4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
68 * Scheduler clock - returns current time in nanosec units.
69 * This is default implementation.
70 * Architectures and sub-architectures can override this.
72 unsigned long long __attribute__((weak
)) sched_clock(void)
74 return (unsigned long long)jiffies
* (1000000000 / HZ
);
78 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
82 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
91 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96 * Some helpers for converting nanosecond timing to jiffy resolution
98 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
99 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
108 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
109 * Timeslices get refilled after they expire.
111 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
112 #define DEF_TIMESLICE (100 * HZ / 1000)
116 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
117 * Since cpu_power is a 'constant', we can use a reciprocal divide.
119 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
121 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
125 * Each time a sched group cpu_power is changed,
126 * we must compute its reciprocal value
128 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
130 sg
->__cpu_power
+= val
;
131 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
135 #define SCALE_PRIO(x, prio) \
136 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
139 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
140 * to time slice values: [800ms ... 100ms ... 5ms]
142 static unsigned int static_prio_timeslice(int static_prio
)
144 if (static_prio
== NICE_TO_PRIO(19))
147 if (static_prio
< NICE_TO_PRIO(0))
148 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
150 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
153 static inline int rt_policy(int policy
)
155 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
160 static inline int task_has_rt_policy(struct task_struct
*p
)
162 return rt_policy(p
->policy
);
166 * This is the priority-queue data structure of the RT scheduling class:
168 struct rt_prio_array
{
169 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
170 struct list_head queue
[MAX_RT_PRIO
];
174 struct load_weight load
;
175 u64 load_update_start
, load_update_last
;
176 unsigned long delta_fair
, delta_exec
, delta_stat
;
179 /* CFS-related fields in a runqueue */
181 struct load_weight load
;
182 unsigned long nr_running
;
188 unsigned long wait_runtime_overruns
, wait_runtime_underruns
;
190 struct rb_root tasks_timeline
;
191 struct rb_node
*rb_leftmost
;
192 struct rb_node
*rb_load_balance_curr
;
193 #ifdef CONFIG_FAIR_GROUP_SCHED
194 /* 'curr' points to currently running entity on this cfs_rq.
195 * It is set to NULL otherwise (i.e when none are currently running).
197 struct sched_entity
*curr
;
198 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
207 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
211 /* Real-Time classes' related field in a runqueue: */
213 struct rt_prio_array active
;
214 int rt_load_balance_idx
;
215 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
219 * This is the main, per-CPU runqueue data structure.
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
226 spinlock_t lock
; /* runqueue lock */
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
232 unsigned long nr_running
;
233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
235 unsigned char idle_at_tick
;
237 unsigned char in_nohz_recently
;
239 struct load_stat ls
; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates
;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
255 unsigned long nr_uninterruptible
;
257 struct task_struct
*curr
, *idle
;
258 unsigned long next_balance
;
259 struct mm_struct
*prev_mm
;
261 u64 clock
, prev_clock_raw
;
264 unsigned int clock_warps
, clock_overflows
;
265 unsigned int clock_unstable_events
;
270 struct sched_domain
*sd
;
272 /* For active balancing */
275 int cpu
; /* cpu of this runqueue */
277 struct task_struct
*migration_thread
;
278 struct list_head migration_queue
;
281 #ifdef CONFIG_SCHEDSTATS
283 struct sched_info rq_sched_info
;
285 /* sys_sched_yield() stats */
286 unsigned long yld_exp_empty
;
287 unsigned long yld_act_empty
;
288 unsigned long yld_both_empty
;
289 unsigned long yld_cnt
;
291 /* schedule() stats */
292 unsigned long sched_switch
;
293 unsigned long sched_cnt
;
294 unsigned long sched_goidle
;
296 /* try_to_wake_up() stats */
297 unsigned long ttwu_cnt
;
298 unsigned long ttwu_local
;
300 struct lock_class_key rq_lock_key
;
303 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
304 static DEFINE_MUTEX(sched_hotcpu_mutex
);
306 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
308 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
311 static inline int cpu_of(struct rq
*rq
)
321 * Update the per-runqueue clock, as finegrained as the platform can give
322 * us, but without assuming monotonicity, etc.:
324 static void __update_rq_clock(struct rq
*rq
)
326 u64 prev_raw
= rq
->prev_clock_raw
;
327 u64 now
= sched_clock();
328 s64 delta
= now
- prev_raw
;
329 u64 clock
= rq
->clock
;
331 #ifdef CONFIG_SCHED_DEBUG
332 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
335 * Protect against sched_clock() occasionally going backwards:
337 if (unlikely(delta
< 0)) {
342 * Catch too large forward jumps too:
344 if (unlikely(delta
> 2*TICK_NSEC
)) {
346 rq
->clock_overflows
++;
348 if (unlikely(delta
> rq
->clock_max_delta
))
349 rq
->clock_max_delta
= delta
;
354 rq
->prev_clock_raw
= now
;
358 static void update_rq_clock(struct rq
*rq
)
360 if (likely(smp_processor_id() == cpu_of(rq
)))
361 __update_rq_clock(rq
);
365 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
366 * See detach_destroy_domains: synchronize_sched for details.
368 * The domain tree of any CPU may only be accessed from within
369 * preempt-disabled sections.
371 #define for_each_domain(cpu, __sd) \
372 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
374 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
375 #define this_rq() (&__get_cpu_var(runqueues))
376 #define task_rq(p) cpu_rq(task_cpu(p))
377 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
380 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
381 * clock constructed from sched_clock():
383 unsigned long long cpu_clock(int cpu
)
385 unsigned long long now
;
389 local_irq_save(flags
);
393 local_irq_restore(flags
);
398 #ifdef CONFIG_FAIR_GROUP_SCHED
399 /* Change a task's ->cfs_rq if it moves across CPUs */
400 static inline void set_task_cfs_rq(struct task_struct
*p
)
402 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
405 static inline void set_task_cfs_rq(struct task_struct
*p
)
410 #ifndef prepare_arch_switch
411 # define prepare_arch_switch(next) do { } while (0)
413 #ifndef finish_arch_switch
414 # define finish_arch_switch(prev) do { } while (0)
417 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
418 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
420 return rq
->curr
== p
;
423 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
427 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
429 #ifdef CONFIG_DEBUG_SPINLOCK
430 /* this is a valid case when another task releases the spinlock */
431 rq
->lock
.owner
= current
;
434 * If we are tracking spinlock dependencies then we have to
435 * fix up the runqueue lock - which gets 'carried over' from
438 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
440 spin_unlock_irq(&rq
->lock
);
443 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
444 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
449 return rq
->curr
== p
;
453 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
457 * We can optimise this out completely for !SMP, because the
458 * SMP rebalancing from interrupt is the only thing that cares
463 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
464 spin_unlock_irq(&rq
->lock
);
466 spin_unlock(&rq
->lock
);
470 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
474 * After ->oncpu is cleared, the task can be moved to a different CPU.
475 * We must ensure this doesn't happen until the switch is completely
481 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
485 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
488 * __task_rq_lock - lock the runqueue a given task resides on.
489 * Must be called interrupts disabled.
491 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
498 spin_lock(&rq
->lock
);
499 if (unlikely(rq
!= task_rq(p
))) {
500 spin_unlock(&rq
->lock
);
501 goto repeat_lock_task
;
507 * task_rq_lock - lock the runqueue a given task resides on and disable
508 * interrupts. Note the ordering: we can safely lookup the task_rq without
509 * explicitly disabling preemption.
511 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
517 local_irq_save(*flags
);
519 spin_lock(&rq
->lock
);
520 if (unlikely(rq
!= task_rq(p
))) {
521 spin_unlock_irqrestore(&rq
->lock
, *flags
);
522 goto repeat_lock_task
;
527 static inline void __task_rq_unlock(struct rq
*rq
)
530 spin_unlock(&rq
->lock
);
533 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
536 spin_unlock_irqrestore(&rq
->lock
, *flags
);
540 * this_rq_lock - lock this runqueue and disable interrupts.
542 static inline struct rq
*this_rq_lock(void)
549 spin_lock(&rq
->lock
);
555 * CPU frequency is/was unstable - start new by setting prev_clock_raw:
557 void sched_clock_unstable_event(void)
562 rq
= task_rq_lock(current
, &flags
);
563 rq
->prev_clock_raw
= sched_clock();
564 rq
->clock_unstable_events
++;
565 task_rq_unlock(rq
, &flags
);
569 * resched_task - mark a task 'to be rescheduled now'.
571 * On UP this means the setting of the need_resched flag, on SMP it
572 * might also involve a cross-CPU call to trigger the scheduler on
577 #ifndef tsk_is_polling
578 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
581 static void resched_task(struct task_struct
*p
)
585 assert_spin_locked(&task_rq(p
)->lock
);
587 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
590 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
593 if (cpu
== smp_processor_id())
596 /* NEED_RESCHED must be visible before we test polling */
598 if (!tsk_is_polling(p
))
599 smp_send_reschedule(cpu
);
602 static void resched_cpu(int cpu
)
604 struct rq
*rq
= cpu_rq(cpu
);
607 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
609 resched_task(cpu_curr(cpu
));
610 spin_unlock_irqrestore(&rq
->lock
, flags
);
613 static inline void resched_task(struct task_struct
*p
)
615 assert_spin_locked(&task_rq(p
)->lock
);
616 set_tsk_need_resched(p
);
620 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
622 #if BITS_PER_LONG == 32
623 if (likely(divident
<= 0xffffffffULL
))
624 return (u32
)divident
/ divisor
;
625 do_div(divident
, divisor
);
629 return divident
/ divisor
;
633 #if BITS_PER_LONG == 32
634 # define WMULT_CONST (~0UL)
636 # define WMULT_CONST (1UL << 32)
639 #define WMULT_SHIFT 32
642 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
643 struct load_weight
*lw
)
647 if (unlikely(!lw
->inv_weight
))
648 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
650 tmp
= (u64
)delta_exec
* weight
;
652 * Check whether we'd overflow the 64-bit multiplication:
654 if (unlikely(tmp
> WMULT_CONST
)) {
655 tmp
= ((tmp
>> WMULT_SHIFT
/2) * lw
->inv_weight
)
658 tmp
= (tmp
* lw
->inv_weight
) >> WMULT_SHIFT
;
661 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
664 static inline unsigned long
665 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
667 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
670 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
676 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
683 * To aid in avoiding the subversion of "niceness" due to uneven distribution
684 * of tasks with abnormal "nice" values across CPUs the contribution that
685 * each task makes to its run queue's load is weighted according to its
686 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
687 * scaled version of the new time slice allocation that they receive on time
691 #define WEIGHT_IDLEPRIO 2
692 #define WMULT_IDLEPRIO (1 << 31)
695 * Nice levels are multiplicative, with a gentle 10% change for every
696 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
697 * nice 1, it will get ~10% less CPU time than another CPU-bound task
698 * that remained on nice 0.
700 * The "10% effect" is relative and cumulative: from _any_ nice level,
701 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
702 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
703 * If a task goes up by ~10% and another task goes down by ~10% then
704 * the relative distance between them is ~25%.)
706 static const int prio_to_weight
[40] = {
707 /* -20 */ 88818, 71054, 56843, 45475, 36380, 29104, 23283, 18626, 14901, 11921,
708 /* -10 */ 9537, 7629, 6103, 4883, 3906, 3125, 2500, 2000, 1600, 1280,
709 /* 0 */ NICE_0_LOAD
/* 1024 */,
710 /* 1 */ 819, 655, 524, 419, 336, 268, 215, 172, 137,
711 /* 10 */ 110, 87, 70, 56, 45, 36, 29, 23, 18, 15,
715 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
717 * In cases where the weight does not change often, we can use the
718 * precalculated inverse to speed up arithmetics by turning divisions
719 * into multiplications:
721 static const u32 prio_to_wmult
[40] = {
722 /* -20 */ 48356, 60446, 75558, 94446, 118058,
723 /* -15 */ 147573, 184467, 230589, 288233, 360285,
724 /* -10 */ 450347, 562979, 703746, 879575, 1099582,
725 /* -5 */ 1374389, 1717986, 2147483, 2684354, 3355443,
726 /* 0 */ 4194304, 5244160, 6557201, 8196502, 10250518,
727 /* 5 */ 12782640, 16025997, 19976592, 24970740, 31350126,
728 /* 10 */ 39045157, 49367440, 61356675, 76695844, 95443717,
729 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
732 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
735 * runqueue iterator, to support SMP load-balancing between different
736 * scheduling classes, without having to expose their internal data
737 * structures to the load-balancing proper:
741 struct task_struct
*(*start
)(void *);
742 struct task_struct
*(*next
)(void *);
745 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
746 unsigned long max_nr_move
, unsigned long max_load_move
,
747 struct sched_domain
*sd
, enum cpu_idle_type idle
,
748 int *all_pinned
, unsigned long *load_moved
,
749 int *this_best_prio
, struct rq_iterator
*iterator
);
751 #include "sched_stats.h"
752 #include "sched_rt.c"
753 #include "sched_fair.c"
754 #include "sched_idletask.c"
755 #ifdef CONFIG_SCHED_DEBUG
756 # include "sched_debug.c"
759 #define sched_class_highest (&rt_sched_class)
761 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
763 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
764 ls
->delta_exec
+= ls
->delta_stat
;
765 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
771 * Update delta_exec, delta_fair fields for rq.
773 * delta_fair clock advances at a rate inversely proportional to
774 * total load (rq->ls.load.weight) on the runqueue, while
775 * delta_exec advances at the same rate as wall-clock (provided
778 * delta_exec / delta_fair is a measure of the (smoothened) load on this
779 * runqueue over any given interval. This (smoothened) load is used
780 * during load balance.
782 * This function is called /before/ updating rq->ls.load
783 * and when switching tasks.
785 static void update_curr_load(struct rq
*rq
, u64 now
)
787 struct load_stat
*ls
= &rq
->ls
;
790 start
= ls
->load_update_start
;
791 ls
->load_update_start
= rq
->clock
;
792 ls
->delta_stat
+= rq
->clock
- start
;
794 * Stagger updates to ls->delta_fair. Very frequent updates
797 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
798 __update_curr_load(rq
, ls
);
802 inc_load(struct rq
*rq
, const struct task_struct
*p
, u64 now
)
804 update_curr_load(rq
, now
);
805 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
809 dec_load(struct rq
*rq
, const struct task_struct
*p
, u64 now
)
811 update_curr_load(rq
, now
);
812 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
815 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
, u64 now
)
818 inc_load(rq
, p
, now
);
821 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
, u64 now
)
824 dec_load(rq
, p
, now
);
827 static void set_load_weight(struct task_struct
*p
)
829 task_rq(p
)->cfs
.wait_runtime
-= p
->se
.wait_runtime
;
830 p
->se
.wait_runtime
= 0;
832 if (task_has_rt_policy(p
)) {
833 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
834 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
839 * SCHED_IDLE tasks get minimal weight:
841 if (p
->policy
== SCHED_IDLE
) {
842 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
843 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
847 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
848 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
852 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, u64 now
)
854 sched_info_queued(p
);
855 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
860 dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
862 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
867 * __normal_prio - return the priority that is based on the static prio
869 static inline int __normal_prio(struct task_struct
*p
)
871 return p
->static_prio
;
875 * Calculate the expected normal priority: i.e. priority
876 * without taking RT-inheritance into account. Might be
877 * boosted by interactivity modifiers. Changes upon fork,
878 * setprio syscalls, and whenever the interactivity
879 * estimator recalculates.
881 static inline int normal_prio(struct task_struct
*p
)
885 if (task_has_rt_policy(p
))
886 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
888 prio
= __normal_prio(p
);
893 * Calculate the current priority, i.e. the priority
894 * taken into account by the scheduler. This value might
895 * be boosted by RT tasks, or might be boosted by
896 * interactivity modifiers. Will be RT if the task got
897 * RT-boosted. If not then it returns p->normal_prio.
899 static int effective_prio(struct task_struct
*p
)
901 p
->normal_prio
= normal_prio(p
);
903 * If we are RT tasks or we were boosted to RT priority,
904 * keep the priority unchanged. Otherwise, update priority
905 * to the normal priority:
907 if (!rt_prio(p
->prio
))
908 return p
->normal_prio
;
913 * activate_task - move a task to the runqueue.
915 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
922 if (p
->state
== TASK_UNINTERRUPTIBLE
)
923 rq
->nr_uninterruptible
--;
925 enqueue_task(rq
, p
, wakeup
, now
);
926 inc_nr_running(p
, rq
, now
);
930 * activate_idle_task - move idle task to the _front_ of runqueue.
932 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
939 if (p
->state
== TASK_UNINTERRUPTIBLE
)
940 rq
->nr_uninterruptible
--;
942 enqueue_task(rq
, p
, 0, now
);
943 inc_nr_running(p
, rq
, now
);
947 * deactivate_task - remove a task from the runqueue.
950 deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
, u64 now
)
952 if (p
->state
== TASK_UNINTERRUPTIBLE
)
953 rq
->nr_uninterruptible
++;
955 dequeue_task(rq
, p
, sleep
, now
);
956 dec_nr_running(p
, rq
, now
);
960 * task_curr - is this task currently executing on a CPU?
961 * @p: the task in question.
963 inline int task_curr(const struct task_struct
*p
)
965 return cpu_curr(task_cpu(p
)) == p
;
968 /* Used instead of source_load when we know the type == 0 */
969 unsigned long weighted_cpuload(const int cpu
)
971 return cpu_rq(cpu
)->ls
.load
.weight
;
974 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
977 task_thread_info(p
)->cpu
= cpu
;
984 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
986 int old_cpu
= task_cpu(p
);
987 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
988 u64 clock_offset
, fair_clock_offset
;
990 clock_offset
= old_rq
->clock
- new_rq
->clock
;
991 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
993 if (p
->se
.wait_start_fair
)
994 p
->se
.wait_start_fair
-= fair_clock_offset
;
995 if (p
->se
.sleep_start_fair
)
996 p
->se
.sleep_start_fair
-= fair_clock_offset
;
998 #ifdef CONFIG_SCHEDSTATS
999 if (p
->se
.wait_start
)
1000 p
->se
.wait_start
-= clock_offset
;
1001 if (p
->se
.sleep_start
)
1002 p
->se
.sleep_start
-= clock_offset
;
1003 if (p
->se
.block_start
)
1004 p
->se
.block_start
-= clock_offset
;
1007 __set_task_cpu(p
, new_cpu
);
1010 struct migration_req
{
1011 struct list_head list
;
1013 struct task_struct
*task
;
1016 struct completion done
;
1020 * The task's runqueue lock must be held.
