_task_cfs_rq(struct task_struct
*p
)
408 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
411 static inline void set_task_cfs_rq(struct task_struct
*p
)
416 #ifndef prepare_arch_switch
417 # define prepare_arch_switch(next) do { } while (0)
419 #ifndef finish_arch_switch
420 # define finish_arch_switch(prev) do { } while (0)
423 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
424 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
426 return rq
->curr
== p
;
429 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
433 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
435 #ifdef CONFIG_DEBUG_SPINLOCK
436 /* this is a valid case when another task releases the spinlock */
437 rq
->lock
.owner
= current
;
440 * If we are tracking spinlock dependencies then we have to
441 * fix up the runqueue lock - which gets 'carried over' from
444 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
446 spin_unlock_irq(&rq
->lock
);
449 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
450 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
455 return rq
->curr
== p
;
459 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
463 * We can optimise this out completely for !SMP, because the
464 * SMP rebalancing from interrupt is the only thing that cares
469 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
470 spin_unlock_irq(&rq
->lock
);
472 spin_unlock(&rq
->lock
);
476 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
480 * After ->oncpu is cleared, the task can be moved to a different CPU.
481 * We must ensure this doesn't happen until the switch is completely
487 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
491 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
494 * __task_rq_lock - lock the runqueue a given task resides on.
495 * Must be called interrupts disabled.
497 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
504 spin_lock(&rq
->lock
);
505 if (unlikely(rq
!= task_rq(p
))) {
506 spin_unlock(&rq
->lock
);
507 goto repeat_lock_task
;
513 * task_rq_lock - lock the runqueue a given task resides on and disable
514 * interrupts. Note the ordering: we can safely lookup the task_rq without
515 * explicitly disabling preemption.
517 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
523 local_irq_save(*flags
);
525 spin_lock(&rq
->lock
);
526 if (unlikely(rq
!= task_rq(p
))) {
527 spin_unlock_irqrestore(&rq
->lock
, *flags
);
528 goto repeat_lock_task
;
533 static inline void __task_rq_unlock(struct rq
*rq
)
536 spin_unlock(&rq
->lock
);
539 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
542 spin_unlock_irqrestore(&rq
->lock
, *flags
);
546 * this_rq_lock - lock this runqueue and disable interrupts.
548 static inline struct rq
*this_rq_lock(void)
555 spin_lock(&rq
->lock
);
561 * We are going deep-idle (irqs are disabled):
563 void sched_clock_idle_sleep_event(void)
565 struct rq
*rq
= cpu_rq(smp_processor_id());
567 spin_lock(&rq
->lock
);
568 __update_rq_clock(rq
);
569 spin_unlock(&rq
->lock
);
570 rq
->clock_deep_idle_events
++;
572 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
575 * We just idled delta nanoseconds (called with irqs disabled):
577 void sched_clock_idle_wakeup_event(u64 delta_ns
)
579 struct rq
*rq
= cpu_rq(smp_processor_id());
580 u64 now
= sched_clock();
582 rq
->idle_clock
+= delta_ns
;
584 * Override the previous timestamp and ignore all
585 * sched_clock() deltas that occured while we idled,
586 * and use the PM-provided delta_ns to advance the
589 spin_lock(&rq
->lock
);
590 rq
->prev_clock_raw
= now
;
591 rq
->clock
+= delta_ns
;
592 spin_unlock(&rq
->lock
);
594 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
597 * resched_task - mark a task 'to be rescheduled now'.
599 * On UP this means the setting of the need_resched flag, on SMP it
600 * might also involve a cross-CPU call to trigger the scheduler on
605 #ifndef tsk_is_polling
606 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
609 static void resched_task(struct task_struct
*p
)
613 assert_spin_locked(&task_rq(p
)->lock
);
615 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
618 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
621 if (cpu
== smp_processor_id())
624 /* NEED_RESCHED must be visible before we test polling */
626 if (!tsk_is_polling(p
))
627 smp_send_reschedule(cpu
);
630 static void resched_cpu(int cpu
)
632 struct rq
*rq
= cpu_rq(cpu
);
635 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
637 resched_task(cpu_curr(cpu
));
638 spin_unlock_irqrestore(&rq
->lock
, flags
);
641 static inline void resched_task(struct task_struct
*p
)
643 assert_spin_locked(&task_rq(p
)->lock
);
644 set_tsk_need_resched(p
);
648 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
650 #if BITS_PER_LONG == 32
651 if (likely(divident
<= 0xffffffffULL
))
652 return (u32
)divident
/ divisor
;
653 do_div(divident
, divisor
);
657 return divident
/ divisor
;
661 #if BITS_PER_LONG == 32
662 # define WMULT_CONST (~0UL)
664 # define WMULT_CONST (1UL << 32)
667 #define WMULT_SHIFT 32
670 * Shift right and round:
672 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
675 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
676 struct load_weight
*lw
)
680 if (unlikely(!lw
->inv_weight
))
681 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
683 tmp
= (u64
)delta_exec
* weight
;
685 * Check whether we'd overflow the 64-bit multiplication:
687 if (unlikely(tmp
> WMULT_CONST
))
688 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
691 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
693 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
696 static inline unsigned long
697 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
699 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
702 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
708 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
715 * To aid in avoiding the subversion of "niceness" due to uneven distribution
716 * of tasks with abnormal "nice" values across CPUs the contribution that
717 * each task makes to its run queue's load is weighted according to its
718 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
719 * scaled version of the new time slice allocation that they receive on time
723 #define WEIGHT_IDLEPRIO 2
724 #define WMULT_IDLEPRIO (1 << 31)
727 * Nice levels are multiplicative, with a gentle 10% change for every
728 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
729 * nice 1, it will get ~10% less CPU time than another CPU-bound task
730 * that remained on nice 0.
732 * The "10% effect" is relative and cumulative: from _any_ nice level,
733 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
734 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
735 * If a task goes up by ~10% and another task goes down by ~10% then
736 * the relative distance between them is ~25%.)
738 static const int prio_to_weight
[40] = {
739 /* -20 */ 88761, 71755, 56483, 46273, 36291,
740 /* -15 */ 29154, 23254, 18705, 14949, 11916,
741 /* -10 */ 9548, 7620, 6100, 4904, 3906,
742 /* -5 */ 3121, 2501, 1991, 1586, 1277,
743 /* 0 */ 1024, 820, 655, 526, 423,
744 /* 5 */ 335, 272, 215, 172, 137,
745 /* 10 */ 110, 87, 70, 56, 45,
746 /* 15 */ 36, 29, 23, 18, 15,
750 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
752 * In cases where the weight does not change often, we can use the
753 * precalculated inverse to speed up arithmetics by turning divisions
754 * into multiplications:
756 static const u32 prio_to_wmult
[40] = {
757 /* -20 */ 48388, 59856, 76040, 92818, 118348,
758 /* -15 */ 147320, 184698, 229616, 287308, 360437,
759 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
760 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
761 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
762 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
763 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
764 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
767 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
770 * runqueue iterator, to support SMP load-balancing between different
771 * scheduling classes, without having to expose their internal data
772 * structures to the load-balancing proper:
776 struct task_struct
*(*start
)(void *);
777 struct task_struct
*(*next
)(void *);
780 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
781 unsigned long max_nr_move
, unsigned long max_load_move
,
782 struct sched_domain
*sd
, enum cpu_idle_type idle
,
783 int *all_pinned
, unsigned long *load_moved
,
784 int *this_best_prio
, struct rq_iterator
*iterator
);
786 #include "sched_stats.h"
787 #include "sched_rt.c"
788 #include "sched_fair.c"
789 #include "sched_idletask.c"
790 #ifdef CONFIG_SCHED_DEBUG
791 # include "sched_debug.c"
794 #define sched_class_highest (&rt_sched_class)
796 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
798 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
799 ls
->delta_exec
+= ls
->delta_stat
;
800 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
806 * Update delta_exec, delta_fair fields for rq.
808 * delta_fair clock advances at a rate inversely proportional to
809 * total load (rq->ls.load.weight) on the runqueue, while
810 * delta_exec advances at the same rate as wall-clock (provided
813 * delta_exec / delta_fair is a measure of the (smoothened) load on this
814 * runqueue over any given interval. This (smoothened) load is used
815 * during load balance.
817 * This function is called /before/ updating rq->ls.load
818 * and when switching tasks.
820 static void update_curr_load(struct rq
*rq
)
822 struct load_stat
*ls
= &rq
->ls
;
825 start
= ls
->load_update_start
;
826 ls
->load_update_start
= rq
->clock
;
827 ls
->delta_stat
+= rq
->clock
- start
;
829 * Stagger updates to ls->delta_fair. Very frequent updates
832 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
833 __update_curr_load(rq
, ls
);
836 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
838 update_curr_load(rq
);
839 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
842 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
844 update_curr_load(rq
);
845 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
848 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
854 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
860 static void set_load_weight(struct task_struct
*p
)
862 p
->se
.wait_runtime
= 0;
864 if (task_has_rt_policy(p
)) {
865 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
866 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
871 * SCHED_IDLE tasks get minimal weight:
873 if (p
->policy
== SCHED_IDLE
) {
874 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
875 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
879 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
880 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
883 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
885 sched_info_queued(p
);
886 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
890 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
892 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
897 * __normal_prio - return the priority that is based on the static prio
899 static inline int __normal_prio(struct task_struct
*p
)
901 return p
->static_prio
;
905 * Calculate the expected normal priority: i.e. priority
906 * without taking RT-inheritance into account. Might be
907 * boosted by interactivity modifiers. Changes upon fork,
908 * setprio syscalls, and whenever the interactivity
909 * estimator recalculates.
911 static inline int normal_prio(struct task_struct
*p
)
915 if (task_has_rt_policy(p
))
916 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
918 prio
= __normal_prio(p
);
923 * Calculate the current priority, i.e. the priority
924 * taken into account by the scheduler. This value might
925 * be boosted by RT tasks, or might be boosted by
926 * interactivity modifiers. Will be RT if the task got
927 * RT-boosted. If not then it returns p->normal_prio.
929 static int effective_prio(struct task_struct
*p
)
931 p
->normal_prio
= normal_prio(p
);
933 * If we are RT tasks or we were boosted to RT priority,
934 * keep the priority unchanged. Otherwise, update priority
935 * to the normal priority:
937 if (!rt_prio(p
->prio
))
938 return p
->normal_prio
;
943 * activate_task - move a task to the runqueue.
945 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
947 if (p
->state
== TASK_UNINTERRUPTIBLE
)
948 rq
->nr_uninterruptible
--;
950 enqueue_task(rq
, p
, wakeup
);
951 inc_nr_running(p
, rq
);
955 * activate_idle_task - move idle task to the _front_ of runqueue.
957 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
961 if (p
->state
== TASK_UNINTERRUPTIBLE
)
962 rq
->nr_uninterruptible
--;
964 enqueue_task(rq
, p
, 0);
965 inc_nr_running(p
, rq
);
969 * deactivate_task - remove a task from the runqueue.
971 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
973 if (p
->state
== TASK_UNINTERRUPTIBLE
)
974 rq
->nr_uninterruptible
++;
976 dequeue_task(rq
, p
, sleep
);
977 dec_nr_running(p
, rq
);
981 * task_curr - is this task currently executing on a CPU?
982 * @p: the task in question.
984 inline int task_curr(const struct task_struct
*p
)
986 return cpu_curr(task_cpu(p
)) == p
;
989 /* Used instead of source_load when we know the type == 0 */
990 unsigned long weighted_cpuload(const int cpu
)
992 return cpu_rq(cpu
)->ls
.load
.weight
;
995 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
998 task_thread_info(p
)->cpu
= cpu
;
1005 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1007 int old_cpu
= task_cpu(p
);
1008 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1009 u64 clock_offset
, fair_clock_offset
;
1011 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1012 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
1014 if (p
->se
.wait_start_fair
)
1015 p
->se
.wait_start_fair
-= fair_clock_offset
;
1016 if (p
->se
.sleep_start_fair
)
1017 p
->se
.sleep_start_fair
-= fair_clock_offset
;
1019 #ifdef CONFIG_SCHEDSTATS
1020 if (p
->se
.wait_start
)
1021 p
->se
.wait_start
-= clock_offset
;
1022 if (p
->se
.sleep_start
)
1023 p
->se
.sleep_start
-= clock_offset
;
1024 if (p
->se
.block_start
)
1025 p
->se
.block_start
-= clock_offset
;
1028 __set_task_cpu(p
, new_cpu
);
1031 struct migration_req
{
1032 struct list_head list
;
1034 struct task_struct
*task
;
1037 struct completion done
;
1041 * The task's runqueue lock must be held.
1042 * Returns true if you have to wait for migration thread.
1045 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1047 struct rq
*rq
= task_rq(p
);
1050 * If the task is not on a runqueue (and not running), then
1051 * it is sufficient to simply update the task's cpu field.
1053 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1054 set_task_cpu(p
, dest_cpu
);
1058 init_completion(&req
->done
);
1060 req
->dest_cpu
= dest_cpu
;
1061 list_add(&req
->list
, &rq
->migration_queue
);
1067 * wait_task_inactive - wait for a thread to unschedule.
1069 * The caller must ensure that the task *will* unschedule sometime soon,
1070 * else this function might spin for a *long* time. This function can't
1071 * be called with interrupts off, or it may introduce deadlock with
1072 * smp_call_function() if an IPI is sent by the same process we are
1073 * waiting to become inactive.
1075 void wait_task_inactive(struct task_struct
*p
)
1077 unsigned long flags
;
1083 * We do the initial early heuristics without holding
1084 * any task-queue locks at all. We'll only try to get
1085 * the runqueue lock when things look like they will
1091 * If the task is actively running on another CPU
1092 * still, just relax and busy-wait without holding
1095 * NOTE! Since we don't hold any locks, it's not
1096 * even sure that "rq" stays as the right runqueue!
1097 * But we don't care, since "task_running()" will
1098 * return false if the runqueue has changed and p
1099 * is actually now running somewhere else!
1101 while (task_running(rq
, p
))
1105 * Ok, time to look more closely! We need the rq
1106 * lock now, to be *sure*. If we're wrong, we'll
1107 * just go back and repeat.
1109 rq
= task_rq_lock(p
, &flags
);
1110 running
= task_running(rq
, p
);
1111 on_rq
= p
->se
.on_rq
;
1112 task_rq_unlock(rq
, &flags
);
1115 * Was it really running after all now that we
1116 * checked with the proper locks actually held?
1118 * Oops. Go back and try again..
1120 if (unlikely(running
)) {
1126 * It's not enough that it's not actively running,
1127 * it must be off the runqueue _entirely_, and not
1130 * So if it wa still runnable (but just not actively
1131 * running right now), it's preempted, and we should
1132 * yield - it could be a while.
1134 if (unlikely(on_rq
)) {
1140 * Ahh, all good. It wasn't running, and it wasn't
1141 * runnable, which means that it will never become
1142 * running in the future either. We're all done!
1147 * kick_process - kick a running thread to enter/exit the kernel
1148 * @p: the to-be-kicked thread
1150 * Cause a process which is running on another CPU to enter
1151 * kernel-mode, without any delay. (to get signals handled.)
1153 * NOTE: this function doesnt have to take the runqueue lock,
1154 * because all it wants to ensure is that the remote task enters
1155 * the kernel. If the IPI races and the task has been migrated
1156 * to another CPU then no harm is done and the purpose has been
1159 void kick_process(struct task_struct
*p
)
1165 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1166 smp_send_reschedule(cpu
);
1171 * Return a low guess at the load of a migration-source cpu weighted
1172 * according to the scheduling class and "nice" value.
1174 * We want to under-estimate the load of migration sources, to
1175 * balance conservatively.
1177 static inline unsigned long source_load(int cpu
, int type
)
1179 struct rq
*rq
= cpu_rq(cpu
);
1180 unsigned long total
= weighted_cpuload(cpu
);
1185 return min(rq
->cpu_load
[type
-1], total
);
1189 * Return a high guess at the load of a migration-target cpu weighted
1190 * according to the scheduling class and "nice" value.
