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
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/security.h>
34 #include <linux/notifier.h>
35 #include <linux/profile.h>
36 #include <linux/suspend.h>
37 #include <linux/vmalloc.h>
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/smp.h>
41 #include <linux/threads.h>
42 #include <linux/timer.h>
43 #include <linux/rcupdate.h>
44 #include <linux/cpu.h>
45 #include <linux/cpuset.h>
46 #include <linux/percpu.h>
47 #include <linux/kthread.h>
48 #include <linux/seq_file.h>
49 #include <linux/syscalls.h>
50 #include <linux/times.h>
51 #include <linux/acct.h>
52 #include <linux/kprobes.h>
55 #include <asm/unistd.h>
58 * Convert user-nice values [ -20 ... 0 ... 19 ]
59 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
62 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
63 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
64 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
67 * 'User priority' is the nice value converted to something we
68 * can work with better when scaling various scheduler parameters,
69 * it's a [ 0 ... 39 ] range.
71 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
72 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
73 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
76 * Some helpers for converting nanosecond timing to jiffy resolution
78 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
79 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
82 * These are the 'tuning knobs' of the scheduler:
84 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
85 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
86 * Timeslices get refilled after they expire.
88 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
89 #define DEF_TIMESLICE (100 * HZ / 1000)
90 #define ON_RUNQUEUE_WEIGHT 30
91 #define CHILD_PENALTY 95
92 #define PARENT_PENALTY 100
94 #define PRIO_BONUS_RATIO 25
95 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
96 #define INTERACTIVE_DELTA 2
97 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
98 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
99 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
102 * If a task is 'interactive' then we reinsert it in the active
103 * array after it has expired its current timeslice. (it will not
104 * continue to run immediately, it will still roundrobin with
105 * other interactive tasks.)
107 * This part scales the interactivity limit depending on niceness.
109 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
110 * Here are a few examples of different nice levels:
112 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
113 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
114 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
115 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
118 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
119 * priority range a task can explore, a value of '1' means the
120 * task is rated interactive.)
122 * Ie. nice +19 tasks can never get 'interactive' enough to be
123 * reinserted into the active array. And only heavily CPU-hog nice -20
124 * tasks will be expired. Default nice 0 tasks are somewhere between,
125 * it takes some effort for them to get interactive, but it's not
129 #define CURRENT_BONUS(p) \
130 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
133 #define GRANULARITY (10 * HZ / 1000 ? : 1)
136 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
137 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
140 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
141 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
144 #define SCALE(v1,v1_max,v2_max) \
145 (v1) * (v2_max) / (v1_max)
148 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
151 #define TASK_INTERACTIVE(p) \
152 ((p)->prio <= (p)->static_prio - DELTA(p))
154 #define INTERACTIVE_SLEEP(p) \
155 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
156 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
158 #define TASK_PREEMPTS_CURR(p, rq) \
159 ((p)->prio < (rq)->curr->prio)
162 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
163 * to time slice values: [800ms ... 100ms ... 5ms]
165 * The higher a thread's priority, the bigger timeslices
166 * it gets during one round of execution. But even the lowest
167 * priority thread gets MIN_TIMESLICE worth of execution time.
170 #define SCALE_PRIO(x, prio) \
171 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
173 static unsigned int static_prio_timeslice(int static_prio
)
175 if (static_prio
< NICE_TO_PRIO(0))
176 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
178 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
181 static inline unsigned int task_timeslice(task_t
*p
)
183 return static_prio_timeslice(p
->static_prio
);
186 #define task_hot(p, now, sd) ((long long) ((now) - (p)->last_ran) \
187 < (long long) (sd)->cache_hot_time)
190 * These are the runqueue data structures:
193 typedef struct runqueue runqueue_t
;
196 unsigned int nr_active
;
197 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
198 struct list_head queue
[MAX_PRIO
];
202 * This is the main, per-CPU runqueue data structure.
204 * Locking rule: those places that want to lock multiple runqueues
205 * (such as the load balancing or the thread migration code), lock
206 * acquire operations must be ordered by ascending &runqueue.
212 * nr_running and cpu_load should be in the same cacheline because
213 * remote CPUs use both these fields when doing load calculation.
215 unsigned long nr_running
;
216 unsigned long raw_weighted_load
;
218 unsigned long cpu_load
[3];
220 unsigned long long nr_switches
;
223 * This is part of a global counter where only the total sum
224 * over all CPUs matters. A task can increase this counter on
225 * one CPU and if it got migrated afterwards it may decrease
226 * it on another CPU. Always updated under the runqueue lock:
228 unsigned long nr_uninterruptible
;
230 unsigned long expired_timestamp
;
231 unsigned long long timestamp_last_tick
;
233 struct mm_struct
*prev_mm
;
234 prio_array_t
*active
, *expired
, arrays
[2];
235 int best_expired_prio
;
239 struct sched_domain
*sd
;
241 /* For active balancing */
245 task_t
*migration_thread
;
246 struct list_head migration_queue
;
249 #ifdef CONFIG_SCHEDSTATS
251 struct sched_info rq_sched_info
;
253 /* sys_sched_yield() stats */
254 unsigned long yld_exp_empty
;
255 unsigned long yld_act_empty
;
256 unsigned long yld_both_empty
;
257 unsigned long yld_cnt
;
259 /* schedule() stats */
260 unsigned long sched_switch
;
261 unsigned long sched_cnt
;
262 unsigned long sched_goidle
;
264 /* try_to_wake_up() stats */
265 unsigned long ttwu_cnt
;
266 unsigned long ttwu_local
;
270 static DEFINE_PER_CPU(struct runqueue
, runqueues
);
273 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
274 * See detach_destroy_domains: synchronize_sched for details.
276 * The domain tree of any CPU may only be accessed from within
277 * preempt-disabled sections.
279 #define for_each_domain(cpu, domain) \
280 for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
282 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
283 #define this_rq() (&__get_cpu_var(runqueues))
284 #define task_rq(p) cpu_rq(task_cpu(p))
285 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
287 #ifndef prepare_arch_switch
288 # define prepare_arch_switch(next) do { } while (0)
290 #ifndef finish_arch_switch
291 # define finish_arch_switch(prev) do { } while (0)
294 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
295 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
297 return rq
->curr
== p
;
300 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
304 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
306 #ifdef CONFIG_DEBUG_SPINLOCK
307 /* this is a valid case when another task releases the spinlock */
308 rq
->lock
.owner
= current
;
310 spin_unlock_irq(&rq
->lock
);
313 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
314 static inline int task_running(runqueue_t
*rq
, task_t
*p
)
319 return rq
->curr
== p
;
323 static inline void prepare_lock_switch(runqueue_t
*rq
, task_t
*next
)
327 * We can optimise this out completely for !SMP, because the
328 * SMP rebalancing from interrupt is the only thing that cares
333 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
334 spin_unlock_irq(&rq
->lock
);
336 spin_unlock(&rq
->lock
);
340 static inline void finish_lock_switch(runqueue_t
*rq
, task_t
*prev
)
344 * After ->oncpu is cleared, the task can be moved to a different CPU.
345 * We must ensure this doesn't happen until the switch is completely
351 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
355 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
358 * __task_rq_lock - lock the runqueue a given task resides on.
359 * Must be called interrupts disabled.
361 static inline runqueue_t
*__task_rq_lock(task_t
*p
)
368 spin_lock(&rq
->lock
);
369 if (unlikely(rq
!= task_rq(p
))) {
370 spin_unlock(&rq
->lock
);
371 goto repeat_lock_task
;
377 * task_rq_lock - lock the runqueue a given task resides on and disable
378 * interrupts. Note the ordering: we can safely lookup the task_rq without
379 * explicitly disabling preemption.
381 static runqueue_t
*task_rq_lock(task_t
*p
, unsigned long *flags
)
387 local_irq_save(*flags
);
389 spin_lock(&rq
->lock
);
390 if (unlikely(rq
!= task_rq(p
))) {
391 spin_unlock_irqrestore(&rq
->lock
, *flags
);
392 goto repeat_lock_task
;
397 static inline void __task_rq_unlock(runqueue_t
*rq
)
400 spin_unlock(&rq
->lock
);
403 static inline void task_rq_unlock(runqueue_t
*rq
, unsigned long *flags
)
406 spin_unlock_irqrestore(&rq
->lock
, *flags
);
409 #ifdef CONFIG_SCHEDSTATS
411 * bump this up when changing the output format or the meaning of an existing
412 * format, so that tools can adapt (or abort)
414 #define SCHEDSTAT_VERSION 12
416 static int show_schedstat(struct seq_file
*seq
, void *v
)
420 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
421 seq_printf(seq
, "timestamp %lu\n", jiffies
);
422 for_each_online_cpu(cpu
) {
423 runqueue_t
*rq
= cpu_rq(cpu
);
425 struct sched_domain
*sd
;
429 /* runqueue-specific stats */
431 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
432 cpu
, rq
->yld_both_empty
,
433 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
434 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
435 rq
->ttwu_cnt
, rq
->ttwu_local
,
436 rq
->rq_sched_info
.cpu_time
,
437 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
439 seq_printf(seq
, "\n");
442 /* domain-specific stats */
444 for_each_domain(cpu
, sd
) {
445 enum idle_type itype
;
446 char mask_str
[NR_CPUS
];
448 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
449 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
450 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
452 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu",
454 sd
->lb_balanced
[itype
],
455 sd
->lb_failed
[itype
],
456 sd
->lb_imbalance
[itype
],
457 sd
->lb_gained
[itype
],
458 sd
->lb_hot_gained
[itype
],
459 sd
->lb_nobusyq
[itype
],
460 sd
->lb_nobusyg
[itype
]);
462 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
463 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
464 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
465 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
466 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
, sd
->ttwu_move_balance
);
474 static int schedstat_open(struct inode
*inode
, struct file
*file
)
476 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
477 char *buf
= kmalloc(size
, GFP_KERNEL
);
483 res
= single_open(file
, show_schedstat
, NULL
);
485 m
= file
->private_data
;
493 struct file_operations proc_schedstat_operations
= {
494 .open
= schedstat_open
,
497 .release
= single_release
,
500 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
501 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
502 #else /* !CONFIG_SCHEDSTATS */
503 # define schedstat_inc(rq, field) do { } while (0)
504 # define schedstat_add(rq, field, amt) do { } while (0)
508 * rq_lock - lock a given runqueue and disable interrupts.
510 static inline runqueue_t
*this_rq_lock(void)
517 spin_lock(&rq
->lock
);
522 #ifdef CONFIG_SCHEDSTATS
524 * Called when a process is dequeued from the active array and given
525 * the cpu. We should note that with the exception of interactive
526 * tasks, the expired queue will become the active queue after the active
527 * queue is empty, without explicitly dequeuing and requeuing tasks in the
528 * expired queue. (Interactive tasks may be requeued directly to the
529 * active queue, thus delaying tasks in the expired queue from running;
530 * see scheduler_tick()).
532 * This function is only called from sched_info_arrive(), rather than
533 * dequeue_task(). Even though a task may be queued and dequeued multiple
534 * times as it is shuffled about, we're really interested in knowing how
535 * long it was from the *first* time it was queued to the time that it
538 static inline void sched_info_dequeued(task_t
*t
)
540 t
->sched_info
.last_queued
= 0;
544 * Called when a task finally hits the cpu. We can now calculate how
545 * long it was waiting to run. We also note when it began so that we
546 * can keep stats on how long its timeslice is.
548 static void sched_info_arrive(task_t
*t
)
550 unsigned long now
= jiffies
, diff
= 0;
551 struct runqueue
*rq
= task_rq(t
);
553 if (t
->sched_info
.last_queued
)
554 diff
= now
- t
->sched_info
.last_queued
;
555 sched_info_dequeued(t
);
556 t
->sched_info
.run_delay
+= diff
;
557 t
->sched_info
.last_arrival
= now
;
558 t
->sched_info
.pcnt
++;
563 rq
->rq_sched_info
.run_delay
+= diff
;
564 rq
->rq_sched_info
.pcnt
++;
568 * Called when a process is queued into either the active or expired
569 * array. The time is noted and later used to determine how long we
570 * had to wait for us to reach the cpu. Since the expired queue will
571 * become the active queue after active queue is empty, without dequeuing
572 * and requeuing any tasks, we are interested in queuing to either. It
573 * is unusual but not impossible for tasks to be dequeued and immediately
574 * requeued in the same or another array: this can happen in sched_yield(),
575 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
578 * This function is only called from enqueue_task(), but also only updates
579 * the timestamp if it is already not set. It's assumed that
580 * sched_info_dequeued() will clear that stamp when appropriate.
582 static inline void sched_info_queued(task_t
*t
)
584 if (!t
->sched_info
.last_queued
)
585 t
->sched_info
.last_queued
= jiffies
;
589 * Called when a process ceases being the active-running process, either
590 * voluntarily or involuntarily. Now we can calculate how long we ran.
592 static inline void sched_info_depart(task_t
*t
)
594 struct runqueue
*rq
= task_rq(t
);
595 unsigned long diff
= jiffies
- t
->sched_info
.last_arrival
;
597 t
->sched_info
.cpu_time
+= diff
;
600 rq
->rq_sched_info
.cpu_time
+= diff
;
604 * Called when tasks are switched involuntarily due, typically, to expiring
605 * their time slice. (This may also be called when switching to or from
606 * the idle task.) We are only called when prev != next.
608 static inline void sched_info_switch(task_t
*prev
, task_t
*next
)
610 struct runqueue
*rq
= task_rq(prev
);
613 * prev now departs the cpu. It's not interesting to record
614 * stats about how efficient we were at scheduling the idle
617 if (prev
!= rq
->idle
)
618 sched_info_depart(prev
);
620 if (next
!= rq
->idle
)
621 sched_info_arrive(next
);
624 #define sched_info_queued(t) do { } while (0)
625 #define sched_info_switch(t, next) do { } while (0)
626 #endif /* CONFIG_SCHEDSTATS */
629 * Adding/removing a task to/from a priority array:
631 static void dequeue_task(struct task_struct
*p
, prio_array_t
*array
)
634 list_del(&p
->run_list
);
635 if (list_empty(array
->queue
+ p
->prio
))
636 __clear_bit(p
->prio
, array
->bitmap
);
639 static void enqueue_task(struct task_struct
*p
, prio_array_t
*array
)
641 sched_info_queued(p
);
642 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
643 __set_bit(p
->prio
, array
->bitmap
);
649 * Put task to the end of the run list without the overhead of dequeue
650 * followed by enqueue.
652 static void requeue_task(struct task_struct
*p
, prio_array_t
*array
)
654 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
657 static inline void enqueue_task_head(struct task_struct
*p
, prio_array_t
*array
)
659 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
660 __set_bit(p
->prio
, array
->bitmap
);
666 * __normal_prio - return the priority that is based on the static
667 * priority but is modified by bonuses/penalties.
669 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
670 * into the -5 ... 0 ... +5 bonus/penalty range.
672 * We use 25% of the full 0...39 priority range so that:
674 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
675 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
677 * Both properties are important to certain workloads.
680 static inline int __normal_prio(task_t
*p
)
684 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
686 prio
= p
->static_prio
- bonus
;
687 if (prio
< MAX_RT_PRIO
)
689 if (prio
> MAX_PRIO
-1)
695 * To aid in avoiding the subversion of "niceness" due to uneven distribution
696 * of tasks with abnormal "nice" values across CPUs the contribution that
697 * each task makes to its run queue's load is weighted according to its
698 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
699 * scaled version of the new time slice allocation that they receive on time
704 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
705 * If static_prio_timeslice() is ever changed to break this assumption then
706 * this code will need modification
708 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
709 #define LOAD_WEIGHT(lp) \
710 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
711 #define PRIO_TO_LOAD_WEIGHT(prio) \
712 LOAD_WEIGHT(static_prio_timeslice(prio))
713 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
714 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
716 static void set_load_weight(task_t
*p
)
718 if (has_rt_policy(p
)) {
720 if (p
== task_rq(p
)->migration_thread
)
722 * The migration thread does the actual balancing.
723 * Giving its load any weight will skew balancing
729 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
731 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
734 static inline void inc_raw_weighted_load(runqueue_t
*rq
, const task_t
*p
)
736 rq
->raw_weighted_load
+= p
->load_weight
;
739 static inline void dec_raw_weighted_load(runqueue_t
*rq
, const task_t
*p
)
741 rq
->raw_weighted_load
-= p
->load_weight
;
744 static inline void inc_nr_running(task_t
*p
, runqueue_t
*rq
)
747 inc_raw_weighted_load(rq
, p
);
750 static inline void dec_nr_running(task_t
*p
, runqueue_t
*rq
)
753 dec_raw_weighted_load(rq
, p
);
757 * Calculate the expected normal priority: i.e. priority
758 * without taking RT-inheritance into account. Might be
759 * boosted by interactivity modifiers. Changes upon fork,
760 * setprio syscalls, and whenever the interactivity
761 * estimator recalculates.
763 static inline int normal_prio(task_t
*p
)
767 if (has_rt_policy(p
))
768 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
770 prio
= __normal_prio(p
);
775 * Calculate the current priority, i.e. the priority
776 * taken into account by the scheduler. This value might
777 * be boosted by RT tasks, or might be boosted by
778 * interactivity modifiers. Will be RT if the task got
779 * RT-boosted. If not then it returns p->normal_prio.
781 static int effective_prio(task_t
*p
)
783 p
->normal_prio
= normal_prio(p
);
785 * If we are RT tasks or we were boosted to RT priority,
786 * keep the priority unchanged. Otherwise, update priority
787 * to the normal priority:
789 if (!rt_prio(p
->prio
))
790 return p
->normal_prio
;
795 * __activate_task - move a task to the runqueue.
797 static void __activate_task(task_t
*p
, runqueue_t
*rq
)
799 prio_array_t
*target
= rq
->active
;
802 target
= rq
->expired
;
803 enqueue_task(p
, target
);
804 inc_nr_running(p
, rq
);
808 * __activate_idle_task - move idle task to the _front_ of runqueue.
810 static inline void __activate_idle_task(task_t
*p
, runqueue_t
*rq
)
812 enqueue_task_head(p
, rq
->active
);
813 inc_nr_running(p
, rq
);
817 * Recalculate p->normal_prio and p->prio after having slept,
818 * updating the sleep-average too:
820 static int recalc_task_prio(task_t
*p
, unsigned long long now
)
822 /* Caller must always ensure 'now >= p->timestamp' */
823 unsigned long sleep_time
= now
- p
->timestamp
;
828 if (likely(sleep_time
> 0)) {
830 * This ceiling is set to the lowest priority that would allow
831 * a task to be reinserted into the active array on timeslice
834 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
836 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
838 * Prevents user tasks from achieving best priority
839 * with one single large enough sleep.
841 p
->sleep_avg
= ceiling
;
843 * Using INTERACTIVE_SLEEP() as a ceiling places a
844 * nice(0) task 1ms sleep away from promotion, and
845 * gives it 700ms to round-robin with no chance of
846 * being demoted. This is more than generous, so
847 * mark this sleep as non-interactive to prevent the
848 * on-runqueue bonus logic from intervening should
849 * this task not receive cpu immediately.
851 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
854 * Tasks waking from uninterruptible sleep are
855 * limited in their sleep_avg rise as they
856 * are likely to be waiting on I/O
858 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
859 if (p
->sleep_avg
>= ceiling
)
861 else if (p
->sleep_avg
+ sleep_time
>=
863 p
->sleep_avg
= ceiling
;
869 * This code gives a bonus to interactive tasks.
871 * The boost works by updating the 'average sleep time'
872 * value here, based on ->timestamp. The more time a
873 * task spends sleeping, the higher the average gets -
874 * and the higher the priority boost gets as well.
876 p
->sleep_avg
+= sleep_time
;
879 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
880 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
883 return effective_prio(p
);
887 * activate_task - move a task to the runqueue and do priority recalculation
889 * Update all the scheduling statistics stuff. (sleep average
890 * calculation, priority modifiers, etc.)