1021 * Returns true if you have to wait for migration thread.
1024 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1026 struct rq
*rq
= task_rq(p
);
1029 * If the task is not on a runqueue (and not running), then
1030 * it is sufficient to simply update the task's cpu field.
1032 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1033 set_task_cpu(p
, dest_cpu
);
1037 init_completion(&req
->done
);
1039 req
->dest_cpu
= dest_cpu
;
1040 list_add(&req
->list
, &rq
->migration_queue
);
1046 * wait_task_inactive - wait for a thread to unschedule.
1048 * The caller must ensure that the task *will* unschedule sometime soon,
1049 * else this function might spin for a *long* time. This function can't
1050 * be called with interrupts off, or it may introduce deadlock with
1051 * smp_call_function() if an IPI is sent by the same process we are
1052 * waiting to become inactive.
1054 void wait_task_inactive(struct task_struct
*p
)
1056 unsigned long flags
;
1062 * We do the initial early heuristics without holding
1063 * any task-queue locks at all. We'll only try to get
1064 * the runqueue lock when things look like they will
1070 * If the task is actively running on another CPU
1071 * still, just relax and busy-wait without holding
1074 * NOTE! Since we don't hold any locks, it's not
1075 * even sure that "rq" stays as the right runqueue!
1076 * But we don't care, since "task_running()" will
1077 * return false if the runqueue has changed and p
1078 * is actually now running somewhere else!
1080 while (task_running(rq
, p
))
1084 * Ok, time to look more closely! We need the rq
1085 * lock now, to be *sure*. If we're wrong, we'll
1086 * just go back and repeat.
1088 rq
= task_rq_lock(p
, &flags
);
1089 running
= task_running(rq
, p
);
1090 on_rq
= p
->se
.on_rq
;
1091 task_rq_unlock(rq
, &flags
);
1094 * Was it really running after all now that we
1095 * checked with the proper locks actually held?
1097 * Oops. Go back and try again..
1099 if (unlikely(running
)) {
1105 * It's not enough that it's not actively running,
1106 * it must be off the runqueue _entirely_, and not
1109 * So if it wa still runnable (but just not actively
1110 * running right now), it's preempted, and we should
1111 * yield - it could be a while.
1113 if (unlikely(on_rq
)) {
1119 * Ahh, all good. It wasn't running, and it wasn't
1120 * runnable, which means that it will never become
1121 * running in the future either. We're all done!
1126 * kick_process - kick a running thread to enter/exit the kernel
1127 * @p: the to-be-kicked thread
1129 * Cause a process which is running on another CPU to enter
1130 * kernel-mode, without any delay. (to get signals handled.)
1132 * NOTE: this function doesnt have to take the runqueue lock,
1133 * because all it wants to ensure is that the remote task enters
1134 * the kernel. If the IPI races and the task has been migrated
1135 * to another CPU then no harm is done and the purpose has been
1138 void kick_process(struct task_struct
*p
)
1144 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1145 smp_send_reschedule(cpu
);
1150 * Return a low guess at the load of a migration-source cpu weighted
1151 * according to the scheduling class and "nice" value.
1153 * We want to under-estimate the load of migration sources, to
1154 * balance conservatively.
1156 static inline unsigned long source_load(int cpu
, int type
)
1158 struct rq
*rq
= cpu_rq(cpu
);
1159 unsigned long total
= weighted_cpuload(cpu
);
1164 return min(rq
->cpu_load
[type
-1], total
);
1168 * Return a high guess at the load of a migration-target cpu weighted
1169 * according to the scheduling class and "nice" value.
1171 static inline unsigned long target_load(int cpu
, int type
)
1173 struct rq
*rq
= cpu_rq(cpu
);
1174 unsigned long total
= weighted_cpuload(cpu
);
1179 return max(rq
->cpu_load
[type
-1], total
);
1183 * Return the average load per task on the cpu's run queue
1185 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1187 struct rq
*rq
= cpu_rq(cpu
);
1188 unsigned long total
= weighted_cpuload(cpu
);
1189 unsigned long n
= rq
->nr_running
;
1191 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1195 * find_idlest_group finds and returns the least busy CPU group within the
1198 static struct sched_group
*
1199 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1201 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1202 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1203 int load_idx
= sd
->forkexec_idx
;
1204 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1207 unsigned long load
, avg_load
;
1211 /* Skip over this group if it has no CPUs allowed */
1212 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1215 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1217 /* Tally up the load of all CPUs in the group */
1220 for_each_cpu_mask(i
, group
->cpumask
) {
1221 /* Bias balancing toward cpus of our domain */
1223 load
= source_load(i
, load_idx
);
1225 load
= target_load(i
, load_idx
);
1230 /* Adjust by relative CPU power of the group */
1231 avg_load
= sg_div_cpu_power(group
,
1232 avg_load
* SCHED_LOAD_SCALE
);
1235 this_load
= avg_load
;
1237 } else if (avg_load
< min_load
) {
1238 min_load
= avg_load
;
1242 group
= group
->next
;
1243 } while (group
!= sd
->groups
);
1245 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1251 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1254 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1257 unsigned long load
, min_load
= ULONG_MAX
;
1261 /* Traverse only the allowed CPUs */
1262 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1264 for_each_cpu_mask(i
, tmp
) {
1265 load
= weighted_cpuload(i
);
1267 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1277 * sched_balance_self: balance the current task (running on cpu) in domains
1278 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1281 * Balance, ie. select the least loaded group.
1283 * Returns the target CPU number, or the same CPU if no balancing is needed.
1285 * preempt must be disabled.
1287 static int sched_balance_self(int cpu
, int flag
)
1289 struct task_struct
*t
= current
;
1290 struct sched_domain
*tmp
, *sd
= NULL
;
1292 for_each_domain(cpu
, tmp
) {
1294 * If power savings logic is enabled for a domain, stop there.
1296 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1298 if (tmp
->flags
& flag
)
1304 struct sched_group
*group
;
1305 int new_cpu
, weight
;
1307 if (!(sd
->flags
& flag
)) {
1313 group
= find_idlest_group(sd
, t
, cpu
);
1319 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1320 if (new_cpu
== -1 || new_cpu
== cpu
) {
1321 /* Now try balancing at a lower domain level of cpu */
1326 /* Now try balancing at a lower domain level of new_cpu */
1329 weight
= cpus_weight(span
);
1330 for_each_domain(cpu
, tmp
) {
1331 if (weight
<= cpus_weight(tmp
->span
))
1333 if (tmp
->flags
& flag
)
1336 /* while loop will break here if sd == NULL */
1342 #endif /* CONFIG_SMP */
1345 * wake_idle() will wake a task on an idle cpu if task->cpu is
1346 * not idle and an idle cpu is available. The span of cpus to
1347 * search starts with cpus closest then further out as needed,
1348 * so we always favor a closer, idle cpu.
1350 * Returns the CPU we should wake onto.
1352 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1353 static int wake_idle(int cpu
, struct task_struct
*p
)
1356 struct sched_domain
*sd
;
1360 * If it is idle, then it is the best cpu to run this task.
1362 * This cpu is also the best, if it has more than one task already.
1363 * Siblings must be also busy(in most cases) as they didn't already
1364 * pickup the extra load from this cpu and hence we need not check
1365 * sibling runqueue info. This will avoid the checks and cache miss
1366 * penalities associated with that.
1368 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1371 for_each_domain(cpu
, sd
) {
1372 if (sd
->flags
& SD_WAKE_IDLE
) {
1373 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1374 for_each_cpu_mask(i
, tmp
) {
1385 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1392 * try_to_wake_up - wake up a thread
1393 * @p: the to-be-woken-up thread
1394 * @state: the mask of task states that can be woken
1395 * @sync: do a synchronous wakeup?
1397 * Put it on the run-queue if it's not already there. The "current"
1398 * thread is always on the run-queue (except when the actual
1399 * re-schedule is in progress), and as such you're allowed to do
1400 * the simpler "current->state = TASK_RUNNING" to mark yourself
1401 * runnable without the overhead of this.
1403 * returns failure only if the task is already active.
1405 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1407 int cpu
, this_cpu
, success
= 0;
1408 unsigned long flags
;
1412 struct sched_domain
*sd
, *this_sd
= NULL
;
1413 unsigned long load
, this_load
;
1417 rq
= task_rq_lock(p
, &flags
);
1418 old_state
= p
->state
;
1419 if (!(old_state
& state
))
1426 this_cpu
= smp_processor_id();
1429 if (unlikely(task_running(rq
, p
)))
1434 schedstat_inc(rq
, ttwu_cnt
);
1435 if (cpu
== this_cpu
) {
1436 schedstat_inc(rq
, ttwu_local
);
1440 for_each_domain(this_cpu
, sd
) {
1441 if (cpu_isset(cpu
, sd
->span
)) {
1442 schedstat_inc(sd
, ttwu_wake_remote
);
1448 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1452 * Check for affine wakeup and passive balancing possibilities.
1455 int idx
= this_sd
->wake_idx
;
1456 unsigned int imbalance
;
1458 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1460 load
= source_load(cpu
, idx
);
1461 this_load
= target_load(this_cpu
, idx
);
1463 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1465 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1466 unsigned long tl
= this_load
;
1467 unsigned long tl_per_task
;
1469 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1472 * If sync wakeup then subtract the (maximum possible)
1473 * effect of the currently running task from the load
1474 * of the current CPU:
1477 tl
-= current
->se
.load
.weight
;
1480 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1481 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1483 * This domain has SD_WAKE_AFFINE and
1484 * p is cache cold in this domain, and
1485 * there is no bad imbalance.
1487 schedstat_inc(this_sd
, ttwu_move_affine
);
1493 * Start passive balancing when half the imbalance_pct
1496 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1497 if (imbalance
*this_load
<= 100*load
) {
1498 schedstat_inc(this_sd
, ttwu_move_balance
);
1504 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1506 new_cpu
= wake_idle(new_cpu
, p
);
1507 if (new_cpu
!= cpu
) {
1508 set_task_cpu(p
, new_cpu
);
1509 task_rq_unlock(rq
, &flags
);
1510 /* might preempt at this point */
1511 rq
= task_rq_lock(p
, &flags
);
1512 old_state
= p
->state
;
1513 if (!(old_state
& state
))
1518 this_cpu
= smp_processor_id();
1523 #endif /* CONFIG_SMP */
1524 activate_task(rq
, p
, 1);
1526 * Sync wakeups (i.e. those types of wakeups where the waker
1527 * has indicated that it will leave the CPU in short order)
1528 * don't trigger a preemption, if the woken up task will run on
1529 * this cpu. (in this case the 'I will reschedule' promise of
1530 * the waker guarantees that the freshly woken up task is going
1531 * to be considered on this CPU.)
1533 if (!sync
|| cpu
!= this_cpu
)
1534 check_preempt_curr(rq
, p
);
1538 p
->state
= TASK_RUNNING
;
1540 task_rq_unlock(rq
, &flags
);
1545 int fastcall
wake_up_process(struct task_struct
*p
)
1547 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1548 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1550 EXPORT_SYMBOL(wake_up_process
);
1552 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1554 return try_to_wake_up(p
, state
, 0);
1558 * Perform scheduler related setup for a newly forked process p.
1559 * p is forked by current.
1561 * __sched_fork() is basic setup used by init_idle() too:
1563 static void __sched_fork(struct task_struct
*p
)
1565 p
->se
.wait_start_fair
= 0;
1566 p
->se
.exec_start
= 0;
1567 p
->se
.sum_exec_runtime
= 0;
1568 p
->se
.delta_exec
= 0;
1569 p
->se
.delta_fair_run
= 0;
1570 p
->se
.delta_fair_sleep
= 0;
1571 p
->se
.wait_runtime
= 0;
1572 p
->se
.sleep_start_fair
= 0;
1574 #ifdef CONFIG_SCHEDSTATS
1575 p
->se
.wait_start
= 0;
1576 p
->se
.sum_wait_runtime
= 0;
1577 p
->se
.sum_sleep_runtime
= 0;
1578 p
->se
.sleep_start
= 0;
1579 p
->se
.block_start
= 0;
1580 p
->se
.sleep_max
= 0;
1581 p
->se
.block_max
= 0;
1584 p
->se
.wait_runtime_overruns
= 0;
1585 p
->se
.wait_runtime_underruns
= 0;
1588 INIT_LIST_HEAD(&p
->run_list
);
1591 #ifdef CONFIG_PREEMPT_NOTIFIERS
1592 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1596 * We mark the process as running here, but have not actually
1597 * inserted it onto the runqueue yet. This guarantees that
1598 * nobody will actually run it, and a signal or other external
1599 * event cannot wake it up and insert it on the runqueue either.
1601 p
->state
= TASK_RUNNING
;
1605 * fork()/clone()-time setup:
1607 void sched_fork(struct task_struct
*p
, int clone_flags
)
1609 int cpu
= get_cpu();
1614 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1616 __set_task_cpu(p
, cpu
);
1619 * Make sure we do not leak PI boosting priority to the child:
1621 p
->prio
= current
->normal_prio
;
1623 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1624 if (likely(sched_info_on()))
1625 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1627 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1630 #ifdef CONFIG_PREEMPT
1631 /* Want to start with kernel preemption disabled. */
1632 task_thread_info(p
)->preempt_count
= 1;
1638 * After fork, child runs first. (default) If set to 0 then
1639 * parent will (try to) run first.
1641 unsigned int __read_mostly sysctl_sched_child_runs_first
= 1;
1644 * wake_up_new_task - wake up a newly created task for the first time.
1646 * This function will do some initial scheduler statistics housekeeping
1647 * that must be done for every newly created context, then puts the task
1648 * on the runqueue and wakes it.
1650 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1652 unsigned long flags
;
1657 rq
= task_rq_lock(p
, &flags
);
1658 BUG_ON(p
->state
!= TASK_RUNNING
);
1659 this_cpu
= smp_processor_id(); /* parent's CPU */
1660 update_rq_clock(rq
);
1663 p
->prio
= effective_prio(p
);
1665 if (!p
->sched_class
->task_new
|| !sysctl_sched_child_runs_first
||
1666 (clone_flags
& CLONE_VM
) || task_cpu(p
) != this_cpu
||
1667 !current
->se
.on_rq
) {
1669 activate_task(rq
, p
, 0);
1672 * Let the scheduling class do new task startup
1673 * management (if any):
1675 p
->sched_class
->task_new(rq
, p
, now
);
1676 inc_nr_running(p
, rq
, now
);
1678 check_preempt_curr(rq
, p
);
1679 task_rq_unlock(rq
, &flags
);
1682 #ifdef CONFIG_PREEMPT_NOTIFIERS
1685 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1686 * @notifier: notifier struct to register
1688 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1690 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1692 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1695 * preempt_notifier_unregister - no longer interested in preemption notifications
1696 * @notifier: notifier struct to unregister
1698 * This is safe to call from within a preemption notifier.
1700 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1702 hlist_del(¬ifier
->link
);
1704 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1706 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1708 struct preempt_notifier
*notifier
;
1709 struct hlist_node
*node
;
1711 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1712 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1716 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1717 struct task_struct
*next
)
1719 struct preempt_notifier
*notifier
;
1720 struct hlist_node
*node
;
1722 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1723 notifier
->ops
->sched_out(notifier
, next
);
1728 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1733 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1734 struct task_struct
*next
)
1741 * prepare_task_switch - prepare to switch tasks
1742 * @rq: the runqueue preparing to switch
1743 * @prev: the current task that is being switched out
1744 * @next: the task we are going to switch to.
1746 * This is called with the rq lock held and interrupts off. It must
1747 * be paired with a subsequent finish_task_switch after the context
1750 * prepare_task_switch sets up locking and calls architecture specific
1754 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1755 struct task_struct
*next
)
1757 fire_sched_out_preempt_notifiers(prev
, next
);
1758 prepare_lock_switch(rq
, next
);
1759 prepare_arch_switch(next
);
1763 * finish_task_switch - clean up after a task-switch
1764 * @rq: runqueue associated with task-switch
1765 * @prev: the thread we just switched away from.
1767 * finish_task_switch must be called after the context switch, paired
1768 * with a prepare_task_switch call before the context switch.
1769 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1770 * and do any other architecture-specific cleanup actions.
1772 * Note that we may have delayed dropping an mm in context_switch(). If
1773 * so, we finish that here outside of the runqueue lock. (Doing it
1774 * with the lock held can cause deadlocks; see schedule() for
1777 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1778 __releases(rq
->lock
)
1780 struct mm_struct
*mm
= rq
->prev_mm
;
1786 * A task struct has one reference for the use as "current".
1787 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1788 * schedule one last time. The schedule call will never return, and
1789 * the scheduled task must drop that reference.
1790 * The test for TASK_DEAD must occur while the runqueue locks are
1791 * still held, otherwise prev could be scheduled on another cpu, die
1792 * there before we look at prev->state, and then the reference would
1794 * Manfred Spraul <manfred@colorfullife.com>
1796 prev_state
= prev
->state
;
1797 finish_arch_switch(prev
);
1798 finish_lock_switch(rq
, prev
);
1799 fire_sched_in_preempt_notifiers(current
);
1802 if (unlikely(prev_state
== TASK_DEAD
)) {
1804 * Remove function-return probe instances associated with this
1805 * task and put them back on the free list.
1807 kprobe_flush_task(prev
);
1808 put_task_struct(prev
);
1813 * schedule_tail - first thing a freshly forked thread must call.
1814 * @prev: the thread we just switched away from.
1816 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1817 __releases(rq
->lock
)
1819 struct rq
*rq
= this_rq();
1821 finish_task_switch(rq
, prev
);
1822 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1823 /* In this case, finish_task_switch does not reenable preemption */
1826 if (current
->set_child_tid
)
1827 put_user(current
->pid
, current
->set_child_tid
);
1831 * context_switch - switch to the new MM and the new
1832 * thread's register state.