1192 static inline unsigned long target_load(int cpu
, int type
)
1194 struct rq
*rq
= cpu_rq(cpu
);
1195 unsigned long total
= weighted_cpuload(cpu
);
1200 return max(rq
->cpu_load
[type
-1], total
);
1204 * Return the average load per task on the cpu's run queue
1206 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1208 struct rq
*rq
= cpu_rq(cpu
);
1209 unsigned long total
= weighted_cpuload(cpu
);
1210 unsigned long n
= rq
->nr_running
;
1212 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1216 * find_idlest_group finds and returns the least busy CPU group within the
1219 static struct sched_group
*
1220 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1222 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1223 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1224 int load_idx
= sd
->forkexec_idx
;
1225 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1228 unsigned long load
, avg_load
;
1232 /* Skip over this group if it has no CPUs allowed */
1233 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1236 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1238 /* Tally up the load of all CPUs in the group */
1241 for_each_cpu_mask(i
, group
->cpumask
) {
1242 /* Bias balancing toward cpus of our domain */
1244 load
= source_load(i
, load_idx
);
1246 load
= target_load(i
, load_idx
);
1251 /* Adjust by relative CPU power of the group */
1252 avg_load
= sg_div_cpu_power(group
,
1253 avg_load
* SCHED_LOAD_SCALE
);
1256 this_load
= avg_load
;
1258 } else if (avg_load
< min_load
) {
1259 min_load
= avg_load
;
1263 group
= group
->next
;
1264 } while (group
!= sd
->groups
);
1266 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1272 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1275 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1278 unsigned long load
, min_load
= ULONG_MAX
;
1282 /* Traverse only the allowed CPUs */
1283 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1285 for_each_cpu_mask(i
, tmp
) {
1286 load
= weighted_cpuload(i
);
1288 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1298 * sched_balance_self: balance the current task (running on cpu) in domains
1299 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1302 * Balance, ie. select the least loaded group.
1304 * Returns the target CPU number, or the same CPU if no balancing is needed.
1306 * preempt must be disabled.
1308 static int sched_balance_self(int cpu
, int flag
)
1310 struct task_struct
*t
= current
;
1311 struct sched_domain
*tmp
, *sd
= NULL
;
1313 for_each_domain(cpu
, tmp
) {
1315 * If power savings logic is enabled for a domain, stop there.
1317 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1319 if (tmp
->flags
& flag
)
1325 struct sched_group
*group
;
1326 int new_cpu
, weight
;
1328 if (!(sd
->flags
& flag
)) {
1334 group
= find_idlest_group(sd
, t
, cpu
);
1340 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1341 if (new_cpu
== -1 || new_cpu
== cpu
) {
1342 /* Now try balancing at a lower domain level of cpu */
1347 /* Now try balancing at a lower domain level of new_cpu */
1350 weight
= cpus_weight(span
);
1351 for_each_domain(cpu
, tmp
) {
1352 if (weight
<= cpus_weight(tmp
->span
))
1354 if (tmp
->flags
& flag
)
1357 /* while loop will break here if sd == NULL */
1363 #endif /* CONFIG_SMP */
1366 * wake_idle() will wake a task on an idle cpu if task->cpu is
1367 * not idle and an idle cpu is available. The span of cpus to
1368 * search starts with cpus closest then further out as needed,
1369 * so we always favor a closer, idle cpu.
1371 * Returns the CPU we should wake onto.
1373 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1374 static int wake_idle(int cpu
, struct task_struct
*p
)
1377 struct sched_domain
*sd
;
1381 * If it is idle, then it is the best cpu to run this task.
1383 * This cpu is also the best, if it has more than one task already.
1384 * Siblings must be also busy(in most cases) as they didn't already
1385 * pickup the extra load from this cpu and hence we need not check
1386 * sibling runqueue info. This will avoid the checks and cache miss
1387 * penalities associated with that.
1389 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1392 for_each_domain(cpu
, sd
) {
1393 if (sd
->flags
& SD_WAKE_IDLE
) {
1394 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1395 for_each_cpu_mask(i
, tmp
) {
1406 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1413 * try_to_wake_up - wake up a thread
1414 * @p: the to-be-woken-up thread
1415 * @state: the mask of task states that can be woken
1416 * @sync: do a synchronous wakeup?
1418 * Put it on the run-queue if it's not already there. The "current"
1419 * thread is always on the run-queue (except when the actual
1420 * re-schedule is in progress), and as such you're allowed to do
1421 * the simpler "current->state = TASK_RUNNING" to mark yourself
1422 * runnable without the overhead of this.
1424 * returns failure only if the task is already active.
1426 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1428 int cpu
, this_cpu
, success
= 0;
1429 unsigned long flags
;
1433 struct sched_domain
*sd
, *this_sd
= NULL
;
1434 unsigned long load
, this_load
;
1438 rq
= task_rq_lock(p
, &flags
);
1439 old_state
= p
->state
;
1440 if (!(old_state
& state
))
1447 this_cpu
= smp_processor_id();
1450 if (unlikely(task_running(rq
, p
)))
1455 schedstat_inc(rq
, ttwu_cnt
);
1456 if (cpu
== this_cpu
) {
1457 schedstat_inc(rq
, ttwu_local
);
1461 for_each_domain(this_cpu
, sd
) {
1462 if (cpu_isset(cpu
, sd
->span
)) {
1463 schedstat_inc(sd
, ttwu_wake_remote
);
1469 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1473 * Check for affine wakeup and passive balancing possibilities.
1476 int idx
= this_sd
->wake_idx
;
1477 unsigned int imbalance
;
1479 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1481 load
= source_load(cpu
, idx
);
1482 this_load
= target_load(this_cpu
, idx
);
1484 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1486 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1487 unsigned long tl
= this_load
;
1488 unsigned long tl_per_task
;
1490 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1493 * If sync wakeup then subtract the (maximum possible)
1494 * effect of the currently running task from the load
1495 * of the current CPU:
1498 tl
-= current
->se
.load
.weight
;
1501 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1502 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1504 * This domain has SD_WAKE_AFFINE and
1505 * p is cache cold in this domain, and
1506 * there is no bad imbalance.
1508 schedstat_inc(this_sd
, ttwu_move_affine
);
1514 * Start passive balancing when half the imbalance_pct
1517 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1518 if (imbalance
*this_load
<= 100*load
) {
1519 schedstat_inc(this_sd
, ttwu_move_balance
);
1525 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1527 new_cpu
= wake_idle(new_cpu
, p
);
1528 if (new_cpu
!= cpu
) {
1529 set_task_cpu(p
, new_cpu
);
1530 task_rq_unlock(rq
, &flags
);
1531 /* might preempt at this point */
1532 rq
= task_rq_lock(p
, &flags
);
1533 old_state
= p
->state
;
1534 if (!(old_state
& state
))
1539 this_cpu
= smp_processor_id();
1544 #endif /* CONFIG_SMP */
1545 update_rq_clock(rq
);
1546 activate_task(rq
, p
, 1);
1548 * Sync wakeups (i.e. those types of wakeups where the waker
1549 * has indicated that it will leave the CPU in short order)
1550 * don't trigger a preemption, if the woken up task will run on
1551 * this cpu. (in this case the 'I will reschedule' promise of
1552 * the waker guarantees that the freshly woken up task is going
1553 * to be considered on this CPU.)
1555 if (!sync
|| cpu
!= this_cpu
)
1556 check_preempt_curr(rq
, p
);
1560 p
->state
= TASK_RUNNING
;
1562 task_rq_unlock(rq
, &flags
);
1567 int fastcall
wake_up_process(struct task_struct
*p
)
1569 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1570 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1572 EXPORT_SYMBOL(wake_up_process
);
1574 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1576 return try_to_wake_up(p
, state
, 0);
1580 * Perform scheduler related setup for a newly forked process p.
1581 * p is forked by current.
1583 * __sched_fork() is basic setup used by init_idle() too:
1585 static void __sched_fork(struct task_struct
*p
)
1587 p
->se
.wait_start_fair
= 0;
1588 p
->se
.exec_start
= 0;
1589 p
->se
.sum_exec_runtime
= 0;
1590 p
->se
.prev_sum_exec_runtime
= 0;
1591 p
->se
.delta_exec
= 0;
1592 p
->se
.delta_fair_run
= 0;
1593 p
->se
.delta_fair_sleep
= 0;
1594 p
->se
.wait_runtime
= 0;
1595 p
->se
.sleep_start_fair
= 0;
1597 #ifdef CONFIG_SCHEDSTATS
1598 p
->se
.wait_start
= 0;
1599 p
->se
.sum_wait_runtime
= 0;
1600 p
->se
.sum_sleep_runtime
= 0;
1601 p
->se
.sleep_start
= 0;
1602 p
->se
.block_start
= 0;
1603 p
->se
.sleep_max
= 0;
1604 p
->se
.block_max
= 0;
1606 p
->se
.slice_max
= 0;
1608 p
->se
.wait_runtime_overruns
= 0;
1609 p
->se
.wait_runtime_underruns
= 0;
1612 INIT_LIST_HEAD(&p
->run_list
);
1615 #ifdef CONFIG_PREEMPT_NOTIFIERS
1616 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1620 * We mark the process as running here, but have not actually
1621 * inserted it onto the runqueue yet. This guarantees that
1622 * nobody will actually run it, and a signal or other external
1623 * event cannot wake it up and insert it on the runqueue either.
1625 p
->state
= TASK_RUNNING
;
1629 * fork()/clone()-time setup:
1631 void sched_fork(struct task_struct
*p
, int clone_flags
)
1633 int cpu
= get_cpu();
1638 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1640 __set_task_cpu(p
, cpu
);
1643 * Make sure we do not leak PI boosting priority to the child:
1645 p
->prio
= current
->normal_prio
;
1647 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1648 if (likely(sched_info_on()))
1649 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1651 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1654 #ifdef CONFIG_PREEMPT
1655 /* Want to start with kernel preemption disabled. */
1656 task_thread_info(p
)->preempt_count
= 1;
1662 * wake_up_new_task - wake up a newly created task for the first time.
1664 * This function will do some initial scheduler statistics housekeeping
1665 * that must be done for every newly created context, then puts the task
1666 * on the runqueue and wakes it.
1668 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1670 unsigned long flags
;
1674 rq
= task_rq_lock(p
, &flags
);
1675 BUG_ON(p
->state
!= TASK_RUNNING
);
1676 this_cpu
= smp_processor_id(); /* parent's CPU */
1677 update_rq_clock(rq
);
1679 p
->prio
= effective_prio(p
);
1681 if (rt_prio(p
->prio
))
1682 p
->sched_class
= &rt_sched_class
;
1684 p
->sched_class
= &fair_sched_class
;
1686 if (task_cpu(p
) != this_cpu
|| !p
->sched_class
->task_new
||
1687 !current
->se
.on_rq
) {
1688 activate_task(rq
, p
, 0);
1691 * Let the scheduling class do new task startup
1692 * management (if any):
1694 p
->sched_class
->task_new(rq
, p
);
1695 inc_nr_running(p
, rq
);
1697 check_preempt_curr(rq
, p
);
1698 task_rq_unlock(rq
, &flags
);
1701 #ifdef CONFIG_PREEMPT_NOTIFIERS
1704 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1705 * @notifier: notifier struct to register
1707 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1709 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1711 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1714 * preempt_notifier_unregister - no longer interested in preemption notifications
1715 * @notifier: notifier struct to unregister
1717 * This is safe to call from within a preemption notifier.
1719 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1721 hlist_del(¬ifier
->link
);
1723 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1725 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1727 struct preempt_notifier
*notifier
;
1728 struct hlist_node
*node
;
1730 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1731 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1735 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1736 struct task_struct
*next
)
1738 struct preempt_notifier
*notifier
;
1739 struct hlist_node
*node
;
1741 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1742 notifier
->ops
->sched_out(notifier
, next
);
1747 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1752 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1753 struct task_struct
*next
)
1760 * prepare_task_switch - prepare to switch tasks
1761 * @rq: the runqueue preparing to switch
1762 * @prev: the current task that is being switched out
1763 * @next: the task we are going to switch to.
1765 * This is called with the rq lock held and interrupts off. It must
1766 * be paired with a subsequent finish_task_switch after the context
1769 * prepare_task_switch sets up locking and calls architecture specific
1773 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1774 struct task_struct
*next
)
1776 fire_sched_out_preempt_notifiers(prev
, next
);
1777 prepare_lock_switch(rq
, next
);
1778 prepare_arch_switch(next
);
1782 * finish_task_switch - clean up after a task-switch
1783 * @rq: runqueue associated with task-switch
1784 * @prev: the thread we just switched away from.
1786 * finish_task_switch must be called after the context switch, paired
1787 * with a prepare_task_switch call before the context switch.
1788 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1789 * and do any other architecture-specific cleanup actions.
1791 * Note that we may have delayed dropping an mm in context_switch(). If
1792 * so, we finish that here outside of the runqueue lock. (Doing it
1793 * with the lock held can cause deadlocks; see schedule() for
1796 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1797 __releases(rq
->lock
)
1799 struct mm_struct
*mm
= rq
->prev_mm
;
1805 * A task struct has one reference for the use as "current".
1806 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1807 * schedule one last time. The schedule call will never return, and
1808 * the scheduled task must drop that reference.
1809 * The test for TASK_DEAD must occur while the runqueue locks are
1810 * still held, otherwise prev could be scheduled on another cpu, die
1811 * there before we look at prev->state, and then the reference would
1813 * Manfred Spraul <manfred@colorfullife.com>
1815 prev_state
= prev
->state
;
1816 finish_arch_switch(prev
);
1817 finish_lock_switch(rq
, prev
);
1818 fire_sched_in_preempt_notifiers(current
);
1821 if (unlikely(prev_state
== TASK_DEAD
)) {
1823 * Remove function-return probe instances associated with this
1824 * task and put them back on the free list.
1826 kprobe_flush_task(prev
);
1827 put_task_struct(prev
);
1832 * schedule_tail - first thing a freshly forked thread must call.
1833 * @prev: the thread we just switched away from.
1835 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1836 __releases(rq
->lock
)
1838 struct rq
*rq
= this_rq();
1840 finish_task_switch(rq
, prev
);
1841 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1842 /* In this case, finish_task_switch does not reenable preemption */
1845 if (current
->set_child_tid
)
1846 put_user(current
->pid
, current
->set_child_tid
);
1850 * context_switch - switch to the new MM and the new
1851 * thread's register state.
1854 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1855 struct task_struct
*next
)
1857 struct mm_struct
*mm
, *oldmm
;
1859 prepare_task_switch(rq
, prev
, next
);
1861 oldmm
= prev
->active_mm
;
1863 * For paravirt, this is coupled with an exit in switch_to to
1864 * combine the page table reload and the switch backend into
1867 arch_enter_lazy_cpu_mode();
1869 if (unlikely(!mm
)) {
1870 next
->active_mm
= oldmm
;
1871 atomic_inc(&oldmm
->mm_count
);
1872 enter_lazy_tlb(oldmm
, next
);
1874 switch_mm(oldmm
, mm
, next
);
1876 if (unlikely(!prev
->mm
)) {
1877 prev
->active_mm
= NULL
;
1878 rq
->prev_mm
= oldmm
;
1881 * Since the runqueue lock will be released by the next
1882 * task (which is an invalid locking op but in the case
1883 * of the scheduler it's an obvious special-case), so we
1884 * do an early lockdep release here:
1886 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1887 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1890 /* Here we just switch the register state and the stack. */
1891 switch_to(prev
, next
, prev
);
1895 * this_rq must be evaluated again because prev may have moved
1896 * CPUs since it called schedule(), thus the 'rq' on its stack
1897 * frame will be invalid.
1899 finish_task_switch(this_rq(), prev
);
1903 * nr_running, nr_uninterruptible and nr_context_switches:
1905 * externally visible scheduler statistics: current number of runnable
1906 * threads, current number of uninterruptible-sleeping threads, total
1907 * number of context switches performed since bootup.