892 static void activate_task(task_t
*p
, runqueue_t
*rq
, int local
)
894 unsigned long long now
;
899 /* Compensate for drifting sched_clock */
900 runqueue_t
*this_rq
= this_rq();
901 now
= (now
- this_rq
->timestamp_last_tick
)
902 + rq
->timestamp_last_tick
;
907 p
->prio
= recalc_task_prio(p
, now
);
910 * This checks to make sure it's not an uninterruptible task
911 * that is now waking up.
913 if (p
->sleep_type
== SLEEP_NORMAL
) {
915 * Tasks which were woken up by interrupts (ie. hw events)
916 * are most likely of interactive nature. So we give them
917 * the credit of extending their sleep time to the period
918 * of time they spend on the runqueue, waiting for execution
919 * on a CPU, first time around:
922 p
->sleep_type
= SLEEP_INTERRUPTED
;
925 * Normal first-time wakeups get a credit too for
926 * on-runqueue time, but it will be weighted down:
928 p
->sleep_type
= SLEEP_INTERACTIVE
;
933 __activate_task(p
, rq
);
937 * deactivate_task - remove a task from the runqueue.
939 static void deactivate_task(struct task_struct
*p
, runqueue_t
*rq
)
941 dec_nr_running(p
, rq
);
942 dequeue_task(p
, p
->array
);
947 * resched_task - mark a task 'to be rescheduled now'.
949 * On UP this means the setting of the need_resched flag, on SMP it
950 * might also involve a cross-CPU call to trigger the scheduler on
955 #ifndef tsk_is_polling
956 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
959 static void resched_task(task_t
*p
)
963 assert_spin_locked(&task_rq(p
)->lock
);
965 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
968 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
971 if (cpu
== smp_processor_id())
974 /* NEED_RESCHED must be visible before we test polling */
976 if (!tsk_is_polling(p
))
977 smp_send_reschedule(cpu
);
980 static inline void resched_task(task_t
*p
)
982 assert_spin_locked(&task_rq(p
)->lock
);
983 set_tsk_need_resched(p
);
988 * task_curr - is this task currently executing on a CPU?
989 * @p: the task in question.
991 inline int task_curr(const task_t
*p
)
993 return cpu_curr(task_cpu(p
)) == p
;
996 /* Used instead of source_load when we know the type == 0 */
997 unsigned long weighted_cpuload(const int cpu
)
999 return cpu_rq(cpu
)->raw_weighted_load
;
1004 struct list_head list
;
1009 struct completion done
;
1013 * The task's runqueue lock must be held.
1014 * Returns true if you have to wait for migration thread.
1016 static int migrate_task(task_t
*p
, int dest_cpu
, migration_req_t
*req
)
1018 runqueue_t
*rq
= task_rq(p
);
1021 * If the task is not on a runqueue (and not running), then
1022 * it is sufficient to simply update the task's cpu field.
1024 if (!p
->array
&& !task_running(rq
, p
)) {
1025 set_task_cpu(p
, dest_cpu
);
1029 init_completion(&req
->done
);
1031 req
->dest_cpu
= dest_cpu
;
1032 list_add(&req
->list
, &rq
->migration_queue
);
1037 * wait_task_inactive - wait for a thread to unschedule.
1039 * The caller must ensure that the task *will* unschedule sometime soon,
1040 * else this function might spin for a *long* time. This function can't
1041 * be called with interrupts off, or it may introduce deadlock with
1042 * smp_call_function() if an IPI is sent by the same process we are
1043 * waiting to become inactive.
1045 void wait_task_inactive(task_t
*p
)
1047 unsigned long flags
;
1052 rq
= task_rq_lock(p
, &flags
);
1053 /* Must be off runqueue entirely, not preempted. */
1054 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1055 /* If it's preempted, we yield. It could be a while. */
1056 preempted
= !task_running(rq
, p
);
1057 task_rq_unlock(rq
, &flags
);
1063 task_rq_unlock(rq
, &flags
);
1067 * kick_process - kick a running thread to enter/exit the kernel
1068 * @p: the to-be-kicked thread
1070 * Cause a process which is running on another CPU to enter
1071 * kernel-mode, without any delay. (to get signals handled.)
1073 * NOTE: this function doesnt have to take the runqueue lock,
1074 * because all it wants to ensure is that the remote task enters
1075 * the kernel. If the IPI races and the task has been migrated
1076 * to another CPU then no harm is done and the purpose has been
1079 void kick_process(task_t
*p
)
1085 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1086 smp_send_reschedule(cpu
);
1091 * Return a low guess at the load of a migration-source cpu weighted
1092 * according to the scheduling class and "nice" value.
1094 * We want to under-estimate the load of migration sources, to
1095 * balance conservatively.
1097 static inline unsigned long source_load(int cpu
, int type
)
1099 runqueue_t
*rq
= cpu_rq(cpu
);
1102 return rq
->raw_weighted_load
;
1104 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1108 * Return a high guess at the load of a migration-target cpu weighted
1109 * according to the scheduling class and "nice" value.
1111 static inline unsigned long target_load(int cpu
, int type
)
1113 runqueue_t
*rq
= cpu_rq(cpu
);
1116 return rq
->raw_weighted_load
;
1118 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1122 * Return the average load per task on the cpu's run queue
1124 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1126 runqueue_t
*rq
= cpu_rq(cpu
);
1127 unsigned long n
= rq
->nr_running
;
1129 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1133 * find_idlest_group finds and returns the least busy CPU group within the
1136 static struct sched_group
*
1137 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1139 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1140 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1141 int load_idx
= sd
->forkexec_idx
;
1142 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1145 unsigned long load
, avg_load
;
1149 /* Skip over this group if it has no CPUs allowed */
1150 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1153 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1155 /* Tally up the load of all CPUs in the group */
1158 for_each_cpu_mask(i
, group
->cpumask
) {
1159 /* Bias balancing toward cpus of our domain */
1161 load
= source_load(i
, load_idx
);
1163 load
= target_load(i
, load_idx
);
1168 /* Adjust by relative CPU power of the group */
1169 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1172 this_load
= avg_load
;
1174 } else if (avg_load
< min_load
) {
1175 min_load
= avg_load
;
1179 group
= group
->next
;
1180 } while (group
!= sd
->groups
);
1182 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1188 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1191 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1194 unsigned long load
, min_load
= ULONG_MAX
;
1198 /* Traverse only the allowed CPUs */
1199 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1201 for_each_cpu_mask(i
, tmp
) {
1202 load
= weighted_cpuload(i
);
1204 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1214 * sched_balance_self: balance the current task (running on cpu) in domains
1215 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1218 * Balance, ie. select the least loaded group.
1220 * Returns the target CPU number, or the same CPU if no balancing is needed.
1222 * preempt must be disabled.
1224 static int sched_balance_self(int cpu
, int flag
)
1226 struct task_struct
*t
= current
;
1227 struct sched_domain
*tmp
, *sd
= NULL
;
1229 for_each_domain(cpu
, tmp
) {
1231 * If power savings logic is enabled for a domain, stop there.
1233 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1235 if (tmp
->flags
& flag
)
1241 struct sched_group
*group
;
1246 group
= find_idlest_group(sd
, t
, cpu
);
1250 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1251 if (new_cpu
== -1 || new_cpu
== cpu
)
1254 /* Now try balancing at a lower domain level */
1258 weight
= cpus_weight(span
);
1259 for_each_domain(cpu
, tmp
) {
1260 if (weight
<= cpus_weight(tmp
->span
))
1262 if (tmp
->flags
& flag
)
1265 /* while loop will break here if sd == NULL */
1271 #endif /* CONFIG_SMP */
1274 * wake_idle() will wake a task on an idle cpu if task->cpu is
1275 * not idle and an idle cpu is available. The span of cpus to
1276 * search starts with cpus closest then further out as needed,
1277 * so we always favor a closer, idle cpu.
1279 * Returns the CPU we should wake onto.
1281 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1282 static int wake_idle(int cpu
, task_t
*p
)
1285 struct sched_domain
*sd
;
1291 for_each_domain(cpu
, sd
) {
1292 if (sd
->flags
& SD_WAKE_IDLE
) {
1293 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1294 for_each_cpu_mask(i
, tmp
) {
1305 static inline int wake_idle(int cpu
, task_t
*p
)
1312 * try_to_wake_up - wake up a thread
1313 * @p: the to-be-woken-up thread
1314 * @state: the mask of task states that can be woken
1315 * @sync: do a synchronous wakeup?
1317 * Put it on the run-queue if it's not already there. The "current"
1318 * thread is always on the run-queue (except when the actual
1319 * re-schedule is in progress), and as such you're allowed to do
1320 * the simpler "current->state = TASK_RUNNING" to mark yourself
1321 * runnable without the overhead of this.
1323 * returns failure only if the task is already active.
1325 static int try_to_wake_up(task_t
*p
, unsigned int state
, int sync
)
1327 int cpu
, this_cpu
, success
= 0;
1328 unsigned long flags
;
1332 unsigned long load
, this_load
;
1333 struct sched_domain
*sd
, *this_sd
= NULL
;
1337 rq
= task_rq_lock(p
, &flags
);
1338 old_state
= p
->state
;
1339 if (!(old_state
& state
))
1346 this_cpu
= smp_processor_id();
1349 if (unlikely(task_running(rq
, p
)))
1354 schedstat_inc(rq
, ttwu_cnt
);
1355 if (cpu
== this_cpu
) {
1356 schedstat_inc(rq
, ttwu_local
);
1360 for_each_domain(this_cpu
, sd
) {
1361 if (cpu_isset(cpu
, sd
->span
)) {
1362 schedstat_inc(sd
, ttwu_wake_remote
);
1368 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1372 * Check for affine wakeup and passive balancing possibilities.
1375 int idx
= this_sd
->wake_idx
;
1376 unsigned int imbalance
;
1378 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1380 load
= source_load(cpu
, idx
);
1381 this_load
= target_load(this_cpu
, idx
);
1383 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1385 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1386 unsigned long tl
= this_load
;
1387 unsigned long tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1390 * If sync wakeup then subtract the (maximum possible)
1391 * effect of the currently running task from the load
1392 * of the current CPU:
1395 tl
-= current
->load_weight
;
1398 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1399 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1401 * This domain has SD_WAKE_AFFINE and
1402 * p is cache cold in this domain, and
1403 * there is no bad imbalance.
1405 schedstat_inc(this_sd
, ttwu_move_affine
);
1411 * Start passive balancing when half the imbalance_pct
1414 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1415 if (imbalance
*this_load
<= 100*load
) {
1416 schedstat_inc(this_sd
, ttwu_move_balance
);
1422 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1424 new_cpu
= wake_idle(new_cpu
, p
);
1425 if (new_cpu
!= cpu
) {
1426 set_task_cpu(p
, new_cpu
);
1427 task_rq_unlock(rq
, &flags
);
1428 /* might preempt at this point */
1429 rq
= task_rq_lock(p
, &flags
);
1430 old_state
= p
->state
;
1431 if (!(old_state
& state
))
1436 this_cpu
= smp_processor_id();
1441 #endif /* CONFIG_SMP */
1442 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1443 rq
->nr_uninterruptible
--;
1445 * Tasks on involuntary sleep don't earn
1446 * sleep_avg beyond just interactive state.
1448 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1452 * Tasks that have marked their sleep as noninteractive get
1453 * woken up with their sleep average not weighted in an
1456 if (old_state
& TASK_NONINTERACTIVE
)
1457 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1460 activate_task(p
, rq
, cpu
== this_cpu
);
1462 * Sync wakeups (i.e. those types of wakeups where the waker
1463 * has indicated that it will leave the CPU in short order)
1464 * don't trigger a preemption, if the woken up task will run on
1465 * this cpu. (in this case the 'I will reschedule' promise of
1466 * the waker guarantees that the freshly woken up task is going
1467 * to be considered on this CPU.)
1469 if (!sync
|| cpu
!= this_cpu
) {
1470 if (TASK_PREEMPTS_CURR(p
, rq
))
1471 resched_task(rq
->curr
);
1476 p
->state
= TASK_RUNNING
;
1478 task_rq_unlock(rq
, &flags
);
1483 int fastcall
wake_up_process(task_t
*p
)
1485 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1486 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1489 EXPORT_SYMBOL(wake_up_process
);
1491 int fastcall
wake_up_state(task_t
*p
, unsigned int state
)
1493 return try_to_wake_up(p
, state
, 0);
1497 * Perform scheduler related setup for a newly forked process p.
1498 * p is forked by current.
1500 void fastcall
sched_fork(task_t
*p
, int clone_flags
)
1502 int cpu
= get_cpu();
1505 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1507 set_task_cpu(p
, cpu
);
1510 * We mark the process as running here, but have not actually
1511 * inserted it onto the runqueue yet. This guarantees that
1512 * nobody will actually run it, and a signal or other external
1513 * event cannot wake it up and insert it on the runqueue either.
1515 p
->state
= TASK_RUNNING
;
1518 * Make sure we do not leak PI boosting priority to the child:
1520 p
->prio
= current
->normal_prio
;
1522 INIT_LIST_HEAD(&p
->run_list
);
1524 #ifdef CONFIG_SCHEDSTATS
1525 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1527 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1530 #ifdef CONFIG_PREEMPT
1531 /* Want to start with kernel preemption disabled. */
1532 task_thread_info(p
)->preempt_count
= 1;
1535 * Share the timeslice between parent and child, thus the
1536 * total amount of pending timeslices in the system doesn't change,
1537 * resulting in more scheduling fairness.
1539 local_irq_disable();
1540 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1542 * The remainder of the first timeslice might be recovered by
1543 * the parent if the child exits early enough.
1545 p
->first_time_slice
= 1;
1546 current
->time_slice
>>= 1;
1547 p
->timestamp
= sched_clock();
1548 if (unlikely(!current
->time_slice
)) {
1550 * This case is rare, it happens when the parent has only
1551 * a single jiffy left from its timeslice. Taking the
1552 * runqueue lock is not a problem.
1554 current
->time_slice
= 1;
1562 * wake_up_new_task - wake up a newly created task for the first time.
1564 * This function will do some initial scheduler statistics housekeeping
1565 * that must be done for every newly created context, then puts the task
1566 * on the runqueue and wakes it.
1568 void fastcall
wake_up_new_task(task_t
*p
, unsigned long clone_flags
)
1570 unsigned long flags
;
1572 runqueue_t
*rq
, *this_rq
;
1574 rq
= task_rq_lock(p
, &flags
);
1575 BUG_ON(p
->state
!= TASK_RUNNING
);
1576 this_cpu
= smp_processor_id();
1580 * We decrease the sleep average of forking parents
1581 * and children as well, to keep max-interactive tasks
1582 * from forking tasks that are max-interactive. The parent
1583 * (current) is done further down, under its lock.
1585 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1586 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1588 p
->prio
= effective_prio(p
);
1590 if (likely(cpu
== this_cpu
)) {
1591 if (!(clone_flags
& CLONE_VM
)) {
1593 * The VM isn't cloned, so we're in a good position to
1594 * do child-runs-first in anticipation of an exec. This
1595 * usually avoids a lot of COW overhead.
1597 if (unlikely(!current
->array
))
1598 __activate_task(p
, rq
);
1600 p
->prio
= current
->prio
;
1601 p
->normal_prio
= current
->normal_prio
;
1602 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1603 p
->array
= current
->array
;
1604 p
->array
->nr_active
++;
1605 inc_nr_running(p
, rq
);
1609 /* Run child last */
1610 __activate_task(p
, rq
);
1612 * We skip the following code due to cpu == this_cpu
1614 * task_rq_unlock(rq, &flags);
1615 * this_rq = task_rq_lock(current, &flags);
1619 this_rq
= cpu_rq(this_cpu
);
1622 * Not the local CPU - must adjust timestamp. This should
1623 * get optimised away in the !CONFIG_SMP case.
1625 p
->timestamp
= (p
->timestamp
- this_rq
->timestamp_last_tick
)
1626 + rq
->timestamp_last_tick
;
1627 __activate_task(p
, rq
);
1628 if (TASK_PREEMPTS_CURR(p
, rq
))
1629 resched_task(rq
->curr
);
1632 * Parent and child are on different CPUs, now get the
1633 * parent runqueue to update the parent's ->sleep_avg:
1635 task_rq_unlock(rq
, &flags
);
1636 this_rq
= task_rq_lock(current
, &flags
);
1638 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1639 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1640 task_rq_unlock(this_rq
, &flags
);
1644 * Potentially available exiting-child timeslices are
1645 * retrieved here - this way the parent does not get
1646 * penalized for creating too many threads.
1648 * (this cannot be used to 'generate' timeslices
1649 * artificially, because any timeslice recovered here
1650 * was given away by the parent in the first place.)
1652 void fastcall
sched_exit(task_t
*p
)
1654 unsigned long flags
;
1658 * If the child was a (relative-) CPU hog then decrease
1659 * the sleep_avg of the parent as well.
1661 rq
= task_rq_lock(p
->parent
, &flags
);
1662 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1663 p
->parent
->time_slice
+= p
->time_slice
;
1664 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1665 p
->parent
->time_slice
= task_timeslice(p
);
1667 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1668 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1669 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1671 task_rq_unlock(rq
, &flags
);
1675 * prepare_task_switch - prepare to switch tasks
1676 * @rq: the runqueue preparing to switch
1677 * @next: the task we are going to switch to.
1679 * This is called with the rq lock held and interrupts off. It must
1680 * be paired with a subsequent finish_task_switch after the context
1683 * prepare_task_switch sets up locking and calls architecture specific
1686 static inline void prepare_task_switch(runqueue_t
*rq
, task_t
*next
)
1688 prepare_lock_switch(rq
, next
);
1689 prepare_arch_switch(next
);
1693 * finish_task_switch - clean up after a task-switch
1694 * @rq: runqueue associated with task-switch
1695 * @prev: the thread we just switched away from.
1697 * finish_task_switch must be called after the context switch, paired
1698 * with a prepare_task_switch call before the context switch.
1699 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1700 * and do any other architecture-specific cleanup actions.
1702 * Note that we may have delayed dropping an mm in context_switch(). If
1703 * so, we finish that here outside of the runqueue lock. (Doing it
1704 * with the lock held can cause deadlocks; see schedule() for
1707 static inline void finish_task_switch(runqueue_t
*rq
, task_t
*prev
)
1708 __releases(rq
->lock
)
1710 struct mm_struct
*mm
= rq
->prev_mm
;
1711 unsigned long prev_task_flags
;
1716 * A task struct has one reference for the use as "current".
1717 * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1718 * calls schedule one last time. The schedule call will never return,
1719 * and the scheduled task must drop that reference.
1720 * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1721 * still held, otherwise prev could be scheduled on another cpu, die
1722 * there before we look at prev->state, and then the reference would
1724 * Manfred Spraul <manfred@colorfullife.com>
1726 prev_task_flags
= prev
->flags
;
1727 finish_arch_switch(prev
);
1728 finish_lock_switch(rq
, prev
);
1731 if (unlikely(prev_task_flags
& PF_DEAD
)) {
1733 * Remove function-return probe instances associated with this
1734 * task and put them back on the free list.
1736 kprobe_flush_task(prev
);
1737 put_task_struct(prev
);
1742 * schedule_tail - first thing a freshly forked thread must call.
1743 * @prev: the thread we just switched away from.
1745 asmlinkage
void schedule_tail(task_t
*prev
)
1746 __releases(rq
->lock
)
1748 runqueue_t
*rq
= this_rq();
1749 finish_task_switch(rq
, prev
);
1750 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1751 /* In this case, finish_task_switch does not reenable preemption */
1754 if (current
->set_child_tid
)
1755 put_user(current
->pid
, current
->set_child_tid
);
1759 * context_switch - switch to the new MM and the new
1760 * thread's register state.