1835 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1836 struct task_struct
*next
)
1838 struct mm_struct
*mm
, *oldmm
;
1840 prepare_task_switch(rq
, prev
, next
);
1842 oldmm
= prev
->active_mm
;
1844 * For paravirt, this is coupled with an exit in switch_to to
1845 * combine the page table reload and the switch backend into
1848 arch_enter_lazy_cpu_mode();
1850 if (unlikely(!mm
)) {
1851 next
->active_mm
= oldmm
;
1852 atomic_inc(&oldmm
->mm_count
);
1853 enter_lazy_tlb(oldmm
, next
);
1855 switch_mm(oldmm
, mm
, next
);
1857 if (unlikely(!prev
->mm
)) {
1858 prev
->active_mm
= NULL
;
1859 rq
->prev_mm
= oldmm
;
1862 * Since the runqueue lock will be released by the next
1863 * task (which is an invalid locking op but in the case
1864 * of the scheduler it's an obvious special-case), so we
1865 * do an early lockdep release here:
1867 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1868 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1871 /* Here we just switch the register state and the stack. */
1872 switch_to(prev
, next
, prev
);
1876 * this_rq must be evaluated again because prev may have moved
1877 * CPUs since it called schedule(), thus the 'rq' on its stack
1878 * frame will be invalid.
1880 finish_task_switch(this_rq(), prev
);
1884 * nr_running, nr_uninterruptible and nr_context_switches:
1886 * externally visible scheduler statistics: current number of runnable
1887 * threads, current number of uninterruptible-sleeping threads, total
1888 * number of context switches performed since bootup.
1890 unsigned long nr_running(void)
1892 unsigned long i
, sum
= 0;
1894 for_each_online_cpu(i
)
1895 sum
+= cpu_rq(i
)->nr_running
;
1900 unsigned long nr_uninterruptible(void)
1902 unsigned long i
, sum
= 0;
1904 for_each_possible_cpu(i
)
1905 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1908 * Since we read the counters lockless, it might be slightly
1909 * inaccurate. Do not allow it to go below zero though:
1911 if (unlikely((long)sum
< 0))
1917 unsigned long long nr_context_switches(void)
1920 unsigned long long sum
= 0;
1922 for_each_possible_cpu(i
)
1923 sum
+= cpu_rq(i
)->nr_switches
;
1928 unsigned long nr_iowait(void)
1930 unsigned long i
, sum
= 0;
1932 for_each_possible_cpu(i
)
1933 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1938 unsigned long nr_active(void)
1940 unsigned long i
, running
= 0, uninterruptible
= 0;
1942 for_each_online_cpu(i
) {
1943 running
+= cpu_rq(i
)->nr_running
;
1944 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1947 if (unlikely((long)uninterruptible
< 0))
1948 uninterruptible
= 0;
1950 return running
+ uninterruptible
;
1954 * Update rq->cpu_load[] statistics. This function is usually called every
1955 * scheduler tick (TICK_NSEC).
1957 static void update_cpu_load(struct rq
*this_rq
)
1959 u64 fair_delta64
, exec_delta64
, idle_delta64
, sample_interval64
, tmp64
;
1960 unsigned long total_load
= this_rq
->ls
.load
.weight
;
1961 unsigned long this_load
= total_load
;
1962 struct load_stat
*ls
= &this_rq
->ls
;
1966 __update_rq_clock(this_rq
);
1967 now
= this_rq
->clock
;
1969 this_rq
->nr_load_updates
++;
1970 if (unlikely(!(sysctl_sched_features
& SCHED_FEAT_PRECISE_CPU_LOAD
)))
1973 /* Update delta_fair/delta_exec fields first */
1974 update_curr_load(this_rq
, now
);
1976 fair_delta64
= ls
->delta_fair
+ 1;
1979 exec_delta64
= ls
->delta_exec
+ 1;
1982 sample_interval64
= this_rq
->clock
- ls
->load_update_last
;
1983 ls
->load_update_last
= this_rq
->clock
;
1985 if ((s64
)sample_interval64
< (s64
)TICK_NSEC
)
1986 sample_interval64
= TICK_NSEC
;
1988 if (exec_delta64
> sample_interval64
)
1989 exec_delta64
= sample_interval64
;
1991 idle_delta64
= sample_interval64
- exec_delta64
;
1993 tmp64
= div64_64(SCHED_LOAD_SCALE
* exec_delta64
, fair_delta64
);
1994 tmp64
= div64_64(tmp64
* exec_delta64
, sample_interval64
);
1996 this_load
= (unsigned long)tmp64
;
2000 /* Update our load: */
2001 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2002 unsigned long old_load
, new_load
;
2004 /* scale is effectively 1 << i now, and >> i divides by scale */
2006 old_load
= this_rq
->cpu_load
[i
];
2007 new_load
= this_load
;
2009 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2016 * double_rq_lock - safely lock two runqueues
2018 * Note this does not disable interrupts like task_rq_lock,
2019 * you need to do so manually before calling.
2021 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2022 __acquires(rq1
->lock
)
2023 __acquires(rq2
->lock
)
2025 BUG_ON(!irqs_disabled());
2027 spin_lock(&rq1
->lock
);
2028 __acquire(rq2
->lock
); /* Fake it out ;) */
2031 spin_lock(&rq1
->lock
);
2032 spin_lock(&rq2
->lock
);
2034 spin_lock(&rq2
->lock
);
2035 spin_lock(&rq1
->lock
);
2041 * double_rq_unlock - safely unlock two runqueues
2043 * Note this does not restore interrupts like task_rq_unlock,
2044 * you need to do so manually after calling.
2046 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2047 __releases(rq1
->lock
)
2048 __releases(rq2
->lock
)
2050 spin_unlock(&rq1
->lock
);
2052 spin_unlock(&rq2
->lock
);
2054 __release(rq2
->lock
);
2058 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2060 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2061 __releases(this_rq
->lock
)
2062 __acquires(busiest
->lock
)
2063 __acquires(this_rq
->lock
)
2065 if (unlikely(!irqs_disabled())) {
2066 /* printk() doesn't work good under rq->lock */
2067 spin_unlock(&this_rq
->lock
);
2070 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2071 if (busiest
< this_rq
) {
2072 spin_unlock(&this_rq
->lock
);
2073 spin_lock(&busiest
->lock
);
2074 spin_lock(&this_rq
->lock
);
2076 spin_lock(&busiest
->lock
);
2081 * If dest_cpu is allowed for this process, migrate the task to it.
2082 * This is accomplished by forcing the cpu_allowed mask to only
2083 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2084 * the cpu_allowed mask is restored.
2086 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2088 struct migration_req req
;
2089 unsigned long flags
;
2092 rq
= task_rq_lock(p
, &flags
);
2093 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2094 || unlikely(cpu_is_offline(dest_cpu
)))
2097 /* force the process onto the specified CPU */
2098 if (migrate_task(p
, dest_cpu
, &req
)) {
2099 /* Need to wait for migration thread (might exit: take ref). */
2100 struct task_struct
*mt
= rq
->migration_thread
;
2102 get_task_struct(mt
);
2103 task_rq_unlock(rq
, &flags
);
2104 wake_up_process(mt
);
2105 put_task_struct(mt
);
2106 wait_for_completion(&req
.done
);
2111 task_rq_unlock(rq
, &flags
);
2115 * sched_exec - execve() is a valuable balancing opportunity, because at
2116 * this point the task has the smallest effective memory and cache footprint.
2118 void sched_exec(void)
2120 int new_cpu
, this_cpu
= get_cpu();
2121 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2123 if (new_cpu
!= this_cpu
)
2124 sched_migrate_task(current
, new_cpu
);
2128 * pull_task - move a task from a remote runqueue to the local runqueue.
2129 * Both runqueues must be locked.
2131 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2132 struct rq
*this_rq
, int this_cpu
)
2134 update_rq_clock(src_rq
);
2135 deactivate_task(src_rq
, p
, 0, src_rq
->clock
);
2136 set_task_cpu(p
, this_cpu
);
2137 activate_task(this_rq
, p
, 0);
2139 * Note that idle threads have a prio of MAX_PRIO, for this test
2140 * to be always true for them.
2142 check_preempt_curr(this_rq
, p
);
2146 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2149 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2150 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2154 * We do not migrate tasks that are:
2155 * 1) running (obviously), or
2156 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2157 * 3) are cache-hot on their current CPU.
2159 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2163 if (task_running(rq
, p
))
2167 * Aggressive migration if too many balance attempts have failed:
2169 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2175 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2176 unsigned long max_nr_move
, unsigned long max_load_move
,
2177 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2178 int *all_pinned
, unsigned long *load_moved
,
2179 int *this_best_prio
, struct rq_iterator
*iterator
)
2181 int pulled
= 0, pinned
= 0, skip_for_load
;
2182 struct task_struct
*p
;
2183 long rem_load_move
= max_load_move
;
2185 if (max_nr_move
== 0 || max_load_move
== 0)
2191 * Start the load-balancing iterator:
2193 p
= iterator
->start(iterator
->arg
);
2198 * To help distribute high priority tasks accross CPUs we don't
2199 * skip a task if it will be the highest priority task (i.e. smallest
2200 * prio value) on its new queue regardless of its load weight
2202 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2203 SCHED_LOAD_SCALE_FUZZ
;
2204 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2205 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2206 p
= iterator
->next(iterator
->arg
);
2210 pull_task(busiest
, p
, this_rq
, this_cpu
);
2212 rem_load_move
-= p
->se
.load
.weight
;
2215 * We only want to steal up to the prescribed number of tasks
2216 * and the prescribed amount of weighted load.
2218 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2219 if (p
->prio
< *this_best_prio
)
2220 *this_best_prio
= p
->prio
;
2221 p
= iterator
->next(iterator
->arg
);
2226 * Right now, this is the only place pull_task() is called,
2227 * so we can safely collect pull_task() stats here rather than
2228 * inside pull_task().
2230 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2233 *all_pinned
= pinned
;
2234 *load_moved
= max_load_move
- rem_load_move
;
2239 * move_tasks tries to move up to max_load_move weighted load from busiest to
2240 * this_rq, as part of a balancing operation within domain "sd".
2241 * Returns 1 if successful and 0 otherwise.
2243 * Called with both runqueues locked.
2245 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2246 unsigned long max_load_move
,
2247 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2250 struct sched_class
*class = sched_class_highest
;
2251 unsigned long total_load_moved
= 0;
2252 int this_best_prio
= this_rq
->curr
->prio
;
2256 class->load_balance(this_rq
, this_cpu
, busiest
,
2257 ULONG_MAX
, max_load_move
- total_load_moved
,
2258 sd
, idle
, all_pinned
, &this_best_prio
);
2259 class = class->next
;
2260 } while (class && max_load_move
> total_load_moved
);
2262 return total_load_moved
> 0;
2266 * move_one_task tries to move exactly one task from busiest to this_rq, as
2267 * part of active balancing operations within "domain".
2268 * Returns 1 if successful and 0 otherwise.
2270 * Called with both runqueues locked.
2272 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2273 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2275 struct sched_class
*class;
2276 int this_best_prio
= MAX_PRIO
;
2278 for (class = sched_class_highest
; class; class = class->next
)
2279 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2280 1, ULONG_MAX
, sd
, idle
, NULL
,
2288 * find_busiest_group finds and returns the busiest CPU group within the
2289 * domain. It calculates and returns the amount of weighted load which
2290 * should be moved to restore balance via the imbalance parameter.
2292 static struct sched_group
*
2293 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2294 unsigned long *imbalance
, enum cpu_idle_type idle
,
2295 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2297 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2298 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2299 unsigned long max_pull
;
2300 unsigned long busiest_load_per_task
, busiest_nr_running
;
2301 unsigned long this_load_per_task
, this_nr_running
;
2303 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2304 int power_savings_balance
= 1;
2305 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2306 unsigned long min_nr_running
= ULONG_MAX
;
2307 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2310 max_load
= this_load
= total_load
= total_pwr
= 0;
2311 busiest_load_per_task
= busiest_nr_running
= 0;
2312 this_load_per_task
= this_nr_running
= 0;
2313 if (idle
== CPU_NOT_IDLE
)
2314 load_idx
= sd
->busy_idx
;
2315 else if (idle
== CPU_NEWLY_IDLE
)
2316 load_idx
= sd
->newidle_idx
;
2318 load_idx
= sd
->idle_idx
;
2321 unsigned long load
, group_capacity
;
2324 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2325 unsigned long sum_nr_running
, sum_weighted_load
;
2327 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2330 balance_cpu
= first_cpu(group
->cpumask
);
2332 /* Tally up the load of all CPUs in the group */
2333 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2335 for_each_cpu_mask(i
, group
->cpumask
) {
2338 if (!cpu_isset(i
, *cpus
))
2343 if (*sd_idle
&& rq
->nr_running
)
2346 /* Bias balancing toward cpus of our domain */
2348 if (idle_cpu(i
) && !first_idle_cpu
) {
2353 load
= target_load(i
, load_idx
);
2355 load
= source_load(i
, load_idx
);
2358 sum_nr_running
+= rq
->nr_running
;
2359 sum_weighted_load
+= weighted_cpuload(i
);
2363 * First idle cpu or the first cpu(busiest) in this sched group
2364 * is eligible for doing load balancing at this and above
2365 * domains. In the newly idle case, we will allow all the cpu's
2366 * to do the newly idle load balance.
2368 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2369 balance_cpu
!= this_cpu
&& balance
) {
2374 total_load
+= avg_load
;
2375 total_pwr
+= group
->__cpu_power
;
2377 /* Adjust by relative CPU power of the group */
2378 avg_load
= sg_div_cpu_power(group
,
2379 avg_load
* SCHED_LOAD_SCALE
);
2381 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2384 this_load
= avg_load
;
2386 this_nr_running
= sum_nr_running
;
2387 this_load_per_task
= sum_weighted_load
;
2388 } else if (avg_load
> max_load
&&
2389 sum_nr_running
> group_capacity
) {
2390 max_load
= avg_load
;
2392 busiest_nr_running
= sum_nr_running
;
2393 busiest_load_per_task
= sum_weighted_load
;
2396 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2398 * Busy processors will not participate in power savings
2401 if (idle
== CPU_NOT_IDLE
||
2402 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2406 * If the local group is idle or completely loaded
2407 * no need to do power savings balance at this domain
2409 if (local_group
&& (this_nr_running
>= group_capacity
||
2411 power_savings_balance
= 0;
2414 * If a group is already running at full capacity or idle,
2415 * don't include that group in power savings calculations
2417 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2422 * Calculate the group which has the least non-idle load.
2423 * This is the group from where we need to pick up the load
2426 if ((sum_nr_running
< min_nr_running
) ||
2427 (sum_nr_running
== min_nr_running
&&
2428 first_cpu(group
->cpumask
) <
2429 first_cpu(group_min
->cpumask
))) {
2431 min_nr_running
= sum_nr_running
;
2432 min_load_per_task
= sum_weighted_load
/
2437 * Calculate the group which is almost near its
2438 * capacity but still has some space to pick up some load
2439 * from other group and save more power
2441 if (sum_nr_running
<= group_capacity
- 1) {
2442 if (sum_nr_running
> leader_nr_running
||
2443 (sum_nr_running
== leader_nr_running
&&
2444 first_cpu(group
->cpumask
) >
2445 first_cpu(group_leader
->cpumask
))) {
2446 group_leader
= group
;
2447 leader_nr_running
= sum_nr_running
;
2452 group
= group
->next
;
2453 } while (group
!= sd
->groups
);
2455 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2458 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2460 if (this_load
>= avg_load
||
2461 100*max_load
<= sd
->imbalance_pct
*this_load
)
2464 busiest_load_per_task
/= busiest_nr_running
;
2466 * We're trying to get all the cpus to the average_load, so we don't
2467 * want to push ourselves above the average load, nor do we wish to
2468 * reduce the max loaded cpu below the average load, as either of these
2469 * actions would just result in more rebalancing later, and ping-pong
2470 * tasks around. Thus we look for the minimum possible imbalance.
2471 * Negative imbalances (*we* are more loaded than anyone else) will
2472 * be counted as no imbalance for these purposes -- we can't fix that
2473 * by pulling tasks to us. Be careful of negative numbers as they'll
2474 * appear as very large values with unsigned longs.
2476 if (max_load
<= busiest_load_per_task
)
2480 * In the presence of smp nice balancing, certain scenarios can have
2481 * max load less than avg load(as we skip the groups at or below
2482 * its cpu_power, while calculating max_load..)
2484 if (max_load
< avg_load
) {
2486 goto small_imbalance
;
2489 /* Don't want to pull so many tasks that a group would go idle */
2490 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2492 /* How much load to actually move to equalise the imbalance */
2493 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2494 (avg_load
- this_load
) * this->__cpu_power
)
2498 * if *imbalance is less than the average load per runnable task
2499 * there is no gaurantee that any tasks will be moved so we'll have
2500 * a think about bumping its value to force at least one task to be
2503 if (*imbalance
+ SCHED_LOAD_SCALE_FUZZ
< busiest_load_per_task
/2) {
2504 unsigned long tmp
, pwr_now
, pwr_move
;
2508 pwr_move
= pwr_now
= 0;
2510 if (this_nr_running
) {
2511 this_load_per_task
/= this_nr_running
;
2512 if (busiest_load_per_task
> this_load_per_task
)
2515 this_load_per_task
= SCHED_LOAD_SCALE
;
2517 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2518 busiest_load_per_task
* imbn
) {
2519 *imbalance
= busiest_load_per_task
;
2524 * OK, we don't have enough imbalance to justify moving tasks,
2525 * however we may be able to increase total CPU power used by
2529 pwr_now
+= busiest
->__cpu_power
*
2530 min(busiest_load_per_task
, max_load
);
2531 pwr_now
+= this->__cpu_power
*
2532 min(this_load_per_task
, this_load
);
2533 pwr_now
/= SCHED_LOAD_SCALE
;
2535 /* Amount of load we'd subtract */
2536 tmp
= sg_div_cpu_power(busiest
,
2537 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2539 pwr_move
+= busiest
->__cpu_power
*
2540 min(busiest_load_per_task
, max_load
- tmp
);
2542 /* Amount of load we'd add */
2543 if (max_load
* busiest
->__cpu_power
<
2544 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2545 tmp
= sg_div_cpu_power(this,
2546 max_load
* busiest
->__cpu_power
);
2548 tmp
= sg_div_cpu_power(this,
2549 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2550 pwr_move
+= this->__cpu_power
*
2551 min(this_load_per_task
, this_load
+ tmp
);
2552 pwr_move
/= SCHED_LOAD_SCALE
;
2554 /* Move if we gain throughput */
2555 if (pwr_move
<= pwr_now
)
2558 *imbalance
= busiest_load_per_task
;
2564 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2565 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2568 if (this == group_leader
&& group_leader
!= group_min
) {
2569 *imbalance
= min_load_per_task
;
2579 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2582 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2583 unsigned long imbalance
, cpumask_t
*cpus
)
2585 struct rq
*busiest
= NULL
, *rq
;
2586 unsigned long max_load
= 0;
2589 for_each_cpu_mask(i
, group
->cpumask
) {
2592 if (!cpu_isset(i
, *cpus
))
2596 wl
= weighted_cpuload(i
);
2598 if (rq
->nr_running
== 1 && wl
> imbalance
)
2601 if (wl
> max_load
) {
2611 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2612 * so long as it is large enough.