1909 unsigned long nr_running(void)
1911 unsigned long i
, sum
= 0;
1913 for_each_online_cpu(i
)
1914 sum
+= cpu_rq(i
)->nr_running
;
1919 unsigned long nr_uninterruptible(void)
1921 unsigned long i
, sum
= 0;
1923 for_each_possible_cpu(i
)
1924 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1927 * Since we read the counters lockless, it might be slightly
1928 * inaccurate. Do not allow it to go below zero though:
1930 if (unlikely((long)sum
< 0))
1936 unsigned long long nr_context_switches(void)
1939 unsigned long long sum
= 0;
1941 for_each_possible_cpu(i
)
1942 sum
+= cpu_rq(i
)->nr_switches
;
1947 unsigned long nr_iowait(void)
1949 unsigned long i
, sum
= 0;
1951 for_each_possible_cpu(i
)
1952 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1957 unsigned long nr_active(void)
1959 unsigned long i
, running
= 0, uninterruptible
= 0;
1961 for_each_online_cpu(i
) {
1962 running
+= cpu_rq(i
)->nr_running
;
1963 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1966 if (unlikely((long)uninterruptible
< 0))
1967 uninterruptible
= 0;
1969 return running
+ uninterruptible
;
1973 * Update rq->cpu_load[] statistics. This function is usually called every
1974 * scheduler tick (TICK_NSEC).
1976 static void update_cpu_load(struct rq
*this_rq
)
1978 u64 fair_delta64
, exec_delta64
, idle_delta64
, sample_interval64
, tmp64
;
1979 unsigned long total_load
= this_rq
->ls
.load
.weight
;
1980 unsigned long this_load
= total_load
;
1981 struct load_stat
*ls
= &this_rq
->ls
;
1984 this_rq
->nr_load_updates
++;
1985 if (unlikely(!(sysctl_sched_features
& SCHED_FEAT_PRECISE_CPU_LOAD
)))
1988 /* Update delta_fair/delta_exec fields first */
1989 update_curr_load(this_rq
);
1991 fair_delta64
= ls
->delta_fair
+ 1;
1994 exec_delta64
= ls
->delta_exec
+ 1;
1997 sample_interval64
= this_rq
->clock
- ls
->load_update_last
;
1998 ls
->load_update_last
= this_rq
->clock
;
2000 if ((s64
)sample_interval64
< (s64
)TICK_NSEC
)
2001 sample_interval64
= TICK_NSEC
;
2003 if (exec_delta64
> sample_interval64
)
2004 exec_delta64
= sample_interval64
;
2006 idle_delta64
= sample_interval64
- exec_delta64
;
2008 tmp64
= div64_64(SCHED_LOAD_SCALE
* exec_delta64
, fair_delta64
);
2009 tmp64
= div64_64(tmp64
* exec_delta64
, sample_interval64
);
2011 this_load
= (unsigned long)tmp64
;
2015 /* Update our load: */
2016 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2017 unsigned long old_load
, new_load
;
2019 /* scale is effectively 1 << i now, and >> i divides by scale */
2021 old_load
= this_rq
->cpu_load
[i
];
2022 new_load
= this_load
;
2024 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2031 * double_rq_lock - safely lock two runqueues
2033 * Note this does not disable interrupts like task_rq_lock,
2034 * you need to do so manually before calling.
2036 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2037 __acquires(rq1
->lock
)
2038 __acquires(rq2
->lock
)
2040 BUG_ON(!irqs_disabled());
2042 spin_lock(&rq1
->lock
);
2043 __acquire(rq2
->lock
); /* Fake it out ;) */
2046 spin_lock(&rq1
->lock
);
2047 spin_lock(&rq2
->lock
);
2049 spin_lock(&rq2
->lock
);
2050 spin_lock(&rq1
->lock
);
2053 update_rq_clock(rq1
);
2054 update_rq_clock(rq2
);
2058 * double_rq_unlock - safely unlock two runqueues
2060 * Note this does not restore interrupts like task_rq_unlock,
2061 * you need to do so manually after calling.
2063 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2064 __releases(rq1
->lock
)
2065 __releases(rq2
->lock
)
2067 spin_unlock(&rq1
->lock
);
2069 spin_unlock(&rq2
->lock
);
2071 __release(rq2
->lock
);
2075 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2077 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2078 __releases(this_rq
->lock
)
2079 __acquires(busiest
->lock
)
2080 __acquires(this_rq
->lock
)
2082 if (unlikely(!irqs_disabled())) {
2083 /* printk() doesn't work good under rq->lock */
2084 spin_unlock(&this_rq
->lock
);
2087 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2088 if (busiest
< this_rq
) {
2089 spin_unlock(&this_rq
->lock
);
2090 spin_lock(&busiest
->lock
);
2091 spin_lock(&this_rq
->lock
);
2093 spin_lock(&busiest
->lock
);
2098 * If dest_cpu is allowed for this process, migrate the task to it.
2099 * This is accomplished by forcing the cpu_allowed mask to only
2100 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2101 * the cpu_allowed mask is restored.
2103 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2105 struct migration_req req
;
2106 unsigned long flags
;
2109 rq
= task_rq_lock(p
, &flags
);
2110 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2111 || unlikely(cpu_is_offline(dest_cpu
)))
2114 /* force the process onto the specified CPU */
2115 if (migrate_task(p
, dest_cpu
, &req
)) {
2116 /* Need to wait for migration thread (might exit: take ref). */
2117 struct task_struct
*mt
= rq
->migration_thread
;
2119 get_task_struct(mt
);
2120 task_rq_unlock(rq
, &flags
);
2121 wake_up_process(mt
);
2122 put_task_struct(mt
);
2123 wait_for_completion(&req
.done
);
2128 task_rq_unlock(rq
, &flags
);
2132 * sched_exec - execve() is a valuable balancing opportunity, because at
2133 * this point the task has the smallest effective memory and cache footprint.
2135 void sched_exec(void)
2137 int new_cpu
, this_cpu
= get_cpu();
2138 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2140 if (new_cpu
!= this_cpu
)
2141 sched_migrate_task(current
, new_cpu
);
2145 * pull_task - move a task from a remote runqueue to the local runqueue.
2146 * Both runqueues must be locked.
2148 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2149 struct rq
*this_rq
, int this_cpu
)
2151 deactivate_task(src_rq
, p
, 0);
2152 set_task_cpu(p
, this_cpu
);
2153 activate_task(this_rq
, p
, 0);
2155 * Note that idle threads have a prio of MAX_PRIO, for this test
2156 * to be always true for them.
2158 check_preempt_curr(this_rq
, p
);
2162 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2165 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2166 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2170 * We do not migrate tasks that are:
2171 * 1) running (obviously), or
2172 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2173 * 3) are cache-hot on their current CPU.
2175 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2179 if (task_running(rq
, p
))
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
< busiest_load_per_task
) {
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
)
2566 *imbalance
= busiest_load_per_task
;
2572 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2573 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2576 if (this == group_leader
&& group_leader
!= group_min
) {
2577 *imbalance
= min_load_per_task
;
2587 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2590 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2591 unsigned long imbalance
, cpumask_t
*cpus
)
2593 struct rq
*busiest
= NULL
, *rq
;
2594 unsigned long max_load
= 0;
2597 for_each_cpu_mask(i
, group
->cpumask
) {
2600 if (!cpu_isset(i
, *cpus
))
2604 wl
= weighted_cpuload(i
);
2606 if (rq
->nr_running
== 1 && wl
> imbalance
)
2609 if (wl
> max_load
) {
2619 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2620 * so long as it is large enough.
2622 #define MAX_PINNED_INTERVAL 512
2625 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2626 * tasks if there is an imbalance.
2628 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2629 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2632 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2633 struct sched_group
*group
;
2634 unsigned long imbalance
;
2636 cpumask_t cpus
= CPU_MASK_ALL
;
2637 unsigned long flags
;
2640 * When power savings policy is enabled for the parent domain, idle
2641 * sibling can pick up load irrespective of busy siblings. In this case,
2642 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2643 * portraying it as CPU_NOT_IDLE.
2645 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2646 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2649 schedstat_inc(sd
, lb_cnt
[idle
]);
2652 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2659 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2663 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2665 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2669 BUG_ON(busiest
== this_rq
);
2671 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2674 if (busiest
->nr_running
> 1) {
2676 * Attempt to move tasks. If find_busiest_group has found
2677 * an imbalance but busiest->nr_running <= 1, the group is
2678 * still unbalanced. ld_moved simply stays zero, so it is
2679 * correctly treated as an imbalance.
2681 local_irq_save(flags
);
2682 double_rq_lock(this_rq
, busiest
);
2683 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2684 imbalance
, sd
, idle
, &all_pinned
);
2685 double_rq_unlock(this_rq
, busiest
);
2686 local_irq_restore(flags
);
2689 * some other cpu did the load balance for us.
2691 if (ld_moved
&& this_cpu
!= smp_processor_id())
2692 resched_cpu(this_cpu
);
2694 /* All tasks on this runqueue were pinned by CPU affinity */
2695 if (unlikely(all_pinned
)) {
2696 cpu_clear(cpu_of(busiest
), cpus
);
2697 if (!cpus_empty(cpus
))
2704 schedstat_inc(sd
, lb_failed
[idle
]);
2705 sd
->nr_balance_failed
++;
2707 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2709 spin_lock_irqsave(&busiest
->lock
, flags
);
2711 /* don't kick the migration_thread, if the curr
2712 * task on busiest cpu can't be moved to this_cpu
2714 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2715 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2717 goto out_one_pinned
;
2720 if (!busiest
->active_balance
) {
2721 busiest
->active_balance
= 1;
2722 busiest
->push_cpu
= this_cpu
;
2725 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2727 wake_up_process(busiest
->migration_thread
);
2730 * We've kicked active balancing, reset the failure
2733 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2736 sd
->nr_balance_failed
= 0;
2738 if (likely(!active_balance
)) {
2739 /* We were unbalanced, so reset the balancing interval */
2740 sd
->balance_interval
= sd
->min_interval
;
2743 * If we've begun active balancing, start to back off. This
2744 * case may not be covered by the all_pinned logic if there
2745 * is only 1 task on the busy runqueue (because we don't call
2748 if (sd
->balance_interval
< sd
->max_interval
)
2749 sd
->balance_interval
*= 2;
2752 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2753 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2758 schedstat_inc(sd
, lb_balanced
[idle
]);
2760 sd
->nr_balance_failed
= 0;
2763 /* tune up the balancing interval */
2764 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2765 (sd
->balance_interval
< sd
->max_interval
))
2766 sd
->balance_interval
*= 2;
2768 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2769 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2775 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2776 * tasks if there is an imbalance.
2778 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2779 * this_rq is locked.
2782 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2784 struct sched_group
*group
;
2785 struct rq
*busiest
= NULL
;
2786 unsigned long imbalance
;
2790 cpumask_t cpus
= CPU_MASK_ALL
;
2793 * When power savings policy is enabled for the parent domain, idle
2794 * sibling can pick up load irrespective of busy siblings. In this case,
2795 * let the state of idle sibling percolate up as IDLE, instead of
2796 * portraying it as CPU_NOT_IDLE.
2798 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2799 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2802 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2804 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2805 &sd_idle
, &cpus
, NULL
);
2807 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2811 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2814 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2818 BUG_ON(busiest
== this_rq
);
2820 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2823 if (busiest
->nr_running
> 1) {
2824 /* Attempt to move tasks */
2825 double_lock_balance(this_rq
, busiest
);
2826 /* this_rq->clock is already updated */
2827 update_rq_clock(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
);
2925 update_rq_clock(busiest_rq
);
2926 update_rq_clock(target_rq
);
2928 /* Search for an sd spanning us and the target CPU. */
2929 for_each_domain(target_cpu
, sd
) {
2930 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2931 cpu_isset(busiest_cpu
, sd
->span
))
2936 schedstat_inc(sd
, alb_cnt
);
2938 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2940 schedstat_inc(sd
, alb_pushed
);
2942 schedstat_inc(sd
, alb_failed
);
2944 spin_unlock(&target_rq
->lock
);
2949 atomic_t load_balancer
;
2951 } nohz ____cacheline_aligned
= {
2952 .load_balancer
= ATOMIC_INIT(-1),
2953 .cpu_mask
= CPU_MASK_NONE
,
2957 * This routine will try to nominate the ilb (idle load balancing)
2958 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2959 * load balancing on behalf of all those cpus. If all the cpus in the system
2960 * go into this tickless mode, then there will be no ilb owner (as there is
2961 * no need for one) and all the cpus will sleep till the next wakeup event
2964 * For the ilb owner, tick is not stopped. And this tick will be used
2965 * for idle load balancing. ilb owner will still be part of
2968 * While stopping the tick, this cpu will become the ilb owner if there
2969 * is no other owner. And will be the owner till that cpu becomes busy
2970 * or if all cpus in the system stop their ticks at which point
2971 * there is no need for ilb owner.
2973 * When the ilb owner becomes busy, it nominates another owner, during the
2974 * next busy scheduler_tick()
2976 int select_nohz_load_balancer(int stop_tick
)
2978 int cpu
= smp_processor_id();
2981 cpu_set(cpu
, nohz
.cpu_mask
);
2982 cpu_rq(cpu
)->in_nohz_recently
= 1;
2985 * If we are going offline and still the leader, give up!
2987 if (cpu_is_offline(cpu
) &&
2988 atomic_read(&nohz
.load_balancer
) == cpu
) {
2989 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2994 /* time for ilb owner also to sleep */
2995 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2996 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2997 atomic_set(&nohz
.load_balancer
, -1);
3001 if (atomic_read(&nohz
.load_balancer
) == -1) {
3002 /* make me the ilb owner */
3003 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3005 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3008 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3011 cpu_clear(cpu
, nohz
.cpu_mask
);
3013 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3014 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3021 static DEFINE_SPINLOCK(balancing
);
3024 * It checks each scheduling domain to see if it is due to be balanced,
3025 * and initiates a balancing operation if so.
3027 * Balancing parameters are set up in arch_init_sched_domains.
3029 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3032 struct rq
*rq
= cpu_rq(cpu
);
3033 unsigned long interval
;
3034 struct sched_domain
*sd
;
3035 /* Earliest time when we have to do rebalance again */
3036 unsigned long next_balance
= jiffies
+ 60*HZ
;
3037 int update_next_balance
= 0;
3039 for_each_domain(cpu
, sd
) {
3040 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3043 interval
= sd
->balance_interval
;
3044 if (idle
!= CPU_IDLE
)
3045 interval
*= sd
->busy_factor
;
3047 /* scale ms to jiffies */
3048 interval
= msecs_to_jiffies(interval
);
3049 if (unlikely(!interval
))
3051 if (interval
> HZ
*NR_CPUS
/10)
3052 interval
= HZ
*NR_CPUS
/10;
3055 if (sd
->flags
& SD_SERIALIZE
) {
3056 if (!spin_trylock(&balancing
))
3060 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3061 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3063 * We've pulled tasks over so either we're no
3064 * longer idle, or one of our SMT siblings is
3067 idle
= CPU_NOT_IDLE
;
3069 sd
->last_balance
= jiffies
;
3071 if (sd
->flags
& SD_SERIALIZE
)
3072 spin_unlock(&balancing
);
3074 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3075 next_balance
= sd
->last_balance
+ interval
;
3076 update_next_balance
= 1;
3080 * Stop the load balance at this level. There is another
3081 * CPU in our sched group which is doing load balancing more
3089 * next_balance will be updated only when there is a need.
3090 * When the cpu is attached to null domain for ex, it will not be
3093 if (likely(update_next_balance
))
3094 rq
->next_balance
= next_balance
;
3098 * run_rebalance_domains is triggered when needed from the scheduler tick.
3099 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3100 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3102 static void run_rebalance_domains(struct softirq_action
*h
)
3104 int this_cpu
= smp_processor_id();
3105 struct rq
*this_rq
= cpu_rq(this_cpu
);
3106 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3107 CPU_IDLE
: CPU_NOT_IDLE
;
3109 rebalance_domains(this_cpu
, idle
);
3113 * If this cpu is the owner for idle load balancing, then do the
3114 * balancing on behalf of the other idle cpus whose ticks are
3117 if (this_rq
->idle_at_tick
&&
3118 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3119 cpumask_t cpus
= nohz
.cpu_mask
;
3123 cpu_clear(this_cpu
, cpus
);
3124 for_each_cpu_mask(balance_cpu
, cpus
) {
3126 * If this cpu gets work to do, stop the load balancing
3127 * work being done for other cpus. Next load
3128 * balancing owner will pick it up.