1763 task_t
* context_switch(runqueue_t
*rq
, task_t
*prev
, task_t
*next
)
1765 struct mm_struct
*mm
= next
->mm
;
1766 struct mm_struct
*oldmm
= prev
->active_mm
;
1768 if (unlikely(!mm
)) {
1769 next
->active_mm
= oldmm
;
1770 atomic_inc(&oldmm
->mm_count
);
1771 enter_lazy_tlb(oldmm
, next
);
1773 switch_mm(oldmm
, mm
, next
);
1775 if (unlikely(!prev
->mm
)) {
1776 prev
->active_mm
= NULL
;
1777 WARN_ON(rq
->prev_mm
);
1778 rq
->prev_mm
= oldmm
;
1781 /* Here we just switch the register state and the stack. */
1782 switch_to(prev
, next
, prev
);
1788 * nr_running, nr_uninterruptible and nr_context_switches:
1790 * externally visible scheduler statistics: current number of runnable
1791 * threads, current number of uninterruptible-sleeping threads, total
1792 * number of context switches performed since bootup.
1794 unsigned long nr_running(void)
1796 unsigned long i
, sum
= 0;
1798 for_each_online_cpu(i
)
1799 sum
+= cpu_rq(i
)->nr_running
;
1804 unsigned long nr_uninterruptible(void)
1806 unsigned long i
, sum
= 0;
1808 for_each_possible_cpu(i
)
1809 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1812 * Since we read the counters lockless, it might be slightly
1813 * inaccurate. Do not allow it to go below zero though:
1815 if (unlikely((long)sum
< 0))
1821 unsigned long long nr_context_switches(void)
1824 unsigned long long sum
= 0;
1826 for_each_possible_cpu(i
)
1827 sum
+= cpu_rq(i
)->nr_switches
;
1832 unsigned long nr_iowait(void)
1834 unsigned long i
, sum
= 0;
1836 for_each_possible_cpu(i
)
1837 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1842 unsigned long nr_active(void)
1844 unsigned long i
, running
= 0, uninterruptible
= 0;
1846 for_each_online_cpu(i
) {
1847 running
+= cpu_rq(i
)->nr_running
;
1848 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1851 if (unlikely((long)uninterruptible
< 0))
1852 uninterruptible
= 0;
1854 return running
+ uninterruptible
;
1860 * double_rq_lock - safely lock two runqueues
1862 * Note this does not disable interrupts like task_rq_lock,
1863 * you need to do so manually before calling.
1865 static void double_rq_lock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1866 __acquires(rq1
->lock
)
1867 __acquires(rq2
->lock
)
1870 spin_lock(&rq1
->lock
);
1871 __acquire(rq2
->lock
); /* Fake it out ;) */
1874 spin_lock(&rq1
->lock
);
1875 spin_lock(&rq2
->lock
);
1877 spin_lock(&rq2
->lock
);
1878 spin_lock(&rq1
->lock
);
1884 * double_rq_unlock - safely unlock two runqueues
1886 * Note this does not restore interrupts like task_rq_unlock,
1887 * you need to do so manually after calling.
1889 static void double_rq_unlock(runqueue_t
*rq1
, runqueue_t
*rq2
)
1890 __releases(rq1
->lock
)
1891 __releases(rq2
->lock
)
1893 spin_unlock(&rq1
->lock
);
1895 spin_unlock(&rq2
->lock
);
1897 __release(rq2
->lock
);
1901 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1903 static void double_lock_balance(runqueue_t
*this_rq
, runqueue_t
*busiest
)
1904 __releases(this_rq
->lock
)
1905 __acquires(busiest
->lock
)
1906 __acquires(this_rq
->lock
)
1908 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1909 if (busiest
< this_rq
) {
1910 spin_unlock(&this_rq
->lock
);
1911 spin_lock(&busiest
->lock
);
1912 spin_lock(&this_rq
->lock
);
1914 spin_lock(&busiest
->lock
);
1919 * If dest_cpu is allowed for this process, migrate the task to it.
1920 * This is accomplished by forcing the cpu_allowed mask to only
1921 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1922 * the cpu_allowed mask is restored.
1924 static void sched_migrate_task(task_t
*p
, int dest_cpu
)
1926 migration_req_t req
;
1928 unsigned long flags
;
1930 rq
= task_rq_lock(p
, &flags
);
1931 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
1932 || unlikely(cpu_is_offline(dest_cpu
)))
1935 /* force the process onto the specified CPU */
1936 if (migrate_task(p
, dest_cpu
, &req
)) {
1937 /* Need to wait for migration thread (might exit: take ref). */
1938 struct task_struct
*mt
= rq
->migration_thread
;
1939 get_task_struct(mt
);
1940 task_rq_unlock(rq
, &flags
);
1941 wake_up_process(mt
);
1942 put_task_struct(mt
);
1943 wait_for_completion(&req
.done
);
1947 task_rq_unlock(rq
, &flags
);
1951 * sched_exec - execve() is a valuable balancing opportunity, because at
1952 * this point the task has the smallest effective memory and cache footprint.
1954 void sched_exec(void)
1956 int new_cpu
, this_cpu
= get_cpu();
1957 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
1959 if (new_cpu
!= this_cpu
)
1960 sched_migrate_task(current
, new_cpu
);
1964 * pull_task - move a task from a remote runqueue to the local runqueue.
1965 * Both runqueues must be locked.
1968 void pull_task(runqueue_t
*src_rq
, prio_array_t
*src_array
, task_t
*p
,
1969 runqueue_t
*this_rq
, prio_array_t
*this_array
, int this_cpu
)
1971 dequeue_task(p
, src_array
);
1972 dec_nr_running(p
, src_rq
);
1973 set_task_cpu(p
, this_cpu
);
1974 inc_nr_running(p
, this_rq
);
1975 enqueue_task(p
, this_array
);
1976 p
->timestamp
= (p
->timestamp
- src_rq
->timestamp_last_tick
)
1977 + this_rq
->timestamp_last_tick
;
1979 * Note that idle threads have a prio of MAX_PRIO, for this test
1980 * to be always true for them.
1982 if (TASK_PREEMPTS_CURR(p
, this_rq
))
1983 resched_task(this_rq
->curr
);
1987 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1990 int can_migrate_task(task_t
*p
, runqueue_t
*rq
, int this_cpu
,
1991 struct sched_domain
*sd
, enum idle_type idle
,
1995 * We do not migrate tasks that are:
1996 * 1) running (obviously), or
1997 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1998 * 3) are cache-hot on their current CPU.
2000 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2004 if (task_running(rq
, p
))
2008 * Aggressive migration if:
2009 * 1) task is cache cold, or
2010 * 2) too many balance attempts have failed.
2013 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
)
2016 if (task_hot(p
, rq
->timestamp_last_tick
, sd
))
2021 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2023 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2024 * load from busiest to this_rq, as part of a balancing operation within
2025 * "domain". Returns the number of tasks moved.
2027 * Called with both runqueues locked.
2029 static int move_tasks(runqueue_t
*this_rq
, int this_cpu
, runqueue_t
*busiest
,
2030 unsigned long max_nr_move
, unsigned long max_load_move
,
2031 struct sched_domain
*sd
, enum idle_type idle
,
2034 prio_array_t
*array
, *dst_array
;
2035 struct list_head
*head
, *curr
;
2036 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, busiest_best_prio
;
2037 int busiest_best_prio_seen
;
2038 int skip_for_load
; /* skip the task based on weighted load issues */
2042 if (max_nr_move
== 0 || max_load_move
== 0)
2045 rem_load_move
= max_load_move
;
2047 this_best_prio
= rq_best_prio(this_rq
);
2048 busiest_best_prio
= rq_best_prio(busiest
);
2050 * Enable handling of the case where there is more than one task
2051 * with the best priority. If the current running task is one
2052 * of those with prio==busiest_best_prio we know it won't be moved
2053 * and therefore it's safe to override the skip (based on load) of
2054 * any task we find with that prio.
2056 busiest_best_prio_seen
= busiest_best_prio
== busiest
->curr
->prio
;
2059 * We first consider expired tasks. Those will likely not be
2060 * executed in the near future, and they are most likely to
2061 * be cache-cold, thus switching CPUs has the least effect
2064 if (busiest
->expired
->nr_active
) {
2065 array
= busiest
->expired
;
2066 dst_array
= this_rq
->expired
;
2068 array
= busiest
->active
;
2069 dst_array
= this_rq
->active
;
2073 /* Start searching at priority 0: */
2077 idx
= sched_find_first_bit(array
->bitmap
);
2079 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2080 if (idx
>= MAX_PRIO
) {
2081 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2082 array
= busiest
->active
;
2083 dst_array
= this_rq
->active
;
2089 head
= array
->queue
+ idx
;
2092 tmp
= list_entry(curr
, task_t
, run_list
);
2097 * To help distribute high priority tasks accross CPUs we don't
2098 * skip a task if it will be the highest priority task (i.e. smallest
2099 * prio value) on its new queue regardless of its load weight
2101 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2102 if (skip_for_load
&& idx
< this_best_prio
)
2103 skip_for_load
= !busiest_best_prio_seen
&& idx
== busiest_best_prio
;
2104 if (skip_for_load
||
2105 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2106 busiest_best_prio_seen
|= idx
== busiest_best_prio
;
2113 #ifdef CONFIG_SCHEDSTATS
2114 if (task_hot(tmp
, busiest
->timestamp_last_tick
, sd
))
2115 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2118 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2120 rem_load_move
-= tmp
->load_weight
;
2123 * We only want to steal up to the prescribed number of tasks
2124 * and the prescribed amount of weighted load.
2126 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2127 if (idx
< this_best_prio
)
2128 this_best_prio
= idx
;
2136 * Right now, this is the only place pull_task() is called,
2137 * so we can safely collect pull_task() stats here rather than
2138 * inside pull_task().
2140 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2143 *all_pinned
= pinned
;
2148 * find_busiest_group finds and returns the busiest CPU group within the
2149 * domain. It calculates and returns the amount of weighted load which should be
2150 * moved to restore balance via the imbalance parameter.
2152 static struct sched_group
*
2153 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2154 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
)
2156 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2157 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2158 unsigned long max_pull
;
2159 unsigned long busiest_load_per_task
, busiest_nr_running
;
2160 unsigned long this_load_per_task
, this_nr_running
;
2162 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2163 int power_savings_balance
= 1;
2164 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2165 unsigned long min_nr_running
= ULONG_MAX
;
2166 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2169 max_load
= this_load
= total_load
= total_pwr
= 0;
2170 busiest_load_per_task
= busiest_nr_running
= 0;
2171 this_load_per_task
= this_nr_running
= 0;
2172 if (idle
== NOT_IDLE
)
2173 load_idx
= sd
->busy_idx
;
2174 else if (idle
== NEWLY_IDLE
)
2175 load_idx
= sd
->newidle_idx
;
2177 load_idx
= sd
->idle_idx
;
2180 unsigned long load
, group_capacity
;
2183 unsigned long sum_nr_running
, sum_weighted_load
;
2185 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2187 /* Tally up the load of all CPUs in the group */
2188 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2190 for_each_cpu_mask(i
, group
->cpumask
) {
2191 runqueue_t
*rq
= cpu_rq(i
);
2193 if (*sd_idle
&& !idle_cpu(i
))
2196 /* Bias balancing toward cpus of our domain */
2198 load
= target_load(i
, load_idx
);
2200 load
= source_load(i
, load_idx
);
2203 sum_nr_running
+= rq
->nr_running
;
2204 sum_weighted_load
+= rq
->raw_weighted_load
;
2207 total_load
+= avg_load
;
2208 total_pwr
+= group
->cpu_power
;
2210 /* Adjust by relative CPU power of the group */
2211 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2213 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2216 this_load
= avg_load
;
2218 this_nr_running
= sum_nr_running
;
2219 this_load_per_task
= sum_weighted_load
;
2220 } else if (avg_load
> max_load
&&
2221 sum_nr_running
> group_capacity
) {
2222 max_load
= avg_load
;
2224 busiest_nr_running
= sum_nr_running
;
2225 busiest_load_per_task
= sum_weighted_load
;
2228 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2230 * Busy processors will not participate in power savings
2233 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2237 * If the local group is idle or completely loaded
2238 * no need to do power savings balance at this domain
2240 if (local_group
&& (this_nr_running
>= group_capacity
||
2242 power_savings_balance
= 0;
2245 * If a group is already running at full capacity or idle,
2246 * don't include that group in power savings calculations
2248 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2253 * Calculate the group which has the least non-idle load.
2254 * This is the group from where we need to pick up the load
2257 if ((sum_nr_running
< min_nr_running
) ||
2258 (sum_nr_running
== min_nr_running
&&
2259 first_cpu(group
->cpumask
) <
2260 first_cpu(group_min
->cpumask
))) {
2262 min_nr_running
= sum_nr_running
;
2263 min_load_per_task
= sum_weighted_load
/
2268 * Calculate the group which is almost near its
2269 * capacity but still has some space to pick up some load
2270 * from other group and save more power
2272 if (sum_nr_running
<= group_capacity
- 1)
2273 if (sum_nr_running
> leader_nr_running
||
2274 (sum_nr_running
== leader_nr_running
&&
2275 first_cpu(group
->cpumask
) >
2276 first_cpu(group_leader
->cpumask
))) {
2277 group_leader
= group
;
2278 leader_nr_running
= sum_nr_running
;
2283 group
= group
->next
;
2284 } while (group
!= sd
->groups
);
2286 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2289 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2291 if (this_load
>= avg_load
||
2292 100*max_load
<= sd
->imbalance_pct
*this_load
)
2295 busiest_load_per_task
/= busiest_nr_running
;
2297 * We're trying to get all the cpus to the average_load, so we don't
2298 * want to push ourselves above the average load, nor do we wish to
2299 * reduce the max loaded cpu below the average load, as either of these
2300 * actions would just result in more rebalancing later, and ping-pong
2301 * tasks around. Thus we look for the minimum possible imbalance.
2302 * Negative imbalances (*we* are more loaded than anyone else) will
2303 * be counted as no imbalance for these purposes -- we can't fix that
2304 * by pulling tasks to us. Be careful of negative numbers as they'll
2305 * appear as very large values with unsigned longs.
2307 if (max_load
<= busiest_load_per_task
)
2311 * In the presence of smp nice balancing, certain scenarios can have
2312 * max load less than avg load(as we skip the groups at or below
2313 * its cpu_power, while calculating max_load..)
2315 if (max_load
< avg_load
) {
2317 goto small_imbalance
;
2320 /* Don't want to pull so many tasks that a group would go idle */
2321 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2323 /* How much load to actually move to equalise the imbalance */
2324 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2325 (avg_load
- this_load
) * this->cpu_power
)
2329 * if *imbalance is less than the average load per runnable task
2330 * there is no gaurantee that any tasks will be moved so we'll have
2331 * a think about bumping its value to force at least one task to be
2334 if (*imbalance
< busiest_load_per_task
) {
2335 unsigned long pwr_now
, pwr_move
;
2340 pwr_move
= pwr_now
= 0;
2342 if (this_nr_running
) {
2343 this_load_per_task
/= this_nr_running
;
2344 if (busiest_load_per_task
> this_load_per_task
)
2347 this_load_per_task
= SCHED_LOAD_SCALE
;
2349 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2350 *imbalance
= busiest_load_per_task
;
2355 * OK, we don't have enough imbalance to justify moving tasks,
2356 * however we may be able to increase total CPU power used by
2360 pwr_now
+= busiest
->cpu_power
*
2361 min(busiest_load_per_task
, max_load
);
2362 pwr_now
+= this->cpu_power
*
2363 min(this_load_per_task
, this_load
);
2364 pwr_now
/= SCHED_LOAD_SCALE
;
2366 /* Amount of load we'd subtract */
2367 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/busiest
->cpu_power
;
2369 pwr_move
+= busiest
->cpu_power
*
2370 min(busiest_load_per_task
, max_load
- tmp
);
2372 /* Amount of load we'd add */
2373 if (max_load
*busiest
->cpu_power
<
2374 busiest_load_per_task
*SCHED_LOAD_SCALE
)
2375 tmp
= max_load
*busiest
->cpu_power
/this->cpu_power
;
2377 tmp
= busiest_load_per_task
*SCHED_LOAD_SCALE
/this->cpu_power
;
2378 pwr_move
+= this->cpu_power
*min(this_load_per_task
, this_load
+ tmp
);
2379 pwr_move
/= SCHED_LOAD_SCALE
;
2381 /* Move if we gain throughput */
2382 if (pwr_move
<= pwr_now
)
2385 *imbalance
= busiest_load_per_task
;
2391 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2392 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2395 if (this == group_leader
&& group_leader
!= group_min
) {
2396 *imbalance
= min_load_per_task
;
2406 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2408 static runqueue_t
*find_busiest_queue(struct sched_group
*group
,
2409 enum idle_type idle
, unsigned long imbalance
)
2411 unsigned long max_load
= 0;
2412 runqueue_t
*busiest
= NULL
, *rqi
;
2415 for_each_cpu_mask(i
, group
->cpumask
) {
2418 if (rqi
->nr_running
== 1 && rqi
->raw_weighted_load
> imbalance
)
2421 if (rqi
->raw_weighted_load
> max_load
) {
2422 max_load
= rqi
->raw_weighted_load
;
2431 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2432 * so long as it is large enough.
2434 #define MAX_PINNED_INTERVAL 512
2436 #define minus_1_or_zero(n) ((n) > 0 ? (n) - 1 : 0)
2438 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2439 * tasks if there is an imbalance.
2441 * Called with this_rq unlocked.
2443 static int load_balance(int this_cpu
, runqueue_t
*this_rq
,
2444 struct sched_domain
*sd
, enum idle_type idle
)
2446 struct sched_group
*group
;
2447 runqueue_t
*busiest
;
2448 unsigned long imbalance
;
2449 int nr_moved
, all_pinned
= 0;
2450 int active_balance
= 0;
2453 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2454 !sched_smt_power_savings
)
2457 schedstat_inc(sd
, lb_cnt
[idle
]);
2459 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
);
2461 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2465 busiest
= find_busiest_queue(group
, idle
, imbalance
);
2467 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2471 BUG_ON(busiest
== this_rq
);
2473 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2476 if (busiest
->nr_running
> 1) {
2478 * Attempt to move tasks. If find_busiest_group has found
2479 * an imbalance but busiest->nr_running <= 1, the group is
2480 * still unbalanced. nr_moved simply stays zero, so it is
2481 * correctly treated as an imbalance.
2483 double_rq_lock(this_rq
, busiest
);
2484 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2485 minus_1_or_zero(busiest
->nr_running
),
2486 imbalance
, sd
, idle
, &all_pinned
);
2487 double_rq_unlock(this_rq
, busiest
);
2489 /* All tasks on this runqueue were pinned by CPU affinity */
2490 if (unlikely(all_pinned
))
2495 schedstat_inc(sd
, lb_failed
[idle
]);
2496 sd
->nr_balance_failed
++;
2498 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2500 spin_lock(&busiest
->lock
);
2502 /* don't kick the migration_thread, if the curr
2503 * task on busiest cpu can't be moved to this_cpu
2505 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2506 spin_unlock(&busiest
->lock
);
2508 goto out_one_pinned
;
2511 if (!busiest
->active_balance
) {
2512 busiest
->active_balance
= 1;
2513 busiest
->push_cpu
= this_cpu
;
2516 spin_unlock(&busiest
->lock
);
2518 wake_up_process(busiest
->migration_thread
);
2521 * We've kicked active balancing, reset the failure
2524 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2527 sd
->nr_balance_failed
= 0;
2529 if (likely(!active_balance
)) {
2530 /* We were unbalanced, so reset the balancing interval */
2531 sd
->balance_interval
= sd
->min_interval
;
2534 * If we've begun active balancing, start to back off. This
2535 * case may not be covered by the all_pinned logic if there
2536 * is only 1 task on the busy runqueue (because we don't call
2539 if (sd
->balance_interval
< sd
->max_interval
)
2540 sd
->balance_interval
*= 2;
2543 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2544 !sched_smt_power_savings
)
2549 schedstat_inc(sd
, lb_balanced
[idle
]);
2551 sd
->nr_balance_failed
= 0;
2554 /* tune up the balancing interval */
2555 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2556 (sd
->balance_interval
< sd
->max_interval
))
2557 sd
->balance_interval
*= 2;
2559 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2565 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2566 * tasks if there is an imbalance.
2568 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2569 * this_rq is locked.