2614 #define MAX_PINNED_INTERVAL 512
2617 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2618 * tasks if there is an imbalance.
2620 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2621 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2624 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2625 struct sched_group
*group
;
2626 unsigned long imbalance
;
2628 cpumask_t cpus
= CPU_MASK_ALL
;
2629 unsigned long flags
;
2632 * When power savings policy is enabled for the parent domain, idle
2633 * sibling can pick up load irrespective of busy siblings. In this case,
2634 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2635 * portraying it as CPU_NOT_IDLE.
2637 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2638 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2641 schedstat_inc(sd
, lb_cnt
[idle
]);
2644 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2651 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2655 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2657 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2661 BUG_ON(busiest
== this_rq
);
2663 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2666 if (busiest
->nr_running
> 1) {
2668 * Attempt to move tasks. If find_busiest_group has found
2669 * an imbalance but busiest->nr_running <= 1, the group is
2670 * still unbalanced. ld_moved simply stays zero, so it is
2671 * correctly treated as an imbalance.
2673 local_irq_save(flags
);
2674 double_rq_lock(this_rq
, busiest
);
2675 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2676 imbalance
, sd
, idle
, &all_pinned
);
2677 double_rq_unlock(this_rq
, busiest
);
2678 local_irq_restore(flags
);
2681 * some other cpu did the load balance for us.
2683 if (ld_moved
&& this_cpu
!= smp_processor_id())
2684 resched_cpu(this_cpu
);
2686 /* All tasks on this runqueue were pinned by CPU affinity */
2687 if (unlikely(all_pinned
)) {
2688 cpu_clear(cpu_of(busiest
), cpus
);
2689 if (!cpus_empty(cpus
))
2696 schedstat_inc(sd
, lb_failed
[idle
]);
2697 sd
->nr_balance_failed
++;
2699 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2701 spin_lock_irqsave(&busiest
->lock
, flags
);
2703 /* don't kick the migration_thread, if the curr
2704 * task on busiest cpu can't be moved to this_cpu
2706 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2707 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2709 goto out_one_pinned
;
2712 if (!busiest
->active_balance
) {
2713 busiest
->active_balance
= 1;
2714 busiest
->push_cpu
= this_cpu
;
2717 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2719 wake_up_process(busiest
->migration_thread
);
2722 * We've kicked active balancing, reset the failure
2725 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2728 sd
->nr_balance_failed
= 0;
2730 if (likely(!active_balance
)) {
2731 /* We were unbalanced, so reset the balancing interval */
2732 sd
->balance_interval
= sd
->min_interval
;
2735 * If we've begun active balancing, start to back off. This
2736 * case may not be covered by the all_pinned logic if there
2737 * is only 1 task on the busy runqueue (because we don't call
2740 if (sd
->balance_interval
< sd
->max_interval
)
2741 sd
->balance_interval
*= 2;
2744 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2745 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2750 schedstat_inc(sd
, lb_balanced
[idle
]);
2752 sd
->nr_balance_failed
= 0;
2755 /* tune up the balancing interval */
2756 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2757 (sd
->balance_interval
< sd
->max_interval
))
2758 sd
->balance_interval
*= 2;
2760 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2761 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2767 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2768 * tasks if there is an imbalance.
2770 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2771 * this_rq is locked.
2774 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2776 struct sched_group
*group
;
2777 struct rq
*busiest
= NULL
;
2778 unsigned long imbalance
;
2782 cpumask_t cpus
= CPU_MASK_ALL
;
2785 * When power savings policy is enabled for the parent domain, idle
2786 * sibling can pick up load irrespective of busy siblings. In this case,
2787 * let the state of idle sibling percolate up as IDLE, instead of
2788 * portraying it as CPU_NOT_IDLE.
2790 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2791 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2794 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2796 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2797 &sd_idle
, &cpus
, NULL
);
2799 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2803 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2806 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2810 BUG_ON(busiest
== this_rq
);
2812 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2815 if (busiest
->nr_running
> 1) {
2816 /* Attempt to move tasks */
2817 double_lock_balance(this_rq
, busiest
);
2818 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2819 imbalance
, sd
, CPU_NEWLY_IDLE
,
2821 spin_unlock(&busiest
->lock
);
2823 if (unlikely(all_pinned
)) {
2824 cpu_clear(cpu_of(busiest
), cpus
);
2825 if (!cpus_empty(cpus
))
2831 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2832 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2833 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2836 sd
->nr_balance_failed
= 0;
2841 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2842 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2843 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2845 sd
->nr_balance_failed
= 0;
2851 * idle_balance is called by schedule() if this_cpu is about to become
2852 * idle. Attempts to pull tasks from other CPUs.
2854 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2856 struct sched_domain
*sd
;
2857 int pulled_task
= -1;
2858 unsigned long next_balance
= jiffies
+ HZ
;
2860 for_each_domain(this_cpu
, sd
) {
2861 unsigned long interval
;
2863 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2866 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2867 /* If we've pulled tasks over stop searching: */
2868 pulled_task
= load_balance_newidle(this_cpu
,
2871 interval
= msecs_to_jiffies(sd
->balance_interval
);
2872 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2873 next_balance
= sd
->last_balance
+ interval
;
2877 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2879 * We are going idle. next_balance may be set based on
2880 * a busy processor. So reset next_balance.
2882 this_rq
->next_balance
= next_balance
;
2887 * active_load_balance is run by migration threads. It pushes running tasks
2888 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2889 * running on each physical CPU where possible, and avoids physical /
2890 * logical imbalances.
2892 * Called with busiest_rq locked.
2894 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2896 int target_cpu
= busiest_rq
->push_cpu
;
2897 struct sched_domain
*sd
;
2898 struct rq
*target_rq
;
2900 /* Is there any task to move? */
2901 if (busiest_rq
->nr_running
<= 1)
2904 target_rq
= cpu_rq(target_cpu
);
2907 * This condition is "impossible", if it occurs
2908 * we need to fix it. Originally reported by
2909 * Bjorn Helgaas on a 128-cpu setup.
2911 BUG_ON(busiest_rq
== target_rq
);
2913 /* move a task from busiest_rq to target_rq */
2914 double_lock_balance(busiest_rq
, target_rq
);
2916 /* Search for an sd spanning us and the target CPU. */
2917 for_each_domain(target_cpu
, sd
) {
2918 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2919 cpu_isset(busiest_cpu
, sd
->span
))
2924 schedstat_inc(sd
, alb_cnt
);
2926 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2928 schedstat_inc(sd
, alb_pushed
);
2930 schedstat_inc(sd
, alb_failed
);
2932 spin_unlock(&target_rq
->lock
);
2937 atomic_t load_balancer
;
2939 } nohz ____cacheline_aligned
= {
2940 .load_balancer
= ATOMIC_INIT(-1),
2941 .cpu_mask
= CPU_MASK_NONE
,
2945 * This routine will try to nominate the ilb (idle load balancing)
2946 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2947 * load balancing on behalf of all those cpus. If all the cpus in the system
2948 * go into this tickless mode, then there will be no ilb owner (as there is
2949 * no need for one) and all the cpus will sleep till the next wakeup event
2952 * For the ilb owner, tick is not stopped. And this tick will be used
2953 * for idle load balancing. ilb owner will still be part of
2956 * While stopping the tick, this cpu will become the ilb owner if there
2957 * is no other owner. And will be the owner till that cpu becomes busy
2958 * or if all cpus in the system stop their ticks at which point
2959 * there is no need for ilb owner.
2961 * When the ilb owner becomes busy, it nominates another owner, during the
2962 * next busy scheduler_tick()
2964 int select_nohz_load_balancer(int stop_tick
)
2966 int cpu
= smp_processor_id();
2969 cpu_set(cpu
, nohz
.cpu_mask
);
2970 cpu_rq(cpu
)->in_nohz_recently
= 1;
2973 * If we are going offline and still the leader, give up!
2975 if (cpu_is_offline(cpu
) &&
2976 atomic_read(&nohz
.load_balancer
) == cpu
) {
2977 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2982 /* time for ilb owner also to sleep */
2983 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2984 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2985 atomic_set(&nohz
.load_balancer
, -1);
2989 if (atomic_read(&nohz
.load_balancer
) == -1) {
2990 /* make me the ilb owner */
2991 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2993 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2996 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2999 cpu_clear(cpu
, nohz
.cpu_mask
);
3001 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3002 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3009 static DEFINE_SPINLOCK(balancing
);
3012 * It checks each scheduling domain to see if it is due to be balanced,
3013 * and initiates a balancing operation if so.
3015 * Balancing parameters are set up in arch_init_sched_domains.
3017 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3020 struct rq
*rq
= cpu_rq(cpu
);
3021 unsigned long interval
;
3022 struct sched_domain
*sd
;
3023 /* Earliest time when we have to do rebalance again */
3024 unsigned long next_balance
= jiffies
+ 60*HZ
;
3026 for_each_domain(cpu
, sd
) {
3027 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3030 interval
= sd
->balance_interval
;
3031 if (idle
!= CPU_IDLE
)
3032 interval
*= sd
->busy_factor
;
3034 /* scale ms to jiffies */
3035 interval
= msecs_to_jiffies(interval
);
3036 if (unlikely(!interval
))
3038 if (interval
> HZ
*NR_CPUS
/10)
3039 interval
= HZ
*NR_CPUS
/10;
3042 if (sd
->flags
& SD_SERIALIZE
) {
3043 if (!spin_trylock(&balancing
))
3047 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3048 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3050 * We've pulled tasks over so either we're no
3051 * longer idle, or one of our SMT siblings is
3054 idle
= CPU_NOT_IDLE
;
3056 sd
->last_balance
= jiffies
;
3058 if (sd
->flags
& SD_SERIALIZE
)
3059 spin_unlock(&balancing
);
3061 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3062 next_balance
= sd
->last_balance
+ interval
;
3065 * Stop the load balance at this level. There is another
3066 * CPU in our sched group which is doing load balancing more
3072 rq
->next_balance
= next_balance
;
3076 * run_rebalance_domains is triggered when needed from the scheduler tick.
3077 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3078 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3080 static void run_rebalance_domains(struct softirq_action
*h
)
3082 int this_cpu
= smp_processor_id();
3083 struct rq
*this_rq
= cpu_rq(this_cpu
);
3084 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3085 CPU_IDLE
: CPU_NOT_IDLE
;
3087 rebalance_domains(this_cpu
, idle
);
3091 * If this cpu is the owner for idle load balancing, then do the
3092 * balancing on behalf of the other idle cpus whose ticks are
3095 if (this_rq
->idle_at_tick
&&
3096 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3097 cpumask_t cpus
= nohz
.cpu_mask
;
3101 cpu_clear(this_cpu
, cpus
);
3102 for_each_cpu_mask(balance_cpu
, cpus
) {
3104 * If this cpu gets work to do, stop the load balancing
3105 * work being done for other cpus. Next load
3106 * balancing owner will pick it up.
3111 rebalance_domains(balance_cpu
, SCHED_IDLE
);
3113 rq
= cpu_rq(balance_cpu
);
3114 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3115 this_rq
->next_balance
= rq
->next_balance
;
3122 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3124 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3125 * idle load balancing owner or decide to stop the periodic load balancing,
3126 * if the whole system is idle.
3128 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3132 * If we were in the nohz mode recently and busy at the current
3133 * scheduler tick, then check if we need to nominate new idle
3136 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3137 rq
->in_nohz_recently
= 0;
3139 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3140 cpu_clear(cpu
, nohz
.cpu_mask
);
3141 atomic_set(&nohz
.load_balancer
, -1);
3144 if (atomic_read(&nohz
.load_balancer
) == -1) {
3146 * simple selection for now: Nominate the
3147 * first cpu in the nohz list to be the next
3150 * TBD: Traverse the sched domains and nominate
3151 * the nearest cpu in the nohz.cpu_mask.
3153 int ilb
= first_cpu(nohz
.cpu_mask
);
3161 * If this cpu is idle and doing idle load balancing for all the
3162 * cpus with ticks stopped, is it time for that to stop?
3164 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3165 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3171 * If this cpu is idle and the idle load balancing is done by
3172 * someone else, then no need raise the SCHED_SOFTIRQ
3174 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3175 cpu_isset(cpu
, nohz
.cpu_mask
))
3178 if (time_after_eq(jiffies
, rq
->next_balance
))
3179 raise_softirq(SCHED_SOFTIRQ
);
3182 #else /* CONFIG_SMP */
3185 * on UP we do not need to balance between CPUs:
3187 static inline void idle_balance(int cpu
, struct rq
*rq
)
3191 /* Avoid "used but not defined" warning on UP */
3192 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3193 unsigned long max_nr_move
, unsigned long max_load_move
,
3194 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3195 int *all_pinned
, unsigned long *load_moved
,
3196 int *this_best_prio
, struct rq_iterator
*iterator
)
3205 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3207 EXPORT_PER_CPU_SYMBOL(kstat
);
3210 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3211 * that have not yet been banked in case the task is currently running.
3213 unsigned long long task_sched_runtime(struct task_struct
*p
)
3215 unsigned long flags
;
3219 rq
= task_rq_lock(p
, &flags
);
3220 ns
= p
->se
.sum_exec_runtime
;
3221 if (rq
->curr
== p
) {
3222 update_rq_clock(rq
);
3223 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3224 if ((s64
)delta_exec
> 0)
3227 task_rq_unlock(rq
, &flags
);
3233 * Account user cpu time to a process.
3234 * @p: the process that the cpu time gets accounted to
3235 * @hardirq_offset: the offset to subtract from hardirq_count()
3236 * @cputime: the cpu time spent in user space since the last update
3238 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3240 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3243 p
->utime
= cputime_add(p
->utime
, cputime
);
3245 /* Add user time to cpustat. */
3246 tmp
= cputime_to_cputime64(cputime
);
3247 if (TASK_NICE(p
) > 0)
3248 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3250 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3254 * Account system cpu time to a process.
3255 * @p: the process that the cpu time gets accounted to
3256 * @hardirq_offset: the offset to subtract from hardirq_count()
3257 * @cputime: the cpu time spent in kernel space since the last update
3259 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3262 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3263 struct rq
*rq
= this_rq();
3266 p
->stime
= cputime_add(p
->stime
, cputime
);
3268 /* Add system time to cpustat. */
3269 tmp
= cputime_to_cputime64(cputime
);
3270 if (hardirq_count() - hardirq_offset
)
3271 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3272 else if (softirq_count())
3273 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3274 else if (p
!= rq
->idle
)
3275 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3276 else if (atomic_read(&rq
->nr_iowait
) > 0)
3277 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3279 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3280 /* Account for system time used */
3281 acct_update_integrals(p
);
3285 * Account for involuntary wait time.
3286 * @p: the process from which the cpu time has been stolen
3287 * @steal: the cpu time spent in involuntary wait
3289 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3291 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3292 cputime64_t tmp
= cputime_to_cputime64(steal
);
3293 struct rq
*rq
= this_rq();
3295 if (p
== rq
->idle
) {
3296 p
->stime
= cputime_add(p
->stime
, steal
);
3297 if (atomic_read(&rq
->nr_iowait
) > 0)
3298 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3300 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3302 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3306 * This function gets called by the timer code, with HZ frequency.
3307 * We call it with interrupts disabled.
3309 * It also gets called by the fork code, when changing the parent's
3312 void scheduler_tick(void)
3314 int cpu
= smp_processor_id();
3315 struct rq
*rq
= cpu_rq(cpu
);
3316 struct task_struct
*curr
= rq
->curr
;
3318 spin_lock(&rq
->lock
);
3319 update_cpu_load(rq
);
3320 if (curr
!= rq
->idle
) /* FIXME: needed? */
3321 curr
->sched_class
->task_tick(rq
, curr
);
3322 spin_unlock(&rq
->lock
);
3325 rq
->idle_at_tick
= idle_cpu(cpu
);
3326 trigger_load_balance(rq
, cpu
);
3330 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3332 void fastcall
add_preempt_count(int val
)
3337 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3339 preempt_count() += val
;
3341 * Spinlock count overflowing soon?
3343 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3346 EXPORT_SYMBOL(add_preempt_count
);
3348 void fastcall
sub_preempt_count(int val
)
3353 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3356 * Is the spinlock portion underflowing?
3358 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3359 !(preempt_count() & PREEMPT_MASK
)))
3362 preempt_count() -= val
;
3364 EXPORT_SYMBOL(sub_preempt_count
);
3369 * Print scheduling while atomic bug:
3371 static noinline
void __schedule_bug(struct task_struct
*prev
)
3373 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3374 prev
->comm
, preempt_count(), prev
->pid
);
3375 debug_show_held_locks(prev
);
3376 if (irqs_disabled())
3377 print_irqtrace_events(prev
);
3382 * Various schedule()-time debugging checks and statistics:
3384 static inline void schedule_debug(struct task_struct
*prev
)
3387 * Test if we are atomic. Since do_exit() needs to call into
3388 * schedule() atomically, we ignore that path for now.
3389 * Otherwise, whine if we are scheduling when we should not be.
3391 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3392 __schedule_bug(prev
);
3394 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3396 schedstat_inc(this_rq(), sched_cnt
);
3400 * Pick up the highest-prio task:
3402 static inline struct task_struct
*
3403 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3405 struct sched_class
*class;
3406 struct task_struct
*p
;
3409 * Optimization: we know that if all tasks are in
3410 * the fair class we can call that function directly:
3412 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3413 p
= fair_sched_class
.pick_next_task(rq
);
3418 class = sched_class_highest
;
3420 p
= class->pick_next_task(rq
);
3424 * Will never be NULL as the idle class always
3425 * returns a non-NULL p:
3427 class = class->next
;
3432 * schedule() is the main scheduler function.