3133 rebalance_domains(balance_cpu
, CPU_IDLE
);
3135 rq
= cpu_rq(balance_cpu
);
3136 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3137 this_rq
->next_balance
= rq
->next_balance
;
3144 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3146 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3147 * idle load balancing owner or decide to stop the periodic load balancing,
3148 * if the whole system is idle.
3150 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3154 * If we were in the nohz mode recently and busy at the current
3155 * scheduler tick, then check if we need to nominate new idle
3158 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3159 rq
->in_nohz_recently
= 0;
3161 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3162 cpu_clear(cpu
, nohz
.cpu_mask
);
3163 atomic_set(&nohz
.load_balancer
, -1);
3166 if (atomic_read(&nohz
.load_balancer
) == -1) {
3168 * simple selection for now: Nominate the
3169 * first cpu in the nohz list to be the next
3172 * TBD: Traverse the sched domains and nominate
3173 * the nearest cpu in the nohz.cpu_mask.
3175 int ilb
= first_cpu(nohz
.cpu_mask
);
3183 * If this cpu is idle and doing idle load balancing for all the
3184 * cpus with ticks stopped, is it time for that to stop?
3186 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3187 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3193 * If this cpu is idle and the idle load balancing is done by
3194 * someone else, then no need raise the SCHED_SOFTIRQ
3196 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3197 cpu_isset(cpu
, nohz
.cpu_mask
))
3200 if (time_after_eq(jiffies
, rq
->next_balance
))
3201 raise_softirq(SCHED_SOFTIRQ
);
3204 #else /* CONFIG_SMP */
3207 * on UP we do not need to balance between CPUs:
3209 static inline void idle_balance(int cpu
, struct rq
*rq
)
3213 /* Avoid "used but not defined" warning on UP */
3214 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3215 unsigned long max_nr_move
, unsigned long max_load_move
,
3216 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3217 int *all_pinned
, unsigned long *load_moved
,
3218 int *this_best_prio
, struct rq_iterator
*iterator
)
3227 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3229 EXPORT_PER_CPU_SYMBOL(kstat
);
3232 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3233 * that have not yet been banked in case the task is currently running.
3235 unsigned long long task_sched_runtime(struct task_struct
*p
)
3237 unsigned long flags
;
3241 rq
= task_rq_lock(p
, &flags
);
3242 ns
= p
->se
.sum_exec_runtime
;
3243 if (rq
->curr
== p
) {
3244 update_rq_clock(rq
);
3245 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3246 if ((s64
)delta_exec
> 0)
3249 task_rq_unlock(rq
, &flags
);
3255 * Account user cpu time to a process.
3256 * @p: the process that the cpu time gets accounted to
3257 * @hardirq_offset: the offset to subtract from hardirq_count()
3258 * @cputime: the cpu time spent in user space since the last update
3260 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3262 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3265 p
->utime
= cputime_add(p
->utime
, cputime
);
3267 /* Add user time to cpustat. */
3268 tmp
= cputime_to_cputime64(cputime
);
3269 if (TASK_NICE(p
) > 0)
3270 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3272 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3276 * Account system cpu time to a process.
3277 * @p: the process that the cpu time gets accounted to
3278 * @hardirq_offset: the offset to subtract from hardirq_count()
3279 * @cputime: the cpu time spent in kernel space since the last update
3281 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3284 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3285 struct rq
*rq
= this_rq();
3288 p
->stime
= cputime_add(p
->stime
, cputime
);
3290 /* Add system time to cpustat. */
3291 tmp
= cputime_to_cputime64(cputime
);
3292 if (hardirq_count() - hardirq_offset
)
3293 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3294 else if (softirq_count())
3295 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3296 else if (p
!= rq
->idle
)
3297 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3298 else if (atomic_read(&rq
->nr_iowait
) > 0)
3299 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3301 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3302 /* Account for system time used */
3303 acct_update_integrals(p
);
3307 * Account for involuntary wait time.
3308 * @p: the process from which the cpu time has been stolen
3309 * @steal: the cpu time spent in involuntary wait
3311 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3313 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3314 cputime64_t tmp
= cputime_to_cputime64(steal
);
3315 struct rq
*rq
= this_rq();
3317 if (p
== rq
->idle
) {
3318 p
->stime
= cputime_add(p
->stime
, steal
);
3319 if (atomic_read(&rq
->nr_iowait
) > 0)
3320 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3322 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3324 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3328 * This function gets called by the timer code, with HZ frequency.
3329 * We call it with interrupts disabled.
3331 * It also gets called by the fork code, when changing the parent's
3334 void scheduler_tick(void)
3336 int cpu
= smp_processor_id();
3337 struct rq
*rq
= cpu_rq(cpu
);
3338 struct task_struct
*curr
= rq
->curr
;
3339 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3341 spin_lock(&rq
->lock
);
3342 __update_rq_clock(rq
);
3344 * Let rq->clock advance by at least TICK_NSEC:
3346 if (unlikely(rq
->clock
< next_tick
))
3347 rq
->clock
= next_tick
;
3348 rq
->tick_timestamp
= rq
->clock
;
3349 update_cpu_load(rq
);
3350 if (curr
!= rq
->idle
) /* FIXME: needed? */
3351 curr
->sched_class
->task_tick(rq
, curr
);
3352 spin_unlock(&rq
->lock
);
3355 rq
->idle_at_tick
= idle_cpu(cpu
);
3356 trigger_load_balance(rq
, cpu
);
3360 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3362 void fastcall
add_preempt_count(int val
)
3367 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3369 preempt_count() += val
;
3371 * Spinlock count overflowing soon?
3373 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3376 EXPORT_SYMBOL(add_preempt_count
);
3378 void fastcall
sub_preempt_count(int val
)
3383 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3386 * Is the spinlock portion underflowing?
3388 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3389 !(preempt_count() & PREEMPT_MASK
)))
3392 preempt_count() -= val
;
3394 EXPORT_SYMBOL(sub_preempt_count
);
3399 * Print scheduling while atomic bug:
3401 static noinline
void __schedule_bug(struct task_struct
*prev
)
3403 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3404 prev
->comm
, preempt_count(), prev
->pid
);
3405 debug_show_held_locks(prev
);
3406 if (irqs_disabled())
3407 print_irqtrace_events(prev
);
3412 * Various schedule()-time debugging checks and statistics:
3414 static inline void schedule_debug(struct task_struct
*prev
)
3417 * Test if we are atomic. Since do_exit() needs to call into
3418 * schedule() atomically, we ignore that path for now.
3419 * Otherwise, whine if we are scheduling when we should not be.
3421 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3422 __schedule_bug(prev
);
3424 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3426 schedstat_inc(this_rq(), sched_cnt
);
3430 * Pick up the highest-prio task:
3432 static inline struct task_struct
*
3433 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3435 struct sched_class
*class;
3436 struct task_struct
*p
;
3439 * Optimization: we know that if all tasks are in
3440 * the fair class we can call that function directly:
3442 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3443 p
= fair_sched_class
.pick_next_task(rq
);
3448 class = sched_class_highest
;
3450 p
= class->pick_next_task(rq
);
3454 * Will never be NULL as the idle class always
3455 * returns a non-NULL p:
3457 class = class->next
;
3462 * schedule() is the main scheduler function.
3464 asmlinkage
void __sched
schedule(void)
3466 struct task_struct
*prev
, *next
;
3473 cpu
= smp_processor_id();
3477 switch_count
= &prev
->nivcsw
;
3479 release_kernel_lock(prev
);
3480 need_resched_nonpreemptible
:
3482 schedule_debug(prev
);
3484 spin_lock_irq(&rq
->lock
);
3485 clear_tsk_need_resched(prev
);
3486 __update_rq_clock(rq
);
3488 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3489 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3490 unlikely(signal_pending(prev
)))) {
3491 prev
->state
= TASK_RUNNING
;
3493 deactivate_task(rq
, prev
, 1);
3495 switch_count
= &prev
->nvcsw
;
3498 if (unlikely(!rq
->nr_running
))
3499 idle_balance(cpu
, rq
);
3501 prev
->sched_class
->put_prev_task(rq
, prev
);
3502 next
= pick_next_task(rq
, prev
);
3504 sched_info_switch(prev
, next
);
3506 if (likely(prev
!= next
)) {
3511 context_switch(rq
, prev
, next
); /* unlocks the rq */
3513 spin_unlock_irq(&rq
->lock
);
3515 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3516 cpu
= smp_processor_id();
3518 goto need_resched_nonpreemptible
;
3520 preempt_enable_no_resched();
3521 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3524 EXPORT_SYMBOL(schedule
);
3526 #ifdef CONFIG_PREEMPT
3528 * this is the entry point to schedule() from in-kernel preemption
3529 * off of preempt_enable. Kernel preemptions off return from interrupt
3530 * occur there and call schedule directly.
3532 asmlinkage
void __sched
preempt_schedule(void)
3534 struct thread_info
*ti
= current_thread_info();
3535 #ifdef CONFIG_PREEMPT_BKL
3536 struct task_struct
*task
= current
;
3537 int saved_lock_depth
;
3540 * If there is a non-zero preempt_count or interrupts are disabled,
3541 * we do not want to preempt the current task. Just return..
3543 if (likely(ti
->preempt_count
|| irqs_disabled()))
3547 add_preempt_count(PREEMPT_ACTIVE
);
3549 * We keep the big kernel semaphore locked, but we
3550 * clear ->lock_depth so that schedule() doesnt
3551 * auto-release the semaphore:
3553 #ifdef CONFIG_PREEMPT_BKL
3554 saved_lock_depth
= task
->lock_depth
;
3555 task
->lock_depth
= -1;
3558 #ifdef CONFIG_PREEMPT_BKL
3559 task
->lock_depth
= saved_lock_depth
;
3561 sub_preempt_count(PREEMPT_ACTIVE
);
3563 /* we could miss a preemption opportunity between schedule and now */
3565 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3568 EXPORT_SYMBOL(preempt_schedule
);
3571 * this is the entry point to schedule() from kernel preemption
3572 * off of irq context.
3573 * Note, that this is called and return with irqs disabled. This will
3574 * protect us against recursive calling from irq.
3576 asmlinkage
void __sched
preempt_schedule_irq(void)
3578 struct thread_info
*ti
= current_thread_info();
3579 #ifdef CONFIG_PREEMPT_BKL
3580 struct task_struct
*task
= current
;
3581 int saved_lock_depth
;
3583 /* Catch callers which need to be fixed */
3584 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3587 add_preempt_count(PREEMPT_ACTIVE
);
3589 * We keep the big kernel semaphore locked, but we
3590 * clear ->lock_depth so that schedule() doesnt
3591 * auto-release the semaphore:
3593 #ifdef CONFIG_PREEMPT_BKL
3594 saved_lock_depth
= task
->lock_depth
;
3595 task
->lock_depth
= -1;
3599 local_irq_disable();
3600 #ifdef CONFIG_PREEMPT_BKL
3601 task
->lock_depth
= saved_lock_depth
;
3603 sub_preempt_count(PREEMPT_ACTIVE
);
3605 /* we could miss a preemption opportunity between schedule and now */
3607 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3611 #endif /* CONFIG_PREEMPT */
3613 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3616 return try_to_wake_up(curr
->private, mode
, sync
);
3618 EXPORT_SYMBOL(default_wake_function
);
3621 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3622 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3623 * number) then we wake all the non-exclusive tasks and one exclusive task.
3625 * There are circumstances in which we can try to wake a task which has already
3626 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3627 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3629 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3630 int nr_exclusive
, int sync
, void *key
)
3632 wait_queue_t
*curr
, *next
;
3634 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3635 unsigned flags
= curr
->flags
;
3637 if (curr
->func(curr
, mode
, sync
, key
) &&
3638 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3644 * __wake_up - wake up threads blocked on a waitqueue.
3646 * @mode: which threads
3647 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3648 * @key: is directly passed to the wakeup function
3650 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3651 int nr_exclusive
, void *key
)
3653 unsigned long flags
;
3655 spin_lock_irqsave(&q
->lock
, flags
);
3656 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3657 spin_unlock_irqrestore(&q
->lock
, flags
);
3659 EXPORT_SYMBOL(__wake_up
);
3662 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3664 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3666 __wake_up_common(q
, mode
, 1, 0, NULL
);
3670 * __wake_up_sync - wake up threads blocked on a waitqueue.
3672 * @mode: which threads
3673 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3675 * The sync wakeup differs that the waker knows that it will schedule
3676 * away soon, so while the target thread will be woken up, it will not
3677 * be migrated to another CPU - ie. the two threads are 'synchronized'
3678 * with each other. This can prevent needless bouncing between CPUs.
3680 * On UP it can prevent extra preemption.
3683 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3685 unsigned long flags
;
3691 if (unlikely(!nr_exclusive
))
3694 spin_lock_irqsave(&q
->lock
, flags
);
3695 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3696 spin_unlock_irqrestore(&q
->lock
, flags
);
3698 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3700 void fastcall
complete(struct completion
*x
)
3702 unsigned long flags
;
3704 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3706 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3708 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3710 EXPORT_SYMBOL(complete
);
3712 void fastcall
complete_all(struct completion
*x
)
3714 unsigned long flags
;
3716 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3717 x
->done
+= UINT_MAX
/2;
3718 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3720 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3722 EXPORT_SYMBOL(complete_all
);
3724 void fastcall __sched
wait_for_completion(struct completion
*x
)
3728 spin_lock_irq(&x
->wait
.lock
);
3730 DECLARE_WAITQUEUE(wait
, current
);
3732 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3733 __add_wait_queue_tail(&x
->wait
, &wait
);
3735 __set_current_state(TASK_UNINTERRUPTIBLE
);
3736 spin_unlock_irq(&x
->wait
.lock
);
3738 spin_lock_irq(&x
->wait
.lock
);
3740 __remove_wait_queue(&x
->wait
, &wait
);
3743 spin_unlock_irq(&x
->wait
.lock
);
3745 EXPORT_SYMBOL(wait_for_completion
);
3747 unsigned long fastcall __sched
3748 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3752 spin_lock_irq(&x
->wait
.lock
);
3754 DECLARE_WAITQUEUE(wait
, current
);
3756 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3757 __add_wait_queue_tail(&x
->wait
, &wait
);
3759 __set_current_state(TASK_UNINTERRUPTIBLE
);
3760 spin_unlock_irq(&x
->wait
.lock
);
3761 timeout
= schedule_timeout(timeout
);
3762 spin_lock_irq(&x
->wait
.lock
);
3764 __remove_wait_queue(&x
->wait
, &wait
);
3768 __remove_wait_queue(&x
->wait
, &wait
);
3772 spin_unlock_irq(&x
->wait
.lock
);
3775 EXPORT_SYMBOL(wait_for_completion_timeout
);
3777 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3783 spin_lock_irq(&x
->wait
.lock
);
3785 DECLARE_WAITQUEUE(wait
, current
);
3787 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3788 __add_wait_queue_tail(&x
->wait
, &wait
);
3790 if (signal_pending(current
)) {
3792 __remove_wait_queue(&x
->wait
, &wait
);
3795 __set_current_state(TASK_INTERRUPTIBLE
);
3796 spin_unlock_irq(&x
->wait
.lock
);
3798 spin_lock_irq(&x
->wait
.lock
);
3800 __remove_wait_queue(&x
->wait
, &wait
);
3804 spin_unlock_irq(&x
->wait
.lock
);
3808 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3810 unsigned long fastcall __sched
3811 wait_for_completion_interruptible_timeout(struct completion
*x
,
3812 unsigned long timeout
)
3816 spin_lock_irq(&x
->wait
.lock
);
3818 DECLARE_WAITQUEUE(wait
, current
);
3820 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3821 __add_wait_queue_tail(&x
->wait
, &wait
);
3823 if (signal_pending(current
)) {
3824 timeout
= -ERESTARTSYS
;
3825 __remove_wait_queue(&x
->wait
, &wait
);
3828 __set_current_state(TASK_INTERRUPTIBLE
);
3829 spin_unlock_irq(&x
->wait
.lock
);
3830 timeout
= schedule_timeout(timeout
);
3831 spin_lock_irq(&x
->wait
.lock
);
3833 __remove_wait_queue(&x
->wait
, &wait
);
3837 __remove_wait_queue(&x
->wait
, &wait
);
3841 spin_unlock_irq(&x
->wait
.lock
);
3844 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3847 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3849 spin_lock_irqsave(&q
->lock
, *flags
);
3850 __add_wait_queue(q
, wait
);
3851 spin_unlock(&q
->lock
);
3855 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3857 spin_lock_irq(&q
->lock
);
3858 __remove_wait_queue(q
, wait
);
3859 spin_unlock_irqrestore(&q
->lock
, *flags
);
3862 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3864 unsigned long flags
;
3867 init_waitqueue_entry(&wait
, current
);
3869 current
->state
= TASK_INTERRUPTIBLE
;
3871 sleep_on_head(q
, &wait
, &flags
);
3873 sleep_on_tail(q
, &wait
, &flags
);
3875 EXPORT_SYMBOL(interruptible_sleep_on
);
3878 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3880 unsigned long flags
;
3883 init_waitqueue_entry(&wait
, current
);
3885 current
->state
= TASK_INTERRUPTIBLE
;
3887 sleep_on_head(q
, &wait
, &flags
);
3888 timeout
= schedule_timeout(timeout
);
3889 sleep_on_tail(q
, &wait
, &flags
);
3893 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3895 void __sched
sleep_on(wait_queue_head_t
*q
)
3897 unsigned long flags
;
3900 init_waitqueue_entry(&wait
, current
);
3902 current
->state
= TASK_UNINTERRUPTIBLE
;
3904 sleep_on_head(q
, &wait
, &flags
);
3906 sleep_on_tail(q
, &wait
, &flags
);
3908 EXPORT_SYMBOL(sleep_on
);
3910 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3912 unsigned long flags
;
3915 init_waitqueue_entry(&wait
, current
);
3917 current
->state
= TASK_UNINTERRUPTIBLE
;
3919 sleep_on_head(q
, &wait
, &flags
);
3920 timeout
= schedule_timeout(timeout
);
3921 sleep_on_tail(q
, &wait
, &flags
);
3925 EXPORT_SYMBOL(sleep_on_timeout
);
3927 #ifdef CONFIG_RT_MUTEXES
3930 * rt_mutex_setprio - set the current priority of a task
3932 * @prio: prio value (kernel-internal form)
3934 * This function changes the 'effective' priority of a task. It does
3935 * not touch ->normal_prio like __setscheduler().