2571 static int load_balance_newidle(int this_cpu
, runqueue_t
*this_rq
,
2572 struct sched_domain
*sd
)
2574 struct sched_group
*group
;
2575 runqueue_t
*busiest
= NULL
;
2576 unsigned long imbalance
;
2580 if (sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2583 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2584 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
, &sd_idle
);
2586 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2590 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
);
2592 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2596 BUG_ON(busiest
== this_rq
);
2598 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2601 if (busiest
->nr_running
> 1) {
2602 /* Attempt to move tasks */
2603 double_lock_balance(this_rq
, busiest
);
2604 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2605 minus_1_or_zero(busiest
->nr_running
),
2606 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2607 spin_unlock(&busiest
->lock
);
2611 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2612 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
)
2615 sd
->nr_balance_failed
= 0;
2620 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2621 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&& !sched_smt_power_savings
)
2623 sd
->nr_balance_failed
= 0;
2628 * idle_balance is called by schedule() if this_cpu is about to become
2629 * idle. Attempts to pull tasks from other CPUs.
2631 static void idle_balance(int this_cpu
, runqueue_t
*this_rq
)
2633 struct sched_domain
*sd
;
2635 for_each_domain(this_cpu
, sd
) {
2636 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2637 if (load_balance_newidle(this_cpu
, this_rq
, sd
)) {
2638 /* We've pulled tasks over so stop searching */
2646 * active_load_balance is run by migration threads. It pushes running tasks
2647 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2648 * running on each physical CPU where possible, and avoids physical /
2649 * logical imbalances.
2651 * Called with busiest_rq locked.
2653 static void active_load_balance(runqueue_t
*busiest_rq
, int busiest_cpu
)
2655 struct sched_domain
*sd
;
2656 runqueue_t
*target_rq
;
2657 int target_cpu
= busiest_rq
->push_cpu
;
2659 if (busiest_rq
->nr_running
<= 1)
2660 /* no task to move */
2663 target_rq
= cpu_rq(target_cpu
);
2666 * This condition is "impossible", if it occurs
2667 * we need to fix it. Originally reported by
2668 * Bjorn Helgaas on a 128-cpu setup.
2670 BUG_ON(busiest_rq
== target_rq
);
2672 /* move a task from busiest_rq to target_rq */
2673 double_lock_balance(busiest_rq
, target_rq
);
2675 /* Search for an sd spanning us and the target CPU. */
2676 for_each_domain(target_cpu
, sd
) {
2677 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2678 cpu_isset(busiest_cpu
, sd
->span
))
2682 if (unlikely(sd
== NULL
))
2685 schedstat_inc(sd
, alb_cnt
);
2687 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2688 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
, NULL
))
2689 schedstat_inc(sd
, alb_pushed
);
2691 schedstat_inc(sd
, alb_failed
);
2693 spin_unlock(&target_rq
->lock
);
2697 * rebalance_tick will get called every timer tick, on every CPU.
2699 * It checks each scheduling domain to see if it is due to be balanced,
2700 * and initiates a balancing operation if so.
2702 * Balancing parameters are set up in arch_init_sched_domains.
2705 /* Don't have all balancing operations going off at once */
2706 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2708 static void rebalance_tick(int this_cpu
, runqueue_t
*this_rq
,
2709 enum idle_type idle
)
2711 unsigned long old_load
, this_load
;
2712 unsigned long j
= jiffies
+ CPU_OFFSET(this_cpu
);
2713 struct sched_domain
*sd
;
2716 this_load
= this_rq
->raw_weighted_load
;
2717 /* Update our load */
2718 for (i
= 0; i
< 3; i
++) {
2719 unsigned long new_load
= this_load
;
2721 old_load
= this_rq
->cpu_load
[i
];
2723 * Round up the averaging division if load is increasing. This
2724 * prevents us from getting stuck on 9 if the load is 10, for
2727 if (new_load
> old_load
)
2728 new_load
+= scale
-1;
2729 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2732 for_each_domain(this_cpu
, sd
) {
2733 unsigned long interval
;
2735 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2738 interval
= sd
->balance_interval
;
2739 if (idle
!= SCHED_IDLE
)
2740 interval
*= sd
->busy_factor
;
2742 /* scale ms to jiffies */
2743 interval
= msecs_to_jiffies(interval
);
2744 if (unlikely(!interval
))
2747 if (j
- sd
->last_balance
>= interval
) {
2748 if (load_balance(this_cpu
, this_rq
, sd
, idle
)) {
2750 * We've pulled tasks over so either we're no
2751 * longer idle, or one of our SMT siblings is
2756 sd
->last_balance
+= interval
;
2762 * on UP we do not need to balance between CPUs:
2764 static inline void rebalance_tick(int cpu
, runqueue_t
*rq
, enum idle_type idle
)
2767 static inline void idle_balance(int cpu
, runqueue_t
*rq
)
2772 static inline int wake_priority_sleeper(runqueue_t
*rq
)
2775 #ifdef CONFIG_SCHED_SMT
2776 spin_lock(&rq
->lock
);
2778 * If an SMT sibling task has been put to sleep for priority
2779 * reasons reschedule the idle task to see if it can now run.
2781 if (rq
->nr_running
) {
2782 resched_task(rq
->idle
);
2785 spin_unlock(&rq
->lock
);
2790 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2792 EXPORT_PER_CPU_SYMBOL(kstat
);
2795 * This is called on clock ticks and on context switches.
2796 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2798 static inline void update_cpu_clock(task_t
*p
, runqueue_t
*rq
,
2799 unsigned long long now
)
2801 unsigned long long last
= max(p
->timestamp
, rq
->timestamp_last_tick
);
2802 p
->sched_time
+= now
- last
;
2806 * Return current->sched_time plus any more ns on the sched_clock
2807 * that have not yet been banked.
2809 unsigned long long current_sched_time(const task_t
*tsk
)
2811 unsigned long long ns
;
2812 unsigned long flags
;
2813 local_irq_save(flags
);
2814 ns
= max(tsk
->timestamp
, task_rq(tsk
)->timestamp_last_tick
);
2815 ns
= tsk
->sched_time
+ (sched_clock() - ns
);
2816 local_irq_restore(flags
);
2821 * We place interactive tasks back into the active array, if possible.
2823 * To guarantee that this does not starve expired tasks we ignore the
2824 * interactivity of a task if the first expired task had to wait more
2825 * than a 'reasonable' amount of time. This deadline timeout is
2826 * load-dependent, as the frequency of array switched decreases with
2827 * increasing number of running tasks. We also ignore the interactivity
2828 * if a better static_prio task has expired:
2830 #define EXPIRED_STARVING(rq) \
2831 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2832 (jiffies - (rq)->expired_timestamp >= \
2833 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2834 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2837 * Account user cpu time to a process.
2838 * @p: the process that the cpu time gets accounted to
2839 * @hardirq_offset: the offset to subtract from hardirq_count()
2840 * @cputime: the cpu time spent in user space since the last update
2842 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
2844 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2847 p
->utime
= cputime_add(p
->utime
, cputime
);
2849 /* Add user time to cpustat. */
2850 tmp
= cputime_to_cputime64(cputime
);
2851 if (TASK_NICE(p
) > 0)
2852 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
2854 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
2858 * Account system cpu time to a process.
2859 * @p: the process that the cpu time gets accounted to
2860 * @hardirq_offset: the offset to subtract from hardirq_count()
2861 * @cputime: the cpu time spent in kernel space since the last update
2863 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2866 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2867 runqueue_t
*rq
= this_rq();
2870 p
->stime
= cputime_add(p
->stime
, cputime
);
2872 /* Add system time to cpustat. */
2873 tmp
= cputime_to_cputime64(cputime
);
2874 if (hardirq_count() - hardirq_offset
)
2875 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
2876 else if (softirq_count())
2877 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
2878 else if (p
!= rq
->idle
)
2879 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
2880 else if (atomic_read(&rq
->nr_iowait
) > 0)
2881 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2883 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2884 /* Account for system time used */
2885 acct_update_integrals(p
);
2889 * Account for involuntary wait time.
2890 * @p: the process from which the cpu time has been stolen
2891 * @steal: the cpu time spent in involuntary wait
2893 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
2895 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2896 cputime64_t tmp
= cputime_to_cputime64(steal
);
2897 runqueue_t
*rq
= this_rq();
2899 if (p
== rq
->idle
) {
2900 p
->stime
= cputime_add(p
->stime
, steal
);
2901 if (atomic_read(&rq
->nr_iowait
) > 0)
2902 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
2904 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
2906 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
2910 * This function gets called by the timer code, with HZ frequency.
2911 * We call it with interrupts disabled.
2913 * It also gets called by the fork code, when changing the parent's
2916 void scheduler_tick(void)
2918 int cpu
= smp_processor_id();
2919 runqueue_t
*rq
= this_rq();
2920 task_t
*p
= current
;
2921 unsigned long long now
= sched_clock();
2923 update_cpu_clock(p
, rq
, now
);
2925 rq
->timestamp_last_tick
= now
;
2927 if (p
== rq
->idle
) {
2928 if (wake_priority_sleeper(rq
))
2930 rebalance_tick(cpu
, rq
, SCHED_IDLE
);
2934 /* Task might have expired already, but not scheduled off yet */
2935 if (p
->array
!= rq
->active
) {
2936 set_tsk_need_resched(p
);
2939 spin_lock(&rq
->lock
);
2941 * The task was running during this tick - update the
2942 * time slice counter. Note: we do not update a thread's
2943 * priority until it either goes to sleep or uses up its
2944 * timeslice. This makes it possible for interactive tasks
2945 * to use up their timeslices at their highest priority levels.
2949 * RR tasks need a special form of timeslice management.
2950 * FIFO tasks have no timeslices.
2952 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
2953 p
->time_slice
= task_timeslice(p
);
2954 p
->first_time_slice
= 0;
2955 set_tsk_need_resched(p
);
2957 /* put it at the end of the queue: */
2958 requeue_task(p
, rq
->active
);
2962 if (!--p
->time_slice
) {
2963 dequeue_task(p
, rq
->active
);
2964 set_tsk_need_resched(p
);
2965 p
->prio
= effective_prio(p
);
2966 p
->time_slice
= task_timeslice(p
);
2967 p
->first_time_slice
= 0;
2969 if (!rq
->expired_timestamp
)
2970 rq
->expired_timestamp
= jiffies
;
2971 if (!TASK_INTERACTIVE(p
) || EXPIRED_STARVING(rq
)) {
2972 enqueue_task(p
, rq
->expired
);
2973 if (p
->static_prio
< rq
->best_expired_prio
)
2974 rq
->best_expired_prio
= p
->static_prio
;
2976 enqueue_task(p
, rq
->active
);
2979 * Prevent a too long timeslice allowing a task to monopolize
2980 * the CPU. We do this by splitting up the timeslice into
2983 * Note: this does not mean the task's timeslices expire or
2984 * get lost in any way, they just might be preempted by
2985 * another task of equal priority. (one with higher
2986 * priority would have preempted this task already.) We
2987 * requeue this task to the end of the list on this priority
2988 * level, which is in essence a round-robin of tasks with
2991 * This only applies to tasks in the interactive
2992 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2994 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
2995 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
2996 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
2997 (p
->array
== rq
->active
)) {
2999 requeue_task(p
, rq
->active
);
3000 set_tsk_need_resched(p
);
3004 spin_unlock(&rq
->lock
);
3006 rebalance_tick(cpu
, rq
, NOT_IDLE
);
3009 #ifdef CONFIG_SCHED_SMT
3010 static inline void wakeup_busy_runqueue(runqueue_t
*rq
)
3012 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3013 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
3014 resched_task(rq
->idle
);
3018 * Called with interrupt disabled and this_rq's runqueue locked.
3020 static void wake_sleeping_dependent(int this_cpu
)
3022 struct sched_domain
*tmp
, *sd
= NULL
;
3025 for_each_domain(this_cpu
, tmp
) {
3026 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3035 for_each_cpu_mask(i
, sd
->span
) {
3036 runqueue_t
*smt_rq
= cpu_rq(i
);
3040 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3043 wakeup_busy_runqueue(smt_rq
);
3044 spin_unlock(&smt_rq
->lock
);
3049 * number of 'lost' timeslices this task wont be able to fully
3050 * utilize, if another task runs on a sibling. This models the
3051 * slowdown effect of other tasks running on siblings:
3053 static inline unsigned long smt_slice(task_t
*p
, struct sched_domain
*sd
)
3055 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
3059 * To minimise lock contention and not have to drop this_rq's runlock we only
3060 * trylock the sibling runqueues and bypass those runqueues if we fail to
3061 * acquire their lock. As we only trylock the normal locking order does not
3062 * need to be obeyed.
3064 static int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
, task_t
*p
)
3066 struct sched_domain
*tmp
, *sd
= NULL
;
3069 /* kernel/rt threads do not participate in dependent sleeping */
3070 if (!p
->mm
|| rt_task(p
))
3073 for_each_domain(this_cpu
, tmp
) {
3074 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3083 for_each_cpu_mask(i
, sd
->span
) {
3091 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3094 smt_curr
= smt_rq
->curr
;
3100 * If a user task with lower static priority than the
3101 * running task on the SMT sibling is trying to schedule,
3102 * delay it till there is proportionately less timeslice
3103 * left of the sibling task to prevent a lower priority
3104 * task from using an unfair proportion of the
3105 * physical cpu's resources. -ck
3107 if (rt_task(smt_curr
)) {
3109 * With real time tasks we run non-rt tasks only
3110 * per_cpu_gain% of the time.
3112 if ((jiffies
% DEF_TIMESLICE
) >
3113 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3116 if (smt_curr
->static_prio
< p
->static_prio
&&
3117 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3118 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3122 spin_unlock(&smt_rq
->lock
);
3127 static inline void wake_sleeping_dependent(int this_cpu
)
3131 static inline int dependent_sleeper(int this_cpu
, runqueue_t
*this_rq
,
3138 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3140 void fastcall
add_preempt_count(int val
)
3145 BUG_ON((preempt_count() < 0));
3146 preempt_count() += val
;
3148 * Spinlock count overflowing soon?
3150 BUG_ON((preempt_count() & PREEMPT_MASK
) >= PREEMPT_MASK
-10);
3152 EXPORT_SYMBOL(add_preempt_count
);
3154 void fastcall
sub_preempt_count(int val
)
3159 BUG_ON(val
> preempt_count());
3161 * Is the spinlock portion underflowing?
3163 BUG_ON((val
< PREEMPT_MASK
) && !(preempt_count() & PREEMPT_MASK
));
3164 preempt_count() -= val
;
3166 EXPORT_SYMBOL(sub_preempt_count
);
3170 static inline int interactive_sleep(enum sleep_type sleep_type
)
3172 return (sleep_type
== SLEEP_INTERACTIVE
||
3173 sleep_type
== SLEEP_INTERRUPTED
);
3177 * schedule() is the main scheduler function.
3179 asmlinkage
void __sched
schedule(void)
3182 task_t
*prev
, *next
;
3184 prio_array_t
*array
;
3185 struct list_head
*queue
;
3186 unsigned long long now
;
3187 unsigned long run_time
;
3188 int cpu
, idx
, new_prio
;
3191 * Test if we are atomic. Since do_exit() needs to call into
3192 * schedule() atomically, we ignore that path for now.
3193 * Otherwise, whine if we are scheduling when we should not be.
3195 if (unlikely(in_atomic() && !current
->exit_state
)) {
3196 printk(KERN_ERR
"BUG: scheduling while atomic: "
3198 current
->comm
, preempt_count(), current
->pid
);
3201 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3206 release_kernel_lock(prev
);
3207 need_resched_nonpreemptible
:
3211 * The idle thread is not allowed to schedule!
3212 * Remove this check after it has been exercised a bit.
3214 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3215 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3219 schedstat_inc(rq
, sched_cnt
);
3220 now
= sched_clock();
3221 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3222 run_time
= now
- prev
->timestamp
;
3223 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3226 run_time
= NS_MAX_SLEEP_AVG
;
3229 * Tasks charged proportionately less run_time at high sleep_avg to
3230 * delay them losing their interactive status
3232 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3234 spin_lock_irq(&rq
->lock
);
3236 if (unlikely(prev
->flags
& PF_DEAD
))
3237 prev
->state
= EXIT_DEAD
;
3239 switch_count
= &prev
->nivcsw
;
3240 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3241 switch_count
= &prev
->nvcsw
;
3242 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3243 unlikely(signal_pending(prev
))))
3244 prev
->state
= TASK_RUNNING
;
3246 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3247 rq
->nr_uninterruptible
++;
3248 deactivate_task(prev
, rq
);
3252 cpu
= smp_processor_id();
3253 if (unlikely(!rq
->nr_running
)) {
3254 idle_balance(cpu
, rq
);
3255 if (!rq
->nr_running
) {
3257 rq
->expired_timestamp
= 0;
3258 wake_sleeping_dependent(cpu
);
3264 if (unlikely(!array
->nr_active
)) {
3266 * Switch the active and expired arrays.
3268 schedstat_inc(rq
, sched_switch
);
3269 rq
->active
= rq
->expired
;
3270 rq
->expired
= array
;
3272 rq
->expired_timestamp
= 0;
3273 rq
->best_expired_prio
= MAX_PRIO
;
3276 idx
= sched_find_first_bit(array
->bitmap
);
3277 queue
= array
->queue
+ idx
;
3278 next
= list_entry(queue
->next
, task_t
, run_list
);
3280 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3281 unsigned long long delta
= now
- next
->timestamp
;
3282 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3285 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3286 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3288 array
= next
->array
;
3289 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3291 if (unlikely(next
->prio
!= new_prio
)) {
3292 dequeue_task(next
, array
);
3293 next
->prio
= new_prio
;
3294 enqueue_task(next
, array
);
3297 next
->sleep_type
= SLEEP_NORMAL
;
3298 if (dependent_sleeper(cpu
, rq
, next
))
3301 if (next
== rq
->idle
)
3302 schedstat_inc(rq
, sched_goidle
);
3304 prefetch_stack(next
);
3305 clear_tsk_need_resched(prev
);
3306 rcu_qsctr_inc(task_cpu(prev
));
3308 update_cpu_clock(prev
, rq
, now
);
3310 prev
->sleep_avg
-= run_time
;
3311 if ((long)prev
->sleep_avg
<= 0)
3312 prev
->sleep_avg
= 0;
3313 prev
->timestamp
= prev
->last_ran
= now
;
3315 sched_info_switch(prev
, next
);
3316 if (likely(prev
!= next
)) {
3317 next
->timestamp
= now
;
3322 prepare_task_switch(rq
, next
);
3323 prev
= context_switch(rq
, prev
, next
);
3326 * this_rq must be evaluated again because prev may have moved
3327 * CPUs since it called schedule(), thus the 'rq' on its stack
3328 * frame will be invalid.
3330 finish_task_switch(this_rq(), prev
);
3332 spin_unlock_irq(&rq
->lock
);
3335 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3336 goto need_resched_nonpreemptible
;
3337 preempt_enable_no_resched();
3338 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3342 EXPORT_SYMBOL(schedule
);
3344 #ifdef CONFIG_PREEMPT
3346 * this is is the entry point to schedule() from in-kernel preemption
3347 * off of preempt_enable. Kernel preemptions off return from interrupt
3348 * occur there and call schedule directly.
3350 asmlinkage
void __sched
preempt_schedule(void)
3352 struct thread_info
*ti
= current_thread_info();
3353 #ifdef CONFIG_PREEMPT_BKL
3354 struct task_struct
*task
= current
;
3355 int saved_lock_depth
;
3358 * If there is a non-zero preempt_count or interrupts are disabled,
3359 * we do not want to preempt the current task. Just return..
3361 if (unlikely(ti
->preempt_count
|| irqs_disabled()))
3365 add_preempt_count(PREEMPT_ACTIVE
);
3367 * We keep the big kernel semaphore locked, but we
3368 * clear ->lock_depth so that schedule() doesnt
3369 * auto-release the semaphore:
3371 #ifdef CONFIG_PREEMPT_BKL
3372 saved_lock_depth
= task
->lock_depth
;
3373 task
->lock_depth
= -1;
3376 #ifdef CONFIG_PREEMPT_BKL
3377 task
->lock_depth
= saved_lock_depth
;
3379 sub_preempt_count(PREEMPT_ACTIVE
);
3381 /* we could miss a preemption opportunity between schedule and now */
3383 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3387 EXPORT_SYMBOL(preempt_schedule
);
3390 * this is is the entry point to schedule() from kernel preemption
3391 * off of irq context.