3434 asmlinkage
void __sched
schedule(void)
3436 struct task_struct
*prev
, *next
;
3444 cpu
= smp_processor_id();
3448 switch_count
= &prev
->nivcsw
;
3450 release_kernel_lock(prev
);
3451 need_resched_nonpreemptible
:
3453 schedule_debug(prev
);
3455 spin_lock_irq(&rq
->lock
);
3456 clear_tsk_need_resched(prev
);
3457 __update_rq_clock(rq
);
3460 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3461 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3462 unlikely(signal_pending(prev
)))) {
3463 prev
->state
= TASK_RUNNING
;
3465 deactivate_task(rq
, prev
, 1, now
);
3467 switch_count
= &prev
->nvcsw
;
3470 if (unlikely(!rq
->nr_running
))
3471 idle_balance(cpu
, rq
);
3473 prev
->sched_class
->put_prev_task(rq
, prev
, now
);
3474 next
= pick_next_task(rq
, prev
);
3476 sched_info_switch(prev
, next
);
3478 if (likely(prev
!= next
)) {
3483 context_switch(rq
, prev
, next
); /* unlocks the rq */
3485 spin_unlock_irq(&rq
->lock
);
3487 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3488 cpu
= smp_processor_id();
3490 goto need_resched_nonpreemptible
;
3492 preempt_enable_no_resched();
3493 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3496 EXPORT_SYMBOL(schedule
);
3498 #ifdef CONFIG_PREEMPT
3500 * this is the entry point to schedule() from in-kernel preemption
3501 * off of preempt_enable. Kernel preemptions off return from interrupt
3502 * occur there and call schedule directly.
3504 asmlinkage
void __sched
preempt_schedule(void)
3506 struct thread_info
*ti
= current_thread_info();
3507 #ifdef CONFIG_PREEMPT_BKL
3508 struct task_struct
*task
= current
;
3509 int saved_lock_depth
;
3512 * If there is a non-zero preempt_count or interrupts are disabled,
3513 * we do not want to preempt the current task. Just return..
3515 if (likely(ti
->preempt_count
|| irqs_disabled()))
3519 add_preempt_count(PREEMPT_ACTIVE
);
3521 * We keep the big kernel semaphore locked, but we
3522 * clear ->lock_depth so that schedule() doesnt
3523 * auto-release the semaphore:
3525 #ifdef CONFIG_PREEMPT_BKL
3526 saved_lock_depth
= task
->lock_depth
;
3527 task
->lock_depth
= -1;
3530 #ifdef CONFIG_PREEMPT_BKL
3531 task
->lock_depth
= saved_lock_depth
;
3533 sub_preempt_count(PREEMPT_ACTIVE
);
3535 /* we could miss a preemption opportunity between schedule and now */
3537 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3540 EXPORT_SYMBOL(preempt_schedule
);
3543 * this is the entry point to schedule() from kernel preemption
3544 * off of irq context.
3545 * Note, that this is called and return with irqs disabled. This will
3546 * protect us against recursive calling from irq.
3548 asmlinkage
void __sched
preempt_schedule_irq(void)
3550 struct thread_info
*ti
= current_thread_info();
3551 #ifdef CONFIG_PREEMPT_BKL
3552 struct task_struct
*task
= current
;
3553 int saved_lock_depth
;
3555 /* Catch callers which need to be fixed */
3556 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3559 add_preempt_count(PREEMPT_ACTIVE
);
3561 * We keep the big kernel semaphore locked, but we
3562 * clear ->lock_depth so that schedule() doesnt
3563 * auto-release the semaphore:
3565 #ifdef CONFIG_PREEMPT_BKL
3566 saved_lock_depth
= task
->lock_depth
;
3567 task
->lock_depth
= -1;
3571 local_irq_disable();
3572 #ifdef CONFIG_PREEMPT_BKL
3573 task
->lock_depth
= saved_lock_depth
;
3575 sub_preempt_count(PREEMPT_ACTIVE
);
3577 /* we could miss a preemption opportunity between schedule and now */
3579 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3583 #endif /* CONFIG_PREEMPT */
3585 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3588 return try_to_wake_up(curr
->private, mode
, sync
);
3590 EXPORT_SYMBOL(default_wake_function
);
3593 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3594 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3595 * number) then we wake all the non-exclusive tasks and one exclusive task.
3597 * There are circumstances in which we can try to wake a task which has already
3598 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3599 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3601 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3602 int nr_exclusive
, int sync
, void *key
)
3604 struct list_head
*tmp
, *next
;
3606 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3607 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3608 unsigned flags
= curr
->flags
;
3610 if (curr
->func(curr
, mode
, sync
, key
) &&
3611 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3617 * __wake_up - wake up threads blocked on a waitqueue.
3619 * @mode: which threads
3620 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3621 * @key: is directly passed to the wakeup function
3623 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3624 int nr_exclusive
, void *key
)
3626 unsigned long flags
;
3628 spin_lock_irqsave(&q
->lock
, flags
);
3629 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3630 spin_unlock_irqrestore(&q
->lock
, flags
);
3632 EXPORT_SYMBOL(__wake_up
);
3635 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3637 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3639 __wake_up_common(q
, mode
, 1, 0, NULL
);
3643 * __wake_up_sync - wake up threads blocked on a waitqueue.
3645 * @mode: which threads
3646 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3648 * The sync wakeup differs that the waker knows that it will schedule
3649 * away soon, so while the target thread will be woken up, it will not
3650 * be migrated to another CPU - ie. the two threads are 'synchronized'
3651 * with each other. This can prevent needless bouncing between CPUs.
3653 * On UP it can prevent extra preemption.
3656 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3658 unsigned long flags
;
3664 if (unlikely(!nr_exclusive
))
3667 spin_lock_irqsave(&q
->lock
, flags
);
3668 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3669 spin_unlock_irqrestore(&q
->lock
, flags
);
3671 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3673 void fastcall
complete(struct completion
*x
)
3675 unsigned long flags
;
3677 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3679 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3681 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3683 EXPORT_SYMBOL(complete
);
3685 void fastcall
complete_all(struct completion
*x
)
3687 unsigned long flags
;
3689 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3690 x
->done
+= UINT_MAX
/2;
3691 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3693 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3695 EXPORT_SYMBOL(complete_all
);
3697 void fastcall __sched
wait_for_completion(struct completion
*x
)
3701 spin_lock_irq(&x
->wait
.lock
);
3703 DECLARE_WAITQUEUE(wait
, current
);
3705 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3706 __add_wait_queue_tail(&x
->wait
, &wait
);
3708 __set_current_state(TASK_UNINTERRUPTIBLE
);
3709 spin_unlock_irq(&x
->wait
.lock
);
3711 spin_lock_irq(&x
->wait
.lock
);
3713 __remove_wait_queue(&x
->wait
, &wait
);
3716 spin_unlock_irq(&x
->wait
.lock
);
3718 EXPORT_SYMBOL(wait_for_completion
);
3720 unsigned long fastcall __sched
3721 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3725 spin_lock_irq(&x
->wait
.lock
);
3727 DECLARE_WAITQUEUE(wait
, current
);
3729 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3730 __add_wait_queue_tail(&x
->wait
, &wait
);
3732 __set_current_state(TASK_UNINTERRUPTIBLE
);
3733 spin_unlock_irq(&x
->wait
.lock
);
3734 timeout
= schedule_timeout(timeout
);
3735 spin_lock_irq(&x
->wait
.lock
);
3737 __remove_wait_queue(&x
->wait
, &wait
);
3741 __remove_wait_queue(&x
->wait
, &wait
);
3745 spin_unlock_irq(&x
->wait
.lock
);
3748 EXPORT_SYMBOL(wait_for_completion_timeout
);
3750 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3756 spin_lock_irq(&x
->wait
.lock
);
3758 DECLARE_WAITQUEUE(wait
, current
);
3760 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3761 __add_wait_queue_tail(&x
->wait
, &wait
);
3763 if (signal_pending(current
)) {
3765 __remove_wait_queue(&x
->wait
, &wait
);
3768 __set_current_state(TASK_INTERRUPTIBLE
);
3769 spin_unlock_irq(&x
->wait
.lock
);
3771 spin_lock_irq(&x
->wait
.lock
);
3773 __remove_wait_queue(&x
->wait
, &wait
);
3777 spin_unlock_irq(&x
->wait
.lock
);
3781 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3783 unsigned long fastcall __sched
3784 wait_for_completion_interruptible_timeout(struct completion
*x
,
3785 unsigned long timeout
)
3789 spin_lock_irq(&x
->wait
.lock
);
3791 DECLARE_WAITQUEUE(wait
, current
);
3793 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3794 __add_wait_queue_tail(&x
->wait
, &wait
);
3796 if (signal_pending(current
)) {
3797 timeout
= -ERESTARTSYS
;
3798 __remove_wait_queue(&x
->wait
, &wait
);
3801 __set_current_state(TASK_INTERRUPTIBLE
);
3802 spin_unlock_irq(&x
->wait
.lock
);
3803 timeout
= schedule_timeout(timeout
);
3804 spin_lock_irq(&x
->wait
.lock
);
3806 __remove_wait_queue(&x
->wait
, &wait
);
3810 __remove_wait_queue(&x
->wait
, &wait
);
3814 spin_unlock_irq(&x
->wait
.lock
);
3817 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3820 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3822 spin_lock_irqsave(&q
->lock
, *flags
);
3823 __add_wait_queue(q
, wait
);
3824 spin_unlock(&q
->lock
);
3828 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3830 spin_lock_irq(&q
->lock
);
3831 __remove_wait_queue(q
, wait
);
3832 spin_unlock_irqrestore(&q
->lock
, *flags
);
3835 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3837 unsigned long flags
;
3840 init_waitqueue_entry(&wait
, current
);
3842 current
->state
= TASK_INTERRUPTIBLE
;
3844 sleep_on_head(q
, &wait
, &flags
);
3846 sleep_on_tail(q
, &wait
, &flags
);
3848 EXPORT_SYMBOL(interruptible_sleep_on
);
3851 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3853 unsigned long flags
;
3856 init_waitqueue_entry(&wait
, current
);
3858 current
->state
= TASK_INTERRUPTIBLE
;
3860 sleep_on_head(q
, &wait
, &flags
);
3861 timeout
= schedule_timeout(timeout
);
3862 sleep_on_tail(q
, &wait
, &flags
);
3866 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3868 void __sched
sleep_on(wait_queue_head_t
*q
)
3870 unsigned long flags
;
3873 init_waitqueue_entry(&wait
, current
);
3875 current
->state
= TASK_UNINTERRUPTIBLE
;
3877 sleep_on_head(q
, &wait
, &flags
);
3879 sleep_on_tail(q
, &wait
, &flags
);
3881 EXPORT_SYMBOL(sleep_on
);
3883 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3885 unsigned long flags
;
3888 init_waitqueue_entry(&wait
, current
);
3890 current
->state
= TASK_UNINTERRUPTIBLE
;
3892 sleep_on_head(q
, &wait
, &flags
);
3893 timeout
= schedule_timeout(timeout
);
3894 sleep_on_tail(q
, &wait
, &flags
);
3898 EXPORT_SYMBOL(sleep_on_timeout
);
3900 #ifdef CONFIG_RT_MUTEXES
3903 * rt_mutex_setprio - set the current priority of a task
3905 * @prio: prio value (kernel-internal form)
3907 * This function changes the 'effective' priority of a task. It does
3908 * not touch ->normal_prio like __setscheduler().
3910 * Used by the rt_mutex code to implement priority inheritance logic.
3912 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3914 unsigned long flags
;
3919 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3921 rq
= task_rq_lock(p
, &flags
);
3922 update_rq_clock(rq
);
3926 on_rq
= p
->se
.on_rq
;
3928 dequeue_task(rq
, p
, 0, now
);
3931 p
->sched_class
= &rt_sched_class
;
3933 p
->sched_class
= &fair_sched_class
;
3938 enqueue_task(rq
, p
, 0, now
);
3940 * Reschedule if we are currently running on this runqueue and
3941 * our priority decreased, or if we are not currently running on
3942 * this runqueue and our priority is higher than the current's
3944 if (task_running(rq
, p
)) {
3945 if (p
->prio
> oldprio
)
3946 resched_task(rq
->curr
);
3948 check_preempt_curr(rq
, p
);
3951 task_rq_unlock(rq
, &flags
);
3956 void set_user_nice(struct task_struct
*p
, long nice
)
3958 int old_prio
, delta
, on_rq
;
3959 unsigned long flags
;
3963 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3966 * We have to be careful, if called from sys_setpriority(),
3967 * the task might be in the middle of scheduling on another CPU.
3969 rq
= task_rq_lock(p
, &flags
);
3970 update_rq_clock(rq
);
3973 * The RT priorities are set via sched_setscheduler(), but we still
3974 * allow the 'normal' nice value to be set - but as expected
3975 * it wont have any effect on scheduling until the task is
3976 * SCHED_FIFO/SCHED_RR:
3978 if (task_has_rt_policy(p
)) {
3979 p
->static_prio
= NICE_TO_PRIO(nice
);
3982 on_rq
= p
->se
.on_rq
;
3984 dequeue_task(rq
, p
, 0, now
);
3985 dec_load(rq
, p
, now
);
3988 p
->static_prio
= NICE_TO_PRIO(nice
);
3991 p
->prio
= effective_prio(p
);
3992 delta
= p
->prio
- old_prio
;
3995 enqueue_task(rq
, p
, 0, now
);
3996 inc_load(rq
, p
, now
);
3998 * If the task increased its priority or is running and
3999 * lowered its priority, then reschedule its CPU:
4001 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4002 resched_task(rq
->curr
);
4005 task_rq_unlock(rq
, &flags
);
4007 EXPORT_SYMBOL(set_user_nice
);
4010 * can_nice - check if a task can reduce its nice value
4014 int can_nice(const struct task_struct
*p
, const int nice
)
4016 /* convert nice value [19,-20] to rlimit style value [1,40] */
4017 int nice_rlim
= 20 - nice
;
4019 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4020 capable(CAP_SYS_NICE
));
4023 #ifdef __ARCH_WANT_SYS_NICE
4026 * sys_nice - change the priority of the current process.
4027 * @increment: priority increment
4029 * sys_setpriority is a more generic, but much slower function that
4030 * does similar things.
4032 asmlinkage
long sys_nice(int increment
)
4037 * Setpriority might change our priority at the same moment.
4038 * We don't have to worry. Conceptually one call occurs first
4039 * and we have a single winner.
4041 if (increment
< -40)
4046 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4052 if (increment
< 0 && !can_nice(current
, nice
))
4055 retval
= security_task_setnice(current
, nice
);
4059 set_user_nice(current
, nice
);
4066 * task_prio - return the priority value of a given task.
4067 * @p: the task in question.
4069 * This is the priority value as seen by users in /proc.
4070 * RT tasks are offset by -200. Normal tasks are centered
4071 * around 0, value goes from -16 to +15.
4073 int task_prio(const struct task_struct
*p
)
4075 return p
->prio
- MAX_RT_PRIO
;
4079 * task_nice - return the nice value of a given task.
4080 * @p: the task in question.
4082 int task_nice(const struct task_struct
*p
)
4084 return TASK_NICE(p
);
4086 EXPORT_SYMBOL_GPL(task_nice
);
4089 * idle_cpu - is a given cpu idle currently?
4090 * @cpu: the processor in question.
4092 int idle_cpu(int cpu
)
4094 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4098 * idle_task - return the idle task for a given cpu.
4099 * @cpu: the processor in question.
4101 struct task_struct
*idle_task(int cpu
)
4103 return cpu_rq(cpu
)->idle
;
4107 * find_process_by_pid - find a process with a matching PID value.
4108 * @pid: the pid in question.
4110 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4112 return pid
? find_task_by_pid(pid
) : current
;
4115 /* Actually do priority change: must hold rq lock. */
4117 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4119 BUG_ON(p
->se
.on_rq
);
4122 switch (p
->policy
) {
4126 p
->sched_class
= &fair_sched_class
;
4130 p
->sched_class
= &rt_sched_class
;
4134 p
->rt_priority
= prio
;
4135 p
->normal_prio
= normal_prio(p
);
4136 /* we are holding p->pi_lock already */
4137 p
->prio
= rt_mutex_getprio(p
);
4142 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4143 * @p: the task in question.
4144 * @policy: new policy.
4145 * @param: structure containing the new RT priority.
4147 * NOTE that the task may be already dead.
4149 int sched_setscheduler(struct task_struct
*p
, int policy
,
4150 struct sched_param
*param
)
4152 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4153 unsigned long flags
;
4156 /* may grab non-irq protected spin_locks */
4157 BUG_ON(in_interrupt());
4159 /* double check policy once rq lock held */
4161 policy
= oldpolicy
= p
->policy
;
4162 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4163 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4164 policy
!= SCHED_IDLE
)
4167 * Valid priorities for SCHED_FIFO and SCHED_RR are
4168 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4169 * SCHED_BATCH and SCHED_IDLE is 0.
4171 if (param
->sched_priority
< 0 ||
4172 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4173 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4175 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4179 * Allow unprivileged RT tasks to decrease priority:
4181 if (!capable(CAP_SYS_NICE
)) {
4182 if (rt_policy(policy
)) {
4183 unsigned long rlim_rtprio
;
4185 if (!lock_task_sighand(p
, &flags
))
4187 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4188 unlock_task_sighand(p
, &flags
);
4190 /* can't set/change the rt policy */
4191 if (policy
!= p
->policy
&& !rlim_rtprio
)
4194 /* can't increase priority */
4195 if (param
->sched_priority
> p
->rt_priority
&&
4196 param
->sched_priority
> rlim_rtprio
)
4200 * Like positive nice levels, dont allow tasks to
4201 * move out of SCHED_IDLE either:
4203 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4206 /* can't change other user's priorities */
4207 if ((current
->euid
!= p
->euid
) &&
4208 (current
->euid
!= p
->uid
))
4212 retval
= security_task_setscheduler(p
, policy
, param
);
4216 * make sure no PI-waiters arrive (or leave) while we are
4217 * changing the priority of the task:
4219 spin_lock_irqsave(&p
->pi_lock
, flags
);
4221 * To be able to change p->policy safely, the apropriate
4222 * runqueue lock must be held.