3937 * Used by the rt_mutex code to implement priority inheritance logic.
3939 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3941 unsigned long flags
;
3945 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3947 rq
= task_rq_lock(p
, &flags
);
3948 update_rq_clock(rq
);
3951 on_rq
= p
->se
.on_rq
;
3953 dequeue_task(rq
, p
, 0);
3956 p
->sched_class
= &rt_sched_class
;
3958 p
->sched_class
= &fair_sched_class
;
3963 enqueue_task(rq
, p
, 0);
3965 * Reschedule if we are currently running on this runqueue and
3966 * our priority decreased, or if we are not currently running on
3967 * this runqueue and our priority is higher than the current's
3969 if (task_running(rq
, p
)) {
3970 if (p
->prio
> oldprio
)
3971 resched_task(rq
->curr
);
3973 check_preempt_curr(rq
, p
);
3976 task_rq_unlock(rq
, &flags
);
3981 void set_user_nice(struct task_struct
*p
, long nice
)
3983 int old_prio
, delta
, on_rq
;
3984 unsigned long flags
;
3987 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3990 * We have to be careful, if called from sys_setpriority(),
3991 * the task might be in the middle of scheduling on another CPU.
3993 rq
= task_rq_lock(p
, &flags
);
3994 update_rq_clock(rq
);
3996 * The RT priorities are set via sched_setscheduler(), but we still
3997 * allow the 'normal' nice value to be set - but as expected
3998 * it wont have any effect on scheduling until the task is
3999 * SCHED_FIFO/SCHED_RR:
4001 if (task_has_rt_policy(p
)) {
4002 p
->static_prio
= NICE_TO_PRIO(nice
);
4005 on_rq
= p
->se
.on_rq
;
4007 dequeue_task(rq
, p
, 0);
4011 p
->static_prio
= NICE_TO_PRIO(nice
);
4014 p
->prio
= effective_prio(p
);
4015 delta
= p
->prio
- old_prio
;
4018 enqueue_task(rq
, p
, 0);
4021 * If the task increased its priority or is running and
4022 * lowered its priority, then reschedule its CPU:
4024 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4025 resched_task(rq
->curr
);
4028 task_rq_unlock(rq
, &flags
);
4030 EXPORT_SYMBOL(set_user_nice
);
4033 * can_nice - check if a task can reduce its nice value
4037 int can_nice(const struct task_struct
*p
, const int nice
)
4039 /* convert nice value [19,-20] to rlimit style value [1,40] */
4040 int nice_rlim
= 20 - nice
;
4042 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4043 capable(CAP_SYS_NICE
));
4046 #ifdef __ARCH_WANT_SYS_NICE
4049 * sys_nice - change the priority of the current process.
4050 * @increment: priority increment
4052 * sys_setpriority is a more generic, but much slower function that
4053 * does similar things.
4055 asmlinkage
long sys_nice(int increment
)
4060 * Setpriority might change our priority at the same moment.
4061 * We don't have to worry. Conceptually one call occurs first
4062 * and we have a single winner.
4064 if (increment
< -40)
4069 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4075 if (increment
< 0 && !can_nice(current
, nice
))
4078 retval
= security_task_setnice(current
, nice
);
4082 set_user_nice(current
, nice
);
4089 * task_prio - return the priority value of a given task.
4090 * @p: the task in question.
4092 * This is the priority value as seen by users in /proc.
4093 * RT tasks are offset by -200. Normal tasks are centered
4094 * around 0, value goes from -16 to +15.
4096 int task_prio(const struct task_struct
*p
)
4098 return p
->prio
- MAX_RT_PRIO
;
4102 * task_nice - return the nice value of a given task.
4103 * @p: the task in question.
4105 int task_nice(const struct task_struct
*p
)
4107 return TASK_NICE(p
);
4109 EXPORT_SYMBOL_GPL(task_nice
);
4112 * idle_cpu - is a given cpu idle currently?
4113 * @cpu: the processor in question.
4115 int idle_cpu(int cpu
)
4117 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4121 * idle_task - return the idle task for a given cpu.
4122 * @cpu: the processor in question.
4124 struct task_struct
*idle_task(int cpu
)
4126 return cpu_rq(cpu
)->idle
;
4130 * find_process_by_pid - find a process with a matching PID value.
4131 * @pid: the pid in question.
4133 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4135 return pid
? find_task_by_pid(pid
) : current
;
4138 /* Actually do priority change: must hold rq lock. */
4140 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4142 BUG_ON(p
->se
.on_rq
);
4145 switch (p
->policy
) {
4149 p
->sched_class
= &fair_sched_class
;
4153 p
->sched_class
= &rt_sched_class
;
4157 p
->rt_priority
= prio
;
4158 p
->normal_prio
= normal_prio(p
);
4159 /* we are holding p->pi_lock already */
4160 p
->prio
= rt_mutex_getprio(p
);
4165 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4166 * @p: the task in question.
4167 * @policy: new policy.
4168 * @param: structure containing the new RT priority.
4170 * NOTE that the task may be already dead.
4172 int sched_setscheduler(struct task_struct
*p
, int policy
,
4173 struct sched_param
*param
)
4175 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4176 unsigned long flags
;
4179 /* may grab non-irq protected spin_locks */
4180 BUG_ON(in_interrupt());
4182 /* double check policy once rq lock held */
4184 policy
= oldpolicy
= p
->policy
;
4185 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4186 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4187 policy
!= SCHED_IDLE
)
4190 * Valid priorities for SCHED_FIFO and SCHED_RR are
4191 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4192 * SCHED_BATCH and SCHED_IDLE is 0.
4194 if (param
->sched_priority
< 0 ||
4195 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4196 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4198 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4202 * Allow unprivileged RT tasks to decrease priority:
4204 if (!capable(CAP_SYS_NICE
)) {
4205 if (rt_policy(policy
)) {
4206 unsigned long rlim_rtprio
;
4208 if (!lock_task_sighand(p
, &flags
))
4210 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4211 unlock_task_sighand(p
, &flags
);
4213 /* can't set/change the rt policy */
4214 if (policy
!= p
->policy
&& !rlim_rtprio
)
4217 /* can't increase priority */
4218 if (param
->sched_priority
> p
->rt_priority
&&
4219 param
->sched_priority
> rlim_rtprio
)
4223 * Like positive nice levels, dont allow tasks to
4224 * move out of SCHED_IDLE either:
4226 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4229 /* can't change other user's priorities */
4230 if ((current
->euid
!= p
->euid
) &&
4231 (current
->euid
!= p
->uid
))
4235 retval
= security_task_setscheduler(p
, policy
, param
);
4239 * make sure no PI-waiters arrive (or leave) while we are
4240 * changing the priority of the task:
4242 spin_lock_irqsave(&p
->pi_lock
, flags
);
4244 * To be able to change p->policy safely, the apropriate
4245 * runqueue lock must be held.
4247 rq
= __task_rq_lock(p
);
4248 /* recheck policy now with rq lock held */
4249 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4250 policy
= oldpolicy
= -1;
4251 __task_rq_unlock(rq
);
4252 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4255 update_rq_clock(rq
);
4256 on_rq
= p
->se
.on_rq
;
4258 deactivate_task(rq
, p
, 0);
4260 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4262 activate_task(rq
, p
, 0);
4264 * Reschedule if we are currently running on this runqueue and
4265 * our priority decreased, or if we are not currently running on
4266 * this runqueue and our priority is higher than the current's
4268 if (task_running(rq
, p
)) {
4269 if (p
->prio
> oldprio
)
4270 resched_task(rq
->curr
);
4272 check_preempt_curr(rq
, p
);
4275 __task_rq_unlock(rq
);
4276 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4278 rt_mutex_adjust_pi(p
);
4282 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4285 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4287 struct sched_param lparam
;
4288 struct task_struct
*p
;
4291 if (!param
|| pid
< 0)
4293 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4298 p
= find_process_by_pid(pid
);
4300 retval
= sched_setscheduler(p
, policy
, &lparam
);
4307 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4308 * @pid: the pid in question.
4309 * @policy: new policy.
4310 * @param: structure containing the new RT priority.
4312 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4313 struct sched_param __user
*param
)
4315 /* negative values for policy are not valid */
4319 return do_sched_setscheduler(pid
, policy
, param
);
4323 * sys_sched_setparam - set/change the RT priority of a thread
4324 * @pid: the pid in question.
4325 * @param: structure containing the new RT priority.
4327 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4329 return do_sched_setscheduler(pid
, -1, param
);
4333 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4334 * @pid: the pid in question.
4336 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4338 struct task_struct
*p
;
4339 int retval
= -EINVAL
;
4345 read_lock(&tasklist_lock
);
4346 p
= find_process_by_pid(pid
);
4348 retval
= security_task_getscheduler(p
);
4352 read_unlock(&tasklist_lock
);
4359 * sys_sched_getscheduler - get the RT priority of a thread
4360 * @pid: the pid in question.
4361 * @param: structure containing the RT priority.
4363 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4365 struct sched_param lp
;
4366 struct task_struct
*p
;
4367 int retval
= -EINVAL
;
4369 if (!param
|| pid
< 0)
4372 read_lock(&tasklist_lock
);
4373 p
= find_process_by_pid(pid
);
4378 retval
= security_task_getscheduler(p
);
4382 lp
.sched_priority
= p
->rt_priority
;
4383 read_unlock(&tasklist_lock
);
4386 * This one might sleep, we cannot do it with a spinlock held ...
4388 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4394 read_unlock(&tasklist_lock
);
4398 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4400 cpumask_t cpus_allowed
;
4401 struct task_struct
*p
;
4404 mutex_lock(&sched_hotcpu_mutex
);
4405 read_lock(&tasklist_lock
);
4407 p
= find_process_by_pid(pid
);
4409 read_unlock(&tasklist_lock
);
4410 mutex_unlock(&sched_hotcpu_mutex
);
4415 * It is not safe to call set_cpus_allowed with the
4416 * tasklist_lock held. We will bump the task_struct's
4417 * usage count and then drop tasklist_lock.
4420 read_unlock(&tasklist_lock
);
4423 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4424 !capable(CAP_SYS_NICE
))
4427 retval
= security_task_setscheduler(p
, 0, NULL
);
4431 cpus_allowed
= cpuset_cpus_allowed(p
);
4432 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4433 retval
= set_cpus_allowed(p
, new_mask
);
4437 mutex_unlock(&sched_hotcpu_mutex
);
4441 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4442 cpumask_t
*new_mask
)
4444 if (len
< sizeof(cpumask_t
)) {
4445 memset(new_mask
, 0, sizeof(cpumask_t
));
4446 } else if (len
> sizeof(cpumask_t
)) {
4447 len
= sizeof(cpumask_t
);
4449 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4453 * sys_sched_setaffinity - set the cpu affinity of a process
4454 * @pid: pid of the process
4455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4456 * @user_mask_ptr: user-space pointer to the new cpu mask
4458 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4459 unsigned long __user
*user_mask_ptr
)
4464 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4468 return sched_setaffinity(pid
, new_mask
);
4472 * Represents all cpu's present in the system
4473 * In systems capable of hotplug, this map could dynamically grow
4474 * as new cpu's are detected in the system via any platform specific
4475 * method, such as ACPI for e.g.
4478 cpumask_t cpu_present_map __read_mostly
;
4479 EXPORT_SYMBOL(cpu_present_map
);
4482 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4483 EXPORT_SYMBOL(cpu_online_map
);
4485 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4486 EXPORT_SYMBOL(cpu_possible_map
);
4489 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4491 struct task_struct
*p
;
4494 mutex_lock(&sched_hotcpu_mutex
);
4495 read_lock(&tasklist_lock
);
4498 p
= find_process_by_pid(pid
);
4502 retval
= security_task_getscheduler(p
);
4506 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4509 read_unlock(&tasklist_lock
);
4510 mutex_unlock(&sched_hotcpu_mutex
);
4516 * sys_sched_getaffinity - get the cpu affinity of a process
4517 * @pid: pid of the process
4518 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4519 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4521 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4522 unsigned long __user
*user_mask_ptr
)
4527 if (len
< sizeof(cpumask_t
))
4530 ret
= sched_getaffinity(pid
, &mask
);
4534 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4537 return sizeof(cpumask_t
);
4541 * sys_sched_yield - yield the current processor to other threads.
4543 * This function yields the current CPU to other tasks. If there are no
4544 * other threads running on this CPU then this function will return.
4546 asmlinkage
long sys_sched_yield(void)
4548 struct rq
*rq
= this_rq_lock();
4550 schedstat_inc(rq
, yld_cnt
);
4551 current
->sched_class
->yield_task(rq
, current
);
4554 * Since we are going to call schedule() anyway, there's
4555 * no need to preempt or enable interrupts:
4557 __release(rq
->lock
);
4558 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4559 _raw_spin_unlock(&rq
->lock
);
4560 preempt_enable_no_resched();
4567 static void __cond_resched(void)
4569 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4570 __might_sleep(__FILE__
, __LINE__
);
4573 * The BKS might be reacquired before we have dropped
4574 * PREEMPT_ACTIVE, which could trigger a second
4575 * cond_resched() call.
4578 add_preempt_count(PREEMPT_ACTIVE
);
4580 sub_preempt_count(PREEMPT_ACTIVE
);
4581 } while (need_resched());
4584 int __sched
cond_resched(void)
4586 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4587 system_state
== SYSTEM_RUNNING
) {
4593 EXPORT_SYMBOL(cond_resched
);
4596 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4597 * call schedule, and on return reacquire the lock.