3392 * Note, that this is called and return with irqs disabled. This will
3393 * protect us against recursive calling from irq.
3395 asmlinkage
void __sched
preempt_schedule_irq(void)
3397 struct thread_info
*ti
= current_thread_info();
3398 #ifdef CONFIG_PREEMPT_BKL
3399 struct task_struct
*task
= current
;
3400 int saved_lock_depth
;
3402 /* Catch callers which need to be fixed*/
3403 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3406 add_preempt_count(PREEMPT_ACTIVE
);
3408 * We keep the big kernel semaphore locked, but we
3409 * clear ->lock_depth so that schedule() doesnt
3410 * auto-release the semaphore:
3412 #ifdef CONFIG_PREEMPT_BKL
3413 saved_lock_depth
= task
->lock_depth
;
3414 task
->lock_depth
= -1;
3418 local_irq_disable();
3419 #ifdef CONFIG_PREEMPT_BKL
3420 task
->lock_depth
= saved_lock_depth
;
3422 sub_preempt_count(PREEMPT_ACTIVE
);
3424 /* we could miss a preemption opportunity between schedule and now */
3426 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3430 #endif /* CONFIG_PREEMPT */
3432 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3435 task_t
*p
= curr
->private;
3436 return try_to_wake_up(p
, mode
, sync
);
3439 EXPORT_SYMBOL(default_wake_function
);
3442 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3443 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3444 * number) then we wake all the non-exclusive tasks and one exclusive task.
3446 * There are circumstances in which we can try to wake a task which has already
3447 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3448 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3450 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3451 int nr_exclusive
, int sync
, void *key
)
3453 struct list_head
*tmp
, *next
;
3455 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3458 curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3459 flags
= curr
->flags
;
3460 if (curr
->func(curr
, mode
, sync
, key
) &&
3461 (flags
& WQ_FLAG_EXCLUSIVE
) &&
3468 * __wake_up - wake up threads blocked on a waitqueue.
3470 * @mode: which threads
3471 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3472 * @key: is directly passed to the wakeup function
3474 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3475 int nr_exclusive
, void *key
)
3477 unsigned long flags
;
3479 spin_lock_irqsave(&q
->lock
, flags
);
3480 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3481 spin_unlock_irqrestore(&q
->lock
, flags
);
3484 EXPORT_SYMBOL(__wake_up
);
3487 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3489 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3491 __wake_up_common(q
, mode
, 1, 0, NULL
);
3495 * __wake_up_sync - wake up threads blocked on a waitqueue.
3497 * @mode: which threads
3498 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3500 * The sync wakeup differs that the waker knows that it will schedule
3501 * away soon, so while the target thread will be woken up, it will not
3502 * be migrated to another CPU - ie. the two threads are 'synchronized'
3503 * with each other. This can prevent needless bouncing between CPUs.
3505 * On UP it can prevent extra preemption.
3508 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3510 unsigned long flags
;
3516 if (unlikely(!nr_exclusive
))
3519 spin_lock_irqsave(&q
->lock
, flags
);
3520 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3521 spin_unlock_irqrestore(&q
->lock
, flags
);
3523 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3525 void fastcall
complete(struct completion
*x
)
3527 unsigned long flags
;
3529 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3531 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3533 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3535 EXPORT_SYMBOL(complete
);
3537 void fastcall
complete_all(struct completion
*x
)
3539 unsigned long flags
;
3541 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3542 x
->done
+= UINT_MAX
/2;
3543 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3545 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3547 EXPORT_SYMBOL(complete_all
);
3549 void fastcall __sched
wait_for_completion(struct completion
*x
)
3552 spin_lock_irq(&x
->wait
.lock
);
3554 DECLARE_WAITQUEUE(wait
, current
);
3556 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3557 __add_wait_queue_tail(&x
->wait
, &wait
);
3559 __set_current_state(TASK_UNINTERRUPTIBLE
);
3560 spin_unlock_irq(&x
->wait
.lock
);
3562 spin_lock_irq(&x
->wait
.lock
);
3564 __remove_wait_queue(&x
->wait
, &wait
);
3567 spin_unlock_irq(&x
->wait
.lock
);
3569 EXPORT_SYMBOL(wait_for_completion
);
3571 unsigned long fastcall __sched
3572 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3576 spin_lock_irq(&x
->wait
.lock
);
3578 DECLARE_WAITQUEUE(wait
, current
);
3580 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3581 __add_wait_queue_tail(&x
->wait
, &wait
);
3583 __set_current_state(TASK_UNINTERRUPTIBLE
);
3584 spin_unlock_irq(&x
->wait
.lock
);
3585 timeout
= schedule_timeout(timeout
);
3586 spin_lock_irq(&x
->wait
.lock
);
3588 __remove_wait_queue(&x
->wait
, &wait
);
3592 __remove_wait_queue(&x
->wait
, &wait
);
3596 spin_unlock_irq(&x
->wait
.lock
);
3599 EXPORT_SYMBOL(wait_for_completion_timeout
);
3601 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3607 spin_lock_irq(&x
->wait
.lock
);
3609 DECLARE_WAITQUEUE(wait
, current
);
3611 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3612 __add_wait_queue_tail(&x
->wait
, &wait
);
3614 if (signal_pending(current
)) {
3616 __remove_wait_queue(&x
->wait
, &wait
);
3619 __set_current_state(TASK_INTERRUPTIBLE
);
3620 spin_unlock_irq(&x
->wait
.lock
);
3622 spin_lock_irq(&x
->wait
.lock
);
3624 __remove_wait_queue(&x
->wait
, &wait
);
3628 spin_unlock_irq(&x
->wait
.lock
);
3632 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3634 unsigned long fastcall __sched
3635 wait_for_completion_interruptible_timeout(struct completion
*x
,
3636 unsigned long timeout
)
3640 spin_lock_irq(&x
->wait
.lock
);
3642 DECLARE_WAITQUEUE(wait
, current
);
3644 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3645 __add_wait_queue_tail(&x
->wait
, &wait
);
3647 if (signal_pending(current
)) {
3648 timeout
= -ERESTARTSYS
;
3649 __remove_wait_queue(&x
->wait
, &wait
);
3652 __set_current_state(TASK_INTERRUPTIBLE
);
3653 spin_unlock_irq(&x
->wait
.lock
);
3654 timeout
= schedule_timeout(timeout
);
3655 spin_lock_irq(&x
->wait
.lock
);
3657 __remove_wait_queue(&x
->wait
, &wait
);
3661 __remove_wait_queue(&x
->wait
, &wait
);
3665 spin_unlock_irq(&x
->wait
.lock
);
3668 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3671 #define SLEEP_ON_VAR \
3672 unsigned long flags; \
3673 wait_queue_t wait; \
3674 init_waitqueue_entry(&wait, current);
3676 #define SLEEP_ON_HEAD \
3677 spin_lock_irqsave(&q->lock,flags); \
3678 __add_wait_queue(q, &wait); \
3679 spin_unlock(&q->lock);
3681 #define SLEEP_ON_TAIL \
3682 spin_lock_irq(&q->lock); \
3683 __remove_wait_queue(q, &wait); \
3684 spin_unlock_irqrestore(&q->lock, flags);
3686 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3690 current
->state
= TASK_INTERRUPTIBLE
;
3697 EXPORT_SYMBOL(interruptible_sleep_on
);
3699 long fastcall __sched
3700 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3704 current
->state
= TASK_INTERRUPTIBLE
;
3707 timeout
= schedule_timeout(timeout
);
3713 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3715 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3719 current
->state
= TASK_UNINTERRUPTIBLE
;
3726 EXPORT_SYMBOL(sleep_on
);
3728 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3732 current
->state
= TASK_UNINTERRUPTIBLE
;
3735 timeout
= schedule_timeout(timeout
);
3741 EXPORT_SYMBOL(sleep_on_timeout
);
3743 #ifdef CONFIG_RT_MUTEXES
3746 * rt_mutex_setprio - set the current priority of a task
3748 * @prio: prio value (kernel-internal form)
3750 * This function changes the 'effective' priority of a task. It does
3751 * not touch ->normal_prio like __setscheduler().
3753 * Used by the rt_mutex code to implement priority inheritance logic.
3755 void rt_mutex_setprio(task_t
*p
, int prio
)
3757 unsigned long flags
;
3758 prio_array_t
*array
;
3762 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3764 rq
= task_rq_lock(p
, &flags
);
3769 dequeue_task(p
, array
);
3774 * If changing to an RT priority then queue it
3775 * in the active array!
3779 enqueue_task(p
, array
);
3781 * Reschedule if we are currently running on this runqueue and
3782 * our priority decreased, or if we are not currently running on
3783 * this runqueue and our priority is higher than the current's
3785 if (task_running(rq
, p
)) {
3786 if (p
->prio
> oldprio
)
3787 resched_task(rq
->curr
);
3788 } else if (TASK_PREEMPTS_CURR(p
, rq
))
3789 resched_task(rq
->curr
);
3791 task_rq_unlock(rq
, &flags
);
3796 void set_user_nice(task_t
*p
, long nice
)
3798 unsigned long flags
;
3799 prio_array_t
*array
;
3801 int old_prio
, delta
;
3803 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3806 * We have to be careful, if called from sys_setpriority(),
3807 * the task might be in the middle of scheduling on another CPU.
3809 rq
= task_rq_lock(p
, &flags
);
3811 * The RT priorities are set via sched_setscheduler(), but we still
3812 * allow the 'normal' nice value to be set - but as expected
3813 * it wont have any effect on scheduling until the task is
3814 * not SCHED_NORMAL/SCHED_BATCH:
3816 if (has_rt_policy(p
)) {
3817 p
->static_prio
= NICE_TO_PRIO(nice
);
3822 dequeue_task(p
, array
);
3823 dec_raw_weighted_load(rq
, p
);
3826 p
->static_prio
= NICE_TO_PRIO(nice
);
3829 p
->prio
= effective_prio(p
);
3830 delta
= p
->prio
- old_prio
;
3833 enqueue_task(p
, array
);
3834 inc_raw_weighted_load(rq
, p
);
3836 * If the task increased its priority or is running and
3837 * lowered its priority, then reschedule its CPU:
3839 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3840 resched_task(rq
->curr
);
3843 task_rq_unlock(rq
, &flags
);
3845 EXPORT_SYMBOL(set_user_nice
);
3848 * can_nice - check if a task can reduce its nice value
3852 int can_nice(const task_t
*p
, const int nice
)
3854 /* convert nice value [19,-20] to rlimit style value [1,40] */
3855 int nice_rlim
= 20 - nice
;
3856 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3857 capable(CAP_SYS_NICE
));
3860 #ifdef __ARCH_WANT_SYS_NICE
3863 * sys_nice - change the priority of the current process.
3864 * @increment: priority increment
3866 * sys_setpriority is a more generic, but much slower function that
3867 * does similar things.
3869 asmlinkage
long sys_nice(int increment
)
3875 * Setpriority might change our priority at the same moment.
3876 * We don't have to worry. Conceptually one call occurs first
3877 * and we have a single winner.
3879 if (increment
< -40)
3884 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
3890 if (increment
< 0 && !can_nice(current
, nice
))
3893 retval
= security_task_setnice(current
, nice
);
3897 set_user_nice(current
, nice
);
3904 * task_prio - return the priority value of a given task.
3905 * @p: the task in question.
3907 * This is the priority value as seen by users in /proc.
3908 * RT tasks are offset by -200. Normal tasks are centered
3909 * around 0, value goes from -16 to +15.
3911 int task_prio(const task_t
*p
)
3913 return p
->prio
- MAX_RT_PRIO
;
3917 * task_nice - return the nice value of a given task.
3918 * @p: the task in question.
3920 int task_nice(const task_t
*p
)
3922 return TASK_NICE(p
);
3924 EXPORT_SYMBOL_GPL(task_nice
);
3927 * idle_cpu - is a given cpu idle currently?
3928 * @cpu: the processor in question.
3930 int idle_cpu(int cpu
)
3932 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
3936 * idle_task - return the idle task for a given cpu.
3937 * @cpu: the processor in question.
3939 task_t
*idle_task(int cpu
)
3941 return cpu_rq(cpu
)->idle
;
3945 * find_process_by_pid - find a process with a matching PID value.
3946 * @pid: the pid in question.
3948 static inline task_t
*find_process_by_pid(pid_t pid
)
3950 return pid
? find_task_by_pid(pid
) : current
;
3953 /* Actually do priority change: must hold rq lock. */
3954 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
3958 p
->rt_priority
= prio
;
3959 p
->normal_prio
= normal_prio(p
);
3960 /* we are holding p->pi_lock already */
3961 p
->prio
= rt_mutex_getprio(p
);
3963 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
3965 if (policy
== SCHED_BATCH
)
3971 * sched_setscheduler - change the scheduling policy and/or RT priority of
3973 * @p: the task in question.
3974 * @policy: new policy.
3975 * @param: structure containing the new RT priority.
3977 int sched_setscheduler(struct task_struct
*p
, int policy
,
3978 struct sched_param
*param
)
3981 int oldprio
, oldpolicy
= -1;
3982 prio_array_t
*array
;
3983 unsigned long flags
;
3986 /* may grab non-irq protected spin_locks */
3987 BUG_ON(in_interrupt());
3989 /* double check policy once rq lock held */
3991 policy
= oldpolicy
= p
->policy
;
3992 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3993 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
3996 * Valid priorities for SCHED_FIFO and SCHED_RR are
3997 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4000 if (param
->sched_priority
< 0 ||
4001 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4002 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4004 if ((policy
== SCHED_NORMAL
|| policy
== SCHED_BATCH
)
4005 != (param
->sched_priority
== 0))
4009 * Allow unprivileged RT tasks to decrease priority:
4011 if (!capable(CAP_SYS_NICE
)) {
4013 * can't change policy, except between SCHED_NORMAL
4016 if (((policy
!= SCHED_NORMAL
&& p
->policy
!= SCHED_BATCH
) &&
4017 (policy
!= SCHED_BATCH
&& p
->policy
!= SCHED_NORMAL
)) &&
4018 !p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
4020 /* can't increase priority */
4021 if ((policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
) &&
4022 param
->sched_priority
> p
->rt_priority
&&
4023 param
->sched_priority
>
4024 p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
)
4026 /* can't change other user's priorities */
4027 if ((current
->euid
!= p
->euid
) &&
4028 (current
->euid
!= p
->uid
))
4032 retval
= security_task_setscheduler(p
, policy
, param
);
4036 * make sure no PI-waiters arrive (or leave) while we are
4037 * changing the priority of the task:
4039 spin_lock_irqsave(&p
->pi_lock
, flags
);
4041 * To be able to change p->policy safely, the apropriate
4042 * runqueue lock must be held.
4044 rq
= __task_rq_lock(p
);
4045 /* recheck policy now with rq lock held */
4046 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4047 policy
= oldpolicy
= -1;
4048 __task_rq_unlock(rq
);
4049 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4054 deactivate_task(p
, rq
);
4056 __setscheduler(p
, policy
, param
->sched_priority
);
4058 __activate_task(p
, rq
);
4060 * Reschedule if we are currently running on this runqueue and
4061 * our priority decreased, or if we are not currently running on
4062 * this runqueue and our priority is higher than the current's
4064 if (task_running(rq
, p
)) {
4065 if (p
->prio
> oldprio
)
4066 resched_task(rq
->curr
);
4067 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4068 resched_task(rq
->curr
);
4070 __task_rq_unlock(rq
);
4071 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4075 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4078 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4081 struct sched_param lparam
;
4082 struct task_struct
*p
;
4084 if (!param
|| pid
< 0)
4086 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4088 read_lock_irq(&tasklist_lock
);
4089 p
= find_process_by_pid(pid
);
4091 read_unlock_irq(&tasklist_lock
);
4095 read_unlock_irq(&tasklist_lock
);
4096 retval
= sched_setscheduler(p
, policy
, &lparam
);
4102 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4103 * @pid: the pid in question.
4104 * @policy: new policy.
4105 * @param: structure containing the new RT priority.
4107 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4108 struct sched_param __user
*param
)
4110 /* negative values for policy are not valid */
4114 return do_sched_setscheduler(pid
, policy
, param
);
4118 * sys_sched_setparam - set/change the RT priority of a thread
4119 * @pid: the pid in question.
4120 * @param: structure containing the new RT priority.
4122 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4124 return do_sched_setscheduler(pid
, -1, param
);
4128 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4129 * @pid: the pid in question.
4131 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4133 int retval
= -EINVAL
;
4140 read_lock(&tasklist_lock
);
4141 p
= find_process_by_pid(pid
);
4143 retval
= security_task_getscheduler(p
);
4147 read_unlock(&tasklist_lock
);
4154 * sys_sched_getscheduler - get the RT priority of a thread
4155 * @pid: the pid in question.
4156 * @param: structure containing the RT priority.
4158 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4160 struct sched_param lp
;
4161 int retval
= -EINVAL
;
4164 if (!param
|| pid
< 0)
4167 read_lock(&tasklist_lock
);
4168 p
= find_process_by_pid(pid
);
4173 retval
= security_task_getscheduler(p
);
4177 lp
.sched_priority
= p
->rt_priority
;
4178 read_unlock(&tasklist_lock
);
4181 * This one might sleep, we cannot do it with a spinlock held ...
4183 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4189 read_unlock(&tasklist_lock
);
4193 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4197 cpumask_t cpus_allowed
;
4200 read_lock(&tasklist_lock
);
4202 p
= find_process_by_pid(pid
);
4204 read_unlock(&tasklist_lock
);
4205 unlock_cpu_hotplug();
4210 * It is not safe to call set_cpus_allowed with the
4211 * tasklist_lock held. We will bump the task_struct's
4212 * usage count and then drop tasklist_lock.
4215 read_unlock(&tasklist_lock
);
4218 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4219 !capable(CAP_SYS_NICE
))
4222 retval
= security_task_setscheduler(p
, 0, NULL
);
4226 cpus_allowed
= cpuset_cpus_allowed(p
);
4227 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4228 retval
= set_cpus_allowed(p
, new_mask
);
4232 unlock_cpu_hotplug();
4236 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4237 cpumask_t
*new_mask
)
4239 if (len
< sizeof(cpumask_t
)) {
4240 memset(new_mask
, 0, sizeof(cpumask_t
));
4241 } else if (len
> sizeof(cpumask_t
)) {
4242 len
= sizeof(cpumask_t
);
4244 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4248 * sys_sched_setaffinity - set the cpu affinity of a process
4249 * @pid: pid of the process
4250 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4251 * @user_mask_ptr: user-space pointer to the new cpu mask
4253 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4254 unsigned long __user
*user_mask_ptr
)
4259 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4263 return sched_setaffinity(pid
, new_mask
);
4267 * Represents all cpu's present in the system
4268 * In systems capable of hotplug, this map could dynamically grow
4269 * as new cpu's are detected in the system via any platform specific
4270 * method, such as ACPI for e.g.
4273 cpumask_t cpu_present_map __read_mostly
;
4274 EXPORT_SYMBOL(cpu_present_map
);
4277 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4278 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4281 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4287 read_lock(&tasklist_lock
);
4290 p
= find_process_by_pid(pid
);
4294 retval
= security_task_getscheduler(p
);
4298 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4301 read_unlock(&tasklist_lock
);
4302 unlock_cpu_hotplug();
4310 * sys_sched_getaffinity - get the cpu affinity of a process
4311 * @pid: pid of the process
4312 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4313 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4315 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4316 unsigned long __user
*user_mask_ptr
)
4321 if (len
< sizeof(cpumask_t
))
4324 ret
= sched_getaffinity(pid
, &mask
);
4328 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4331 return sizeof(cpumask_t
);
4335 * sys_sched_yield - yield the current processor to other threads.
4337 * this function yields the current CPU by moving the calling thread
4338 * to the expired array. If there are no other threads running on this
4339 * CPU then this function will return.