4224 rq
= __task_rq_lock(p
);
4225 /* recheck policy now with rq lock held */
4226 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4227 policy
= oldpolicy
= -1;
4228 __task_rq_unlock(rq
);
4229 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4232 on_rq
= p
->se
.on_rq
;
4234 update_rq_clock(rq
);
4235 deactivate_task(rq
, p
, 0, rq
->clock
);
4238 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4240 activate_task(rq
, p
, 0);
4242 * Reschedule if we are currently running on this runqueue and
4243 * our priority decreased, or if we are not currently running on
4244 * this runqueue and our priority is higher than the current's
4246 if (task_running(rq
, p
)) {
4247 if (p
->prio
> oldprio
)
4248 resched_task(rq
->curr
);
4250 check_preempt_curr(rq
, p
);
4253 __task_rq_unlock(rq
);
4254 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4256 rt_mutex_adjust_pi(p
);
4260 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4263 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4265 struct sched_param lparam
;
4266 struct task_struct
*p
;
4269 if (!param
|| pid
< 0)
4271 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4276 p
= find_process_by_pid(pid
);
4278 retval
= sched_setscheduler(p
, policy
, &lparam
);
4285 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4286 * @pid: the pid in question.
4287 * @policy: new policy.
4288 * @param: structure containing the new RT priority.
4290 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4291 struct sched_param __user
*param
)
4293 /* negative values for policy are not valid */
4297 return do_sched_setscheduler(pid
, policy
, param
);
4301 * sys_sched_setparam - set/change the RT priority of a thread
4302 * @pid: the pid in question.
4303 * @param: structure containing the new RT priority.
4305 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4307 return do_sched_setscheduler(pid
, -1, param
);
4311 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4312 * @pid: the pid in question.
4314 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4316 struct task_struct
*p
;
4317 int retval
= -EINVAL
;
4323 read_lock(&tasklist_lock
);
4324 p
= find_process_by_pid(pid
);
4326 retval
= security_task_getscheduler(p
);
4330 read_unlock(&tasklist_lock
);
4337 * sys_sched_getscheduler - get the RT priority of a thread
4338 * @pid: the pid in question.
4339 * @param: structure containing the RT priority.
4341 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4343 struct sched_param lp
;
4344 struct task_struct
*p
;
4345 int retval
= -EINVAL
;
4347 if (!param
|| pid
< 0)
4350 read_lock(&tasklist_lock
);
4351 p
= find_process_by_pid(pid
);
4356 retval
= security_task_getscheduler(p
);
4360 lp
.sched_priority
= p
->rt_priority
;
4361 read_unlock(&tasklist_lock
);
4364 * This one might sleep, we cannot do it with a spinlock held ...
4366 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4372 read_unlock(&tasklist_lock
);
4376 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4378 cpumask_t cpus_allowed
;
4379 struct task_struct
*p
;
4382 mutex_lock(&sched_hotcpu_mutex
);
4383 read_lock(&tasklist_lock
);
4385 p
= find_process_by_pid(pid
);
4387 read_unlock(&tasklist_lock
);
4388 mutex_unlock(&sched_hotcpu_mutex
);
4393 * It is not safe to call set_cpus_allowed with the
4394 * tasklist_lock held. We will bump the task_struct's
4395 * usage count and then drop tasklist_lock.
4398 read_unlock(&tasklist_lock
);
4401 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4402 !capable(CAP_SYS_NICE
))
4405 retval
= security_task_setscheduler(p
, 0, NULL
);
4409 cpus_allowed
= cpuset_cpus_allowed(p
);
4410 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4411 retval
= set_cpus_allowed(p
, new_mask
);
4415 mutex_unlock(&sched_hotcpu_mutex
);
4419 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4420 cpumask_t
*new_mask
)
4422 if (len
< sizeof(cpumask_t
)) {
4423 memset(new_mask
, 0, sizeof(cpumask_t
));
4424 } else if (len
> sizeof(cpumask_t
)) {
4425 len
= sizeof(cpumask_t
);
4427 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4431 * sys_sched_setaffinity - set the cpu affinity of a process
4432 * @pid: pid of the process
4433 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4434 * @user_mask_ptr: user-space pointer to the new cpu mask
4436 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4437 unsigned long __user
*user_mask_ptr
)
4442 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4446 return sched_setaffinity(pid
, new_mask
);
4450 * Represents all cpu's present in the system
4451 * In systems capable of hotplug, this map could dynamically grow
4452 * as new cpu's are detected in the system via any platform specific
4453 * method, such as ACPI for e.g.
4456 cpumask_t cpu_present_map __read_mostly
;
4457 EXPORT_SYMBOL(cpu_present_map
);
4460 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4461 EXPORT_SYMBOL(cpu_online_map
);
4463 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4464 EXPORT_SYMBOL(cpu_possible_map
);
4467 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4469 struct task_struct
*p
;
4472 mutex_lock(&sched_hotcpu_mutex
);
4473 read_lock(&tasklist_lock
);
4476 p
= find_process_by_pid(pid
);
4480 retval
= security_task_getscheduler(p
);
4484 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4487 read_unlock(&tasklist_lock
);
4488 mutex_unlock(&sched_hotcpu_mutex
);
4494 * sys_sched_getaffinity - get the cpu affinity of a process
4495 * @pid: pid of the process
4496 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4497 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4499 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4500 unsigned long __user
*user_mask_ptr
)
4505 if (len
< sizeof(cpumask_t
))
4508 ret
= sched_getaffinity(pid
, &mask
);
4512 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4515 return sizeof(cpumask_t
);
4519 * sys_sched_yield - yield the current processor to other threads.
4521 * This function yields the current CPU to other tasks. If there are no
4522 * other threads running on this CPU then this function will return.
4524 asmlinkage
long sys_sched_yield(void)
4526 struct rq
*rq
= this_rq_lock();
4528 schedstat_inc(rq
, yld_cnt
);
4529 if (unlikely(rq
->nr_running
== 1))
4530 schedstat_inc(rq
, yld_act_empty
);
4532 current
->sched_class
->yield_task(rq
, current
);
4535 * Since we are going to call schedule() anyway, there's
4536 * no need to preempt or enable interrupts:
4538 __release(rq
->lock
);
4539 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4540 _raw_spin_unlock(&rq
->lock
);
4541 preempt_enable_no_resched();
4548 static void __cond_resched(void)
4550 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4551 __might_sleep(__FILE__
, __LINE__
);
4554 * The BKS might be reacquired before we have dropped
4555 * PREEMPT_ACTIVE, which could trigger a second
4556 * cond_resched() call.
4559 add_preempt_count(PREEMPT_ACTIVE
);
4561 sub_preempt_count(PREEMPT_ACTIVE
);
4562 } while (need_resched());
4565 int __sched
cond_resched(void)
4567 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4568 system_state
== SYSTEM_RUNNING
) {
4574 EXPORT_SYMBOL(cond_resched
);
4577 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4578 * call schedule, and on return reacquire the lock.
4580 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4581 * operations here to prevent schedule() from being called twice (once via
4582 * spin_unlock(), once by hand).
4584 int cond_resched_lock(spinlock_t
*lock
)
4588 if (need_lockbreak(lock
)) {
4594 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4595 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4596 _raw_spin_unlock(lock
);
4597 preempt_enable_no_resched();
4604 EXPORT_SYMBOL(cond_resched_lock
);
4606 int __sched
cond_resched_softirq(void)
4608 BUG_ON(!in_softirq());
4610 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4618 EXPORT_SYMBOL(cond_resched_softirq
);
4621 * yield - yield the current processor to other threads.
4623 * This is a shortcut for kernel-space yielding - it marks the
4624 * thread runnable and calls sys_sched_yield().
4626 void __sched
yield(void)
4628 set_current_state(TASK_RUNNING
);
4631 EXPORT_SYMBOL(yield
);
4634 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4635 * that process accounting knows that this is a task in IO wait state.
4637 * But don't do that if it is a deliberate, throttling IO wait (this task
4638 * has set its backing_dev_info: the queue against which it should throttle)
4640 void __sched
io_schedule(void)
4642 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4644 delayacct_blkio_start();
4645 atomic_inc(&rq
->nr_iowait
);
4647 atomic_dec(&rq
->nr_iowait
);
4648 delayacct_blkio_end();
4650 EXPORT_SYMBOL(io_schedule
);
4652 long __sched
io_schedule_timeout(long timeout
)
4654 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4657 delayacct_blkio_start();
4658 atomic_inc(&rq
->nr_iowait
);
4659 ret
= schedule_timeout(timeout
);
4660 atomic_dec(&rq
->nr_iowait
);
4661 delayacct_blkio_end();
4666 * sys_sched_get_priority_max - return maximum RT priority.
4667 * @policy: scheduling class.
4669 * this syscall returns the maximum rt_priority that can be used
4670 * by a given scheduling class.
4672 asmlinkage
long sys_sched_get_priority_max(int policy
)
4679 ret
= MAX_USER_RT_PRIO
-1;
4691 * sys_sched_get_priority_min - return minimum RT priority.
4692 * @policy: scheduling class.
4694 * this syscall returns the minimum rt_priority that can be used
4695 * by a given scheduling class.
4697 asmlinkage
long sys_sched_get_priority_min(int policy
)
4715 * sys_sched_rr_get_interval - return the default timeslice of a process.
4716 * @pid: pid of the process.
4717 * @interval: userspace pointer to the timeslice value.
4719 * this syscall writes the default timeslice value of a given process
4720 * into the user-space timespec buffer. A value of '0' means infinity.
4723 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4725 struct task_struct
*p
;
4726 int retval
= -EINVAL
;
4733 read_lock(&tasklist_lock
);
4734 p
= find_process_by_pid(pid
);
4738 retval
= security_task_getscheduler(p
);
4742 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4743 0 : static_prio_timeslice(p
->static_prio
), &t
);
4744 read_unlock(&tasklist_lock
);
4745 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4749 read_unlock(&tasklist_lock
);
4753 static const char stat_nam
[] = "RSDTtZX";
4755 static void show_task(struct task_struct
*p
)
4757 unsigned long free
= 0;
4760 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4761 printk("%-13.13s %c", p
->comm
,
4762 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4763 #if BITS_PER_LONG == 32
4764 if (state
== TASK_RUNNING
)
4765 printk(" running ");
4767 printk(" %08lx ", thread_saved_pc(p
));
4769 if (state
== TASK_RUNNING
)
4770 printk(" running task ");
4772 printk(" %016lx ", thread_saved_pc(p
));
4774 #ifdef CONFIG_DEBUG_STACK_USAGE
4776 unsigned long *n
= end_of_stack(p
);
4779 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4782 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4784 if (state
!= TASK_RUNNING
)
4785 show_stack(p
, NULL
);
4788 void show_state_filter(unsigned long state_filter
)
4790 struct task_struct
*g
, *p
;
4792 #if BITS_PER_LONG == 32
4794 " task PC stack pid father\n");
4797 " task PC stack pid father\n");
4799 read_lock(&tasklist_lock
);
4800 do_each_thread(g
, p
) {
4802 * reset the NMI-timeout, listing all files on a slow
4803 * console might take alot of time:
4805 touch_nmi_watchdog();
4806 if (!state_filter
|| (p
->state
& state_filter
))
4808 } while_each_thread(g
, p
);
4810 touch_all_softlockup_watchdogs();
4812 #ifdef CONFIG_SCHED_DEBUG
4813 sysrq_sched_debug_show();
4815 read_unlock(&tasklist_lock
);
4817 * Only show locks if all tasks are dumped:
4819 if (state_filter
== -1)
4820 debug_show_all_locks();
4823 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4825 idle
->sched_class
= &idle_sched_class
;
4829 * init_idle - set up an idle thread for a given CPU
4830 * @idle: task in question
4831 * @cpu: cpu the idle task belongs to
4833 * NOTE: this function does not set the idle thread's NEED_RESCHED
4834 * flag, to make booting more robust.
4836 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4838 struct rq
*rq
= cpu_rq(cpu
);
4839 unsigned long flags
;
4842 idle
->se
.exec_start
= sched_clock();
4844 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4845 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4846 __set_task_cpu(idle
, cpu
);
4848 spin_lock_irqsave(&rq
->lock
, flags
);
4849 rq
->curr
= rq
->idle
= idle
;
4850 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4853 spin_unlock_irqrestore(&rq
->lock
, flags
);
4855 /* Set the preempt count _outside_ the spinlocks! */
4856 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4857 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4859 task_thread_info(idle
)->preempt_count
= 0;
4862 * The idle tasks have their own, simple scheduling class:
4864 idle
->sched_class
= &idle_sched_class
;
4868 * In a system that switches off the HZ timer nohz_cpu_mask
4869 * indicates which cpus entered this state. This is used
4870 * in the rcu update to wait only for active cpus. For system
4871 * which do not switch off the HZ timer nohz_cpu_mask should
4872 * always be CPU_MASK_NONE.
4874 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4877 * Increase the granularity value when there are more CPUs,
4878 * because with more CPUs the 'effective latency' as visible
4879 * to users decreases. But the relationship is not linear,
4880 * so pick a second-best guess by going with the log2 of the
4883 * This idea comes from the SD scheduler of Con Kolivas:
4885 static inline void sched_init_granularity(void)
4887 unsigned int factor
= 1 + ilog2(num_online_cpus());
4888 const unsigned long gran_limit
= 100000000;
4890 sysctl_sched_granularity
*= factor
;
4891 if (sysctl_sched_granularity
> gran_limit
)
4892 sysctl_sched_granularity
= gran_limit
;
4894 sysctl_sched_runtime_limit
= sysctl_sched_granularity
* 4;
4895 sysctl_sched_wakeup_granularity
= sysctl_sched_granularity
/ 2;
4900 * This is how migration works:
4902 * 1) we queue a struct migration_req structure in the source CPU's
4903 * runqueue and wake up that CPU's migration thread.
4904 * 2) we down() the locked semaphore => thread blocks.
4905 * 3) migration thread wakes up (implicitly it forces the migrated
4906 * thread off the CPU)
4907 * 4) it gets the migration request and checks whether the migrated
4908 * task is still in the wrong runqueue.
4909 * 5) if it's in the wrong runqueue then the migration thread removes
4910 * it and puts it into the right queue.
4911 * 6) migration thread up()s the semaphore.
4912 * 7) we wake up and the migration is done.
4916 * Change a given task's CPU affinity. Migrate the thread to a
4917 * proper CPU and schedule it away if the CPU it's executing on
4918 * is removed from the allowed bitmask.
4920 * NOTE: the caller must have a valid reference to the task, the
4921 * task must not exit() & deallocate itself prematurely. The
4922 * call is not atomic; no spinlocks may be held.
4924 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4926 struct migration_req req
;
4927 unsigned long flags
;
4931 rq
= task_rq_lock(p
, &flags
);
4932 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4937 p
->cpus_allowed
= new_mask
;
4938 /* Can the task run on the task's current CPU? If so, we're done */
4939 if (cpu_isset(task_cpu(p
), new_mask
))
4942 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4943 /* Need help from migration thread: drop lock and wait. */
4944 task_rq_unlock(rq
, &flags
);
4945 wake_up_process(rq
->migration_thread
);
4946 wait_for_completion(&req
.done
);
4947 tlb_migrate_finish(p
->mm
);
4951 task_rq_unlock(rq
, &flags
);
4955 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4958 * Move (not current) task off this cpu, onto dest cpu. We're doing
4959 * this because either it can't run here any more (set_cpus_allowed()
4960 * away from this CPU, or CPU going down), or because we're
4961 * attempting to rebalance this task on exec (sched_exec).
4963 * So we race with normal scheduler movements, but that's OK, as long
4964 * as the task is no longer on this CPU.
4966 * Returns non-zero if task was successfully migrated.
4968 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4970 struct rq
*rq_dest
, *rq_src
;
4973 if (unlikely(cpu_is_offline(dest_cpu
)))
4976 rq_src
= cpu_rq(src_cpu
);
4977 rq_dest
= cpu_rq(dest_cpu
);
4979 double_rq_lock(rq_src
, rq_dest
);
4980 /* Already moved. */
4981 if (task_cpu(p
) != src_cpu
)
4983 /* Affinity changed (again). */
4984 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4987 on_rq
= p
->se
.on_rq
;
4989 update_rq_clock(rq_src
);
4990 deactivate_task(rq_src
, p
, 0, rq_src
->clock
);
4992 set_task_cpu(p
, dest_cpu
);
4994 activate_task(rq_dest
, p
, 0);
4995 check_preempt_curr(rq_dest
, p
);
4999 double_rq_unlock(rq_src
, rq_dest
);
5004 * migration_thread - this is a highprio system thread that performs
5005 * thread migration by bumping thread off CPU then 'pushing' onto
5008 static int migration_thread(void *data
)
5010 int cpu
= (long)data
;
5014 BUG_ON(rq
->migration_thread
!= current
);
5016 set_current_state(TASK_INTERRUPTIBLE
);
5017 while (!kthread_should_stop()) {
5018 struct migration_req
*req
;
5019 struct list_head
*head
;
5021 spin_lock_irq(&rq
->lock
);
5023 if (cpu_is_offline(cpu
)) {
5024 spin_unlock_irq(&rq
->lock
);
5028 if (rq
->active_balance
) {
5029 active_load_balance(rq
, cpu
);
5030 rq
->active_balance
= 0;
5033 head
= &rq
->migration_queue
;
5035 if (list_empty(head
)) {
5036 spin_unlock_irq(&rq
->lock
);
5038 set_current_state(TASK_INTERRUPTIBLE
);
5041 req
= list_entry(head
->next
, struct migration_req
, list
);
5042 list_del_init(head
->next
);
5044 spin_unlock(&rq
->lock
);
5045 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5048 complete(&req
->done
);
5050 __set_current_state(TASK_RUNNING
);
5054 /* Wait for kthread_stop */
5055 set_current_state(TASK_INTERRUPTIBLE
);
5056 while (!kthread_should_stop()) {
5058 set_current_state(TASK_INTERRUPTIBLE
);
5060 __set_current_state(TASK_RUNNING
);
5064 #ifdef CONFIG_HOTPLUG_CPU
5066 * Figure out where task on dead CPU should go, use force if neccessary.