4599 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4600 * operations here to prevent schedule() from being called twice (once via
4601 * spin_unlock(), once by hand).
4603 int cond_resched_lock(spinlock_t
*lock
)
4607 if (need_lockbreak(lock
)) {
4613 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4614 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4615 _raw_spin_unlock(lock
);
4616 preempt_enable_no_resched();
4623 EXPORT_SYMBOL(cond_resched_lock
);
4625 int __sched
cond_resched_softirq(void)
4627 BUG_ON(!in_softirq());
4629 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4637 EXPORT_SYMBOL(cond_resched_softirq
);
4640 * yield - yield the current processor to other threads.
4642 * This is a shortcut for kernel-space yielding - it marks the
4643 * thread runnable and calls sys_sched_yield().
4645 void __sched
yield(void)
4647 set_current_state(TASK_RUNNING
);
4650 EXPORT_SYMBOL(yield
);
4653 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4654 * that process accounting knows that this is a task in IO wait state.
4656 * But don't do that if it is a deliberate, throttling IO wait (this task
4657 * has set its backing_dev_info: the queue against which it should throttle)
4659 void __sched
io_schedule(void)
4661 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4663 delayacct_blkio_start();
4664 atomic_inc(&rq
->nr_iowait
);
4666 atomic_dec(&rq
->nr_iowait
);
4667 delayacct_blkio_end();
4669 EXPORT_SYMBOL(io_schedule
);
4671 long __sched
io_schedule_timeout(long timeout
)
4673 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4676 delayacct_blkio_start();
4677 atomic_inc(&rq
->nr_iowait
);
4678 ret
= schedule_timeout(timeout
);
4679 atomic_dec(&rq
->nr_iowait
);
4680 delayacct_blkio_end();
4685 * sys_sched_get_priority_max - return maximum RT priority.
4686 * @policy: scheduling class.
4688 * this syscall returns the maximum rt_priority that can be used
4689 * by a given scheduling class.
4691 asmlinkage
long sys_sched_get_priority_max(int policy
)
4698 ret
= MAX_USER_RT_PRIO
-1;
4710 * sys_sched_get_priority_min - return minimum RT priority.
4711 * @policy: scheduling class.
4713 * this syscall returns the minimum rt_priority that can be used
4714 * by a given scheduling class.
4716 asmlinkage
long sys_sched_get_priority_min(int policy
)
4734 * sys_sched_rr_get_interval - return the default timeslice of a process.
4735 * @pid: pid of the process.
4736 * @interval: userspace pointer to the timeslice value.
4738 * this syscall writes the default timeslice value of a given process
4739 * into the user-space timespec buffer. A value of '0' means infinity.
4742 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4744 struct task_struct
*p
;
4745 int retval
= -EINVAL
;
4752 read_lock(&tasklist_lock
);
4753 p
= find_process_by_pid(pid
);
4757 retval
= security_task_getscheduler(p
);
4761 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4762 0 : static_prio_timeslice(p
->static_prio
), &t
);
4763 read_unlock(&tasklist_lock
);
4764 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4768 read_unlock(&tasklist_lock
);
4772 static const char stat_nam
[] = "RSDTtZX";
4774 static void show_task(struct task_struct
*p
)
4776 unsigned long free
= 0;
4779 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4780 printk("%-13.13s %c", p
->comm
,
4781 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4782 #if BITS_PER_LONG == 32
4783 if (state
== TASK_RUNNING
)
4784 printk(" running ");
4786 printk(" %08lx ", thread_saved_pc(p
));
4788 if (state
== TASK_RUNNING
)
4789 printk(" running task ");
4791 printk(" %016lx ", thread_saved_pc(p
));
4793 #ifdef CONFIG_DEBUG_STACK_USAGE
4795 unsigned long *n
= end_of_stack(p
);
4798 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4801 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4803 if (state
!= TASK_RUNNING
)
4804 show_stack(p
, NULL
);
4807 void show_state_filter(unsigned long state_filter
)
4809 struct task_struct
*g
, *p
;
4811 #if BITS_PER_LONG == 32
4813 " task PC stack pid father\n");
4816 " task PC stack pid father\n");
4818 read_lock(&tasklist_lock
);
4819 do_each_thread(g
, p
) {
4821 * reset the NMI-timeout, listing all files on a slow
4822 * console might take alot of time:
4824 touch_nmi_watchdog();
4825 if (!state_filter
|| (p
->state
& state_filter
))
4827 } while_each_thread(g
, p
);
4829 touch_all_softlockup_watchdogs();
4831 #ifdef CONFIG_SCHED_DEBUG
4832 sysrq_sched_debug_show();
4834 read_unlock(&tasklist_lock
);
4836 * Only show locks if all tasks are dumped:
4838 if (state_filter
== -1)
4839 debug_show_all_locks();
4842 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4844 idle
->sched_class
= &idle_sched_class
;
4848 * init_idle - set up an idle thread for a given CPU
4849 * @idle: task in question
4850 * @cpu: cpu the idle task belongs to
4852 * NOTE: this function does not set the idle thread's NEED_RESCHED
4853 * flag, to make booting more robust.
4855 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4857 struct rq
*rq
= cpu_rq(cpu
);
4858 unsigned long flags
;
4861 idle
->se
.exec_start
= sched_clock();
4863 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4864 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4865 __set_task_cpu(idle
, cpu
);
4867 spin_lock_irqsave(&rq
->lock
, flags
);
4868 rq
->curr
= rq
->idle
= idle
;
4869 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4872 spin_unlock_irqrestore(&rq
->lock
, flags
);
4874 /* Set the preempt count _outside_ the spinlocks! */
4875 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4876 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4878 task_thread_info(idle
)->preempt_count
= 0;
4881 * The idle tasks have their own, simple scheduling class:
4883 idle
->sched_class
= &idle_sched_class
;
4887 * In a system that switches off the HZ timer nohz_cpu_mask
4888 * indicates which cpus entered this state. This is used
4889 * in the rcu update to wait only for active cpus. For system
4890 * which do not switch off the HZ timer nohz_cpu_mask should
4891 * always be CPU_MASK_NONE.
4893 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4897 * This is how migration works:
4899 * 1) we queue a struct migration_req structure in the source CPU's
4900 * runqueue and wake up that CPU's migration thread.
4901 * 2) we down() the locked semaphore => thread blocks.
4902 * 3) migration thread wakes up (implicitly it forces the migrated
4903 * thread off the CPU)
4904 * 4) it gets the migration request and checks whether the migrated
4905 * task is still in the wrong runqueue.
4906 * 5) if it's in the wrong runqueue then the migration thread removes
4907 * it and puts it into the right queue.
4908 * 6) migration thread up()s the semaphore.
4909 * 7) we wake up and the migration is done.
4913 * Change a given task's CPU affinity. Migrate the thread to a
4914 * proper CPU and schedule it away if the CPU it's executing on
4915 * is removed from the allowed bitmask.
4917 * NOTE: the caller must have a valid reference to the task, the
4918 * task must not exit() & deallocate itself prematurely. The
4919 * call is not atomic; no spinlocks may be held.
4921 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4923 struct migration_req req
;
4924 unsigned long flags
;
4928 rq
= task_rq_lock(p
, &flags
);
4929 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4934 p
->cpus_allowed
= new_mask
;
4935 /* Can the task run on the task's current CPU? If so, we're done */
4936 if (cpu_isset(task_cpu(p
), new_mask
))
4939 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4940 /* Need help from migration thread: drop lock and wait. */
4941 task_rq_unlock(rq
, &flags
);
4942 wake_up_process(rq
->migration_thread
);
4943 wait_for_completion(&req
.done
);
4944 tlb_migrate_finish(p
->mm
);
4948 task_rq_unlock(rq
, &flags
);
4952 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4955 * Move (not current) task off this cpu, onto dest cpu. We're doing
4956 * this because either it can't run here any more (set_cpus_allowed()
4957 * away from this CPU, or CPU going down), or because we're
4958 * attempting to rebalance this task on exec (sched_exec).
4960 * So we race with normal scheduler movements, but that's OK, as long
4961 * as the task is no longer on this CPU.
4963 * Returns non-zero if task was successfully migrated.
4965 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4967 struct rq
*rq_dest
, *rq_src
;
4970 if (unlikely(cpu_is_offline(dest_cpu
)))
4973 rq_src
= cpu_rq(src_cpu
);
4974 rq_dest
= cpu_rq(dest_cpu
);
4976 double_rq_lock(rq_src
, rq_dest
);
4977 /* Already moved. */
4978 if (task_cpu(p
) != src_cpu
)
4980 /* Affinity changed (again). */
4981 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4984 on_rq
= p
->se
.on_rq
;
4986 deactivate_task(rq_src
, p
, 0);
4988 set_task_cpu(p
, dest_cpu
);
4990 activate_task(rq_dest
, p
, 0);
4991 check_preempt_curr(rq_dest
, p
);
4995 double_rq_unlock(rq_src
, rq_dest
);
5000 * migration_thread - this is a highprio system thread that performs
5001 * thread migration by bumping thread off CPU then 'pushing' onto
5004 static int migration_thread(void *data
)
5006 int cpu
= (long)data
;
5010 BUG_ON(rq
->migration_thread
!= current
);
5012 set_current_state(TASK_INTERRUPTIBLE
);
5013 while (!kthread_should_stop()) {
5014 struct migration_req
*req
;
5015 struct list_head
*head
;
5017 spin_lock_irq(&rq
->lock
);
5019 if (cpu_is_offline(cpu
)) {
5020 spin_unlock_irq(&rq
->lock
);
5024 if (rq
->active_balance
) {
5025 active_load_balance(rq
, cpu
);
5026 rq
->active_balance
= 0;
5029 head
= &rq
->migration_queue
;
5031 if (list_empty(head
)) {
5032 spin_unlock_irq(&rq
->lock
);
5034 set_current_state(TASK_INTERRUPTIBLE
);
5037 req
= list_entry(head
->next
, struct migration_req
, list
);
5038 list_del_init(head
->next
);
5040 spin_unlock(&rq
->lock
);
5041 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5044 complete(&req
->done
);
5046 __set_current_state(TASK_RUNNING
);
5050 /* Wait for kthread_stop */
5051 set_current_state(TASK_INTERRUPTIBLE
);
5052 while (!kthread_should_stop()) {
5054 set_current_state(TASK_INTERRUPTIBLE
);
5056 __set_current_state(TASK_RUNNING
);
5060 #ifdef CONFIG_HOTPLUG_CPU
5062 * Figure out where task on dead CPU should go, use force if neccessary.
5063 * NOTE: interrupts should be disabled by the caller
5065 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5067 unsigned long flags
;
5074 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5075 cpus_and(mask
, mask
, p
->cpus_allowed
);
5076 dest_cpu
= any_online_cpu(mask
);
5078 /* On any allowed CPU? */
5079 if (dest_cpu
== NR_CPUS
)
5080 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5082 /* No more Mr. Nice Guy. */
5083 if (dest_cpu
== NR_CPUS
) {
5084 rq
= task_rq_lock(p
, &flags
);
5085 cpus_setall(p
->cpus_allowed
);
5086 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5087 task_rq_unlock(rq
, &flags
);
5090 * Don't tell them about moving exiting tasks or
5091 * kernel threads (both mm NULL), since they never
5094 if (p
->mm
&& printk_ratelimit())
5095 printk(KERN_INFO
"process %d (%s) no "
5096 "longer affine to cpu%d\n",
5097 p
->pid
, p
->comm
, dead_cpu
);
5099 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5104 * While a dead CPU has no uninterruptible tasks queued at this point,
5105 * it might still have a nonzero ->nr_uninterruptible counter, because
5106 * for performance reasons the counter is not stricly tracking tasks to
5107 * their home CPUs. So we just add the counter to another CPU's counter,
5108 * to keep the global sum constant after CPU-down:
5110 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5112 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5113 unsigned long flags
;
5115 local_irq_save(flags
);
5116 double_rq_lock(rq_src
, rq_dest
);
5117 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5118 rq_src
->nr_uninterruptible
= 0;
5119 double_rq_unlock(rq_src
, rq_dest
);
5120 local_irq_restore(flags
);
5123 /* Run through task list and migrate tasks from the dead cpu. */
5124 static void migrate_live_tasks(int src_cpu
)
5126 struct task_struct
*p
, *t
;
5128 write_lock_irq(&tasklist_lock
);
5130 do_each_thread(t
, p
) {
5134 if (task_cpu(p
) == src_cpu
)
5135 move_task_off_dead_cpu(src_cpu
, p
);
5136 } while_each_thread(t
, p
);
5138 write_unlock_irq(&tasklist_lock
);
5142 * Schedules idle task to be the next runnable task on current CPU.
5143 * It does so by boosting its priority to highest possible and adding it to
5144 * the _front_ of the runqueue. Used by CPU offline code.
5146 void sched_idle_next(void)
5148 int this_cpu
= smp_processor_id();
5149 struct rq
*rq
= cpu_rq(this_cpu
);
5150 struct task_struct
*p
= rq
->idle
;
5151 unsigned long flags
;
5153 /* cpu has to be offline */
5154 BUG_ON(cpu_online(this_cpu
));
5157 * Strictly not necessary since rest of the CPUs are stopped by now
5158 * and interrupts disabled on the current cpu.
5160 spin_lock_irqsave(&rq
->lock
, flags
);
5162 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5164 /* Add idle task to the _front_ of its priority queue: */
5165 activate_idle_task(p
, rq
);
5167 spin_unlock_irqrestore(&rq
->lock
, flags
);
5171 * Ensures that the idle task is using init_mm right before its cpu goes
5174 void idle_task_exit(void)
5176 struct mm_struct
*mm
= current
->active_mm
;
5178 BUG_ON(cpu_online(smp_processor_id()));
5181 switch_mm(mm
, &init_mm
, current
);
5185 /* called under rq->lock with disabled interrupts */
5186 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5188 struct rq
*rq
= cpu_rq(dead_cpu
);
5190 /* Must be exiting, otherwise would be on tasklist. */
5191 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5193 /* Cannot have done final schedule yet: would have vanished. */
5194 BUG_ON(p
->state
== TASK_DEAD
);
5199 * Drop lock around migration; if someone else moves it,
5200 * that's OK. No task can be added to this CPU, so iteration is
5202 * NOTE: interrupts should be left disabled --dev@
5204 spin_unlock(&rq
->lock
);
5205 move_task_off_dead_cpu(dead_cpu
, p
);
5206 spin_lock(&rq
->lock
);
5211 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5212 static void migrate_dead_tasks(unsigned int dead_cpu
)
5214 struct rq
*rq
= cpu_rq(dead_cpu
);
5215 struct task_struct
*next
;
5218 if (!rq
->nr_running
)
5220 update_rq_clock(rq
);
5221 next
= pick_next_task(rq
, rq
->curr
);
5224 migrate_dead(dead_cpu
, next
);
5228 #endif /* CONFIG_HOTPLUG_CPU */
5230 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5232 static struct ctl_table sd_ctl_dir
[] = {
5234 .procname
= "sched_domain",
5240 static struct ctl_table sd_ctl_root
[] = {
5242 .ctl_name
= CTL_KERN
,
5243 .procname
= "kernel",
5245 .child
= sd_ctl_dir
,
5250 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5252 struct ctl_table
*entry
=
5253 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5256 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5262 set_table_entry(struct ctl_table
*entry
,
5263 const char *procname
, void *data
, int maxlen
,
5264 mode_t mode
, proc_handler
*proc_handler
)
5266 entry
->procname
= procname
;
5268 entry
->maxlen
= maxlen
;
5270 entry
->proc_handler
= proc_handler
;
5273 static struct ctl_table
*
5274 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5276 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5278 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5279 sizeof(long), 0644, proc_doulongvec_minmax
);
5280 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5281 sizeof(long), 0644, proc_doulongvec_minmax
);
5282 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5283 sizeof(int), 0644, proc_dointvec_minmax
);
5284 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5285 sizeof(int), 0644, proc_dointvec_minmax
);
5286 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5287 sizeof(int), 0644, proc_dointvec_minmax
);
5288 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5289 sizeof(int), 0644, proc_dointvec_minmax
);
5290 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5291 sizeof(int), 0644, proc_dointvec_minmax
);
5292 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5293 sizeof(int), 0644, proc_dointvec_minmax
);
5294 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5295 sizeof(int), 0644, proc_dointvec_minmax
);
5296 set_table_entry(&table
[10], "cache_nice_tries",
5297 &sd
->cache_nice_tries
,
5298 sizeof(int), 0644, proc_dointvec_minmax
);
5299 set_table_entry(&table
[12], "flags", &sd
->flags
,
5300 sizeof(int), 0644, proc_dointvec_minmax
);
5305 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5307 struct ctl_table
*entry
, *table
;
5308 struct sched_domain
*sd
;
5309 int domain_num
= 0, i
;
5312 for_each_domain(cpu
, sd
)
5314 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5317 for_each_domain(cpu
, sd
) {
5318 snprintf(buf
, 32, "domain%d", i
);
5319 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5321 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5328 static struct ctl_table_header
*sd_sysctl_header
;
5329 static void init_sched_domain_sysctl(void)
5331 int i
, cpu_num
= num_online_cpus();
5332 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5335 sd_ctl_dir
[0].child
= entry
;
5337 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5338 snprintf(buf
, 32, "cpu%d", i
);
5339 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5341 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5343 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5346 static void init_sched_domain_sysctl(void)
5352 * migration_call - callback that gets triggered when a CPU is added.