4341 asmlinkage
long sys_sched_yield(void)
4343 runqueue_t
*rq
= this_rq_lock();
4344 prio_array_t
*array
= current
->array
;
4345 prio_array_t
*target
= rq
->expired
;
4347 schedstat_inc(rq
, yld_cnt
);
4349 * We implement yielding by moving the task into the expired
4352 * (special rule: RT tasks will just roundrobin in the active
4355 if (rt_task(current
))
4356 target
= rq
->active
;
4358 if (array
->nr_active
== 1) {
4359 schedstat_inc(rq
, yld_act_empty
);
4360 if (!rq
->expired
->nr_active
)
4361 schedstat_inc(rq
, yld_both_empty
);
4362 } else if (!rq
->expired
->nr_active
)
4363 schedstat_inc(rq
, yld_exp_empty
);
4365 if (array
!= target
) {
4366 dequeue_task(current
, array
);
4367 enqueue_task(current
, target
);
4370 * requeue_task is cheaper so perform that if possible.
4372 requeue_task(current
, array
);
4375 * Since we are going to call schedule() anyway, there's
4376 * no need to preempt or enable interrupts:
4378 __release(rq
->lock
);
4379 _raw_spin_unlock(&rq
->lock
);
4380 preempt_enable_no_resched();
4387 static inline void __cond_resched(void)
4389 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4390 __might_sleep(__FILE__
, __LINE__
);
4393 * The BKS might be reacquired before we have dropped
4394 * PREEMPT_ACTIVE, which could trigger a second
4395 * cond_resched() call.
4397 if (unlikely(preempt_count()))
4399 if (unlikely(system_state
!= SYSTEM_RUNNING
))
4402 add_preempt_count(PREEMPT_ACTIVE
);
4404 sub_preempt_count(PREEMPT_ACTIVE
);
4405 } while (need_resched());
4408 int __sched
cond_resched(void)
4410 if (need_resched()) {
4417 EXPORT_SYMBOL(cond_resched
);
4420 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4421 * call schedule, and on return reacquire the lock.
4423 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4424 * operations here to prevent schedule() from being called twice (once via
4425 * spin_unlock(), once by hand).
4427 int cond_resched_lock(spinlock_t
*lock
)
4431 if (need_lockbreak(lock
)) {
4437 if (need_resched()) {
4438 _raw_spin_unlock(lock
);
4439 preempt_enable_no_resched();
4447 EXPORT_SYMBOL(cond_resched_lock
);
4449 int __sched
cond_resched_softirq(void)
4451 BUG_ON(!in_softirq());
4453 if (need_resched()) {
4454 __local_bh_enable();
4462 EXPORT_SYMBOL(cond_resched_softirq
);
4466 * yield - yield the current processor to other threads.
4468 * this is a shortcut for kernel-space yielding - it marks the
4469 * thread runnable and calls sys_sched_yield().
4471 void __sched
yield(void)
4473 set_current_state(TASK_RUNNING
);
4477 EXPORT_SYMBOL(yield
);
4480 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4481 * that process accounting knows that this is a task in IO wait state.
4483 * But don't do that if it is a deliberate, throttling IO wait (this task
4484 * has set its backing_dev_info: the queue against which it should throttle)
4486 void __sched
io_schedule(void)
4488 struct runqueue
*rq
= &__raw_get_cpu_var(runqueues
);
4490 atomic_inc(&rq
->nr_iowait
);
4492 atomic_dec(&rq
->nr_iowait
);
4495 EXPORT_SYMBOL(io_schedule
);
4497 long __sched
io_schedule_timeout(long timeout
)
4499 struct runqueue
*rq
= &__raw_get_cpu_var(runqueues
);
4502 atomic_inc(&rq
->nr_iowait
);
4503 ret
= schedule_timeout(timeout
);
4504 atomic_dec(&rq
->nr_iowait
);
4509 * sys_sched_get_priority_max - return maximum RT priority.
4510 * @policy: scheduling class.
4512 * this syscall returns the maximum rt_priority that can be used
4513 * by a given scheduling class.
4515 asmlinkage
long sys_sched_get_priority_max(int policy
)
4522 ret
= MAX_USER_RT_PRIO
-1;
4533 * sys_sched_get_priority_min - return minimum RT priority.
4534 * @policy: scheduling class.
4536 * this syscall returns the minimum rt_priority that can be used
4537 * by a given scheduling class.
4539 asmlinkage
long sys_sched_get_priority_min(int policy
)
4556 * sys_sched_rr_get_interval - return the default timeslice of a process.
4557 * @pid: pid of the process.
4558 * @interval: userspace pointer to the timeslice value.
4560 * this syscall writes the default timeslice value of a given process
4561 * into the user-space timespec buffer. A value of '0' means infinity.
4564 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4566 int retval
= -EINVAL
;
4574 read_lock(&tasklist_lock
);
4575 p
= find_process_by_pid(pid
);
4579 retval
= security_task_getscheduler(p
);
4583 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4584 0 : task_timeslice(p
), &t
);
4585 read_unlock(&tasklist_lock
);
4586 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4590 read_unlock(&tasklist_lock
);
4594 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4596 if (list_empty(&p
->children
)) return NULL
;
4597 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4600 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4602 if (p
->sibling
.prev
==&p
->parent
->children
) return NULL
;
4603 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4606 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4608 if (p
->sibling
.next
==&p
->parent
->children
) return NULL
;
4609 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4612 static void show_task(task_t
*p
)
4616 unsigned long free
= 0;
4617 static const char *stat_nam
[] = { "R", "S", "D", "T", "t", "Z", "X" };
4619 printk("%-13.13s ", p
->comm
);
4620 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4621 if (state
< ARRAY_SIZE(stat_nam
))
4622 printk(stat_nam
[state
]);
4625 #if (BITS_PER_LONG == 32)
4626 if (state
== TASK_RUNNING
)
4627 printk(" running ");
4629 printk(" %08lX ", thread_saved_pc(p
));
4631 if (state
== TASK_RUNNING
)
4632 printk(" running task ");
4634 printk(" %016lx ", thread_saved_pc(p
));
4636 #ifdef CONFIG_DEBUG_STACK_USAGE
4638 unsigned long *n
= end_of_stack(p
);
4641 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4644 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4645 if ((relative
= eldest_child(p
)))
4646 printk("%5d ", relative
->pid
);
4649 if ((relative
= younger_sibling(p
)))
4650 printk("%7d", relative
->pid
);
4653 if ((relative
= older_sibling(p
)))
4654 printk(" %5d", relative
->pid
);
4658 printk(" (L-TLB)\n");
4660 printk(" (NOTLB)\n");
4662 if (state
!= TASK_RUNNING
)
4663 show_stack(p
, NULL
);
4666 void show_state(void)
4670 #if (BITS_PER_LONG == 32)
4673 printk(" task PC pid father child younger older\n");
4677 printk(" task PC pid father child younger older\n");
4679 read_lock(&tasklist_lock
);
4680 do_each_thread(g
, p
) {
4682 * reset the NMI-timeout, listing all files on a slow
4683 * console might take alot of time:
4685 touch_nmi_watchdog();
4687 } while_each_thread(g
, p
);
4689 read_unlock(&tasklist_lock
);
4690 mutex_debug_show_all_locks();
4694 * init_idle - set up an idle thread for a given CPU
4695 * @idle: task in question
4696 * @cpu: cpu the idle task belongs to
4698 * NOTE: this function does not set the idle thread's NEED_RESCHED
4699 * flag, to make booting more robust.
4701 void __devinit
init_idle(task_t
*idle
, int cpu
)
4703 runqueue_t
*rq
= cpu_rq(cpu
);
4704 unsigned long flags
;
4706 idle
->timestamp
= sched_clock();
4707 idle
->sleep_avg
= 0;
4709 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4710 idle
->state
= TASK_RUNNING
;
4711 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4712 set_task_cpu(idle
, cpu
);
4714 spin_lock_irqsave(&rq
->lock
, flags
);
4715 rq
->curr
= rq
->idle
= idle
;
4716 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4719 spin_unlock_irqrestore(&rq
->lock
, flags
);
4721 /* Set the preempt count _outside_ the spinlocks! */
4722 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4723 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4725 task_thread_info(idle
)->preempt_count
= 0;
4730 * In a system that switches off the HZ timer nohz_cpu_mask
4731 * indicates which cpus entered this state. This is used
4732 * in the rcu update to wait only for active cpus. For system
4733 * which do not switch off the HZ timer nohz_cpu_mask should
4734 * always be CPU_MASK_NONE.
4736 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4740 * This is how migration works:
4742 * 1) we queue a migration_req_t structure in the source CPU's
4743 * runqueue and wake up that CPU's migration thread.
4744 * 2) we down() the locked semaphore => thread blocks.
4745 * 3) migration thread wakes up (implicitly it forces the migrated
4746 * thread off the CPU)
4747 * 4) it gets the migration request and checks whether the migrated
4748 * task is still in the wrong runqueue.
4749 * 5) if it's in the wrong runqueue then the migration thread removes
4750 * it and puts it into the right queue.
4751 * 6) migration thread up()s the semaphore.
4752 * 7) we wake up and the migration is done.
4756 * Change a given task's CPU affinity. Migrate the thread to a
4757 * proper CPU and schedule it away if the CPU it's executing on
4758 * is removed from the allowed bitmask.
4760 * NOTE: the caller must have a valid reference to the task, the
4761 * task must not exit() & deallocate itself prematurely. The
4762 * call is not atomic; no spinlocks may be held.
4764 int set_cpus_allowed(task_t
*p
, cpumask_t new_mask
)
4766 unsigned long flags
;
4768 migration_req_t req
;
4771 rq
= task_rq_lock(p
, &flags
);
4772 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4777 p
->cpus_allowed
= new_mask
;
4778 /* Can the task run on the task's current CPU? If so, we're done */
4779 if (cpu_isset(task_cpu(p
), new_mask
))
4782 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4783 /* Need help from migration thread: drop lock and wait. */
4784 task_rq_unlock(rq
, &flags
);
4785 wake_up_process(rq
->migration_thread
);
4786 wait_for_completion(&req
.done
);
4787 tlb_migrate_finish(p
->mm
);
4791 task_rq_unlock(rq
, &flags
);
4795 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4798 * Move (not current) task off this cpu, onto dest cpu. We're doing
4799 * this because either it can't run here any more (set_cpus_allowed()
4800 * away from this CPU, or CPU going down), or because we're
4801 * attempting to rebalance this task on exec (sched_exec).
4803 * So we race with normal scheduler movements, but that's OK, as long
4804 * as the task is no longer on this CPU.
4806 * Returns non-zero if task was successfully migrated.
4808 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4810 runqueue_t
*rq_dest
, *rq_src
;
4813 if (unlikely(cpu_is_offline(dest_cpu
)))
4816 rq_src
= cpu_rq(src_cpu
);
4817 rq_dest
= cpu_rq(dest_cpu
);
4819 double_rq_lock(rq_src
, rq_dest
);
4820 /* Already moved. */
4821 if (task_cpu(p
) != src_cpu
)
4823 /* Affinity changed (again). */
4824 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4827 set_task_cpu(p
, dest_cpu
);
4830 * Sync timestamp with rq_dest's before activating.
4831 * The same thing could be achieved by doing this step
4832 * afterwards, and pretending it was a local activate.
4833 * This way is cleaner and logically correct.
4835 p
->timestamp
= p
->timestamp
- rq_src
->timestamp_last_tick
4836 + rq_dest
->timestamp_last_tick
;
4837 deactivate_task(p
, rq_src
);
4838 activate_task(p
, rq_dest
, 0);
4839 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
4840 resched_task(rq_dest
->curr
);
4844 double_rq_unlock(rq_src
, rq_dest
);
4849 * migration_thread - this is a highprio system thread that performs
4850 * thread migration by bumping thread off CPU then 'pushing' onto
4853 static int migration_thread(void *data
)
4856 int cpu
= (long)data
;
4859 BUG_ON(rq
->migration_thread
!= current
);
4861 set_current_state(TASK_INTERRUPTIBLE
);
4862 while (!kthread_should_stop()) {
4863 struct list_head
*head
;
4864 migration_req_t
*req
;
4868 spin_lock_irq(&rq
->lock
);
4870 if (cpu_is_offline(cpu
)) {
4871 spin_unlock_irq(&rq
->lock
);
4875 if (rq
->active_balance
) {
4876 active_load_balance(rq
, cpu
);
4877 rq
->active_balance
= 0;
4880 head
= &rq
->migration_queue
;
4882 if (list_empty(head
)) {
4883 spin_unlock_irq(&rq
->lock
);
4885 set_current_state(TASK_INTERRUPTIBLE
);
4888 req
= list_entry(head
->next
, migration_req_t
, list
);
4889 list_del_init(head
->next
);
4891 spin_unlock(&rq
->lock
);
4892 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4895 complete(&req
->done
);
4897 __set_current_state(TASK_RUNNING
);
4901 /* Wait for kthread_stop */
4902 set_current_state(TASK_INTERRUPTIBLE
);
4903 while (!kthread_should_stop()) {
4905 set_current_state(TASK_INTERRUPTIBLE
);
4907 __set_current_state(TASK_RUNNING
);
4911 #ifdef CONFIG_HOTPLUG_CPU
4912 /* Figure out where task on dead CPU should go, use force if neccessary. */
4913 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*tsk
)
4916 unsigned long flags
;
4922 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
4923 cpus_and(mask
, mask
, tsk
->cpus_allowed
);
4924 dest_cpu
= any_online_cpu(mask
);
4926 /* On any allowed CPU? */
4927 if (dest_cpu
== NR_CPUS
)
4928 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4930 /* No more Mr. Nice Guy. */
4931 if (dest_cpu
== NR_CPUS
) {
4932 rq
= task_rq_lock(tsk
, &flags
);
4933 cpus_setall(tsk
->cpus_allowed
);
4934 dest_cpu
= any_online_cpu(tsk
->cpus_allowed
);
4935 task_rq_unlock(rq
, &flags
);
4938 * Don't tell them about moving exiting tasks or
4939 * kernel threads (both mm NULL), since they never
4942 if (tsk
->mm
&& printk_ratelimit())
4943 printk(KERN_INFO
"process %d (%s) no "
4944 "longer affine to cpu%d\n",
4945 tsk
->pid
, tsk
->comm
, dead_cpu
);
4947 if (!__migrate_task(tsk
, dead_cpu
, dest_cpu
))
4952 * While a dead CPU has no uninterruptible tasks queued at this point,
4953 * it might still have a nonzero ->nr_uninterruptible counter, because
4954 * for performance reasons the counter is not stricly tracking tasks to
4955 * their home CPUs. So we just add the counter to another CPU's counter,
4956 * to keep the global sum constant after CPU-down:
4958 static void migrate_nr_uninterruptible(runqueue_t
*rq_src
)
4960 runqueue_t
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
4961 unsigned long flags
;
4963 local_irq_save(flags
);
4964 double_rq_lock(rq_src
, rq_dest
);
4965 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
4966 rq_src
->nr_uninterruptible
= 0;
4967 double_rq_unlock(rq_src
, rq_dest
);
4968 local_irq_restore(flags
);
4971 /* Run through task list and migrate tasks from the dead cpu. */
4972 static void migrate_live_tasks(int src_cpu
)
4974 struct task_struct
*tsk
, *t
;
4976 write_lock_irq(&tasklist_lock
);
4978 do_each_thread(t
, tsk
) {
4982 if (task_cpu(tsk
) == src_cpu
)
4983 move_task_off_dead_cpu(src_cpu
, tsk
);
4984 } while_each_thread(t
, tsk
);
4986 write_unlock_irq(&tasklist_lock
);
4989 /* Schedules idle task to be the next runnable task on current CPU.
4990 * It does so by boosting its priority to highest possible and adding it to
4991 * the _front_ of runqueue. Used by CPU offline code.
4993 void sched_idle_next(void)
4995 int cpu
= smp_processor_id();
4996 runqueue_t
*rq
= this_rq();
4997 struct task_struct
*p
= rq
->idle
;
4998 unsigned long flags
;
5000 /* cpu has to be offline */
5001 BUG_ON(cpu_online(cpu
));
5003 /* Strictly not necessary since rest of the CPUs are stopped by now
5004 * and interrupts disabled on current cpu.
5006 spin_lock_irqsave(&rq
->lock
, flags
);
5008 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5009 /* Add idle task to _front_ of it's priority queue */
5010 __activate_idle_task(p
, rq
);
5012 spin_unlock_irqrestore(&rq
->lock
, flags
);
5015 /* Ensures that the idle task is using init_mm right before its cpu goes
5018 void idle_task_exit(void)
5020 struct mm_struct
*mm
= current
->active_mm
;
5022 BUG_ON(cpu_online(smp_processor_id()));
5025 switch_mm(mm
, &init_mm
, current
);
5029 static void migrate_dead(unsigned int dead_cpu
, task_t
*tsk
)
5031 struct runqueue
*rq
= cpu_rq(dead_cpu
);
5033 /* Must be exiting, otherwise would be on tasklist. */
5034 BUG_ON(tsk
->exit_state
!= EXIT_ZOMBIE
&& tsk
->exit_state
!= EXIT_DEAD
);
5036 /* Cannot have done final schedule yet: would have vanished. */
5037 BUG_ON(tsk
->flags
& PF_DEAD
);
5039 get_task_struct(tsk
);
5042 * Drop lock around migration; if someone else moves it,
5043 * that's OK. No task can be added to this CPU, so iteration is
5046 spin_unlock_irq(&rq
->lock
);
5047 move_task_off_dead_cpu(dead_cpu
, tsk
);
5048 spin_lock_irq(&rq
->lock
);
5050 put_task_struct(tsk
);
5053 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5054 static void migrate_dead_tasks(unsigned int dead_cpu
)
5057 struct runqueue
*rq
= cpu_rq(dead_cpu
);
5059 for (arr
= 0; arr
< 2; arr
++) {
5060 for (i
= 0; i
< MAX_PRIO
; i
++) {
5061 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5062 while (!list_empty(list
))
5063 migrate_dead(dead_cpu
,
5064 list_entry(list
->next
, task_t
,
5069 #endif /* CONFIG_HOTPLUG_CPU */
5072 * migration_call - callback that gets triggered when a CPU is added.
5073 * Here we can start up the necessary migration thread for the new CPU.
5075 static int __cpuinit
migration_call(struct notifier_block
*nfb
,
5076 unsigned long action
,
5079 int cpu
= (long)hcpu
;
5080 struct task_struct
*p
;
5081 struct runqueue
*rq
;
5082 unsigned long flags
;
5085 case CPU_UP_PREPARE
:
5086 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5089 p
->flags
|= PF_NOFREEZE
;
5090 kthread_bind(p
, cpu
);
5091 /* Must be high prio: stop_machine expects to yield to it. */
5092 rq
= task_rq_lock(p
, &flags
);
5093 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5094 task_rq_unlock(rq
, &flags
);
5095 cpu_rq(cpu
)->migration_thread
= p
;
5098 /* Strictly unneccessary, as first user will wake it. */
5099 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5101 #ifdef CONFIG_HOTPLUG_CPU
5102 case CPU_UP_CANCELED
:
5103 if (!cpu_rq(cpu
)->migration_thread
)
5105 /* Unbind it from offline cpu so it can run. Fall thru. */
5106 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5107 any_online_cpu(cpu_online_map
));
5108 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5109 cpu_rq(cpu
)->migration_thread
= NULL
;
5112 migrate_live_tasks(cpu
);
5114 kthread_stop(rq
->migration_thread
);
5115 rq
->migration_thread
= NULL
;
5116 /* Idle task back to normal (off runqueue, low prio) */
5117 rq
= task_rq_lock(rq
->idle
, &flags
);
5118 deactivate_task(rq
->idle
, rq
);
5119 rq
->idle
->static_prio
= MAX_PRIO
;
5120 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5121 migrate_dead_tasks(cpu
);
5122 task_rq_unlock(rq
, &flags
);
5123 migrate_nr_uninterruptible(rq
);
5124 BUG_ON(rq
->nr_running
!= 0);
5126 /* No need to migrate the tasks: it was best-effort if
5127 * they didn't do lock_cpu_hotplug(). Just wake up
5128 * the requestors. */
5129 spin_lock_irq(&rq
->lock
);
5130 while (!list_empty(&rq
->migration_queue
)) {
5131 migration_req_t
*req
;
5132 req
= list_entry(rq
->migration_queue
.next
,
5133 migration_req_t
, list
);
5134 list_del_init(&req
->list
);
5135 complete(&req
->done
);
5137 spin_unlock_irq(&rq
->lock
);
5144 /* Register at highest priority so that task migration (migrate_all_tasks)
5145 * happens before everything else.