5067 * NOTE: interrupts should be disabled by the caller
5069 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5071 unsigned long flags
;
5078 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5079 cpus_and(mask
, mask
, p
->cpus_allowed
);
5080 dest_cpu
= any_online_cpu(mask
);
5082 /* On any allowed CPU? */
5083 if (dest_cpu
== NR_CPUS
)
5084 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5086 /* No more Mr. Nice Guy. */
5087 if (dest_cpu
== NR_CPUS
) {
5088 rq
= task_rq_lock(p
, &flags
);
5089 cpus_setall(p
->cpus_allowed
);
5090 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5091 task_rq_unlock(rq
, &flags
);
5094 * Don't tell them about moving exiting tasks or
5095 * kernel threads (both mm NULL), since they never
5098 if (p
->mm
&& printk_ratelimit())
5099 printk(KERN_INFO
"process %d (%s) no "
5100 "longer affine to cpu%d\n",
5101 p
->pid
, p
->comm
, dead_cpu
);
5103 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5108 * While a dead CPU has no uninterruptible tasks queued at this point,
5109 * it might still have a nonzero ->nr_uninterruptible counter, because
5110 * for performance reasons the counter is not stricly tracking tasks to
5111 * their home CPUs. So we just add the counter to another CPU's counter,
5112 * to keep the global sum constant after CPU-down:
5114 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5116 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5117 unsigned long flags
;
5119 local_irq_save(flags
);
5120 double_rq_lock(rq_src
, rq_dest
);
5121 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5122 rq_src
->nr_uninterruptible
= 0;
5123 double_rq_unlock(rq_src
, rq_dest
);
5124 local_irq_restore(flags
);
5127 /* Run through task list and migrate tasks from the dead cpu. */
5128 static void migrate_live_tasks(int src_cpu
)
5130 struct task_struct
*p
, *t
;
5132 write_lock_irq(&tasklist_lock
);
5134 do_each_thread(t
, p
) {
5138 if (task_cpu(p
) == src_cpu
)
5139 move_task_off_dead_cpu(src_cpu
, p
);
5140 } while_each_thread(t
, p
);
5142 write_unlock_irq(&tasklist_lock
);
5146 * Schedules idle task to be the next runnable task on current CPU.
5147 * It does so by boosting its priority to highest possible and adding it to
5148 * the _front_ of the runqueue. Used by CPU offline code.
5150 void sched_idle_next(void)
5152 int this_cpu
= smp_processor_id();
5153 struct rq
*rq
= cpu_rq(this_cpu
);
5154 struct task_struct
*p
= rq
->idle
;
5155 unsigned long flags
;
5157 /* cpu has to be offline */
5158 BUG_ON(cpu_online(this_cpu
));
5161 * Strictly not necessary since rest of the CPUs are stopped by now
5162 * and interrupts disabled on the current cpu.
5164 spin_lock_irqsave(&rq
->lock
, flags
);
5166 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5168 /* Add idle task to the _front_ of its priority queue: */
5169 activate_idle_task(p
, rq
);
5171 spin_unlock_irqrestore(&rq
->lock
, flags
);
5175 * Ensures that the idle task is using init_mm right before its cpu goes
5178 void idle_task_exit(void)
5180 struct mm_struct
*mm
= current
->active_mm
;
5182 BUG_ON(cpu_online(smp_processor_id()));
5185 switch_mm(mm
, &init_mm
, current
);
5189 /* called under rq->lock with disabled interrupts */
5190 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5192 struct rq
*rq
= cpu_rq(dead_cpu
);
5194 /* Must be exiting, otherwise would be on tasklist. */
5195 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5197 /* Cannot have done final schedule yet: would have vanished. */
5198 BUG_ON(p
->state
== TASK_DEAD
);
5203 * Drop lock around migration; if someone else moves it,
5204 * that's OK. No task can be added to this CPU, so iteration is
5206 * NOTE: interrupts should be left disabled --dev@
5208 spin_unlock(&rq
->lock
);
5209 move_task_off_dead_cpu(dead_cpu
, p
);
5210 spin_lock(&rq
->lock
);
5215 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5216 static void migrate_dead_tasks(unsigned int dead_cpu
)
5218 struct rq
*rq
= cpu_rq(dead_cpu
);
5219 struct task_struct
*next
;
5222 if (!rq
->nr_running
)
5224 update_rq_clock(rq
);
5225 next
= pick_next_task(rq
, rq
->curr
);
5228 migrate_dead(dead_cpu
, next
);
5232 #endif /* CONFIG_HOTPLUG_CPU */
5234 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5236 static struct ctl_table sd_ctl_dir
[] = {
5238 .procname
= "sched_domain",
5244 static struct ctl_table sd_ctl_root
[] = {
5246 .procname
= "kernel",
5248 .child
= sd_ctl_dir
,
5253 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5255 struct ctl_table
*entry
=
5256 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5259 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5265 set_table_entry(struct ctl_table
*entry
,
5266 const char *procname
, void *data
, int maxlen
,
5267 mode_t mode
, proc_handler
*proc_handler
)
5269 entry
->procname
= procname
;
5271 entry
->maxlen
= maxlen
;
5273 entry
->proc_handler
= proc_handler
;
5276 static struct ctl_table
*
5277 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5279 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5281 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5282 sizeof(long), 0644, proc_doulongvec_minmax
);
5283 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5284 sizeof(long), 0644, proc_doulongvec_minmax
);
5285 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5286 sizeof(int), 0644, proc_dointvec_minmax
);
5287 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5288 sizeof(int), 0644, proc_dointvec_minmax
);
5289 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5290 sizeof(int), 0644, proc_dointvec_minmax
);
5291 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5292 sizeof(int), 0644, proc_dointvec_minmax
);
5293 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5294 sizeof(int), 0644, proc_dointvec_minmax
);
5295 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5296 sizeof(int), 0644, proc_dointvec_minmax
);
5297 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5298 sizeof(int), 0644, proc_dointvec_minmax
);
5299 set_table_entry(&table
[10], "cache_nice_tries",
5300 &sd
->cache_nice_tries
,
5301 sizeof(int), 0644, proc_dointvec_minmax
);
5302 set_table_entry(&table
[12], "flags", &sd
->flags
,
5303 sizeof(int), 0644, proc_dointvec_minmax
);
5308 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5310 struct ctl_table
*entry
, *table
;
5311 struct sched_domain
*sd
;
5312 int domain_num
= 0, i
;
5315 for_each_domain(cpu
, sd
)
5317 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5320 for_each_domain(cpu
, sd
) {
5321 snprintf(buf
, 32, "domain%d", i
);
5322 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5324 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5331 static struct ctl_table_header
*sd_sysctl_header
;
5332 static void init_sched_domain_sysctl(void)
5334 int i
, cpu_num
= num_online_cpus();
5335 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5338 sd_ctl_dir
[0].child
= entry
;
5340 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5341 snprintf(buf
, 32, "cpu%d", i
);
5342 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5344 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5346 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5349 static void init_sched_domain_sysctl(void)
5355 * migration_call - callback that gets triggered when a CPU is added.
5356 * Here we can start up the necessary migration thread for the new CPU.
5358 static int __cpuinit
5359 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5361 struct task_struct
*p
;
5362 int cpu
= (long)hcpu
;
5363 unsigned long flags
;
5367 case CPU_LOCK_ACQUIRE
:
5368 mutex_lock(&sched_hotcpu_mutex
);
5371 case CPU_UP_PREPARE
:
5372 case CPU_UP_PREPARE_FROZEN
:
5373 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5376 kthread_bind(p
, cpu
);
5377 /* Must be high prio: stop_machine expects to yield to it. */
5378 rq
= task_rq_lock(p
, &flags
);
5379 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5380 task_rq_unlock(rq
, &flags
);
5381 cpu_rq(cpu
)->migration_thread
= p
;
5385 case CPU_ONLINE_FROZEN
:
5386 /* Strictly unneccessary, as first user will wake it. */
5387 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5390 #ifdef CONFIG_HOTPLUG_CPU
5391 case CPU_UP_CANCELED
:
5392 case CPU_UP_CANCELED_FROZEN
:
5393 if (!cpu_rq(cpu
)->migration_thread
)
5395 /* Unbind it from offline cpu so it can run. Fall thru. */
5396 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5397 any_online_cpu(cpu_online_map
));
5398 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5399 cpu_rq(cpu
)->migration_thread
= NULL
;
5403 case CPU_DEAD_FROZEN
:
5404 migrate_live_tasks(cpu
);
5406 kthread_stop(rq
->migration_thread
);
5407 rq
->migration_thread
= NULL
;
5408 /* Idle task back to normal (off runqueue, low prio) */
5409 rq
= task_rq_lock(rq
->idle
, &flags
);
5410 update_rq_clock(rq
);
5411 deactivate_task(rq
, rq
->idle
, 0, rq
->clock
);
5412 rq
->idle
->static_prio
= MAX_PRIO
;
5413 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5414 rq
->idle
->sched_class
= &idle_sched_class
;
5415 migrate_dead_tasks(cpu
);
5416 task_rq_unlock(rq
, &flags
);
5417 migrate_nr_uninterruptible(rq
);
5418 BUG_ON(rq
->nr_running
!= 0);
5420 /* No need to migrate the tasks: it was best-effort if
5421 * they didn't take sched_hotcpu_mutex. Just wake up
5422 * the requestors. */
5423 spin_lock_irq(&rq
->lock
);
5424 while (!list_empty(&rq
->migration_queue
)) {
5425 struct migration_req
*req
;
5427 req
= list_entry(rq
->migration_queue
.next
,
5428 struct migration_req
, list
);
5429 list_del_init(&req
->list
);
5430 complete(&req
->done
);
5432 spin_unlock_irq(&rq
->lock
);
5435 case CPU_LOCK_RELEASE
:
5436 mutex_unlock(&sched_hotcpu_mutex
);
5442 /* Register at highest priority so that task migration (migrate_all_tasks)
5443 * happens before everything else.
5445 static struct notifier_block __cpuinitdata migration_notifier
= {
5446 .notifier_call
= migration_call
,
5450 int __init
migration_init(void)
5452 void *cpu
= (void *)(long)smp_processor_id();
5455 /* Start one for the boot CPU: */
5456 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5457 BUG_ON(err
== NOTIFY_BAD
);
5458 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5459 register_cpu_notifier(&migration_notifier
);
5467 /* Number of possible processor ids */
5468 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5469 EXPORT_SYMBOL(nr_cpu_ids
);
5471 #undef SCHED_DOMAIN_DEBUG
5472 #ifdef SCHED_DOMAIN_DEBUG
5473 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5478 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5482 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5487 struct sched_group
*group
= sd
->groups
;
5488 cpumask_t groupmask
;
5490 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5491 cpus_clear(groupmask
);
5494 for (i
= 0; i
< level
+ 1; i
++)
5496 printk("domain %d: ", level
);
5498 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5499 printk("does not load-balance\n");
5501 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5506 printk("span %s\n", str
);
5508 if (!cpu_isset(cpu
, sd
->span
))
5509 printk(KERN_ERR
"ERROR: domain->span does not contain "
5511 if (!cpu_isset(cpu
, group
->cpumask
))
5512 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5516 for (i
= 0; i
< level
+ 2; i
++)
5522 printk(KERN_ERR
"ERROR: group is NULL\n");
5526 if (!group
->__cpu_power
) {
5528 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5532 if (!cpus_weight(group
->cpumask
)) {
5534 printk(KERN_ERR
"ERROR: empty group\n");
5537 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5539 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5542 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5544 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5547 group
= group
->next
;
5548 } while (group
!= sd
->groups
);
5551 if (!cpus_equal(sd
->span
, groupmask
))
5552 printk(KERN_ERR
"ERROR: groups don't span "
5560 if (!cpus_subset(groupmask
, sd
->span
))
5561 printk(KERN_ERR
"ERROR: parent span is not a superset "
5562 "of domain->span\n");
5567 # define sched_domain_debug(sd, cpu) do { } while (0)
5570 static int sd_degenerate(struct sched_domain
*sd
)
5572 if (cpus_weight(sd
->span
) == 1)
5575 /* Following flags need at least 2 groups */
5576 if (sd
->flags
& (SD_LOAD_BALANCE
|
5577 SD_BALANCE_NEWIDLE
|
5581 SD_SHARE_PKG_RESOURCES
)) {
5582 if (sd
->groups
!= sd
->groups
->next
)
5586 /* Following flags don't use groups */
5587 if (sd
->flags
& (SD_WAKE_IDLE
|
5596 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5598 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5600 if (sd_degenerate(parent
))
5603 if (!cpus_equal(sd
->span
, parent
->span
))
5606 /* Does parent contain flags not in child? */
5607 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5608 if (cflags
& SD_WAKE_AFFINE
)
5609 pflags
&= ~SD_WAKE_BALANCE
;
5610 /* Flags needing groups don't count if only 1 group in parent */
5611 if (parent
->groups
== parent
->groups
->next
) {
5612 pflags
&= ~(SD_LOAD_BALANCE
|
5613 SD_BALANCE_NEWIDLE
|
5617 SD_SHARE_PKG_RESOURCES
);
5619 if (~cflags
& pflags
)
5626 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5627 * hold the hotplug lock.
5629 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5631 struct rq
*rq
= cpu_rq(cpu
);
5632 struct sched_domain
*tmp
;
5634 /* Remove the sched domains which do not contribute to scheduling. */
5635 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5636 struct sched_domain
*parent
= tmp
->parent
;
5639 if (sd_parent_degenerate(tmp
, parent
)) {
5640 tmp
->parent
= parent
->parent
;
5642 parent
->parent
->child
= tmp
;
5646 if (sd
&& sd_degenerate(sd
)) {
5652 sched_domain_debug(sd
, cpu
);
5654 rcu_assign_pointer(rq
->sd
, sd
);
5657 /* cpus with isolated domains */
5658 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5660 /* Setup the mask of cpus configured for isolated domains */
5661 static int __init
isolated_cpu_setup(char *str
)
5663 int ints
[NR_CPUS
], i
;
5665 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5666 cpus_clear(cpu_isolated_map
);
5667 for (i
= 1; i
<= ints
[0]; i
++)
5668 if (ints
[i
] < NR_CPUS
)
5669 cpu_set(ints
[i
], cpu_isolated_map
);
5673 __setup ("isolcpus=", isolated_cpu_setup
);
5676 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5677 * to a function which identifies what group(along with sched group) a CPU
5678 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5679 * (due to the fact that we keep track of groups covered with a cpumask_t).
5681 * init_sched_build_groups will build a circular linked list of the groups
5682 * covered by the given span, and will set each group's ->cpumask correctly,
5683 * and ->cpu_power to 0.
5686 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5687 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5688 struct sched_group
**sg
))
5690 struct sched_group
*first
= NULL
, *last
= NULL
;
5691 cpumask_t covered
= CPU_MASK_NONE
;
5694 for_each_cpu_mask(i
, span
) {
5695 struct sched_group
*sg
;
5696 int group
= group_fn(i
, cpu_map
, &sg
);
5699 if (cpu_isset(i
, covered
))
5702 sg
->cpumask
= CPU_MASK_NONE
;
5703 sg
->__cpu_power
= 0;
5705 for_each_cpu_mask(j
, span
) {
5706 if (group_fn(j
, cpu_map
, NULL
) != group
)
5709 cpu_set(j
, covered
);
5710 cpu_set(j
, sg
->cpumask
);
5721 #define SD_NODES_PER_DOMAIN 16
5726 * find_next_best_node - find the next node to include in a sched_domain
5727 * @node: node whose sched_domain we're building
5728 * @used_nodes: nodes already in the sched_domain
5730 * Find the next node to include in a given scheduling domain. Simply
5731 * finds the closest node not already in the @used_nodes map.
5733 * Should use nodemask_t.
5735 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5737 int i
, n
, val
, min_val
, best_node
= 0;
5741 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5742 /* Start at @node */
5743 n
= (node
+ i
) % MAX_NUMNODES
;
5745 if (!nr_cpus_node(n
))
5748 /* Skip already used nodes */
5749 if (test_bit(n
, used_nodes
))
5752 /* Simple min distance search */
5753 val
= node_distance(node
, n
);
5755 if (val
< min_val
) {
5761 set_bit(best_node
, used_nodes
);
5766 * sched_domain_node_span - get a cpumask for a node's sched_domain
5767 * @node: node whose cpumask we're constructing
5768 * @size: number of nodes to include in this span
5770 * Given a node, construct a good cpumask for its sched_domain to span. It
5771 * should be one that prevents unnecessary balancing, but also spreads tasks
5774 static cpumask_t
sched_domain_node_span(int node
)
5776 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5777 cpumask_t span
, nodemask
;
5781 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5783 nodemask
= node_to_cpumask(node
);
5784 cpus_or(span
, span
, nodemask
);
5785 set_bit(node
, used_nodes
);
5787 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5788 int next_node
= find_next_best_node(node
, used_nodes
);
5790 nodemask
= node_to_cpumask(next_node
);
5791 cpus_or(span
, span
, nodemask
);
5798 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5801 * SMT sched-domains:
5803 #ifdef CONFIG_SCHED_SMT
5804 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5805 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5807 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5808 struct sched_group
**sg
)
5811 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5817 * multi-core sched-domains:
5819 #ifdef CONFIG_SCHED_MC
5820 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5821 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5824 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5825 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5826 struct sched_group
**sg
)
5829 cpumask_t mask
= cpu_sibling_map
[cpu
];
5830 cpus_and(mask
, mask
, *cpu_map
);
5831 group
= first_cpu(mask
);
5833 *sg
= &per_cpu(sched_group_core
, group
);
5836 #elif defined(CONFIG_SCHED_MC)
5837 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5838 struct sched_group
**sg
)
5841 *sg
= &per_cpu(sched_group_core
, cpu
);
5846 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5847 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5849 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5850 struct sched_group
**sg
)
5853 #ifdef CONFIG_SCHED_MC
5854 cpumask_t mask
= cpu_coregroup_map(cpu
);
5855 cpus_and(mask
, mask
, *cpu_map
);
5856 group
= first_cpu(mask
);
5857 #elif defined(CONFIG_SCHED_SMT)
5858 cpumask_t mask
= cpu_sibling_map
[cpu
];
5859 cpus_and(mask
, mask
, *cpu_map
);
5860 group
= first_cpu(mask
);
5865 *sg
= &per_cpu(sched_group_phys
, group
);
5871 * The init_sched_build_groups can't handle what we want to do with node
5872 * groups, so roll our own. Now each node has its own list of groups which
5873 * gets dynamically allocated.
5875 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5876 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5878 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5879 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5881 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5882 struct sched_group
**sg
)
5884 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5887 cpus_and(nodemask
, nodemask
, *cpu_map
);
5888 group
= first_cpu(nodemask
);
5891 *sg
= &per_cpu(sched_group_allnodes
, group
);
5895 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5897 struct sched_group
*sg
= group_head
;
5903 for_each_cpu_mask(j
, sg
->cpumask
) {
5904 struct sched_domain
*sd
;
5906 sd
= &per_cpu(phys_domains
, j
);
5907 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5909 * Only add "power" once for each
5915 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5918 if (sg
!= group_head
)
5924 /* Free memory allocated for various sched_group structures */
5925 static void free_sched_groups(const cpumask_t
*cpu_map
)
5929 for_each_cpu_mask(cpu
, *cpu_map
) {
5930 struct sched_group
**sched_group_nodes
5931 = sched_group_nodes_bycpu
[cpu
];
5933 if (!sched_group_nodes
)
5936 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5937 cpumask_t nodemask
= node_to_cpumask(i
);
5938 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5940 cpus_and(nodemask
, nodemask
, *cpu_map
);
5941 if (cpus_empty(nodemask
))
5951 if (oldsg
!= sched_group_nodes
[i
])
5954 kfree(sched_group_nodes
);
5955 sched_group_nodes_bycpu
[cpu
] = NULL
;
5959 static void free_sched_groups(const cpumask_t
*cpu_map
)
5965 * Initialize sched groups cpu_power.