5353 * Here we can start up the necessary migration thread for the new CPU.
5355 static int __cpuinit
5356 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5358 struct task_struct
*p
;
5359 int cpu
= (long)hcpu
;
5360 unsigned long flags
;
5364 case CPU_LOCK_ACQUIRE
:
5365 mutex_lock(&sched_hotcpu_mutex
);
5368 case CPU_UP_PREPARE
:
5369 case CPU_UP_PREPARE_FROZEN
:
5370 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5373 kthread_bind(p
, cpu
);
5374 /* Must be high prio: stop_machine expects to yield to it. */
5375 rq
= task_rq_lock(p
, &flags
);
5376 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5377 task_rq_unlock(rq
, &flags
);
5378 cpu_rq(cpu
)->migration_thread
= p
;
5382 case CPU_ONLINE_FROZEN
:
5383 /* Strictly unneccessary, as first user will wake it. */
5384 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5387 #ifdef CONFIG_HOTPLUG_CPU
5388 case CPU_UP_CANCELED
:
5389 case CPU_UP_CANCELED_FROZEN
:
5390 if (!cpu_rq(cpu
)->migration_thread
)
5392 /* Unbind it from offline cpu so it can run. Fall thru. */
5393 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5394 any_online_cpu(cpu_online_map
));
5395 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5396 cpu_rq(cpu
)->migration_thread
= NULL
;
5400 case CPU_DEAD_FROZEN
:
5401 migrate_live_tasks(cpu
);
5403 kthread_stop(rq
->migration_thread
);
5404 rq
->migration_thread
= NULL
;
5405 /* Idle task back to normal (off runqueue, low prio) */
5406 rq
= task_rq_lock(rq
->idle
, &flags
);
5407 update_rq_clock(rq
);
5408 deactivate_task(rq
, rq
->idle
, 0);
5409 rq
->idle
->static_prio
= MAX_PRIO
;
5410 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5411 rq
->idle
->sched_class
= &idle_sched_class
;
5412 migrate_dead_tasks(cpu
);
5413 task_rq_unlock(rq
, &flags
);
5414 migrate_nr_uninterruptible(rq
);
5415 BUG_ON(rq
->nr_running
!= 0);
5417 /* No need to migrate the tasks: it was best-effort if
5418 * they didn't take sched_hotcpu_mutex. Just wake up
5419 * the requestors. */
5420 spin_lock_irq(&rq
->lock
);
5421 while (!list_empty(&rq
->migration_queue
)) {
5422 struct migration_req
*req
;
5424 req
= list_entry(rq
->migration_queue
.next
,
5425 struct migration_req
, list
);
5426 list_del_init(&req
->list
);
5427 complete(&req
->done
);
5429 spin_unlock_irq(&rq
->lock
);
5432 case CPU_LOCK_RELEASE
:
5433 mutex_unlock(&sched_hotcpu_mutex
);
5439 /* Register at highest priority so that task migration (migrate_all_tasks)
5440 * happens before everything else.
5442 static struct notifier_block __cpuinitdata migration_notifier
= {
5443 .notifier_call
= migration_call
,
5447 int __init
migration_init(void)
5449 void *cpu
= (void *)(long)smp_processor_id();
5452 /* Start one for the boot CPU: */
5453 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5454 BUG_ON(err
== NOTIFY_BAD
);
5455 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5456 register_cpu_notifier(&migration_notifier
);
5464 /* Number of possible processor ids */
5465 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5466 EXPORT_SYMBOL(nr_cpu_ids
);
5468 #undef SCHED_DOMAIN_DEBUG
5469 #ifdef SCHED_DOMAIN_DEBUG
5470 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5475 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5479 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5484 struct sched_group
*group
= sd
->groups
;
5485 cpumask_t groupmask
;
5487 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5488 cpus_clear(groupmask
);
5491 for (i
= 0; i
< level
+ 1; i
++)
5493 printk("domain %d: ", level
);
5495 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5496 printk("does not load-balance\n");
5498 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5503 printk("span %s\n", str
);
5505 if (!cpu_isset(cpu
, sd
->span
))
5506 printk(KERN_ERR
"ERROR: domain->span does not contain "
5508 if (!cpu_isset(cpu
, group
->cpumask
))
5509 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5513 for (i
= 0; i
< level
+ 2; i
++)
5519 printk(KERN_ERR
"ERROR: group is NULL\n");
5523 if (!group
->__cpu_power
) {
5525 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5529 if (!cpus_weight(group
->cpumask
)) {
5531 printk(KERN_ERR
"ERROR: empty group\n");
5534 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5536 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5539 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5541 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5544 group
= group
->next
;
5545 } while (group
!= sd
->groups
);
5548 if (!cpus_equal(sd
->span
, groupmask
))
5549 printk(KERN_ERR
"ERROR: groups don't span "
5557 if (!cpus_subset(groupmask
, sd
->span
))
5558 printk(KERN_ERR
"ERROR: parent span is not a superset "
5559 "of domain->span\n");
5564 # define sched_domain_debug(sd, cpu) do { } while (0)
5567 static int sd_degenerate(struct sched_domain
*sd
)
5569 if (cpus_weight(sd
->span
) == 1)
5572 /* Following flags need at least 2 groups */
5573 if (sd
->flags
& (SD_LOAD_BALANCE
|
5574 SD_BALANCE_NEWIDLE
|
5578 SD_SHARE_PKG_RESOURCES
)) {
5579 if (sd
->groups
!= sd
->groups
->next
)
5583 /* Following flags don't use groups */
5584 if (sd
->flags
& (SD_WAKE_IDLE
|
5593 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5595 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5597 if (sd_degenerate(parent
))
5600 if (!cpus_equal(sd
->span
, parent
->span
))
5603 /* Does parent contain flags not in child? */
5604 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5605 if (cflags
& SD_WAKE_AFFINE
)
5606 pflags
&= ~SD_WAKE_BALANCE
;
5607 /* Flags needing groups don't count if only 1 group in parent */
5608 if (parent
->groups
== parent
->groups
->next
) {
5609 pflags
&= ~(SD_LOAD_BALANCE
|
5610 SD_BALANCE_NEWIDLE
|
5614 SD_SHARE_PKG_RESOURCES
);
5616 if (~cflags
& pflags
)
5623 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5624 * hold the hotplug lock.
5626 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5628 struct rq
*rq
= cpu_rq(cpu
);
5629 struct sched_domain
*tmp
;
5631 /* Remove the sched domains which do not contribute to scheduling. */
5632 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5633 struct sched_domain
*parent
= tmp
->parent
;
5636 if (sd_parent_degenerate(tmp
, parent
)) {
5637 tmp
->parent
= parent
->parent
;
5639 parent
->parent
->child
= tmp
;
5643 if (sd
&& sd_degenerate(sd
)) {
5649 sched_domain_debug(sd
, cpu
);
5651 rcu_assign_pointer(rq
->sd
, sd
);
5654 /* cpus with isolated domains */
5655 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5657 /* Setup the mask of cpus configured for isolated domains */
5658 static int __init
isolated_cpu_setup(char *str
)
5660 int ints
[NR_CPUS
], i
;
5662 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5663 cpus_clear(cpu_isolated_map
);
5664 for (i
= 1; i
<= ints
[0]; i
++)
5665 if (ints
[i
] < NR_CPUS
)
5666 cpu_set(ints
[i
], cpu_isolated_map
);
5670 __setup ("isolcpus=", isolated_cpu_setup
);
5673 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5674 * to a function which identifies what group(along with sched group) a CPU
5675 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5676 * (due to the fact that we keep track of groups covered with a cpumask_t).
5678 * init_sched_build_groups will build a circular linked list of the groups
5679 * covered by the given span, and will set each group's ->cpumask correctly,
5680 * and ->cpu_power to 0.
5683 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5684 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5685 struct sched_group
**sg
))
5687 struct sched_group
*first
= NULL
, *last
= NULL
;
5688 cpumask_t covered
= CPU_MASK_NONE
;
5691 for_each_cpu_mask(i
, span
) {
5692 struct sched_group
*sg
;
5693 int group
= group_fn(i
, cpu_map
, &sg
);
5696 if (cpu_isset(i
, covered
))
5699 sg
->cpumask
= CPU_MASK_NONE
;
5700 sg
->__cpu_power
= 0;
5702 for_each_cpu_mask(j
, span
) {
5703 if (group_fn(j
, cpu_map
, NULL
) != group
)
5706 cpu_set(j
, covered
);
5707 cpu_set(j
, sg
->cpumask
);
5718 #define SD_NODES_PER_DOMAIN 16
5723 * find_next_best_node - find the next node to include in a sched_domain
5724 * @node: node whose sched_domain we're building
5725 * @used_nodes: nodes already in the sched_domain
5727 * Find the next node to include in a given scheduling domain. Simply
5728 * finds the closest node not already in the @used_nodes map.
5730 * Should use nodemask_t.
5732 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5734 int i
, n
, val
, min_val
, best_node
= 0;
5738 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5739 /* Start at @node */
5740 n
= (node
+ i
) % MAX_NUMNODES
;
5742 if (!nr_cpus_node(n
))
5745 /* Skip already used nodes */
5746 if (test_bit(n
, used_nodes
))
5749 /* Simple min distance search */
5750 val
= node_distance(node
, n
);
5752 if (val
< min_val
) {
5758 set_bit(best_node
, used_nodes
);
5763 * sched_domain_node_span - get a cpumask for a node's sched_domain
5764 * @node: node whose cpumask we're constructing
5765 * @size: number of nodes to include in this span
5767 * Given a node, construct a good cpumask for its sched_domain to span. It
5768 * should be one that prevents unnecessary balancing, but also spreads tasks
5771 static cpumask_t
sched_domain_node_span(int node
)
5773 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5774 cpumask_t span
, nodemask
;
5778 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5780 nodemask
= node_to_cpumask(node
);
5781 cpus_or(span
, span
, nodemask
);
5782 set_bit(node
, used_nodes
);
5784 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5785 int next_node
= find_next_best_node(node
, used_nodes
);
5787 nodemask
= node_to_cpumask(next_node
);
5788 cpus_or(span
, span
, nodemask
);
5795 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5798 * SMT sched-domains:
5800 #ifdef CONFIG_SCHED_SMT
5801 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5802 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5804 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5805 struct sched_group
**sg
)
5808 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5814 * multi-core sched-domains:
5816 #ifdef CONFIG_SCHED_MC
5817 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5818 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5821 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5822 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5823 struct sched_group
**sg
)
5826 cpumask_t mask
= cpu_sibling_map
[cpu
];
5827 cpus_and(mask
, mask
, *cpu_map
);
5828 group
= first_cpu(mask
);
5830 *sg
= &per_cpu(sched_group_core
, group
);
5833 #elif defined(CONFIG_SCHED_MC)
5834 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5835 struct sched_group
**sg
)
5838 *sg
= &per_cpu(sched_group_core
, cpu
);
5843 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5844 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5846 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5847 struct sched_group
**sg
)
5850 #ifdef CONFIG_SCHED_MC
5851 cpumask_t mask
= cpu_coregroup_map(cpu
);
5852 cpus_and(mask
, mask
, *cpu_map
);
5853 group
= first_cpu(mask
);
5854 #elif defined(CONFIG_SCHED_SMT)
5855 cpumask_t mask
= cpu_sibling_map
[cpu
];
5856 cpus_and(mask
, mask
, *cpu_map
);
5857 group
= first_cpu(mask
);
5862 *sg
= &per_cpu(sched_group_phys
, group
);
5868 * The init_sched_build_groups can't handle what we want to do with node
5869 * groups, so roll our own. Now each node has its own list of groups which
5870 * gets dynamically allocated.
5872 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5873 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5875 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5876 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5878 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5879 struct sched_group
**sg
)
5881 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5884 cpus_and(nodemask
, nodemask
, *cpu_map
);
5885 group
= first_cpu(nodemask
);
5888 *sg
= &per_cpu(sched_group_allnodes
, group
);
5892 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5894 struct sched_group
*sg
= group_head
;
5900 for_each_cpu_mask(j
, sg
->cpumask
) {
5901 struct sched_domain
*sd
;
5903 sd
= &per_cpu(phys_domains
, j
);
5904 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5906 * Only add "power" once for each
5912 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5915 if (sg
!= group_head
)
5921 /* Free memory allocated for various sched_group structures */
5922 static void free_sched_groups(const cpumask_t
*cpu_map
)
5926 for_each_cpu_mask(cpu
, *cpu_map
) {
5927 struct sched_group
**sched_group_nodes
5928 = sched_group_nodes_bycpu
[cpu
];
5930 if (!sched_group_nodes
)
5933 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5934 cpumask_t nodemask
= node_to_cpumask(i
);
5935 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5937 cpus_and(nodemask
, nodemask
, *cpu_map
);
5938 if (cpus_empty(nodemask
))
5948 if (oldsg
!= sched_group_nodes
[i
])
5951 kfree(sched_group_nodes
);
5952 sched_group_nodes_bycpu
[cpu
] = NULL
;
5956 static void free_sched_groups(const cpumask_t
*cpu_map
)
5962 * Initialize sched groups cpu_power.
5964 * cpu_power indicates the capacity of sched group, which is used while
5965 * distributing the load between different sched groups in a sched domain.
5966 * Typically cpu_power for all the groups in a sched domain will be same unless
5967 * there are asymmetries in the topology. If there are asymmetries, group
5968 * having more cpu_power will pickup more load compared to the group having
5971 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5972 * the maximum number of tasks a group can handle in the presence of other idle
5973 * or lightly loaded groups in the same sched domain.
5975 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5977 struct sched_domain
*child
;
5978 struct sched_group
*group
;
5980 WARN_ON(!sd
|| !sd
->groups
);
5982 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5987 sd
->groups
->__cpu_power
= 0;
5990 * For perf policy, if the groups in child domain share resources
5991 * (for example cores sharing some portions of the cache hierarchy
5992 * or SMT), then set this domain groups cpu_power such that each group
5993 * can handle only one task, when there are other idle groups in the
5994 * same sched domain.
5996 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5998 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5999 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6004 * add cpu_power of each child group to this groups cpu_power
6006 group
= child
->groups
;
6008 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6009 group
= group
->next
;
6010 } while (group
!= child
->groups
);
6014 * Build sched domains for a given set of cpus and attach the sched domains
6015 * to the individual cpus
6017 static int build_sched_domains(const cpumask_t
*cpu_map
)
6021 struct sched_group
**sched_group_nodes
= NULL
;
6022 int sd_allnodes
= 0;
6025 * Allocate the per-node list of sched groups
6027 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6029 if (!sched_group_nodes
) {
6030 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6033 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6037 * Set up domains for cpus specified by the cpu_map.