5147 static struct notifier_block __cpuinitdata migration_notifier
= {
5148 .notifier_call
= migration_call
,
5152 int __init
migration_init(void)
5154 void *cpu
= (void *)(long)smp_processor_id();
5155 /* Start one for boot CPU. */
5156 migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5157 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5158 register_cpu_notifier(&migration_notifier
);
5164 #undef SCHED_DOMAIN_DEBUG
5165 #ifdef SCHED_DOMAIN_DEBUG
5166 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5171 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5175 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5180 struct sched_group
*group
= sd
->groups
;
5181 cpumask_t groupmask
;
5183 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5184 cpus_clear(groupmask
);
5187 for (i
= 0; i
< level
+ 1; i
++)
5189 printk("domain %d: ", level
);
5191 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5192 printk("does not load-balance\n");
5194 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
5198 printk("span %s\n", str
);
5200 if (!cpu_isset(cpu
, sd
->span
))
5201 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
5202 if (!cpu_isset(cpu
, group
->cpumask
))
5203 printk(KERN_ERR
"ERROR: domain->groups does not contain CPU%d\n", cpu
);
5206 for (i
= 0; i
< level
+ 2; i
++)
5212 printk(KERN_ERR
"ERROR: group is NULL\n");
5216 if (!group
->cpu_power
) {
5218 printk(KERN_ERR
"ERROR: domain->cpu_power not set\n");
5221 if (!cpus_weight(group
->cpumask
)) {
5223 printk(KERN_ERR
"ERROR: empty group\n");
5226 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5228 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5231 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5233 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5236 group
= group
->next
;
5237 } while (group
!= sd
->groups
);
5240 if (!cpus_equal(sd
->span
, groupmask
))
5241 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5247 if (!cpus_subset(groupmask
, sd
->span
))
5248 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
5254 #define sched_domain_debug(sd, cpu) {}
5257 static int sd_degenerate(struct sched_domain
*sd
)
5259 if (cpus_weight(sd
->span
) == 1)
5262 /* Following flags need at least 2 groups */
5263 if (sd
->flags
& (SD_LOAD_BALANCE
|
5264 SD_BALANCE_NEWIDLE
|
5267 if (sd
->groups
!= sd
->groups
->next
)
5271 /* Following flags don't use groups */
5272 if (sd
->flags
& (SD_WAKE_IDLE
|
5280 static int sd_parent_degenerate(struct sched_domain
*sd
,
5281 struct sched_domain
*parent
)
5283 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5285 if (sd_degenerate(parent
))
5288 if (!cpus_equal(sd
->span
, parent
->span
))
5291 /* Does parent contain flags not in child? */
5292 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5293 if (cflags
& SD_WAKE_AFFINE
)
5294 pflags
&= ~SD_WAKE_BALANCE
;
5295 /* Flags needing groups don't count if only 1 group in parent */
5296 if (parent
->groups
== parent
->groups
->next
) {
5297 pflags
&= ~(SD_LOAD_BALANCE
|
5298 SD_BALANCE_NEWIDLE
|
5302 if (~cflags
& pflags
)
5309 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5310 * hold the hotplug lock.
5312 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5314 runqueue_t
*rq
= cpu_rq(cpu
);
5315 struct sched_domain
*tmp
;
5317 /* Remove the sched domains which do not contribute to scheduling. */
5318 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5319 struct sched_domain
*parent
= tmp
->parent
;
5322 if (sd_parent_degenerate(tmp
, parent
))
5323 tmp
->parent
= parent
->parent
;
5326 if (sd
&& sd_degenerate(sd
))
5329 sched_domain_debug(sd
, cpu
);
5331 rcu_assign_pointer(rq
->sd
, sd
);
5334 /* cpus with isolated domains */
5335 static cpumask_t __devinitdata cpu_isolated_map
= CPU_MASK_NONE
;
5337 /* Setup the mask of cpus configured for isolated domains */
5338 static int __init
isolated_cpu_setup(char *str
)
5340 int ints
[NR_CPUS
], i
;
5342 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5343 cpus_clear(cpu_isolated_map
);
5344 for (i
= 1; i
<= ints
[0]; i
++)
5345 if (ints
[i
] < NR_CPUS
)
5346 cpu_set(ints
[i
], cpu_isolated_map
);
5350 __setup ("isolcpus=", isolated_cpu_setup
);
5353 * init_sched_build_groups takes an array of groups, the cpumask we wish
5354 * to span, and a pointer to a function which identifies what group a CPU
5355 * belongs to. The return value of group_fn must be a valid index into the
5356 * groups[] array, and must be >= 0 and < NR_CPUS (due to the fact that we
5357 * keep track of groups covered with a cpumask_t).
5359 * init_sched_build_groups will build a circular linked list of the groups
5360 * covered by the given span, and will set each group's ->cpumask correctly,
5361 * and ->cpu_power to 0.
5363 static void init_sched_build_groups(struct sched_group groups
[], cpumask_t span
,
5364 int (*group_fn
)(int cpu
))
5366 struct sched_group
*first
= NULL
, *last
= NULL
;
5367 cpumask_t covered
= CPU_MASK_NONE
;
5370 for_each_cpu_mask(i
, span
) {
5371 int group
= group_fn(i
);
5372 struct sched_group
*sg
= &groups
[group
];
5375 if (cpu_isset(i
, covered
))
5378 sg
->cpumask
= CPU_MASK_NONE
;
5381 for_each_cpu_mask(j
, span
) {
5382 if (group_fn(j
) != group
)
5385 cpu_set(j
, covered
);
5386 cpu_set(j
, sg
->cpumask
);
5397 #define SD_NODES_PER_DOMAIN 16
5400 * Self-tuning task migration cost measurement between source and target CPUs.
5402 * This is done by measuring the cost of manipulating buffers of varying
5403 * sizes. For a given buffer-size here are the steps that are taken:
5405 * 1) the source CPU reads+dirties a shared buffer
5406 * 2) the target CPU reads+dirties the same shared buffer
5408 * We measure how long they take, in the following 4 scenarios:
5410 * - source: CPU1, target: CPU2 | cost1
5411 * - source: CPU2, target: CPU1 | cost2
5412 * - source: CPU1, target: CPU1 | cost3
5413 * - source: CPU2, target: CPU2 | cost4
5415 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5416 * the cost of migration.
5418 * We then start off from a small buffer-size and iterate up to larger
5419 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5420 * doing a maximum search for the cost. (The maximum cost for a migration
5421 * normally occurs when the working set size is around the effective cache
5424 #define SEARCH_SCOPE 2
5425 #define MIN_CACHE_SIZE (64*1024U)
5426 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5427 #define ITERATIONS 1
5428 #define SIZE_THRESH 130
5429 #define COST_THRESH 130
5432 * The migration cost is a function of 'domain distance'. Domain
5433 * distance is the number of steps a CPU has to iterate down its
5434 * domain tree to share a domain with the other CPU. The farther
5435 * two CPUs are from each other, the larger the distance gets.
5437 * Note that we use the distance only to cache measurement results,
5438 * the distance value is not used numerically otherwise. When two
5439 * CPUs have the same distance it is assumed that the migration
5440 * cost is the same. (this is a simplification but quite practical)
5442 #define MAX_DOMAIN_DISTANCE 32
5444 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5445 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5447 * Architectures may override the migration cost and thus avoid
5448 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5449 * virtualized hardware:
5451 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5452 CONFIG_DEFAULT_MIGRATION_COST
5459 * Allow override of migration cost - in units of microseconds.
5460 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5461 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5463 static int __init
migration_cost_setup(char *str
)
5465 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5467 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5469 printk("#ints: %d\n", ints
[0]);
5470 for (i
= 1; i
<= ints
[0]; i
++) {
5471 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5472 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5477 __setup ("migration_cost=", migration_cost_setup
);
5480 * Global multiplier (divisor) for migration-cutoff values,
5481 * in percentiles. E.g. use a value of 150 to get 1.5 times
5482 * longer cache-hot cutoff times.
5484 * (We scale it from 100 to 128 to long long handling easier.)
5487 #define MIGRATION_FACTOR_SCALE 128
5489 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5491 static int __init
setup_migration_factor(char *str
)
5493 get_option(&str
, &migration_factor
);
5494 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5498 __setup("migration_factor=", setup_migration_factor
);
5501 * Estimated distance of two CPUs, measured via the number of domains
5502 * we have to pass for the two CPUs to be in the same span:
5504 static unsigned long domain_distance(int cpu1
, int cpu2
)
5506 unsigned long distance
= 0;
5507 struct sched_domain
*sd
;
5509 for_each_domain(cpu1
, sd
) {
5510 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5511 if (cpu_isset(cpu2
, sd
->span
))
5515 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5517 distance
= MAX_DOMAIN_DISTANCE
-1;
5523 static unsigned int migration_debug
;
5525 static int __init
setup_migration_debug(char *str
)
5527 get_option(&str
, &migration_debug
);
5531 __setup("migration_debug=", setup_migration_debug
);
5534 * Maximum cache-size that the scheduler should try to measure.
5535 * Architectures with larger caches should tune this up during
5536 * bootup. Gets used in the domain-setup code (i.e. during SMP
5539 unsigned int max_cache_size
;
5541 static int __init
setup_max_cache_size(char *str
)
5543 get_option(&str
, &max_cache_size
);
5547 __setup("max_cache_size=", setup_max_cache_size
);
5550 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5551 * is the operation that is timed, so we try to generate unpredictable
5552 * cachemisses that still end up filling the L2 cache:
5554 static void touch_cache(void *__cache
, unsigned long __size
)
5556 unsigned long size
= __size
/sizeof(long), chunk1
= size
/3,
5558 unsigned long *cache
= __cache
;
5561 for (i
= 0; i
< size
/6; i
+= 8) {
5564 case 1: cache
[size
-1-i
]++;
5565 case 2: cache
[chunk1
-i
]++;
5566 case 3: cache
[chunk1
+i
]++;
5567 case 4: cache
[chunk2
-i
]++;
5568 case 5: cache
[chunk2
+i
]++;
5574 * Measure the cache-cost of one task migration. Returns in units of nsec.
5576 static unsigned long long measure_one(void *cache
, unsigned long size
,
5577 int source
, int target
)
5579 cpumask_t mask
, saved_mask
;
5580 unsigned long long t0
, t1
, t2
, t3
, cost
;
5582 saved_mask
= current
->cpus_allowed
;
5585 * Flush source caches to RAM and invalidate them:
5590 * Migrate to the source CPU:
5592 mask
= cpumask_of_cpu(source
);
5593 set_cpus_allowed(current
, mask
);
5594 WARN_ON(smp_processor_id() != source
);
5597 * Dirty the working set:
5600 touch_cache(cache
, size
);
5604 * Migrate to the target CPU, dirty the L2 cache and access
5605 * the shared buffer. (which represents the working set
5606 * of a migrated task.)
5608 mask
= cpumask_of_cpu(target
);
5609 set_cpus_allowed(current
, mask
);
5610 WARN_ON(smp_processor_id() != target
);
5613 touch_cache(cache
, size
);
5616 cost
= t1
-t0
+ t3
-t2
;
5618 if (migration_debug
>= 2)
5619 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5620 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5622 * Flush target caches to RAM and invalidate them:
5626 set_cpus_allowed(current
, saved_mask
);
5632 * Measure a series of task migrations and return the average
5633 * result. Since this code runs early during bootup the system
5634 * is 'undisturbed' and the average latency makes sense.
5636 * The algorithm in essence auto-detects the relevant cache-size,
5637 * so it will properly detect different cachesizes for different
5638 * cache-hierarchies, depending on how the CPUs are connected.
5640 * Architectures can prime the upper limit of the search range via
5641 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5643 static unsigned long long
5644 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5646 unsigned long long cost1
, cost2
;
5650 * Measure the migration cost of 'size' bytes, over an
5651 * average of 10 runs:
5653 * (We perturb the cache size by a small (0..4k)
5654 * value to compensate size/alignment related artifacts.
5655 * We also subtract the cost of the operation done on
5661 * dry run, to make sure we start off cache-cold on cpu1,
5662 * and to get any vmalloc pagefaults in advance:
5664 measure_one(cache
, size
, cpu1
, cpu2
);
5665 for (i
= 0; i
< ITERATIONS
; i
++)
5666 cost1
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu2
);
5668 measure_one(cache
, size
, cpu2
, cpu1
);
5669 for (i
= 0; i
< ITERATIONS
; i
++)
5670 cost1
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu1
);
5673 * (We measure the non-migrating [cached] cost on both
5674 * cpu1 and cpu2, to handle CPUs with different speeds)
5678 measure_one(cache
, size
, cpu1
, cpu1
);
5679 for (i
= 0; i
< ITERATIONS
; i
++)
5680 cost2
+= measure_one(cache
, size
- i
*1024, cpu1
, cpu1
);
5682 measure_one(cache
, size
, cpu2
, cpu2
);
5683 for (i
= 0; i
< ITERATIONS
; i
++)
5684 cost2
+= measure_one(cache
, size
- i
*1024, cpu2
, cpu2
);
5687 * Get the per-iteration migration cost:
5689 do_div(cost1
, 2*ITERATIONS
);
5690 do_div(cost2
, 2*ITERATIONS
);
5692 return cost1
- cost2
;
5695 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5697 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5698 unsigned int max_size
, size
, size_found
= 0;
5699 long long cost
= 0, prev_cost
;
5703 * Search from max_cache_size*5 down to 64K - the real relevant
5704 * cachesize has to lie somewhere inbetween.
5706 if (max_cache_size
) {
5707 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5708 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5711 * Since we have no estimation about the relevant
5714 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5715 size
= MIN_CACHE_SIZE
;
5718 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5719 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
5724 * Allocate the working set:
5726 cache
= vmalloc(max_size
);
5728 printk("could not vmalloc %d bytes for cache!\n", 2*max_size
);
5729 return 1000000; // return 1 msec on very small boxen
5732 while (size
<= max_size
) {
5734 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
5740 if (max_cost
< cost
) {
5746 * Calculate average fluctuation, we use this to prevent
5747 * noise from triggering an early break out of the loop:
5749 fluct
= abs(cost
- prev_cost
);
5750 avg_fluct
= (avg_fluct
+ fluct
)/2;
5752 if (migration_debug
)
5753 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5755 (long)cost
/ 1000000,
5756 ((long)cost
/ 100000) % 10,
5757 (long)max_cost
/ 1000000,
5758 ((long)max_cost
/ 100000) % 10,
5759 domain_distance(cpu1
, cpu2
),
5763 * If we iterated at least 20% past the previous maximum,
5764 * and the cost has dropped by more than 20% already,
5765 * (taking fluctuations into account) then we assume to
5766 * have found the maximum and break out of the loop early:
5768 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
5769 if (cost
+avg_fluct
<= 0 ||
5770 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
5772 if (migration_debug
)
5773 printk("-> found max.\n");
5777 * Increase the cachesize in 10% steps:
5779 size
= size
* 10 / 9;
5782 if (migration_debug
)
5783 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5784 cpu1
, cpu2
, size_found
, max_cost
);
5789 * A task is considered 'cache cold' if at least 2 times
5790 * the worst-case cost of migration has passed.
5792 * (this limit is only listened to if the load-balancing
5793 * situation is 'nice' - if there is a large imbalance we
5794 * ignore it for the sake of CPU utilization and
5795 * processing fairness.)
5797 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
5800 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
5802 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
5803 unsigned long j0
, j1
, distance
, max_distance
= 0;
5804 struct sched_domain
*sd
;
5809 * First pass - calculate the cacheflush times:
5811 for_each_cpu_mask(cpu1
, *cpu_map
) {
5812 for_each_cpu_mask(cpu2
, *cpu_map
) {
5815 distance
= domain_distance(cpu1
, cpu2
);
5816 max_distance
= max(max_distance
, distance
);
5818 * No result cached yet?
5820 if (migration_cost
[distance
] == -1LL)
5821 migration_cost
[distance
] =
5822 measure_migration_cost(cpu1
, cpu2
);
5826 * Second pass - update the sched domain hierarchy with
5827 * the new cache-hot-time estimations:
5829 for_each_cpu_mask(cpu
, *cpu_map
) {
5831 for_each_domain(cpu
, sd
) {
5832 sd
->cache_hot_time
= migration_cost
[distance
];
5839 if (migration_debug
)
5840 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5848 if (system_state
== SYSTEM_BOOTING
) {
5849 printk("migration_cost=");
5850 for (distance
= 0; distance
<= max_distance
; distance
++) {
5853 printk("%ld", (long)migration_cost
[distance
] / 1000);
5858 if (migration_debug
)
5859 printk("migration: %ld seconds\n", (j1
-j0
)/HZ
);
5862 * Move back to the original CPU. NUMA-Q gets confused
5863 * if we migrate to another quad during bootup.
5865 if (raw_smp_processor_id() != orig_cpu
) {
5866 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
5867 saved_mask
= current
->cpus_allowed
;
5869 set_cpus_allowed(current
, mask
);
5870 set_cpus_allowed(current
, saved_mask
);
5877 * find_next_best_node - find the next node to include in a sched_domain
5878 * @node: node whose sched_domain we're building
5879 * @used_nodes: nodes already in the sched_domain
5881 * Find the next node to include in a given scheduling domain. Simply
5882 * finds the closest node not already in the @used_nodes map.
5884 * Should use nodemask_t.
5886 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5888 int i
, n
, val
, min_val
, best_node
= 0;
5892 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5893 /* Start at @node */
5894 n
= (node
+ i
) % MAX_NUMNODES
;
5896 if (!nr_cpus_node(n
))
5899 /* Skip already used nodes */
5900 if (test_bit(n
, used_nodes
))
5903 /* Simple min distance search */
5904 val
= node_distance(node
, n
);
5906 if (val
< min_val
) {
5912 set_bit(best_node
, used_nodes
);
5917 * sched_domain_node_span - get a cpumask for a node's sched_domain
5918 * @node: node whose cpumask we're constructing
5919 * @size: number of nodes to include in this span
5921 * Given a node, construct a good cpumask for its sched_domain to span. It
5922 * should be one that prevents unnecessary balancing, but also spreads tasks
5925 static cpumask_t
sched_domain_node_span(int node
)
5928 cpumask_t span
, nodemask
;
5929 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5932 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5934 nodemask
= node_to_cpumask(node
);
5935 cpus_or(span
, span
, nodemask
);
5936 set_bit(node
, used_nodes
);
5938 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5939 int next_node
= find_next_best_node(node
, used_nodes
);
5940 nodemask
= node_to_cpumask(next_node
);
5941 cpus_or(span
, span
, nodemask
);
5948 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5950 * At the moment, CONFIG_SCHED_SMT is never defined, but leave it in so we
5951 * can switch it on easily if needed.
5953 #ifdef CONFIG_SCHED_SMT
5954 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5955 static struct sched_group sched_group_cpus
[NR_CPUS
];
5956 static int cpu_to_cpu_group(int cpu
)
5962 #ifdef CONFIG_SCHED_MC
5963 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5964 static struct sched_group
*sched_group_core_bycpu
[NR_CPUS
];
5967 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5968 static int cpu_to_core_group(int cpu
)
5970 return first_cpu(cpu_sibling_map
[cpu
]);
5972 #elif defined(CONFIG_SCHED_MC)
5973 static int cpu_to_core_group(int cpu
)
5979 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5980 static struct sched_group
*sched_group_phys_bycpu
[NR_CPUS
];
5981 static int cpu_to_phys_group(int cpu
)
5983 #if defined(CONFIG_SCHED_MC)
5984 cpumask_t mask
= cpu_coregroup_map(cpu
);
5985 return first_cpu(mask
);
5986 #elif defined(CONFIG_SCHED_SMT)
5987 return first_cpu(cpu_sibling_map
[cpu
]);
5995 * The init_sched_build_groups can't handle what we want to do with node
5996 * groups, so roll our own. Now each node has its own list of groups which
5997 * gets dynamically allocated.