5967 * cpu_power indicates the capacity of sched group, which is used while
5968 * distributing the load between different sched groups in a sched domain.
5969 * Typically cpu_power for all the groups in a sched domain will be same unless
5970 * there are asymmetries in the topology. If there are asymmetries, group
5971 * having more cpu_power will pickup more load compared to the group having
5974 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5975 * the maximum number of tasks a group can handle in the presence of other idle
5976 * or lightly loaded groups in the same sched domain.
5978 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5980 struct sched_domain
*child
;
5981 struct sched_group
*group
;
5983 WARN_ON(!sd
|| !sd
->groups
);
5985 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5990 sd
->groups
->__cpu_power
= 0;
5993 * For perf policy, if the groups in child domain share resources
5994 * (for example cores sharing some portions of the cache hierarchy
5995 * or SMT), then set this domain groups cpu_power such that each group
5996 * can handle only one task, when there are other idle groups in the
5997 * same sched domain.
5999 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6001 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6002 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6007 * add cpu_power of each child group to this groups cpu_power
6009 group
= child
->groups
;
6011 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6012 group
= group
->next
;
6013 } while (group
!= child
->groups
);
6017 * Build sched domains for a given set of cpus and attach the sched domains
6018 * to the individual cpus
6020 static int build_sched_domains(const cpumask_t
*cpu_map
)
6024 struct sched_group
**sched_group_nodes
= NULL
;
6025 int sd_allnodes
= 0;
6028 * Allocate the per-node list of sched groups
6030 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6032 if (!sched_group_nodes
) {
6033 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6036 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6040 * Set up domains for cpus specified by the cpu_map.
6042 for_each_cpu_mask(i
, *cpu_map
) {
6043 struct sched_domain
*sd
= NULL
, *p
;
6044 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6046 cpus_and(nodemask
, nodemask
, *cpu_map
);
6049 if (cpus_weight(*cpu_map
) >
6050 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6051 sd
= &per_cpu(allnodes_domains
, i
);
6052 *sd
= SD_ALLNODES_INIT
;
6053 sd
->span
= *cpu_map
;
6054 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6060 sd
= &per_cpu(node_domains
, i
);
6062 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6066 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6070 sd
= &per_cpu(phys_domains
, i
);
6072 sd
->span
= nodemask
;
6076 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6078 #ifdef CONFIG_SCHED_MC
6080 sd
= &per_cpu(core_domains
, i
);
6082 sd
->span
= cpu_coregroup_map(i
);
6083 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6086 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6089 #ifdef CONFIG_SCHED_SMT
6091 sd
= &per_cpu(cpu_domains
, i
);
6092 *sd
= SD_SIBLING_INIT
;
6093 sd
->span
= cpu_sibling_map
[i
];
6094 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6097 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6101 #ifdef CONFIG_SCHED_SMT
6102 /* Set up CPU (sibling) groups */
6103 for_each_cpu_mask(i
, *cpu_map
) {
6104 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6105 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6106 if (i
!= first_cpu(this_sibling_map
))
6109 init_sched_build_groups(this_sibling_map
, cpu_map
,
6114 #ifdef CONFIG_SCHED_MC
6115 /* Set up multi-core groups */
6116 for_each_cpu_mask(i
, *cpu_map
) {
6117 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6118 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6119 if (i
!= first_cpu(this_core_map
))
6121 init_sched_build_groups(this_core_map
, cpu_map
,
6122 &cpu_to_core_group
);
6126 /* Set up physical groups */
6127 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6128 cpumask_t nodemask
= node_to_cpumask(i
);
6130 cpus_and(nodemask
, nodemask
, *cpu_map
);
6131 if (cpus_empty(nodemask
))
6134 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6138 /* Set up node groups */
6140 init_sched_build_groups(*cpu_map
, cpu_map
,
6141 &cpu_to_allnodes_group
);
6143 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6144 /* Set up node groups */
6145 struct sched_group
*sg
, *prev
;
6146 cpumask_t nodemask
= node_to_cpumask(i
);
6147 cpumask_t domainspan
;
6148 cpumask_t covered
= CPU_MASK_NONE
;
6151 cpus_and(nodemask
, nodemask
, *cpu_map
);
6152 if (cpus_empty(nodemask
)) {
6153 sched_group_nodes
[i
] = NULL
;
6157 domainspan
= sched_domain_node_span(i
);
6158 cpus_and(domainspan
, domainspan
, *cpu_map
);
6160 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6162 printk(KERN_WARNING
"Can not alloc domain group for "
6166 sched_group_nodes
[i
] = sg
;
6167 for_each_cpu_mask(j
, nodemask
) {
6168 struct sched_domain
*sd
;
6170 sd
= &per_cpu(node_domains
, j
);
6173 sg
->__cpu_power
= 0;
6174 sg
->cpumask
= nodemask
;
6176 cpus_or(covered
, covered
, nodemask
);
6179 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6180 cpumask_t tmp
, notcovered
;
6181 int n
= (i
+ j
) % MAX_NUMNODES
;
6183 cpus_complement(notcovered
, covered
);
6184 cpus_and(tmp
, notcovered
, *cpu_map
);
6185 cpus_and(tmp
, tmp
, domainspan
);
6186 if (cpus_empty(tmp
))
6189 nodemask
= node_to_cpumask(n
);
6190 cpus_and(tmp
, tmp
, nodemask
);
6191 if (cpus_empty(tmp
))
6194 sg
= kmalloc_node(sizeof(struct sched_group
),
6198 "Can not alloc domain group for node %d\n", j
);
6201 sg
->__cpu_power
= 0;
6203 sg
->next
= prev
->next
;
6204 cpus_or(covered
, covered
, tmp
);
6211 /* Calculate CPU power for physical packages and nodes */
6212 #ifdef CONFIG_SCHED_SMT
6213 for_each_cpu_mask(i
, *cpu_map
) {
6214 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6216 init_sched_groups_power(i
, sd
);
6219 #ifdef CONFIG_SCHED_MC
6220 for_each_cpu_mask(i
, *cpu_map
) {
6221 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6223 init_sched_groups_power(i
, sd
);
6227 for_each_cpu_mask(i
, *cpu_map
) {
6228 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6230 init_sched_groups_power(i
, sd
);
6234 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6235 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6238 struct sched_group
*sg
;
6240 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6241 init_numa_sched_groups_power(sg
);
6245 /* Attach the domains */
6246 for_each_cpu_mask(i
, *cpu_map
) {
6247 struct sched_domain
*sd
;
6248 #ifdef CONFIG_SCHED_SMT
6249 sd
= &per_cpu(cpu_domains
, i
);
6250 #elif defined(CONFIG_SCHED_MC)
6251 sd
= &per_cpu(core_domains
, i
);
6253 sd
= &per_cpu(phys_domains
, i
);
6255 cpu_attach_domain(sd
, i
);
6262 free_sched_groups(cpu_map
);
6267 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6269 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6271 cpumask_t cpu_default_map
;
6275 * Setup mask for cpus without special case scheduling requirements.
6276 * For now this just excludes isolated cpus, but could be used to
6277 * exclude other special cases in the future.
6279 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6281 err
= build_sched_domains(&cpu_default_map
);
6286 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6288 free_sched_groups(cpu_map
);
6292 * Detach sched domains from a group of cpus specified in cpu_map
6293 * These cpus will now be attached to the NULL domain
6295 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6299 for_each_cpu_mask(i
, *cpu_map
)
6300 cpu_attach_domain(NULL
, i
);
6301 synchronize_sched();
6302 arch_destroy_sched_domains(cpu_map
);
6306 * Partition sched domains as specified by the cpumasks below.
6307 * This attaches all cpus from the cpumasks to the NULL domain,
6308 * waits for a RCU quiescent period, recalculates sched
6309 * domain information and then attaches them back to the
6310 * correct sched domains
6311 * Call with hotplug lock held
6313 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6315 cpumask_t change_map
;
6318 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6319 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6320 cpus_or(change_map
, *partition1
, *partition2
);
6322 /* Detach sched domains from all of the affected cpus */
6323 detach_destroy_domains(&change_map
);
6324 if (!cpus_empty(*partition1
))
6325 err
= build_sched_domains(partition1
);
6326 if (!err
&& !cpus_empty(*partition2
))
6327 err
= build_sched_domains(partition2
);
6332 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6333 int arch_reinit_sched_domains(void)
6337 mutex_lock(&sched_hotcpu_mutex
);
6338 detach_destroy_domains(&cpu_online_map
);
6339 err
= arch_init_sched_domains(&cpu_online_map
);
6340 mutex_unlock(&sched_hotcpu_mutex
);
6345 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6349 if (buf
[0] != '0' && buf
[0] != '1')
6353 sched_smt_power_savings
= (buf
[0] == '1');
6355 sched_mc_power_savings
= (buf
[0] == '1');
6357 ret
= arch_reinit_sched_domains();
6359 return ret
? ret
: count
;
6362 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6366 #ifdef CONFIG_SCHED_SMT
6368 err
= sysfs_create_file(&cls
->kset
.kobj
,
6369 &attr_sched_smt_power_savings
.attr
);
6371 #ifdef CONFIG_SCHED_MC
6372 if (!err
&& mc_capable())
6373 err
= sysfs_create_file(&cls
->kset
.kobj
,
6374 &attr_sched_mc_power_savings
.attr
);
6380 #ifdef CONFIG_SCHED_MC
6381 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6383 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6385 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6386 const char *buf
, size_t count
)
6388 return sched_power_savings_store(buf
, count
, 0);
6390 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6391 sched_mc_power_savings_store
);
6394 #ifdef CONFIG_SCHED_SMT
6395 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6397 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6399 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6400 const char *buf
, size_t count
)
6402 return sched_power_savings_store(buf
, count
, 1);
6404 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6405 sched_smt_power_savings_store
);
6409 * Force a reinitialization of the sched domains hierarchy. The domains
6410 * and groups cannot be updated in place without racing with the balancing
6411 * code, so we temporarily attach all running cpus to the NULL domain
6412 * which will prevent rebalancing while the sched domains are recalculated.
6414 static int update_sched_domains(struct notifier_block
*nfb
,
6415 unsigned long action
, void *hcpu
)
6418 case CPU_UP_PREPARE
:
6419 case CPU_UP_PREPARE_FROZEN
:
6420 case CPU_DOWN_PREPARE
:
6421 case CPU_DOWN_PREPARE_FROZEN
:
6422 detach_destroy_domains(&cpu_online_map
);
6425 case CPU_UP_CANCELED
:
6426 case CPU_UP_CANCELED_FROZEN
:
6427 case CPU_DOWN_FAILED
:
6428 case CPU_DOWN_FAILED_FROZEN
:
6430 case CPU_ONLINE_FROZEN
:
6432 case CPU_DEAD_FROZEN
:
6434 * Fall through and re-initialise the domains.
6441 /* The hotplug lock is already held by cpu_up/cpu_down */
6442 arch_init_sched_domains(&cpu_online_map
);
6447 void __init
sched_init_smp(void)
6449 cpumask_t non_isolated_cpus
;
6451 mutex_lock(&sched_hotcpu_mutex
);
6452 arch_init_sched_domains(&cpu_online_map
);
6453 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6454 if (cpus_empty(non_isolated_cpus
))
6455 cpu_set(smp_processor_id(), non_isolated_cpus
);
6456 mutex_unlock(&sched_hotcpu_mutex
);
6457 /* XXX: Theoretical race here - CPU may be hotplugged now */
6458 hotcpu_notifier(update_sched_domains
, 0);
6460 init_sched_domain_sysctl();
6462 /* Move init over to a non-isolated CPU */
6463 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6465 sched_init_granularity();
6468 void __init
sched_init_smp(void)
6470 sched_init_granularity();
6472 #endif /* CONFIG_SMP */
6474 int in_sched_functions(unsigned long addr
)
6476 /* Linker adds these: start and end of __sched functions */
6477 extern char __sched_text_start
[], __sched_text_end
[];
6479 return in_lock_functions(addr
) ||
6480 (addr
>= (unsigned long)__sched_text_start
6481 && addr
< (unsigned long)__sched_text_end
);
6484 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6486 cfs_rq
->tasks_timeline
= RB_ROOT
;
6487 cfs_rq
->fair_clock
= 1;
6488 #ifdef CONFIG_FAIR_GROUP_SCHED
6493 void __init
sched_init(void)
6495 u64 now
= sched_clock();
6496 int highest_cpu
= 0;
6500 * Link up the scheduling class hierarchy:
6502 rt_sched_class
.next
= &fair_sched_class
;
6503 fair_sched_class
.next
= &idle_sched_class
;
6504 idle_sched_class
.next
= NULL
;
6506 for_each_possible_cpu(i
) {
6507 struct rt_prio_array
*array
;
6511 spin_lock_init(&rq
->lock
);
6512 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6515 init_cfs_rq(&rq
->cfs
, rq
);
6516 #ifdef CONFIG_FAIR_GROUP_SCHED
6517 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6518 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6520 rq
->ls
.load_update_last
= now
;
6521 rq
->ls
.load_update_start
= now
;
6523 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6524 rq
->cpu_load
[j
] = 0;
6527 rq
->active_balance
= 0;
6528 rq
->next_balance
= jiffies
;
6531 rq
->migration_thread
= NULL
;
6532 INIT_LIST_HEAD(&rq
->migration_queue
);
6534 atomic_set(&rq
->nr_iowait
, 0);
6536 array
= &rq
->rt
.active
;
6537 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6538 INIT_LIST_HEAD(array
->queue
+ j
);
6539 __clear_bit(j
, array
->bitmap
);
6542 /* delimiter for bitsearch: */
6543 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6546 set_load_weight(&init_task
);
6548 #ifdef CONFIG_PREEMPT_NOTIFIERS
6549 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6553 nr_cpu_ids
= highest_cpu
+ 1;
6554 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6557 #ifdef CONFIG_RT_MUTEXES
6558 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6562 * The boot idle thread does lazy MMU switching as well:
6564 atomic_inc(&init_mm
.mm_count
);
6565 enter_lazy_tlb(&init_mm
, current
);
6568 * Make us the idle thread. Technically, schedule() should not be
6569 * called from this thread, however somewhere below it might be,
6570 * but because we are the idle thread, we just pick up running again
6571 * when this runqueue becomes "idle".
6573 init_idle(current
, smp_processor_id());
6575 * During early bootup we pretend to be a normal task:
6577 current
->sched_class
= &fair_sched_class
;
6580 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6581 void __might_sleep(char *file
, int line
)
6584 static unsigned long prev_jiffy
; /* ratelimiting */
6586 if ((in_atomic() || irqs_disabled()) &&
6587 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6588 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6590 prev_jiffy
= jiffies
;
6591 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6592 " context at %s:%d\n", file
, line
);
6593 printk("in_atomic():%d, irqs_disabled():%d\n",
6594 in_atomic(), irqs_disabled());
6595 debug_show_held_locks(current
);
6596 if (irqs_disabled())
6597 print_irqtrace_events(current
);
6602 EXPORT_SYMBOL(__might_sleep
);
6605 #ifdef CONFIG_MAGIC_SYSRQ
6606 void normalize_rt_tasks(void)
6608 struct task_struct
*g
, *p
;
6609 unsigned long flags
;
6613 read_lock_irq(&tasklist_lock
);
6614 do_each_thread(g
, p
) {
6616 p
->se
.wait_runtime
= 0;
6617 p
->se
.exec_start
= 0;
6618 p
->se
.wait_start_fair
= 0;
6619 p
->se
.sleep_start_fair
= 0;
6620 #ifdef CONFIG_SCHEDSTATS
6621 p
->se
.wait_start
= 0;
6622 p
->se
.sleep_start
= 0;
6623 p
->se
.block_start
= 0;
6625 task_rq(p
)->cfs
.fair_clock
= 0;
6626 task_rq(p
)->clock
= 0;
6630 * Renice negative nice level userspace
6633 if (TASK_NICE(p
) < 0 && p
->mm
)
6634 set_user_nice(p
, 0);
6638 spin_lock_irqsave(&p
->pi_lock
, flags
);
6639 rq
= __task_rq_lock(p
);
6642 * Do not touch the migration thread:
6644 if (p
== rq
->migration_thread
)
6648 on_rq
= p
->se
.on_rq
;
6650 update_rq_clock(task_rq(p
));
6651 deactivate_task(task_rq(p
), p
, 0, task_rq(p
)->clock
);
6653 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6655 activate_task(task_rq(p
), p
, 0);
6656 resched_task(rq
->curr
);
6661 __task_rq_unlock(rq
);
6662 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6663 } while_each_thread(g
, p
);
6665 read_unlock_irq(&tasklist_lock
);
6668 #endif /* CONFIG_MAGIC_SYSRQ */
6672 * These functions are only useful for the IA64 MCA handling.
6674 * They can only be called when the whole system has been
6675 * stopped - every CPU needs to be quiescent, and no scheduling
6676 * activity can take place. Using them for anything else would
6677 * be a serious bug, and as a result, they aren't even visible
6678 * under any other configuration.
6682 * curr_task - return the current task for a given cpu.
6683 * @cpu: the processor in question.
6685 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6687 struct task_struct
*curr_task(int cpu
)
6689 return cpu_curr(cpu
);
6693 * set_curr_task - set the current task for a given cpu.
6694 * @cpu: the processor in question.
6695 * @p: the task pointer to set.
6697 * Description: This function must only be used when non-maskable interrupts
6698 * are serviced on a separate stack. It allows the architecture to switch the
6699 * notion of the current task on a cpu in a non-blocking manner. This function
6700 * must be called with all CPU's synchronized, and interrupts disabled, the
6701 * and caller must save the original value of the current task (see
6702 * curr_task() above) and restore that value before reenabling interrupts and
6703 * re-starting the system.
6705 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6707 void set_curr_task(int cpu
, struct task_struct
*p
)