6039 for_each_cpu_mask(i
, *cpu_map
) {
6040 struct sched_domain
*sd
= NULL
, *p
;
6041 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6043 cpus_and(nodemask
, nodemask
, *cpu_map
);
6046 if (cpus_weight(*cpu_map
) >
6047 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6048 sd
= &per_cpu(allnodes_domains
, i
);
6049 *sd
= SD_ALLNODES_INIT
;
6050 sd
->span
= *cpu_map
;
6051 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6057 sd
= &per_cpu(node_domains
, i
);
6059 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6063 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6067 sd
= &per_cpu(phys_domains
, i
);
6069 sd
->span
= nodemask
;
6073 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6075 #ifdef CONFIG_SCHED_MC
6077 sd
= &per_cpu(core_domains
, i
);
6079 sd
->span
= cpu_coregroup_map(i
);
6080 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6083 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6086 #ifdef CONFIG_SCHED_SMT
6088 sd
= &per_cpu(cpu_domains
, i
);
6089 *sd
= SD_SIBLING_INIT
;
6090 sd
->span
= cpu_sibling_map
[i
];
6091 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6094 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6098 #ifdef CONFIG_SCHED_SMT
6099 /* Set up CPU (sibling) groups */
6100 for_each_cpu_mask(i
, *cpu_map
) {
6101 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6102 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6103 if (i
!= first_cpu(this_sibling_map
))
6106 init_sched_build_groups(this_sibling_map
, cpu_map
,
6111 #ifdef CONFIG_SCHED_MC
6112 /* Set up multi-core groups */
6113 for_each_cpu_mask(i
, *cpu_map
) {
6114 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6115 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6116 if (i
!= first_cpu(this_core_map
))
6118 init_sched_build_groups(this_core_map
, cpu_map
,
6119 &cpu_to_core_group
);
6123 /* Set up physical groups */
6124 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6125 cpumask_t nodemask
= node_to_cpumask(i
);
6127 cpus_and(nodemask
, nodemask
, *cpu_map
);
6128 if (cpus_empty(nodemask
))
6131 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6135 /* Set up node groups */
6137 init_sched_build_groups(*cpu_map
, cpu_map
,
6138 &cpu_to_allnodes_group
);
6140 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6141 /* Set up node groups */
6142 struct sched_group
*sg
, *prev
;
6143 cpumask_t nodemask
= node_to_cpumask(i
);
6144 cpumask_t domainspan
;
6145 cpumask_t covered
= CPU_MASK_NONE
;
6148 cpus_and(nodemask
, nodemask
, *cpu_map
);
6149 if (cpus_empty(nodemask
)) {
6150 sched_group_nodes
[i
] = NULL
;
6154 domainspan
= sched_domain_node_span(i
);
6155 cpus_and(domainspan
, domainspan
, *cpu_map
);
6157 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6159 printk(KERN_WARNING
"Can not alloc domain group for "
6163 sched_group_nodes
[i
] = sg
;
6164 for_each_cpu_mask(j
, nodemask
) {
6165 struct sched_domain
*sd
;
6167 sd
= &per_cpu(node_domains
, j
);
6170 sg
->__cpu_power
= 0;
6171 sg
->cpumask
= nodemask
;
6173 cpus_or(covered
, covered
, nodemask
);
6176 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6177 cpumask_t tmp
, notcovered
;
6178 int n
= (i
+ j
) % MAX_NUMNODES
;
6180 cpus_complement(notcovered
, covered
);
6181 cpus_and(tmp
, notcovered
, *cpu_map
);
6182 cpus_and(tmp
, tmp
, domainspan
);
6183 if (cpus_empty(tmp
))
6186 nodemask
= node_to_cpumask(n
);
6187 cpus_and(tmp
, tmp
, nodemask
);
6188 if (cpus_empty(tmp
))
6191 sg
= kmalloc_node(sizeof(struct sched_group
),
6195 "Can not alloc domain group for node %d\n", j
);
6198 sg
->__cpu_power
= 0;
6200 sg
->next
= prev
->next
;
6201 cpus_or(covered
, covered
, tmp
);
6208 /* Calculate CPU power for physical packages and nodes */
6209 #ifdef CONFIG_SCHED_SMT
6210 for_each_cpu_mask(i
, *cpu_map
) {
6211 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6213 init_sched_groups_power(i
, sd
);
6216 #ifdef CONFIG_SCHED_MC
6217 for_each_cpu_mask(i
, *cpu_map
) {
6218 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6220 init_sched_groups_power(i
, sd
);
6224 for_each_cpu_mask(i
, *cpu_map
) {
6225 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6227 init_sched_groups_power(i
, sd
);
6231 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6232 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6235 struct sched_group
*sg
;
6237 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6238 init_numa_sched_groups_power(sg
);
6242 /* Attach the domains */
6243 for_each_cpu_mask(i
, *cpu_map
) {
6244 struct sched_domain
*sd
;
6245 #ifdef CONFIG_SCHED_SMT
6246 sd
= &per_cpu(cpu_domains
, i
);
6247 #elif defined(CONFIG_SCHED_MC)
6248 sd
= &per_cpu(core_domains
, i
);
6250 sd
= &per_cpu(phys_domains
, i
);
6252 cpu_attach_domain(sd
, i
);
6259 free_sched_groups(cpu_map
);
6264 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6266 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6268 cpumask_t cpu_default_map
;
6272 * Setup mask for cpus without special case scheduling requirements.
6273 * For now this just excludes isolated cpus, but could be used to
6274 * exclude other special cases in the future.
6276 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6278 err
= build_sched_domains(&cpu_default_map
);
6283 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6285 free_sched_groups(cpu_map
);
6289 * Detach sched domains from a group of cpus specified in cpu_map
6290 * These cpus will now be attached to the NULL domain
6292 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6296 for_each_cpu_mask(i
, *cpu_map
)
6297 cpu_attach_domain(NULL
, i
);
6298 synchronize_sched();
6299 arch_destroy_sched_domains(cpu_map
);
6303 * Partition sched domains as specified by the cpumasks below.
6304 * This attaches all cpus from the cpumasks to the NULL domain,
6305 * waits for a RCU quiescent period, recalculates sched
6306 * domain information and then attaches them back to the
6307 * correct sched domains
6308 * Call with hotplug lock held
6310 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6312 cpumask_t change_map
;
6315 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6316 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6317 cpus_or(change_map
, *partition1
, *partition2
);
6319 /* Detach sched domains from all of the affected cpus */
6320 detach_destroy_domains(&change_map
);
6321 if (!cpus_empty(*partition1
))
6322 err
= build_sched_domains(partition1
);
6323 if (!err
&& !cpus_empty(*partition2
))
6324 err
= build_sched_domains(partition2
);
6329 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6330 static int arch_reinit_sched_domains(void)
6334 mutex_lock(&sched_hotcpu_mutex
);
6335 detach_destroy_domains(&cpu_online_map
);
6336 err
= arch_init_sched_domains(&cpu_online_map
);
6337 mutex_unlock(&sched_hotcpu_mutex
);
6342 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6346 if (buf
[0] != '0' && buf
[0] != '1')
6350 sched_smt_power_savings
= (buf
[0] == '1');
6352 sched_mc_power_savings
= (buf
[0] == '1');
6354 ret
= arch_reinit_sched_domains();
6356 return ret
? ret
: count
;
6359 #ifdef CONFIG_SCHED_MC
6360 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6362 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6364 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6365 const char *buf
, size_t count
)
6367 return sched_power_savings_store(buf
, count
, 0);
6369 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6370 sched_mc_power_savings_store
);
6373 #ifdef CONFIG_SCHED_SMT
6374 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6376 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6378 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6379 const char *buf
, size_t count
)
6381 return sched_power_savings_store(buf
, count
, 1);
6383 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6384 sched_smt_power_savings_store
);
6387 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6391 #ifdef CONFIG_SCHED_SMT
6393 err
= sysfs_create_file(&cls
->kset
.kobj
,
6394 &attr_sched_smt_power_savings
.attr
);
6396 #ifdef CONFIG_SCHED_MC
6397 if (!err
&& mc_capable())
6398 err
= sysfs_create_file(&cls
->kset
.kobj
,
6399 &attr_sched_mc_power_savings
.attr
);
6406 * Force a reinitialization of the sched domains hierarchy. The domains
6407 * and groups cannot be updated in place without racing with the balancing
6408 * code, so we temporarily attach all running cpus to the NULL domain
6409 * which will prevent rebalancing while the sched domains are recalculated.
6411 static int update_sched_domains(struct notifier_block
*nfb
,
6412 unsigned long action
, void *hcpu
)
6415 case CPU_UP_PREPARE
:
6416 case CPU_UP_PREPARE_FROZEN
:
6417 case CPU_DOWN_PREPARE
:
6418 case CPU_DOWN_PREPARE_FROZEN
:
6419 detach_destroy_domains(&cpu_online_map
);
6422 case CPU_UP_CANCELED
:
6423 case CPU_UP_CANCELED_FROZEN
:
6424 case CPU_DOWN_FAILED
:
6425 case CPU_DOWN_FAILED_FROZEN
:
6427 case CPU_ONLINE_FROZEN
:
6429 case CPU_DEAD_FROZEN
:
6431 * Fall through and re-initialise the domains.
6438 /* The hotplug lock is already held by cpu_up/cpu_down */
6439 arch_init_sched_domains(&cpu_online_map
);
6444 void __init
sched_init_smp(void)
6446 cpumask_t non_isolated_cpus
;
6448 mutex_lock(&sched_hotcpu_mutex
);
6449 arch_init_sched_domains(&cpu_online_map
);
6450 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6451 if (cpus_empty(non_isolated_cpus
))
6452 cpu_set(smp_processor_id(), non_isolated_cpus
);
6453 mutex_unlock(&sched_hotcpu_mutex
);
6454 /* XXX: Theoretical race here - CPU may be hotplugged now */
6455 hotcpu_notifier(update_sched_domains
, 0);
6457 init_sched_domain_sysctl();
6459 /* Move init over to a non-isolated CPU */
6460 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6464 void __init
sched_init_smp(void)
6467 #endif /* CONFIG_SMP */
6469 int in_sched_functions(unsigned long addr
)
6471 /* Linker adds these: start and end of __sched functions */
6472 extern char __sched_text_start
[], __sched_text_end
[];
6474 return in_lock_functions(addr
) ||
6475 (addr
>= (unsigned long)__sched_text_start
6476 && addr
< (unsigned long)__sched_text_end
);
6479 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6481 cfs_rq
->tasks_timeline
= RB_ROOT
;
6482 cfs_rq
->fair_clock
= 1;
6483 #ifdef CONFIG_FAIR_GROUP_SCHED
6488 void __init
sched_init(void)
6490 u64 now
= sched_clock();
6491 int highest_cpu
= 0;
6495 * Link up the scheduling class hierarchy:
6497 rt_sched_class
.next
= &fair_sched_class
;
6498 fair_sched_class
.next
= &idle_sched_class
;
6499 idle_sched_class
.next
= NULL
;
6501 for_each_possible_cpu(i
) {
6502 struct rt_prio_array
*array
;
6506 spin_lock_init(&rq
->lock
);
6507 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6510 init_cfs_rq(&rq
->cfs
, rq
);
6511 #ifdef CONFIG_FAIR_GROUP_SCHED
6512 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6513 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6515 rq
->ls
.load_update_last
= now
;
6516 rq
->ls
.load_update_start
= now
;
6518 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6519 rq
->cpu_load
[j
] = 0;
6522 rq
->active_balance
= 0;
6523 rq
->next_balance
= jiffies
;
6526 rq
->migration_thread
= NULL
;
6527 INIT_LIST_HEAD(&rq
->migration_queue
);
6529 atomic_set(&rq
->nr_iowait
, 0);
6531 array
= &rq
->rt
.active
;
6532 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6533 INIT_LIST_HEAD(array
->queue
+ j
);
6534 __clear_bit(j
, array
->bitmap
);
6537 /* delimiter for bitsearch: */
6538 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6541 set_load_weight(&init_task
);
6543 #ifdef CONFIG_PREEMPT_NOTIFIERS
6544 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6548 nr_cpu_ids
= highest_cpu
+ 1;
6549 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6552 #ifdef CONFIG_RT_MUTEXES
6553 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6557 * The boot idle thread does lazy MMU switching as well:
6559 atomic_inc(&init_mm
.mm_count
);
6560 enter_lazy_tlb(&init_mm
, current
);
6563 * Make us the idle thread. Technically, schedule() should not be
6564 * called from this thread, however somewhere below it might be,
6565 * but because we are the idle thread, we just pick up running again
6566 * when this runqueue becomes "idle".
6568 init_idle(current
, smp_processor_id());
6570 * During early bootup we pretend to be a normal task:
6572 current
->sched_class
= &fair_sched_class
;
6575 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6576 void __might_sleep(char *file
, int line
)
6579 static unsigned long prev_jiffy
; /* ratelimiting */
6581 if ((in_atomic() || irqs_disabled()) &&
6582 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6583 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6585 prev_jiffy
= jiffies
;
6586 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6587 " context at %s:%d\n", file
, line
);
6588 printk("in_atomic():%d, irqs_disabled():%d\n",
6589 in_atomic(), irqs_disabled());
6590 debug_show_held_locks(current
);
6591 if (irqs_disabled())
6592 print_irqtrace_events(current
);
6597 EXPORT_SYMBOL(__might_sleep
);
6600 #ifdef CONFIG_MAGIC_SYSRQ
6601 void normalize_rt_tasks(void)
6603 struct task_struct
*g
, *p
;
6604 unsigned long flags
;
6608 read_lock_irq(&tasklist_lock
);
6609 do_each_thread(g
, p
) {
6611 p
->se
.wait_runtime
= 0;
6612 p
->se
.exec_start
= 0;
6613 p
->se
.wait_start_fair
= 0;
6614 p
->se
.sleep_start_fair
= 0;
6615 #ifdef CONFIG_SCHEDSTATS
6616 p
->se
.wait_start
= 0;
6617 p
->se
.sleep_start
= 0;
6618 p
->se
.block_start
= 0;
6620 task_rq(p
)->cfs
.fair_clock
= 0;
6621 task_rq(p
)->clock
= 0;
6625 * Renice negative nice level userspace
6628 if (TASK_NICE(p
) < 0 && p
->mm
)
6629 set_user_nice(p
, 0);
6633 spin_lock_irqsave(&p
->pi_lock
, flags
);
6634 rq
= __task_rq_lock(p
);
6637 * Do not touch the migration thread:
6639 if (p
== rq
->migration_thread
)
6643 update_rq_clock(rq
);
6644 on_rq
= p
->se
.on_rq
;
6646 deactivate_task(rq
, p
, 0);
6647 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6649 activate_task(rq
, p
, 0);
6650 resched_task(rq
->curr
);
6655 __task_rq_unlock(rq
);
6656 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6657 } while_each_thread(g
, p
);
6659 read_unlock_irq(&tasklist_lock
);
6662 #endif /* CONFIG_MAGIC_SYSRQ */
6666 * These functions are only useful for the IA64 MCA handling.
6668 * They can only be called when the whole system has been
6669 * stopped - every CPU needs to be quiescent, and no scheduling
6670 * activity can take place. Using them for anything else would
6671 * be a serious bug, and as a result, they aren't even visible
6672 * under any other configuration.
6676 * curr_task - return the current task for a given cpu.
6677 * @cpu: the processor in question.
6679 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6681 struct task_struct
*curr_task(int cpu
)
6683 return cpu_curr(cpu
);
6687 * set_curr_task - set the current task for a given cpu.
6688 * @cpu: the processor in question.
6689 * @p: the task pointer to set.
6691 * Description: This function must only be used when non-maskable interrupts
6692 * are serviced on a separate stack. It allows the architecture to switch the
6693 * notion of the current task on a cpu in a non-blocking manner. This function
6694 * must be called with all CPU's synchronized, and interrupts disabled, the
6695 * and caller must save the original value of the current task (see
6696 * curr_task() above) and restore that value before reenabling interrupts and
6697 * re-starting the system.
6699 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6701 void set_curr_task(int cpu
, struct task_struct
*p
)