5999 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6000 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6002 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6003 static struct sched_group
*sched_group_allnodes_bycpu
[NR_CPUS
];
6005 static int cpu_to_allnodes_group(int cpu
)
6007 return cpu_to_node(cpu
);
6009 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6011 struct sched_group
*sg
= group_head
;
6017 for_each_cpu_mask(j
, sg
->cpumask
) {
6018 struct sched_domain
*sd
;
6020 sd
= &per_cpu(phys_domains
, j
);
6021 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6023 * Only add "power" once for each
6029 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6032 if (sg
!= group_head
)
6037 /* Free memory allocated for various sched_group structures */
6038 static void free_sched_groups(const cpumask_t
*cpu_map
)
6044 for_each_cpu_mask(cpu
, *cpu_map
) {
6045 struct sched_group
*sched_group_allnodes
6046 = sched_group_allnodes_bycpu
[cpu
];
6047 struct sched_group
**sched_group_nodes
6048 = sched_group_nodes_bycpu
[cpu
];
6050 if (sched_group_allnodes
) {
6051 kfree(sched_group_allnodes
);
6052 sched_group_allnodes_bycpu
[cpu
] = NULL
;
6055 if (!sched_group_nodes
)
6058 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6059 cpumask_t nodemask
= node_to_cpumask(i
);
6060 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6062 cpus_and(nodemask
, nodemask
, *cpu_map
);
6063 if (cpus_empty(nodemask
))
6073 if (oldsg
!= sched_group_nodes
[i
])
6076 kfree(sched_group_nodes
);
6077 sched_group_nodes_bycpu
[cpu
] = NULL
;
6080 for_each_cpu_mask(cpu
, *cpu_map
) {
6081 if (sched_group_phys_bycpu
[cpu
]) {
6082 kfree(sched_group_phys_bycpu
[cpu
]);
6083 sched_group_phys_bycpu
[cpu
] = NULL
;
6085 #ifdef CONFIG_SCHED_MC
6086 if (sched_group_core_bycpu
[cpu
]) {
6087 kfree(sched_group_core_bycpu
[cpu
]);
6088 sched_group_core_bycpu
[cpu
] = NULL
;
6095 * Build sched domains for a given set of cpus and attach the sched domains
6096 * to the individual cpus
6098 static int build_sched_domains(const cpumask_t
*cpu_map
)
6101 struct sched_group
*sched_group_phys
= NULL
;
6102 #ifdef CONFIG_SCHED_MC
6103 struct sched_group
*sched_group_core
= NULL
;
6106 struct sched_group
**sched_group_nodes
= NULL
;
6107 struct sched_group
*sched_group_allnodes
= NULL
;
6110 * Allocate the per-node list of sched groups
6112 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6114 if (!sched_group_nodes
) {
6115 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6118 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6122 * Set up domains for cpus specified by the cpu_map.
6124 for_each_cpu_mask(i
, *cpu_map
) {
6126 struct sched_domain
*sd
= NULL
, *p
;
6127 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6129 cpus_and(nodemask
, nodemask
, *cpu_map
);
6132 if (cpus_weight(*cpu_map
)
6133 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6134 if (!sched_group_allnodes
) {
6135 sched_group_allnodes
6136 = kmalloc(sizeof(struct sched_group
)
6139 if (!sched_group_allnodes
) {
6141 "Can not alloc allnodes sched group\n");
6144 sched_group_allnodes_bycpu
[i
]
6145 = sched_group_allnodes
;
6147 sd
= &per_cpu(allnodes_domains
, i
);
6148 *sd
= SD_ALLNODES_INIT
;
6149 sd
->span
= *cpu_map
;
6150 group
= cpu_to_allnodes_group(i
);
6151 sd
->groups
= &sched_group_allnodes
[group
];
6156 sd
= &per_cpu(node_domains
, i
);
6158 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6160 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6163 if (!sched_group_phys
) {
6165 = kmalloc(sizeof(struct sched_group
) * NR_CPUS
,
6167 if (!sched_group_phys
) {
6168 printk (KERN_WARNING
"Can not alloc phys sched"
6172 sched_group_phys_bycpu
[i
] = sched_group_phys
;
6176 sd
= &per_cpu(phys_domains
, i
);
6177 group
= cpu_to_phys_group(i
);
6179 sd
->span
= nodemask
;
6181 sd
->groups
= &sched_group_phys
[group
];
6183 #ifdef CONFIG_SCHED_MC
6184 if (!sched_group_core
) {
6186 = kmalloc(sizeof(struct sched_group
) * NR_CPUS
,
6188 if (!sched_group_core
) {
6189 printk (KERN_WARNING
"Can not alloc core sched"
6193 sched_group_core_bycpu
[i
] = sched_group_core
;
6197 sd
= &per_cpu(core_domains
, i
);
6198 group
= cpu_to_core_group(i
);
6200 sd
->span
= cpu_coregroup_map(i
);
6201 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6203 sd
->groups
= &sched_group_core
[group
];
6206 #ifdef CONFIG_SCHED_SMT
6208 sd
= &per_cpu(cpu_domains
, i
);
6209 group
= cpu_to_cpu_group(i
);
6210 *sd
= SD_SIBLING_INIT
;
6211 sd
->span
= cpu_sibling_map
[i
];
6212 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6214 sd
->groups
= &sched_group_cpus
[group
];
6218 #ifdef CONFIG_SCHED_SMT
6219 /* Set up CPU (sibling) groups */
6220 for_each_cpu_mask(i
, *cpu_map
) {
6221 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6222 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6223 if (i
!= first_cpu(this_sibling_map
))
6226 init_sched_build_groups(sched_group_cpus
, this_sibling_map
,
6231 #ifdef CONFIG_SCHED_MC
6232 /* Set up multi-core groups */
6233 for_each_cpu_mask(i
, *cpu_map
) {
6234 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6235 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6236 if (i
!= first_cpu(this_core_map
))
6238 init_sched_build_groups(sched_group_core
, this_core_map
,
6239 &cpu_to_core_group
);
6244 /* Set up physical groups */
6245 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6246 cpumask_t nodemask
= node_to_cpumask(i
);
6248 cpus_and(nodemask
, nodemask
, *cpu_map
);
6249 if (cpus_empty(nodemask
))
6252 init_sched_build_groups(sched_group_phys
, nodemask
,
6253 &cpu_to_phys_group
);
6257 /* Set up node groups */
6258 if (sched_group_allnodes
)
6259 init_sched_build_groups(sched_group_allnodes
, *cpu_map
,
6260 &cpu_to_allnodes_group
);
6262 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6263 /* Set up node groups */
6264 struct sched_group
*sg
, *prev
;
6265 cpumask_t nodemask
= node_to_cpumask(i
);
6266 cpumask_t domainspan
;
6267 cpumask_t covered
= CPU_MASK_NONE
;
6270 cpus_and(nodemask
, nodemask
, *cpu_map
);
6271 if (cpus_empty(nodemask
)) {
6272 sched_group_nodes
[i
] = NULL
;
6276 domainspan
= sched_domain_node_span(i
);
6277 cpus_and(domainspan
, domainspan
, *cpu_map
);
6279 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6281 printk(KERN_WARNING
"Can not alloc domain group for "
6285 sched_group_nodes
[i
] = sg
;
6286 for_each_cpu_mask(j
, nodemask
) {
6287 struct sched_domain
*sd
;
6288 sd
= &per_cpu(node_domains
, j
);
6292 sg
->cpumask
= nodemask
;
6294 cpus_or(covered
, covered
, nodemask
);
6297 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6298 cpumask_t tmp
, notcovered
;
6299 int n
= (i
+ j
) % MAX_NUMNODES
;
6301 cpus_complement(notcovered
, covered
);
6302 cpus_and(tmp
, notcovered
, *cpu_map
);
6303 cpus_and(tmp
, tmp
, domainspan
);
6304 if (cpus_empty(tmp
))
6307 nodemask
= node_to_cpumask(n
);
6308 cpus_and(tmp
, tmp
, nodemask
);
6309 if (cpus_empty(tmp
))
6312 sg
= kmalloc_node(sizeof(struct sched_group
),
6316 "Can not alloc domain group for node %d\n", j
);
6321 sg
->next
= prev
->next
;
6322 cpus_or(covered
, covered
, tmp
);
6329 /* Calculate CPU power for physical packages and nodes */
6330 #ifdef CONFIG_SCHED_SMT
6331 for_each_cpu_mask(i
, *cpu_map
) {
6332 struct sched_domain
*sd
;
6333 sd
= &per_cpu(cpu_domains
, i
);
6334 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6337 #ifdef CONFIG_SCHED_MC
6338 for_each_cpu_mask(i
, *cpu_map
) {
6340 struct sched_domain
*sd
;
6341 sd
= &per_cpu(core_domains
, i
);
6342 if (sched_smt_power_savings
)
6343 power
= SCHED_LOAD_SCALE
* cpus_weight(sd
->groups
->cpumask
);
6345 power
= SCHED_LOAD_SCALE
+ (cpus_weight(sd
->groups
->cpumask
)-1)
6346 * SCHED_LOAD_SCALE
/ 10;
6347 sd
->groups
->cpu_power
= power
;
6351 for_each_cpu_mask(i
, *cpu_map
) {
6352 struct sched_domain
*sd
;
6353 #ifdef CONFIG_SCHED_MC
6354 sd
= &per_cpu(phys_domains
, i
);
6355 if (i
!= first_cpu(sd
->groups
->cpumask
))
6358 sd
->groups
->cpu_power
= 0;
6359 if (sched_mc_power_savings
|| sched_smt_power_savings
) {
6362 for_each_cpu_mask(j
, sd
->groups
->cpumask
) {
6363 struct sched_domain
*sd1
;
6364 sd1
= &per_cpu(core_domains
, j
);
6366 * for each core we will add once
6367 * to the group in physical domain
6369 if (j
!= first_cpu(sd1
->groups
->cpumask
))
6372 if (sched_smt_power_savings
)
6373 sd
->groups
->cpu_power
+= sd1
->groups
->cpu_power
;
6375 sd
->groups
->cpu_power
+= SCHED_LOAD_SCALE
;
6379 * This has to be < 2 * SCHED_LOAD_SCALE
6380 * Lets keep it SCHED_LOAD_SCALE, so that
6381 * while calculating NUMA group's cpu_power
6383 * numa_group->cpu_power += phys_group->cpu_power;
6385 * See "only add power once for each physical pkg"
6388 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6391 sd
= &per_cpu(phys_domains
, i
);
6392 if (sched_smt_power_savings
)
6393 power
= SCHED_LOAD_SCALE
* cpus_weight(sd
->groups
->cpumask
);
6395 power
= SCHED_LOAD_SCALE
;
6396 sd
->groups
->cpu_power
= power
;
6401 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6402 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6404 init_numa_sched_groups_power(sched_group_allnodes
);
6407 /* Attach the domains */
6408 for_each_cpu_mask(i
, *cpu_map
) {
6409 struct sched_domain
*sd
;
6410 #ifdef CONFIG_SCHED_SMT
6411 sd
= &per_cpu(cpu_domains
, i
);
6412 #elif defined(CONFIG_SCHED_MC)
6413 sd
= &per_cpu(core_domains
, i
);
6415 sd
= &per_cpu(phys_domains
, i
);
6417 cpu_attach_domain(sd
, i
);
6420 * Tune cache-hot values:
6422 calibrate_migration_costs(cpu_map
);
6427 free_sched_groups(cpu_map
);
6431 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6433 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6435 cpumask_t cpu_default_map
;
6439 * Setup mask for cpus without special case scheduling requirements.
6440 * For now this just excludes isolated cpus, but could be used to
6441 * exclude other special cases in the future.
6443 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6445 err
= build_sched_domains(&cpu_default_map
);
6450 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6452 free_sched_groups(cpu_map
);
6456 * Detach sched domains from a group of cpus specified in cpu_map
6457 * These cpus will now be attached to the NULL domain
6459 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6463 for_each_cpu_mask(i
, *cpu_map
)
6464 cpu_attach_domain(NULL
, i
);
6465 synchronize_sched();
6466 arch_destroy_sched_domains(cpu_map
);
6470 * Partition sched domains as specified by the cpumasks below.
6471 * This attaches all cpus from the cpumasks to the NULL domain,
6472 * waits for a RCU quiescent period, recalculates sched
6473 * domain information and then attaches them back to the
6474 * correct sched domains
6475 * Call with hotplug lock held
6477 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6479 cpumask_t change_map
;
6482 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6483 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6484 cpus_or(change_map
, *partition1
, *partition2
);
6486 /* Detach sched domains from all of the affected cpus */
6487 detach_destroy_domains(&change_map
);
6488 if (!cpus_empty(*partition1
))
6489 err
= build_sched_domains(partition1
);
6490 if (!err
&& !cpus_empty(*partition2
))
6491 err
= build_sched_domains(partition2
);
6496 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6497 int arch_reinit_sched_domains(void)
6502 detach_destroy_domains(&cpu_online_map
);
6503 err
= arch_init_sched_domains(&cpu_online_map
);
6504 unlock_cpu_hotplug();
6509 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6513 if (buf
[0] != '0' && buf
[0] != '1')
6517 sched_smt_power_savings
= (buf
[0] == '1');
6519 sched_mc_power_savings
= (buf
[0] == '1');
6521 ret
= arch_reinit_sched_domains();
6523 return ret
? ret
: count
;
6526 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6529 #ifdef CONFIG_SCHED_SMT
6531 err
= sysfs_create_file(&cls
->kset
.kobj
,
6532 &attr_sched_smt_power_savings
.attr
);
6534 #ifdef CONFIG_SCHED_MC
6535 if (!err
&& mc_capable())
6536 err
= sysfs_create_file(&cls
->kset
.kobj
,
6537 &attr_sched_mc_power_savings
.attr
);
6543 #ifdef CONFIG_SCHED_MC
6544 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6546 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6548 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
, const char *buf
, size_t count
)
6550 return sched_power_savings_store(buf
, count
, 0);
6552 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6553 sched_mc_power_savings_store
);
6556 #ifdef CONFIG_SCHED_SMT
6557 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6559 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6561 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
, const char *buf
, size_t count
)
6563 return sched_power_savings_store(buf
, count
, 1);
6565 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6566 sched_smt_power_savings_store
);
6570 #ifdef CONFIG_HOTPLUG_CPU
6572 * Force a reinitialization of the sched domains hierarchy. The domains
6573 * and groups cannot be updated in place without racing with the balancing
6574 * code, so we temporarily attach all running cpus to the NULL domain
6575 * which will prevent rebalancing while the sched domains are recalculated.
6577 static int update_sched_domains(struct notifier_block
*nfb
,
6578 unsigned long action
, void *hcpu
)
6581 case CPU_UP_PREPARE
:
6582 case CPU_DOWN_PREPARE
:
6583 detach_destroy_domains(&cpu_online_map
);
6586 case CPU_UP_CANCELED
:
6587 case CPU_DOWN_FAILED
:
6591 * Fall through and re-initialise the domains.
6598 /* The hotplug lock is already held by cpu_up/cpu_down */
6599 arch_init_sched_domains(&cpu_online_map
);
6605 void __init
sched_init_smp(void)
6608 arch_init_sched_domains(&cpu_online_map
);
6609 unlock_cpu_hotplug();
6610 /* XXX: Theoretical race here - CPU may be hotplugged now */
6611 hotcpu_notifier(update_sched_domains
, 0);
6614 void __init
sched_init_smp(void)
6617 #endif /* CONFIG_SMP */
6619 int in_sched_functions(unsigned long addr
)
6621 /* Linker adds these: start and end of __sched functions */
6622 extern char __sched_text_start
[], __sched_text_end
[];
6623 return in_lock_functions(addr
) ||
6624 (addr
>= (unsigned long)__sched_text_start
6625 && addr
< (unsigned long)__sched_text_end
);
6628 void __init
sched_init(void)
6633 for_each_possible_cpu(i
) {
6634 prio_array_t
*array
;
6637 spin_lock_init(&rq
->lock
);
6639 rq
->active
= rq
->arrays
;
6640 rq
->expired
= rq
->arrays
+ 1;
6641 rq
->best_expired_prio
= MAX_PRIO
;
6645 for (j
= 1; j
< 3; j
++)
6646 rq
->cpu_load
[j
] = 0;
6647 rq
->active_balance
= 0;
6649 rq
->migration_thread
= NULL
;
6650 INIT_LIST_HEAD(&rq
->migration_queue
);
6652 atomic_set(&rq
->nr_iowait
, 0);
6654 for (j
= 0; j
< 2; j
++) {
6655 array
= rq
->arrays
+ j
;
6656 for (k
= 0; k
< MAX_PRIO
; k
++) {
6657 INIT_LIST_HEAD(array
->queue
+ k
);
6658 __clear_bit(k
, array
->bitmap
);
6660 // delimiter for bitsearch
6661 __set_bit(MAX_PRIO
, array
->bitmap
);
6665 set_load_weight(&init_task
);
6667 * The boot idle thread does lazy MMU switching as well:
6669 atomic_inc(&init_mm
.mm_count
);
6670 enter_lazy_tlb(&init_mm
, current
);
6673 * Make us the idle thread. Technically, schedule() should not be
6674 * called from this thread, however somewhere below it might be,
6675 * but because we are the idle thread, we just pick up running again
6676 * when this runqueue becomes "idle".
6678 init_idle(current
, smp_processor_id());
6681 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6682 void __might_sleep(char *file
, int line
)
6684 #if defined(in_atomic)
6685 static unsigned long prev_jiffy
; /* ratelimiting */
6687 if ((in_atomic() || irqs_disabled()) &&
6688 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6689 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6691 prev_jiffy
= jiffies
;
6692 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6693 " context at %s:%d\n", file
, line
);
6694 printk("in_atomic():%d, irqs_disabled():%d\n",
6695 in_atomic(), irqs_disabled());
6700 EXPORT_SYMBOL(__might_sleep
);
6703 #ifdef CONFIG_MAGIC_SYSRQ
6704 void normalize_rt_tasks(void)
6706 struct task_struct
*p
;
6707 prio_array_t
*array
;
6708 unsigned long flags
;
6711 read_lock_irq(&tasklist_lock
);
6712 for_each_process(p
) {
6716 spin_lock_irqsave(&p
->pi_lock
, flags
);
6717 rq
= __task_rq_lock(p
);
6721 deactivate_task(p
, task_rq(p
));
6722 __setscheduler(p
, SCHED_NORMAL
, 0);
6724 __activate_task(p
, task_rq(p
));
6725 resched_task(rq
->curr
);
6728 __task_rq_unlock(rq
);
6729 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6731 read_unlock_irq(&tasklist_lock
);
6734 #endif /* CONFIG_MAGIC_SYSRQ */
6738 * These functions are only useful for the IA64 MCA handling.
6740 * They can only be called when the whole system has been
6741 * stopped - every CPU needs to be quiescent, and no scheduling
6742 * activity can take place. Using them for anything else would
6743 * be a serious bug, and as a result, they aren't even visible
6744 * under any other configuration.
6748 * curr_task - return the current task for a given cpu.
6749 * @cpu: the processor in question.
6751 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6753 task_t
*curr_task(int cpu
)
6755 return cpu_curr(cpu
);
6759 * set_curr_task - set the current task for a given cpu.
6760 * @cpu: the processor in question.
6761 * @p: the task pointer to set.
6763 * Description: This function must only be used when non-maskable interrupts
6764 * are serviced on a separate stack. It allows the architecture to switch the
6765 * notion of the current task on a cpu in a non-blocking manner. This function
6766 * must be called with all CPU's synchronized, and interrupts disabled, the
6767 * and caller must save the original value of the current task (see
6768 * curr_task() above) and restore that value before reenabling interrupts and
6769 * re-starting the system.
6771 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6773 void set_curr_task(int cpu
, task_t
*p
)